diff --git a/ICCV/2025/VideoVAE+_ Large Motion Video Autoencoding with Cross-modal Video VAE/d40a1523-5eb7-42eb-ab5f-29118f57653b_content_list.json b/ICCV/2025/VideoVAE+_ Large Motion Video Autoencoding with Cross-modal Video VAE/d40a1523-5eb7-42eb-ab5f-29118f57653b_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..d7e231a3b685a4f778ea7d598a28d442787f4fce --- /dev/null +++ b/ICCV/2025/VideoVAE+_ Large Motion Video Autoencoding with Cross-modal Video VAE/d40a1523-5eb7-42eb-ab5f-29118f57653b_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:63990649fbdf3589755754c1dd96ce492a78d4c484bfd5d0905a74ebbb30be7b +size 73745 diff --git a/ICCV/2025/VideoVAE+_ Large Motion Video Autoencoding with Cross-modal Video VAE/d40a1523-5eb7-42eb-ab5f-29118f57653b_model.json b/ICCV/2025/VideoVAE+_ Large Motion Video Autoencoding with Cross-modal Video VAE/d40a1523-5eb7-42eb-ab5f-29118f57653b_model.json new file mode 100644 index 0000000000000000000000000000000000000000..a32d045ee28f78d2e406583dd0361386aa89cbbb --- /dev/null +++ b/ICCV/2025/VideoVAE+_ Large Motion Video Autoencoding with Cross-modal Video VAE/d40a1523-5eb7-42eb-ab5f-29118f57653b_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:eb95b894199e02d729e27eb77a9faa2a550c2d38acef61edf2e01babf19b54a8 +size 90311 diff --git a/ICCV/2025/VideoVAE+_ Large Motion Video Autoencoding with Cross-modal Video VAE/d40a1523-5eb7-42eb-ab5f-29118f57653b_origin.pdf b/ICCV/2025/VideoVAE+_ Large Motion Video Autoencoding with Cross-modal Video VAE/d40a1523-5eb7-42eb-ab5f-29118f57653b_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..052fc1c20e9b069c9f737981aefc2f119e56ac2c --- /dev/null +++ b/ICCV/2025/VideoVAE+_ Large Motion Video Autoencoding with Cross-modal Video VAE/d40a1523-5eb7-42eb-ab5f-29118f57653b_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8c0d1ac214b3151c07c84061e95070a90bee0271bf62bfb2e9d4f381daabb98d +size 3870485 diff --git a/ICCV/2025/VideoVAE+_ Large Motion Video Autoencoding with Cross-modal Video VAE/full.md b/ICCV/2025/VideoVAE+_ Large Motion Video Autoencoding with Cross-modal Video VAE/full.md new file mode 100644 index 0000000000000000000000000000000000000000..df9181617e1e21557ecfa5e73af06f3f5480b001 --- /dev/null +++ b/ICCV/2025/VideoVAE+_ Large Motion Video Autoencoding with Cross-modal Video VAE/full.md @@ -0,0 +1,290 @@ +# VideoVAE+: Large Motion Video Autoencoding with Cross-modal Video VAE + +Yazhou Xing* Yang Fei* Yingqing He*† Jingye Chen Jiaxin Xie Xiaowei Chi Qifeng Chen† The Hong Kong University of Science and Technology + +# Abstract + +Learning a robust video Variational Autoencoder (VAE) is essential for reducing video redundancy and facilitating efficient video generation. Directly applying image VAEs to individual frames in isolation results in temporal inconsistencies and fails to compress temporal redundancy effectively. Existing works on Video VAEs compress temporal redundancy but struggle to handle videos with large motion effectively. They suffer from issues such as severe image blur and loss of detail in scenarios with large motion. In this paper, we present a powerful video VAE named VideoVAE+ that effectively reconstructs videos with large motion. First, we investigate two architecture choices and propose our simple yet effective architecture with better spatiotemporal joint modeling performance. Second, we propose to leverage the textual information in existing text-to-video datasets and incorporate text guidance during training. The textual guidance is optional during inference. We find that this design enhances the reconstruction quality and preservation of detail. Finally, our models achieve strong performance compared with various baseline approaches in both general videos and large motion videos, demonstrating its effectiveness on the challenging large motion scenarios. + +# 1. Introduction + +Given the significant attention in the field of video generation, Latent Video Diffusion Models (LVDMs) [4, 5, 14, 37] have emerged as a popular framework. They have been successfully applied to powerful text-to-video models such as VideoCrafter [7, 8], Sora [6], CogVideoX [31] and Hunyuan Video [27]. Different from directly generating video pixels, LVDMs generate latent representations of videos in a compact latent space. This is achieved by first training a Video VAE to encode videos into this latent space. Thus, Video VAE, as a key and fundamental component of LVDMs, has attracted great attention recently. An effective Video VAE can help to reduce the training costs + +of video diffusion models while improving the final quality of the generated videos. Initially, a series of studies adopted the image VAE from Stable Diffusion [21] for video generation tasks, such asAnimateDiff [13], MagicVideo [37], VideoCrafter1 [7], VideoCrafter2 [8], and Align-Your-Latents [5]. However, directly adopting an image VAE and compressing video on a frame-by-frame basis leads to temporal flickering due to the lack of temporal correlation. Additionally, the information redundancy along the temporal dimension is not reduced, leading to low training efficiency for subsequent latent video diffusion models. From the introduction of Sora, which compresses videos both temporally and spatially through a Video VAE, a series of studies have emerged that aim to replicate Sora and train their own Video VAEs, including Open Sora [36], Open Sora Plan [17], CV-VAE [35], CogVideo X [31], Easy animate [30], and Cosmos Tokenizer [20]. Despite their progress in temporal compression, current video VAEs still struggle to handle videos with large motion. They suffer from many problems in large motion scenarios, including motion ghost, low-level temporal flickering, blurring (faces, hands, edges, texts), and motion stuttering (lack of correct temporal transition). + +In this work, we investigate the architecture of Video VAE and propose simple yet effective Video VAE models that have robust reconstruction performance on both general videos and videos with large motion. First, we examine two designs for spatial and temporal compression, including simultaneous spatial-temporal (ST) compression and sequential ST compression. We observed that simultaneous ST compression achieves better low-level temporal smoothness and texture stability, while sequential ST compression achieves better motion recovery, particularly in scenarios of large motion. Thus, we propose our architecture that integrates the advantages of both methods and enables effective video detail and motion reconstruction. + +Second, we integrate extra information from another modality, e.g., text, into the Video VAE models, and propose the first text-guided Video VAE. The motivation is as follows: (1) We observed that the normally used datasets for text-to-video generation contain text-video pairs, thus + +# Latent Channels $= 4$ + +![](images/6c3c17624a1f4be710b91106bd6ab57cd09565688e921576c3c7ac1b2044cae9.jpg) +(0) Ground Truth +(0) +(2) +(3) + +![](images/e64c15d8c9e0f48fb382835f9cd66deda54f8b6e7d0dd88b4d2d67f7fc93b462.jpg) +(1) Open Sora Plan + +![](images/2e4a9367554f96a97254d43933d1443b8ff1d1ceb08c538cc4eea651814c33b5.jpg) +(2) Open Sora + +![](images/93eeaa96c3e520734a4906833e2a86613a7fdd266f69c87535d6563e8f25822d.jpg) + +![](images/4c4115855c6e09d81b3bfb801d7bea0c70a3bbf37e246a0fd051e2b9c81900d5.jpg) +(3) CV-VAE + +![](images/f8ce4d18b07153ffccc9bb9f13e98eb883accc0da2b850b0e33b9b40242c8f22.jpg) +(4) Ours + +# Latent Channels = 16 + +![](images/a577adeddedc3e6a587ea82aaad946eed02f06c297611f3e6e331bdea3e9dcd7.jpg) +(5) Cosmos Tokenizer +(5) +(7) +(8) + +![](images/4ecc7bc5a32efa1a08adccc810eb19f951475be6702f74ac4996d90d569dbc47.jpg) +(6) CV-VAE + +![](images/f83917b5576bbef99644d938bf4d85be5466f96196b66360342f2d86cfc49d2c.jpg) +(7)CogVideo X +(9) Ours + +![](images/393baa49985d71a7c4bf338165fc6bf5cf752f8ebefe09a7b81f18b912ee4e53.jpg) + +![](images/9e16d4d6c0de8db5db8b081f3bd179447beec40f6c56c68de671e8b6cc0662cd.jpg) +(8) EasyAnimate +Figure 1. Our reconstruction results compared with a line of three recent strong baseline approaches. The ground truth frame is (0). Our model significantly outperforms previous methods, especially under large motion scenarios such as people doing sports. + +![](images/797d8eb15daaf95d7d568952bb9136bf3a0281d4c5c1c3b9a6e88a4b9167c8f4.jpg) + +the text descriptions for training VideoVAE models are already existing and easy to obtain. (2) Since video compression is inherently lossy, it is challenging to recover all details from the video latent representations without additional supplementary information. As a result, achieving high-quality reconstruction becomes increasingly difficult at higher compression ratios, as finer video details are more likely to be lost. To this end, we conduct experiment on this design, and we discover that leveraging information from other modalities, e.g., text, enhances video reconstruction performance, particularly improving detail reconstruction and boosting performance under higher compression ratio + +settings. It also helps the model converge during training. + +To evaluate Video VAE models in scenarios with large motion, we construct an additional test set containing videos with large motion. Through extensive experiments and comparisons, we demonstrate the superior performance of our models under various settings. + +In summary, our contributions are as follows: + +- We propose an effective and robust Video VAE, conduct extensive experiments, and achieve state-of-the-art. +- We propose a simple yet effective spatiotemporal modeling architecture for Video VAE. +- We propose the first text-guided video VAE that leverages + +the information from other modalities, i.e., text guidance, and shows the improvement of textual guidance. + +# 2. Related Work + +Video Variational Autoencoder Video Variational Autoencoders (VAEs) [16] can be broadly categorized into discrete and continuous types. Discrete video VAEs compress videos into discrete tokens by learning a codebook for quantization and have achieved state-of-the-art performance in video reconstruction, as demonstrated by models like MAGVIT-v2 [33]. However, these VAEs are not suitable for Latent Video Diffusion Models (LVDMs) [15] due to the lack of necessary gradients for backpropagation, which hinders smooth optimization. + +In contrast, continuous Video VAEs compress videos into continuous latent representations that are widely adopted in LVDMs. In earlier video generation studies, including Stable Video Diffusion [4], the Video VAE was directly adapted from the image VAE used in Stable Diffusion [21], achieving a compression ratio of $1 \times 8 \times 8$ by processing each frame independently. To further reduce the temporal redundancy, more recent studies [17, 30, 31, 35, 36] have trained their VAEs to achieve a more efficient compression ratio of $4 \times 8 \times 8$ . + +Despite these advancements, all of the aforementioned video VAEs struggle with accurately reconstructing videos with large motions due primarily to their limited ability to handle the temporal dimension effectively. A high-quality Video VAE that can robustly reconstruct videos with significant motion is critical in the LVDM pipeline, as it ensures efficient latent space compression, maintains temporal coherence and reduces computational overhead [25]. Without a robust VAE, large motions in videos can lead to poor latent representations, negatively impacting the quality and overall performance of the LVDMs. + +Latent Video Diffusion Models Latent Video Diffusion Models (LVDMs), such as Sora [6], OpenSora [36], Open Sora Plan [17], VideoCrafter1 [7], VideoCrafter2 [8], Latte[19], CogVideoX [31], DynamiCrafter [29] and Vidu [3] have recently achieved significant success in video generation tasks. The general pipeline for these LVDMs consists of two primary steps. First, the raw video is compressed into a latent space via a video Variational Autoencoder (VAE), significantly reducing computational complexity. In the second step, a diffusion model operates within this latent space, learning the desired transformations. The performance of LVDMs is critically dependent on video VAEs, as the quality of the generated video is heavily influenced by the latent space representation and the encoding-decoding capabilities of the VAE. + +In image generation tasks, Stable Diffusion [21] has excelled, largely due to its efficient VAE that reconstructs di + +verse image types with high fidelity. However, no existing VAE in video generation achieves comparable quality, particularly due to challenges in compressing the temporal dimension. This limitation hinders the performance of LVDMs, especially in high-motion scenarios. + +# 3. Method + +# 3.1. Overview + +The video autoencoding problem can be defined as follows. Let $\mathbf{X} \in \mathbb{R}^{C \times T \times H \times W}$ represent a video or image tensor, where $C$ , $T$ , $H$ , and $W$ denote the number of channels, frame(s), height, and width, respectively. We want to train an encoder $\mathcal{E}$ that compresses the input tensor $\mathbf{X}$ into a compact latent representation $\mathbf{Z} \in \mathbb{R}^{C' \times T' \times H' \times W'}$ . The learned compact latent $\mathbf{Z}$ can be further reconstructed back to RGB space with decoder $\mathcal{D}$ : + +$$ +\mathbf {Z} = \mathcal {E} (\mathbf {X}), \hat {\mathbf {X}} = \mathcal {D} (\mathbf {Z}). \tag {1} +$$ + +Our goal is to design and learn such an autoencoder that can reduce the spatial and temporal dimension of video data in latent space and reconstruct the video with highly spatial and temporal fidelity, especially for large-motion scenarios. + +We first examine two inherited video VAE designs from the pre-trained Stable Diffusion model. We then combine the best of two designs and propose our spatiotemporal modeling that can reconstruct high-dynamic contents with fine details. We then investigate the text-conditioned video autoencoding and propose an effective text-guided video VAE architecture. Moreover, we propose a joint image and video compression training method, that enables text-aided joint image and video autoencoding. Our method does not rely on causal convolution as adopted by prior works. Finally, we carefully study the effects of different loss functions on the reconstruction performance and present the state-of-the-art video VAE architecture. + +# 3.2. Towards Better Spatiotemporal Modeling + +Designing a video VAE that is inherited from a pre-trained 2D spatial VAE is a good practice to leverage the spatial compression prior. There are typically two options to inflate a 2D spatial VAE to its 3D video counterpart. + +Option 1: Simultaneous Spatiotemporal Compression One common way to inherit the weight from pre-trained 2D VAE is to inflate the 2D spatial blocks to 3D temporal blocks and simultaneously do the spatial and temporal compression. We first examine this design. Specifically, we replace the 2D convolution in SD VAE with 3D convolution of kernel size (1,3,3), whose weights are initialized from the 2D convolution. Then we add an additional temporal convolution layer with kernel size (3,3,3) to learn spatiotemporal + +![](images/a044eedcb6a1d818746978974732b7d97c50112d27eaadc3d233460caf2c2586.jpg) + +![](images/45c2dd4a39b8d1be74c147d62c973fd9abe7d10631130006edef191111de1ef2.jpg) +Figure 2. We first investigate two preliminary solutions including simultaneous compression and sequential compression. Simultaneous modeling is achieved by inflating pre-trained 2D spatial VAE to 3D VAE, and simultaneously compress both spatial and temporal information. Sequential modeling indicates first compressing the spatial dimension with a spatial-only 2D encoder and then compressing the temporal information with a temporal encoder. The decoding is symmetrical to the encoding architecture. Our solution is firstly compressing on the spatial dimension, with temporal-aware modelling, and then compressing on the temporal dimension, with a temporal encoder. The decoding is symmetrical to the encoding architecture. Additionally, our VAE also benefits from easy-obtained cross-modality information, i.e., textual guidance, to provide more information to reconstruct fine details. + +patterns. In this middle block of the inflated VAE, we inflate the 2D attention to 3D attention and we also include a temporal attention to capture both the spatial and temporal information. We keep other components unchanged to maximally leverage the learned prior of SD VAE. + +# Option 2: Sequential Spatiotemporal Compression + +Another reasonable way to cooperate the SD VAE to video VAE is to keep the SD VAE unchanged: first utilize the SD VAE to compress the input video frame-by-frame, and then learn a temporal autoencoding process to further compress the temporal redundancy, as shown in Fig. 2. Specifically, we adopt a lightweight temporal autoencoder for temporal compression. The encoder consists of one convolutional layer to process the input, and two or three 3D ResNet blocks with convolutional downsampling layers to compress the temporal redundancy. Notably, we design the decoder to be asymmetric as the encoder, i.e., there will be two 3D ResNet blocks following each upsampling layer in the decoder. Through this asymmetric design, our decoder can potentially gain some hallucination ability beyond the + +reconstruction. + +Surprisingly, we find this sequential spatiotemporal design can better compress and recover the dynamic of the input video than option 1, but is not good at recovering spatial details, which is proved by consistent improvement under large-motion video autoencoding as shown in Fig. 4. + +Our Solution We find that simultaneous spatiotemporal compression leads to better detail-recovering capability, and the sequential spatiotemporal compression will exceed at motion-recovering ability. Thus, we propose to combine the best of two worlds, and introduce the two-stage spatiotemporal modeling for video VAE. As the first stage, we inflate the 2D convolution to 3D convolution with kernel size (1,3,3), and similarly to option 1, we add additional temporal convolution layers through 3D convolution. We also include a temporal attention in the middle block of the encoder and decoder, after the spatial cross attention layers. We denote our first-stage model as a temporal-aware spatial autoencoder. Different from option 1, we only compress the spatial information and do not compress the temporal + +![](images/4f7d559c4a5390d40c3a6fcf248bdf84cfda86a43327953e405ab754242d11be.jpg) +Figure 3. The architecture of our temporal-aware spatial autoencoder. We expand the 2D convolution of SD VAE [21] to 3D convolution and append one additional 3D convolution as temporal convolution after the expanded 3D convolution, which forms the STBlock3D. We also inject the cross-attention layers for cross-modal learning with textual conditions. + +information at the first stage, but introduce another temporal encoder to further encode the temporal dimensions, which serves as the second stage compression. We follow the same design of option 2 for our temporal encoder and decoder. After that, we decode the reconstructed latent of the second stage to the RGB space, with the inflated 3D decoder. We jointly train the inflated 3D VAE and the temporal autoencoder. The main idea is illustrated in Fig. 2 and Fig. 3. + +**Formulation** Recall $\mathbf{X} \in \mathbb{R}^{C \times T \times H \times W}$ represent a video, where $C, T, H$ , and $W$ denote the number of channels, frames, height, and width, respectively. The $i$ -th frame of the video is denoted as $\mathbf{x}_i \in \mathbb{R}^{C \times H \times W}$ . The temporal-aware spatial encoder encodes $\mathbf{X}$ into a latent representation $\mathbf{Z}_1 \in \mathbb{R}^{c \times T \times h \times w}$ , where $c$ is the number of latent channels, and $h = \frac{H}{8}$ , $w = \frac{W}{8}$ , as formulated by: + +$$ +\mathbf {Z} _ {1} = \mathcal {E} _ {1} (\mathbf {X}). \tag {2} +$$ + +Next, the temporal autoencoder encodes $\mathbf{Z}_1$ into $\mathbf{Z}_2 \in \mathbb{R}^{c' \times t \times h \times w}$ , where $c'$ is the number of latent channels for $\mathbf{Z}_2$ and $t = \frac{T}{4}$ , as given by: + +$$ +\mathbf {Z} _ {2} = \mathcal {E} _ {2} \left(\mathbf {Z} _ {1}\right). \tag {3} +$$ + +Reconstruction is achieved by decoding $\mathbf{Z}_2$ back into the original video space, $\hat{\mathbf{X}}\in \mathbb{R}^{C\times T\times H\times W}$ , through the following inverse process: + +$$ +\hat {\mathbf {X}} = \mathcal {D} _ {1} \left(\mathcal {D} _ {2} \left(\mathbf {Z} _ {2}\right)\right) = \mathcal {D} _ {1} \left(\mathbf {Z} _ {1}\right). \tag {4} +$$ + +# 3.3. Cross-modal Modeling + +Since textual information is a native component for text-to-video generation datasets, we examine if the textual information can improve the autoencoding process of the model. To achieve that, we split the feature maps into patches as tokens after each ResNet block in the encoder and decoder, and compute the cross attention by taking visual tokens as query (Q) and value (V), the text embeddings as key (K). + +We try to keep the patch size trackable for each layer. Specifically, we use patch size to $8 \times 8$ , $4 \times 4$ , $2 \times 2$ , and $1 \times 1$ for each layer in the temporal-aware spatial autoencoder respectively. We directly use each pixel as one patch in the temporal autoencoder. We adopt LayerNorm as the normalization function. We use Flan-T5 [12] as the text embedder. A projection convolution is applied to the result, which is then added to the input via a residual connection. + +# 3.4. Joint Image and Video Compression + +In contrast to existing architectures such as MagVitV2 [33], OD-VAE [9], and OPS-VAE [36], which use Causalconv3D layers, we rely primarily on standard Conv3D layer. + +A notable feature of our architecture is the ability to mask out the temporal autoencoder, allowing the first-stage model to operate as a standalone image compressor. During training, our model is flexible to take both image and video as input: when the current batch is composed of images, we will disable the temporal convolution and temporal attention layers, as well as the temporal autoencoder. We train our model on both the image dataset and video dataset to let the model learn the image and video compression ability simultaneously. Besides, training on more high-quality images can also help improve the video autoencoding performance. We quantitatively evaluate the performance of our joint image and video compression and compare it with alternative solutions in Table 1. + +# 3.5. Loss Functions + +We use the reconstruction loss, the KL divergence loss, and the video adversarial loss (3D GAN loss) to optimize our model. The reconstruction loss, $\mathcal{L}_{\mathrm{recon}}$ , ensures that the generated frames are perceptually and structurally similar to the input frames. It combines a pixel-wise error term with a perceptual loss, weighted by a hyperparameter. The KL divergence loss, $\mathcal{L}_{\mathrm{KL}}$ , regularizes the latent space by encouraging it to conform to a prior distribution, ensuring smoothness and continuity in the learned latent representations. Given the hierarchical structure of our latent space, we only regularize the innermost latent $\mathbf{Z}_2$ , with dimensions $\frac{T}{4} \times \frac{H}{8} \times \frac{W}{8}$ , where $T$ , $H$ , and $W$ represent the temporal, height, and width dimensions, respectively. The 3D GAN loss, $\mathcal{L}_{\mathrm{GAN}}$ , is introduced to enhance the realism of the generated video sequences, leveraging a discriminator to distinguish between + +![](images/6fa0ba916225e0df3d1174431940505f0f0856e54212f59bd978e3afd8675740.jpg) +Figure 4. Comparisons among simultaneous spatiotemporal modeling, sequential spatiotemporal modeling and our proposed solution. More results can be found in the supplement. + +![](images/7e28eb150bc409215c604672a2a965f65ede55224cfc0485bba3d74566f30310.jpg) + +![](images/49248394d64fe8bf47316ef3cb2af119bb592846b4d43a6a7efb034d9174624c.jpg) + +![](images/cb5e9e39fde8b0ab5fa6c7621bb42bf92dd9e66cb1c835918ecf4ff500ee4d75.jpg) + +real and generated sequences. The total loss function is expressed as: + +$$ +\mathcal {L} _ {\text {t o t a l}} = \mathcal {L} _ {\text {r e c o n}} + \lambda_ {\mathrm {K L}} \mathcal {L} _ {\mathrm {K L}} + \lambda_ {\mathrm {G A N}} \mathcal {L} _ {\mathrm {G A N}}. \tag {5} +$$ + +# 4. Experiments + +# 4.1. Experimental Setup + +Training Details We train our models on three datasets: the public Panda2M [10] with 2.28M videos, MM-Trailer [11] with 3.11M videos, and an internal text-video dataset with 0.48M videos. In total the training video number is 5.87M videos. To start training, we initialize our 4-channel and 16-channel Video VAE models from SD1.4 [21] and SD3.5 [1], respectively. For both models, we enable the video GAN loss after 50K warmup steps. We initially train the 4-channel and 16-channel latent Video VAEs for 230K and 310K steps, respectively. Subsequently, we conduct joint image-video training, empirically using an 8:2 video-to-image ratio to balance video and image reconstruction. The training images are randomly selected from our training video datasets. By masking the temporal dimension and bypassing the temporal autoencoder, we treat these images as independent static frames, allowing the model to learn from both temporal and spatial information. The 4-channel and 16-channel latent Video VAEs undergo additional joint training for 100K and 185K steps, respectively. For the cross-modal VAE, both models were initialized with their pre-trained weights. We trained the 4-channel cross-modal VAE on video-text pairs for 215,000 steps and the 16-channel variant for 180,000 steps, enabling the model to learn robust alignment between visual and textual modalities. + +Evaluation Datasets We use three test sets for evaluation: the WebVid-10M [2] test set, the Inter4K [22] test set, and a self-constructed large motion test set. Specifically, the WebVid test set contains 1,000 videos. Each video is center cropped and resized into $256 \times 256$ resolution. The Inter4K test set consists of 500 videos with a different resolution setting of $480 \times 864$ for a comprehensive evaluation. The self-constructed large motion test set is used to better assess the model's ability to handle challenging large motion patterns. The construction details are as follows: We filter videos from WebVid-10M test set and Inter4K test set with large motions based on average optical flow scores. 80 videos with top score are selected. The final average optical flow of three evaluation sets are 12.585 (Large motion), 5.905 (WebVid), and 9.1339 (Inter4K), indicating the motion level is aligned with our expectations. We randomly select a clip of 16 consecutive frames from each video with the same frame stride as the training which is 4. The FPS of the evaluated video is the FPS of original video divide by the sampling frame stride. + +Evaluation Metrics We use both pixel-level and perceptual-level video reconstruction metrics including PSNR, SSIM, and LPIPS [34] to quantitatively measure the reconstruction performance of video semantics and details. We also calculate the distribution metrics including FVD and KVD [26], to measure the quality and realism of the reconstructed distributions. Additionally, we conduct a user study based on MOS score to evaluate the temporal smoothness and overall video quality from real users. + +
ModelDownsample Factor#ChannelsWebVid Test Set [2]Inter4K Test Set [22]Large Motion Test Set
PSNR (↑)SSIM (↑)LPIPS (↓)FVD (↓)PSNR (↑)SSIM (↑)LPIPS (↓)FVD (↓)PSNR (↑)SSIM (↑)LPIPS (↓)FVD (↓)
Open-Sora-Plan (OD VAE [9])4x8x8429.16460.83340.078940.061628.66900.83810.090638.396027.56970.80450.1065219.4829
IV-VAE [28]4x8x8428.70000.82530.083238.445728.75990.83980.090628.629327.28380.80140.1053197.0937
Open-Sora (OPS VAE [36])4x8x8429.37530.82840.124090.886329.27210.84310.131655.147027.75860.80320.1540355.5645
CV-VAE [35]4x8x8428.67950.81540.107270.716027.74370.81240.128450.900626.94560.78490.1411293.7747
VidTok [24]4x8x8430.25170.86790.064533.996629.24000.85200.086533.691327.29220.80730.1044193.3726
Video VAE w/o Joint Training (Ours)4x8x8430.20910.86560.056624.298228.90480.85430.068818.123627.39170.80780.0867141.1166
Video VAE (Ours)4x8x8430.31400.86760.053822.439128.92270.85650.066517.686527.62360.81360.0841138.6341
Cross-Modal VAE (Ours)4x8x8430.34860.86790.053018.105829.02950.85850.064915.034527.50090.80920.0820122.4046
Cosmos-Tokenizer [20]4x8x81631.25450.88610.103024.713131.20020.89570.107119.378130.16190.86750.1194109.3825
CogVideoX-VAE [31]4x8x81632.89400.92080.050414.055932.51220.92290.053210.854531.09060.89780.068568.6781
EasyAnimate-VAE [30]4x8x81632.12330.90850.040511.953331.50660.90480.05729.779730.52130.88460.059866.0293
CV-VAE [35]4x8x81632.27660.90800.054616.847831.61290.90600.064215.184630.71360.88680.072691.4076
Open-Sora-Plan (WF-VAE-L [18])4x8x81633.10310.92140.04009.319232.52110.91530.05057.903231.15650.89580.060262.4018
IV-VAE [28]4x8x81633.18870.92270.038110.293033.20880.92700.04088.400231.92770.90860.049359.5411
HunyuanVAE [27]4x8x81633.43810.92710.03238.774433.23390.92720.03729.201032.14410.91180.044551.4671
VidTok [24]4x8x81634.34690.93640.03379.664933.40850.93050.04278.243231.93320.90800.053452.9496
Video VAE w/o Joint Training (Ours)4x8x81633.88440.93340.03448.550332.94160.92970.04098.624631.84710.90730.049946.6997
Video VAE w/ causal conv (Ours)4x8x81633.30310.92330.03829.791133.06450.92660.04159.562332.03130.90870.049450.4350
Video VAE (Ours)4x8x81634.15580.93620.02717.719533.31840.93280.03166.343232.15030.91220.040941.8797
Cross-Modal VAE (Ours)4x8x81634.44300.94030.02716.356033.77600.93710.03156.827832.29080.91740.040843.4941
Video VAE (Ours)8x8x8427.84450.80800.076035.276026.89210.80330.095334.465724.98300.74380.1214217.8798
Cross-Modal VAE (Ours)8x8x8428.58820.82600.071234.469727.30750.81880.088328.938825.79570.76490.1112202.7046
+ +Table 1. Quantitative comparison with state-of-the-art methods. + +
Model#chTemporal Smoothness ↑Overall Quality ↑
Open-Sora-Plan43.843.64
Open-Sora43.162.74
CV-VAE42.842.46
Our VAE44.64.54
Cosmos-Tokenizer163.182.54
CogVideoX-VAE163.863.48
EasyAnimate-VAE164.064.02
CV-VAE163.363.00
Our VAE164.54.4
+ +Table 2. User study based on MOS scores for comparisons with baselines. + +
ModelFactorUCF101WebVid
FVD↓KVD↓FVD↓KVD↓
Vc21x8x8498.7023.38570.4848.45
Vc2 + ours4x8x8479.6920.18529.7537.39
+ +Table 3. Applying our VAE on the video generation task. + +# 4.2. Comparison with State-of-the-arts + +We compare our Video VAE models with the state-of-the-art methods: Open-Sora-Plan [17], Open-Sora [36], CVVAE [35], IV-VAE [28], VidTok [24] on the 4-channel setting, and Cosmos-Tokenizer [20], CogVideoX [31], EasyAnimate [30], CV-VAE [35], HunyuanVAE [27], IV-VAE [28], WF-VAE [18], VidTok [24] on the 16-channel setting. + +Quantitative Evaluation We compare our method with more than 10 baselines on three test sets, as shown in Table 1. Among these, our 4-channel latent Video VAE demonstrates superior performance across most of datasets and metrics. Specifically, our model achieves the best reconstruction quality on the WebVid test set, shown as more than 1dB improvements over baselines and a significant improvement on the LPIPS metrics, which indicates our re + +construction is both with high-fidelity and better perceptual quality. A similar conclusion can be made on the Inter4K test set. On the large motion test set, our model maintains strong performance with a significant PSNR, SSIM, LPIPS and FVD improvement, showcasing its superiority in handling complex motion scenarios. + +For models with 16-channel latent space, our model consistently outperforms these baselines across all test sets. For example, on the WebVid test set, our model achieves more than 2dB in terms of PSNR, significantly higher than Cosmos-Tokenizer and CogVideoX. Moreover, our model achieves the best SSIM and LPIPS, demonstrating substantial improvements in both fidelity and perceptual quality. + +Qualitative Evaluation We provide qualitative comparisons with the baselines in Fig. 1. Our method demonstrates significantly improved motion recovery, greatly reducing ghosting artifacts even in rapid motion scenarios. In contrast, Open-Sora-Plan and CV-VAE struggle to reconstruct fast-moving objects, leading to ghosting artifacts. Additionally, Open-Sora VAE introduces color reconstruction errors, as seen in the clothing of the moving figure. Increasing the latent channels to 16 improves motion reconstruction across all baselines, but noticeable detail errors remain. Our 16-channel model further mitigates these errors, resulting in more accurate detail reconstruction. + +User study We conduct a user study to further compare our method against baselines. We conduct the user study using Mean Opinion Score (MOS) [23]. We randomly sample 10 videos and ask the users to rank the videos from different methods according to two criteria: temporal smoothness and overall video quality. The results are shown in Table 2. Our participants include researchers, engineers, and university students—approximately $50\%$ male and $50\%$ fe + +
ModelPSNR (↑)SSIM (↑)LPIPS (↓)
Simultaneous24.05930.73150.1293
Sequential23.36810.69170.1481
Ours24.67220.72340.1162
+ +Table 4. Ablation study comparing simultaneous modeling, sequential modeling, and ours on the large-motion test set. + +male, mostly aged 20-30. Questionnaires were carefully designed to ensure a clear understanding of evaluation criteria. Our methods obtain the highest MOS on both the temporal smoothness and the overall video quality. + +# 4.3. Compatibility with video generation models + +We investigate the compatibility of our VAE's latent space with video diffusion models. For this assessment, we employ the widely recognized open-source video generation model VideoCrafter2 (VC2) [8]. Specifically, we finetuned VC2 using both our VAE and VC2's original VAE on the UCF dataset. The comparison results measured by Fréchet Video Distance (FVD) and Kernel Video Distance (KVD) are presented in Table 3. Our model achieves superior generation results while maintaining a higher compression rate. Additionally, we finetuned VC2 with our VAE on a private dataset, then evaluated both the modified and original VC2 on 1,000 WebVid videos. These comparative results are shown in Table 3. The findings further validate our VAE's effectiveness in video generation applications. + +# 4.4. Ablation Study + +Spatiotemporal Compression Architecture The visual ablation and quantitative ablation results are shown in Fig. 4 and Table 4. We compare different spatiotemporal compression strategies, including simultaneous and sequential spatiotemporal compression, and our proposed solution. These architecture variants are tested on the large motion test set to determine which model handles challenging scenarios most effectively, as shown in Table 4. + +Cross-modal guidance We provide visual ablation results on the cross-modal training on the 4-channel model, as shown in Fig. 5. The textual guidance can improve the fine-grained detail recovery, such as the text on the man's clothes, and the eye of the woman. The quantitative ablation results are shown in Table 1, from where we can see a clear improvement between "Cross-Modal VAE (Ours)" and "Video VAE (Ours)" under all three difference settings: $4 \times 8 \times 8$ -4 channels, $4 \times 8 \times 8$ -16 channels, and $8 \times 8 \times 8$ -4 channels. Additionally, we analyse this design on the convergence of training. The figure of convergence curve during training is documented in the Section C of the supplementary materials. From the figure we can see that adopting textual guidance can speed up the converge of the model training. + +![](images/ce586868ad69e5a3116545f8cf3a5e6af4c555fcb64eb41b6c15c63ea5d28fc5.jpg) + +![](images/40145049f93f29e67e4068bd4dad90144721d70965ffdcf100df1a3584074934.jpg) +Figure 5. Visual ablation of the proposed cross-modal guidance on the 4-ch models. The fine-grained reconstruction of small regions such as the definition of the text on the cloth, and the woman's eye in the second case are improved. We also visualize the learned attention map which shows the focused semantics of the video. + +Joint Training We evaluate the effectiveness of our image-video joint training on video reconstruction on both 4-channel and 16-channel settings, as shown in Table 1 ("Video VAE w/o Joint Training (Ours)" v.s. "Video VAE (Ours)"). We also compare our joint training strategy with causal convolution-based methods [9, 32, 36] in Table 1 ("Video VAE w/ causal conv (Ours)"). The joint training can further boost the performance of video reconstruction in both the 4-channel and 16-channel experiments. The visual results are documented in the Section A of the supplementary materials. We also show the image reconstruction results and compare our models with SD 1.4 and SD 3.5, as shown in the Section A of the supplementary materials. Our performance is better than the SD 1.4 and almost comparable with SD 3.5. Overall, these results demonstrate that our joint image-video training strategy allows the model to retain strong image reconstruction capability while improve the performance of video reconstruction. + +# 5. Conclusion + +We propose VideoVAE+ to address the challenging problem of large motion video autoencoding. VideoVAE+ introduces two key innovations: an improved spatiotemporal modeling architecture and the incorporation of textual guidance. To enhance reconstruction quality and versatility, we employ joint image-video training. Extensive experiments on various datasets and settings demonstrate the superior performance of our model over state-of-the-art baselines, particularly in challenging large motion scenarios. + +Acknowledgement This research was supported by the Research Grant Council of the Hong Kong Special Administrative Region under grant number 16212623. + +# References + +[1] Stability AI. Stable diffusion 3.5 large, 2023. 6 +[2] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1728-1738, 2021. 6, 7 +[3] Fan Bao, Chengong Xiang, Gang Yue, Guande He, Hongzhou Zhu, Kaiwen Zheng, Min Zhao, Shilong Liu, Yaole Wang, and Jun Zhu. Vidu: a highly consistent, dynamic and skilled text-to-video generator with diffusion models. arXiv preprint arXiv:2405.04233, 2024. 3 +[4] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 1, 3 +[5] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22563-22575, 2023. 1 +[6] Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, Clarence Ng, Ricky Wang, and Aditya Ramesh. Video generation models as world simulators. 2024. 1, 3 +[7] Haoxin Chen, Menghan Xia, Yingqing He, Yong Zhang, Xiaodong Cun, Shaoshu Yang, Jinbo Xing, Yaofang Liu, Qifeng Chen, Xintao Wang, et al. Videocrafter1: Open diffusion models for high-quality video generation. arXiv preprint arXiv:2310.19512, 2023. 1, 3 +[8] Haoxin Chen, Yong Zhang, Xiaodong Cun, Menghan Xia, Xintao Wang, Chao Weng, and Ying Shan. Videocraft2: Overcoming data limitations for high-quality video diffusion models, 2024. 1, 3, 8 +[9] Liuhan Chen, Zongjian Li, Bin Lin, Bin Zhu, Qian Wang, Shenghai Yuan, Xing Zhou, Xinhua Cheng, and Li Yuan. Od-vae: An omni-dimensional video compressor for improving latent video diffusion model, 2024. 5, 7, 8 +[10] Tsai-Shien Chen, Aliaksandr Siarohin, Willi Menapace, Ekaterina Deyneka, Hsiang-wei Chao, Byung Eun Jeon, Yuwei Fang, Hsin-Ying Lee, Jian Ren, Ming-Hsuan Yang, et al. Panda-70m: Captioning 70m videos with multiple cross-modality teachers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13320-13331, 2024. 6 +[11] Xiaowei Chi, Yatian Wang, Aosong Cheng, Pengjun Fang, Zeyue Tian, Yingqing He, Zhaoyang Liu, Xingqun Qi, Jiahao Pan, Rongyu Zhang, Mengfei Li, Ruibin Yuan, Yanbing Jiang, Wei Xue, Wenhan Luo, Qifeng Chen, Shanghang Zhang, Qifeng Liu, and Yike Guo. Mmtrail: A multimodal trailer video dataset with language and music descriptions, 2024. 6 +[12] Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Yunxuan Li, Xuezhi Wang, Mostafa + +Dehghani, Siddhartha Brahma, et al. Scaling instructionfinetuned language models. Journal of Machine Learning Research, 25(70):1-53, 2024. 5 +[13] Yuwei Guo, Ceyuan Yang, Anyi Rao, Zhengyang Liang, Yaohui Wang, Yu Qiao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Animatediff: Animate your personalized text-to-image diffusion models without specific tuning. In The Twelfth International Conference on Learning Representations. 1 +[14] Yingqing He, Tianyu Yang, Yong Zhang, Ying Shan, and Qifeng Chen. Latent video diffusion models for high-fidelity video generation with arbitrary lengths. arXiv preprint arXiv:2211.13221, 2022. 1 +[15] Yingqing He, Tianyu Yang, Yong Zhang, Ying Shan, and Qifeng Chen. Latent video diffusion models for high-fidelity long video generation, 2023. 3 +[16] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. stat, 1050:1, 2014. 3 +[17] PKU-Yuan Lab and Tuzhan AI etc. Open-sora-plan, 2024. 1,3,7 +[18] Zongjian Li, Bin Lin, Yang Ye, Liuhan Chen, Xinhua Cheng, Shenghai Yuan, and Li Yuan. Wf-vae: Enhancing video vae by wavelet-driven energy flow for latent video diffusion model. arXiv preprint arXiv:2411.17459, 2024. 7 +[19] Xin Ma, Yaohui Wang, Gengyun Jia, Xinyuan Chen, Ziwei Liu, Yuan-Fang Li, Cunjian Chen, and Yu Qiao. Latte: Latent diffusion transformer for video generation. arXiv preprint arXiv:2401.03048, 2024. 3 +[20] NVIDIA. Cosmos tokenizer: A suite of image and video neural tokenizers, 2024. 1, 7 +[21] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 3, 5, 6 +[22] Alexandros Stergiou and Ronald Poppe. Adapool: Exponential adaptive pooling for information-retaining downsampling. arXiv preprint, 2021. 6, 7 +[23] Robert C Streijl, Stefan Winkler, and David S Hands. Mean opinion score (mos) revisited: methods and applications, limitations and alternatives. Multimedia Systems, 22(2):213-227, 2016. 7 +[24] Anni Tang, Tianyu He, Junliang Guo, Xinle Cheng, Li Song, and Jiang Bian. Vidtok: A versatile and open-source video tokenizer. arXiv preprint arXiv:2412.13061, 2024. 7 +[25] The Movie Gen team @ Meta. Movie gen: A cast of media foundation models. 2024. 3 +[26] Thomas Unterthiner, Sjoerd Van Steenkiste, Karol Kurach, Raphaël Marinier, Marcin Michalski, and Sylvain Gelly. Fvd: A new metric for video generation. 2019. 6 +[27] Zijian Zhang Rox Min Zuozhuo Dai Jin Zhou Jiangfeng Xiong Xin Li Bo Wu Jianwei Zhang Kathrina Wu Qin Lin Aladdin Wang Andong Wang Changlin Li Duojun Huang Fang Yang Hao Tan Hongmei Wang Jacob Song Jiawang Bai Jianbing Wu Jinbao Xue Joey Wang Junkun Yuan Kai Wang Mengyang Liu Pengyu Li Shuai Li Weiyuan Wang Wenqing Yu Xinchi Deng Yang Li Yanxin Long Yi Chen Yutao Cui + +Yuanbo Peng Zhentao Yu Zhiyu He Zhiyong Xu Zixiang Zhou Zunnan Xu Yangyu Tao Qinglin Lu Songtao Liu Dax Zhou Hongfa Wang Yong Yang Di Wang Yuhong Liu Weijie Kong, Qi Tian and along with Caesar Zhong Jie Jiang. Hunyuanvideo: A systematic framework for large video generative models, 2024. 1, 7 +[28] Pingyu Wu, Kai Zhu, Yu Liu, Liming Zhao, Wei Zhai, Yang Cao, and Zheng-Jun Zha. Improved video vae for latent video diffusion model. arXiv preprint arXiv:2411.06449, 2024. 7 +[29] Jinbo Xing, Menghan Xia, Yong Zhang, Haoxin Chen, Wangbo Yu, Hanyuan Liu, Xintao Wang, Tien-Tsin Wong, and Ying Shan. Dynamiccafter: Animating open-domain images with video diffusion priors, 2023. 3 +[30] Jiaqi Xu, Xinyi Zou, Kunzhe Huang, Yunkuo Chen, Bo Liu, MengLi Cheng, Xing Shi, and Jun Huang. Easyanimate: A high-performance long video generation method based on transformer architecture, 2024. 1, 3, 7 +[31] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, Da Yin, Xiaotao Gu, Yuxuan Zhang, Weihan Wang, Yean Cheng, Ting Liu, Bin Xu, Yuxiao Dong, and Jie Tang. Cogvideox: Text-to-video diffusion models with an expert transformer, 2024. 1, 3, 7 +[32] Lijun Yu, Yong Cheng, Kihyuk Sohn, José Lezama, Han Zhang, Huiwen Chang, Alexander G Hauptmann, Ming-Hsuan Yang, Yuan Hao, Irfan Essa, et al. Magvit: Masked generative video transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10459-10469, 2023. 8 +[33] Lijun Yu, José Lezama, Nitesh B Gundavarapu, Luca Versari, Kihyuk Sohn, David Minnen, Yong Cheng, Agrim Gupta, Xiuye Gu, Alexander G Hauptmann, et al. Language model beats diffusion-tokenizer is key to visual generation. International Conference on Learning Representations, 2024. 3, 5 +[34] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 6 +[35] Sijie Zhao, Yong Zhang, Xiaodong Cun, Shaoshu Yang, Muyao Niu, Xiaoyu Li, Wenbo Hu, and Ying Shan. Cv-vae: A compatible video vae for latent generative video models. arXiv preprint arXiv:2405.20279, 2024. 1, 3, 7 +[36] Zangwei Zheng, Xiangyu Peng, Tianji Yang, Chenhui Shen, Shenggui Li, Hongxin Liu, Yukun Zhou, Tianyi Li, and Yang You. Open-sora: Democratizing efficient video production for all, 2024. 1, 3, 5, 7, 8 +[37] Daquan Zhou, Weimin Wang, Hanshu Yan, Weiwei Lv, Yizhe Zhu, and Jiashi Feng. Magicvideo: Efficient video generation with latent diffusion models. arXiv preprint arXiv:2211.11018, 2022. 1 \ No newline at end of file diff --git a/ICCV/2025/VideoVAE+_ Large Motion Video Autoencoding with Cross-modal Video VAE/images.zip b/ICCV/2025/VideoVAE+_ Large Motion Video Autoencoding with Cross-modal Video VAE/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..0c999c0f0b0f7461000929c9162b8f845b5e5625 --- /dev/null +++ b/ICCV/2025/VideoVAE+_ Large Motion Video Autoencoding with Cross-modal Video VAE/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bd680bd541da48d92484fd1da7cdd0ea97fb0d302d10a2df2e3824de04694967 +size 842225 diff --git a/ICCV/2025/VideoVAE+_ Large Motion Video Autoencoding with Cross-modal Video VAE/layout.json b/ICCV/2025/VideoVAE+_ Large Motion Video Autoencoding with Cross-modal Video VAE/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..6072a7aebe8b2bb730dc529dabb930fdc8255cff --- /dev/null +++ b/ICCV/2025/VideoVAE+_ Large Motion Video Autoencoding with Cross-modal Video VAE/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:50f80334eae0ea241e0fb9081a98aa37a9a1ef5f73cb591771838cd268fee8ef +size 336578 diff --git a/ICCV/2025/ViewSRD_ 3D Visual Grounding via Structured Multi-View Decomposition/28fcbc42-0fe9-4106-81ab-90efedfc5cc9_content_list.json b/ICCV/2025/ViewSRD_ 3D Visual Grounding via Structured Multi-View Decomposition/28fcbc42-0fe9-4106-81ab-90efedfc5cc9_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..a2edaf4703513dafd10c3bf7ca49f0c8d503b3a9 --- /dev/null +++ b/ICCV/2025/ViewSRD_ 3D Visual Grounding via Structured Multi-View Decomposition/28fcbc42-0fe9-4106-81ab-90efedfc5cc9_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:14a61c707d78d75a26849edc3eff2995111679f32b9575efc7cecc84ec5c8d08 +size 85977 diff --git a/ICCV/2025/ViewSRD_ 3D Visual Grounding via Structured Multi-View Decomposition/28fcbc42-0fe9-4106-81ab-90efedfc5cc9_model.json b/ICCV/2025/ViewSRD_ 3D Visual Grounding via Structured Multi-View Decomposition/28fcbc42-0fe9-4106-81ab-90efedfc5cc9_model.json new file mode 100644 index 0000000000000000000000000000000000000000..899b679c50c5326e4261a4ae7b2468520bf351d5 --- /dev/null +++ b/ICCV/2025/ViewSRD_ 3D Visual Grounding via Structured Multi-View Decomposition/28fcbc42-0fe9-4106-81ab-90efedfc5cc9_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:868c1e98f89fea0064bcd1e04751e28459720d84986aa3ac126e8e1a431dece2 +size 109632 diff --git a/ICCV/2025/ViewSRD_ 3D Visual Grounding via Structured Multi-View Decomposition/28fcbc42-0fe9-4106-81ab-90efedfc5cc9_origin.pdf b/ICCV/2025/ViewSRD_ 3D Visual Grounding via Structured Multi-View Decomposition/28fcbc42-0fe9-4106-81ab-90efedfc5cc9_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..9fd5e814a8b2df5b8169a798842573285a7e1d6d --- /dev/null +++ b/ICCV/2025/ViewSRD_ 3D Visual Grounding via Structured Multi-View Decomposition/28fcbc42-0fe9-4106-81ab-90efedfc5cc9_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:16f3b0381766f803ff35abe79da815986bc06e8ccde45294183054b3332ab9bf +size 1523595 diff --git a/ICCV/2025/ViewSRD_ 3D Visual Grounding via Structured Multi-View Decomposition/full.md b/ICCV/2025/ViewSRD_ 3D Visual Grounding via Structured Multi-View Decomposition/full.md new file mode 100644 index 0000000000000000000000000000000000000000..9bc925f3cabf6ba04f89b0b3bf85009b30d67ff3 --- /dev/null +++ b/ICCV/2025/ViewSRD_ 3D Visual Grounding via Structured Multi-View Decomposition/full.md @@ -0,0 +1,358 @@ +# ViewSRD: 3D Visual Grounding via Structured Multi-View Decomposition + +Ronggang Huang $^{1*}$ , Haoxin Yang $^{1*†}$ , Yan Cai $^{1}$ , Xuemiao Xu $^{12345†}$ , Huaidong Zhang $^{1}$ , Shengfeng He $^{6}$ + +$^{1}$ South China University of Technology $^{2}$ Guangdong Engineering Center for Large Model and GenAI Technology + +$^{3}$ State Key Laboratory of Subtropical Building and Urban Science + +4 Ministry of Education Key Laboratory of Big Data and Intelligent Robot + +5 Guangdong Provincial Key Lab of Computational Intelligence and Cyberspace Information + +$^{6}$ Singapore Management University + +# Abstract + +3D visual grounding aims to identify and localize objects in a 3D space based on textual descriptions. However, existing methods struggle with disentangling targets from anchors in complex multi-anchor queries and resolving inconsistencies in spatial descriptions caused by perspective variations. To tackle these challenges, we propose ViewSRD, a framework that formulates 3D visual grounding as a structured multi-view decomposition process. First, the Simple Relation Decoupling (SRD) module restructures complex multi-anchor queries into a set of targeted single-anchor statements, generating a structured set of perspective-aware descriptions that clarify positional relationships. These decomposed representations serve as the foundation for the Multi-view Textual-Scene Interaction (Multi-TSI) module, which integrates textual and scene features across multiple viewpoints using shared, Cross-modal Consistent View Tokens (CCVTs) to preserve spatial correlations. Finally, a Textual-Scene Reasoning module synthesizes multi-view predictions into a unified and robust 3D visual grounding. Experiments on 3D visual grounding datasets show that ViewSRD significantly outperforms state-of-the-art methods, particularly in complex queries requiring precise spatial differentiation. Code is available at https://github.com/visualjason/ViewSRD. + +# 1. Introduction + +3D Visual Grounding (3DVG) aims to establish semantic correspondences between natural language descriptions and target objects in a 3D space [19, 44]. This task has gained significant attention in applications such as visual language navigation [25, 57], intelligent agents [4, 49], and autonomous vehicles [9, 13]. + +![](images/946d478baee8357da051cab472b739f69fde08a54f1f66f44e3dd17a68c57079.jpg) +Figure 1. (a) Previous 3DVG methods struggle with ambiguities from complex multi-anchor queries and perspective shifts. (b) ViewSRD addresses this by using the SRD module to simplify queries and the CCVTs to capture viewpoint variations in both scene and textual modal, boosting cross-modal feature interaction and enhancing grounding accuracy. + +Traditional single-view approaches rely on 2D sampled images to extract scene information [38] or construct scene graphs from textual descriptions [42]. However, these methods are inherently limited by their dependence on single-view cues, as language descriptions often presuppose specific viewpoints. To overcome this limitation, recent research has explored multi-view 3DVG, integrating multiple perspectives to enhance robustness [8, 31, 58]. Some methods process distinct descriptions for different viewpoints via manual annotation and learning [15, 37], while others incorporate spatial modules to encode relative spatial coordinates under specific perspectives [5]. However, they typically address only isolated aspects of the problem, limiting their effectiveness in handling complex multi-view scenarios. + +Despite their potential, existing 3DVG models struggle to disentangle targets from anchors in multi-anchor textual descriptions [5, 14, 32]. Large language models (LLMs) often have difficulty interpreting such descriptions [17, 51], yet resolving these ambiguities is crucial for improving grounding accuracy [20]. Compounding this challenge, inconsist- + +cies between textual descriptions and spatial relationships arise when viewpoints change. As illustrated in Fig. 1, an object described as being to the right of another—such as "The nightstand is to the right of the bed"—from a front-facing view may appear on the left when observed from the opposite direction. These perspective-induced inconsistencies make it significantly harder for models to establish accurate correspondences between textual descriptions and visual information, further degrading performance. Ultimately, both the inherent complexity of multi-anchor queries and the challenges introduced by perspective shifts hinder the accurate interpretation of positional relationships in 3DVG, limiting the overall effectiveness of existing systems. + +To tackle these challenges, we propose ViewSRD, a framework that formulates 3D visual grounding as a structured multi-view decomposition process. By leveraging the Simple Relation Decoupling (SRD) module, ViewSRD effectively disentangles target-anchor relationships in the complex multi-anchor queries, while the Multi-view Textual-Scene Interaction (Multi-TSI) module integrates multi-view information to enhance grounding accuracy. As illustrated in Fig. 1(b), ViewSRD first applies the SRD module to decompose complex multi-anchor queries into a set of simpler single-anchor queries, isolating interactions between the target and its anchors. This structured decomposition allows the model to more effectively learn positional relationships from textual descriptions. The Multi-TSI module then fuses textual and scene features across multiple viewpoints using Cross-modal Consistent View Tokens (CCVTs), which explicitly encode viewpoint information as learnable cue for both textual and scene module. This mechanism ensures that the model accurately captures spatial interactions, even under perspective shifts. Finally, the Textual-Scene Reasoning module aggregates these multi-view features to accurately predict the final 3D VG results. Extensive experiments have validated the efficacy of our proposed ViewSRD across different 3DVG benchmarks, demonstrating its superior performance across diverse scenarios. In summary, our contributions are fourfold: + +- We propose ViewSRD, a framework that formulates 3D visual grounding as a structured multi-view decomposition process, effectively handling complex multi-anchor queries and mitigating text-visual inconsistencies across different perspectives. +- We introduce the Simple Relation Decoupling (SRD) module, which restructures complex multi-anchor queries into simpler single-anchor statements, disentangling target-anchor relationships. This structured decomposition enables the model to extract more effective textual features for grounding. +- We develop the Multi-view Textual-Scene Interaction (Multi-TSI) module to explicitly encode viewpoint infor + +mation using cross-modal consistent view tokens. This mechanism ensures alignment between textual descriptions and visual features across different perspectives, reducing spatial ambiguities. + +- We conduct extensive evaluations on 3D visual grounding datasets, where ViewSRD achieves state-of-the-art performance, yielding superior performance over prior work. + +# 2. Related Work + +3D Visual Grounding. 3D computer vision has made great progress in various fields [7, 26, 29, 33, 45, 55, 56], the 3D visual grounding (3DVG) task involves identifying a target object in a 3D scene based on a natural language description [19, 44]. Pioneering datasets such as ScanRefer [6] and ReferIt3D [1], built on ScanNet [10], have driven progress in this field. Recent advancements like MVT [18] address view inconsistency by developing a view-robust multi-modal representation. Other works [23, 27, 47, 53] explore multimodal situated reasoning but lack a dedicated focus on handling the high semantic complexity of natural language in 3D grounding, particularly in disentangling intricate sentence structures. + +Despite these advancements, the complexity of natural language descriptions remains a significant challenge in grounding tasks. Referring expressions often require reasoning over multiple anchor objects to precisely identify the target, making it crucial to disentangle and interpret intricate linguistic structures and spatial dependencies. Our method addresses this challenge by decoupling complex queries into simpler statements, improving the extraction of key relational information. Additionally, by leveraging view tokens, ViewSRD learns more accurate associations between textual descriptions and multi-view information. + +Language Comprehension. Understanding referential language in 3DVG requires models to not only parse spatial descriptions but also interpret object relationships within a scene. Scene graphs, where objects serve as nodes and relationships form directed edges, have been widely used for tasks such as image retrieval and caption evaluation [19]. Traditional approaches employ scene graphs to enhance query comprehension [42], with efforts to convert sentences into structured representations [12, 34, 43] or generate grounded scene graphs for images [22, 24, 50]. However, these methods primarily focus on static, well-defined relationships and struggle with the dynamic, context-dependent nature of natural language. In datasets such as $\mathrm{Nr3D}$ [1], the complexity of interwoven spatial relationships and ambiguous references makes direct scene graph construction challenging. To address this, we propose leveraging Large Language Models (LLMs) [28, 40, 46] to enhance semantic understanding and spatial reasoning, reducing reliance on rigid structures while improving language comprehension. + +3D Multi-View Learning. 3D vision research has largely + +![](images/e37681de2108e9a0f8c5bbec750f4eef8488e8483ad3a69ccda17c0c0dc2f3dc.jpg) +Figure 2. Overview of ViewSRD. We begin by employing the Simple Relation Decoupling (SRD) module to decompose complex multi-anchor queries into multiple simpler single-anchor queries. Next, text and scene features are extracted separately using the text encoder and scene encoder. To explicitly incorporate scene information into the model, we fuse Cross-modal Consistent View Tokens (CCVTs) with these extracted features. The Multi-view Textual-Scene Interaction (Multi-TSI) module then facilitates a comprehensive interaction between textual and scene information, the 3DVG prediction results are finally generated by the Textual-Scene Reasoning Module. + +focused on generating 2D projections from multiple viewpoints. While LLM-based grounding methods integrate multi-view images, they struggle with accurately identifying the primary viewpoint and demonstrating reliability, as discussed in [17, 51]. MVT [18] maps 3D scenes into multiple perspectives to enhance cross-view feature aggregation but lacks a mechanism to weigh each view's contribution, limiting performance in complex scenes. Similarly, ViewRefer [15] utilizes multi-view prototypes for cross-view interactions but lacks explicit training guidance on view importance. Mikasa [5] incorporates relative spatial coordinate information and a scene-aware module to improve object grounding but does not fully resolve view weighting challenges. In contrast, we propose Cross-modal Consistent View Tokens, which guide the model to dynamically adjust representation spaces and assess whether spatial relationships in decoupled sentences exhibit view dependency. This mechanism enables more reliable multi-view reasoning, improving performance in complex scenes. + +# 3. ViewSRD + +In the context of 3D point cloud scenes, the term multiview refers to observing a shared scene representation (e.g., XYZ+RGB format) from different simulated viewpoints by rotating the scene around its central axis or camera viewpoints. Each view provides a partial observation of the same 3D environment, resulting in varying object appearances, + +occlusions, and spatial configurations across views. This multi-view setup introduces significant challenges for 3D visual grounding: (1) language-grounded spatial relations must remain consistent across view-dependent variations, and (2) object referents may be partially or completely invisible in certain views. + +To tackle these challenges, we propose ViewSRD, a structured multi-view 3D visual grounding framework. The overall framework of our method is illustrated in Fig. 2. ViewSRD comprises two key components. The first component is the Simple Relation Decoupling (SRD) module, which decomposes multi-anchor queries into a series of single-anchor queries by leveraging the powerful language processing capabilities of LLMs and predefined prompt templates. This decomposition enables more precise inference of relative relationships between objects, improving the model's ability to capture spatial interactions. The second is the Multi-view Textual-Scene Interaction (Multi-TSI) module, which mitigates viewpoint dependency by integrating a shared, cross-modal consistent view token into both the language and visual models. These tokens facilitate feature interaction across perspectives, allowing the visual and textual models to align cross-modal viewpoint information more effectively. + +# 3.1. Simple Relation Decoupling Module + +The Simple Relation Decoupling (SRD) module is designed to structurally decompose a multi-anchor query into mul + +![](images/77876b31e223121186458180f7215c934c21964215ab7e2d277a674b46fc75f7.jpg) +Figure 3. Overview of the SRD Module pipeline. + +tiple simpler single-anchor queries, enhancing the text encoder's ability to comprehend and process relational information. As illustrated in Fig. 3, the SRD module first predicts the target and anchor labels within a sentence, assigning them as the subject and object in the simplified sentence, respectively. This restructuring forms the foundation for generating a structured prompt, which is then fed into an LLM to produce a set of simplified queries. To maintain semantic integrity, we employ a Sentence Matching algorithm which is described in detail in the supplementary material. that filters and retains the most relevant simplified queries, ensuring that the refined queries faithfully preserve the original meaning while improving clarity and interpretability. By disentangling object relationships between the target and multiple anchors, the SRD module enables more precise relational reasoning, enhancing 3DVG performance. + +Target and anchors digging. We pre-train a classifier $Clas$ to identify the word in a sentence that corresponds to the {Target} object. Given an input sentence, $Clas$ first determines which word belongs to the target. Subsequently, we assess whether other words in the sentence appear in the predefined anchor set, $A_{lab} = \{A_{lab1}, A_{lab2}, \ldots\}$ , provided by the dataset. If a word matches an entry in this set, it is classified as an {Anchor} object. For more details about the $Clas$ , please refer to the supplementary materials. + +Decoupled multi-anchor queries. In practical referring queries, multiple anchors frequently co-occur within the same sentence, and the spatial relationship of the target is inherently tied to the anchor labels. In such cases, spatial + +descriptions involving multiple objects and their attributes often become entangled. For instance, in the $\{\text{Query}\}$ illustrated in Fig. 3, the object "couch", which is near the target "pillow", may dominate the spatial description, thereby weakening the relationship between the target and other anchors. To address the coupling issue in such queries, we design a set of prompt templates based on prior target and anchor digging, the process is shown in Fig. 3. Leveraging the reasoning capabilities of LLMs, we decompose complex multi-anchor queries into simpler single-anchor queries. This decoupling process clarifies the positional relationships between the target and its anchors in 3DVG, enhancing the model's spatial understanding. + +We define a total of $k$ structured $\{Example\}$ derived from our pre-designed templates, such as "The target is on the anchor", with additional examples provided in the supplementary materials. For each anchor, the model generates $k$ candidate queries, where $k$ denotes the number of generated examples. To ensure the selected sentence best aligns with the original query, we apply a sentence-matching algorithm that evaluates both label consistency and semantic consistency. The final ranking is determined by a weighted average of these two scores. For further details, please refer to the supplementary materials. + +# 3.2. Textual Aggregation + +Given a complex multi-anchor query sentence, the SRD module decomposes it into $(I + 1)$ sentences, where $I$ represents the number of anchors in the original sentence. Each anchor contributes to a shorter, simplified sentence, while the original complex query remains as a longer reference sentence. To extract meaningful linguistic representations, we employ BERT [11] as the text encoder to extract $(I + 1)$ sentence features, generating a language feature set $\{F_{T_0}, F_{T_1}, \ldots, F_{T_I}\}$ , where $F_{T_0}$ corresponds to the original complex query, and the remaining elements represent the features of the decoupled simpler queries. To enable the model to effectively learn from diverse sentence representations, we introduce a textual feature aggregation strategy. We randomly sample one feature from the language feature set as the main feature $F_{\mathrm{main}}$ , while treating the remaining features as auxiliary features $F_{\mathrm{aux}}$ . The final aggregated feature is computed as: + +$$ +F _ {\text {a g g}} = \alpha F _ {\text {m i n}} + (1 - \alpha) \cdot \frac {1}{I} \sum_ {i = 1} ^ {I} F _ {\text {a u x} _ {i}}, \tag {1} +$$ + +where $\alpha$ is uniformly sampled from $\{0, 0.1, 0.3, 0.5\}$ during training and fixed at 0.5 during validation. This adaptive fusion strategy ensures smooth feature integration, enhancing the model's robustness in language-conditioned 3DVG. + +# 3.3. Multi-view Textual-Scene Interaction Module + +Cross-modal Consistent View Tokens. Previous methods have largely overlooked the inconsistency in textual descriptions arising from perspective shifts in multi-view VG, making it challenging for models to accurately interpret these variations [3, 41]. To address this limitation, we introduce a series of learnable and shared Cross-modal Consistent View Tokens(CCVTs), which are integrated into both the textual and scene modules. By incorporating these tokens, both models are explicitly guided with relevant perspective information, enabling them to more effectively capture and understand the transformations and interactions induced by viewpoint changes. + +Formally, let $\mathcal{V} = \{V_n | n = 1, 2, \dots, N; V_n \in \mathbb{R}^D\}$ represents the set of CCVTs, where $N$ denotes the number of viewpoints and $D$ represents the dimensionality of CCVTs. The CCVTs are jointly optimized with our proposed textual and scene modules. Once trained, their values remain fixed during inference, serving as a stable reference that enhances the model's ability to comprehend multi-view scenarios and resolve perspective-induced inconsistencies. + +Multi-view Textual Module. To effectively integrate sentence features from text encoders with viewpoint features extracted from CCVTs, we introduce the Multi-view Textual Module, which employs a cross-attention mechanism [39] to seamlessly encode viewpoint features $\mathcal{V}$ into the textual feature space through multi-head attention operation. + +Since each sentence inherently carries distinct viewpoint information, it is crucial to embed perspective-aware features into textual representations effectively. To achieve this, we first compute the normalized dot product between each view token and the 0th token of each sentence's language feature $\{F_{T_0}, F_{T_1}, \ldots, F_{T_I}\}$ , as the 0th token $F^0$ typically aggregates the most salient semantic information. We take the average of these dot products across different sentences and compute a corresponding probability distribution using the softmax function. This probability is then used to reweight the view token, adaptively increasing its contribution when the description aligns with the viewpoint and reducing it when the description does not match. The refined viewpoint token is formulated as: + +$$ +\mathcal {V} = \operatorname {S o f t m a x} \left(\frac {1}{I} \sum_ {i = 0} ^ {I} \frac {F _ {T _ {i}} ^ {0} \mathcal {V} ^ {T}}{\| F _ {T _ {i}} ^ {0} \| \cdot \| \mathcal {V} \|}\right) \mathcal {V}. \tag {2} +$$ + +Subsequently, the aggregated features $F_{agg}$ , as introduced in Section 3.2, serve as the query, while the viewpoint features $\mathcal{V}$ act as the key and value in the attention computation. The textual feature enriched with viewpoint embeddings, denoted as $F_q'$ , is formulated as: + +$$ +F _ {q} ^ {\prime} = \operatorname {S o f t m a x} \left(\frac {\left(W _ {q} F _ {a g g}\right) \left(W _ {k} \mathcal {V}\right) ^ {T}}{\sqrt {D}}\right) W _ {v} \mathcal {V}, \tag {3} +$$ + +where $W_{q}$ , $W_{k}$ , and $W_{\nu}$ are learnable linear projection matrices. Following this, $F_{q}^{\prime}$ undergoes an additional self-attention operation to further refine the textual features, ensuring that the encoded representations effectively capture perspective-dependent information. + +Multi-view Scene Module. To effectively capture object features across diverse scenes, we introduce a Multi-View Scene Module that extracts and refines scene representations from multiple viewpoints. To achieve this, we employ PointNet++ [35] as the scene encoder, computing scene features $F_{V_n}$ for each viewpoint, where $n \in N$ denotes the scene index across $N$ viewpoints. Each scene feature $F_{V_n}$ consists of object-level representations, expressed as $\{F_{o1}, F_{o2}, \dots, F_{oi}\}$ , where $i$ corresponds to the number of objects present in the scene. + +To explicitly inform the model of the current scene, we concatenate our CCVTs $V_{n}$ with the extracted scene features, forming the input representation: + +$$ +\mathbf {X} _ {n} = \left\{F _ {V _ {n}}, V _ {n} \right\} = \left\{F _ {o 1}, F _ {o 2}, \dots , F _ {o i}, V _ {n} \right\}. \tag {4} +$$ + +These combined feature representations are then processed through several Transformer layers [39], denoted as $\mathrm{Trans}(\cdot)$ , which enhances the relational encoding between objects and viewpoints. This mechanism ensures that both global scene context and fine-grained object details are effectively captured: + +$$ +\mathbf {Z} _ {n} ^ {(l + 1)} = \operatorname {T r a n s} ^ {(l)} \left(\mathbf {Z} _ {n} ^ {(l)}\right), \tag {5} +$$ + +where the initial input to the Transformer is $\mathbf{Z}_n^{(0)} = \mathbf{X}_n$ , and $\mathbf{Z}_n^{(L)}$ represents the refined features after $L$ Transformer layers. + +At the final Transformer layer, the output consists of both [object] tokens and [view] tokens. Since the transformed features $F_{V_n}'$ encapsulate both object-specific and viewpoint information, we retain only the [object] tokens for the subsequent grounding task: + +$$ +F _ {V _ {n}} ^ {\prime} = \left\{F _ {o 1} ^ {\prime}, F _ {o 2} ^ {\prime}, \dots , F _ {o i} ^ {\prime} \right\}. \tag {6} +$$ + +This design ensures that object representations are enriched with multi-view contextual information while maintaining their distinct semantic properties for accurate 3DVG. + +# 3.4. Textual-Scene Reasoning Module + +With the above formulation, we obtain the view-interactive textual features $F_{q}^{\prime}$ and scene features $F_{V}^{\prime}$ , where $F_{V}^{\prime} = \{F_{V_{n}}^{\prime} \mid n = 1,2,\dots,N\}$ , each enriched with viewpoint information. These features are then processed through the proposed Textual-Scene Reasoning Module to generate the final prediction. This module primarily consists of a Transformer with a cross-attention mechanism [39], where $F_{V}^{\prime}$ serves as the query, while $F_{q}^{\prime}$ functions as the key and + +![](images/eab320a4aa093e1a0a09eb54348963095a829c995355ccfeda5b59a77e1aa50e.jpg) +Ground Truth + +![](images/5e1275eb13913bf36b808131b4383c19e25874041f10d0499bc4543c1ebed529.jpg) +incorrect predictions +MVT + +![](images/b3544930bfe2050df8eb84b270fb226a2830f396252a9b0455fe4e54ca2a5c81.jpg) +correct predictions + +![](images/5fb571b271ea56608fd7c7c6f62d0d3bb2faf8118e6a911463d938c27d9d1c1a.jpg) +Mentioned Object + +![](images/c149c4de3a8c338b77009e2de4bc646b7704f071a6738cc18d6d896b22d30dbd.jpg) +Ground Truth + +![](images/b49289c54111f079b81ab5b8963f6f7120532f80730b0fbe8865d3312ed86137.jpg) +Complex multi-anchor query: Find the chair that is at the desk with a computer monitor on it. + +![](images/95cb992526d27d3664042219440eaa408c6038157d9f13a432871fad19da47e9.jpg) +CoT3DRef + +![](images/f5f5f0d2610bf5ecbc6f2630be133044fb29858f4af05dad6193523d8c897efb.jpg) +ViewSRD + +![](images/f7c22f04be7dbb0aeaff017a1870c824d5f5e00d4038cacaf53277ef875edc60.jpg) +Simple Queris +The chair is with monitor on it. + +![](images/530ed46376e10e77852b4c228f60b9c9155e8c19ee09b48476902b3e0a29fb30.jpg) +The chair is at the desk. + +![](images/11f835dbb2f1d03af174defc265118f3fbcfe6a78a0161aadc407577c4fc29ad.jpg) +Complex multi-anchor query: There are 2 trash cans under the sink. Select the smaller one on the right. + +![](images/065bb2a6f238b1db2299fd4590de1198da22b9f6679cac2bed2b93d40b1ca310.jpg) + +![](images/7a445e61e35c1244f29b91bab6f708e8005259ba4817be3f50b66089fccac252.jpg) + +![](images/6c0d4b646259957f37bb2a501ddb6608a95684c85f41c2b89ad6438dccac49d9.jpg) + +![](images/6006126a568261dc266c962781defe6534b40da65d487b4d87df9c38592960e1.jpg) +The trash cans are under the sink. + +![](images/12bbc812fbb669456efeb24d4dc777a16d2ea5488a8f7e902a61212f7d9b2599.jpg) +The target is the smaller trash can on the right. + +![](images/8170dc158b7d08c3685aa3ee25b7206c7b3655b7d6d666d875c00ef82ef6cf7a.jpg) +Complex multi-anchor query: +(a) +Figure 4. Visualization Results of the 3D Visual Grounding Results. For the presented 3D scenes, we utilize green, red, blue, and yellow boxes to represent the ground truth, incorrect predictions, correct predictions, and the mentioned objects, respectively. Columns (e) and (f) present the decomposed simple queries derived from the complex queries. + +![](images/a44851b1ae326977c23cf76679dae8435ede9c0e6e9f8346a3c242e36dbece2c.jpg) +The nightstand to the right of +(b) + +![](images/56b449e314adaa9d4b9d19d7cd620f1a263bdad65604d60fc6a2c101f2dc7405.jpg) +of the bed, closest to the wall with the framed picture. +(c) + +![](images/32a87899ed66c4fa58909669e57355887b7f6484ac9284eae8fecd4b69e93333.jpg) +(d) + +![](images/879be34d395ac7f61c54a22ee7f75b9ae45ad4f5554cb46d17b49cf2ef5e3a08.jpg) +The nightstand is to the right of the bed. +(e) + +![](images/7e8945f5de11d68a090c92b9585b0a6d4e7d11e1fe63e5cef509be786cdc58f0.jpg) +The nightstand is closest to the wall with the framed picture. +(f) + +value, facilitating fine-grained alignment between textual and visual representations. Additionally, a View Aggregation mechanism integrates information across multiple viewpoints by computing both the average and maximum values of the output features. Finally, a Prediction Head projects the aggregated features into the result space, enabling a view-aware 3D Visual Grounding model capable of effectively reasoning across multiple perspectives. + +# 3.5. Overall Loss Functions + +Following prior research [18, 36], three distinct loss functions are applied on ViewSRD. These include a referential loss $\mathcal{L}_{Ref}$ derived from grounding predictions, an object-level loss $\mathcal{L}_{Object}$ capturing object shape and center and a sentence-level loss $\mathcal{L}_{Sent}$ designed to identify the target and anchor phrases within the $F_{\mathrm{agg}}$ . Similarly to [2], we extend the referential loss $\mathcal{L}_{Ref}$ to localize both the target and the anchors, which we term as parallel referential loss $\mathcal{L}_{\mathrm{ref}}^P$ , where both the target and anchors are localized simultaneously. For details of these losses, please refer to supplementary materials. The total loss function is defined as: + +$$ +\mathcal {L} = \lambda_ {O b j} \mathcal {L} _ {O b j e c t} + \lambda_ {R e f} \mathcal {L} _ {R e f} ^ {P} + \lambda_ {S e n t} \mathcal {L} _ {S e n t}. \tag {7} +$$ + +# 4. Experiments + +# 4.1. Experiment Settings + +Datasets. Nr3D [1] contains 45,503 human utterances referencing 707 indoor scenes from ScanNet [10], cover + +ing 76 object categories with multiple same-class distractors. Sr3D [1] includes 83,572 template-based sentences in a "target-spatial relation-anchor" format, offering a simpler setup with similar distractors. ScanRefer [6] provides 51,583 free-form descriptions for 11,046 objects across 800 ScanNet scenes, incorporating spatial and attribute-level references to support 3DVG. + +Evaluation Metrics. For $\mathrm{Nr3D}$ and $\mathrm{Sr3D}$ , grounding accuracy is measured by the percentage of correctly matched boxes [18, 36]. For ScanRefer, we report Acc@0.25 and Acc@0.5, i.e., the percentage of predicted boxes with IoU exceeding 0.25 or 0.5, respectively [42]. + +Implementation Details. All experiments are implemented in PyTorch and run on a single RTX 4090 GPU. We use AdamW [30] with a learning rate of 0.0005. The number of input views is set to $N = 4$ . We set $\lambda_{Ref}, \lambda_{Obj}, \lambda_{Sent} = 1.0, 0.5, 0.5$ . For the SRD module, we adopt DeepSeekR1 [28], which balances performance and reproducibility, and can runs on a RTX 4090. + +# 4.2.3D Visual Grounding Results + +We compare ViewSRD with recent state-of-the-art approaches to evaluate its effectiveness on 3DVG. Fig. 4 illustrates complex query cases from $\mathrm{Nr3D}$ [1], including ground truth boxes, predictions from MVT [18], CoT3DRef [2], and ViewSRD, along with the original queries and the simplified sentences produced by the SRD module. In multi-anchor scenarios (e.g., involving "bed", "table", and "chair"), MVT and CoT3DRef often misalign predictions + +Table 1. Performance $\left( \% \right)$ comparison on $\mathrm{{Nr3D}}\left\lbrack 1\right\rbrack$ and Sr3D [1]. + +
MethodNr3DSr3D
OverallEasyHardView Dep.View Indep.OverallEasyHardView Dep.View Indep.
3DVG-Transformer [54]40.848.534.834.843.751.454.244.944.651.7
LanguageRefer [36]43.951.036.641.745.056.058.949.349.256.3
TransRefer3D [16]42.148.536.036.544.957.460.550.249.957.7
SAT [48]49.256.342.446.950.457.961.250.049.258.3
MVT [18]55.161.349.154.355.464.566.958.858.464.7
ViewRefer [15]56.063.049.755.156.867.068.962.152.267.7
MiKASA [5]64.469.759.465.464.075.278.667.370.475.4
CoT3DRef [2]64.470.059.261.965.773.275.267.967.673.5
ViewSRD (ours)69.975.364.868.670.676.078.370.669.076.2
+ +due to challenges in spatial reasoning. In contrast, ViewSRD correctly grounds targets by decomposing complex queries and leveraging robust spatial relationships between target-anchor pairs. Moreover, under viewpoint shifts, CoT3DRef struggles to maintain alignment, whereas ViewSRD reliably grounds targets by capturing spatial relations invariant to viewpoint changes (e.g., "The trash cans are under the sink"). + +Quantitative results on $\mathrm{Nr3D}$ (Table 1) show that ViewSRD achieves a $5.2\%$ accuracy gain over the best prior method, CoT3DRef, under identical settings. Under view-dependent evaluation, it further outperforms CoT3DRef by $6.7\%$ , demonstrating the effectiveness of CCVTs in aligning textual and visual spaces and modeling viewpoint-sensitive relations through query decomposition. To assess generalization, we also evaluate on Sr3D [1] (Table 1). ViewSRD achieves the highest accuracy of $76.2\%$ in the View-Independent setting, with additional gains of $+2.8\%$ and $+2.7\%$ in the View-Independence and Hard scenarios, respectively. These results confirm the robustness and generalizability of our approach across diverse scenario. + +# 4.3. Analysis of Anchors + +In this section, we analyze the impact of the number of anchors in a query on 3DVG performance. The results presented in Table 2 underscore the effectiveness of our approach, particularly in multi-anchor scenarios, where our method successfully disentangles spatial relationships by explicitly modeling target-anchor interactions. In contrast, existing methods such as MVT [18] and CoT3DRef [2], which do not account for the necessity of spatial relationship decoupling, exhibit a notable performance decline in multi-anchor queries compared to single-anchor cases. Notably, our approach achieves higher accuracy in multi-anchor queries than in single-anchor ones, demonstrating that when properly processed, multi-anchor information enhances 3DVG performance rather than introducing ambiguity. These findings validate the efficacy of ViewSRD in effectively leveraging complex spatial relationships for improved grounding accuracy. + +Table 2. Performance (%) comparison on Nr3D [1] with new criterions Multi-Anc and Single-Anc. + +
ModelMulti-AncSingle-AncOverall
MVT [18]52.656.655.1
CoT3DRef [2]63.165.264.4
ViewSRD71.569.569.9
+ +Table 3. Performance $(\%)$ of SRD module improves MVT [18], BUTD-DETR [21] and EDA [42] on ScanRefer [6] dataset. + +
MethodUnique (19%)Multiple (81%)Overall
0.250.50.250.50.250.5
MVT [18]77.766.531.925.340.833.3
MVT+SRD78.667.234.127.142.1 (3.2%↑)34.3(3.0%↑)
BUTD-DETR [21]82.864.944.733.950.438.6
BUTD-DETR+SRD85.066.245.334.257.9 (14.9%↑)45.7 (18.4%↑)
EDA [42]80.465.335.625.143.632.3
EDA+SRD81.067.336.428.344.4 (1.8%↑)35.3 (9.3%↑)
ViewSRD82.168.237.429.045.436.0
+ +# 4.4. SRD Enhances Other 3DVG Methods. + +Our SRD module is inherently model-agnostic, operating independently of the training process by focusing exclusively on decoupling complex multi-anchor queries into simpler single-anchor queries. This decoupling mechanism reduces ambiguity in multi-anchor descriptions, enhances target grounding, and serves as a model-independent preprocessing step, ensuring seamless compatibility with various 3DVG methods to improve performance without modifying existing architectures. As demonstrated in Table 3, integrating SRD into MVT [18], BUTD-DETR [21] and EDA [42] consistently leads to performance improvements. These results highlight SRD's ability to refine query interpretation by effectively disentangling target-anchor relationships, thereby reducing errors introduced by complex linguistic structures. These improvements reinforce the critical role of SRD module in enhancing accuracy of 3DVG. + +Table 4. Ablation studies on Nr3D [1]. All components contribute to final performance(%) + +
ComponentOverallEasyHardView Dep.View Indep.
w/o CCVTs.62.268.556.160.163.2
w/o Textual M.68.073.562.667.668.1
w/o Scene M.64.670.558.963.864.9
w/o SRD M.68.673.064.866.570.0
w/o Weight.69.074.264.066.570.2
LLM-Aug.69.174.563.768.069.5
ViewSRD69.975.364.868.670.6
+ +Table 5. Ablation of view numbers on Nr3D [1]. + +
View NumberOverallEasyHardView Dep.View Indep.
TrainTest
4166.071.760.564.067.0
4268.975.163.066.969.9
4469.975.364.868.670.6
1164.470.958.160.866.2
2267.773.062.566.168.4
8868.474.163.067.468.9
+ +# 4.5. Ablation Study + +Analysis of ViewSRD Components. To assess the contribution of each component within ViewSRD, we conducted detailed ablation studies on the Nr3D dataset [1]. Starting from the full model, we systematically removed key modules one at a time to evaluate their individual impact. The results, presented in Table 4, demonstrate that each component plays a crucial role in enhancing model performance across different scenarios. Notably, the removal of the CCVTs leads to the most significant performance degradation. This is primarily because, without the view token, the model lacks explicit viewpoint information, impairing its ability to distinguish between different perspectives. Similarly, removing either the textual module or the scene module results in a noticeable decline, underscoring the necessity of cross-modal interaction. When view-alignment weighting is disabled (w/o Weight), performance drops by $0.9\%$ , showing that dynamic alignment of view features is critical for performance under view-dependent conditions. Removing the SRD module leads to performance degradation, confirming the benefit of multi-anchor query decoupling. We also compare it with an LLM-based augmentation method from Multi3DRefer [52] and find that SRD achieves greater gains, highlighting the advantage of structured query decomposition over generic augmentation. + +Analysis of Multi-View Modeling. We evaluate the effect of varying view counts on 3DVG performance using the Nr3D dataset. As shown in Table 5, testing with more views consistently improves accuracy when the model is trained with four views, highlighting the benefit of aggregating complementary spatial cues from multiple perspectives. When training and testing with the same number of + +Table 6. Accuracy comparison when replacing different LLMs in SRD module on Nr3D [1]. + +
LLM decouplerAccuracy
OpenChat [40]69.6%
DeepSeek-R1 [28]69.9%
Qwen-Plus [46]70.5%
Qwen-Turbo [46]70.7%
+ +views, performance improves from $64.4\%$ (1 view) to $67.7\%$ (2 views), but plateaus at $68.4\%$ with 8 views, suggesting diminishing returns. Notably, four views offer a strong tradeoff, capturing diverse spatial information with minimal redundancy and maintaining computational efficiency. This also suggests that uniformly attending to many views may dilute focus on key perspectives. Future work will explore adaptive view selection. + +Analysis of SRD's LLM Decoupler. In this paper, we employ the open-source DeepSeek-R1 [28] as the LLM in the SRD module and further investigate the impact of different LLMs on the final performance of 3DVG. As shown in Table 6, different LLM decouplers exhibit varying levels of effectiveness in the sentence decoupling task. Models with stronger decoupling capabilities yield better results. For instance, OpenChat [40] and DeepSeek-R1 [28] achieve accuracies of $69.6\%$ and $69.9\%$ , respectively, while models designed with enhanced sentence decoupling capabilities, such as Qwen-Plus [46] and Qwen-Turbo [46], achieve $70.5\%$ and $70.7\%$ , with Qwen-Turbo demonstrating the highest performance. These results indicate that as an LLM's ability to disentangle complex sentence structures improves, it becomes more effective at isolating and extracting relevant information, ultimately leading to significant gains in 3DVG accuracy. + +# 5. Conclusion + +In this paper, we introduce ViewSRD, a framework that disentangles target-anchor relationships via the Simple Relation Decoupling (SRD) module and enhances multi-view understanding through the Multi-view Textual-Scene Interaction (Multi-TSI) module. By decomposing complex multi-anchor queries into simpler single-anchor sentences, SRD clarifies positional relationships, while Multi-TSI integrates textual and scene features across viewpoints using cross-modal consistent view tokens (CCVTs) to capture spatial interactions. Extensive experiments demonstrate ViewSRD's state-of-the-art performance in 3DVG. + +A limitation of ViewSRD is its assumption that complex queries can be fully decomposed without overlapping relationships. While the decomposition into overlapping relations does not degrade performance, it diminishes the intended benefits of simplification. Future work will explore adaptive query to better preserve contextual dependencies. + +Acknowledgements. This work is supported by Guangdong Provincial Natural Science Foundation for Outstanding Youth Team Project (No. 2024B1515040010), NSFC Key Project (No. U23A20391), China National Key R&D Program (Grant No. 2023YFE0202700, 2024YFB4709200), Key-Area Research and Development Program of Guangzhou City (No. 2023B01J0022), Guangdong Natural Science Funds for Distinguished Young Scholars (Grant 2023B1515020097), the National Research Foundation, Singapore under its AI Singapore Programme (AISG Award No.: AISG3-GV-2023-011), the Singapore Ministry of Education AcRF Tier 1 Grant (Grant No.: MSS25C004), and the Lee Kong Chian Fellowships. + +# References + +[1] Panos Achlioptas, Ahmed Abdelreehm, Fei Xia, Mohamed Elhoseiny, and Leonidas Guibas. Referit3d: Neural listeners for fine-grained 3d object identification in real-world scenes. In ECCV, pages 422-440, 2020. 2, 6, 7, 8 +[2] Eslam Mohamed Bakr, Mohamed Ayman, Mahmoud Ahmed, Habib Slim, and Mohamed Elhoseiny. Cot3dref: Chain-of-thoughts data-efficient 3d visual grounding. The Twelfth International Conference on Learning Representations, 2024. 6, 7 +[3] Christopher Beckham, Martin Weiss, Florian Golemo, Sina Honari, Derek Nowrouzehrai, and Christopher Pal. Visual question answering from another perspective: Clevr mental rotation tests. Pattern Recognition, 136:109209, 2023. 5 +[4] Francisco Maria Calisto, João Fernandes, Margarida Morais, Carlos Santiago, João Maria Abrantes, Nuno Nunes, and Jacinto C Nascimento. Assertiveness-based agent communication for a personalized medicine on medical imaging diagnosis. In Proceedings of the 2023 CHI conference on human factors in computing systems, pages 1-20, 2023. 1 +[5] Chun-Peng Chang, Shaoxiang Wang, Alain Pagani, and Didier Stricker. Mikasa: Multi-key-anchor & scene-aware transformer for 3d visual grounding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14131-14140, 2024. 1, 3, 7 +[6] Dave Zhenyu Chen, Angel X Chang, and Matthias Nießner. Scanrefer: 3d object localization in rgb-d scans using natural language. In ECCV, pages 202-221. Springer, 2020. 2, 6, 7 +[7] Weihong Chen, Xuemiao Xu, Haoxin Yang, Yi Xie, Peng Xiao, Cheng Xu, Huaidong Zhang, and Pheng-Ann Heng. Scjd: Sparse correlation and joint distillation for efficient 3d human pose estimation. arXiv preprint arXiv:2503.14097, 2025. 2 +[8] Tushar Choudhary, Vikrant Dewangan, Shivam Chandhok, Shubham Priyadarshan, Anushka Jain, Arun K Singh, Siddharth Srivastava, Krishna Murthy Jatavallabhula, and K Madhava Krishna. Talk2bev: Language-enhanced bird's-eye view maps for autonomous driving. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 16345-16352. IEEE, 2024. 1 +[9] Can Cui, Yunsheng Ma, Xu Cao, Wenqian Ye, and Ziran Wang. Drive as you speak: Enabling human-like interac + +tion with large language models in autonomous vehicles. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 902-909, 2024. 1 +[10] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In CVPR, pages 5828-5839, 2017. 2, 6 +[11] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 4 +[12] Azade Farshad, Yousef Yeganeh, Yu Chi, Chengzhi Shen, Böjrn Ommer, and Nassir Navab. Scenegenie: Scene graph guided diffusion models for image synthesis. In ICCV, pages 88-98, 2023. 2 +[13] Shuo Feng, Haowei Sun, Xintao Yan, Haojie Zhu, Zhengxia Zou, Shengyin Shen, and Henry X Liu. Dense reinforcement learning for safety validation of autonomous vehicles. Nature, 615(7953):620-627, 2023. 1 +[14] Liang Geng and Jianqin Yin. Viewinfer3d: 3d visual grounding based on embodied viewpoint inference. IEEE Robotics and Automation Letters, 2024. 1 +[15] Zoey Guo, Yiwen Tang, Ray Zhang, Dong Wang, Zhigang Wang, Bin Zhao, and Xuelong Li. Viewrefer: Grasp the multi-view knowledge for 3d visual grounding. In ICCV, pages 15372-15383, 2023. 1, 3, 7 +[16] Dailan He, Yusheng Zhao, Junyu Luo, Tianrui Hui, Shaofei Huang, Aixi Zhang, and Si Liu. Transrefer3d: Entity-andrelation aware transformer for fine-grained 3d visual grounding. In ACM MM, pages 2344-2352, 2021. 7 +[17] Yining Hong, Haoyu Zhen, Peihao Chen, Shuhong Zheng, Yilun Du, Zhenfang Chen, and Chuang Gan. 3d-llm: Injecting the 3d world into large language models. Advances in Neural Information Processing Systems, 36:20482-20494, 2023. 1, 3 +[18] Shijia Huang, Yilun Chen, Jiaya Jia, and Liwei Wang. Multiview transformer for 3d visual grounding. In CVPR, pages 15524-15533, 2022. 2, 3, 6, 7 +[19] Yufeng Huang, Jiji Tang, Zhuo Chen, Rongsheng Zhang, Xinfeng Zhang, Weijie Chen, Zeng Zhao, Zhou Zhao, Tangjie Lv, Zhipeng Hu, et al. Structure-clip: Towards scene graph knowledge to enhance multi-modal structured representations. In AAAI, pages 2417–2425, 2024. 1, 2 +[20] Zixin Huang, Xuesong Tao, and Xinyuan Liu. Nan-detr: noising multi-anchor makes detr better for object detection. Frontiers in Neurorobotics, 18:1484088, 2024. 1 +[21] Ayush Jain, Nikolaos Gkanatsios, Ishita Mediratta, and Katerina Fragkiadaki. Bottom up top down detection transformers for language grounding in images and point clouds. In European Conference on Computer Vision, pages 417-433. Springer, 2022. 7 +[22] Xinjie Jiang, Chenxi Zheng, Xuemiao Xu, Bangzhen Liu, Weiying Zheng, Huaidong Zhang, and Shengfeng He. Vrdone: One-stage video visual relation detection. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 1437-1446, 2024. 2 +[23] Jingdan Kang, Haoxin Yang, Yan Cai, Huaidong Zhang, Xuemiao Xu, Yong Du, and Shengfeng He. Sita: Structurally + +imperceptible and transferable adversarial attacks for stylized image generation. IEEE Transactions on Information Forensics and Security, 2025. 2 +[24] Sanjoy Kundu and Sathyanarayanan N Aakur. Is-ggt: Iterative scene graph generation with generative transformers. In CVPR, pages 6292-6301, 2023. 2 +[25] Jialu Li and Mohit Bansal. Panogen: Text-conditioned panoramic environment generation for vision-and-language navigation. Advances in Neural Information Processing Systems, 36, 2024. 1 +[26] Yihong Lin, Xuemiao Xu, Huaidong Zhang, Cheng Xu, Weijie Li, Yi Xie, Jing Qin, and Shengfeng He. Delving into invisible semantics for generalized one-shot neural human rendering. IEEE Transactions on Visualization and Computer Graphics, 2025. 2 +[27] Xiongkun Linghu, Jiangyong Huang, Xuesong Niu, Xiaojian Shawn Ma, Baoxiong Jia, and Siyuan Huang. Multimodal situated reasoning in 3d scenes. Advances in Neural Information Processing Systems, 37:140903-140936, 2025. 2 +[28] Aixin Liu, Bei Feng, Bing Xue, Bingxuan Wang, Bochao Wu, Chengda Lu, Chenggang Zhao, Chengqi Deng, Chenyu Zhang, Chong Ruan, et al. Deepseek-v3 technical report. arXiv preprint arXiv:2412.19437, 2024. 2, 6, 8 +[29] Bangzhen Liu, Chenxi Zheng, Xuemiao Xu, Cheng Xu, Huaidong Zhang, and Shengfeng He. Rotation-adaptive point cloud domain generalization via intricate orientation learning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025. 2 +[30] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6 +[31] Ruiyuan Lyu, Jingli Lin, Tai Wang, Xiaohan Mao, Yilun Chen, Runsen Xu, Haifeng Huang, Chenming Zhu, Dahua Lin, and Jiangmiao Pang. Mmscan: A multi-modal 3d scene dataset with hierarchical grounded language annotations. Advances in Neural Information Processing Systems, 37:50898-50924, 2025. 1 +[32] Yunze Man, Liang-Yan Gui, and Yu-Xiong Wang. Situational awareness matters in 3d vision language reasoning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13678-13688, 2024. 1 +[33] Honghan Pan, Bangzhen Liu, Xuemiao Xu, Chenxi Zheng, Yongwei Nie, and Shengfeng He. Gaussian prompter: Linking 2d prompts for 3d gaussian segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025. 2 +[34] Itthisak Phueaksri, Marc A Kastner, Yasutomo Kawanishi, Takahiro Komamizu, and Ichiro Ide. An approach to generate a caption for an image collection using scene graph generation. IEEE Access, 2023. 2 +[35] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 30, 2017. 5 +[36] Junha Roh, Karthik Desingh, Ali Farhadi, and Dieter Fox. Languagerefer: Spatial-language model for 3d visual grounding. In Conference on Robot Learning, pages 1046-1056. PMLR, 2022. 6, 7 + +[37] Xiangxi Shi, Zhonghua Wu, and Stefan Lee. Aware visual grounding in 3d scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14056-14065, 2024. 1 +[38] Yashar Talebirad and Amirhossein Nadiri. Multi-agent collaboration: Harnessing the power of intelligent llm agents. arXiv preprint arXiv:2306.03314, 2023. 1 +[39] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 5 +[40] Guan Wang, Sijie Cheng, Xianyuan Zhan, Xiangang Li, Sen Song, and Yang Liu. Openchat: Advancing open-source language models with mixed-quality data. arXiv preprint arXiv:2309.11235, 2023. 2, 8 +[41] Yuan Wang, Yali Li, and Shengjin Wang. G^3-lq: Marrying hyperbolic alignment with explicit semantic-geometric modeling for 3d visual grounding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13917–13926, 2024. 5 +[42] Yanmin Wu, Xinhua Cheng, Renrui Zhang, Zesen Cheng, and Jian Zhang. Eda: Explicit text-decoupling and dense alignment for 3d visual grounding. In CVPR, pages 19231-19242, 2023. 1, 2, 6, 7 +[43] Cheng Xu, Wei Qu, Xuemiao Xu, and Xueling Liu. Multiscale flow-based occluding effect and content separation for cartoon animations. IEEE Transactions on Visualization and Computer Graphics, 29(9):4001-4014, 2022. 2 +[44] Can Xu, Yuehui Han, Rui Xu, Le Hui, Jin Xie, and Jian Yang. Multi-attribute interactions matter for 3d visual grounding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17253-17262, 2024. 1, 2 +[45] Yingjie Xu, Bangzhen Liu, Hao Tang, Bailin Deng, and Shengfeng He. Learning with unreliability: Fast few-shot voxel radiance fields with relative geometric consistency. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20342-20351, 2024. 2 +[46] An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, et al. Qwen2.5 technical report. arXiv preprint arXiv:2412.15115, 2024. 2, 8 +[47] Haoxin Yang, Xuemiao Xu, Cheng Xu, Huaidong Zhang, Jing Qin, Yi Wang, Pheng-Ann Heng, and Shengfeng He. G2face: High-fidelity reversible face anonymization via generative and geometric priors. IEEE Transactions on Information Forensics and Security, 2024. 2 +[48] Zhengyuan Yang, Songyang Zhang, Liwei Wang, and Jiebo Luo. Sat: 2d semantics assisted training for 3d visual grounding. In ICCV, pages 1856-1866, 2021. 7 +[49] Zhao Yang, Jiaxuan Liu, Yucheng Han, Xin Chen, Zebiao Huang, Bin Fu, and Gang Yu. Appagent: Multimodal agents as smartphone users. arXiv preprint arXiv:2312.13771, 2023. 1 +[50] Qifan Yu, Juncheng Li, Yu Wu, Siliang Tang, Wei Ji, and Yueting Zhuang. Visually-prompted language model for fine-grained scene graph generation in an open world. In ICCV, pages 21560-21571, 2023. 2 + +[51] Zhihao Yuan, Jinke Ren, Chun-Mei Feng, Hengshuang Zhao, Shuguang Cui, and Zhen Li. Visual programming for zero-shot open-vocabulary 3d visual grounding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20623-20633, 2024. 1, 3 +[52] Yiming Zhang, ZeMing Gong, and Angel X Chang. Multi3drefer: Grounding text description to multiple 3d objects. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15225-15236, 2023. 8 +[53] Yuqi Zhang, Han Luo, and Yinjie Lei. Towards clip-driven language-free 3d visual grounding via 2d-3d relational enhancement and consistency. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13063–13072, 2024. 2 +[54] Lichen Zhao, Daigang Cai, Lu Sheng, and Dong Xu. 3dvg-transformer: Relation modeling for visual grounding on point clouds. In ICCV, pages 2928-2937, 2021. 7 +[55] Chenxi Zheng, Yihong Lin, Bangzhen Liu, Xuemiao Xu, Yongwei Nie, and Shengfeng He. Recdreamer: Consistent text-to-3d generation via uniform score distillation. In The Thirteenth International Conference on Learning Representations. 2 +[56] Chenxi Zheng, Bangzhen Liu, Xuemiao Xu, Huaidong Zhang, and Shengfeng He. Learning an interpretable stylized subspace for 3d-aware animatable artforms. IEEE Transactions on Visualization and Computer Graphics, 31(2):1465-1477, 2024. 2 +[57] Gengze Zhou, Yicong Hong, and Qi Wu. Navgpt: Explicit reasoning in vision-and-language navigation with large language models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 7641-7649, 2024. 1 +[58] Ziyu Zhu, Zhuofan Zhang, Xiaojian Ma, Xuesong Niu, Yixin Chen, Baoxiong Jia, Zhidong Deng, Siyuan Huang, and Qing Li. Unifying 3d vision-language understanding via promptable queries. In European Conference on Computer Vision, pages 188–206. Springer, 2024. 1 \ No newline at end of file diff --git a/ICCV/2025/ViewSRD_ 3D Visual Grounding via Structured Multi-View Decomposition/images.zip b/ICCV/2025/ViewSRD_ 3D Visual Grounding via Structured Multi-View Decomposition/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..d181ba10d86979e29e372a624ff61d5169cc6828 --- /dev/null +++ b/ICCV/2025/ViewSRD_ 3D Visual Grounding via Structured Multi-View Decomposition/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:331b556dd5e0d0899d57ee3a2a3371547bc20aef264dea2ab5c8f49976fc4e92 +size 597715 diff --git a/ICCV/2025/ViewSRD_ 3D Visual Grounding via Structured Multi-View Decomposition/layout.json b/ICCV/2025/ViewSRD_ 3D Visual Grounding via Structured Multi-View Decomposition/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..4d56f5b73977cb82fb0489124df843f9ef3793fa --- /dev/null +++ b/ICCV/2025/ViewSRD_ 3D Visual Grounding via Structured Multi-View Decomposition/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4832766d2dd66fd806c05681f29d4616c0b0ab323edeecf4b60f3f7f0a03fb75 +size 443768 diff --git a/ICCV/2025/VisHall3D_ Monocular Semantic Scene Completion from Reconstructing the Visible Regions to Hallucinating the Invisible Regions/329f4613-dbba-4972-811d-2a326b45a72a_content_list.json b/ICCV/2025/VisHall3D_ Monocular Semantic Scene Completion from Reconstructing the Visible Regions to Hallucinating the Invisible Regions/329f4613-dbba-4972-811d-2a326b45a72a_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..4355e3377b21424ea11d4a768efa062b181d7742 --- /dev/null +++ b/ICCV/2025/VisHall3D_ Monocular Semantic Scene Completion from Reconstructing the Visible Regions to Hallucinating the Invisible Regions/329f4613-dbba-4972-811d-2a326b45a72a_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9712ea716a60c73e7c97ed46a7ae26a769dc8fa30ebdee85ee179bd8fa8163aa +size 81781 diff --git a/ICCV/2025/VisHall3D_ Monocular Semantic Scene Completion from Reconstructing the Visible Regions to Hallucinating the Invisible Regions/329f4613-dbba-4972-811d-2a326b45a72a_model.json b/ICCV/2025/VisHall3D_ Monocular Semantic Scene Completion from Reconstructing the Visible Regions to Hallucinating the Invisible Regions/329f4613-dbba-4972-811d-2a326b45a72a_model.json new file mode 100644 index 0000000000000000000000000000000000000000..937d059378cc85c1eec5cffbc9ac34aa2b4ec86a --- /dev/null +++ b/ICCV/2025/VisHall3D_ Monocular Semantic Scene Completion from Reconstructing the Visible Regions to Hallucinating the Invisible Regions/329f4613-dbba-4972-811d-2a326b45a72a_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d9641e538db9c739e1dd7d05e0e70a2206f14fe00cb549163c7e6d209d7bec46 +size 100832 diff --git a/ICCV/2025/VisHall3D_ Monocular Semantic Scene Completion from Reconstructing the Visible Regions to Hallucinating the Invisible Regions/329f4613-dbba-4972-811d-2a326b45a72a_origin.pdf b/ICCV/2025/VisHall3D_ Monocular Semantic Scene Completion from Reconstructing the Visible Regions to Hallucinating the Invisible Regions/329f4613-dbba-4972-811d-2a326b45a72a_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..b4e55e6f33c1964663a3c01c7c171aea4c2bc69c --- /dev/null +++ b/ICCV/2025/VisHall3D_ Monocular Semantic Scene Completion from Reconstructing the Visible Regions to Hallucinating the Invisible Regions/329f4613-dbba-4972-811d-2a326b45a72a_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dacc5f304dc65f47e144d7f231bb6320d67bff8856a8023bdfabd5adee203c31 +size 1954753 diff --git a/ICCV/2025/VisHall3D_ Monocular Semantic Scene Completion from Reconstructing the Visible Regions to Hallucinating the Invisible Regions/full.md b/ICCV/2025/VisHall3D_ Monocular Semantic Scene Completion from Reconstructing the Visible Regions to Hallucinating the Invisible Regions/full.md new file mode 100644 index 0000000000000000000000000000000000000000..b15b845d6d4499233cec2dfdaf12fc38b4b89a96 --- /dev/null +++ b/ICCV/2025/VisHall3D_ Monocular Semantic Scene Completion from Reconstructing the Visible Regions to Hallucinating the Invisible Regions/full.md @@ -0,0 +1,330 @@ +# VisHall3D: Monocular Semantic Scene Completion from Reconstructing the Visible Regions to Hallucinating the Invisible Regions + +Haoang Lu Yuanqi Su* Xiaoning Zhang Longjun Gao Yu Xue Le Wang Xi'an Jiaotong University, China https://github.com/luang321/vishall3d + +# Abstract + +This paper introduces VisHall3D, a novel two-stage framework for monocular semantic scene completion that aims to address the issues of feature entanglement and geometric inconsistency prevalent in existing methods. VisHall3D decomposes the scene completion task into two stages: reconstructing the visible regions (vision) and inferring the invisible regions (hallucination). In the first stage, VisFrontierNet, a visibility-aware projection module, is introduced to accurately trace the visual frontier while preserving fine-grained details. In the second stage, OcclusionMAE, a hallucination network, is employed to generate plausible geometries for the invisible regions using a noise injection mechanism. By decoupling scene completion into these two distinct stages, VisHall3D effectively mitigates feature entanglement and geometric inconsistency, leading to significantly improved reconstruction quality. + +The effectiveness of VisHall3D is validated through extensive experiments on two challenging benchmarks: SemanticKITTI and SSCBench-KITTI-360. VisHall3D achieves state-of-the-art performance, outperforming previous methods by a significant margin and paves the way for more accurate and reliable scene understanding in autonomous driving and other applications. + +# 1. Introduction + +Monocular Semantic Scene Completion (Monocular SSC) [3, 20, 29, 42, 43] has emerged as a promising solution for enabling autonomous vehicles to perceive and understand their surroundings, as it can reconstruct complete 3D scenes using only single RGB images. This cost-effective and flexible approach has the potential to revolutionize 3D perception in autonomous driving. However, existing Monocular SSC methods still suffer from two major challenges: feature entanglement and geometric inconsistency. + +These issues stem from the inherent distinction between visible and invisible regions in the 3D scene. The visible regions, or the visual frontier, correspond to the surface points of the 3D scene captured in the input image. When estimating 3D occupancy grids, the ground truth is typically obtained by accumulating data from multiple adjacent Lidar frames, inevitably including some voxels that are not visible in the current image. As a result, the occupancy estimation task faces two distinct challenges: tracing the visual frontier and generating the unseen voxels, as shown in Fig.1. + +Existing methods, such as MonoScene [3], Occdepth [29], and NDC-Scene [42], among others [12, 20, 37, 43, 47], have made significant progress in Monocular SSC. However, by treating SSC as a single-stage task, they fail to recognize the inherent distinction between visible and invisible regions, leading to feature entanglement and geometric inconsistency. Feature entanglement occurs when the features learned for visible and invisible regions are mixed together, while geometric inconsistency arises when the reconstructed visible and invisible regions are misaligned or have conflicting structures. + +To address these challenges, we propose VisHall3D, a novel two-stage framework that explicitly separates the tasks of reconstructing visible regions (vision) and inferring invisible regions (hallucination). In the first stage, VisHall3D uses VisFrontierNet, a visibility-aware projection module, to accurately trace the visual frontier while preserving fine-grained details. By explicitly modeling the boundary between visible and invisible regions, VisFrontierNet helps to mitigate feature entanglement. In the second stage, VisHall3D employs OcclusionMAE, a hallucination network that generates plausible geometries for invisible regions using a noise injection mechanism. By decomposing SSC into these two distinct stages, VisHall3D effectively addresses feature entanglement and geometric inconsistency, leading to significantly improved reconstruction quality, as shown in Fig.1. + +We validate the effectiveness of VisHall3D through extensive experiments on SemanticKITTI [1] and SSCBenchKITTI-360 [21, 25], achieving state-of-the-art performance. + +![](images/6f58826bb3d0d1908539ec20f881ae8bb0cd2b2800ed51fa42a378f268f6d5b7.jpg) +(a) Visualization of the Visible and Invisible Parts in a Scene +Figure 1. Visualization of the decoupling of prediction for visible regions and hallucination for invisible regions: (a) Division of visible and invisible areas: OOV (Out-of-view) voxels are in black, and occluded FOV (Field-of-View) voxels are in dark; (b) Pipeline of current mainstream methods; (c) Pipeline of our approach and its decoupling strategy for prediction and hallucination. + +![](images/0404c0aab474dd6145ebd831181b7adcfad7f85e2b0098da6c6808424df48ff4.jpg) +(c) Ours VisHall3D + +# Our main contributions are: + +- A novel two-stage framework, VisHall3D, for Monocular SSC that explicitly separates vision and hallucination processes to mitigate feature entanglement and geometric inconsistency. +- VisFrontierNet, a visibility-aware projection module that accurately traces the visual frontier by modeling the boundary between visible and invisible regions. +- OcclusionMAE, a hallucination network that generates plausible geometries for invisible regions using a noise injection mechanism. +- VisHall3D sets a new standard for Monocular SSC, paving the way for more accurate and reliable scene understanding in various applications. + +# 2. Related Work + +# 2.1. Semantic Scene Completion + +3D Semantic Scene Completion aims to generate dense 3D semantic voxel grids from incomplete observations, as first defined in SSCNet [34]. Early SSC works typically relied on geometric inputs such as LiDAR [6, 31, 40] or depth maps [18, 19, 34]. However, recent studies have explored reconstructing entire SSC scenes using visual-only inputs, starting from multi-view methods [11, 22], then progressing to stereo-based approaches [14], and finally to monocular + +frameworks [3, 42]. + +Among multi-view methods, BEVFormer [22] and TPV-Former [11] leverage diverse 3D feature representations to reduce computational overhead and enhance network capability. FB-OCC [24] and OccTransformer [26], on the other hand, focus on improving the transition from 2D to 3D. + +Monoscene [3] pioneered single-image SSC, while subsequent works introduced NDC coordinates [42], depth prediction [20], horizontal direction emphasis [37], and voxel-to-graph structures [41] to optimize 2D-to-3D transformation. The latest advancements, Symphonies [12] and CG-Former [43], refine the approach with object-aware and context-aware optimizations. However, these methods still suffer from feature entanglement and geometric inconsistency due to the inherent ambiguity in monocular 3D reconstruction. + +# 2.2.3D From a Single Image + +3D reconstruction from a single image is an ill-posed problem due to the lack of explicit depth information. Early monocular 3D tasks focused on reconstructing coarse information within the visible region by extending 2D methods [28, 48], exploiting geometric knowledge of objects [13, 23, 44, 45], or treating it as a Perspective-n-Points (PnP) problem [5, 27]. + +The emergence of monocular semantic scene comple + +![](images/7d6792b208a1916ec7f41f4a9458344af4d92da01a4561058c8c9f826a95a1fb.jpg) +Figure 2. Overview of our method and the structure of each module. (a) The pipeline of our proposed VisFrontierNet, including unsigned distance (uDistance) Encoding and Feature Refinement. (b) The detailed architecture of the OcclusionMAE. The unsigned distance is calculated for the identified visual frontier which is then encoded and combined with image feature for the 3D representation. + +tion (SSC) [3, 36, 47] has expanded the scope of monocular 3D tasks to reconstructing both visible and invisible regions with detailed shapes and semantics. This requires models to possess not only visual perception capabilities but also the ability to hallucinate missing information. However, existing methods often struggle to effectively handle the distinction between visible and invisible regions, leading to suboptimal results. + +# 2.3. Multi-stage Refinement + +Multi-stage refinement strategy has been widely adopted in computer vision, particularly in object detection [2, 4, 35, 49], to refine proposals and enhance accuracy. Inspired by this, we propose to decouple the tasks of visual perception and hallucination in monocular SSC. We first generate a coarse prediction of the visible regions, and then leverage Hallucinating MAE [9] to hallucinate and optimize the entire scene. + +# 3. VisHall3D: VisFrontierNet and Occlusion-MAE + +We propose VisHall3D, a two-stage framework for generating a 3D occupancy grid from a single image: (1) refinement of the visual frontier by VisFrontierNet and (2) completion of the invisible voxels through OcclusionMAE, as shown in Fig.2. The visual frontier refers to the set of voxels directly visible from the camera's viewpoint, corresponding to the surface points of the 3D scene captured in the input image, as shown in Fig.3. + +In Sec.3.1, we provide an overview of the proposed model, including its overall architecture and key compo + +nents. We then delve into the details of VisFrontierNet in Sec.3.2, explaining its role in refining the visual frontier. Subsequently, in Sec.3.3, we present OcclusionMAE and its approach to completing the invisible voxels. Each subsection offers a comprehensive explanation of the respective modules and their roles within the framework. + +# 3.1. Method Overview + +In our approach, we divide the 3D occupancy grid voxels into visible, occluded, and out-of-view categories based on visibility (Fig.1). Visible voxels are directly observed; occluded voxels are hidden by obstructions; out-of-view voxels lie beyond the camera's field of view. + +Our proposed method seamlessly integrates feature lifting and 3D generation to reconstruct accurate 3D occupancy from a single input image. + +![](images/82bd2e435c8c3e02793219f17e7eaa20bd54aa8eadd1270f454e062f527cc89d.jpg) +Figure 3. Illustration of the visual frontier from side view. The solid line gives the estimated depth. + +Estimating Visible Voxel Occupancy: For visible voxels, we leverage the direct correspondence between the 2D image pixels and their corresponding 3D coordinates in the occupancy grid. Given the estimated depth map $D$ , we determine the visible frontier in the 3D space and identify the voxels in front of the frontier as visible voxels. + +To capture the frontier's geometry, we compute a truncated unsigned distance map $D_{u}$ representing voxel distances to the frontier. This map is encoded using 3D convolutions and fused with image features projected into 3D space using voxel-projected bounding boxes $bbox_{v}$ and RoI pooling, injecting visual information into the 3D volume. + +We then refine the visual frontier by lifting 2D features into 3D using a depth probability map $D_{p}$ and scattering them onto a discretized view frustum. Finally, 3D deformable attention refines frontier features, focusing the network on informative regions for better reconstruction. + +Inferring Occluded and Out-of-View Voxel Occupancy: For occluded and out-of-view voxels, the direct visual evidence is not available, requiring a more sophisticated generation process. We propose OcclusionMAE, a denoising Masked Autoencoder that generates the occupancy for the entire set of voxels by taking the noisy prediction on the visible voxels from VisFrontierNet as input. + +The denoising procedure in OcclusionMAE involves adding noise to the visible voxel predictions and injecting the noise level into the network via adaptive batch normalization. OcclusionMAE employs a 3D U-Net architecture to process the modulated features and generate the final occupancy grid, effectively aligning the results with the image semantics while accounting for the noisy nature of the visible voxel predictions. + +# 3.2. Visual Frontier Propagation via VisFrontierNet + +VisFrontierNet is designed to estimate the occupancy of visible voxels by leveraging the correspondence between 2D image pixels and their 3D coordinates in the occupancy grid. Given the 3D grids $G$ , our goal is to predict the label of each voxel $v \in G$ from an input image. The main steps of VisFrontierNet include: (1) identifying the visual frontier, (2) representing and encoding it, and (3) refining its feature. + +We first transform the coordinates of each voxel into the camera frame; then project them onto the image plane, obtaining the corresponding coordinates $(x_v, y_v, d_v)$ for each voxel $v$ . Let $\pi$ denote the projection process, it maps each voxel $v$ to a corresponding 2D coordinate $(x_v, y_v)$ on the image plane. + +$$ +\left(x _ {v}, y _ {v}, d _ {v}\right) = \pi (v) \tag {1} +$$ + +Here, we also obtain the depth value $d_v$ for each voxel. + +Visual Frontier Identification. The estimated depth map $D$ gives a visual frontier that determines the visible and occluded voxels. The voxels in front of the visual frontier are considered visible, as shown in Fig. 3. + +$$ +V := \{v | v \in G, d _ {v} < D \left(x _ {v}, y _ {v}\right) + \theta_ {d} \} \tag {2} +$$ + +By comparing the projected depth $d_v$ with the estimated depth of the projected point $D(x_v, y_v)$ , we determine + +whether the voxel lies in front of or behind the visual frontier. A relaxation factor $\theta_{d}$ is added to accommodate potential errors in depth estimation. The filtered visible voxels $V$ allow our network to focus on estimating the correct lifting from the 2D image plane into the 3D space. + +Unsigned Distance Function and Encoding. To capture the geometric characteristics of the visual frontier, we introduce the truncated unsigned distance map $(D_u)$ . It measures the unsigned distance $dist_v = |d_v - D(x_v, y_v)|$ from each voxel to the visual frontier, and is defined as: + +$$ +D _ {u} = \left\{ \begin{array}{l l} 2 - 2 \sigma (\gamma * \text {d i s t} _ {v}), & \text {d i s t} _ {v} < \theta \\ 0, & \text {O t h e r w i s e} \end{array} \right. \tag {3} +$$ + +where $\sigma$ is the sigmoid function, $\gamma$ is a controlling factor set to 10 for sharp decaying, and $\theta$ controls the truncated region (set to 1), as shown in Fig.3 (a). + +![](images/4c4f1f7528e31bcfa15b6a7b74bfcda1b14c57dafee94b5da26bceaf8c4253c3.jpg) +(a) + +![](images/858c2678dd78c8181e361d277a987f9081b2d928aced873ee594dbd678973cfd.jpg) +(b) +Figure 4. Visual comparison of Hard Lifting and our Unsigned Distance approach. (a) Illustration of Hard Lifting. (b) Geometric information captured by our Unsigned Distance approach. + +Through this approach, we address the challenge of sparse voxels in distant regions, a limitation inherent in the hard lifting of geometric information adopted by previous methods [12, 20], where each pixel corresponds to a single voxel, as illustrated in the Fig. 4. + +The unsigned distance map $D_{u}$ is then fed into a 3D convolutional module for geometric encoding $\mathrm{Enc}_g$ , which captures the geometric properties of the visual frontier. To inject the image features into the 3D space, we combine the geometric encoding with the image feature for each voxel using RoI pooling [7]: + +$$ +F _ {3 D, 1} (v) = \operatorname {E n c} _ {g} (v) \oplus \operatorname {R o I} \left(F _ {2 D}, b b o x _ {v}\right) \tag {4} +$$ + +where $F_{2D}$ denotes the image feature map, and $bbox_{v}$ is the projected bounding box for voxel $v$ . In practice, We use ResNet-50 [8] as the backbone for extracting multi-scale features from the input image. Similar to Symphonies [12], we utilize MaskDINO [16]'s neck to combine multi-scale image features into a single one $F_{2D}$ . + +Feature Refinement for Visual Frontier. After encoding both the geometric and visual features for the frontier, + +![](images/bd5024ffe1204d991f12c20741036eba87d28f63c2719aa4d7e225f80b8cbd92.jpg) +Figure 5. Qualitative visual comparison with Monoscene[3] and Ground Truth (GT) on the SemanticKITTI[1] dataset. + +![](images/057fdc1fa042d497b99d0add7f9158aa0261bd5831bc6dae499016f8f9a4bce0.jpg) + +a refinement procedure is utilized to modify the representation for visual frontier according to image features. We lift the 2D features $F_{2D}$ into the 3D space using a depth probability map $D_p$ , which represents the likelihood of each pixel belonging to different depth ranges. The lifted features are then scattered onto a discretized grid on the view frustum [30], as shown in Fig.2. + +To refine the features for the visual frontier, we employ 3D deformable attention [17, 49]. It learns to dynamically adjust the receptive field and sampling locations based on the input features, enabling the network to focus on the most informative regions and adapt to the specific characteristics of the visual frontier, producing refined 3D features $F_{3D,2}$ . + +$$ +F _ {3 D, 2} \leftarrow \operatorname {D e f o r m 3 D} \left(F _ {3 D, 1}, F _ {2 D}, D _ {p}\right) \tag {5} +$$ + +# 3.3. Generation via Denoising MAE + +OcclusionMAE is a denoising Masked Autoencoder designed to generate the occupancy for the entire set of voxels, taking the noisy prediction $O_V$ on the visible voxels from VisFrontierNet as input. + +Noise Adding to Visible Voxel Predictions. Considering the noisy nature of the predicted occupancy on visible voxels, we introduce a denoising procedure. For each visible voxel in $V$ , we randomly assign it a value from its neighboring voxels within a horizontal range $R_{h}$ and a depth range $R_{d}$ . Here, considering the spatial resolution of the scene (256 voxels in the horizontal and depth directions, but only 32 voxels in the vertical direction), we focus on adding noise in the horizontal and depth directions while leaving the vertical direction unchanged. The noise adding procedure is described as follows. + +$$ +\tilde {O} _ {V} = \operatorname {A d d N o i s e} \left(O _ {V}, t \left(R _ {h}, R _ {d}\right)\right) \tag {6} +$$ + +Where, $t$ is the noise level. It is important to note that the added noise does not alter the original semantics (e.g., changing a vehicle to a person) but only perturbs geometric information (e.g., adjusting voxel positions). This is because we do not want the network to generate non-existent classes. + +Noise Level Injecting via Adaptive Batch Normalization. The noise level $t$ is injected into the network using adaptive batch normalization [10]. Let $F$ denote the features extracted from $\tilde{O}_V$ through patch embedding and 3D convolution. The adaptive batch normalization works as follows: + +$$ +a d a. B N (F (v)) = \gamma \cdot \frac {F (v) - \mu}{\sqrt {\delta^ {2} + \epsilon}} + \beta \tag {7} +$$ + +where $\mu$ and $\delta^2$ are the mean and variance of $F$ , respectively. The features from the noise level $t$ are used to generate the parameters $\gamma$ and $\beta$ . This allows the network to adapt to different noise levels and generate more accurate occupancy predictions. + +Generating Occupancy Grid with 3D U-Net. The resulting 3D feature is concatenated with refined context feature derived from VisFrontierNet $F_{3D,2}(v)$ and fed into a 3D U-Net to produce the final occupancy prediction. The 3D U-Net processes the modulated features and generates the occupancy grid for the entire set of voxels, effectively aligning the results with the image semantics while accounting for the noisy nature of the visible voxel predictions. + +The denoising procedure can be summarized as follows. + +$$ +O = \text {O c c l u s i o n M A E} (\bar {O} _ {V}, t) \tag {8} +$$ + +By introducing a denoising procedure and injecting the + +
MethodDateIoU↑ mIoU↑roadsidewalkparkingother-gmd.buildingcartruckbicyclemotorcycleother-veh.vegetationtrunkterrainpersonbicyclistmotorcyclistfencepoletraf-sign
Stereo camera-based methods
StereoScene[14]IJCAI202443.3415.3661.9031.2030.7010.7024.2022.802.803.402.406.1023.808.4027.002.902.200.5016.507.007.20
Monocular temporal methods
VoxFormer-T[20]CVPR202343.2113.4154.1026.9025.107.3023.5021.703.601.901.604.1024.408.1024.201.601.100.0013.106.605.70
HASSC-T[36]CVPR202442.8714.3855.3029.6025.9011.3023.1023.002.901.901.504.9024.809.8026.501.403.000.0014.307.007.10
HTCL[15]ECCV202444.2317.0964.4034.8033.8012.4025.9027.305.701.802.205.4025.3010.8031.201.103.100.9021.109.008.30
H2GFormer-T[37]AAAI202443.5214.6057.9030.4030.006.9024.0023.705.200.601.205.0025.2010.7025.801.100.100.0014.607.509.30
Monocular single-frame methods
MonoScene[3]CVPR202334.1611.0854.7027.1024.805.7014.4018.803.300.500.704.4014.902.4019.501.001.400.4011.103.302.10
VoxFormer-S[20]CVPR202342.9512.2053.9025.3021.105.6019.8020.803.502.600.703.7022.407.5021.301.402.600.2011.105.104.90
TPVFormer[11]CVPR202334.2511.2655.1027.2027.406.5014.8019.203.701.000.502.3013.902.6020.401.102.400.3011.002.901.50
SurroundOcc[38]ICCV202334.7211.8656.9028.3030.206.8015.2020.601.401.601.204.4014.903.4019.301.402.000.1011.303.902.40
OccFormer[46]ICCV202334.5312.3255.9030.3031.506.5015.7021.601.201.501.703.2016.803.9021.302.201.100.2011.903.803.70
IAMSSC[39]T-ITS202443.7412.3754.0025.5024.706.9019.2021.303.801.100.603.9022.705.8019.401.502.900.5011.905.304.10
DepthSSC[41]arXiV202444.5813.1155.6427.2525.725.7820.4621.943.741.350.984.1723.377.6421.561.342.790.2812.945.876.23
HASSC-S[36]CVPR202443.4013.34------------------
Symphonize[12]CVPR202442.1915.0458.4029.3026.9011.7024.7023.603.203.602.605.6024.2010.0023.103.201.902.0016.107.708.00
H2GFormer-S[37]AAAI202444.2013.7256.4028.6026.504.9022.8023.404.800.800.904.1024.609.1023.801.202.500.1013.306.406.30
MonoOcc-L[47]ICRA2024-15.6359.1030.9027.109.8022.9023.907.204.502.407.7025.009.8026.102.804.700.6016.907.308.40
CGFormer[43]NIPS202444.4116.6364.3034.2034.1012.1025.8026.104.303.701.302.7024.5011.2029.301.703.600.4018.708.709.30
OursICCV202546.5017.4664.6034.1032.0012.5026.9026.707.502.903.306.2027.3012.5028.002.305.101.9019.509.209.20
+ +Table 1. Quantitative results on the hidden test set of SemanticKITTI [1], where the highest and second-highest scores for each metric are highlighted in bold and underline, respectively. + +noise level into the network, OcclusionMAE learns to generate accurate occupancy predictions. + +# 3.4. Training Losses + +To effectively train VisFrontierNet and OcclusionMAE, we employ a combination of several loss functions that capture different aspects of the 3D reconstruction problem. The primary loss is a category frequency-weighted cross-entropy loss $\mathcal{L}_{ce}$ , which guides the learning of both networks. Additionally, we introduce the Scene-Class Affinity Loss, inspired by MonoScene [3], to separately constrain the recall, precision, and specificity of the scene's geometry and semantics, denoted as $\mathcal{L}_{geo}$ and $\mathcal{L}_{sem}$ , respectively. Finally, a depth loss $\mathcal{L}_d$ is applied to ensure the coincidence of the probabilistic depth distribution $D_{p}$ and the depth map $D$ . + +# 4. Experimental Evaluations + +In this section, we evaluate the performance of VisHall3D on two mainstream outdoor SSC (Semantic Scene Completion) datasets: SemanticKITTI [1] and SSCBench-KITTI-360 [21, 25]. We compare VisHall3D with state-of-the-art (SOTA) methods, including those utilizing stereo vision or temporal information. Furthermore, we conduct abla + +tion studies to investigate the impact of each module on the model's performance. + +# 4.1. Dataset and Evaluation Metrics + +SemanticKITTI [1] and SSCBench-KITTI-360 [21, 25] are two widely-used outdoor SSC datasets. Both datasets adopt a scene range of $51.2\mathrm{m} \times 51.2\mathrm{m} \times 6.4\mathrm{m}$ , with each voxel having an edge length of $0.2\mathrm{m}$ , resulting in a scene resolution of $256 \times 256 \times 32$ . SemanticKITTI consists of 10 training sequences (3,834 samples), 1 validation sequence (815 samples), and 11 test sequences (3,992 samples), with 20 valid classes and 1 invalid class. SSCBench-KITTI-360 provides 7 training sequences (8,487 samples), 1 validation sequence (1,812 samples), and 1 test sequence (2,566 samples), with 19 valid classes. + +Following previous methods[12, 20, 36, 37, 43, 47], we use MobileStereoNet[33] for depth prediction. + +# 4.2. Comparisons + +We present the comparison of VisHall3D with SOTA methods on the SemanticKITTI [1] and SSCBench-KITTI-360 [21, 25] datasets in Tab.1 and Tab.2, respectively. + +On the SemanticKITTI dataset, VisHall3D achieves an mIoU of $17.46\%$ , outperforming all existing monocular + +
MethodDateIoU↑ mIoU↑carbicyclemotorcycletruckother-veh.personroadparkingsidewalkother-grd.buildingfencevegetationterrainpoletraf.-signother-struct.other-object.
LiDAR-based methods
SSCNet [34]CVPR201753.5816.9531.950.000.1710.290.000.0765.7017.3341.243.2244.416.7743.7228.870.780.758.690.67
LMSCNet [32]3DV 202047.3513.6520.910.000.000.260.580.0062.9513.5133.510.2043.670.3340.0126.800.000.003.630.00
Monocular camera-based methods
MonoScene [3]CVPR202337.8712.3119.340.430.588.022.030.8648.3511.3828.133.3232.893.5326.1516.756.925.674.203.09
TPVFormer [11]CVPR202340.2213.6421.561.091.378.062.572.3852.9911.9931.073.7834.834.8030.0817.527.465.865.482.70
OccFormer [46]ICCV202340.2713.8122.580.660.269.893.822.7754.3013.4431.533.5536.424.8031.0019.517.778.516.954.60
VoxFormer [20]CVPR202338.7611.9117.841.160.894.562.061.6347.019.6727.212.8931.384.9728.9914.696.516.923.792.43
IAMSSC [39]T-ITS202441.8012.9718.532.451.765.123.923.0947.5510.5628.354.1231.536.2829.1715.248.297.016.354.19
DepthSSC [41]arXiV202440.8514.2821.902.364.3011.514.562.9250.8812.8930.272.4937.335.2229.6121.595.977.715.243.51
Symphonies [12]CVPR202444.1218.5830.021.855.9025.0712.068.2054.9413.8332.766.9335.118.5838.3311.5214.019.5714.4411.28
CGFormer[43]NIPS202448.0720.0529.853.423.9617.596.706.6363.8517.1540.725.5342.738.2238.8024.0416.2417.4510.186.77
OursICCV202549.1220.9530.771.916.6017.998.728.6764.3518.8341.534.4843.879.0739.7524.9416.5220.6610.307.99
+ +Table 2. Quantitative results on the test set of SSCBench-KITTI-360 [21, 25], where the highest and second-highest scores for each metric are highlighted in bold and underline, respectively. + +
MethodParams(M)↓Memory(M)↓Times(s)↓
Monoscene[3]149.519,0410.49
CGFormer[43]122.419,3300.41
Ours127.822,5970.34
+ +Table 3. Computational Complexity and Memory Usage + +methods. Notably, VisHall3D surpasses the latest stereo method StereoScene [14], as well as the top-performing temporal method HTCL [15] in terms of both IoU and mIoU. This demonstrates the effectiveness of our decoupled two-stage framework in capturing both the visible and invisible regions of the scene. + +On the SSCBench-KITTI-360 dataset, VisHall3D sets a new state-of-the-art with an mIoU of $20.95\%$ , outperforming the previous best method CGFormer [43] by $0.90\%$ . Remarkably, VisHall3D even surpasses early LiDAR-based methods such as SCSCNet [34] and LMSCNet [32], showcasing the potential of monocular methods in capturing complex 3D scenes. + +Fig. 5 presents a qualitative comparison of VisHall3D with MonoScene [3] and the ground truth on the SemanticKITTI dataset. VisHall3D generates more accurate and complete scene reconstructions, especially in the invisible regions occluded by foreground objects. This visual comparison further validates the superiority of our method in capturing the global scene context and hallucinating plausible geometries. + +We also compare VisHall3D against SOTA methods in terms of computational complexity in Tab. 3. Despite our two-stage framework, VisHall3D maintains parameters comparable to CGFormer, while outperforming it in mIoU + +and IoU. The slight memory increase is justified by performance gains. Under identical conditions, VisHall3D runs faster than CGFormer and Monoscene while delivering superior reconstruction quality. + +# 4.3. Ablation Studies + +In line with other works [3, 12, 20, 43], our ablation studies are primarily conducted on the SemanticKITTI [1] validation set. These studies encompass three key aspects: (1) the overall architectural components, (2) the division of visible and invisible regions, and (3) the noise incorporated in the OcclusionMAE. + +Overall Architectural Components. Tab. 4 presents our component-wise analysis of the network architecture. The baseline model retains the backbone and neck but adopts a direct hard-assign lifting approach from [12, 20], where each pixel corresponds to a single voxel. Its architecture can be considered as a variation of Symphonies [12] where the Symphonies Decoder is replaced with a 3D UNet. + +The introduction of unsigned distance function for the visual frontier not only enhances the mIoU but also significantly improves the IoU by $4.33\%$ , likely due to its superior preservation of geometric information for the visual frontier in the predicted depth map. The unsigned distance function provides a more informative representation of the visual frontier compared to binary masks, enabling the network to better capture the visibility relationships between voxels. Subsequently, by incorporating the OcclusionMAE, we divided the network into two parts: the first part maintains the original structure but predicts voxels for visible regions only, while the second part, the OcclusionMAE, hal + +
MethodIoU↑mIoU↑Params (M)Memory (M)
Baseline42.1114.5657.215260
VisFrontierNet w/o Feature Refinement46.44 (+4.33)15.11 (+0.55)74.317349
+ OcclusionMAE w/o Denoising46.38 (-0.06)15.99 (+0.88)86.718746
+ Feature Refinement45.88 (-0.50)16.59 (+0.60)125.521246
+ Denoising46.14 (+0.26)17.06 (+0.47)127.822597
+ +lucinates the entire scene, including invisible regions. Although this does not improve the IoU, it substantially boosts the mIoU by $0.88\%$ . + +Further, we integrated the feature refinement based on 3D deformable attention to enhance the model's perception of visible regions. While this results in a slight decrease in IoU by $0.50\%$ , the mIoU continues to improve by $0.60\%$ , reaching 16.59. Finally, to enhance the capability of the OcclusionMAE, we introduced a denoising strategy, which further improves the mIoU by $0.47\%$ . + +Interestingly, VisFrontierNet without feature refinement achieves the highest IoU, showing that its unsigned distance function alone offers a strong geometric prior for accurate visible region prediction. + +Table 4. Ablation study of the architectural components on the validation set of SemanticKITTI [1], where the model initially used the hard assignment method from [12, 20] for lifting 2D features to 3D. + +
InvisibilityThreshold θd(m)IoU↑mIoU↑
OOV-43.9815.89
OOV + Occ.1.545.8716.65
OOV + Occ.2.545.8316.92
OOV + Occ.3.546.1417.06
OOV + Occ.4.545.9416.95
+ +Visibility Identification. Decoupling visible and invisible scenes is one of our most critical components, and the key lies in how to delineate what constitutes the visible and invisible regions. Tab. 5 presents our ablation study on the division of visible and invisible regions. + +The baseline approach solely relies on the out of view checking to distinguish visible from invisible areas, considering only whether voxels project onto the image plane while entirely ignoring occlusion issues. As shown in Tab. 5, this naive implementation performs significantly worse than methods that account for occlusion. In occlusion-aware methods, we fine-tuned the threshold and found that $3.5\mathrm{m}$ gives the best results, suggesting moderate visibility relaxation helps handle occlusion and depth uncertainty. However, larger thresholds (e.g., $4.5\mathrm{m}$ ) introduce + +noise and harm performance. + +Table 5. Ablation study on visible and invisible region division on SemanticKITTI [1]'s validation set, where OOV and Occ. refer to using the out of view and occlusion respectively. + +
Rh(Voxel)Rd(Voxel)IoU↑mIoU↑
0046.0316.32
0245.9216.79
0345.8316.92
0445.7316.66
1346.1916.52
1446.1216.69
+ +Table 6. Ablation study on the noise level in OcclusionMAE on SemanticKITTI [1], where $R_{h}$ and $R_{d}$ refer to the noise range in the horizontal and depth directions, respectively (conducted with the threshold $\theta_{d}$ of $2.5\mathrm{m}$ ). + +Noise in the OcclusionMAE. To achieve effective hallucination, we introduce a certain level of noise into the OcclusionMAE, which also enhances the network's robustness. Tab. 6 presents our study on the noise parameters in the Hallucinating MAE. These parameters consist of two components: the maximum sampling distance in the horizontal direction and the depth direction. When no noise is applied (both parameters are set to 0), the model performs worst on mIoU. In contrast, the best mIoU is achieved when the horizontal noise is set to 0 and the depth noise to 3. + +Interestingly, the impact of increasing horizontal noise on the network is significantly greater than that of depth noise, likely due to the inherent inaccuracies in depth estimation. This finding suggests that the model is more sensitive to the perturbation in the horizontal direction, and a careful balance between the two noise components is crucial for optimal performance. + +# 5. Conclusion + +In this paper, we propose VisHall3D, a two-stage framework for monocular semantic scene completion that tackles feature entanglement and geometric inconsistency by explicitly separating visible-region reconstruction and invisible-region hallucination, and demonstrate its effectiveness with extensive experiments. + +# Acknowledgements + +This paper is supported by the National Natural Science Foundation of China (No.62473306, N0.U24B20181, No.62441616) and Science and Technology Research and Development Plan of China State Railway Group Co., Ltd. (No. RITS2023KF03). + +# References + +[1] Jens Behley, Martin Garbade, Andres Milioto, Jan Quenzel, Sven Behnke, Cyril Stachniss, and Jurgen Gall. Semantickitti: A dataset for semantic scene understanding of lidar sequences. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9297-9307, 2019. 1, 5, 6, 7, 8 +[2] Zhaowei Cai and Nuno Vasconcelos. Cascade r-cnn: High quality object detection and instance segmentation. IEEE transactions on pattern analysis and machine intelligence, 43(5):1483-1498, 2019. 3 +[3] Anh-Quan Cao and Raoul De Charette. Monoscene: Monocular 3d semantic scene completion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3991–4001, 2022. 1, 2, 3, 5, 6, 7 +[4] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 3 +[5] Hansheng Chen, Pichao Wang, Fan Wang, Wei Tian, Lu Xiong, and Hao Li. Epro-pnp: Generalized end-to-end probabilistic perspective-n-points for monocular object pose estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2781-2790, 2022. 2 +[6] Ran Cheng, Christopher Agia, Yuan Ren, Xinhai Li, and Liu Bingbing. S3cnet: A sparse semantic scene completion network for lidar point clouds. In Conference on Robot Learning, pages 2148-2161. PMLR, 2021. 2 +[7] Ross Girshick. Fast r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 1440-1448, 2015. 4 +[8] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 4 +[9] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 3 +[10] Xun Huang and Serge Belongie. Arbitrary style transfer in real-time with adaptive instance normalization. In Proceedings of the IEEE international conference on computer vision, pages 1501-1510, 2017. 5 +[11] Yuanhui Huang, Wenzhao Zheng, Yunpeng Zhang, Jie Zhou, and Jiwen Lu. Tri-perspective view for vision-based 3d semantic occupancy prediction. In Proceedings of + +the IEEE/CVF conference on computer vision and pattern recognition, pages 9223-9232, 2023. 2, 6, 7 +[12] Haoyi Jiang, Tianheng Cheng, Naiyu Gao, Haoyang Zhang, Tianwei Lin, Wenyu Liu, and Xinggang Wang. Symphonize 3d semantic scene completion with contextual instance queries. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20258-20267, 2024. 1, 2, 4, 6, 7, 8 +[13] Mykola Lavreniuk. Spidepth: Strengthened pose information for self-supervised monocular depth estimation. arXiv preprint arXiv:2404.12501, 2024. 2 +[14] Bohan Li, Yasheng Sun, Zhujin Liang, Dalong Du, Zhuanghui Zhang, Xiaofeng Wang, Yunnan Wang, Xin Jin, and Wenjun Zeng. Bridging stereo geometry and bev representation with reliable mutual interaction for semantic scene completion. arXiv preprint arXiv:2303.13959, 2023. 2, 6, 7 +[15] Bohan Li, Jiajun Deng, Wenyao Zhang, Zhujin Liang, Dalong Du, Xin Jin, and Wenjun Zeng. Hierarchical temporal context learning for camera-based semantic scene completion. arXiv preprint arXiv:2407.02077, 2024. 6, 7 +[16] Feng Li, Hao Zhang, Huaizhe Xu, Shilong Liu, Lei Zhang, Lionel M Ni, and Heung-Yeung Shum. Mask dino: Towards a unified transformer-based framework for object detection and segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3041-3050, 2023. 4 +[17] Hongyang Li, Hao Zhang, Zhaoyang Zeng, Shilong Liu, Feng Li, Tianhe Ren, and Lei Zhang. Dfa3d: 3d deformable attention for 2d-to-3d feature lifting. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6684–6693, 2023. 5 +[18] Jie Li, Yu Liu, Xia Yuan, Chunxia Zhao, Roland Siegwart, Ian Reid, and Cesar Cadena. Depth based semantic scene completion with position importance aware loss. IEEE Robotics and Automation Letters, 5(1):219-226, 2019. 2 +[19] Jie Li, Kai Han, Peng Wang, Yu Liu, and Xia Yuan. Anisotropic convolutional networks for 3d semantic scene completion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3351-3359, 2020. 2 +[20] Yiming Li, Zhiding Yu, Christopher Choy, Chaowei Xiao, Jose M Alvarez, Sanja Fidler, Chen Feng, and Anima Anandkumar. Voxformer: Sparse voxel transformer for camera-based 3d semantic scene completion. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9087-9098, 2023. 1, 2, 4, 6, 7, 8 +[21] Yiming Li, Sihang Li, Xinhao Liu, Moonjun Gong, Kenan Li, Nuo Chen, Zijun Wang, Zhiheng Li, Tao Jiang, Fisher Yu, Yue Wang, Hang Zhao, Zhiding Yu, and Chen Feng. Sscbench: A large-scale 3d semantic scene completion benchmark for autonomous driving. In 2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2024. 1, 6, 7 +[22] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Yu Qiao, and Jifeng Dai. Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. 2 + +[23] Zhuoling Li, Zhan Qu, Yang Zhou, Jianzhuang Liu, Haoqian Wang, and Lihui Jiang. Diversity matters: Fully exploiting depth clues for reliable monocular 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2791-2800, 2022. 2 +[24] Zhiqi Li, Zhiding Yu, David Austin, Mingsheng Fang, Shiyi Lan, Jan Kautz, and Jose M Alvarez. Fb-occ: 3d occupancy prediction based on forward-backward view transformation. arXiv preprint arXiv:2307.01492, 2023. 2 +[25] Yiyi Liao, Jun Xie, and Andreas Geiger. Kitti-360: A novel dataset and benchmarks for urban scene understanding in 2d and 3d. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(3):3292-3310, 2022. 1, 6, 7 +[26] Jian Liu, Sipeng Zhang, Chuixin Kong, Wenyuan Zhang, Yuhang Wu, Yikang Ding, Borun Xu, Ruibo Ming, Donglai Wei, and Xianming Liu. Occtransformer: Improving befoformer for 3d camera-only occupancy prediction. arXiv preprint arXiv:2402.18140, 2024. 2 +[27] Zongdai Liu, Dingfu Zhou, Feixiang Lu, Jin Fang, and Liangjun Zhang. Autoshape: Real-time shape-aware monocular 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15641-15650, 2021. 2 +[28] Shujie Luo, Hang Dai, Ling Shao, and Yong Ding. M3dssd: Monocular 3d single stage object detector. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6145-6154, 2021. 2 +[29] Ruihang Miao, Weizhou Liu, Mingrui Chen, Zheng Gong, Weixin Xu, Chen Hu, and Shuchang Zhou. Occdepth: A depth-aware method for 3d semantic scene completion. arXiv preprint arXiv:2302.13540, 2023. 1 +[30] Jonah Philion and Sanja Fidler. Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In Proceedings of the European Conference on Computer Vision, 2020. 5 +[31] Christoph B Rist, David Emmerichs, Markus Enzweiler, and Dariu M Gavrila. Semantic scene completion using local deep implicit functions on lidar data. IEEE transactions on pattern analysis and machine intelligence, 44(10):7205-7218, 2021. 2 +[32] Luis Roldao, Raoul De Charette, and Anne Verroust-Blondet. Lmsnet: Lightweight multiscale 3d semantic completion. In 2020 International Conference on 3D Vision (3DV), pages 111–119. IEEE, 2020. 7 +[33] Faranak Shamsafar, Samuel Woerz,afia Rahim, and Andreas Zell. Mobilestereonet: Towards lightweight deep networks for stereo matching. In Proceedings of the IEEE/cvf winter conference on applications of computer vision, pages 2417-2426, 2022. 6 +[34] Shuran Song, Fisher Yu, Andy Zeng, Angel X Chang, Manolis Savva, and Thomas Funkhouser. Semantic scene completion from a single depth image. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1746-1754, 2017. 2, 7 +[35] Peize Sun, Rufeng Zhang, Yi Jiang, Tao Kong, Chenfeng Xu, Wei Zhan, Masayoshi Tomizuka, Lei Li, Zehuan Yuan, Changhu Wang, et al. Sparse r-cnn: End-to-end object detection with learnable proposals. In Proceedings of + +the IEEE/CVF conference on computer vision and pattern recognition, pages 14454-14463, 2021. 3 +[36] Song Wang, Jiawei Yu, Wentong Li, Wenyu Liu, Xiaolu Liu, Junbo Chen, and Jianke Zhu. Not all voxels are equal: Hardness-aware semantic scene completion with self-distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14792-14801, 2024. 3, 6 +[37] Yu Wang and Chao Tong. H2gformer: Horizontal-to-global voxel transformer for 3d semantic scene completion. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 5722-5730, 2024. 1, 2, 6 +[38] Yi Wei, Linqing Zhao, Wenzhao Zheng, Zheng Zhu, Jie Zhou, and Jiwen Lu. Surroundocc: Multi-camera 3d occupancy prediction for autonomous driving. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 21729-21740, 2023. 6 +[39] Haihong Xiao, Hongbin Xu, Wenxiong Kang, and Yuqiong Li. Instance-aware monocular 3d semantic scene completion. IEEE Transactions on Intelligent Transportation Systems, 25(7):6543-6554, 2024. 6, 7 +[40] Xu Yan, Jiantao Gao, Jie Li, Ruimao Zhang, Zhen Li, Rui Huang, and Shuguang Cui. Sparse single sweep lidar point cloud segmentation via learning contextual shape priors from scene completion. In Proceedings of the AAAI conference on artificial intelligence, pages 3101-3109, 2021. 2 +[41] Jiawei Yao and Jusheng Zhang. Depthssc: Depth-spatial alignment and dynamic voxel resolution for monocular 3d semantic scene completion. arXiv preprint arXiv:2311.17084, 2023. 2, 6, 7 +[42] Jiawei Yao, Chuming Li, Keqiang Sun, Yingjie Cai, Hao Li, Wanli Ouyang, and Hongsheng Li. Ndc-scene: Boost monocular 3d semantic scene completion in normalized device coordinates space. In 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 9421-9431. IEEE Computer Society, 2023. 1, 2 +[43] Zhu Yu, Runmin Zhang, Jiacheng Ying, Junchen Yu, Xiaohai Hu, Lun Luo, Si-Yuan Cao, and Hui-liang Shen. Context and geometry aware voxel transformer for semantic scene completion. In Advances in Neural Information Processing Systems, pages 1531-1555, 2024. 1, 2, 6, 7 +[44] Renrui Zhang, Han Qiu, Tai Wang, Ziyu Guo, Ziteng Cui, Yu Qiao, Hongsheng Li, and Peng Gao. Monodetr: Depth-guided transformer for monocular 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9155-9166, 2023. 2 +[45] Yunpeng Zhang, Jiwen Lu, and Jie Zhou. Objects are different: Flexible monocular 3d object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3289-3298, 2021. 2 +[46] Yunpeng Zhang, Zheng Zhu, and Dalong Du. Occformer: Dual-path transformer for vision-based 3d semantic occupancy prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9433-9443, 2023. 6, 7 +[47] Yupeng Zheng, Xiang Li, Pengfei Li, Yuhang Zheng, Bu Jin, Chengliang Zhong, Xiaoxiao Long, Hao Zhao, and Qichao + +Zhang. Monoocc: Digging into monocular semantic occupancy prediction. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 18398-18405. IEEE, 2024. 1, 3, 6 +[48] Xingyi Zhou, Dequan Wang, and Philipp Krahenbuhl. Objects as points. arXiv preprint arXiv:1904.07850, 2019. 2 +[49] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020.3, 5 \ No newline at end of file diff --git a/ICCV/2025/VisHall3D_ Monocular Semantic Scene Completion from Reconstructing the Visible Regions to Hallucinating the Invisible Regions/images.zip b/ICCV/2025/VisHall3D_ Monocular Semantic Scene Completion from Reconstructing the Visible Regions to Hallucinating the Invisible Regions/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..d0ef5a0e32879117fa67321c2e347ccded37eab6 --- /dev/null +++ b/ICCV/2025/VisHall3D_ Monocular Semantic Scene Completion from Reconstructing the Visible Regions to Hallucinating the Invisible Regions/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:54b18dba3aa6b32fd10f59dccf430b0e64beb952f8a014de640cced11d941426 +size 946045 diff --git a/ICCV/2025/VisHall3D_ Monocular Semantic Scene Completion from Reconstructing the Visible Regions to Hallucinating the Invisible Regions/layout.json b/ICCV/2025/VisHall3D_ Monocular Semantic Scene Completion from Reconstructing the Visible Regions to Hallucinating the Invisible Regions/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..9802b3ab9f33cb4f0e3c22778e86c96c0bbb1628 --- /dev/null +++ b/ICCV/2025/VisHall3D_ Monocular Semantic Scene Completion from Reconstructing the Visible Regions to Hallucinating the Invisible Regions/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ba5c535e832c1330d59a59fb30f8524ecfb2d9fb569267281ae25adf9be23f29 +size 374280 diff --git a/ICCV/2025/VisNumBench_ Evaluating Number Sense of Multimodal Large Language Models/27f14a06-5371-423d-833b-707308fc0b78_content_list.json b/ICCV/2025/VisNumBench_ Evaluating Number Sense of Multimodal Large Language Models/27f14a06-5371-423d-833b-707308fc0b78_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..bd0d9771bc08406a919ca20b2422aaefe20b5e7b --- /dev/null +++ b/ICCV/2025/VisNumBench_ Evaluating Number Sense of Multimodal Large Language Models/27f14a06-5371-423d-833b-707308fc0b78_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:87aaf83470504c3e797ab7b45fb8f0c5334f2981da7ceaa447244122eabe1625 +size 85744 diff --git a/ICCV/2025/VisNumBench_ Evaluating Number Sense of Multimodal Large Language Models/27f14a06-5371-423d-833b-707308fc0b78_model.json b/ICCV/2025/VisNumBench_ Evaluating Number Sense of Multimodal Large Language Models/27f14a06-5371-423d-833b-707308fc0b78_model.json new file mode 100644 index 0000000000000000000000000000000000000000..8783389575162847825d3ae321d54d6f5f425c24 --- /dev/null +++ b/ICCV/2025/VisNumBench_ Evaluating Number Sense of Multimodal Large Language Models/27f14a06-5371-423d-833b-707308fc0b78_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:80f748221ccc6264be0b3d61c4e2b028e05bda22b5948c4b908f4f82e09de3ed +size 126384 diff --git a/ICCV/2025/VisNumBench_ Evaluating Number Sense of Multimodal Large Language Models/27f14a06-5371-423d-833b-707308fc0b78_origin.pdf b/ICCV/2025/VisNumBench_ Evaluating Number Sense of Multimodal Large Language Models/27f14a06-5371-423d-833b-707308fc0b78_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..36f20af82f361c13070273efb49d1e86c90ab695 --- /dev/null +++ b/ICCV/2025/VisNumBench_ Evaluating Number Sense of Multimodal Large Language Models/27f14a06-5371-423d-833b-707308fc0b78_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c88532ab36e9ff1421488aca98767bf3bfbc3d56325b9a0d1504f849ff3d486c +size 4646681 diff --git a/ICCV/2025/VisNumBench_ Evaluating Number Sense of Multimodal Large Language Models/full.md b/ICCV/2025/VisNumBench_ Evaluating Number Sense of Multimodal Large Language Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..510fcfe2af12d9e2cb3e49b19ee42b88e0b8a200 --- /dev/null +++ b/ICCV/2025/VisNumBench_ Evaluating Number Sense of Multimodal Large Language Models/full.md @@ -0,0 +1,454 @@ +# VisNumBench: Evaluating Number Sense of Multimodal Large Language Models + +Tengjin Weng $^{1,2}$ , Jingyi Wang $^{3}$ , Wenhao Jiang $^{2*}$ , Zhong Ming $^{1,2,4*}$ + +1College of Computer Science and Software Engineering, Shenzhen University + +$^{2}$ Guangdong Laboratory of Artificial Intelligence and Digital Economy (SZ) + +3Shenzhen International Graduate School, Tsinghua University + +4Shenzhen Technology University + +wtjdsb@gmail.com, jingyi-w24@mails.tsinghua.edu.cn, cswhjiang@gmail.com, mingz@szu.edu.cn + +# Abstract + +Can Multimodal Large Language Models (MLLMs) develop an intuitive number sense similar to humans? Targeting this problem, we introduce Visual Number Benchmark (VisNumBench) to evaluate the number sense abilities of MLLMs across a wide range of visual numerical tasks. VisNumBench consists of about 1,900 multiple-choice question-answer pairs derived from both synthetic and real-world visual data, covering seven visual numerical attributes and four types of visual numerical estimation tasks. Our experiments on VisNumBench led to the following key findings: (i) The 17 MLLMs we tested—including open-source models such as Qwen2.5-VL and InternVL2.5, as well as proprietary models like GPT-4o and Gemini 2.0 Flash—perform significantly below human levels in number sense-related tasks. (ii) Multimodal mathematical models and multimodal chain-of-thought (CoT) models did not exhibit significant improvements in number sense abilities. (iii) Stronger MLLMs with larger parameter sizes and broader general abilities demonstrate modest gains in number sense abilities. We believe VisNumBench will serve as a valuable resource for the research community, encouraging further advancements in enhancing MLLMs' number sense abilities. Code and dataset are available at https://wwwwttjjj.github.io/VisNumBench/. + +# 1. Introduction + +Number sense is an innate cognitive ability shared by both humans and animals through the approximate number system [14]. It enables individuals to perceive, process, and manipulate numerical information intuitively. By fostering a deeper understanding of abstract number concepts, it facilitates the grasp of complex mathematical theories and their practical application in real-world scenarios. Figure 1 illus + +![](images/54e2ce99daf85f648464938b3bed9b38be50d955e9d45a83917d751c105f4ff5.jpg) +Figure 1. Explanations of number sense: how humans intuitively perceive and estimate values of angle, quantity, length, and scale. + +trates the human ability to perceive and estimate numerical quantities. For instance, a person can quickly recognize a group of five bananas and intuitively estimate that there are about two more groups of the same size. By leveraging this innate number sense and grouping strategy, one can infer that the total number of bananas is approximately fifteen. + +Multimodal Large Language Models (MLLMs) have made remarkable strides in tackling complex multimodal tasks [2, 7, 24, 27]. Recent research has focused on enhancing their mathematical and scientific reasoning capabilities by incorporating external tools [31, 50]. To assess these abilities, numerous benchmarks [10, 18, 20, 29, 30, 32] have been developed to evaluate the performance of MLLMs on mathematical reasoning and numerical interpretation tasks. While existing benchmarks effectively assess structured numerical reasoning problems, they primarily emphasize abstract symbolic computation, mathematical problem-solving, or interpreting numerical data in textual contexts. However, these evaluations overlook a critical aspect of human-like numerical cognition: intuitive number sense. + +# VisNumBench-Synthetic + +Angle Perception +(Range Estimation) +The estimated angle +between two lines? +(d) (60, 70) degrees +![](images/c89c4aa7659b625b4814ba0cfc4f200898df7210d1457719075f1bceb7830ff3.jpg) +(a) (90, 100) degrees + +![](images/659246b9425b3c2260a97c275daa486a849431d5e971a91536ba64f3fc25e1b6.jpg) +Scale Perception +(Range Estimation) + +The estimated value +(d) (40,50) +![](images/4f8842b49718652a2dfe935cd4e0e96097b87c617a49bc7aabb339e70dfbef63.jpg) +(b) (30, 40) degrees +(a) (10, 20) +(b) (0, 10) +(c) (30, 40) + +![](images/b4ba77050e2c73b3fc5eb6b1b1f9024f8fbcc0143f4b48e957f056a4ad6f77fb.jpg) +Length Perception +(Multiplicative Estimation) + +![](images/1655a1e0cb00046e65db5ff14818ff31e3bd3d8000797157d4d0479cb5d0621b.jpg) +line A is X times larger + +![](images/6c722918b7d1dae5d1dc4c65b4165cd11b99a718317a92074bf29317243d7b57.jpg) + +![](images/72252c78b104b61ad4d36562dca80a21527f64c1d1c94a135decf9cce5770440.jpg) + +![](images/a89db02c0c1cfeda69121c56f24a40d878ffa226bded7947e18b951a658e6cc4.jpg) +(c) 28 +(d) 20 + +(d) 4.0 +![](images/7fdcfb47960a0c360ccdb79fb297bf18d8315f5331bc3e35f78fc30e5ef6d8e4.jpg) +(b) 1.3 + +Quantity Perception +(Value Estimation) +The number of +cuboids in the figure? +![](images/7ea445bfcd244dd6754e8c457c7e33848adcd3e3f7a403635d57be3ba64561c6.jpg) +(a) 42 +(b) 50 + +Depth Perception +(Value Comparison) +The order of points from closest to +Area Perception +(Value Estimation) +The area of shape right is 4, +the area of shape left? +![](images/d3e3511c6a9c6f6ba141c0a0905148a4c833ef67d4131d029df9a723492bee19.jpg) +(a) 17 +(b) 6 +(c) 2 +(d) 10 + +![](images/8c4d4dc1446ed92cad19ed3078c62f909d1980610f637ca61969f2402d5c590f.jpg) +farthest for the view + +![](images/e93c8229e3affbd8f183211717a07189dcdc77a11c47b1420eef11404faeb186.jpg) + +![](images/b01bf66e6d1403418f7228e3335df1fa79656e812e653e2a81f634f75a9b3a9b.jpg) + +![](images/43aa46bced2d9f7f4fac7807fb6e3f5d8198467b291afbf4559bae343ccd8756.jpg) + +![](images/b8e9601ce8288775441bce627af636b96b42f3cc8251a82687164638ab29230e.jpg) +(c) BCAD +DBCA + +![](images/3aa725c91483ba2c9e97b8a451de275b71a13b6e976cac7c4bc1d66423eac1ee.jpg) + +![](images/c521c95b29ee461f55abdc5bcf9dea1e59aad32450211b730d73b38d2ed32849.jpg) + +# VisNumBench-Real + +Angle Perception +(Value Estimation) +The estimated opening angle of the laptop? +![](images/4b1186e2f115a75a54bc1c7ba6d0071bce32ab42204a7a5cc72bbea5e4485766.jpg) +(c) (170, 180) degrees + +![](images/e0070766ef25299f09730e7356ce706dfe7df47a01ebcd04a0ad66ae11807128.jpg) +Scale Perception +(Value Estimation) +The estimated time of clock is pointing to? + +![](images/ba2a7977eab9431a6185dd0d8b0cec80ce1168bd5af15d38cbf989550fd7bb26.jpg) +Length Perception +(Range Estimation) +The estimated ratio of the water +ght to the total height of the cup? + +![](images/713b58e2f57a05382daf0756d9bea44ee22d8580f92fc59c75b90de14f5a6d7b.jpg) +Quantity Perception +(Value Comparison) +Which color of flowers appears the most? + +![](images/5fc901ad8f3805bff4b52aae450fecd2297662ad9e2b2aac937e50b33caa146b.jpg) +Depth Perception +(Multiplicative Estimation) + +![](images/1ca3badf18ae823378a6611b87e85c6dfb0162ab900061809eb2ac4cdedff7c4.jpg) + +![](images/455bdbf39309d46a9d26bbc86b9f0daf91ca1554f673a3ead99e33b5be8235e0.jpg) + +![](images/e27f64743b4172d43c3d5e29aa0061542b3bb74dbbf5f13af4f91904476ce4b8.jpg) +(c) 4.3 + +![](images/c7fc5475aa67197d68a1884d651e92db614a5a1a8e22617256dc241c561b7175.jpg) +(d) Object D + +![](images/ef1d3a0bccca690a31face69f41133edbc933f7a8a427c64da8d774bf43d63ca.jpg) +Qwen-2.5-72B +Figure 2. Examples from VisNumBench and responses from MLLMs. VisNumBench is divided into two subsets: VisNumBench-Synthetic and VisNumBench-Real. It focuses on seven key visual numerical attributes: angle, scale, length, quantity, depth, area, and volume. Even state-of-the-art MLLMs often struggle to answer the questions in VisNumBench accurately. + +![](images/c5b6dc355dcf6160c1f27a27282f7f9bc2b2e89d824042f7ae0dbffae4911719.jpg) +(d) 6.0 +InternVL-2.5-78B + +The distance from point A to the viewer is X times than point B? +Volume Perception +(Value Comparison) +Which object has the largest +volume? +![](images/61de78a42334de471f271c06a3f0b872d5619df89af4e73b1aa71fa891a0c9c5.jpg) +(a) Object A +(b) Object B +(c) Object C + +# Range Estimation + +Which of the following ranges is a + +reasonable estimate for Angle A? + +(a) (150, 160) degrees +(b) (50, 60) degrees + +(b) (50, 60) degrees (c) (90, 100) degrees + +(d)(10,29) degrees. + +# Value Estimation + +Which of the following options is a + +reasonable estimate of the Angle B? + +(a)10 degrees + +(b)112 degrees (c) 50 degrees + +(c)60 degrees (d)30 degrees + +(a) So degrees + +![](images/dccde4829e56f7c54b32b150362bfbb3f1e8cd3ed58006f80bb3cb244081367f.jpg) +There are four angles in +the figure: Angle A, Angle B. +B, Angle C, Angle D. + +# Value Comparison + +Which angle is the largest among angles +A, B, C, and D? +(a) Angle A +(b) Angle B +(c) Angle C +(d) Angle D + +# Multiplicative Estimation + +Angle C is X times larger than Angel D. +Which of the following options is a +reasonable estimate for $x?$ +(a) 1.3 +(6)1.9 +(d)4.3 + +Unlike humans, who effortlessly estimate quantities, perceive proportions, and grasp numerical relationships at a glance, MLLMs often depend on explicit reasoning steps rather than perceptual intuition. This limitation raises fundamental questions about whether current models genuinely comprehend numerical concepts or merely manipulate them based on learned patterns in text and images. + +In this work, we introduce the Visual Number Benchmark (VisNumBench), inspired by human number sense abilities. As illustrated in Figure 2, VisNumBench is struc + +Figure 3. Illustrations of four distinct visual numerical estimation tasks: range estimation, value comparison, value estimation, and multiplicative estimation. +Results on VisNumBench-Synthetic +Figure 4. Evaluation results of MLLMs on the VisNumBench. The performance of MLLMs on VisNumBench is significantly poor in terms of accuracy, whereas human performance is nearly perfect. +![](images/25920f2779014d2f5e62179c2eda2f242cd9f0dc748e839425e02aa1aa619b7b.jpg) +Random +LLaV +—LLaV +InternVL2-5-78B + +Results on VisNumBench-Real +![](images/e7f88ba0b0eb9343b2a07459380a462062ce36bb4396c924f2f57741d8795b8b.jpg) +Qwe +Gemin +GPT-4 +Human + +tured into two components based on different visual scenarios: VisNumBench-Synthetic and VisNumBench-Real. VisNumBench-Synthetic comprises controlled, synthetic images in which numerical relationships are explicitly defined. VisNumBench-Real contains real-world images, providing a more complex and less controlled environment. VisNumBench targets seven key dimensions of visual nu + +Table 1. Dataset statistics of VisNumBench based on various visual numerical attributes. + +
VisNumBenchAngleLengthScaleQuantityDepthAreaVolumeTotal
VisNumBench-Synthetic170181140196135189-1011
VisNumBench-Real149162143147154-147902
Answer Format4/5 options3/4 options4 options3/4 options4 options4/5 options3/4 options3/4/5 options
+ +merical attributes through four distinct types of visual numerical estimation tasks, as depicted in Figure 3. + +We evaluated 17 MLLMs on VisNumBench and found that even state-of-the-art models perform poorly on our proposed benchmark, which is shown in Figure 4. Furthermore, our experiments reveal that adopting multimodal mathematical models and multimodal Chain-of-Thought (CoT) models did not lead to substantial performance improvements. However, the performance of the latest models is better than the previous models from the same family. For example, Qwen2.5VL [42] performs better than Qwen2VL [45]. It seems that optimization on data, training techniques, and model architecture will help models improve their number sense ability. In this work, we aim to advance MLLMs toward higher levels of intelligence by developing models that enhance visual number sense abilities. The main contributions of this paper are listed as follows: + +1. We introduce VisNumBench, a comprehensive benchmark integrating diverse data sources and an automated evaluation framework to assess the numerical sense abilities of MLLMs across various visual numerical tasks. +2. We conduct a comprehensive evaluation of various MLLMs on VisNumBench, finding that even the most advanced models still exhibit limited numerical sense. +3. Further experiments on historical models from the same family show that their numerical sense abilities have improved over time. To enhance this ability within a short period, more specialized optimizations in data, training techniques, and model architecture may be required. + +# 2. Related Work + +# 2.1. Multimodal Large Language Models + +Recent advancements in MLLMs have demonstrated exceptional capabilities across a wide range of tasks. Leveraging multimodal pre-training, MLLMs have achieved outstanding performance in both open-source models [1, 4, 5, 46, 52] and proprietary models [3, 35, 37]. Consequently, these models have been widely adopted in various domains, including mathematical reasoning [56], chart understanding [33], medical image analysis [23], and text-rich image comprehension [58]. Their growing success has spurred the development of an increasing number of benchmarks to assess performance across diverse visual and linguistic tasks. + +# 2.2. Benchmarks for MLLMs + +Advancements in MLLMs have led to the development of numerous benchmarks aimed at evaluating model performance across a broad spectrum of general multimodal tasks. Several recent studies [15, 16, 20-22, 28, 53, 54] have introduced more comprehensive reasoning and perception multimodal benchmarks that provide extensive and holistic assessments. Beyond general multimodal tasks, specialized benchmarks have been created to evaluate the mathematical reasoning capabilities of MLLMs. Tasks such as visual reasoning with numbers, arithmetic problem-solving, and algebraic manipulation play a crucial role in assessing both the numerical proficiency and higher-order cognitive abilities of MLLMs. Prominent benchmarks, including MATHVISTA [30], Math-Vision [44], MathOdyssey [13], and SMART-840 [9], are designed to test models on a diverse range of mathematical challenges, such as word problems, equation-solving, and complex multi-step reasoning. + +These benchmarks aim to assess the ability of models to understand and process mathematical content in both images and text, as well as their ability to apply mathematical operations in reasoning contexts. However, evaluating MLLMs is crucial not only for measuring their proficiency in traditional mathematical reasoning but also for understanding their ability to handle more tangible, real-world dependent mathematical perception tasks. Humans typically acquire basic mathematical knowledge through intuitive number sense, which they then apply to real-world scenarios. This intuitive understanding of numbers and their relationships is a fundamental aspect of human cognition, enabling the seamless application of mathematical concepts in everyday life. While MLLMs can process complex mathematical problems, they may struggle with tasks that require this intuitive number sense. Therefore, existing benchmarks should not only evaluate the proficiency of models in traditional mathematical reasoning but also assess their ability to apply mathematical concepts in real-world contexts. This would help bridge the gap between abstract problem-solving and practical application. + +# 2.3. Number Sense of MLLMs + +In the context of MLLMs, previous research [11, 25, 47] has focused on tasks that rely on ordinal regression to evaluate number sense. Examples of such tasks include Age + +![](images/cc03952f6266c2dea88dfba535e6e548a10d601bc2642f4050ff00a4a6530f07.jpg) +Figure 5. An illustration of the construction steps for images and questions in VisNumBench-Synthetic and VisNumBench-Real. + +Estimation [38], Historical Image Dating [36], and Image Aesthetics Assessment [39]. These studies typically focus on estimating or ranking numerical attributes based on visual input, such as predicting the age of a person in an image or determining the historical context of a photograph. While these tasks assess the ability of the model to interpret numerical cues and sequences, they often focus on specific domains and may overlook the broader, more generalizable concept of number sense. + +Assessing the number sense of MLLMs refers to evaluating their number sense abilities in a variety of scenarios. This includes tasks such as interpreting measurements, performing approximate calculations, comparing quantities, and identifying numerical relationships in different contexts. Such evaluations provide a better understanding of how models perform on tasks that require flexible and contextual reasoning about numbers while also enhancing their broad applicability and intelligence. + +# 3. VisNumBench + +# 3.1. Overview + +We introduce VisNumBench, a benchmark specifically designed to directly evaluate the intuitive numerical abilities of MLLMs. Each instance in VisNumBench comprises a figure, a multiple-choice question, and a corresponding answer label. The dataset statistics of VisNumBench are presented in Table 1. + +Compared with previous benchmarks, VisNumBench has the following novel features: + +- Comprehensive Scenario Integration: VisNumBench incorporates both controlled synthetic figures and intricate real-world scenes, enabling a thorough evaluation of the number sense abilities of MLLMs. +- Multidimensional Visual Numerical Attributes: VisNumBench encompasses seven fundamental aspects of number sense—angle, length, scale, quantity, depth, area, and volume—ensuring a rigorous and comprehensive evaluation of the numerical capabilities of MLLMs. +- Comprehensive Visual Numerical Estimation Tasks: + +VisNumBench encompasses four distinct modes of visual numerical estimation—value comparison, value estimation, range estimation, and multiplicative estimation. These diverse tasks enable a thorough evaluation of MLLMs' ability to estimate numerical values across different visual numerical categories. + +# 3.2. Data Collection Process + +# 3.2.1. Source of Images + +The development of VisNumBench required gathering figures from diverse sources, as depicted in part (a) of Figure 5. + +- Python Program. We developed a series of Python scripts based on Matplotlib to generate figures by randomly sampling parameters, which are also stored for future use. This approach allows precise control over various numerical properties, ensuring a well-balanced data distribution and minimizing potential biases. +- Existing Datasets. To enhance the diversity of the proposed benchmark and leverage existing high-quality data, we incorporated figures from multiple well-established datasets [16, 20, 30, 41, 48, 57], covering a broad range of numerical and spatial perception scenarios. +- Web Collection. To incorporate more natural and diverse visual data, we collected figures with numerical information from public web sources [12, 17, 43]. These figures were carefully curated and filtered to ensure relevance and clarity before designing the corresponding questions. +- Images Taken by Us. To better reflect real-world conditions, we manually constructed various number sense scenarios and captured images using a camera. These scenes encompassed real-world measurements, physical counting tasks, and estimation challenges under natural lighting and occlusion conditions. + +# 3.2.2. QA Pair Construction and Quality Control + +As shown in parts (b) and (c) in Figure 5, we involve QA pairs for images from different sources. For the figures generated by the Python program, we manually design different questions based on the specific characteristics of each + +Table 2. Accuracies of MLLMs on the VisNumBench-Synthetic (%) dataset. Dark gray and light gray indicate the best and the second best results among all models, respectively. + +
AngleLengthScaleQuantityDepthAreaAverage
Random24.4425.4125.0025.0025.0023.6824.76
Open-source MLLMs
Phi-3.5-vision19.4140.8841.4326.5326.6739.1532.34
LLaVA-v1.5-7B31.1830.3934.2933.1626.6721.1629.38
LLaVA-v1.5-13B35.8830.9432.1436.7333.3324.3432.15
LLaVA-v1.6-34B40.0045.3040.0046.9464.4433.3344.31
LLaVA-Onevision-7B25.8851.3842.8638.7834.8144.4439.96
LLaVA-Onevision-72B24.7161.3362.1450.0048.1558.7350.84
InternVL2.5-8B26.4741.9949.2934.6941.4846.0339.66
InternVL2.5-38B39.4159.6759.2954.0860.7461.3855.59
InternVL2.5-78B35.2959.6768.5742.8661.4872.4956.18
Janus-Pro-7B31.7643.6545.7135.7133.3336.5137.69
Qwen2.5-VL-3B30.0049.1750.7132.1442.2251.8542.43
Qwen2.5-VL-7B23.5353.5955.0039.2948.8958.2046.19
Qwen2.5-VL-72B37.0659.6765.0057.6561.4870.3758.46
API-based models
GPT-4o35.2943.0954.2937.2454.0743.3943.72
Gemini 1.5 Flash26.4747.4144.4026.0223.7041.2733.33
Gemini 2.0 Flash31.1857.4681.4355.1051.1170.9057.57
Gemini 1.5 Pro34.1239.2347.1440.8258.5248.1544.02
Human90.0096.00100.0096.0098.0092.0095.33
+ +figure and generate corresponding annotations using the parameters saved. We can design the question such as: "In the ABCD sections of the pie chart, which section has the largest/smallest area?" Based on the parameters saved, the answer would be $\frac{C}{D}$ . For images from other sources, we manually designed questions and annotated them based on different numerical attributes of the images. In addition, we can generate different numerical estimation tasks based on the answer type. For example, for the question: "Which of the following options is a reasonable estimate of the number of cubes in the figure?", when the answer is "(50, 75)", it corresponds to a range estimation task. On the contrary, if the answer is "62", it belongs to the value estimation task. + +We employ a combination of automated and manual methods to design distractors. For numerical answers, we generate alternative options that are easily confusable with the correct answer. Some distractors are constructed based on their inherent properties (e.g., areas A, B, and D in part (c) of Figure 5). These distractors are designed to align with human perceptual biases, appearing plausible yet distinguishable from the correct answer. + +To ensure the high quality of VisNumBench, we meticulously reviewed all collected data and filtered out any ambiguous or unclear entries. More details of the data con + +struction process can be found in Appendix B. + +# 4. Experiments + +We evaluate 17 well-known MLLMs from 8 model families, including 13 open-source models: Phi-3.5-vision [1], LLaVA-v1.5 (7B, 13B), and LLaVA-v1.6-34B [26], LLaVA-Onevision (7B, 72B) [23], Qwen2.5-VL (3B, 7B, 72B) [42], InternVL2.5 (8B, 38B, 78B) [5], and Janus-Pro-7B [4]. Additionally, we assess 4 proprietary models: GPT-4o [19], Gemini 1.5 Flash, Gemini 2.0 Flash, and Gemini 1.5 Pro [37]. + +We randomly selected 600 samples (50 QA pairs from each numerical attribute), with 300 sourced from VisNumBench-Synthetic and 300 from VisNumBench-Real. Human evaluators independently answered each question and provided assessments. Accuracy $(\%)$ is reported for all experimental results, and all the results are provided in Tables 2 and 3. + +# 4.1. Evaluation Results and Analysis + +From Tables 2 and 3, we observe that the performance of MLLMs is not comparable to that of humans. Among the seven types of questions, quantity-related tasks appear to + +Table 3. Accuracies of MLLMs on the VisNumBench-Real (%) dataset. Dark gray and Light gray indicate the best and second-best results among all models, respectively. + +
AngleLengthScaleQuantityDepthVolumeAverage
Random25.0025.0025.0027.8325.0025.4025.54
Open-source MLLMs
Phi-3.5-vision30.2037.6527.9748.3048.7029.9337.25
LLaVA-v1.5-7B22.8232.7225.8736.7325.3227.2128.49
LLaVA-v1.5-13B28.8643.2129.3746.9449.3541.5040.02
LLaVA-v1.6-34B28.8654.9423.0868.0363.6463.2750.55
LLaVA-Onevision-7B18.1244.4420.2864.6344.8150.3440.58
LLaVA-Onevision-72B17.4557.4144.7674.8348.7061.2250.78
InternVL2.5-8B28.8634.5715.3864.6349.3547.6240.13
InternVL2.5-38B30.2051.8526.5783.6761.0458.5052.11
InternVL2.5-78B36.9158.6448.9579.5952.6062.5956.54
Janus-Pro-7B22.8232.1035.6648.9835.7130.6134.26
Qwen2.5-VL-3B30.2044.4435.6651.7043.5149.6642.57
Qwen2.5-VL-7B24.1638.8932.1759.1848.7042.8641.02
Qwen2.5-VL-72B34.2350.6243.3680.2752.6059.1853.33
API-based models
GPT-4o27.5230.2537.0660.5435.7147.6239.58
Gemini 1.5 Flash14.7735.8026.5757.1424.6843.5433.70
Gemini 2.0 Flash38.9348.7774.1481.6346.1051.7056.54
Gemini 1.5 Pro30.2045.6827.9768.0364.2955.1048.67
Human96.00100.00100.0098.0096.0094.0097.33
+ +be the easiest, while angle-related tasks are the most difficult. This is possibly because the amount of training data available for quantity-related tasks is significantly greater than that for angle-related tasks. By comparing the evaluations on synthetic and real images, we find that the performance of the same model does not exhibit significant variance. Thus, in terms of numerical reasoning ability, both synthetic and real images present similar challenges for existing MLLMs. Moreover, we observe that the best open-sourced model performs comparably to the best closed-source models. + +More detailed analyses and discussions are provided in the following subsections. + +# 4.1.1. Performance on VisNumBench-Synthetic + +Table 2 presents the results for various MLLMs on the VisNumBench-Synthetic dataset. Among the open-source models, Qwen2.5-VL-72B achieves the best performance, with an average accuracy of $58.46\%$ . InternVL2.5-38B, InternVL2.5-78B, and LLaVA-v1.6-34B also demonstrate strong performance, each achieving either the best or the second-best accuracy in at least two tasks. LLaVA-v1.6-34B attains the highest accuracy in angle and depth estimation; however, its overall average accuracy is only $44.31\%$ . + +LLaVA-Onevision-72B also performs well, achieving the highest accuracy in length estimation at $61.33\%$ . In general, models with larger parameter sizes tend to exhibit superior performance, aligning with the intuition that larger models can better capture complex numerical relationships and fine-grained visual patterns. + +In the API-based models, Gemini 2.0 Flash demonstrates the best performance, achieving an average accuracy of $57.57\%$ . In contrast, GPT-4o and Gemini 1.5 Pro exhibit comparable performance, albeit with lower average accuracies. Gemini 1.5 Flash yields the weakest performance, with an average accuracy of $33.33\%$ . Notably, certain open-source models perform on par with or even surpass proprietary models, suggesting that the disparity in numerical reasoning capabilities between open-source and closed-source models is minimal. + +# 4.1.2. Performance on VisNumBench-Real + +Accuracy on the VisNumBench-Real dataset shows similar trends. InternVL2.5-78B and Gemini 2.0 Flash stand out with an average accuracy of $56.54\%$ , achieving nearoptimal results across multiple tasks, as shown in Table 3. InternVL2.5-38B attained an exceptionally high accuracy of $83.67\%$ on the quantity task, while LLaVA-v1.6-34B + +![](images/649d79fd6115cabeb761054ef4d6d5b77b953feebde1129d222c653db86bef32.jpg) + +![](images/ebab1d392163a0d285a0c607205573471f8bf165fa8c17f662570a857898c4f8.jpg) + +![](images/79a4bcb7130e050321a1bbdbd6167d424b5ee69d1e799fd486bb5de2a34f48a8.jpg) +Figure 6. Confusion matrices (\%) of MLLMs on the VisNumBench-Synthetic and VisNumBench-Real datasets across different visual numerical estimation tasks. + +excelled in the volume task, achieving the highest scores. Qwen2.5-VL-72B demonstrated relatively balanced performance, yielding a suboptimal average accuracy of $53.33\%$ . + +Surprisingly, Gemini 1.5 Pro achieved the highest accuracy in depth estimation, reaching $64.29\%$ . However, its overall average accuracy remained unsatisfactory. Other proprietary models, such as GPT-4o and Gemini 1.5 Flash, exhibited relatively weaker performance. In general, the accuracy on VisNumBench-Real is lower than that on VisNumBench-Synthetic, likely due to the increased complexity and variability of real-world images. + +# 4.1.3. Performance on Different Visual Numerical Estimation Tasks + +As we analyze performance across different numerical estimation tasks, Figure 6 reveals that in the synthetic scenario, Gemini 2.0 Flash and Qwen2.5-VL-72B achieve the highest performance across all visual numerical estimation tasks, particularly in range estimation, value estimation, and value comparison, where their accuracies consistently exceed $60\%$ . In contrast, GPT-4o exhibits the lowest performance in all tasks, especially in value comparison and multiplicative estimation. In the real-world scenario, most models achieved their best performance in value comparison tasks, which are also the easiest for humans. Although Gemini 2.0 Flash and InternVL2.5-78B continue to perform well in most tasks, their performance in multiplicative estimation has declined compared to the synthetic scenario. Additionally, GPT-4o continues to perform poorly across all tasks, particularly in multiplicative estimation and value comparison, where it falls significantly behind other models. + +Notably, in the multiplicative estimation task, LLaVAv1.6-34B outperforms all other models by a significant margin. This suggests that certain models may be more special + +ized for specific types of tasks, and further fine-tuning could enhance performance across different tasks. + +# 4.2. Further Analysis + +How do math-special models perform on the VisNum-Bench? To investigate the number sense abilities of mathspecial models, we introduce two multimodal mathematical models: (1) InternVL2-8B-MPO [49], initialized from InternVL2-8B [8] and fine-tuned on the large-scale multimodal reasoning preference dataset MMPR [49], achieving an accuracy of $65.65\%$ on MathVista; (2) Math-LLaVA-13B [40], initialized from LLaVA-v1.5-13B and fine-tuned on the MathV360K [40] dataset. As shown in Figure 7, InternVL2-8B-MPO achieved a $1.0\%$ improvement in synthetic scenarios and a $0.3\%$ increase in real-world scenarios. Its enhancements are task-specific rather than universally effective across different number sense challenges. In contrast, Math-LLaVA-13B exhibited a polarized performance trend: while it improved by $3.7\%$ on the synthetic dataset, its accuracy declined by $6.9\%$ in real-world scenarios. This suggests that although the model benefits from training on synthetic data, it struggles to generalize to the complexity and variability of real-world number sense tasks. Relying solely on synthetic data may be insufficient to enhance number sense capabilities in real-world applications. Additional strategies, such as incorporating more diverse real-world training data or refining model architectures, may be necessary to achieve meaningful improvements. + +How do the multimodal reasoning models perform? To examine whether reasoning techniques can enhance the number sense abilities of MLLMs, we evaluate two multimodal reasoning models: Llama-3.2V-11B-cot² [51] and R1-OneVision-7B³. Llama-3.2V-11B-cot is trained us + +![](images/924fde89d6d9c652e24a037eababcc72eac5c9407576678c19233cb4eee5a3ee.jpg) +VisNumBench-Synthetic +Figure 7. Improvements brought by multimodal mathematical models (InternVL2-8B-MPO and Math-LLaVA-13B) and multimodal CoT models (Llama-3.2V-11B-cot and R1-OneVision-7B). Table 10 and Table 11 in the appendix provide detailed results. + +Table 4. Comparisons of the performance of models from the Qwen-VL family and the InternVL family in synthetic and real-world scenes. Table 12 in the appendix provides detailed results. + +
Average (Synthetic)Average (Real)
Qwen2-VL-2B31.8524.94
Qwen2.5-VL-3B42.24(↑+10.39)42.57(↑+17.63)
Qwen2-VL-7B41.2541.91
Qwen2.5-VL-7B46.19(↑+4.94)41.02(↓-0.89)
Qwen2-VL-72B54.2046.56
Qwen2.5-VL-72B58.46(↑+4.26)53.33(↑+6.77)
InternVL2-8B39.5639.58
InternVL2.5-8B39.66(↑+0.10)40.13(↑+0.55)
InternVL2-40B45.5045.12
InternVL2.5-38B55.59(↑+10.09)52.11(↑+6.99)
+ +ing LLaVA-o1-100k [51], achieving a $6.2\%$ performance improvement on MathVista compared to Llama-3.2-11B-Vision-Instruct [34]. R1-OneVision-7B, trained with a rule-based reinforcement learning technique, attains an accuracy of $44.06\%$ on Mathverse [55]. Accordingly, we assess these models on our benchmark. The results, presented in Figure 7, indicate that neither Llama-3.2V-11B-cot nor R1-OneVision-7B achieved the expected performance gains. On the contrary, their accuracy dropped significantly—except for a modest $1.6\%$ improvement by Llama-3.2V-11B-cot in synthetic scenarios—especially in real-world settings. These findings suggest that developing reasoning techniques specifically tailored for number sense abilities may be necessary. + +What helps improve the performance? To determine + +the factors contributing to the improvement of number sense ability in MLLMs, we evaluate historical models from the same family over time, specifically the Qwen-VL family and the InternVL family. The results are presented in Table 4. As observed, the performance of the latest models generally surpasses that of their predecessors. By comparing Qwen2-VL [45] with Qwen2.5-VL [42], as well as InternVL2 with InternVL2.5 [6], we observe improvements in several aspects: (1) data scale and quality, (2) a more powerful encoder, (3) model architecture, and (4) training strategy. These findings suggest that further exploration in these directions is essential for enhancing the number sense abilities of MLLMs. + +# 5. Conclusion + +In this work, we introduce VisNumBench, a novel benchmark designed to evaluate MLLMs on core number sense abilities that are inadequately addressed by existing evaluation benchmarks. Our assessment of 17 MLLMs uncovers substantial deficiencies in their capacity to demonstrate human-like number sense. Even the most advanced models still demonstrate limited numerical sense abilities. Further experiments on historical models from the same family show that to enhance this ability within a short period, more specialized optimizations in data, training techniques, and model architecture may be required. + +# 6. Acknowledge + +This research was supported by the National Natural Science Foundation of China under Grant No. 62272315. + +# References + +[1] Marah Abdin, Jyoti Aneja, Hany Awadalla, Ahmed Awadallah, Ammar Ahmad Awan, Nguyen Bach, Amit Bahree, Arash Bakhtiari, Jianmin Bao, Harkirat Behl, et al. Phi-3 technical report: A highly capable language model locally on your phone. arXiv preprint arXiv:2404.14219, 2024. 3, 5 +[2] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in Neural Information Processing Systems, 35:23716-23736, 2022. 1 +[3] AI Anthropic. Claude 3.5 sonnet model card addendum. Claude-3.5 Model Card, 3, 2024. 3 +[4] Xiaokang Chen, Zhiyu Wu, Xingchao Liu, Zizheng Pan, Wen Liu, Zhenda Xie, Xingkai Yu, and Chong Ruan. Januspro: Unified multimodal understanding and generation with data and model scaling, 2025. 3, 5 +[5] Zhe Chen, Weiyun Wang, Yue Cao, Yangzhou Liu, Zhangwei Gao, Erfei Cui, Jinguo Zhu, Shenglong Ye, Hao Tian, Zhaoyang Liu, et al. Expanding performance boundaries of open-source multimodal models with model, data, and test-time scaling. arXiv preprint arXiv:2412.05271, 2024. 3, 5 +[6] Zhe Chen, Weiyun Wang, Yue Cao, Yangzhou Liu, Zhangwei Gao, Erfei Cui, Jinguo Zhu, Shenglong Ye, Hao Tian, Zhaoyang Liu, et al. Expanding performance boundaries of open-source multimodal models with model, data, and test-time scaling. arXiv preprint arXiv:2412.05271, 2024.8 +[7] Zhe Chen, Weiyun Wang, Hao Tian, Shenglong Ye, Zhangwei Gao, Erfei Cui, Wenwen Tong, Kongzhi Hu, Jiapeng Luo, Zheng Ma, et al. How far are we to gpt-4v? closing the gap to commercial multimodal models with open-source suites. Science China Information Sciences, 67(12):220101, 2024. 1 +[8] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24185–24198, 2024. 7 +[9] Anoop Cherian, Kuan-Chuan Peng, Suhas Lohit, Joanna Matthiesen, Kevin Smith, and Joshua B Tenenbaum. Evaluating large vision-and-language models on children's mathematical olympiads. arXiv preprint arXiv:2406.15736, 2024. 3 +[10] Adam Dahlgren Lindström and Savitha Sam Abraham. CLEVR-Math: A dataset for compositional language, visual and mathematical reasoning. In 16th International Workshop on Neural-Symbolic Learning and Reasoning, NeSy 2022, Windsor, UK, september 28-30, 2022. CEUR-WS, 2022. 1 +[11] Yao Du, Qiang Zhai, Weihang Dai, and Xiaomeng Li. Teach clip to develop a number sense for ordinal regression. arXiv preprint arXiv:2408.03574, 2024. 3 +[12] echarts. echarts. https://echarts.apache.org/zh/index.html. Accessed: 2025-01-26. 4, 1 +[13] Meng Fang, Xiangpeng Wan, Fei Lu, Fei Xing, and Kai Zou. Mathodyssey: Benchmarking mathematical problem + +solving skills in large language models using odyssey math data. arXiv preprint arXiv:2406.18321, 2024. 3 +[14] Lisa Feigenson, Stanislas Dehaene, and Elizabeth Spelke. Core systems of number. Trends in cognitive sciences, 8(7): 307-314, 2004. 1 +[15] Chaoyou Fu, Peixian Chen, Yunhang Shen, Yulei Qin, Mengdan Zhang, Xu Lin, Jinrui Yang, Xiawu Zheng, Ke Li, Xing Sun, et al. Mme: A comprehensive evaluation benchmark for multimodal large language models. arXiv preprint arXiv:2306.13394, 2023. 3 +[16] Xingyu Fu, Yushi Hu, Bangzheng Li, Yu Feng, Haoyu Wang, Xudong Lin, Dan Roth, Noah A Smith, Wei-Chiu Ma, and Ranjay Krishna. Blink: Multimodal large language models can see but not perceive. In European Conference on Computer Vision, pages 148-166. Springer, 2025. 3, 4, 1, 2 +[17] Google. Google images. https://images.google.com/. Accessed: 2025-01-26. 4, 1 +[18] He Hu, Yucheng Zhou, Lianzhong You, Hongbo Xu, Qian-ning Wang, Zheng Lian, Fei Richard Yu, Fei Ma, and Laizhong Cui. Emobench-m: Benchmarking emotional intelligence for multimodal large language models. arXiv preprint arXiv:2502.04424, 2025. 1 +[19] Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Weli-hinda, Alan Hayes, Alec Radford, et al. Gpt-4o system card. arXiv preprint arXiv:2410.21276, 2024. 5 +[20] Ryo Kamoi, Yusen Zhang, Sarkar Snigdha Sarathi Das, Ranran Haoran Zhang, and Rui Zhang. Visonlyqa: Large vision language models still struggle with visual perception of geometric information. arXiv preprint arXiv:2412.00947, 2024. 1, 3, 4 +[21] Bohao Li, Yuying Ge, Yixiao Ge, Guangzhi Wang, Rui Wang, Ruimao Zhang, and Ying Shan. Seed-bench-2: Benchmarking multimodal large language models. arXiv preprint arXiv:2311.17092, 2023. +[22] Bohao Li, Rui Wang, Guangzhi Wang, Yuying Ge, Yixiao Ge, and Ying Shan. Seed-bench: Benchmarking multimodal llms with generative comprehension. arXiv preprint arXiv:2307.16125, 2023. 3 +[23] Chunyuan Li, Cliff Wong, Sheng Zhang, Naoto Usuyama, Haotian Liu, Jianwei Yang, Tristan Naumann, Hoifung Poon, and Jianfeng Gao. Llava-med: Training a large language-and-vision assistant for biomedicine in one day. Advances in Neural Information Processing Systems, 36, 2024. 3, 5 +[24] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. arXiv preprint arXiv:2301.12597, 2023. 1 +[25] Wanhua Li, Xiaoke Huang, Zheng Zhu, Yansong Tang, Xiu Li, Jie Zhou, and Jiwen Lu. Ordinalclip: Learning rank prompts for language-guided ordinal regression. Advances in Neural Information Processing Systems, 35:35313-35325, 2022. 3 +[26] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. arXiv preprint arXiv:2304.08485, 2023. 5 + +[27] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26296-26306, 2024. 1 +[28] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, et al. MMBench: Is your multi-modal model an all-around player? arXiv preprint arXiv:2307.06281, 2023. 3 +[29] Pan Lu, Ran Gong, Shibiao Jiang, Liang Qiu, Siyuan Huang, Xiaodan Liang, and Song-Chun Zhu. Inter-GPS: Interpretable geometry problem solving with formal language and symbolic reasoning. In The 59th Annual Meeting of the Association for Computational Linguistics (ACL), 2021. 1 +[30] Pan Lu, Hritik Bansal, Tony Xia, Jiacheng Liu, Chunyuan Li, Hannaneh Hajishirzi, Hao Cheng, Kai-Wei Chang, Michel Galley, and Jianfeng Gao. Mathvista: Evaluating mathematical reasoning of foundation models in visual contexts. arXiv preprint arXiv:2310.02255, 2023. 1, 3, 4 +[31] Pan Lu, Baolin Peng, Hao Cheng, Michel Galley, Kai-Wei Chang, Ying Nian Wu, Song-Chun Zhu, and Jianfeng Gao. Chameleon: Plug-and-play compositional reasoning with large language models. In The 37th Conference on Neural Information Processing Systems (NeurIPS), 2023. 1 +[32] Ahmed Masry, Xuan Long Do, Jia Qing Tan, Shafiq Joty, and Enamul Hoque. ChartQA: A benchmark for question answering about charts with visual and logical reasoning. In Findings of the Association for Computational Linguistics: ACL 2022, pages 2263-2279, 2022. 1 +[33] Ahmed Masry, Parsa Kavehzadeh, Xuan Long Do, Enamul Hoque, and Shafiq Joty. UniChart: A universal vision-language pretrained model for chart comprehension and reasoning. arXiv preprint arXiv:2305.14761, 2023. 3 +[34] Meta AI. Llama 3.2 vision instruct (11b). https://huggingface.co/meta-llama/Llama-3.2-11B-Vision-Instruct, 2024. Accessed: 2025-07.8 +[35] OpenAI. GPT-4V(ision) system card, 2023. 3 +[36] Frank Palermo, James Hays, and Alexei A Efros. Dating historical color images. In Computer Vision-ECCV 2012: 12th European Conference on Computer Vision, Florence, Italy, October 7-13, 2012, Proceedings, Part VI 12, pages 499-512. Springer, 2012. 4 +[37] Sundar Pichai and Demis Hassabis. Our next-generation model: Gemini 1.5. ai, 2024. 3, 5 +[38] Karl Ricanek and Tamirat Tesafaye. Morph: A longitudinal image database of normal adult age-progression. In 7th international conference on automatic face and gesture recognition (FGR06), pages 341-345. IEEE, 2006. 4 +[39] Rossano Schifanella, Miriam Redi, and Luca Maria Aiello. An image is worth more than a thousand favorites: Surfacing the hidden beauty of flickr pictures. In Proceedings of the international AAAI conference on web and social media, pages 397-406, 2015. 4 +[40] Wenhao Shi, Zhiqiang Hu, Yi Bin, Junhua Liu, Yang Yang, See-Kiong Ng, Lidong Bing, and Roy Ka-Wei Lee. Math-llava: Bootstrapping mathematical reasoning for multimodal large language models. arXiv preprint arXiv:2406.17294, 2024. 7 + +[41] Nathan Silberman, Derek Hoiem, Pushmeet Kohli, and Rob Fergus. Indoor segmentation and support inference from rgbd images. In Computer Vision-ECCV 2012: 12th European Conference on Computer Vision, Florence, Italy, October 7-13, 2012, Proceedings, Part V 12, pages 746-760. Springer, 2012. 4, 1 +[42] Qwen Team. Qwen2.5-vl, 2025. 3, 5, 8 +[43] WallpapersCraft. Desktop wallpapers hd, free desktop backgrounds. https://wallpaperscraft.com/. Accessed: 2025-01-26. 4, 1 +[44] Ke Wang, Junting Pan, Weikang Shi, Zimu Lu, Mingjie Zhan, and Hongsheng Li. Measuring multimodal mathematical reasoning with math-vision dataset. arXiv preprint arXiv:2402.14804, 2024.3 +[45] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 3, 8 +[46] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 3 +[47] Rui Wang, Peipei Li, Huaibo Huang, Chunshui Cao, Ran He, and Zhaofeng He. Learning-to-rank meets language: Boosting language-driven ordering alignment for ordinal classification. Advances in Neural Information Processing Systems, 36, 2023. 3 +[48] Wenshan Wang, Delong Zhu, Xiangwei Wang, Yaoyu Hu, Yuheng Qiu, Chen Wang, Yafei Hu, Ashish Kapoor, and Sebastian Scherer. Tartanair: A dataset to push the limits of visual slam. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 4909-4916. IEEE, 2020. 4, 1 +[49] Weiyun Wang, Zhe Chen, Wenhai Wang, Yue Cao, Yangzhou Liu, Zhangwei Gao, Jinguo Zhu, Xizhou Zhu, Lewei Lu, Yu Qiao, et al. Enhancing the reasoning ability of multimodal large language models via mixed preference optimization. arXiv preprint arXiv:2411.10442, 2024. 7 +[50] Xiaoxuan Wang, Ziniu Hu, Pan Lu, Yanqiao Zhu, Jieyu Zhang, Satyen Subramaniam, Arjun R Loomba, Shichang Zhang, Yizhou Sun, and Wei Wang. SciBench: Evaluating college-level scientific problem-solving abilities of large language models. arXiv preprint arXiv:2307.10635, 2023. 1 +[51] Guowei Xu, Peng Jin, Li Hao, Yibing Song, Lichao Sun, and Li Yuan. Llava-ol: Let vision language models reason step-by-step. arXiv preprint arXiv:2411.10440, 2024. 7, 8 +[52] Le Xue, Manli Shu, Anas Awadalla, Jun Wang, An Yan, Senthil Purushwalkam, Honglu Zhou, Viraj Prabhu, Yutong Dai, Michael S Ryoo, et al. xgen-mm (blip-3): A family of open large multimodal models. arXiv preprint arXiv:2408.08872, 2024. 3 +[53] Jihan Yang, Shusheng Yang, Anjali W Gupta, Rilyn Han, Li Fei-Fei, and Saining Xie. Thinking in space: How multimodal large language models see, remember, and recall spaces. arXiv preprint arXiv:2412.14171, 2024. 3 + +[54] Weihao Yu, Zhengyuan Yang, Linjie Li, Jianfeng Wang, Kevin Lin, Zicheng Liu, Xinchao Wang, and Lijuan Wang. Mm-vet: Evaluating large multimodal models for integrated capabilities. arXiv preprint arXiv:2308.02490, 2023. 3 +[55] Renrui Zhang, Dongzhi Jiang, Yichi Zhang, Haokun Lin, Ziyu Guo, Pengshuo Qiu, Aojun Zhou, Pan Lu, Kai-Wei Chang, Yu Qiao, et al. Mathverse: Does your multi-modal llm truly see the diagrams in visual math problems? In European Conference on Computer Vision, pages 169–186. Springer, 2024. 8 +[56] Renrui Zhang, Xinyu Wei, Dongzhi Jiang, Yichi Zhang, Ziyu Guo, Chengzhuo Tong, Jiaming Liu, Aojun Zhou, Bin Wei, Shanghang Zhang, et al. Mavis: Mathematical visual instruction tuning. CoRR, 2024. 3 +[57] Yingying Zhang, Desen Zhou, Siqin Chen, Shenghua Gao, and Yi Ma. Single-image crowd counting via multi-column convolutional neural network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 589-597, 2016. 4, 1 +[58] Yanzhe Zhang, Ruiyi Zhang, Jiumiang Gu, Yufan Zhou, Nedim Lipka, Diyi Yang, and Tong Sun. LLaVAR: Enhanced visual instruction tuning for text-rich image understanding. arXiv preprint arXiv:2306.17107, 2023. 3 \ No newline at end of file diff --git a/ICCV/2025/VisNumBench_ Evaluating Number Sense of Multimodal Large Language Models/images.zip b/ICCV/2025/VisNumBench_ Evaluating Number Sense of Multimodal Large Language Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..d75110796f171b49f15eda5e12f0f6596bb490b5 --- /dev/null +++ b/ICCV/2025/VisNumBench_ Evaluating Number Sense of Multimodal Large Language Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:73cce710d1099635967d3ab3306c33099d3e8cd5d194054d887db543de6e6465 +size 719584 diff --git a/ICCV/2025/VisNumBench_ Evaluating Number Sense of Multimodal Large Language Models/layout.json b/ICCV/2025/VisNumBench_ Evaluating Number Sense of Multimodal Large Language Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..b93afd705fa169456f9967231fb7c36f501f3a33 --- /dev/null +++ b/ICCV/2025/VisNumBench_ Evaluating Number Sense of Multimodal Large Language Models/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:63b52ace55518673df9e433300d3a3c8165b34c65a391280dccec31dacc446c5 +size 519076 diff --git a/ICCV/2025/VisRL_ Intention-Driven Visual Perception via Reinforced Reasoning/f2419d2f-e576-4a4f-bc19-8d0454e83639_content_list.json b/ICCV/2025/VisRL_ Intention-Driven Visual Perception via Reinforced Reasoning/f2419d2f-e576-4a4f-bc19-8d0454e83639_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..ea9afc6dbb2f07a33633160f486ca5bbb88312c1 --- /dev/null +++ b/ICCV/2025/VisRL_ Intention-Driven Visual Perception via Reinforced Reasoning/f2419d2f-e576-4a4f-bc19-8d0454e83639_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:da9812a29a788e6a3aa789170184ca222729e700c1a37843bbeb5d42bd8ae173 +size 90115 diff --git a/ICCV/2025/VisRL_ Intention-Driven Visual Perception via Reinforced Reasoning/f2419d2f-e576-4a4f-bc19-8d0454e83639_model.json b/ICCV/2025/VisRL_ Intention-Driven Visual Perception via Reinforced Reasoning/f2419d2f-e576-4a4f-bc19-8d0454e83639_model.json new file mode 100644 index 0000000000000000000000000000000000000000..5a652ab52593bf7c81915694fa1266986e0d8ef0 --- /dev/null +++ b/ICCV/2025/VisRL_ Intention-Driven Visual Perception via Reinforced Reasoning/f2419d2f-e576-4a4f-bc19-8d0454e83639_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:094218b9f21defeba310c012f5dde4161365b74651497ec659a45d3173ea6b3e +size 113725 diff --git a/ICCV/2025/VisRL_ Intention-Driven Visual Perception via Reinforced Reasoning/f2419d2f-e576-4a4f-bc19-8d0454e83639_origin.pdf b/ICCV/2025/VisRL_ Intention-Driven Visual Perception via Reinforced Reasoning/f2419d2f-e576-4a4f-bc19-8d0454e83639_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..fe6b7237df3284c0e97d73827b4e4d3c2d724e7a --- /dev/null +++ b/ICCV/2025/VisRL_ Intention-Driven Visual Perception via Reinforced Reasoning/f2419d2f-e576-4a4f-bc19-8d0454e83639_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6c61948650b7a983989eb07d97ab746a3b1d6e9bac2efbd27c039c6797065580 +size 1876097 diff --git a/ICCV/2025/VisRL_ Intention-Driven Visual Perception via Reinforced Reasoning/full.md b/ICCV/2025/VisRL_ Intention-Driven Visual Perception via Reinforced Reasoning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..0bd1937064133ee87a27dabfd13b7c1c61fa77d5 --- /dev/null +++ b/ICCV/2025/VisRL_ Intention-Driven Visual Perception via Reinforced Reasoning/full.md @@ -0,0 +1,352 @@ +# VisRL: Intention-Driven Visual Perception via Reinforced Reasoning + +Zhangquan Chen $^{1*}$ Xufang Luo $^{2\dagger}$ Dongsheng Li $^{2}$ $^{1}$ Tsinghua University, China $^{2}$ Microsoft Research Asia, China + +# Abstract + +Visual understanding is inherently intention-driven—humans selectively focus on different regions of a scene based on their goals. Recent advances in large multimodal models (LMMs) enable flexible expression of such intentions through natural language, allowing queries to guide visual reasoning processes. Frameworks like Visual Chain-of-Thought have demonstrated the benefit of incorporating explicit reasoning steps, where the model predicts a focus region before answering a query. However, existing approaches rely heavily on supervised training with annotated intermediate bounding boxes, which severely limits scalability due to the combinatorial explosion of intention-region pairs. To overcome this limitation, we propose VIsRL, the first framework that applies reinforcement learning (RL) to the problem of intention-driven visual perception. VIsRL optimizes the entire visual reasoning process using only reward signals. By treating intermediate focus selection as an internal decision optimized through trial-and-error, our method eliminates the need for costly region annotations while aligning more closely with how humans learn to perceive the world. Extensive experiments across multiple benchmarks show that VIsRL consistently outperforms strong baselines, demonstrating both its effectiveness and its strong generalization across different LMMs. Our code is available at https://github.com/zhangquanchen/VisRL. + +# 1. Introduction + +Visual understanding is a fundamental problem in computer vision, enabling machines to interpret and interact with their surroundings [22, 41, 59]. Traditional methods of visual perception often process entire scenes uniformly [9, 11, 25, 47, 69], without considering the intent behind a given task. However, human perception is inherently intention-driven — people focus on different aspects of a scene depending on their goals. For example, when entering a room, a person searching for a television remote will + +![](images/9277f6a29146dab9bb3e8da61be0e35331b6ceddcc15bd75e5d434df8ab6a139.jpg) +Figure 1. Illustration of using RL to optimize the visual reasoning process. SFT trains with densely annotated training data for several epochs. V1sRL leverages self-generated data and self-provided rewards to iteratively update the model using step-level DPO. This RL process removes the need for bounding box annotations, enabling a more human-like, intention-driven visual perception. + +scan tables and couches, while someone checking the time will look at the walls for a clock. This context-dependent approach to visual perception suggests that intelligent models should also adapt their focus based on the task at hand. This leads to the problem of intention-driven visual perception, where the goal is to dynamically determine the most relevant regions of an image based on a given query or task [51]. + +With the advent of large multimodal models, the intention in perception tasks can now be expressed in a highly flexible way, which is natural language. Common LMMs, such as LLaVA [34] and Qwen-VL [56], first encode visual signals into tokens, and then both visual and text tokens are jointly processed by the large language models to produce final outputs. Despite showing powerful ability in many tasks [12, 26, 32, 57, 68, 70], this kind of one-pass end-to-end LMMs still suffer from hallucinations [20] and do not explicitly address the intention-driven visual perception problem. Further extending this paradigm, instead of treating the entire process as a single black-box inference, recent works have proposed frameworks like Visual Chain-of-Thought (Visual CoT [48]). These methods introduce an explicit reasoning step where the model first predicts a bounding box + +![](images/34e4b7de59ed68f6323e6b4fb1a9594e9422f153f9a2d36601f5b9ab4256dc7a.jpg) +Figure 2. The schematic illustration of our VIsRL framework. VIsRL first utilizes a small amount of data for SFT warm-up, but in the subsequent RL training phase, it can leverage large-scale data without bounding box annotations. The RL phase of VIsRL consists of iterative cycles of data generation and optimization, and $k$ in the figure indicates the iteration index. The data generation process does not rely on external models or annotations; instead, it employs the model itself for data synthesis and scoring. The optimization step adopts step-level DPO to ensure the model learns each step of the reasoning process. In summary, VIsRL enables intention-driven visual perception by leveraging RL to learn from task rewards without requiring annotations and external helps. + +representing the critical region to focus on, crops the image to extract this region, and then feeds the cropped visual input back to the multimodal model. By conditioning the final answer on both the query and the selected focus area, this approach not only improves interpretability but also helps leverage the multi-turn in-context learning capabilities of the underlying LMM. + +However, despite the advantages brought by methods like Visual CoT, these approaches also impose extremely high requirements on training data. Existing methods rely on supervised learning to teach models to produce intermediate reasoning steps. Specifically, for each intention or query, the training process requires corresponding bounding box annotations to guide the model in identifying the correct focus area. This dependence on exhaustive annotation is impractical, as the same image can correspond to vastly different regions depending on the request. As a result, the annotation complexity grows combinatorially with the diversity of queries, making it impossible to cover all potential cases in a scalable manner. + +In this work, we propose a novel learning framework VisRL that optimizes the entire reasoning process based solely on rewards and feedback from the task itself, rather than dense annotations (shown in Figure 1). Specifically, we treat the success or failure of the task as a reward signal and apply reinforcement learning (RL) to train the model. Our approach requires no bounding box annotations, thereby addressing the scalability challenge posed by annotation requirements. Furthermore, this learning paradigm aligns more closely with how humans acquire perceptual skills — humans do not learn to focus on specific regions through meticulously annotated training data for each task, but rather through trial-and-error interaction with the environment, gradually developing the ability to adaptively zoom into relevant regions. By leveraging this reward-driven learning strategy, our framework enables intention-driven visual per + +ception in a more flexible, scalable, and human-like manner. + +VisRL adopts an iterative DPO framework to complete the RL process for visual reasoning. This framework consists of multiple cycles of data generation and model optimization. During the data generation phase, we introduce a diversity controller to ensure that the generated bounding boxes cover a wide variety of potential focus regions. Additionally, we apply a filtering mechanism to select questions with appropriate difficulty levels for the current model and to identify the most effective preference pairs associated with each question. During the model optimization phase, we apply a step-level DPO algorithm, ensuring that the model learns to optimize every step of the visual reasoning process. + +Our contributions can be summarized as follows. + +- We present VIsRL, the first framework to apply RL to the problem of intention-driven visual perception. By addressing the data annotation bottleneck, VIsRL establishes a learning process that is much closer to human-like visual understanding. +- We design a tailored data generation pipeline for V1sRL, incorporating both a diversity controller to enhance visual exploration and further use a step-level DPO algorithm to fully exploit the collected data. +- Extensive experiments across multiple benchmarks demonstrate that VIsRL consistently outperforms strong baselines. Moreover, our results show that the effectiveness of VIsRL generalizes well across different multimodal models, highlighting its broad applicability. + +# 2. Related Work + +Multi-modal Large Language Models With the advancement of large language models (LLMs) [1, 2, 4, 6, 7, 17, 19, 24, 42, 43, 52, 54, 55, 65, 71], multi-modal large language models which integrate vision and language modalities have also experienced rapid development. This progress + +enables AI systems to better perceive and understand the real-world interplay between visual and textual information. Notable methods like LLaVA [34] aligns image tokens with pre-trained LLMs by training a projector, while other approaches utilize a Q-Former [29, 30] to learn image embeddings via learnable queries after extracting image features. These LMMs provide strong base models for V1sRL, as they have the capability to process both visual and language data simultaneously, enabling the completion of the reasoning process. + +Intention-Driven Visual Models Recently, several methods have attempted to enhance models' intention-driven visual perception capabilities. VisCoT [48] employs a multi-turn interpretable processing mechanism with bounding boxes that dynamically focus on visual inputs. Similarly, SpatialCoT [37] achieves spatial grounding through spatial coordinates, while SegLLM [58] leverages mask-labeled data to enable reasoning about complex user segmentation intentions. V* (SEAL) [61] provides an LLM-guided search mechanism for efficient visual querying. Besides, both MLLM-TPO [64] and VisionLLM v2 [60] achieve intention-driven perception by training decoders for specific downstream tasks. Additionally, MVoT [28] introduces a text-image-text reasoning paradigm by training on interleaved data. These methods generally follow the supervised learning (SL) paradigm, therefore heavily relying on dense-labeled data (e.g., bounding boxes, spatial coordinates, masks, multi-round reasoning conversations), which constrains their ability to scale further. Instead, VɪsRL uses rewards as learning supervision and gets rid of the requirement on dense annotations. On the other hand, several methods employ tool-usage for enhancement. Specifically, AURORA [5] leverages specialized detection models, while Plug-and-Play [8] adopts a multi-agent framework. Besides, ViRReq [51] leverages the knowledge base to decompose visual recognition. However, these approaches rely on external models or knowledge, but not focusing on enhancing the intrinsic capabilities of the models themselves, while VɪsRL tries to completing the task via learning to reasoning by the model itself. + +Multimodal Models with Chain-of-Thoughts Chain-of- Thought (CoT) reasoning plays an important role for LMMs. The methods can be broadly categorized into two types. (1) Text-thought methods [10, 15, 23, 53, 63] elicit textual CoT reasoning of LLMs in visual reasoning tasks by introducing text thinking tokens inspired by [21], guiding towards the final response. Differently, VisRL emphasizes that visual information should be involved in the reasoning process to fully leverage the strengths of multimodal models, rather than relying solely on language tokens for reasoning. (2) Multi-modal-thought methods involves multimodal information in the reasoning process. The Mind's Eye [62] elicits + +spatial reasoning of LLMs by visualizing their reasoning traces. Additionally, [37, 48, 58] first generate visual marks (e.g. bounding boxes, spatial coordinates, masks), and subsequently perform CoT reasoning based on these fine-grained visual marks. These works still uses SL for optimization while V1sRL first explores RL in this direction. Besides, recent repositories [38, 49] also tried to enhance the bounding boxes (bboxes) generation through RL, but they use ground truth bboxes to give rewards, which is still limited by dense annotation data. + +# 3. Methodology + +Preliminary. Reinforcement Learning [14] stands out as a highly effective method for significantly bolstering the robustness, factual accuracy, and safety of large language models [40]. The method consists of two key training stages, namely the reward model training and the policy model training. To avoid this complex training pipeline, [45] proposed Direct Preference Optimization (DPO). DPO streamlines the process by directly leveraging pair-wise preference data to optimize the policy model with an equivalent optimization objective. Specifically, given an input prompt $x$ , and a preference data pair $(y_{win}, y_{lose})$ , DPO aims to maximize the probability of the preferred output $y_{win}$ and minimize that of the undesirable output $y_{lose}$ . The optimization objective is formulated as: + +$$ +\mathcal {L} _ {D P O} (\theta) = - \mathbb {E} _ {(x, y _ {\text {w i n}}, y _ {\text {l o s e}}) \sim D} (P (\theta)), \tag {1} +$$ + +where $D$ is the pair-wise preference dataset, and $P_{\theta}$ is: + +$$ +P (\theta) = \log \sigma \left(\beta \log \frac {\pi_ {\theta} (y _ {w i n} \mid x)}{\pi_ {r e f} (y _ {w i n} \mid x)} - \beta \log \frac {\pi_ {\theta} (y _ {l o s e} \mid x)}{\pi_ {r e f} (y _ {l o s e} \mid x)}\right). \tag {2} +$$ + +$\sigma$ is the sigmoid function, $\pi_{\theta}(\cdot \mid x)$ is the policy model to be optimized, $\pi_{ref}(\cdot \mid x)$ is the reference model kept unchanged during training, and the hyperparameter $\beta$ serves to regulate the proximity of the policy model to the reference model. + +Method overview. As shown in Figure 2, V1sRL adopts DPO to optimize the entire visual reasoning process due to its simplicity. Our method leverages the final task success or failure as the outcome reward, and the grades of intermediate steps as the process reward, guiding the model to gradually refine its reasoning process through reinforcement learning. This reasoning process is divided into two steps. In the first step, the model generates a bounding box $B$ representing the focused area based on the given query or question $Q$ and the original image $I$ . In the second step, the cropped region $I^s$ corresponding to the bounding box, together with the original image and the question, is fed into the multimodal model to produce the final response $R$ . + +In Section 3.1, we will describe the data generation process required to support this two-stage reasoning pipeline. + +In Section 3.2, we will introduce the optimization strategy, explaining how stepwise DPO is applied to enable the model to acquire intention-driven visual perception ability. + +Notably, this data generation and optimization process will be iterated multiple times. The improved model can collect better data, and the better data will further refine the model. The initial model before RL training is denoted with $\mathcal{M}_{RL}^{0}$ , and the model which is updated with $k$ iterations is denoted with $\mathcal{M}_{RL}^{k}$ . + +![](images/8bc8df437b802502c957b43718bc828ef264e8b16c27139ae84ce884bb549df2.jpg) +Figure 3. The schematic illustration of our data generation pipeline. Here $\mathcal{M}_{RL}^{k}$ denotes the model updated with $k$ iterations of data generation and optimization, and $\mathcal{M}_{org}$ is the original model. VIsRL uses $\mathcal{M}_{RL}^{k}$ to generate samples, and use $\mathcal{M}_{org}$ to provide rewards. Hence, different versions of a single model are used in this self-evolution data generation process, and no bounding box annotations and external models are introduced in this process. + +# 3.1. Data Generation + +Before RL training, we use SFT as a warm-up stage. This stage changes the original model $\mathcal{M}_{org}$ into $\mathcal{M}_{SFT}$ , so $\mathcal{M}_{SFT}$ is $\mathcal{M}_{RL}^{0}$ . This stage requires bounding box annotations, so we directly use samples from VisCoT dataset [48]. Our intent here is just making the model have the basic ability to generate bounding boxes in specific format. Therefore, the used data amount is relatively small here (30k in ours and 438k in VisCoT). The second part of our source datasets consists of question-answer pairs from nine source datasets that span five distinct domains, with the majority of them being Visual Question Answering (VQA) and Image Captioning datasets. These are specifically used for constructing preference data for RL training, therefore requiring no bounding box annotations. We use an additional 180k data here. In Alg. 1, we explain how we construct preference data based on our source datasets in details. + +Self-evolution. The critical parts for enabling effective RL learning is generating CoT data, and providing reward signals. Unlike previous manual annotation methods, which is infeasible for providing widely enough coverage for all possible cases, or approaches that rely on more powerful external models, which is more like distillation rather than allowing the model to learn on its own, we adopt a strategy that does not depend on external capabilities—instead, leveraging the model's own self-evolution. To achieve this, we sample from + +Algorithm 1: Self-Evolution Data Generation +Input: One sample from the source dataset including question $Q$ , image $I$ , and response $R$ . The original model $\mathcal{M}_{org}$ to provide criticism. The model updated with $k$ RL iterations $\mathcal{M}_{RL}^{k}$ to generate bounding boxes (b-boxes) and responses. // Hyperparams in experiments: $N = 5$ , $\mathcal{T} = 0.2$ , $\mathcal{T}_{\mathrm{max}}^b = \mathcal{T}_{\mathrm{max}}^r = 0.8$ , $\mathcal{T}_{\mathrm{min}}^b = \mathcal{T}_{\mathrm{min}}^r = 0.2$ +1 Candidate preference pool $P \gets \emptyset$ . +2 for $i = 1$ to N do +3 $\begin{array}{l} B_{i,1} = \mathcal{M}_{RL}^{k}(Q,I), B_{i,2} = \mathcal{M}_{RL}^{k}(Q,I); \\ \text{if IoU} (B_{i,1},B_{i,2}) > \mathcal{T}\text{then} \\ \boxed{B_{i,2} = \mathcal{R}_S(\mathcal{U}_Z(B_{i,1})) // \text{Refer to Eqn.(3).}} \\ \text{Increase b-boxes diversity.} \\ \text{/ Crop sub-images and generate responses} \\ I_{i,1}^s \leftarrow \text{crop}(I,B_{i,1}), I_{i,2}^s \leftarrow \text{crop}(I,B_{i,2}); \\ R_{i,1} = \mathcal{M}_{RL}^{k}(Q,I,I_{i,1}^s), R_{i,2} = \mathcal{M}_{RL}^{k}(Q,I,I_{i,2}^s); \\ // \text{Score b-boxes and responses} \\ s_{i,1}^{b} = \mathcal{M}_{org}(Q,I_{i,1}^{s},R), s_{i,2}^{b} = \mathcal{M}_{org}(Q,I_{i,2}^{s},R); \\ s_{i,1}^{r} = \mathcal{M}_{org}(Q,R_{i,1},R), s_{i,2}^{r} = \mathcal{M}_{org}(Q,R_{i,2},R); \\ p_{i,1} = (Q,I,B_{i,1},I_{i,1}^{s},R_{i,1},s_{i,1}^{b},s_{i,1}^{r}), \\ p_{i,2} = (Q,I,B_{i,2},I_{i,2}^{s},R_{i,2},s_{i,2}^{b},s_{i,2}^{r}) \\ P \gets P \cup \{p_{i,1},p_{i,2}\}; \end{array}$ +4 +5 +6 +7 +8 +9 +10 +11 +12 +13 +14 +15 +16 +17 +18 +19 +20 +return (p_win,pclose) when they both have values // filter the sample that is too hard/easy for current model + +$\mathcal{M}_{SFT}$ multiple times to generate CoT data with as much diversity as possible, while use $\mathcal{M}_{org}$ to provide criticism. This approach not only showcases the intrinsic capabilities of the model, but also facilitates a more effective adjustment of the predicted probability distribution toward a stable state through self-generated data. + +Sample generation and rewarding. For each input question and image $(Q, I)$ , we first sample from $\mathcal{M}_{SFT}$ twice to obtain the bounding boxes $B_1$ and $B_2$ , respectively. Then, we introduce a diversity controller to ensure diversity of bounding boxes. Specifically, we update $B_2$ according to the Intersection over Union (IoU) and the rejecting threshold value $\mathcal{T}$ between $B_1$ and $B_2$ , which is formulated as: + +$$ +\hat {B} _ {2} = \left\{ \begin{array}{l l} B _ {2}, & \operatorname {I o U} \left(B _ {1}, B _ {2}\right) < \mathcal {T} \\ \mathcal {R} _ {\mathcal {S}} \left(\mathcal {U} _ {\mathcal {I}} \left(B _ {1}\right)\right), & \operatorname {I o U} \left(B _ {1}, B _ {2}\right) \geq \mathcal {T} \end{array} , \right. \tag {3} +$$ + +where $\mathcal{U}_{\mathcal{I}}(B_1)$ is a set of bounding boxes that are outside of $B_{1}$ but within image $I$ . The operator $\mathcal{R}_S(\cdot)$ represents a random selection from the set $\mathcal{U}_{\mathcal{I}}(B_1)$ , and has an area differing by no more than $S$ compared to $B_{1}$ . + +Based on the bounding boxes $B_{1}$ and $\hat{B}_{2}$ , we crop the sub-image $I_{1}^{s}$ and $I_{2}^{s}$ from $I$ , respectively. Then, we input $(Q, I, I_{1}^{s})$ and $(Q, I, I_{2}^{s})$ to the model to obtain final + +responses $R_{1}$ and $R_{2}$ , separately. At this point, we have completed the sampling of two distinct CoT reasoning paths. + +Then, we should evaluate whether the paths are good or not. VIsRL uses the original model before SFT stage $\mathcal{M}_{org}$ as a critic to score the pairs $B_{1}, B_{2}$ and $R_{1}, R_{2}$ . Scores for the bounding boxes and the final responses are denoted with $s_1^b$ , $s_2^b$ and $s_1^r$ , $s_2^r$ , respectively. For more details including prompt designs, please refer to Sec. 3 in the Supp. Mat.. + +Data filtering. For each input question and image, repeating the above data generation process $N$ times, we will obtain a set of $2N$ candidates $P = \{p_1,p_2,\dots ,p_{2N}\}$ , where each $p$ contains original question, image, one bounding box, cropped sub-image, the final response and two scores, i.e., $p_i = (Q,I,B_i,I_i^s,R_i,s_i^b,s_i^r)$ . + +Moreover, to ensure the validity of preference data pairs in the candidate set $P$ , we apply filtering by setting win and lose thresholds, $T_{max}^{b}$ and $T_{min}^{b}$ for bounding boxes, $T_{max}^{r}$ and $T_{min}^{r}$ for responses: + +$$ +P _ {w i n} = \left\{p _ {i} \mid s _ {i} ^ {b} \geq \mathcal {T} _ {\max } ^ {b} \text {a n d} s _ {i} ^ {r} \geq \mathcal {T} _ {\max } ^ {r} \right\} \tag {4} +$$ + +$$ +P _ {l o s e} = \left\{p _ {i} \mid s _ {i} ^ {b} < \mathcal {T} _ {\min } ^ {b} \text {a n d} s _ {i} ^ {r} < \mathcal {T} _ {\min } ^ {r} \right\} +$$ + +For the current question and image $(Q, I)$ , if the win set and the loss set is not empty, this question and image and its corresponding generated data are preserved. The intuition here is that we apply a filter to the questions in the dataset, selecting those with a difficulty level suitable for the current model. Questions that are too difficult will prevent the model from generating meaningful answers, while questions that are too easy will result in the model producing correct answers across the board. Both of these types of questions will be filtered out during the current training round. However, as the model's capabilities are updated and strengthened, less data are filtered in the next iteration, as indicated by Data Num. of VisRL-Full vs. VisRL-Full-Iter1 in Tab. 6. + +Then, for each preserved question, we select the most representative path from each set to obtain the win-lose preference pairs, denoted as $(p_{win}, p_{lose}) = (\mathcal{C}_{max}(P_{win}), \mathcal{C}_{min}(P_{lose}))$ . In the case of $\mathcal{C}_{max}$ , it is required that both $s^b$ and $s^r$ of $p_i$ are the maximum over the other elements in the set $P_{win}$ . Conversely, $\mathcal{C}_{min}$ requires that both scores are the minimum in the set $P_{lose}$ . If the condition of being simultaneously maximum/minimum on both $s^b$ and $s^r$ is not met, the data point will also be discarded. Finally, we obtain the preference dataset $D_P = \{(p_{win_1}, p_{lose_1}), (p_{win_2}, p_{lose_2})\ldots\}$ + +# 3.2. Reinforced Visual Reasoning + +Similar with [27], V1sRL uses a step-level DPO method. It is divided into two stages. Stage 1 involves optimizing the bounding box, while stage 2 focuses on the joint optimization of the bounding box and the final response. For stage 1, given + +the input question-image $x = (Q, I)$ , the objective is: + +$$ +\mathcal {L} _ {\text {s t a g e 1}} (\theta) = - \mathbb {E} _ {(x, B _ {\text {w i n}}, B _ {\text {l o s e}}) \sim D _ {P}} \left(P _ {B} (\theta)\right), \tag {5} +$$ + +where each pair-wise preference paths in $D_P$ consists of bounding boxes $B_{win}$ and $B_{lose}$ , the formulation of bounding box preference probability is: + +$$ +P _ {B} (\theta) = \log \sigma \left(\beta_ {1} \log \frac {\pi_ {\theta} (B _ {w i n} \mid x)}{\pi_ {r e f} (B _ {w i n} \mid x)} - \beta_ {1} \log \frac {\pi_ {\theta} (B _ {l o s e} \mid x)}{\pi_ {r e f} (B _ {l o s e} \mid x)}\right). +$$ + +After stage1, the policy model updated from $\pi_{\theta}$ to $\pi_{\hat{\theta}}$ , while the reference model is updated to $\pi_{\hat{r e f}}$ . Then, for stage2, we further consider the cropped-image from $B_{win}$ to make CoT inference, that is $\hat{x} = (Q,I,I^{s}_{win})$ , then the formulation of response preference probability is: + +$$ +P _ {R} (\hat {\theta}) = \log \sigma \left(\beta_ {2} \log \frac {\pi_ {\hat {\theta}} \left(R _ {w i n} \mid \hat {x}\right)}{\pi_ {\hat {r e f}} \left(R _ {w i n} \mid \hat {x}\right)} - \beta_ {2} \log \frac {\pi_ {\hat {\theta}} \left(R _ {l o s e} \mid \hat {x}\right)}{\pi_ {\hat {r e f}} \left(R _ {l o s e} \mid \hat {x}\right)}\right), +$$ + +where $R_{win}$ and $R_{lose}$ are the final preference responses in $D_P$ . Based on Eq. 6 and Eq. 7, the objective for jointly optimizing the bounding boxes and the responses in stage 2 can be formulated as: + +$$ +\mathcal {L} _ {\text {s t a g e 2}} (\hat {\theta}) = - \left(\lambda_ {B} \mathcal {L} _ {B} (\hat {\theta}) + \lambda_ {R} \mathcal {L} _ {R} (\hat {\theta})\right), \tag {8} +$$ + +where: + +$$ +\mathcal {L} _ {B} (\hat {\theta}) = \mathbb {E} _ {(x, B _ {w i n}, B _ {l o s e}) \sim D _ {P}} \left(P _ {B} (\hat {\theta})\right), \tag {9} +$$ + +$$ +\mathcal {L} _ {R} (\hat {\theta}) = \mathbb {E} _ {(\hat {x}, R _ {w i n}, R _ {l o s e}) \sim D _ {P}} \left(P _ {R} (\hat {\theta})\right). \tag {10} +$$ + +# 4. Experiments + +# 4.1. Comparisons with Baselines + +We evaluate our method with several state-of-the-art methods on an array of different categories as follows. More details are in Supp. Mat.. As shown in Tab. 1, we categorize methods into three groups based on the types of LLM and vision encoder, then evaluate them on comprehensive benchmarks (MME, MMBench) as well as hallucination benchmarks (POPE). In all cases, our method achieves either the best or second-best performance, demonstrating the robust and well-rounded improvement over the baseline. In contrast, other methods exhibit performance drops on specific benchmarks (e.g., SEAL on MMBench, VisCoT on MME). We attribute this phenomenon to the limitations inherent in their training approaches – data-driven SFT struggles to generalize effectively, while tool-usage methods suffer from intrinsic shortcomings on certain datasets. Notably, our approach achieves highly comprehensive and promising results while using only $30\mathrm{k}$ dense-labeled (w. bounding boxes) samples – significantly fewer than other methods, and without relying on any external capabilities. In particular, under Vicuna-7B + +Table 1. The evaluation of different baselines on MME [18], MMBench [36], and POPE [31] datasets. Datasets marked with [D] are dense-labeled datasets (e.g., CoT data). In different methods, [B] denotes the base model, [D] represents data-driven SFT methods, and [T] refers to tool-usage methods (e.g., agents). The best is highlighted and the second-best is underlined. Remark: the data number considered here includes only the data used to enhance specific model capabilities and pretraining data, excluding general instruction-tuning dataset. + +
MethodLLMVision EncoderMMEMMBenchPOPEDataset Num.
LLaVA [B] [34]Vicuna-7B [13]CLIP-ViT-L-14-224 [44]1051.234.476.5558K
SEAL [D] [61]Vicuna-7BCLIP-ViT-L-14-2241128.933.182.4558K + 387K [D]
LLaVA + P2G [T] [8]Vicuna-7BCLIP-ViT-L-14-2241223.0——558K + 427K [D]
LLaVA + VIsRLVicuna-7BCLIP-ViT-L-14-2241183.837.578.2558K + 30K [D]+180K
LLaVA + VIsRL-Iter1Vicuna-7BCLIP-ViT-L-14-2241238.338.680.4180K
LLaVA-1.5 [B] [33]Vicuna-7BCLIP-ViT-L-14-3361510.764.385.8558K
VisCoT [D] [48]Vicuna-7BCLIP-ViT-L-14-3361453.667.986.0558K + 376K [D]
LLaVA-1.5 + VIsRLVicuna-7BCLIP-ViT-L-14-3361526.370.187.5558K + 30K [D]+180K
LLaVA-1.5 + VIsRL-Iter1Vicuna-7BCLIP-ViT-L-14-3361560.071.788.8180K
LLaVA-NeXT [B] [35]Vicuna-7B-1.5 [72]CLIP-ViT-L-14-3361611.172.3—558K
VisionLLM v2 [D] [60]Vicuna-7B-1.5CLIP-ViT-L-14-3361512.577.187.5892K
Insight-V-LLaVA [T] [15]Vicuna-7B-1.5CLIP-ViT-L-14-3361583.981.7—558K + 215K [D]
LLaVA-NeXT + VIsRLVicuna-7B-1.5CLIP-ViT-L-14-3361619.278.888.4558K + 30K [D]+180K
LLaVA-NeXT + VIsRL-Iter1Vicuna-7B-1.5CLIP-ViT-L-14-3361637.080.089.3180K
+ +Table 2. Performance on the different benchmarks. Base model is evaluated without CoT while others leverage CoT during inference. The best results from different LMMs are highlighted. + +
Training PhaseBaseSFTSFT+RL1SFT+RL1+RL2
LLaVA-1.5-7B [33]0.4440.5380.5680.605
LLaVA-NeXT-7B [35]0.4590.5560.5930.628
Llama-3.2-V-11B [39]0.6350.7240.7490.786
MiniCPM-o-2.6-8B [66]0.4280.5190.5440.591
PaliGemma2-10B [50]0.3770.4790.5070.562
Yi-VL-6B [67]0.3890.4920.5230.573
Qwen2.5-VL-7B [3]0.6400.7300.7630.796
+ +and CLIP-ViT-L-14-336, our method outperforms VisCoT across all benchmarks – the representative data-driven SFT approach. Specifically, our method outperforms VisCoT by $5.00\%$ (1526.3 vs. 1453.6) on MME and by $1.74\%$ (87.5 vs. 86.0) on the hallucination benchmark POPE. Moreover, after 1 iteration (Base Model + V1sRL–Iter1), V1sRL improves performance by $1\%$ to $4\%$ across all benchmarks. + +# 4.2. Results on Visual CoT Benchmark + +In this section, we comprehensively investigate different training phases in VIsRL across various base LMMs. We use Visual CoT benchmark here, which primarily focuses on scenarios where the LMM needs to concentrate on specific regions within a complete image. + +Settings. For SFT, the objective is to regularize the model with the capability of outputting bounding box, while RL is to enhance the model's visual perception capabilities via rewards. To train the LMM with outputting bounding box in SFT phase, we add a CoT prompt ("Please first provide the bounding box coordinate of the region, then refer to the corresponding sub-image to answer the question better.") to the question, asking the model to identify the most informative region of the image. As shown in Tab. 3, we found that the model has already been capable of outputting bounding boxes under SFT with the dataset of $30k$ . Then, we sequentially proceed with further training using stage1 (RL1) and + +stage2 (RL2) as described in Sec. 3.2. + +Results. As shown in Tab. 2, there remains substantial room for improvement after SFT. Tab. 1 in Supp. Mat. details the results, indicates that there is a decline on some datasets even with the use of visual CoT (e.g. InfogVQA), which further corroborates the validity of our revised SFT strategy. Besides, we found that the difference in results obtained from SFT with $30\mathrm{k}$ or $438\mathrm{k}$ is not significant (i.e., VisCoT (0.574) vs. LLaVA-1.5-7B-Ours-SFT (0.538)). These suggests that: on the one hand, the improvement in model capability is more attributed to the introduction of CoT rather than the SFT memory. On the other hand, the model's capability has already achieved saturated in a data-driven SFT manner and fails to further generalize, which is also the reason why Visual CoT fails in some OOD scenarios. However, our RL method can achieve comprehensive enhancement, with the promising improvement of up to $49.07\%$ (PaliGemma2-10B: 0.377 vs. 0.562), and the minimum improvement of $23.78\%$ (Llama-3.2-Vision-11B: 0.635 vs. 0.786). Meanwhile, we does not rely on a large amount of dense-labeled data (i.e. with bounding box annotation), but still has learned more essential visual perception. + +Detection Ability. Our approach is grounded in CoT for visual reasoning, thereby placing significant emphasis on the accuracy of intermediate bounding boxes. To substantiate the enhancement in bounding box precision following the implementation of our RL method, we present the detection performance in Tab. 4. Specifically, we compute the IoU between the predicted CoT bounding boxes and the corresponding GT bounding boxes, deeming a prediction correct if the IoU value surpasses 0.5. It is evident that, when using the same base model – LLaVA-1.5-7B, the performance after our SFT with 30k data is somewhat inferior to that of Visual CoT which leverages 438k data. However, during the RL phase, we get rid of bounding box annotation data and exclusively utilized 180k simple question-answer + +Table 3. Ratio of successful bounding box outputs of different SFT data number in terms of Qwen2.5-VL-7B. We evaluate on Visual CoT benchmark with 8281 data. + +
SFT Data Num.10k20k30k50k100k
Ratio28.14%95.48%99.87%99.87%99.87%
+ +![](images/3e52be193cc8b79c46bc32a582af7c47a766f94590a0ba60c8c7468929d6e336.jpg) +Figure 4. Performance of our VisRL over multiple iterations, attributing to the intertwined improvement of data quality and model capability during the iterative process. The accuracy is calculated as the average value over the 11 datasets listed in Tab. 2. + +pairs (further processed via our data generation pipeline in Sec. 3.1 to construct preference data), thereby achieving an accuracy improvement of $15.61\%$ (0.437 vs. 0.378) over Visual CoT. Moreover, even with only RL1, we still attained an approximate $8.47\%$ (0.410 vs. 0.378) improvement. Notably, on the DocVQA dataset (which was not included in the RL training phase), RL1 achieved a remarkable $39.71\%$ (0.190 vs. 0.136) improvement, while RL2 accomplished an impressive $73.68\%$ (0.231 vs. 0.136) improvement. We attribute these substantial gains to the robust generalization capability of our RL method. + +# 4.3. Performance across Multiple RL Iterations + +Fig. 4 illustrates the performance variation curves of our method over multiple iterations on LLa-1.5-7B and Qwen2.5-VL-7B. It can be observed that after each iteration (i.e., data is regenerated and VIsRL training is recond. ducted), performance improves significantly, regardless of whether only RL1 is applied or the full process of RL1+RL2 is used. The performance gains range from a minimum of $0.3\%$ to a maximum of $1.8\%$ . Since Qwen2.5-VL-7B is closer to the performance upper bound, its growth is relatively slower. Notably, after 4 iterations, LLa-1.5-7B even surpasses Llama-3.2-Vision-11B, as indicated by the blue curve. This further validates the promising potential of our method and lays the experimental foundation for future optimizations in online algorithms. + +# 4.4. Visualization + +This section presents the qualitative performance of our VisRL in Fig. 5, highlighting the accuracy of our method in identifying critical regions within images and then aid + +in CoT reasoning. Compared to VisCoT (red), our VISRL (green) demonstrates greater performance in both localizing the regions of interest and generating the final response, as evidenced by the ground truth (GT) comparison (blue). + +![](images/9ee861e8ed98cf13ece618bb90d5f4d455fd8e216d232f7d832e48415ebd8f05.jpg) +Question: What piece of furniture is to the right of the curtains? +GT Answer: bed + +![](images/360d1df4760d502f2d48965b02049591921ac63f7ba1852526a94d0e03a61611.jpg) +Question: what names are the ad? GT Answer: klaviarabend wilhelm backhaus, wolfgang amadeus mozart, franz schubert, johannes brahms, ludwig van beethoven +LLaVA-1.5 Answer [score: 0.2]: Wilhelm Backhaus. VisCoT Answer [score: 0.4]: The names on the ad are Wilhelm Backhaus and Johannes Brahm. VisRI Answer [score: 1.0]: franz schubert, Johannes brahm, wilhelm backhaus, wolfgang amadeus mozart, ludwig van beethoven + +![](images/2798f2299e3ba0145307cfeef9aab3bb9204599590eb35a59265e37695378915.jpg) +LLAVA-1.5 Answer [score: 0.0]: Chair VisCoT Answer [score: 0.0]: Chair VisRL Answer [score: 1.0]: bed +Question: What is providing shade for the man in the image? +GT Answer: The man is sitting under a giant umbrella that is providing shade. +LLaVA-1.5 Answer [score: 0.0]: The man in the image is sitting under a awning. +VisCoT Answer [score: 0.0]: The man is sitting under a large awning, which is providing shade for him. +VisRL Answer [score: 1.0]: The man is standing under an umbrella, which is providing shade. The umbrella is black and white and has the words "New Amsterdam" on it. +Figure 5. Visualization of LLaVa-1.5 vs. VisCoT vs. VisRL (based on LLaVa-1.5). GT bounding boxes are shown in blue, VisCoT-generated bounding boxes are shown in red, while Ours-generated bounding boxes are in green. The scores are evaluated by the GPT40. Our method consistently delivers the best results across various benchmarks. More visualizations are in the Supp. Mat.. + +# 4.5. Ablation Study + +Different stages. In Tab. 2, we have conducted the ablation study on the usage of different training stages (SFT, SFT+RL1, SFT+RL1+RL2), demonstrating a consistent performance improvement from SFT to RL1 and then to RL2. Furthermore, in Tab. 5, we present the results of training with RL2-Only after SFT with bounding boxes. As observed, applying RL2-Only training already yields a promising improvement compared to SFT-then-CoT (0.759 vs. 0.730). However, it still results in significant performance drops on certain benchmarks compared to the original model - Qwen2.5-VL-7B (e.g., DocVQA, DUDE, InfogVQA). This suggests that directly optimizing both bounding-box detection and response generation in a joint manner poses certain challenges. Therefore, introducing RL1 first to optimize bounding-box detection separately is necessary. + +Different training strategy. We attempted to directly learn the final response in Tab. 5, i.e., without the CoT setting. In this experiment, we used the same 30k dataset $D_P$ . Specifically, SFT was trained solely on the chosen response (w/o bounding box); DPO [45] was trained on the win/lose preference response; Kahneman-Tversky Optimisation (KTO) [16] assigned "true" and "false" labels to "win" and "lose" response, respectively; and Proximal Policy Optimization (PPO) [46] sampled 15k preference data to train the reward + +Table 4. Detection performance (Top-1 Accuracy@0.5) on the various benchmark, where both "Ours" and "Visual-CoT" similarly utilize LLaVA-1.5-7B as the base model. The amount of dense-labeled data with bounding box annotations is indicated in []. The best is highlighted. + +
MethodDoc/TextChartGeneral VQARelation ReasoningFine-grained
DocVQATextCapsTextVQADUDESROIEInfogVQAFlickr30kGQAOpen imagesVSRCUBAvg
VisCoT [438k] [48]0.1360.4130.4680.0500.1570.0720.4960.4200.5760.6960.6700.378
VisRL-SFT [30k]0.1330.4060.4600.0480.1460.0690.4950.4210.5750.6910.6730.374
VisRL-SFT+RL10.1900.4330.4710.0810.2160.1330.5180.4500.5810.7330.7010.410
VisRL-SFT+RL1+RL20.2310.4640.5040.1240.2520.1720.5190.4870.5900.7450.7150.437
+ +Table 5. Different training strategies for directly fitting the final response in terms of Qwen2.5-VL-7B. Besides, we also ablated on RL2-Only. The red background indicates a decline in performance compared to the original model, while the best is highlighted. + +
MethodDoc/VQADoc/TextChartGeneral VQARelation ReasoningFine-grained
TextCapsTextVQADUDESROIEInfogVQAFlickr30kGQAOpen imagesVSRCUBAvg
Ori0.8360.7600.8470.6060.7890.6850.6010.4670.2890.5810.5830.640
SFT0.7780.7510.8560.5660.7130.6310.6270.5240.3450.6260.7360.650
DPO [45]0.7910.7930.8370.5990.7760.6780.6700.4700.2980.5940.7070.656
KTO [16]0.7970.7900.8330.5970.7800.6940.6710.4690.2990.5990.7070.658
PPO [46]0.7940.7910.8370.5970.7870.7000.6720.4720.3020.6060.7130.661
VisRL-RL2-Only0.8160.7690.8970.6020.8000.6810.6370.6520.7860.8270.8820.759
VisRL-Full0.8740.8190.8970.6400.8290.7530.6750.7000.8140.8640.8920.796
+ +model, while the remaining 15k data were used to train PPO. The results indicate that both previous RL and SFT methods exhibit limited effectiveness in directly fitting the final response, with some benchmarks even showing performance degradation (e.g. DocVQA, DUDE, etc.). In contrast, our training approach shown in Tab. 2, which incorporates the CoT reasoning in RL method, consistently improves model performance across all benchmarks. + +Data Generation Pipeline. Our data generation process follows a self-evolution paradigm and adopts an actor-critic framework, where the actor is the SFTed model and the critics refer to the pre-SFT model. In this study, we conduct ablation experiments on several modules within the pipeline. As shown in Tab. 6, our objective is to maximize the proportion of positive instances in the constructed preference dataset, specifically WP-LN (win-positive, lose-negative), while minimizing the proportion of negative instances, WN-LP (win-negative, lose-positive). Experimental results indicate that under our pipeline, the proportion of positive instances is $64.64\%$ . After the first and second iterations – where the model undergoes full V1sRL training and then participates in data construction again – this proportion increases by $3.18\%$ and $2.30\%$ , respectively. Beside, the number of valid data increases from 30k to 33k to 35k. This is because, through iterative training, the model's capability improves to overcome bottlenecks where obtaining the correct answer was previously outside of its ability. Specifically, some questions are inherently too difficult for model previously, meaning that no matter how many times the model responds, the answer remains incorrect, making it impossible to construct win/lose preference. + +When replacing the critics with GPT-4o, the proportion of positive instances remains largely unchanged (64.64% vs. 65.31%), but this comes at the cost of increased token consumption and reduced response speed. Conversely, when substituting the critics with the SFTed model itself, the + +Table 6. Ablation of our data generation pipeline in terms of Qwen2.5-VL-7B. We evaluate the data quality by comparing the IoU (Top-1 Accuracy@0.5) between the annotated preference data's bounding boxes and the GT bounding boxes. "W" and "L" represent the win and loss in the data annotation, respectively, while "P" and "N" indicate positive or negative. Ideally, we expect WP-LN to be as large as possible, as highlighted in green. + +
WP-LPWP-LNWN-LPWN-LNData Num.
w GPT-4o-2024-11-200.00%65.31%1.32%33.37%47k
w SFTed Model0.00%54.68%0.00%45.32%3k
w/o Bounding Box Critics5.42%31.02%10.04%53.51%86k
w/o Diversity Controller4.53%52.02%4.68%38.77%19k
V1sRL-Full0.43%64.64%1.64%33.29%30k
V1sRL-Full-Iter10.45%67.82%0.82%30.91%33k
V1sRL-Full-Iter20.47%70.12%0.00%29.41%35k
+ +data volume decreases to 1/10 of the original (3k vs. 30k), accompanied by a $9.96\%$ reduction in the proportion of positive instances. Furthermore, omitting the evaluation of the generated bounding boxes results in a $33.62\%$ drop in the proportion of positive instances, while the absence of the diversity controller leads to a $12.62\%$ reduction. Considering these factors, we adopt the approach described in Sec. 3.1. + +# 5. Conclusion + +In this paper, we propose VIsRL, a framework for learning intention-driven visual perception abilities from task feedback. This approach enables the model to undergo RL through self-evolution when trained on simple data with only final responses. Specifically, we introduce a novel pipeline for generating CoT preference data based on the model's own actor-critic process, eliminating the need for external models or human annotations. Using these self-constructed data, we further optimize visual perception in two RL stages: (1) independently optimizing generated bounding box; (2) jointly optimizing both generated bounding box and response. Extensive experiments on various benchmarks and LMMs demonstrate the effectiveness of the proposed framework, establishing a solid foundation for future exploration. + +# References + +[1] Ebtesam Almazrouei, Ruxandra Cojocaru, Michele Baldo, Quentin Malartic, Hamza Alobeidli, Daniele Mazzotta, Guilherme Penedo, Giulia Campesan, Mugariya Farooq, Maitha Alhammadi, et al. Alghafa evaluation benchmark for arabic language models. In Proceedings of ArabicNLP 2023, pages 244-275, 2023. 2 +[2] Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, et al. Qwen technical report. arXiv preprint arXiv:2309.16609, 2023. 2 +[3] Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, et al. Qwen2. 5-vl technical report. arXiv preprint arXiv:2502.13923, 2025. 6 +[4] Xiao Bi, Deli Chen, Guanting Chen, Shanhuang Chen, Damai Dai, Chengqi Deng, Honghui Ding, Kai Dong, Qiushi Du, Zhe Fu, et al. Deepseek llm: Scaling open-source language models with longtermism. arXiv preprint arXiv:2401.02954, 2024. 2 +[5] Mahtab Bigverdi, Zelun Luo, Cheng-Yu Hsieh, Ethan Shen, Dongping Chen, Linda G Shapiro, and Ranjay Krishna. Perception tokens enhance visual reasoning in multimodal language models. arXiv preprint arXiv:2412.03548, 2024. 3 +[6] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 2 +[7] Zheng Cai, Maosong Cao, Haojiong Chen, Kai Chen, Keyu Chen, Xin Chen, Xun Chen, Zehui Chen, Zhi Chen, Pei Chu, et al. Internl m2 technical report. arXiv preprint arXiv:2403.17297, 2024. 2 +[8] Jiaxing Chen, Yuxuan Liu, Dehu Li, Xiang An, Weimo Deng, Ziyong Feng, Yongle Zhao, and Yin Xie. Plug-and-play grounding of reasoning in multimodal large language models. arXiv preprint arXiv:2403.19322, 2024. 3, 6 +[9] Lin Chen, Jinsong Li, Xiaoyi Dong, Pan Zhang, Conghui He, Jiaqi Wang, Feng Zhao, and Dahua Lin. Sharegpt4v: Improving large multi-modal models with better captions. In European Conference on Computer Vision, pages 370-387. Springer, 2024. 1 +[10] Liang Chen, Lei Li, Haozhe Zhao, Yifan Song, and Vinci. R1-v: Reinforcing super generalization ability in vision-language models with less than $3. https://github.com/Deep-Agent/R1-V, 2025. Accessed: 2025-02-02. 3 +[11] Lin Chen, Xilin Wei, Jinsong Li, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Zehui Chen, Haodong Duan, Zhenyu Tang, Li Yuan, et al. Sharegpt4video: Improving video understanding and generation with better captions. Advances in Neural Information Processing Systems, 37:19472-19495, 2025. 1 +[12] Yukang Chen, Shengju Qian, Haotian Tang, Xin Lai, Zhi-jian Liu, Song Han, and Jiaya Jia. Longlora: Efficient fin-tuning of long-context large language models. arXiv preprint arXiv:2309.12307, 2023. 1 + +[13] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E Gonzalez, et al. Vicuna: An open-source chatbot impressing gpt-4 with $90\%$ * chatgpt quality, march 2023. URL https://lmsys.org/blog/2023-03-30-vicuna, 3(5), 2023.6 +[14] Paul F Christiano, Jan Leike, Tom Brown, Miljan Martic, Shane Legg, and Dario Amodei. Deep reinforcement learning from human preferences. Advances in neural information processing systems, 30, 2017. 3 +[15] Yuhao Dong, Zuyan Liu, Hai-Long Sun, Jingkang Yang, Winston Hu, Yongming Rao, and Ziwei Liu. Insight-v: Exploring long-chain visual reasoning with multimodal large language models. arXiv preprint arXiv:2411.14432, 2024. 3, 6 +[16] Kawin Ethayarajh, Winnie Xu, Niklas Muennighoff, Dan Jurafsky, and Douwe Kiela. Kto: Model alignment as prospect theoretic optimization. arXiv preprint arXiv:2402.01306, 2024.7,8 +[17] Luciano Floridi and Massimo Chiriatti. Gpt-3: Its nature, scope, limits, and consequences. *Minds and Machines*, 30: 681–694, 2020. 2 +[18] Chaoyou Fu, Peixian Chen, Yunhang Shen, Yulei Qin, Mengdan Zhang, Xu Lin, Jinrui Yang, Xiawu Zheng, Ke Li, Xing Sun, Yunsheng Wu, and Rongrong Ji. Mme: A comprehensive evaluation benchmark for multimodal large language models, 2024. 6 +[19] Team GLM, Aohan Zeng, Bin Xu, Bowen Wang, Chenhui Zhang, Da Yin, Dan Zhang, Diego Rojas, Guanyu Feng, Hanlin Zhao, et al. Chatglm: A family of large language models from glm-130b to glm-4 all tools. arXiv preprint arXiv:2406.12793, 2024. 2 +[20] Anisha Gunjal, Jihan Yin, and Erhan Bas. Detecting and preventing hallucinations in large vision language models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 18135-18143, 2024. 1 +[21] Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, et al. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948, 2025.3 +[22] Yanming Guo, Yu Liu, Ard Oerlemans, Songyang Lao, Song Wu, and Michael S Lew. Deep learning for visual understanding: A review. Neurocomputing, 187:27-48, 2016. 1 +[23] Jiaming Ji, Jiayi Zhou, Hantao Lou, Boyuan Chen, Donghai Hong, Xuyao Wang, Wenqi Chen, Kaile Wang, Rui Pan, Jiahao Li, Mohan Wang, Josef Dai, Tianyi Qiu, Hua Xu, Dong Li, Weipeng Chen, Jun Song, Bo Zheng, and Yaodong Yang. Align anything: Training all-modality models to follow instructions with language feedback. 2024. 3 +[24] Fengqing Jiang. Identifying and mitigating vulnerabilities in llm-integrated applications. Master's thesis, University of Washington, 2024. 2 +[25] Yang Jiao, Shaoxiang Chen, Zequn Jie, Jingjing Chen, Lin Ma, and Yu-Gang Jiang. Lumen: Unleashing versatile vision-centric capabilities of large multimodal models. Advances in Neural Information Processing Systems, 37:81461-81488, 2025. 1 + +[26] Xin Lai, Zhuotao Tian, Yukang Chen, Yanwei Li, Yuhui Yuan, Shu Liu, and Jiaya Jia. Lisa: Reasoning segmentation via large language model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9579-9589, 2024. 1 +[27] Xin Lai, Zhuotao Tian, Yukang Chen, Senqiao Yang, Xiangru Peng, and Jiaya Jia. Step-dpo: Step-wise preference optimization for long-chain reasoning of llms. arXiv preprint arXiv:2406.18629, 2024. 5 +[28] Chengzhu Li, Wenshan Wu, Huanyu Zhang, Yan Xia, Shaoguang Mao, Li Dong, Ivan Vulic, and Furu Wei. Imagine while reasoning in space: Multimodal visualization-of-thought. arXiv preprint arXiv:2501.07542, 2025. 3 +[29] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International conference on machine learning, pages 12888-12900. PMLR, 2022. 3 +[30] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730-19742. PMLR, 2023. 3 +[31] Yifan Li, Yifan Du, Kun Zhou, Jinpeng Wang, Wayne Xin Zhao, and Ji-Rong Wen. Evaluating object hallucination in large vision-language models. arXiv preprint arXiv:2305.10355, 2023. 6 +[32] Yanwei Li, Chengyao Wang, and Jiaya Jia. Llama-vid: An image is worth 2 tokens in large language models. In European Conference on Computer Vision, pages 323–340. Springer, 2024. 1 +[33] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning, 2023. 6 +[34] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning, 2023. 1, 3, 6 +[35] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava-last: Improved reasoning,OCR,and world knowledge,2024.6 +[36] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, et al. Mmbench: Is your multi-modal model an all-around player? In European conference on computer vision, pages 216-233. Springer, 2024. 6 +[37] Yuecheng Liu, Dafeng Chi, Shiguang Wu, Zhangguang Zhang, Yaochen Hu, Lingfeng Zhang, Yingxue Zhang, Shuang Wu, Tongtong Cao, Guowei Huang, et al. Spatialcot: Advancing spatial reasoning through coordinate alignment and chain-of-thought for embodied task planning. arXiv preprint arXiv:2501.10074, 2025. 3 +[38] Ziyu Liu, Zeyi Sun, Yuhang Zang, Xiaoyi Dong, Yuhang Cao, Haodong Duan, Dahua Lin, and Jiaqi Wang. Visual-rft: Visual reinforcement fine-tuning. arXiv preprint arXiv:2503.01785, 2025.3 +[39] Llama Meta. 3.2: Revolutionizing edge ai and vision with open, customizable models, 2024. URL: https://ai.meta. com/blog/llama-3-2-connect-2024-vision-edge-mobile-devices. 6 + +[40] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. Advances in neural information processing systems, 35:27730-27744, 2022. 3 +[41] Thomas J Palmeri and Isabel Gauthier. Visual object understanding. Nature Reviews Neuroscience, 5(4):291-303, 2004. 1 +[42] Alec Radford, Karthik Narasimhan, Tim Salimans, Ilya Sutskever, et al. Improving language understanding by generative pre-training. 2018. 2 +[43] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 2 +[44] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PmLR, 2021. 6 +[45] Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model. Advances in Neural Information Processing Systems, 36, 2024. 3, 7, 8 +[46] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347, 2017. 7, 8 +[47] Hao Shao, Yuxuan Hu, Letian Wang, Guanglu Song, Steven L Waslander, Yu Liu, and Hongsheng Li. Lmdrive: Closed-loop end-to-end driving with large language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15120-15130, 2024. 1 +[48] Hao Shao, Shengju Qian, Han Xiao, Guanglu Song, Zhuofan Zong, Letian Wang, Yu Liu, and Hongsheng Li. Visual cot: Unleashing chain-of-thought reasoning in multi-modal language models, 2024. 1, 3, 4, 6, 8 +[49] Haozhan Shen, Zilun Zhang, Qianqian Zhang, Ruochen Xu, and Tiancheng Zhao. Vlm-r1: A stable and generalizable r1-style large vision-language model. https://github.com/om-ai-lab/VLM-R1, 2025. Accessed: 2025-02-15. 3 +[50] Andreas Steiner, André Susano Pinto, Michael Tschannen, Daniel Keysers, Xiao Wang, Yonatan Bitton, Alexey Gritsenko, Matthias Minderer, Anthony Sherbondy, Shangbang Long, et al. Paligemma 2: A family of versatile vlms for transfer. arXiv preprint arXiv:2412.03555, 2024. 6 +[51] Chufeng Tang, Lingxi Xie, Xiaopeng Zhang, Xiaolin Hu, and Qi Tian. Visual recognition by request. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15265-15274, 2023. 1, 3 +[52] InternLM Team. Internlm: A multilingual language model with progressively enhanced capabilities, 2023. 2 +[53] Omkar Thawakar, Dinura Dissanayake, Ketan More, Ritesh Thawkar, Ahmed Heakl, Noor Ahsan, Yuhao Li, Mohammed + +Zumri, Jean Lahoud, Rao Muhammad Anwer, Hisham Cholakkal, Ivan Laptev, Mubarak Shah, Fahad Shahbaz Khan, and Salman Khan. Llamav-o1: Rethinking step-by-step visual reasoning in llms, 2025. 3 +[54] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 2 +[55] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 2 +[56] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 1 +[57] Wenhai Wang, Zhe Chen, Xiaokang Chen, Jiannan Wu, Xizhou Zhu, Gang Zeng, Ping Luo, Tong Lu, Jie Zhou, Yu Qiao, et al. Visionllm: Large language model is also an open-ended decoder for vision-centric tasks. Advances in Neural Information Processing Systems, 36:61501-61513, 2023. 1 +[58] XuDong Wang, Shaolun Zhang, Shufan Li, Konstantinos Kallidromitis, Kehan Li, Yusuke Kato, Kazuki Kozuka, and Trevor Darrell. Segllm: Multi-round reasoning segmentation. arXiv preprint arXiv:2410.18923, 2024. 3 +[59] Jeremy M Wolfe and Todd S Horowitz. Five factors that guide attention in visual search. Nature human behaviour, 1 (3):0058, 2017. 1 +[60] Jiannan Wu, Muyan Zhong, Sen Xing, Zeqiang Lai, Zhaoyang Liu, Zhe Chen, Wenhai Wang, Xizhou Zhu, Lewei Lu, Tong Lu, et al. Visionllm v2: An end-to-end generalist multimodal large language model for hundreds of vision-language tasks. Advances in Neural Information Processing Systems, 37:69925-69975, 2025. 3, 6 +[61] Penghao Wu and Saining Xie. V?: Guided visual search as a core mechanism in multimodal llms. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13084-13094, 2024. 3, 6 +[62] Wenshan Wu, Shaoguang Mao, Yadong Zhang, Yan Xia, Li Dong, Lei Cui, and Furu Wei. Mind's eye of llms: visualization-of-thought elicits spatial reasoning in large language models. Advances in Neural Information Processing Systems, 37:90277-90317, 2025. 3 +[63] Xueqing Wu, Yuheng Ding, Bingxuan Li, Pan Lu, Da Yin, Kai-Wei Chang, and Nanyun Peng. Visco: Benchmarking fine-grained critique and correction towards self-improvement in visual reasoning. arXiv preprint arXiv:2412.02172, 2024. 3 +[64] Ziang Yan, Zhilin Li, Yinan He, Chenting Wang, Kunchang Li, Xinhao Li, Xiangyu Zeng, Zilei Wang, Yali Wang, Yu Qiao, et al. Task preference optimization: Improving multimodal large language models with vision task alignment. arXiv preprint arXiv:2412.19326, 2024. 3 +[65] Aiyuan Yang, Bin Xiao, Bingning Wang, Borong Zhang, Ce Bian, Chao Yin, Chenxu Lv, Da Pan, Dian Wang, Dong Yan, + +et al. Baichuan 2: Open large-scale language models. arXiv preprint arXiv:2309.10305, 2023. 2 +[66] Yuan Yao, Tianyu Yu, Ao Zhang, Chongyi Wang, Junbo Cui, Hongji Zhu, Tianchi Cai, Haoyu Li, Weilin Zhao, Zhihui He, et al. Minicpm-v: A gpt-4v level mllm on your phone. arXiv preprint arXiv:2408.01800, 2024. 6 +[67] Alex Young, Bei Chen, Chao Li, Chengen Huang, Ge Zhang, Guanwei Zhang, Guoyin Wang, Heng Li, Jiangcheng Zhu, Jianqun Chen, et al. Yi: Open foundation models by 01. ai. arXiv preprint arXiv:2403.04652, 2024. 6 +[68] Hang Zhang, Xin Li, and Lidong Bing. Video-llama: An instruction-tuned audio-visual language model for video understanding. arXiv preprint arXiv:2306.02858, 2023. 1 +[69] Jiacheng Zhang, Yang Jiao, Shaoxiang Chen, Jingjing Chen, and Yu-Gang Jiang. Eventhallusion: Diagnosing event hallucinations in video llms. arXiv preprint arXiv:2409.16597, 2024. 1 +[70] Shilong Zhang, Peize Sun, Shoufa Chen, Minn Xiao, Wenqi Shao, Wenwei Zhang, Yu Liu, Kai Chen, and Ping Luo. Gpt4roi: Instruction tuning large language model on regionof-interest, 2024. In URL https://openreview.net/forum.1 +[71] Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher Dewan, Mona Diab, Xian Li, Xi Victoria Lin, et al. Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068, 2022. 2 +[72] Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric Xing, et al. Judging llm-as-a-judge with mt-bench and chatbot arena. Advances in Neural Information Processing Systems, 36:46595-46623, 2023. 6 \ No newline at end of file diff --git a/ICCV/2025/VisRL_ Intention-Driven Visual Perception via Reinforced Reasoning/images.zip b/ICCV/2025/VisRL_ Intention-Driven Visual Perception via Reinforced Reasoning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..d6c2612529e395c1807055bab609619d9be2b281 --- /dev/null +++ b/ICCV/2025/VisRL_ Intention-Driven Visual Perception via Reinforced Reasoning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e925c7828149a1d90bb1af4c9e9a5692a5c997484bd28d481bb18427c5322873 +size 599448 diff --git a/ICCV/2025/VisRL_ Intention-Driven Visual Perception via Reinforced Reasoning/layout.json b/ICCV/2025/VisRL_ Intention-Driven Visual Perception via Reinforced Reasoning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..1242a05c858b7abbfde276ba9021ef384025cdc3 --- /dev/null +++ b/ICCV/2025/VisRL_ Intention-Driven Visual Perception via Reinforced Reasoning/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9f5ce02a0faa16c79e28405746ed03f5da956d9d2629fd4593bdd97539281f6b +size 460568 diff --git a/ICCV/2025/Vision-Language Interactive Relation Mining for Open-Vocabulary Scene Graph Generation/f8f524f5-9035-412a-9d49-1b195879895c_content_list.json b/ICCV/2025/Vision-Language Interactive Relation Mining for Open-Vocabulary Scene Graph Generation/f8f524f5-9035-412a-9d49-1b195879895c_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..80a1651f7867e81ae25ab1a48ae27496e7bcad57 --- /dev/null +++ b/ICCV/2025/Vision-Language Interactive Relation Mining for Open-Vocabulary Scene Graph Generation/f8f524f5-9035-412a-9d49-1b195879895c_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f7ac96507a9d146d36af08457cff7e656a586813761a9d23953a48d6cc67c4ff +size 79215 diff --git a/ICCV/2025/Vision-Language Interactive Relation Mining for Open-Vocabulary Scene Graph Generation/f8f524f5-9035-412a-9d49-1b195879895c_model.json b/ICCV/2025/Vision-Language Interactive Relation Mining for Open-Vocabulary Scene Graph Generation/f8f524f5-9035-412a-9d49-1b195879895c_model.json new file mode 100644 index 0000000000000000000000000000000000000000..c30fdf628a5adbb0c1cba4989a0564d4fbe3779e --- /dev/null +++ b/ICCV/2025/Vision-Language Interactive Relation Mining for Open-Vocabulary Scene Graph Generation/f8f524f5-9035-412a-9d49-1b195879895c_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:68d35d39779450126bb252498d27f8b4bbb169aaf834809b6031f9765b8be8b4 +size 96275 diff --git a/ICCV/2025/Vision-Language Interactive Relation Mining for Open-Vocabulary Scene Graph Generation/f8f524f5-9035-412a-9d49-1b195879895c_origin.pdf b/ICCV/2025/Vision-Language Interactive Relation Mining for Open-Vocabulary Scene Graph Generation/f8f524f5-9035-412a-9d49-1b195879895c_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..3f783725c3320aa695993617affb6b9726f01526 --- /dev/null +++ b/ICCV/2025/Vision-Language Interactive Relation Mining for Open-Vocabulary Scene Graph Generation/f8f524f5-9035-412a-9d49-1b195879895c_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a7a07bf2bb3cd87f594a0a6c40d94bad2fe4daa951470e01a6ef530e77a3228b +size 10156362 diff --git a/ICCV/2025/Vision-Language Interactive Relation Mining for Open-Vocabulary Scene Graph Generation/full.md b/ICCV/2025/Vision-Language Interactive Relation Mining for Open-Vocabulary Scene Graph Generation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..d11a8cd1305fe65d1934494b1c47f976aa4a9674 --- /dev/null +++ b/ICCV/2025/Vision-Language Interactive Relation Mining for Open-Vocabulary Scene Graph Generation/full.md @@ -0,0 +1,339 @@ +# Vision-Language Interactive Relation Mining for Open-Vocabulary Scene Graph Generation + +Yukuan Min $^{1}$ , Muli Yang $^{\dagger,2}$ , Jinhao Zhang $^{1}$ , Yuxuan Wang $^{1}$ , Among Wu $^{1}$ , Cheng Deng $^{\dagger,1}$ + +$^{1}$ Xidian University, China + +$^{2}$ A*STAR, Singapore + +ykmin@stu.xidian.edu.cn, yangml@i2r.a-star.edu.sg, {zhangjinhao, wangyuxuan}@stu.xidian.edu.cn + +amwu@xidian.edu.cn, chdeng@mail.xidian.edu.cn + +# Abstract + +To promote the deployment of scenario understanding in the real world, Open-Vocabulary Scene Graph Generation (OV-SGG) has attracted much attention recently, aiming to generalize beyond the limited number of relation categories labeled during training and detect those unseen relations during inference. Towards OV-SGG, one feasible solution is to leverage the large-scale pre-trained vision-language models (VLMs) containing plentiful category-level content to capture accurate correspondences between images and text. However, due to the lack of quadratic relation-aware knowledge in VLMs, directly using the category-level correspondence in the base dataset could not sufficiently represent generalized relations involved in open world. Therefore, designing an effective open-vocabulary relation mining framework is challenging and meaningful. To this end, we propose a novel Vision-Language Interactive Relation Mining model (VL-IRM) for OV-SGG, which explores learning generalized relation-aware knowledge through multimodal interaction. Specifically, first, to enhance the generalization of the relation text to visual content, we present a generative relation model to make the text modality explore possible open-ended relations based on visual content. Then, we employ visual modality to guide the relation text for spatial and semantic extension. Extensive experiments demonstrate the superior OV-SGG performance of our method. + +# 1. Introduction + +To accelerate the understanding for visual content, Scene Graph Generation (SGG) has attracted much attention recently, aiming to predict the predicate relation between mul + +![](images/c03c4a6a5120e13771ae822441a9a63be98034affee444886133b26496b410da.jpg) +Figure 1. Illustration of the interactive visual-language model for OV-SGG. (a) Previous methods directly use category-level correspondence to adapt the vision-language models to open-vocabulary relations. (b) In this work, we consider the lack of quadratic relation-aware knowledge in VLMs, and construct an Interactive Relation Mining model for OV-SGG. + +tiple objects. Though existing methods have been verified to be effective, they usually follow a closed-set assumption, i.e., the training and testing data share the same predicate categories. However, the real scenario is open and full of unknown relations, limiting the application of close-set based methods. To this end, Open-Vocabulary Scene Graph Generation (OV-SGG) is proposed, whose goal is to leverage pre-trained models to extend the predicate space and improve the performance of identifying unknown relations. + +To tackle the OV-SGG problem and enhance the practical utility of models, a common strategy is to exploit the visual-text matching knowledge in the Vision-Language + +![](images/14892d62b2c3712c7f9485ad72bd1cfa2cbb407c3ce59600027d4c3abe0283ee.jpg) +Figure 2. Experimental results of OV-SGG on Visual Genome (VG) dataset [14] for VLM GLIP [19]. R and zR denote the recall metric in the base and novel relations. "Frozen Text" denotes only training the visual part of VLM and froze the text embedding. "Conventional" denotes directly training the VLM on the base dataset [43]. + +Models (VLMs) to expand the semantic space of the relations [5, 8, 21, 43]. Existing works usually construct extra weakly supervised pre-training on image-text pair datasets such as COCO [5, 43], using the pseudo relation label from the caption through the language parser. Moreover, recent advancements propose using an instruction prompt sequence, thus the model could more efficiently utilize the image-text pair knowledge of pre-trained VLMs or Large Language Models (LLMs) [18, 21]. + +Despite the remarkable achievements in the development of OV-SGG models, few works have addressed OV-SGG from the perspective of the VLM's structure. Since existing pre-trained VLMs lack relation-aware knowledge [5], directly building a VLM for OV-SGG is challenging. We demonstrate this phenomenon in Fig. 1 and construct experiments to certify in Fig. 2. On the one hand, the VLM without prior relation-aware knowledge performs poorly in directly generalizing open-vocabulary relations through existing text-embedding mappings. On the other hand, VLM would overfit the relation-aware knowledge in the base set, resulting in poor performance in unseen relations. Inspired by the construction process of existing VLMs [19, 25], we build a Vision-Language Interactive Relation Mining model (VL-IRM) to adapt to the OV-SGG task. We exploit the generative relation model based on visual content to allow the text modality to provide open-vocabulary relation knowledge to the VLM. And we use visual modalities to guide relational text expansion to alleviate VLM's overconfidence of the relation-aware knowledge in the base set. + +Specifically, as depicted in Fig. 1, we utilize an interactive relation mining model for the OV-SGG task. As shown in Fig. 3, our VL-IRM contains two major components: a relation encoder for visual relation feature extraction and a language model to generate relation names. Specifically, we first introduce a generative relation recognition module + +to make the text modality explore possible open-ended relations based on visual content. Then, we design a hierarchical relation extension module. On the one hand, for spatial extension, we exploit pseudo labels to employ the visual modality to guide the relation text supplement the sparse labels. On the other hand, based on the distribution estimation theories [28], we dynamically quantify the semantic bias in VLM and build an unbiased learning strategy for various relations based on the results, achieving semantic extension. + +The contributions can be summarized as follows, + +- We consider a new perspective for OV-SGG, i.e., optimizing the structure of the VLM. Unlike existing methods, this approach does not rely on a large amount of additional pre-training data or carefully set instruction prompts. +- We develop a novel interactive relation mining model for OV-SGG, which analyzes and solves the existing challenges of applying VLM on the OV-SGG task. +- Compared with existing OV-SGG models, VL-IRM achieves comparable results on the Visual Genome (VG) [14] dataset. + +# 2. Related Work + +Open-Vocabulary Scene Graph Generation. Recently, inspired by the recent advances in learning visual models from language supervision in image recognition tasks [10, 11, 30, 39], some works [5, 8, 21, 43] have extended the success in open-vocabulary image recognition to scene graph generation, i.e., Open-Vocabulary Scene Graph Generation (OV-SGG). SVRP [8] first proposes a novel open-vocabulary detector upon visual-language pretraining by formulating the classification as a binary matching between the input queries and the referring captions. To leverage the capability of CLIP [30], GEN-VLKT [22] and HOI-CLIP [29] turn to learn the visual region feature from classification-oriented models by knowledge distillation. Different from distilling frozen vision encoder, OVSGTR [5] proposes a two-step framework for image-caption grounding pre-training and end-to-end detection fine-tuning, using a new deformable DETR [45]. PGSG [21] proposes to extend the relation vocabulary with large-scale scene graph sequences generated from the images by applying the supervision of the vision-language model. Existing OV-SGG works typically rely on additional pre-training or carefully designed instructions to construct relation-aware knowledge in VLMs. In contrast, our method considers directly optimizing the structure of the existing VLMs, guiding the VLM to automatically construct open-vocabulary relation knowledge on a limited base training set. + +Multimodal Language Model. The multimodal language model has gained increasing attention recently for numerous vision-language tasks. Inspired by the remarkable capability of the language models (LM) [2-4, 6, + +![](images/19736a2e8df361981251e7474e4f93442e32f28d032723bb749b0ef9f35b8aeb.jpg) +Figure 3. Overview of our proposed VL-IRM. We construct an interactive relation mining model to extract open-vocabulary relation knowledge. The generative relation recognition module makes the text modality generalize the region-level visual content. The hierarchical relation extension module uses the visual modality's knowledge to extend the relation text. + +33, 35], previous works distill the knowledge from the LM by training the model with extensive image-text pairs [1, 17, 24, 25], resulting in enhanced performance across various vision-and-language tasks such as Vision Question Answering (VQA) and image captioning. BLIP-2 [17] proposes a generic and efficient pre-training strategy that bootstraps vision-language pre-training from frozen pretrained image encoders and frozen large language models. GroundingDINO [25] enables the model grounding objects in the images and supports a wider array of vision-language tasks including visual grounding and image captioning. Structure-CLIP [9] has been proposed for learning a fine-grained alignment for image region and language data and enables the LM reasoning from the images. The success of the multimodal language models and generative models [6, 33, 36, 37] on downstream tasks in various cases provides an exemplar for constructing a generative relation VLM to identify open-vocabulary relations in an image. + +# 3. Method + +# 3.1. OV-SGG Architecture + +As shown in Fig. 3, the OV-SGG architecture comprises three primary components: an image encoder $\mathrm{Enc}_I$ (e.g., Swin Transformer backbone [26]) for image feature extraction, a text encoder $\mathrm{Enc}_L$ (e.g., BERT [12]) for text feature extraction, and a relation classifier for relationship recognition. To reduce computation, the weights of both the image backbone and the text encoder are frozen during training. Given an image I, the bounding box features $\bar{O} \in \mathbb{R}^{\bar{N} \times d}$ is extracted by $\mathrm{Enc}_I$ . And the text of relation categories, (e.g., + +'on, in, wears,...walking') is fed into the text encoder $\mathrm{Enc}_L$ to get the text embeddings $T\in \mathbb{R}^{T\times d}$ + +We first utilize a transformer-based interaction classifier to encode the scene graphs in I. Specifically, taking the top-k most likely to be valid objects $\tilde{O}^{\prime}\in \mathbb{R}^{k\times d}$ as input, we construct relation representations $f_{i\rightarrow j}$ for all possible subject-object pairs $(\tilde{\mathbf{o}}_i,\tilde{\mathbf{o}}_j)$ : + +$$ +f _ {i \rightarrow j} = c a t \left[ f _ {i \rightarrow j} ^ {\text {v i s u a l}}, f _ {i \rightarrow j} ^ {\text {s p a t i a l}} \right], +$$ + +$$ +f _ {i \rightarrow j} ^ {\text {v i s u a l}} = M \left[\left(\tilde {\mathbf {o}} _ {i} - \tilde {\mathbf {o}} _ {j}\right), \left(\tilde {\mathbf {o}} _ {i} + \tilde {\mathbf {o}} _ {j}\right)\right], \tag {1} +$$ + +$$ +f _ {i \rightarrow j} ^ {s p a t i a l} = c a t [ {\bf b} _ {i}, {\bf b} _ {j}, d x, d y, d i s, \theta , A _ {i}, A _ {j}, I, U ], +$$ + +where $M$ are the mapping functions implemented as 2-layer multi-layer perceptrons, $(ct_i^x, ct_i^y)$ and $(ct_j^x, ct_j^y)$ are the normalized center coordinates of two involved objects, $dx = ct_i^x - ct_j^x$ , $dy = ct_i^y - ct_j^y$ , $dis = \sqrt{dx^2 + dy^2}$ , $\theta = \arctan \left(\frac{dy}{dx}\right)$ , $\mathbf{b}_i, \mathbf{b}_j, A_i, A_j, I, U$ denote the normalized bounding boxes, the areas of the subject, the object, their intersection, and union boxes, respectively. Then we get the relation feature through the relation encoder $\mathrm{Enc}_r$ : + +$$ +V _ {r} = \operatorname {E n c} _ {r} \left(f _ {i \rightarrow j}\right). \tag {2} +$$ + +The text embeddings for all categories are extracted by feeding the text of the relation categories into the text encoder [16]. The text embedding $t_c \in \mathbb{R}^d$ for class c is: + +$$ +\boldsymbol {t} _ {c} = \operatorname {E n c} _ {L} \left(\boldsymbol {W} _ {c}\right), \forall c \in \text {r e l a t i o n c a t e g o r i s}, \tag {3} +$$ + +where $W_{c}$ is the class token vector. Therefore, the relation logit can be extracted by the cosine similarities w.r.t $\{t_{c}\}$ : + +$$ +\hat {r} _ {i} = \cos \left(V _ {r}, t _ {c}\right). \tag {4} +$$ + +Pre-trained VLMs can achieve performance on open-vocabulary detection tasks [5, 43]. However, due to the lack of quadratic relation-aware knowledge in VLMs, it is difficult for VLMs to generalize open-vocabulary relations without references. Furthermore, VLMs without relation-aware knowledge may be overconfident in the knowledge of the base training set, thereby failing to achieve recognition of the open-vocabulary relations. To further improve the performance of VLM on OV-SGG, we propose to address the existing challenge of OV-SGG by enhancing the visual-text modality interaction to construct generalizable relation-aware knowledge. On the one hand, to improve the generalization ability of VLM for open-vocabulary relations, we construct a generative relation recognition module to make the text modality explore possible open-ended relations based on visual content. On the other hand, we construct a hierarchical relation extension module to extend the relation text from visual modality, alleviating the overconfidence of VLM on the base training set. + +# 3.2. Generative Relation Recognition + +To enhance the relation-aware knowledge in VLM, we construct a generative relation recognition framework to generate diverse visual relations. + +Relation Generation. To generate the open-vocabulary relations, we build a visual-text interaction generation model for generating relation names. We propose directly linking the relation predictor with a language model, and activating both the image encoder and the language model as trainable components, as shown in Fig. 3. Particularly, we employ an encoder-decoder-based conditional language generation model, where visual representation serves as input for the language encoder, and the associated text (i.e., the text of the relations) serves as the generation target for the language decoder. We use the relation features matched with ground truth through top-k highest confidence scores. We then feed the relation features to a projection layer to align its dimensions with the language model. We construct an encoder comprised of blocks with self-attention layers and feed-forward networks. As for the decoder, we use cross-attention layers to make the text embedding interface with the relation embedding from the encoder. The decoder employs an autoregressive self-attention mechanism, and its final output is processed through a dense layer. Specifically, given the relation feature $V_{r}$ , $V_{r}$ are used as the condition for generating text in the language model. The probability of the next word $w_{i}$ is predicted by the preceding words $w_{i}^{1}, \ldots, w_{i}^{t - 1}$ in text sequence, which can be formulated as: + +$$ +P \left(w _ {i} ^ {t} \mid V _ {r}, w _ {i} ^ {1}, \dots , w _ {i} ^ {t - 1}\right). \tag {5} +$$ + +Context Generation. Previous OV-SGG work directly constructs text embeddings of relation names and maps them to the text modality [5, 43]. However, in the visual + +modality, the relation recognition in SGG often relies on relevant context information such as entity box position and entity features [31, 41] for reasoning. Simply using relation names to construct text modality features of the relations may not be sufficient to represent them. To enhance the adaptation of text embedding to the visual modality, inspired by prompt learning [44], we generate context prompt tokens $C t_{i}$ to fit the model's relation text space into the visual modality. As illustrated in Fig. 3, given a relation category denoted as $\hat{c}_i$ , we use a few prefix tokens $C t_{i}$ at the beginning to automatically determine how to describe the relation. Then, we generate the text embedding for each relation label through the pre-trained text-encoder [43] $\mathrm{Enc}_{L}$ : + +$$ +T = \operatorname {E n c} _ {L} \left(\operatorname {C t} _ {i}, \boldsymbol {W} _ {c}\right), \tag {6} +$$ + +the embeddings of all relation categories are denoted as $T \in \mathbb{R}^{N \times d}$ , where $N$ denotes the number of relation categories, and $d$ represents the dimension of text embedding from the text encoder. Specifically, the context vector is set as a learnable vector while $W_{c}$ is kept as fixed. + +# 3.3. Hierarchical Relation Extension + +Based on the generative relation module, we construct a hierarchical relation extension module to further extend relation-aware knowledge in the base training set to open-vocabulary relations. + +Space-wise Extension. Previous work has shown the label sparsity problem in the SGG task [13, 42], i.e., unannotated relations are carelessly assigned to "background". We argue that this problem is more serious in the label-constrained OV-SGG setting, the VLM would also focus on specific areas in the image due to the sparse relation labels, resulting in overconfidence in the basic dataset at the spatial level. Previous works, such as PGSG [21] and OVSGTR [5], use pre-trained grounding models to generate aligned bounding boxes with relation phrases from captions, and serve as pseudo labels for training the model. However, given that captions may not comprehensively describe every object in an image, the count of relations in the caption could be lower than the actual number of relations present in the corresponding image. Moreover, since there is no relation-aware prior knowledge in the VLM, the generated pseudo-labels may not be appropriate for the target task. Unlike the previous strategies, we propose to use our pretrained relation recognition model to generate pseudo labels based on the visual relation logits to extend the relationship-aware knowledge of VLM at the spatial level. + +Semantic-wise Extension. To mitigate the overfitting of VLM to the base dataset from the semantical level, we propose to construct a semantic-unbiased VLM. Specifically, we quantify the visual model's learned knowledge in various relations. Based on the learning situation, we perform a dynamically fitting strategy in the text modality. Inspired + +by the classic distribution estimation algorithm Logit Adjustment (LA) [28], the relation class-probability can be expressed as $\mathbb{P}(\hat{\mathbf{r}}\mid o_i,o_j)\propto \mathbb{P}(o_i,o_j\mid \hat{\mathbf{r}})\cdot \mathbb{P}(\hat{\mathbf{r}})$ . Then, we have the following conclusion: + +$$ +\begin{array}{l} \mathbb {P} (\hat {\mathbf {r}} \mid o _ {i}, o _ {j}) \propto \mathbb {P} (o _ {i}, o _ {j} \mid \hat {\mathbf {r}}) \cdot \mathbb {P} _ {\text {t r a i n}} (\hat {\mathbf {r}}) / \mathbb {P} _ {\text {t r a i n}} (\hat {\mathbf {r}}) \\ \propto \mathbb {P} _ {t r a i n} (\hat {\mathbf {r}} \mid o _ {i}, o _ {j}) / \mathbb {P} _ {t r a i n} (\hat {\mathbf {r}}) \\ \end{array} +$$ + +where $\mathbb{P}_{train}(\hat{\mathbf{r}}) = \frac{N_C}{\sum N_C}$ denotes the frequency of relational predicate $C$ . However, due to the gap between VLM and visual recognition models, simply estimating the bias through sample distribution may not be accurate. Thus, we employ the Normalized Weighted Geometric Mean (NWGM) approximation to dynamically estimate the predicate class prior, which is widely used as a tool for capturing more stable feature statistics [34]. We estimate the dynamic predicate class prior $\mathbb{P}_{train}(\hat{\mathbf{r}}\mid \theta)$ by the model via the training set, i.e., $\mathbb{P}_{train}(\hat{\mathbf{r}}\mid \theta) = \mathbb{E}_{r\in \mathcal{D}_{train}}\mathbb{P}_{train}(\hat{\mathbf{r}}\mid p;\theta)$ , where $r$ denotes the predicate annotation of the relation triplets. Then, the NWGM approximation for the model $F(x)$ can be formulated as: + +$$ +\mathbb {E} (\operatorname {s i g m o i d} \left(F \left(\boldsymbol {x} _ {i}\right)\right)) \stackrel {N W G M} {\approx} \operatorname {s i g m o i d} \left(\mathbb {E} (F \left(\boldsymbol {x} _ {i}\right))\right), \tag {8} +$$ + +then we can maintain the logit to capture the data bias in the entire training set, and update it after each batch: + +$$ +\hat {r} _ {i} ^ {b i a s} \leftarrow m \hat {r} _ {i} ^ {b i a s} + (1 - m) \frac {1}{B} \sum^ {B} \hat {r} _ {i}, \tag {9} +$$ + +where $B$ denotes the batch size. To perform dynamic fitting, we adjust the relation logits by adding the bias logit with the origin logit and weighting the generation loss in each relation class based on the bias logit: + +$$ +\begin{array}{l} \hat {r} _ {i} ^ {\text {d e b i a s}} = \hat {r} _ {i} + \tau \hat {r} _ {i} ^ {\text {b i a s}}, \tag {10} \\ L _ {g e n} = \left(1 - \tau \hat {r} _ {i c} ^ {b i a s}\right) \cdot L _ {g e n}, \\ \end{array} +$$ + +where $\tau$ is a hyperparameter, $\hat{r}_{ic}^{bias}$ denotes the $c$ -th term in $\hat{r}_i^{bias}$ , $c$ is the relation label for $\hat{r}_i$ . By adding the bias logit with the origin logit, the classifier dynamically demands a larger margin in the highly biased relation categories and a smaller margin in the hardly biased relation categories, thus alleviating the VLM's overconfidence in a specific common base relation. + +# 3.4. Training Objectives + +During the training stage, we follow the previous VLM-based methods [5, 21, 43] to use a multi-task loss consisting of two parts: the loss for associate visual and text features and $\mathcal{L}_{VL}$ and the loss for grounding the objects $\mathcal{L}_{pos}$ . Specifically, based on the relation $\logit\hat{r}_i^{debias}$ , we can calculate a binary cross entropy loss with given ground truths: + +$$ +\begin{array}{l} \mathcal {L} _ {\mathrm {b c e}} = - y \log \sigma \left(\hat {r} _ {i} ^ {\text {d e b i a s}}\right) \tag {11} \\ - (1 - y) \log \left(1 - \sigma \left(\hat {r} _ {i} ^ {\text {d e b i a s}}\right)\right), \\ \end{array} +$$ + +![](images/0edb75b8c4f34569eecc9b9e6e7633ff0ff45ea1bd24bf7526a353014bcb8892.jpg) +Figure 4. Illustration for our semantic-wise extension strategy. The bias in the various relations $\hat{r}_i^{bias}$ is dynamically estimated and momentum-updated by the relation logit of each sample. The balanced logit $\hat{r}_i^{debias}$ avoids the VLM being overwhelmed by common relations. + +where $\sigma$ refers to the sigmoid function, $y$ is a vector where "1" indexes positive tokens. Moreover, as for the generative relation language model, we apply the language modeling loss to train the model. Specifically, we use cross-entropy loss for each word in the generated text, and the language modeling loss is: + +$$ +\mathcal {L} _ {\text {c o n s}} = \sum_ {i = 1} ^ {M} \left(- \sum_ {t = 1} ^ {N} \log P \left(w _ {i} ^ {t} \mid V _ {r}, w _ {i} ^ {1}, \dots , w _ {i} ^ {t - 1}\right)\right), \tag {12} +$$ + +the goal is to minimize the negative likelihood of the correct word predictions [6], which encourages the model to generate accurate and relevant text based on the object features. Thus the sum loss can be formulated as: + +$$ +\mathcal {L} _ {V L} = \mathcal {L} _ {\mathrm {b c e}} + \left(1 - \tau \hat {r} _ {i c} ^ {\text {b i a s}}\right) \cdot \lambda \mathcal {L} _ {\text {c o n s}}, \tag {13} +$$ + +where $\lambda$ is a hyperparameter. As for the regression loss of the bounding box $\mathcal{L}_{pos}$ , for grounding each box prediction B in scene graphs, we utilize GIOU and L1 loss for calculating the distance between B and the ground truth $\mathrm{B}_{gt}$ : + +$$ +\mathcal {L} _ {p o s} = \operatorname {G I O U} (\mathbf {B}, \mathbf {B} _ {g t}) + \| \mathbf {B} - \mathbf {B} _ {g t} \| _ {1}. \tag {14} +$$ + +During inference, we directly generate the relation categories from our model. Specifically, for the relation feature $V_{r}$ of each scene graph, we generate the relation with the generative relation language model conditioned on $V_{r}$ : + +$$ +W = P \left(W \mid V _ {r}\right), W = \left\{w _ {i} ^ {1}, \dots , w _ {i} ^ {t} \right\}. \tag {15} +$$ + +# 4. Experiments + +# 4.1. Datasets and Experimental Settings + +Dataset. We evaluate our approach on the common SGG datasets: Visual Genome (VG) [14] and OpenImage V6 (OIv6) [15]. As for VG dataset, We considered two different settings in the PGSG [21] and OVSGTR [5]. For the 50 relation categories in VG150, we select $50\%$ [21] / 30 + +
MethodBackboneTraining DatasetNovel+baseNovel
R@50/100mR@50/100R@50/100mR@50/100
VS3 [43]GLIPVG11.0 / 12.85.1 / 5.7- / -0.0 / 0.0
SGTR [20]BLIPVG12.6 / 18.23.5 / 5.4- / -0.0 / 0.0
PGSG [21]BLIPVG15.8 / 19.16.2 / 8.3- / -3.7 / 5.2
OursGLIPVG13.7 / 19.77.5 / 10.610.1 / 14.35.2 / 7.5
VS3* [43]GLIPVG15.6 / 17.3- / -0.0 / 0.0- / -
OVSGTR* [5]GroundingDINOVG17.7 / 20.06.4 / 7.50.3 / 0.4- / -
Ours*GLIPVG14.3 / 20.48.7 / 12.79.0 / 12.6- / -
OVSGTR* [5]GroundingDINOVG, COCO, Flickr30k, SBU20.5 / 23.93.9 / 4.613.5 / 16.2- / -
OVSGTR+Ours*GroundingDINOVG, COCO, Flickr30k, SBU21.1 / 25.04.2 / 4.914.2 / 17.7- / -
+ +Table 1. Experimental results of open-vocabulary scene graph generation of the SGDet task on VG. * denotes the method is tested under the OVSGTR's setting with $30\%$ novel relation categories, the others are under the PGSG's setting with $50\%$ novel relation categories. Our method achieves comparable performance to prior models, without requiring access to various instruction prompts or additional pretraining. + +
MethodBackboneTraining DatasetNovel+baseNovel
R@50/100mR@50/100R@50/100mR@50/100
SVRP [8]Faster-RCNNVG33.5 / 35.98.3 / 10.8- / -- / -
CaCao [40]CLIPVG,CC3M,COCO caption- / -10.3 / 12.6- / -- / -
PGSG [21]BLIPVG26.9 / 33.910.8 / 13.9- / -5.2 / 7.7
OursGLIPVG27.8 / 37.511.1 / 14.511.9 / 18.35.9 / 8.2
+ +Table 2. Experimental results of open-vocabulary scene graph generation of the PredCls task on VG. + +
MethodNovel+BaseNovel
R@50/100mR@50/100R@50/100mR@50/100
SGTR [20]36.1 / 38.411.0 / 16.7- / -0.0 / 0.0
PGSG [21]41.3 / 43.320.8 / 23.0- / -3.8 / 8.9
Ours46.3 / 59.419.8 / 24.614.6 / 18.74.5 / 10.1
+ +$\%$ [5] of them as novel relation categories. During training, only base relation annotation is available. For OIV6, we following the setting in the PGSG [21] and select $50\%$ of them as novel relation categories. + +Evaluation protocols and metrics. Since the task evaluation of OV-SGG requires the score of relation triplets based on the relation logits for ranking [5, 43], to assess the effectiveness of the open-end generative results, we propose a new setting: We calculate the similarity score between the text embeddings of generated and annotated category names as the relation logit. In this study, we select the text encoder constructed in the model, i.e., BERT. In this way, we can get the logit of the generative results for evaluation. + +During testing, as in previous works [7, 31], the SGG models generally focus on two common SGG tasks: (1) Predicate Classification (PredCls): models are required to predict the relationships for all object pairs by using both the ground-truth bounding boxes and classes. (2) Scene Graph Detection (SGDet): gives the images, the models are required to localize the object bounding boxes, recognize objects, and predict their relationships. We evaluate the model under the popular metric Recall@K (R@K) and mean Recall@K (mR@K). + +Table 3. Experimental results of open-vocabulary scene graph generation of the SGDet task on OIV6. + +
MethodNovel+BaseNovel
R@50R@100R@50R@100
IMP [38]0.770.940.000.00
VCTREE [31]1.041.170.000.00
TDE [32]1.001.100.000.00
VS3-SwinT [43]5.887.200.000.00
OVSGTR-T [5]13.5316.369.2011.19
OVSGTR-T+Ours14.8718.2712.2715.26
+ +Table 4. Experimental results of open-vocabulary relation and detection SGG setting on VG150 test set. + +Implementation details. We use pre-trained GLIP [19] models to initialize our model and keep the visual backbone and text encoder frozen. Other modules like the interaction classifier are initialized randomly. We retain the top 36 object detections per image for pairwise relation recognition. We adopt FlanT5-base [6] as relation language model and initialize the weights from it. The optimizer is AdamW [27] with a weight decay of 0.05. The learning rate is set to $1 \times 10^{-5}$ for the parameters of the main backbone, and $2 \times 10^{-5}$ for the language model. + +# 4.2. Compared with State-of-the-arts + +Open-Vocabulary SGG. Following previous works [5, 21], the Open-Vocabulary SGG (OV-SGG) setting requires the model not to see novel relation categories during training. Specifically, we assess the model's capacity to predict diverse relations in an open-vocabulary setting on VG and OIV6, presenting the results in Tabs. 1 to 3. Remarkably, we observe that even with only base training data, our model demonstrates commendable performance on OV-SGG. This result suggests that by enhancing the interaction + +
MethodNovel+Base
R@50R@100mR@50mR@100
Frozen10.6615.364.756.60
w/o Context11.3316.675.896.68
w Context12.9418.487.179.60
+ +![](images/37d68c24e07d692d7217cb07fec37e5c3ce3eee308eee5dadc7786b74c7921a3.jpg) +Figure 5. Effect of the learning weight of the generative relation language model. + +of text-image modalities, VLM can directly generate open-vocabulary relations for OV-SGG during inference. Moreover, our method can be combined with existing pre-trained OV-SGG models to achieve better performance. As shown in Tab. 1, OVSGTR + Ours achieves a significant gain on the performance of the VG test set. This demonstrates that our method can also help the existing SGG model to extract more general relation features. + +Open-Vocabulary Relation and Detection SGG. The open-vocabulary relation and detection SGG setting follows previous work [5], and augments the SGG to a fully open-vocabulary domain. In this setting, both novel object and relation categories are omitted during the training phase. We report the performance of the OvD+R-SGG setting in terms of "Base+Novel" and "Novel Relation" in Tab. 4, by combing with OVSGTR, our method outperforms all competitors. Notably, when compared to the recent OVSGTR [5], our method achieves gain of up to $3.07\%$ for zR@50 and $4.07\%$ for zR@100. + +# 4.3. Ablation Study + +Context Generation. Towards OV-SGG, we believe that the generative context prompt plays a role in generating more precise outputs. This is because the learnable context tokens explore a broader format of descriptions during encoding. This contributes to a more adaptable and nuanced understanding of the relation being recognized. As shown in Tab. 5, the learnable context tokens improve the performance for encoding the relation. + +![](images/5b2a6fe07d53ec69b9175ec2694916b6688210f7d79cffdd529ed11e6daf9c6c.jpg) +Figure 6. Ablation study of dynamic fitting module on VG150 test set. + +Table 5. Ablation study of context prompt on VG150 test set. + +
MethodNovel+baseNovel
R@100mR@100R@100mR@100
w/o Debias16.246.6210.234.36
LA [28]12.586.5610.473.21
PGLA [23]13.429.1210.765.63
Dynamic Fitting18.489.6013.456.06
+ +Table 6. Ablation study on different bias estimation strategies. + +
MethodNovel+BaseNovel
R@100mR@100R@100mR@100
w/o Pseudo label18.489.6013.456.06
Only fg19.1210.3813.786.58
Bg & fg19.7110.6014.327.51
+ +Table 7. Ablation study of pseudo labeling on VG150 test set. + +Relation Generation. We conduct ablation experiments to evaluate the effectiveness of the relation language generator in our method. Fig. 5 illustrates the impact of the weight $\lambda$ of the language modeling loss $\mathcal{L}_{\text{cons}}$ . we vary $\lambda$ from 0 to 10. Here, $\lambda = 0$ signifies predicting the relation without the language model as in VS3 [43], with a performance plateau at $\lambda = 2$ . The results show that without the generated results from the language model, a significant decline in performance for novel relations exists, showing the effectiveness of the proposed method in predicting novel categories. + +Space-wise Extension. Due to the limited base relation categories, we found that the count of relation pairs in the annotated base dataset is generally much lower than the actual number of relation pairs in the images. To address the limitation, we incorporated RelGen trained on the base relation categories to supplement additional relation pairs in the images. We use a confidence score greater than 0.3 to filter all possible relation pairs (i.e., $\mathrm{Bg} \& \mathrm{fg}$ ). In Tab. 7, "Only fg" means we only consider the pseudo-labels of the annotated relation pairs. The performance on the VG dataset indicates that the strategy is important for enhancing the comprehensiveness of the pseudo-label generation process. + +Semantic-wise Extension. As shown in Fig. 6, the dy + +![](images/9f70b06da00b2cf689f13b55768acbae769abf9410f18475531e6d847abfa4eb.jpg) +Figure 7. Qualitative results of our model on VG test set. The blue arrows refer to novel relation categories which are not included in the training set. + +![](images/919b05b1c16ace670911f312420366dc3bd7762ae90c09283b6492c5f2a5ad7d.jpg) + +![](images/887ee63f04276e787539dd6cc10d7649540de1e7b20ba0c749952c28584b0aee.jpg) + +![](images/56bfab7d1822fda4db5148495804857dc50057bfa5e4f54ad831a5cc4f86ec2e.jpg) +Figure 8. The generated relation labels from relation prediction on VG test set. A varied set of descriptive relation labels showcases the language model's capability to produce diverse and linguistically rich vocabulary. It is worth noticing that the red relation categories do not exist in the VG dataset. + +![](images/ce924b0d32f56ca5712028fd73a944137a7175ab83eceb49231d98cdc8942fe0.jpg) + +![](images/eb5985823b46da728081333b31f717b627349f6e9f40e70a018bf6c6c807a2a1.jpg) +Figure 9. Comparison of qualitative results on VG test set. + +![](images/5028bf2f611abeb68e9c60bfcd0f813bc2ba820d766f697aa018ba9fdef9f07e.jpg) + +![](images/d910e915cd08d4945b9d1d1c87400e0113d6c06c6f26d3f71a6f1511fac1486d.jpg) + +namic fitting module could alleviate the model's semantic bias towards the common predicates. By dynamically adjusting the learning strategy for various relations, mR@100 and zR@100 metrics significantly increase. We observe the notable improvement on the zR@100 metric by over $2\%$ , showing the significance of the dynamic fitting. When $\tau$ is too large, the model pays too much attention to the few relations while losing the general pattern of the common relations. To better trade off the performance on different relations, we set $\tau$ to 0.7. Furthermore, as shown in Table 6, by comparing our dynamic debiasing strategy with the traditional static debiasing method, we can see that dynamic estimating the bias improves the performance of R@k and mR@k, demonstrating that our method more accurately quantifies and removes semantic bias in VLM. + +# 4.4. Qualitative Results + +Visualization. To verify our method's capability of making reliable relation recognition, we show the scene graphs generated by our method in Fig. 7 and Fig. 9. We can observe that our method can help generate more general open- + +vocabulary relation predicates while remaining faithful to the image content. Specifically, without the full supervision of novel categories, our model can provide novel relationship predictions (e.g., "from" and 'part of'). The results demonstrate the generative relation language model's ability to detect a wider variety of relation categories. + +Detailed Results. In Fig. 8, we show the detailed results of the generated relations with respect to the top-100 relation prediction of the PredCls task on the VG test set (i.e., R@100). These visual results not only affirm the reliability of our framework but also demonstrate its ability to help capture potential reasonable relations. For example, our method captures a novel reasonable relation "putting on" beyond fixed vocabulary constraints on the VG dataset. These results demonstrate that our model has a more general relation recognition ability. + +# 5. Conclusion + +This work proposes a novel vision-language interactive relation mining model for OV-SGG. Specifically, by introducing a generative relation recognition model, our model achieves generating open-vocabulary relation names. In addition, a hierarchical extension module is adopted to further extend the relations. Our method achieves comparable results with previous models on OV-SGG settings. In the future, we will further analyze more effective methods to enhance the visual-text interaction. + +Acknowledgement. Our work is supported in part by the National Key R&D Program of China (No. 2023YFC3305600), Joint Fund of Ministry of Education of China (8091B022149, 8091B02072404), National Natural Science Foundation of China (62132016), and Natural Science Basic Research Program of Shaanxi (2020JC-23). + +# References + +[1] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023. 3 +[2] Jinhe Bi, Yujun Wang, Haokun Chen, Xun Xiao, Artur Hecker, Volker Tresp, and Yunpu Ma. Llava steering: Visual instruction tuning with $500\mathrm{x}$ fewer parameters through modality linear representation-steering. In ACL, 2025. 2 +[3] Jinhe Bi, Yifan Wang, Danqi Yan, Xun Xiao, Artur Hecker, Volker Tresp, and Yunpu Ma. Prism: Self-pruning intrinsic selection method for training-free multimodal data selection, 2025. +[4] Tom B Brown. Language models are few-shot learners. arXiv preprint arXiv:2005.14165, 2020. 2 +[5] Zuyao Chen, Jinlin Wu, Zhen Lei, Zhaoxiang Zhang, and Changwen Chen. Expanding scene graph boundaries: Fully open-vocabulary scene graph generation via visual-concept alignment and retention. arXiv preprint arXiv:2311.10988, 2023. 2, 4, 5, 6, 7 +[6] Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Yunxuan Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, et al. Scaling instructionfinetuned language models. Journal of Machine Learning Research, 25(70):1-53, 2024. 2, 3, 5, 6 +[7] Xingning Dong, Tian Gan, Xuemeng Song, Jianlong Wu, Yuan Cheng, and Liqiang Nie. Stacked hybrid-attention and group collaborative learning for unbiased scene graph generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19427-19436, 2022. 6 +[8] Tao He, Lianli Gao, Jingkuan Song, and Yuan-Fang Li. Towards open-vocabulary scene graph generation with prompt-based finetuning. In European Conference on Computer Vision, pages 56-73. Springer, 2022. 2, 6 +[9] Yufeng Huang, Jiji Tang, Zhuo Chen, Rongsheng Zhang, Xinfeng Zhang, Weijie Chen, Zeng Zhao, Zhou Zhao, Tangjie Lv, Zhipeng Hu, et al. Structure-clip: Towards scene graph knowledge to enhance multi-modal structured representations. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2417-2425, 2024. 3 +[10] Dat Huynh, Jason Kuen, Zhe Lin, Jiumiang Gu, and Ehsan Elhamifar. Open-vocabulary instance segmentation via robust cross-modal pseudo-labeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7020-7031, 2022. 2 +[11] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International conference on machine learning, pages 4904-4916. PMLR, 2021. 2 +[12] Jacob Devlin Ming-Wei Chang Kenton and Lee Kristina Toutanova. Bert: Pre-training of deep bidirectional trans + +formers for language understanding. In Proceedings of NAACL-HLT, pages 4171-4186, 2019. 3 +[13] Kibum Kim, Kanghoon Yoon, Yeonjun In, Jinyoung Moon, Donghyun Kim, and Chanyoung Park. Adaptive self-training framework for fine-grained scene graph generation. In The Twelfth International Conference on Learning Representations, 2023. 4 +[14] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision, 123:32-73, 2017. 2, 5 +[15] Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan Popov, Matteo Malloci, Alexander Kolesnikov, et al. The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale. International journal of computer vision, 128(7):1956-1981, 2020. 5 +[16] Dongxu Li, Junnan Li, Hongdong Li, Juan Carlos Niebles, and Steven CH Hoi. Align and prompt: Video-and-language pre-training with entity prompts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4953-4963, 2022. 3 +[17] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730-19742. PMLR, 2023. 3 +[18] Lin Li, Jun Xiao, Guikun Chen, Jian Shao, Yueting Zhuang, and Long Chen. Zero-shot visual relation detection via composite visual cues from large language models. Advances in Neural Information Processing Systems, 36:50105-50116, 2023. 2 +[19] Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, et al. Grounded language-image pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10965-10975, 2022. 2, 6 +[20] Rongjie Li, Songyang Zhang, and Xuming He. Sgtr: End-to-end scene graph generation with transformer. In proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 19486-19496, 2022. 6 +[21] Rongjie Li, Songyang Zhang, Dahua Lin, Kai Chen, and Xuming He. From pixels to graphs: Open-vocabulary scene graph generation with vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 28076-28086, 2024. 2, 4, 5, 6 +[22] Yue Liao, Aixi Zhang, Miao Lu, Yongliang Wang, Xiaobo Li, and Si Liu. Gen-vlkt: Simplify association and enhance interaction understanding for hoi detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20123-20132, 2022. 2 +[23] Hengyue Liu and Bir Bhanu. Repsgg: novel representations of entities and relationships for scene graph generation. + +IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 7 +[24] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 3 +[25] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Qing Jiang, Chunyuan Li, Jianwei Yang, Hang Su, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 2, 3 +[26] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 3 +[27] I Loshchilov. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6 +[28] Aditya Krishna Menon, Sadeep Jayasumana, Ankit Singh Rawat, Himanshu Jain, Andreas Veit, and Sanjiv Kumar. Long-tail learning via logit adjustment. arXiv preprint arXiv:2007.07314, 2020. 2, 5, 7 +[29] Shan Ning, Longtian Qiu, Yongfei Liu, and Xuming He. Hoiclip: Efficient knowledge transfer for hoi detection with vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23507-23517, 2023. 2 +[30] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2 +[31] Kaihua Tang, Hanwang Zhang, Baoyuan Wu, Wenhan Luo, and Wei Liu. Learning to compose dynamic tree structures for visual contexts. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6619-6628, 2019. 4, 6 +[32] Kaihua Tang, Yulei Niu, Jianqiang Huang, Jiaxin Shi, and Hanwang Zhang. Unbiased scene graph generation from biased training. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3716-3725, 2020. 6 +[33] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 3 +[34] Tan Wang, Jianqiang Huang, Hanwang Zhang, and Qianru Sun. Visual commonsense r-cnn. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10760-10770, 2020. 5 +[35] Chenghao Xu, Guangtao Lyu, Jiexi Yan, Muli Yang, and Cheng Deng. Llm knows body language, too: Translating speech voices into human gestures. In ACL, pages 5004-5013, 2024. 3 +[36] Chenghao Xu, Jiexi Yan, Muli Yang, and Cheng Deng. Rethinking noise sampling in class-imbalanced diffusion models. IEEE TIP, 2024. 3 + +[37] Chenghao Xu, Jiexi Yan, and Cheng Deng. Keep and extent: Unified knowledge embedding for few-shot image generation. IEEE TIP, 2025. 3 +[38] Danfei Xu, Yuke Zhu, Christopher B Choy, and Li Fei-Fei. Scene graph generation by iterative message passing. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5410-5419, 2017. 6 +[39] Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyedhosseini, and Yonghui Wu. Coca: Contrastive captioners are image-text foundation models. arXiv preprint arXiv:2205.01917, 2022. 2 +[40] Qifan Yu, Juncheng Li, Yu Wu, Siliang Tang, Wei Ji, and Yueting Zhuang. Visually-prompted language model for fine-grained scene graph generation in an open world. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 21560-21571, 2023. 6 +[41] Rowan Zellers, Mark Yatskar, Sam Thomson, and Yejin Choi. Neural motifs: Scene graph parsing with global context. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5831-5840, 2018. 4 +[42] Ao Zhang, Yuan Yao, Qianyu Chen, Wei Ji, Zhiyuan Liu, Maosong Sun, and Tat-Seng Chua. Fine-grained scene graph generation with data transfer. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXVII, pages 409-424. Springer, 2022. 4 +[43] Yong Zhang, Yingwei Pan, Ting Yao, Rui Huang, Tao Mei, and Chang-Wen Chen. Learning to generate language-supervised and open-vocabulary scene graph using pretrained visual-semantic space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2915–2924, 2023. 2, 4, 5, 6, 7 +[44] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, 130(9):2337-2348, 2022. 4 +[45] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. In International Conference on Learning Representations. 2 \ No newline at end of file diff --git a/ICCV/2025/Vision-Language Interactive Relation Mining for Open-Vocabulary Scene Graph Generation/images.zip b/ICCV/2025/Vision-Language Interactive Relation Mining for Open-Vocabulary Scene Graph Generation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..b05d2886edfaf2903c18a427a4c6ff1692d96da0 --- /dev/null +++ b/ICCV/2025/Vision-Language Interactive Relation Mining for Open-Vocabulary Scene Graph Generation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:eba3919dd25a442e76465af717c11735f37a1a817a1c14a11e02b10ee222527b +size 598464 diff --git a/ICCV/2025/Vision-Language Interactive Relation Mining for Open-Vocabulary Scene Graph Generation/layout.json b/ICCV/2025/Vision-Language Interactive Relation Mining for Open-Vocabulary Scene Graph Generation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..7fbbd9e0d4badc777a036599c07fccb94e55ea24 --- /dev/null +++ b/ICCV/2025/Vision-Language Interactive Relation Mining for Open-Vocabulary Scene Graph Generation/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9ec624ae556ba748a85bbdf30301e6c2c78a6762218ad13b1848ba76dadb7f46 +size 381221 diff --git a/ICCV/2025/Vision-Language Models Can't See the Obvious/d8b7426a-cb65-4738-af2c-9b996a867e18_content_list.json b/ICCV/2025/Vision-Language Models Can't See the Obvious/d8b7426a-cb65-4738-af2c-9b996a867e18_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..7efaeed911440eaa07a7f88573b7cf2e4c6de526 --- /dev/null +++ b/ICCV/2025/Vision-Language Models Can't See the Obvious/d8b7426a-cb65-4738-af2c-9b996a867e18_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d056746b81b4c5a7b0025daea0e312719c02712b508e9282c9e296950f44eaca +size 90488 diff --git a/ICCV/2025/Vision-Language Models Can't See the Obvious/d8b7426a-cb65-4738-af2c-9b996a867e18_model.json b/ICCV/2025/Vision-Language Models Can't See the Obvious/d8b7426a-cb65-4738-af2c-9b996a867e18_model.json new file mode 100644 index 0000000000000000000000000000000000000000..45ade99399d5a0890f6b71971295ecc26d36e128 --- /dev/null +++ b/ICCV/2025/Vision-Language Models Can't See the Obvious/d8b7426a-cb65-4738-af2c-9b996a867e18_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e7ba382bf90fba3182408bccf9b71b360e64a11e04aa761f81152ca62e533c38 +size 113285 diff --git a/ICCV/2025/Vision-Language Models Can't See the Obvious/d8b7426a-cb65-4738-af2c-9b996a867e18_origin.pdf b/ICCV/2025/Vision-Language Models Can't See the Obvious/d8b7426a-cb65-4738-af2c-9b996a867e18_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..225c9e597f6ad9e84abd6668c6ca70f2e8ec1b31 --- /dev/null +++ b/ICCV/2025/Vision-Language Models Can't See the Obvious/d8b7426a-cb65-4738-af2c-9b996a867e18_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6941698b3aa2b0292e9d80fb7be84cc6df9a32681e92e6a934601239d2f7c259 +size 1043177 diff --git a/ICCV/2025/Vision-Language Models Can't See the Obvious/full.md b/ICCV/2025/Vision-Language Models Can't See the Obvious/full.md new file mode 100644 index 0000000000000000000000000000000000000000..4d493894b5646849cacec63cde8d79a73529ee10 --- /dev/null +++ b/ICCV/2025/Vision-Language Models Can't See the Obvious/full.md @@ -0,0 +1,321 @@ +# Vision-Language Models Can't See the Obvious + +Yasser Dahou* + +Ngoc Dung Huynh*† + +Phuc H. Le-Khac + +Wamiq Reyaz Para + +Ankit Singh + +Sanath Narayan + +Technology Innovation Institute, Abu Dhabi, UAE + +† Deakin University, Australia https://salbench.github.io + +# Abstract + +We present Saliency Benchmark (SalBench), a novel benchmark designed to assess the capability of Large Vision-Language Models (LVLM) in detecting visually salient features that are readily apparent to humans, such as a large circle amidst a grid of smaller ones. This benchmark focuses on low-level features including color, intensity, and orientation, which are fundamental to human visual processing. Our SalBench consists of images that highlight rare, unusual, or unexpected elements within scenes, and naturally draw human attention. It comprises three novel tasks for evaluating the perceptual capabilities of LVLM: Odd-One-Out Detection, Referring Odd-One-Out, and Visual Referring Odd-One-Out. We perform a comprehensive evaluation of state-of-the-art LVLM using SalBench and our findings reveal a surprising limitation: LVLM struggle to identify seemingly obvious visual anomalies, with even the advanced GPT-4o achieving only $47.6\%$ accuracy on such a simple task. SalBench will be an important step in measuring the capabilities of LVLM that align with the subtle definition of human attention. + +# 1. Introduction + +Large Vision-Language Models (LVLM) have emerged as a central focus in recent computer vision research [4, 12, 31, 32, 44, 60]. The primary advantage of these models lies in their ability to reason about images using the Large Language Model (LLM) knowledge about the world, and solve tasks that surpass the capabilities of traditional vision models. LVLMs enable a better reasoning about the visual content using the prior world knowledge. + +Current LVLM exhibit impressive performance on existing benchmarks [11, 17, 27, 34, 55, 57], which evaluate + +# Detection + +Identify the feature(s) in which one object notably differs from the others. + +# Referring + +In which way does the special object located at differ from the rest. + +![](images/7c45c97e9ac31929da96de435af0049c1fcc8a0f372a1e68500ed3c7af93e388.jpg) +Color, pattern, shape + +![](images/c3a3e9e3572ea7afe58bdea8420c95230de3fdded9f6851e24c7f16f9f050135.jpg) +Color, shape + +![](images/0e53786b9c164165b7f7fad02ce3c5fd23cb1b3c6475e2b583ecc66a09330e5c.jpg) +Orientation + +![](images/55847ba90f1fbbe828ce70a4b6d7a805493dff0b617074357a9e43d8cf10e35b.jpg) +Size + +For natural images, choices: Orientation, Color, Focus, Shape, Size, Location, Pattern. For synthetic images, choices: Orientation, Color, Size. + +# Visual Referring + +In which way does the special object highlighted by red box differ from the rest. + +![](images/36f6c17cbb49655370ab76a2a192c23286fa941a86f96a16f93a05495534a7a6.jpg) +Orientation, location + +![](images/1a6f173eba60fe2334107b8d004b2cc30762157955e86100e1dd14a51606b188.jpg) +Color +Figure 1. Our SalBench evaluates the perceptual capabilities of vision-language models. It comprises natural and synthetic images, which consist of a single salient target among multiple distractors. In natural images, the target can vary from the rest in orientation, color, focus, shape, size, location and pattern, while it differs only in orientation, color or size in synthetic images. SalBench evaluates the odd-one-out understanding under three tasks: Detection, Referring and Visual Referring. In detection task, the odd features must be directly predicted. Moreover, the box coordinates of the target are provided as context in text (referring task) or as a highlighted box in the image (visual referring task) for predicting the odd features of the target. Best viewed zoomed-in. + +on a range of capabilities, from general visual question answering [17] to tasks requiring college-level subject knowledge and critical reasoning, such as MMMU [57]. With the main focus of these benchmarks being high-level complex tasks, it raises a fundamental question: can LVLM perform equally well on simple perceptual tasks, such as identifying a black dot on a white background? This aligns with Moravec's paradox [40], which suggests that high-level reasoning tasks are computationally simpler for artificial intelligence systems than low-level perceptual and sensorimotor skills. Consequently, we may find that LVLM excel at complex tasks present in existing benchmarks while strug- + +gling with seemingly simple perceptual tasks that humans perform effortlessly. To this end, we propose a benchmark for quantifying the alignment of vision-language models on low-level perceptual tasks of the human attention. + +While LVLM effectively capture high-level features such as cars and humans, they are likely to struggle to represent crucial aspects of human visual attention that have been extensively studied in neuroscience. Visual search, which is a fundamental process shaping human attention [24, 49] involves the brain's parallel processing of regions that differ significantly in one feature dimension, such as color, intensity, or orientation. These low-level features serve as basic mechanisms of the human visual system. By examining the LVLM performance on simple saliency-driven images, we aim to gain insights into the current state of these models relative to human perceptual abilities and identify areas for necessary improvements. Our key contributions are: + +- We propose SalBench as an open-source benchmark, for evaluating and aiding the improvement of the perceptual capabilities of LVLM. To this end, we augment the P3/O3 datasets [24], which comprise of images with a single distinctive target among many similar distractors, with language instructions and create three novel tasks: Odd-One-Out Detection, Referring Odd-One-Out, and Visual Referring Odd-One-Out, as shown in Figure 1. Our benchmark aims to serve as a tool for assessing the progress in aligning LVLM with human visual attention. +- We conduct a comprehensive analysis of LVLM performance on SalBench, uncovering striking discrepancies between these advanced models and human visual capabilities. Our findings show that even state-of-the-art LVLM, including GPT-4o [42], struggle with basic saliency detection tasks that are trivial for humans. +- We provide insights into the limitations of LVLM in processing low-level visual features, highlighting the need for improved alignment between these models and human visual attention mechanisms. Our work demonstrates the importance of incorporating neuroscience principles into the development of future vision-language models. + +# 2. Related Works + +Many vision-language benchmarks have been introduced in the literature for evaluating and comparing LVLM on various tasks. The early benchmarks were mostly single-task oriented and can be broadly classified into captioning, general visual-question answering (VQA), and text-centric VQA. The captioning task is used to measure the LVLM's caption generation quality and is mostly evaluated on COCO [30] and NoCaps [3] benchmarks using BLEU, CIDER, ROUGE metric scores. Similarly, general VQA benchmarks like VQAv2 [17], GQA [21], ScienceQA [33], OK-VQA [44], VizWiz [18], Pope [28] typically ask general questions about the objects/scene + +in the image. Furthermore, text-centric VQA benchmarks such as OCRVQA [39], TextVQA [45], ST-VQA [8], DocVQA [36], ChartQA [35], InfoVQA [37], and AI2D [22] additionally focus on the vision-language model's ability to detect text in the image (Optical Character Recognition) and then answer the questions. + +With the increasing capabilities of vision-language models in handling different types of tasks, the benchmarks have also evolved to evaluate the LVLM in a fine-grained manner. To this end, more recently, MMBench [32], MME [16], MM-Star [11], and MMMU [56] are curated to test the LVLM on a mix of different aspects, such as reasoning, perception, knowledge, chart interpretation, etc. Similarly, MathVista [34], MathVision [50] assess the mathematical reasoning ability in visual contexts. Differently, vision-centric benchmarks like MMVP [48], RealWorldQA [53], CV-Bench [47] curate questions that can be answered correctly only in the presence of the corresponding visual input. These vision-centric benchmarks ensure that the LVLM relies on its multimodal understanding capability for responding rather than the 'world knowledge' learned by the LLM. In particular, MMVP employs "CLIP-blind pair" images that CLIP [43] struggles to encode properly. While RealWorldQA evaluates the basic real-world spatial understanding capabilities of vision-language models, CV-Bench repurposes standard vision benchmarks like COCO [30], ADE20K [59], Omni3D [10] to assess their performance in classic vision tasks (such as spatial relationship, object count, depth order) within a multimodal context. + +In contrast to the aforementioned benchmarks that test the vision-language models on high-level reasoning tasks, our SalBench strives to evaluate them on their low-level saliency perception capability. SalBench is a vision-centric benchmark, which repurposes the P3/O3 dataset [24] with language instructions to assess the LVLM ability to detect odd patterns in images that are visually salient and obvious to the human eye. Such an evaluation quantifies the vision-language model's ability to identify the distinctive target among numerous similar distractors, in terms of color, orinetation, shape, size, etc. Consequently, our SalBench is designed with an aim to measure the vision-language models' alignment with the human visual attention mechanism. Next, we describe our SalBench in detail. + +# 3. The Saliency Benchmark: SalBench + +We propose a saliency benchmark that assesses the capacity of foundational vision-language models to detect odd patterns that are salient for most humans. To achieve this, we augment the publicly available P3/O3 datasets [24] with language instructions. The P3 dataset comprises 2514 images (810 color, 864 orientation, and 840 size search arrays) arranged in a $7 \times 7$ grid [6, 52]. Each image contains distractors and a target, with pixel jitter applied to pre + +![](images/1efe8579f50b29f576cf5bc3daa12417f7ef42cdde89598a64ca261d603accfb.jpg) +Figure 2. Example responses from state-of-the-art vision-language models. The model responses are in green, enclosed in boxes, while failure reasons are in red. These results highlight that the models often fail to recognize the prominent and salient features, which would naturally capture the attention of an average human observer. Best viewed zoomed-in. + +![](images/0388f741717df41091ff53721ac9ac14df5db4addf48b55f6369ae54a299e146.jpg) + +![](images/705e28dcba6c2ccf589b6ec1127c66ba4af4fc50ddda87f33855c4c78f13da68.jpg) + +vent perceptual grouping. On the other hand, the O3 dataset consists of 2001 real-world images featuring multiple similar objects (distractors) and a distinctive singleton (target). The target typically belongs to the same general category as the distractors but stands out in one or more feature dimensions (e.g., color, shape, size). O3 encompasses nearly 400 common object types, strongly emphasizing color singletons (37% differ by color alone, 47% by color and additional features). Other prominent distinguishing features include texture (33%), shape (26%), size (19%), and orientation (8%). The dataset provides a range of complexity, with distractor counts varying from 2 to over 50, offering a diverse challenge for saliency detection algorithms [24]. + +# 3.1. Task Creation + +As discussed previously, in order to enhance the utility of the P3/O3 datasets as a benchmark for vision-language models, we propose to augment them with language instructions. Prompting the LVLM directly to predict the odd-pattern often results in incorrect responses from state-of-the-art LVLM. Few example responses for such a direct prompting are illustrated in Figure 2. The example conversations show that additional prompting or aid is required for the models to improve their responses. Consequently, to ensure better understanding of the model's perceptual capabilities, we create three novel tasks: Odd-One-Out Detection, Referring Odd-One-Out, and Visual Referring Odd-One-Out, as illustrated in Figure 1. We describe each task as follows: + +(i) Odd-One-Out Detection: The basic task in this bench + +mark is Odd-One-Out Detection. In this task, given an input image, the model must identify which salient pattern from a predefined list of classes is present. For natural images, the list of features comprises: orientation, color, focus, shape, size, location, and pattern. This set of seven classes reflects the complexity of real-world scenes, where multiple factors such as size, shape, and location can combine to make an object stand out. The diversity of these classes captures saliency in natural images, making the task more challenging for models due to the potential interplay of various salient features. On the other hand, for synthetically generated images, the task is constrained to three patterns: color, shape, and orientation. This limitation is intentional to maintain control over the variables influencing saliency. By restricting the odd-pattern to a unique basic class rather than a combination of features, we can ensure that the model's focus is on a specific, isolated pattern. This simplification facilitates a clear assessment of the model's ability to detect saliency based on individual features. + +(ii) Referring Odd-One-Out: This task builds on Odd-One-Out Detection by incorporating spatial information in the language prompt. The model receives an image and bounding box coordinates indicating the odd object's location via text. We test the model's ability to integrate textual and visual information, requiring it to focus on a specific image region using the language instruction and determine how the object there differs from the rest. +(iii) Visual Referring Odd-One-Out: This is a variant of the Referring Odd-One-Out task. Instead of using bounding box coordinates in text, this task visually highlights the odd object, typically with a red box. The model must identify the distinguishing features of this highlighted object compared to the other objects in the scene. This task evaluates the model's visual attention capabilities, requiring it to focus on a visually emphasized region-of-interest and determine the salient pattern in which the highlighted object stands out from its surroundings. + +# 4. Evaluation + +This section presents the evaluation results of the various models on SalBench. The analysis of model performances across zero-shot and few-shot learning settings reveals key insights into their capabilities and limitations. + +Baselines: In our experiments, we evaluate a selection of state-of-the-art models spanning a range of parameter scales to benchmark the performance across model size variations. The lineup includes PaliGemma [7], Phi3, and Phi3.5 [1]; LLava 1.6 [31]; Idefics2 and Idefics3 [25]; VILA [29]; MiniCPM [20]; Qwen2-VL [51]; Molmo [14]; InternVL2 [46]; Llama3.2-Vision [38]; NVLM [13]; Claude [5]; and GPT-4o [2]. These models are considered the best LVLM according to the existing vision-language benchmarks. By encompassing a wide range of parameter + +counts, from smaller and more efficient configurations to large-scale models, we can systematically analyze the impact of model size on task performance. + +Setting: The salient targets can be different in one or more patterns, we devise the task as multi-label classification and report the accuracy of exact matches and the average F1 score over all categories. We evaluate model performance under both zero-shot and few-shot settings. For models that support multiple image inputs, we specifically conduct few-shot evaluation with 3-shots and 5-shots settings. While the zero-shot evaluation assesses the model's ability to generalize without exposure to specific examples, we expect that such a task isn't common and is likely to pose issues to most models. Hence, we also employ the few-shot evaluation to examine how the models leverage examples from the benchmark itself to solve the task. The default prompts for zero-shot evaluation of the three tasks are as described below: + +(i) Detection: Context: Given this list of low-level visual features defined according to feature integration theory: Orientation, Color, Focus, Shape, Size, Location, Pattern. Task: Examine the provided image and identify the feature(s) in which one object notably differs from the others. Write out all applicable features separated by comma. + +(ii) Referring: Context: This image depicts a scene with {num distractor} {object category}. Among those, one object category at location given by this bounding box $(x_{min},y_{min},x_{max},y_{max})$ , is different from the others. Given this list of low-level visual features defined according to feature integration theory: Orientation, Color, Focus, Shape, Size, Location, Pattern. Task: In which way does the special object differ from the rest. Write out all applicable features separated by comma. + +(iii) Visual Referring: Context: This image depicts a scene with $\{\mathrm{num~distractor}\}$ {object category}. Among those, one $\{\mathrm{object~category}\}$ highlighted in a red box is different from the others. Task: Given this list of low-level visual features defined according to feature integration theory: Orientation, Color, Focus, Shape, Size, Location, Pattern. In which way does the special object differ from the rest. Write out all applicable feature(s) separated by comma. + +Additional details of the few-shot prompts are provided in the supplementary material. + +# 4.1. Results + +In Table 1, we analyze the performance of various large vision-language models (LVLM) on our saliency benchmark, focusing on both synthetic and natural images splits (coming from P3 and O3, respectively). The synthetic and natural splits are denoted by SYN and NAT, respectively in the table. The tasks evaluated are detection, referring, and + +
MethodShotDetectionReferringVisual Referring
NATSYNNATSYNNATSYN
Claude-sonnet048.286.751.190.353.987.7
NVLM-D-72B [13]041.577.542.073.737.351.7
Molmo-7B[14]032.067.232.438.033.128.4
Molmo-72B[14]040.683.341.265.436.773.6
Llama3.2-Vision-11B [15]032.148.729.152.429.752.4
PaliGemma-3B-448[7]027.642.01.29.52.34.8
Phi3-4B[1]032.141.232.855.332.847.2
334.133.532.027.132.138.5
531.117.032.118.932.246.7
Phi3.5-Vision-3.5B[1]023.235.027.553.727.563.5
323.319.528.841.028.820.8
525.229.330.811.130.819.0
LLava 1.6-7B[31]024.516.321.410.120.816.6
37.016.415.28.817.817.0
511.416.410.99.19.717.0
Idefics2-8B[26]019.564.329.636.633.849.5
321.166.328.434.231.139.6
534.767.228.342.630.934.5
Idefics3-8B[25]024.328.424.352.822.119.2
326.940.326.920.6721.940.6
522.321.422.318.120.958.3
VILA-1.5-8B[29]023.540.013.023.715.817.0
325.117.028.821.228.817.0
523.217.030.820.730.817.0
Qwen2-VL-1.5B[51]019.226.322.120.620.920.2
325.223.321.421.820.216.3
525.323.821.716.520.917.7
Qwen2-VL-7B[51]032.555.732.534.235.257.4
335.653.836.017.034.164.2
537.254.937.217.729.372.0
Qwen2-VL-72B[51]041.688.844.693.641.774.7
343.989.343.693.143.285.9
543.989.944.992.642.387.9
InternVL-4B[12]026.641.529.863.430.752.2
327.717.027.425.329.541.7
533.417.028.139.130.452.5
InternVL-2-8B[12]020.058.723.071.924.823.0
330.552.324.251.731.764.4
527.843.925.053.731.450.5
GPT-4o047.689.247.388.442.673.5
338.988.437.587.735.786.7
541.986.039.889.138.487.4
+ +Table 1. Performance comparison of vision-language models on the three tasks of SalBench for natural and synthetic image splits. Natural and synthetic splits are denoted by NAT and SYN, respectively. The tasks are also evaluated under zero-shot and few-shot settings. The performance is reported in terms of F1 scores. We observe that performance of all models on these low-level perceptual tasks are lower in comparison to the standard vision-language benchmarks that test the models on high-level complex tasks. This shows that our SalBench offers another dimension for comprehensively evaluating LVLM, that is not present in the existing benchmarks in the literature. + +visual referring, measured using matching accuracy and F1 score metrics. Only F1 score metric is reported in Table 1 for clarity. The performance comparison with matching accuracy metric is provided in the supplementary material. + +Significant Performance Drops: All models exhibit a clear drop in performance on saliency tasks compared to their typical results on standard vision-language bench + +marks. This is evident in both matching accuracy and F1 scores across all tasks. For example, Qwen2-VL-72B achieves high scores on the synthetic split up to $89.9\%$ F1 in the Detection task at the 5-shot setting. However, on the natural images split, its performance decreases significantly, achieving only $43.9\%$ F1 in the Detection task at 5-shot. This indicates that models struggle with saliency tasks, especially in complex, real-world scenarios. + +Better Performance on Synthetic Data: Models generally perform better on the synthetic images split than on the natural split. For instance, GPT-4o obtains an F1 score of $70.9\%$ in the Detection task at 5-shot on synthetic split. In contrast, on the natural images split, GPT-4o achieves only $47.6\%$ F1 in the same task and shot setting. This significant performance gap, often between $30\%$ and $40\%$ , suggests that models find it easier to process simplified synthetic images, where only one visual attribute varies, compared to the complexity of multi-label real-world natural images. + +Model Size Influence: Larger models tend to outperform smaller ones on saliency tasks. In particular, Qwen2-VL-72B consistently achieves higher scores than its smaller variants, Qwen2-VL-7B and Qwen2-VL-1.5B. For the detection task on the synthetic split, Qwen2-VL-72B reaches $89.9\%$ F1 score in the detection task at 5-shot, while Qwen2-VL-7B achieves $54.9\%$ and Qwen2-VL-1.5B only $23.8\%$ . On the natural split, Qwen2-VL-72B attains $43.9\%$ F1 score in detection at 5-shot, compared to $37.2\%$ for Qwen2-VL-7B and $25.3\%$ for Qwen2-VL-1.5B. This trend indicates that increased model capacity allows for better capture of complex patterns necessary for saliency tasks, a behavior not always observed in standard benchmarks. Similarly, GPT-4o shows strong performance, particularly in the Visual Referring tasks, indicating that larger models contribute to better saliency detection and reasoning capabilities over images. This trend suggests that increased model capacity allows for the capture of more complex patterns and relationships necessary for these tasks. This behavior is usually not seen in standard benchmarks such as VQAv2, as we see a marginal performance gap between large and small models. + +Limited Impact of Few-Shot Learning: Increasing the number of shots does not consistently improve performance across models and tasks. Some models show comparable or even decreased performance with more shots. For example, Phi3.5-Vsion-3.5 on the synthetic split in the Visual Referring task has a F1 score of $63.5\%$ at 0-shot, drops to $19.0\%$ at 5-shot. On the natural images split, GPT-4o performance in the Detection task decreases from $47.6\%$ F1 score at 0-shot to $41.9\%$ at 5-shot. This suggests that few-shot learning does not uniformly enhance performance on saliency tasks and that models may not effectively leverage additional examples in this context. The inability to generalize from limited examples suggests a lack in knowledge + +about saliency and visual attention for LVLM being worse when attempting to reason about the image. + +Variability Across Tasks: The performance across tasks and splits for different models varies according to their visual grounding capability. For instance, Idefics3-8B achieves $21.4\%$ F1 accuracy in Detection at 5-shot on the synthetic split but increases to $58.3\%$ in Visual Referring. However, on the natural images split, the same model attains $22.3\%$ in Detection at 5-shot but decreases to $20.9\%$ in Visual Referring. This shows that it is likely easier for the model to perform visual grounding in synthetic images, compared to natural images. Differently, Phi3-4B shows improvements for Visual Referring, in comparison to Detection task on both splits, indicating better visual grounding capabilities. Furthermore, models with lower visual grounding capability often have higher performance for Detection task, compared to Visual Referring. + +Challenges with Real-World Images: The performance drop from synthetic to real-world images is significant across models. Qwen2-VL-72B's score in the Detection task decreases from $89.9\%$ on synthetic images split to $43.9\%$ on natural split at 5-shot. GPT-4o shows a similar decline, from $86.0\%$ on synthetic split to $41.9\%$ on natural split. This suggests that models have difficulty handling the complexity and variability of real-world images, where multiple attributes and contextual factors are involved. + +Inconsistent Few-Shot Performance Gains: Few-shot learning yields inconsistent and unclear performance gains. While some models show slight improvements, others do not benefit or even perform worse with additional shots. For example, GPT-4o on the natural images split achieves $47.6\%$ F1 at 0-shot in Detection, drops to $38.9\%$ at 3-shot, and then increases to $41.9\%$ at 5-shot. Qwen2-VL-72B remains relatively stable across shot settings on the synthetic split, with F1 scores around $89\%$ in Detection, while for the Visual Referring task, the performance improves by $13.2\%$ . This inconsistency indicates that current models might not effectively utilize few-shot example contexts in saliency tasks. + +# 4.2. Knowledge Testing of the Backbones + +To identify where the shortcoming of vision-language models comes from in the saliency setting, we devise a test for individual components: LMMs and vision backbones. + +Could LLMs solve FIT exams? We evaluated several Large Language Models (LLMs) on their understanding of Feature Integration Theory (FIT) [49] using a test comprising 50 questions collected from the internet. The larger models demonstrated good knowledge in this area. GPT-40 achieved the highest score with $97.5\%$ , closely followed by Qwen2-72B-Instruct at $95.0\%$ . Among the smaller models, Llama3-8B-Instruct attained a score of $82.5\%$ . The Vicuna models achieved a score of $67.5\%$ . These results indicate + +
ModelsSYNNAT
Top1Top3Top1Top3
Random24.653.214.636.3
Siglip-so400m-patch14-38455.387.90.057.2
CLIP-ViT-Base-Patch3242.171.712.059.8
CLIP-ViT-Base-Patch1641.678.40.940.8
CLIP-ViT-Large-Patch1441.278.60.031.6
CLIP-ViT-Large-Patch14-33647.579.20.036.3
+ +Table 2. Zero-shot retrieval accuracy for natural and synthetic images using various vision backbones. While the retrieval scores are reasonable for the synthetic images, the lower scores for natural images indicate the need for a stronger vision backbone in LVLM to capture low-level saliency information. + +that even smaller LLMs possess an understanding of the visual stimuli concepts covered in FIT. + +Visual representations retrieval: We investigate whether the feature output from the vision backbone are descriptive and discriminative enough about the saliency task. We compute the cosine similarity between the image embedding and the seven sentences for natural images and three sentences for the synthetic images, with embeddings of the form: "One object is different in {category} compared to the others". We take the top- $k$ highest cosine similarity and count it as a match if any category in the top- $k$ is present in the ground truth. The accuracy is reported in Table 2. Unlike knowledge test from LLM, we see clear gap in performance of vision model. While all models scores above random baseline, they scores relatively low, suggesting that in some cases, the information provided by the vision encoder are not discriminative enough for such low-level task. This highlights the need for stronger vision backbones for LVLM, capturing low-level spatial saliency information. + +# 5. Training on saliency data + +Data Generation: To construct the synthetic dataset, we employed web icons arranged in grid formations and systematically introduced controlled perturbations by altering the color, orientation, or size of randomly selected objects within each grid. This methodology yielded one million image-caption pairs for the alignment stage. For the supervised fine-tuning stage, we synthesized additional images with corresponding questions and answers based on the introduced odd-patterns. The questions were designed to be either open-ended, enhancing the models' saliency understanding capabilities, or multiple-choice, aligning with the benchmark format. + +Training: We selected LLama3.1 8B [15] and Qwen2-7B [54] as the LLMs. For the vision encoders, we incorporated both CLIP [43] and SigLip [58], resulting in four model variants. We followed a Llava like training recipe [31]. During the alignment stage, we use one million + +![](images/e5da6686ac12f6289ee7fc1f6ea4c11af8c661e6461343732e1d17da2b75dc12.jpg) +Figure 3. Performance comparison of trained LVLM on saliency generated data, using two LLMs (Qwen2.5-7B and LLama3.1-8B) paired with two vision encoders (Siglip and Clip) across three tasks: Detection, Referring, and Visual Referring. The F1 scores suggest that training on such data does not help improve the performance on SalBench. + +saliency data with two million image-caption pairs sourced from natural image datasets to promote robust multimodal representation learning. In the instruction-tuning stage, we integrated two million data points from the recently introduced Cambrian dataset [47] with the generated one million instruction saliency data. We employed the same hyperparameters and training configurations as specified in Llava [31] to ensure consistency and facilitate fair comparison of results. + +Performance: As shown in Table 3, despite training on in-distribution saliency data, the performance of all model variants across the P3 tasks remains low. Detection scores are around $16 - 19\%$ , with LLama3.1-8B paired with CLIP achieving the highest at $18.9\%$ . For the referring task, scores are even lower and more variable, ranging from $11.0\%$ to $18.5\%$ . The visual reference task shows slightly better results, particularly for LLama3.1-8B with SigLip, which attains a score of $28.7\%$ , yet this is still insufficient for practical applications. These consistently low scores indicate that the models struggle to capture the salient features within the images, even when trained on data specifically designed to highlight these aspects. The lack of improvement suggests that the current architectures and training recipe are not suitable for capturing the of saliency information in images. + +# 6. Analysis + +In this section, we study the factors contributing to the poor performance of LVLM on the visual attention benchmark. + +# 6.1. Accuracy Breakdown by Task and Difficulty + +Bias towards color: As seen in Table 3, among the low-level features what were asked to identify, color is the only + +
ModelCategoryLevelAccuracy
DetectionReferringVisual Referring
Qwen2-VL-72BOrientationEasy98.697.188.3
Medium95.896.883.0
Hard95.796.271.7
GPT-4oOrientationEasy96.296.597.6
Medium96.289.294.7
Hard98.688.596.2
Qwen2-VL-72BSizeEasy94.298.862.3
Medium50.574.331.9
Hard46.062.033.3
GPT-4oSizeEasy93.396.942.5
Medium59.159.139.1
Hard36.835.311.3
Qwen2-VL-72BColorEasy100.0100.099.1
Medium100.0100.098.6
Hard60.180.273.5
GPT-4oColorEasy99.899.096.1
Medium100.0100.093.3
Hard66.166.257.9
+ +Table 3. Accuracy across different difficulty levels on synthetic images. It can be seen that orientation and color are robust over the three levels, whereas size scores drop for both models. + +
# DistractorsOrientationColorFocusShapeSizeLocationPatternAvg F1
Qwen2-VL-72B
<718.492.818.653.439.730.357.944.5
7-159.393.713.055.837.831.641.240.4
15-2523.595.811.159.139.50.037.538.1
>2516.796.115.052.236.520.025.437.4
GPT-4o
<737.188.238.149.435.723.657.047.0
7-1546.290.140.047.532.627.553.348.2
15-2543.891.545.247.531.012.545.145.2
>2540.589.848.448.130.44.829.441.6
+ +Table 4. Accuracy for varying the number of distractors for real images. The performance linearly decreases when adding more distractors to the scene. + +features directly provided by the data in the form of RGB images. Other features such as size, and shape need to be encoded in higher level representation. We break down the scores in supplementary by category, and observe that all models perform higher when the target class is color. + +On the synthetic images (originating from P3 dataset) of the SalBench, we assess two best models (Qwen2-VL-72B and GPT-4o) across the categories at different levels. The difficulty levels within each category are defined based on specific attributes of the target object relative to distractors: + +- Orientation: Difficulty is determined by the angular difference in rotation between the target object and the distractors. Hard: Rotation differences ranging from 0 to 30 degrees, where minimal rotational disparity renders visual cues less distinguishable. Medium: Rotation differences between 30 to 60 degrees. Easy: Rotation differ + +ences from 60 to 90 degrees, facilitating easier identification due to clear angular differences. + +- Size: Levels are determined based on the area ratio between the target object and the distractors, calculated using their heights and widths. Hard: Ratio between 0.5 and 1.5, where the target and distractors are close in size. Medium: Ratio between 0.3 and 0.5 (target smaller than distractors) or between 1.5 and 3.0 (target larger than distractors), representing moderate size differences. Easy: Ratio less than 0.3 or greater than 3.0, indicating that the target is much smaller or much larger than the distractors, making it easily distinguishable. +- **Color:** Difficulty is assessed by calculating the euclidean distance between the RGB color values of the target and distractors. The RGB values are extracted from hexadecimal color codes, to compute the color distance. **Hard:** Color Distance less than or equal to 50, indicating high color similarity and challenging discrimination. **Medium:** Color Distance between 50 and 100, indicating moderate color differences. **Easy:** Color Distance greater than 100, where the target's color clearly contrasts with that of the distractors, making identification easier. + +Results and Analysis: The performance metrics for both models are detailed in Table 3. The following observations highlight the models' capabilities and limitations across difficulty levels: + +Orientation: Qwen2-VL-72B and GPT-4o demonstrate high accuracy in the Detection and Referring tasks across all difficulty levels, indicating robust performance in recognizing and describing objects with rotational variations. Specifically, Qwen2-VL-72B maintains Detection accuracy above $95\%$ even at the hard level. However, in the Visual Referring task, Qwen2-VL-72B exhibits a decline from $88.3\%$ accuracy at the easy level to $71.7\%$ at the hard level. In contrast, GPT-4o sustains consistently high Visual Referring accuracy, exceeding $94\%$ across all difficulty levels. + +Size: The models' performance deteriorates as the difficulty increases. For Qwen2-VL-72B, Detection accuracy drops from $94.2\%$ at the easy level to $46.0\%$ at the hard level. Similarly, Visual Referring accuracy drops from $62.3\%$ to $33.3\%$ . GPT-4o follows a comparable trend, with Detection accuracy decreasing from $93.3\%$ to $36.8\%$ and Visual Referring accuracy from $42.5\%$ to $11.3\%$ . This suggests both models struggle when size disparities are minimal, highlighting a limitation in perceiving subtle scale differences. + +Color: Both models achieve near-perfect accuracies at the easy and medium levels. Qwen2-VL-72B attains $100\%$ accuracy in Detection and Referring tasks, while GPT-4o records accuracies exceeding $99\%$ . However, at the hard level—characterized by minimal color differences—there is a notable decline. Qwen2-VL-72B's Detection accuracy decreases to $60.1\%$ , and GPT-4o's to $66.1\%$ . + +![](images/70acdd963dfafcdd1038653608eaabd543b62de7c531c7b470dbab2195c4643c.jpg) +Gpt-4o $\rightarrow$ Describe the scene and mention if you notice anything unusual or suspicious. + +![](images/56af023d465789003c6cc052a28980c053fdc1be896b2957ce9c744563252708.jpg) + +![](images/62a4504669a235a4255242cc7b9af0991b9fba1df5226302d929edaf423857ff.jpg) +- The lack of reaction from the two men suggests either unawareness or intentional disregard. + +Figure 4. Perception change test. GPT-4o fails to identify that the individuals talking to the old man were switched. +![](images/204c10a0f89b9fc2e41b286fc47a2a1501e0b5e0ddb3ef878f34f88583c7d672.jpg) +The scene shows an outdoor interaction between a formally dressed man with white hair and another man holding papers. Two individuals carrying a large board pass between them, briefly obstructing their view. Despite the interruption, the conversation continues unaffected. Observations: +This setup might be part of an experiment, possibly testing attention or perception during interruptions. + +![](images/1619d4ca136975764ec41b123fd92d32a6fec57cae3a5e5d6a19c9ea8fcc0dc4.jpg) +- The white-haired man seems to be giving directions. + +![](images/087b2366135e5e24f56a046cd7f9b50e25c92c03b9edfac2b09409b157b2b1c1.jpg) +- The board carriers' interruption could be deliberate, possibly as a distraction. + +# 6.2. Impact of Distractor Quantity on Detection + +We study how the number of distractors in an image affects the performance of the model, on the natural images Detection task. The aim is to understand the models' robustness to increased visual complexity. As shown in Table 4, both models exhibit a decline in average F1 scores as the number of distractors increases. For Qwen2-VL-72B, the average F1 score decreases from $44.5\%$ when there are fewer than 7 distractors to $37.4\%$ when there are more than 25 distractors. GPT-4o shows a similar trend, with its average F1 score dropping from $47.0\%$ to $41.6\%$ over the same range. Although GPT-4o consistently outperforms Qwen2-VL-72B across most distractor quantities, the performance gap narrows as the number of distractors increases. This convergence indicates that both models are similarly challenged by increased visual complexity. + +In fact, both models maintain high accuracy in the Color category regardless of distractor quantity. However, performance declines are evident in categories requiring spatial reasoning and attention to specific object attributes. In the Location category, Qwen2-VL-72B's F1 score drops dramatically from $30.3\%$ with fewer than 7 distractors to $20.0\%$ with more than 25 distractors. Interestingly, in the Shape category, both models maintain relatively stable performance across different distractor quantities. Overall, these findings highlight that while certain visual features like color and, to some extent, shape remain reliable even, increased numbers of distractors hurt the performance. + +# 6.3. Why detecting saliency is important? + +While SalBench might not contribute at measuring LVLM explicit ability for downstream applications [11, 17, 27, 34, 55, 57]. Detecting visual saliency is a fundamental aspect of human perception, for vision-language models replicating + +this capability is critical for several reasons: + +Applications in Robotics and Autonomous Agents: In robotics, the ability to detect salient features is crucial for navigation, object recognition, and interaction within dynamic environments. Recently introduced works leverage LVLM for such applications [9, 23]. As shown in Figure 4, the individual holding the paper was switched, and GPT-4o did not detect this change, we better have this ability for robots with LVLM as engines not to miss salient events. Incorporating the saliency detection capabilities could enable robots and agents to prioritize important stimuli, focus on objects of interest, and respond appropriately to unexpected changes, thereby enhancing their autonomy and operational efficiency. Hence, using SalBench could give insights about the robustness of LVLM for real world deployment. + +LVLM agents: Detecting saliency is important in using these models as agents for website navigation. Designers use principles from FIT to make important elements, such as call-to-action buttons, stand out by contrasting their features (e.g., color, size, or shape). When LVLM are deployed as agents to interact with web environments [19, 41], effective saliency detection becomes needed for efficient navigation and interaction. For instance, an agent tasked with automating web browsing or performing tasks like form filling, data extraction, or content moderation must quickly identify and focus on salient elements such as buttons, links, advertisements, or notifications. By accurately detecting visually prominent features, the agent can operate faster and more effectively, closely mimicking human browsing behavior and enhancing overall performance. + +# 7. Conclusion + +We introduced a saliency benchmark, called SalBench, aimed at evaluating the abilities of large vision-language models (LVLM) in detecting low-level visually-salient features such as color, orientation, and size, which are the building blocks of the human's visual cortex. Through three tasks, Odd-One-Out Detection, Referring Odd-One-Out, and Visual Referring Odd-One-Out, we assessed how well LVLM align with human visual attention mechanisms using both synthetic and natural images. Our evaluation revealed clear limitations in current LVLM, with even advanced models like GPT-4o achieving only $46\%$ accuracy on simple saliency detection tasks. This highlights a clear gap between LVLM and human perceptual capabilities in processing fundamental visual features. SalBench serves as a tool for benchmarking and improving the perceptual alignment of LVLM with human attention mechanisms. Enhancing these models' low-level perceptual abilities is essential for advancing toward more human-like visual understanding of images. + +# References + +[1] Marah Abdin, Sam Ade Jacobs, Ammar Ahmad Awan, Jyoti Aneja, Ahmed Awadallah, Hany Awadalla, Nguyen Bach, Amit Bahree, Arash Bakhtiari, Harkirat Behl, et al. Phi-3 technical report: A highly capable language model locally on your phone. arXiv preprint arXiv:2404.14219, 2024. 3, 4 +[2] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 3 +[3] Harsh Agrawal, Karan Desai, Yufei Wang, Xinlei Chen, Rishabh Jain, Mark Johnson, Dhruv Batra, Devi Parikh, Stefan Lee, and Peter Anderson. Nocaps: Novel object captioning at scale. In Proceedings of the IEEE/CVF international conference on computer vision, pages 8948-8957, 2019. 2 +[4] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in neural information processing systems, 35:23716-23736, 2022. 1 +[5] Anthropic. Claude 3.5 sonnet. https://www.anthropic.com/news/claude-3-5-sonnet,2024.3 +[6] SP Arun. Turning visual search time on its head. Vision research, 74:86-92, 2012. 2 +[7] Lucas Beyer, Andreas Steiner, André Susano Pinto, Alexander Kolesnikov, Xiao Wang, Daniel Salz, Maxim Neumann, Ibrahim Alabdulmohsin, Michael Tschannen, Emanuele Bugliarello, et al. Paligemma: A versatile 3b vlm for transfer. arXiv preprint arXiv:2407.07726, 2024. 3, 4 +[8] Ali Furkan Biten, Ruben Tito, Andres Mafla, Lluis Gomez, Marçal Rusinol, Ernest Valveny, CV Jawahar, and Dimosthenis Karatzas. Scene text visual question answering. In Proceedings of the IEEE/CVF international conference on computer vision, 2019. 2 +[9] Kevin Black, Noah Brown, Danny Driess, Adnan E-mail, Michael Equi, Chelsea Finn, Niccolo Fusai, Lachy Groom, Karol Hausman, Brian Ichter, et al. $\pi_{-}0$ : A vision-language-action flow model for general robot control. arXiv preprint arXiv:2410.24164, 2024. 8 +[10] Garrick Brazil, Abhinav Kumar, Julian Straub, Nikhila Ravi, Justin Johnson, and Georgia Gkioxari. Omni3d: A large benchmark and model for 3d object detection in the wild. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13154-13164, 2023. 2 + +[11] Lin Chen, Jinsong Li, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Zehui Chen, Haodong Duan, Jiaqi Wang, Yu Qiao, Dahua Lin, et al. Are we on the right way for evaluating large vision-language models? arXiv preprint arXiv:2403.20330, 2024. 1, 2, 8 +[12] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Intervl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24185-24198, 2024. 1, 4 +[13] Wenliang Dai, Nayeon Lee, Boxin Wang, Zhuoling Yang, Zihan Liu, Jon Barker, Tuomas Rintamaki, Mohammad Shoeybi, Bryan Catanzaro, and Wei Ping. Nvlm: Open frontier-class multimodal llms. arXiv preprint arXiv:2409.11402, 2024. 3, 4 +[14] Matt Deitke, Christopher Clark, Sangho Lee, Rohun Tripathi, Yue Yang, Jae Sung Park, Mohammadreza Salehi, Niklas Muennighoff, Kyle Lo, Luca Soldaini, et al. Molmo and pixmo: Open weights and open data for state-of-the-art multimodal models. arXiv preprint arXiv:2409.17146, 2024. 3, 4 +[15] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 4, 6 +[16] Chaoyou Fu, Peixian Chen, Yunhang Shen, Yulei Qin, Mengdan Zhang, Xu Lin, Jinrui Yang, Xiawu Zheng, Ke Li, Xing Sun, Yunsheng Wu, and Rongrong Ji. Mme: A comprehensive evaluation benchmark for multimodal large language models. arXiv preprint arXiv:2306.13394, 2023. 2 +[17] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6904-6913, 2017. 1, 2, 8 +[18] Danna Gurari, Qing Li, Abigale J Stangl, Anhong Guo, Chi Lin, Kristen Grauman, Jiebo Luo, and Jeffrey P Bigham. Vizwiz grand challenge: Answering visual questions from blind people. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3608-3617, 2018. 2 +[19] Wenyi Hong, Weihan Wang, Qingsong Lv, Jiazheng Xu, Wenmeng Yu, Junhui Ji, Yan Wang, Zihan Wang, Yuxiao Dong, Ming Ding, et al. Cogagent: A visual language model for gui agents. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14281-14290, 2024. 8 + +[20] Shengding Hu, Yuge Tu, Xu Han, Chaoqun He, Ganqu Cui, Xiang Long, Zhi Zheng, Yewei Fang, Yuxiang Huang, Weilin Zhao, et al. Minicpm: Unveiling the potential of small language models with scalable training strategies. 2024. URL https://doi.org/10.48550/arXiv, 2404.3 +[21] Drew A Hudson and Christopher D Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6700-6709, 2019. 2 +[22] Aniruddha Kembhavi, Mike Salvato, Eric Kolve, Minjoon Seo, Hannaneh Hajishirzi, and Ali Farhadi. A diagram is worth a dozen images. In ECCV, 2016. 2 +[23] Moo Jin Kim, Karl Pertsch, Siddharth Karamcheti, Ted Xiao, Ashwin Balakrishna, Suraj Nair, Rafael Rafailov, Ethan Foster, Grace Lam, Pannag Sanketi, et al. Openvla: An open-source vision-language-action model. arXiv preprint arXiv:2406.09246, 2024.8 +[24] Iuliia Kotseruba, Calden Wloka, Amir Rasouli, and John K Tsotsos. Do saliency models detect odd-one-out targets? new datasets and evaluations. arXiv preprint arXiv:2005.06583, 2020. 2, 3 +[25] Hugo Laurençon, Andrés Marafioti, Victor Sanh, and Léo Tronchon. Building and better understanding vision-language models: insights and future directions. arXiv preprint arXiv:2408.12637, 2024. 3, 4 +[26] Hugo Laurençon, Léo Tronchon, Matthieu Cord, and Victor Sanh. What matters when building vision-language models? arXiv preprint arXiv:2405.02246, 2024. 4 +[27] Bohao Li, Yuying Ge, Yixiao Ge, Guangzhi Wang, Rui Wang, Ruimao Zhang, and Ying Shan. Seedbench: Benchmarking multimodal large language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 1, 8 +[28] Yifan Li, Yifan Du, Kun Zhou, Jinpeng Wang, Wayne Xin Zhao, and Ji-Rong Wen. Evaluating object hallucination in large vision-language models. In The 2023 Conference on Empirical Methods in Natural Language Processing, 2023. 2 +[29] Ji Lin, Hongxu Yin, Wei Ping, Pavlo Molchanov, Mohammad Shoeybi, and Song Han. Vila: On pretraining for visual language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26689-26699, 2024. 3, 4 +[30] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV + +2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 2 +[31] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 1, 3, 4, 6 +[32] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, et al. Mmbench: Is your multi-modal model an all-around player? arXiv preprint arXiv:2307.06281, 2023. 1, 2 +[33] Pan Lu, Swaroop Mishra, Tony Xia, Liang Qiu, Kai-Wei Chang, Song-Chun Zhu, Oyvind Tafjord, Peter Clark, and Ashwin Kalyan. Learn to explain: Multimodal reasoning via thought chains for science question answering. In The 36th Conference on Neural Information Processing Systems (NeurIPS), 2022. 2 +[34] Pan Lu, Hritik Bansal, Tony Xia, Jiacheng Liu, Chunyuan Li, Hannaneh Hajishirzi, Hao Cheng, Kai-Wei Chang, Michel Galley, and Jianfeng Gao. Mathvista: Evaluating math reasoning in visual contexts with gpt-4v, bard, and other large multimodal models. arXiv e-prints, 2023. 1, 2, 8 +[35] Ahmed Masry, Do Xuan Long, Jia Qing Tan, Shafiq Joty, and Enamul Hoque. Chartqa: A benchmark for question answering about charts with visual and logical reasoning. arXiv preprint arXiv:2203.10244, 2022. 2 +[36] Minesh Mathew, Dimosthenis Karatzas, and CV Jawahar. Docvqa: A dataset for vqa on document images. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, 2021. 2 +[37] Minesh Mathew, Viraj Bagal, Ruben Tito, Dimosthenis Karatzas, Ernest Valveny, and CV Jawahar. Infographicvqa. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, 2022. 2 +[38] Meta. Llama 3.2: Revolutionizing edge ai and vision with open, customizable modelsy. https://ai.meta.com/blog/llama-3-2-connect-2024-vision-edge-mobile-devices/, 2024.3 +[39] Anand Mishra, Shashank Shekhar, Ajeet Kumar Singh, and Anirban Chakraborty. Ocr-vqa: Visual question answering by reading text in images. In IC-DAR, 2019. 2 +[40] Hans Moravec. Mind children: The future of robot and human intelligence. Harvard University Press, 1988. 1 +[41] Runliang Niu, Jindong Li, Shiqi Wang, Yali Fu, Xiyu Hu, Xueyuan Leng, He Kong, Yi Chang, and Qi Wang. + +Screenagent: A vision language model-driven computer control agent. arXiv preprint arXiv:2402.07945, 2024.8 +[42] OpenAI. Gpt-4o. https://platform.openai.com/docs/models#gpt-4o, 2024.2 +[43] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 6 +[44] Dustin Schwenk, Apoorv Khandelwal, Christopher Clark, Kenneth Marino, and Roozbeh Mottaghi. Aokvqa: A benchmark for visual question answering using world knowledge. In European conference on computer vision, pages 146-162. Springer, 2022. 1, 2 +[45] Amanpreet Singh, Vivek Natarjan, Meet Shah, Yu Jiang, Xinlei Chen, Devi Parikh, and Marcus Rohrbach. Towards vqa models that can read. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8317-8326, 2019. 2 +[46] InternVL team. Internvl2: Better than the best — expanding performance boundaries of open-source multimodal models with the progressive scaling strategy. https://internvl.github.io/blog/2024-07-02-InternVL-2.0, 2024.3 +[47] Shengbang Tong, Ellis Brown, Penghao Wu, Sanghyun Woo, Manoj Middepogu, Sai Charitha Akula, Jihan Yang, Shusheng Yang, Adithya Iyer, Xichen Pan, et al. Cambrian-1: A fully open, vision-centric exploration of multimodal llms. arXiv preprint arXiv:2406.16860, 2024. 2, 6 +[48] Shengbang Tong, Zhuang Liu, Yuexiang Zhai, Yi Ma, Yann LeCun, and Saining Xie. Eyes wide shut? exploring the visual shortcomings of multimodal llms. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9568-9578, 2024. 2 +[49] Anne M Treisman and Garry Gelade. A feature-integration theory of attention. Cognitive psychology, 12(1):97-136, 1980. 2, 5 +[50] Ke Wang, Junting Pan, Weikang Shi, Zimu Lu, Mingjie Zhan, and Hongsheng Li. Measuring multimodal mathematical reasoning with math-vision dataset, 2024. 2 +[51] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 3, 4 + +[52] Jeremy M Wolfe and Todd S Horowitz. Five factors that guide attention in visual search. Nature human behaviour, 1(3):0058, 2017. 2 +[53] xAI. Grok-1.5v. https://x.ai/blog/grok-1.5v, 2024.2 +[54] An Yang, Baosong Yang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Zhou, Chengpeng Li, Chengyuan Li, Dayiheng Liu, Fei Huang, et al. Qwen2 technical report. arXiv preprint arXiv:2407.10671, 2024. 6 +[55] Weihao Yu, Zhengyuan Yang, Linjie Li, Jianfeng Wang, Kevin Lin, Zicheng Liu, Xinchao Wang, and Lijuan Wang. Mm-vet: Evaluating large multimodal models for integrated capabilities. arXiv preprint arXiv:2308.02490, 2023.1, 8 +[56] Xiang Yue, Yuansheng Ni, Kai Zhang, Tianyu Zheng, Ruoqi Liu, Ge Zhang, Samuel Stevens, Dongfu Jiang, Weiming Ren, Yuxuan Sun, Cong Wei, Botao Yu, Ruibin Yuan, Renliang Sun, Ming Yin, Boyuan Zheng, Zhenzhu Yang, Yibo Liu, Wenhao Huang, Huan Sun, Yu Su, and Wenhu Chen. Mmmu: A massive multi-discipline multimodal understanding and reasoning benchmark for expert agi. In Proceedings of CVPR, 2024. 2 +[57] Xiang Yue, Yuansheng Ni, Kai Zhang, Tianyu Zheng, Ruoqi Liu, Ge Zhang, Samuel Stevens, Dongfu Jiang, Weiming Ren, Yuxuan Sun, et al. Mmmu: A massive multi-discipline multimodal understanding and reasoning benchmark for expert agi. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9556-9567, 2024. 1, 8 +[58] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11975-11986, 2023. 6 +[59] Bolei Zhou, Hang Zhao, Xavier Puig, Tete Xiao, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Semantic understanding of scenes through the ade20k dataset. International Journal of Computer Vision, 127:302-321, 2019. 2 +[60] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 1 \ No newline at end of file diff --git a/ICCV/2025/Vision-Language Models Can't See the Obvious/images.zip b/ICCV/2025/Vision-Language Models Can't See the Obvious/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..c713639f83ed91aa80127e9a01a56d8adc405ef8 --- /dev/null +++ b/ICCV/2025/Vision-Language Models Can't See the Obvious/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9b1419ae3ab133b5dbe07835cb7c88300e67770a0255db54f1895b43fd405b98 +size 396500 diff --git a/ICCV/2025/Vision-Language Models Can't See the Obvious/layout.json b/ICCV/2025/Vision-Language Models Can't See the Obvious/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..ea1b70d69f645f122ef84021b91055b43afc5f01 --- /dev/null +++ b/ICCV/2025/Vision-Language Models Can't See the Obvious/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9afb070fcf05553887d4d5b76d54f6c5b98559b8aa41b10d7c1fde37970c9842 +size 408485 diff --git a/ICCV/2025/Vision-Language Neural Graph Featurization for Extracting Retinal Lesions/7f2c87cd-d7a3-42c9-9013-568610396af7_content_list.json b/ICCV/2025/Vision-Language Neural Graph Featurization for Extracting Retinal Lesions/7f2c87cd-d7a3-42c9-9013-568610396af7_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..0431e27433b9456879a5ae30896542467a1f976f --- /dev/null +++ b/ICCV/2025/Vision-Language Neural Graph Featurization for Extracting Retinal Lesions/7f2c87cd-d7a3-42c9-9013-568610396af7_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0b822ce8bda287048eee538275aceffcd4b027ac5664103b5ffc810ce5aa9184 +size 84507 diff --git a/ICCV/2025/Vision-Language Neural Graph Featurization for Extracting Retinal Lesions/7f2c87cd-d7a3-42c9-9013-568610396af7_model.json b/ICCV/2025/Vision-Language Neural Graph Featurization for Extracting Retinal Lesions/7f2c87cd-d7a3-42c9-9013-568610396af7_model.json new file mode 100644 index 0000000000000000000000000000000000000000..7bee999a85d451c5dde3c1615b13b9667146c0f2 --- /dev/null +++ b/ICCV/2025/Vision-Language Neural Graph Featurization for Extracting Retinal Lesions/7f2c87cd-d7a3-42c9-9013-568610396af7_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2a400b8e0c566767dad42c1d24f0e3f966be2f84d9114ef77a81d097944232d5 +size 102170 diff --git a/ICCV/2025/Vision-Language Neural Graph Featurization for Extracting Retinal Lesions/7f2c87cd-d7a3-42c9-9013-568610396af7_origin.pdf b/ICCV/2025/Vision-Language Neural Graph Featurization for Extracting Retinal Lesions/7f2c87cd-d7a3-42c9-9013-568610396af7_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..d670732f288b362567b67b4cc51baab31490c97c --- /dev/null +++ b/ICCV/2025/Vision-Language Neural Graph Featurization for Extracting Retinal Lesions/7f2c87cd-d7a3-42c9-9013-568610396af7_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:77314c7f3fa186ec2b89c5d40d4ead71bc7ac5e96a47db9265b02bd5a768dd52 +size 22759290 diff --git a/ICCV/2025/Vision-Language Neural Graph Featurization for Extracting Retinal Lesions/full.md b/ICCV/2025/Vision-Language Neural Graph Featurization for Extracting Retinal Lesions/full.md new file mode 100644 index 0000000000000000000000000000000000000000..8d341fef25147eccc2884c04e0ef6d4fa0bb0161 --- /dev/null +++ b/ICCV/2025/Vision-Language Neural Graph Featurization for Extracting Retinal Lesions/full.md @@ -0,0 +1,393 @@ +# Vision-Language Neural Graph Featurization for Extracting Retinal Lesions + +Taimur Hassan1, Anabia Sohail2, Muzammal Naseer2, Naoufel Werghi2 + +1Abu Dhabi University, UAE 2Khalifa University, UAE + +# Abstract + +Retinopathy comprises a group of retinal disorders that can lead to severe visual impairments or blindness. The heterogeneous morphology of lesions poses a significant challenge in developing robust diagnostic systems. Supervised approaches rely on large labeled datasets and often struggle with generalization. To address these limitations, we propose an unsupervised vision-language neural graph featurization method. This method first segments fundus images into a set of superpixels via Simple Linear Iterative Clustering (SLIC). The superpixels are then decomposed into an undirected graph where each superpixel serves as a node, and spatially adjacent nodes are connected by edges. A Hamiltonian path systematically traverses the graph and iteratively updates and propagates node and edge latent space embeddings throughout the graph until convergence is achieved. Then, a normalized cut separates the converged embeddings into two clusters within a latent space that represent the lesion and healthy superpixels of the input scans. The lesion superpixels are further classified into lesion categories using a prompt-based zero-shot vision-language model. The proposed method is rigorously tested on four public datasets, dubbed ODIR, FIVES, BIOMISA, and IDRiD, achieving F1-scores of 0.89, 0.92, 0.93, and 0.92, respectively, with significant performance gains over state-of-the-art methods. + +# 1. Introduction + +Retinopathy represents a group of retinal diseases that can cause severe visual impairments and even blindness if left untreated. The common types of retinal diseases are diabetic retinopathy, diabetic macular edema, and age-related macular degeneration [30, 38, 39]. These retinal diseases represent escalating global health concerns and can be diagnosed by analyzing the presence of retinal lesions. Fundus imaging examination [9] is pivotal for detecting retinal lesions, yet their precise segmentation and recognition remain a challenging task due to significant pathological variations contained within the fundus scans. While supervised + +![](images/ce8bdedef40a98bd66c8f4751ca26cf695dc3286161fa20c6836233a3f8a185a.jpg) +(A) + +![](images/8c7f46ff70c0762f11cd4d20b8032654063a3fa6701240a8ad5dbfae01381f23.jpg) +(B) + +![](images/20892a45a67f6f4c9ce71ba0b0391aaf6ea324c96037c04069398f404d105ba5.jpg) +(C) +Figure 1. Comparison of the proposed framework with DMSF [10], the state-of-the-art method. (A) Fundus image, (B) Ground truth, (C) Lesion classification using the proposed framework, and (D) Lesion classification using the state-of-the-art method. Notice particularly the extracted lesions within the pink bounding boxes in (C) compared to the ground truth (B). The proposed framework is able to extract the smallest lesion pathologies accurately, contrary to the state-of-the-art method [10] (D). + +![](images/8212f8e1f6723fa723b0bed4e9a0b2d0383232595cb6b81d50b1602bda49538d.jpg) +(D) + +learning [18] has been extensively explored toward extracting retinal lesions from the fundus scans, these methods suffer three significant limitations that hinder their deployment in real-world clinical settings to extract retinal lesions effectively. The first limitation is related to their requirement of being trained on a large-scale and well-annotated dataset in order to extract retinal lesions with decent performance. The second limitation is their requirement to be fine-tuned explicitly on the multi-vendor scans to effectively extract retinal lesions from the scans acquired from different machinery at the inference stage. The third limitation is related to their inability to accurately recognize the formation of small lesions, especially microaneurysms (see Figure 1). Due to these limitations, the state-of-the-art (SOTA) models proposed for screening retinal diseases are not scalable to be deployed in the clinical setting [5]. + +To overcome these limitations, we propose an unsupervised neural graph featurization method that extracts retinal lesions from the fundus scans irrespective of the scan specifications, the dataset variations, or the vendor artifacts. Moreover, the proposed method eliminates the need for extensive training, fine-tuning, or ground truth labels to segment retinal lesions from fundus scans. Additionally, it inherently demonstrates robustness to scanner variations by employing Simple Linear Iterative Clustering (SLIC) [1] to generate small superpixels from the scans. These superpixels + +are then represented as nodes in an undirected graph, where spatially adjacent nodes are connected by edges, enabling effective lesion segmentation. + +A Hamiltonian path systematically traverses this graph and iteratively updates and propagates node and edge latent space embeddings throughout the graph until convergence is achieved. Then, a normalized cut separates the converged embeddings into two clusters within a latent space that represent the lesion and healthy superpixel of the input scans. The lesion superpixels are further classified into their appropriate lesion category using a simple prompt-based zero-shot vision language model. + +# 2. Related Work + +Deep learning models have been extensively utilized for segmenting retinal lesions from the fundus images [18, 36]. In the subsequent sections, we discuss various deep learning, foundation, and graph-based approaches that are proposed in the literature to extract retinal lesions from digital retinal images. + +# 2.1. Deep Learning Approaches + +The deep learning models proposed for retinal lesions segmentation are mostly fully supervised [24]. For example, Huang et al. [15] proposed RTNet [15], which incorporates a global transformer block to capture small lesion patterns while employing attention mechanisms in a hierarchical manner. The main limitation of the fully supervised models is their need to be trained using extensive ground truth labels across large-scale datasets to produce decent performance at the inference stage [23, 27]. Recently, these limitations were, to some extent, addressed through the self-supervised and unsupervised learning approaches [20, 33], but the lesion segmentation performance for these approaches was not competitive with the supervised learning approaches [16]. + +# 2.2. Foundation Models for Retinopathy + +Recent advancements in medical imaging have led to a surge in the development of foundation models aimed at enhancing generalization across various tasks and modalities. Toward this end, Zhao et al. [37] introduced Biomed-Parse, a comprehensive biomedical foundation model capable of performing segmentation, detection, and recognition across nine imaging modalities. In the context of retinopathy, Silva-Rodriguez et al. [29] developed a foundation language image model of the retina (FLAIR), designed for performing universal retinal fundus image understanding tasks. + +# 2.3. Graph-based Approaches + +Existing methods for fundus image analysis often overlook the explicit similarity between retinal lesion categories + +across different scans, which leads to their inaccurate recognition. However, graph-based approaches, due to their inherent capacity to analyze homogeneity and heterogeneity between connected nodes, can significantly overcome these limitations [5, 6, 32]. Toward this end, many researchers have proposed graph-based approaches to perform classification and segmentation of retinal lesions and pathologies robustly. For example, Liu et al. [22] proposed a novel label-aware dual Graph Neural Network (GNN) for multi-label fundus image classification. Similarly, Lei et al. [19] proposed a hybrid Graph Neural Network-fused Capsule Network for diabetic retinopathy grading. Salam et al. [26] adapted a k-GNN architecture combined with SIFT descriptors for feature extraction. Xu et al. [35] combined CNNs, GNNs, and Vision Transformers [8] for retinal vessel segmentation. + +# 3. Contributions + +Although extensive research has been conducted on extracting retinal lesions from the fundus scans. But the majority of these methods are based on supervised learning, which requires large-scale and well-annotated datasets for training. Furthermore, the supervised learning methods require extensive re-training and fine-tuning rounds to adapt to different datasets and scanner specifications, which makes them an infeasible and impractical option to be deployed in real-world clinical settings. Recent methods based on GNNs and vision-language models (VLMs) have shown potential toward overcoming the limitations of traditional supervised learning methods. Notably, neural graph approaches have revealed promising results across various domains; however, their application in medical imaging, particularly for analyzing retinal pathologies, remains limited. For instance, prior research has leveraged k-NN graphs as message-passing networks for nuclei classification in whole-slide images [12]. Another study [13] employed neural graph-based techniques to refine nuclei classification in histology images, relying on initial detections as a prerequisite. In contrast to these approaches, the proposed method is unsupervised and does not require any exhaustive training or fine-tuning rounds, as well as ground truth labels to extract retinal lesions from the fundus scans. Apart from this, the lesion segmentation capability of the proposed method is intrinsically resistant to the scanner specifications, the dataset variations, and the vendor artifacts due to its ability to decompose the input fundus scan into small superpixels via SLIC [1], where these superpixels are converted into an undirected graph in which each superpixel serve as a node, and spatially adjacent nodes are connected by edges. Afterward, a Hamiltonian path systematically traverses the graph and iteratively updates and propagates node and edge latent space embeddings throughout the graph until convergence is achieved. Then, a normalized cut separates the converged + +embeddings into two clusters within a latent space that represent the lesion and healthy superpixels within the input scans. The lesion superpixels are further classified into their appropriate lesion category using a simple prompt-based zero-shot vision language model. It should also be noted here that the main contribution of this work is the unsupervised neural graph featurization (NGF) scheme that robustly segments the retinal lesions (especially the small and complex ones that are not clearly visible even to the naked eye). The other components, like SLIC [1], Hamiltonian Paths, prompts, and a vision-language model, are auxiliary and were added just to build the complete system. To summarize, the main contributions of this work are as follows: + +1. We propose a novel unsupervised framework that decomposes an input fundus scan into small superpixels, where each superpixel is clustered into a lesion and a non-lesion (healthy) group through the proposed NGF scheme. +2. The proposed NGF scheme enables even a simple graph-cut method to differentiate between lesion and non-lesion regions across various datasets effectively. +3. Despite its unsupervised nature, our approach demonstrates computational efficiency and achieves significant performance gains compared to state-of-the-art supervised learning approaches across multiple datasets. + +# 4. Proposed Method + +The proposed approach consists of four phases: (a) Lesion Segmentation, (b) Adjacency Network Construction, (c) Neural Graph Featurization, and (d) Lesion Classification. Figure 2 presents a detailed workflow diagram of the proposed framework. + +# 4.1. Lesion Segmentation + +In the initial stage, we use the Simple Linear Iterative Clustering (SLIC) [1] to decompose the input fundus scans into small superpixels. SLIC [1] works by grouping scan pixels into homogeneous superpixels, while preserving boundaries and reducing computational complexity. Superpixels are formed based on color, spatial, and intensity similarities, enhancing lesion detection [1]. The decomposition of the input fundus scan into small superpixels serves as a starting point for the neural graph featurization scheme to identify lesion and non-lesion regions within the input scan accurately. + +# 4.2. Adjacency Network Construction + +To effectively represent the spatial relationships and the structural connectivity between the superpixels, we construct an adjacency network, denoted as $\mathbb{A}$ . This process begins with the generation of a fully connected graph, where each node corresponds to a distinct superpixel within the input scan. To impose a structured traversal order, we derive + +![](images/3c50a90ff42a728f73f011289353075f61572a049bb13745248116e25a694588.jpg) + +![](images/5246900480ef276422aa0c655acc18a662cd62252c408281b162330c220a763e.jpg) + +![](images/d0e7f59bf1714423faf4ec962259fc4877308ab71fd8480932656cc3fba22e8f.jpg) + +![](images/5e6949e3d275f686cd68c745d50130079c44bdb7187e3a7e6719e4bce6dfbe31.jpg) +Figure 2. An overview of the proposed vision-language neural graph featurization approach. First of all, the small superpixels from the input fundus scan are computed using SLIC [1]. Afterward, these superpixels are decomposed into an undirected graph where each superpixel serves as a node, and spatially adjacent nodes are connected by undirected edges. A Hamiltonian path systematically traverses the graph and iteratively updates and propagates node and edge latent space embeddings throughout the graph until convergence is achieved. Then, a normalized cut separates the converged embeddings into two clusters within a latent space that represent the lesion and healthy superpixel of the input scans. The lesion superpixels are further classified into their appropriate lesion category using a simple prompt-based zero-shot vision language model. Zoom in for best visualization. + +a traversal path that extends across the entire spatial dimension of the superpixel decomposed scan. This path serves as the foundation for our adjacency network, ensuring that the spatial connectivity among superpixels is systematically encoded. The constructed adjacency network is depicted in Figure 3, and a detailed explanation of its formulation steps is provided in the following sub-sections. + +# 4.2.1. Complete Graph Construction + +We construct a complete graph $\mathbb{G}(V,E)$ , where $V$ represents the set of nodes, each corresponding to a superpixel within the input scan, and $E$ denotes the set of edges connecting these nodes (superpixels). In this fully connected structure, every node is linked to all other nodes. Each node $v$ is initialized with a deep feature vector, $h_v^0$ , which is extracted from the corresponding superpixel patch $(s_p^v)$ extracted out of the input scan and resized to $224 \times 224 \times 3$ (see Figure 3-D). + +# 4.2.2. Path Generation + +Following the construction of the complete graph $\mathbb{G}$ , we derive a Hamiltonian path $\mathbb{P}$ using a breadth-first search strategy, initiated from an arbitrarily selected node. This path spans the entire graph, ensuring that each node is visited exactly once. Although traversing $\mathbb{P}$ over a graph can be non-trivial [4]. But such concerns can be addressed by constraining graphs using anatomical priors, spatial closeness, and intensity similarities [4]. The directed nature of $\mathbb{P}$ allows for the systematic definition of upstream and downstream nodes relative to any given node $v$ . Specifically, upstream nodes are those that succeed $v$ along $\mathbb{P}$ , while down + +![](images/21b145793e3bec304bfed2363e76d413a68570006c3173e9b56a88a63533b071.jpg) +(A) + +![](images/e11173dec24fd597f0ac6439e0fc5850f2cc6a7e54a5f854dcc4b444e48f78a1.jpg) +(B) +Figure 3. Illustration of the adjacency network construction in which: (A) superpixels are decomposed into graph nodes. (B) These nodes are connected to each other via undirected edges to form a complete graph. (C) The nodes are traversed in a graph using the Hamiltonian path $\mathbb{P}$ (shown in green) that starts from $v_{1}$ and visits each node $(v_{2}, v_{3}, v_{4})$ within the graph only once. (D) For each node in the graph (e.g., $v_{2}$ ), the superpixel patch $(s_{p}^{v_{2}})$ is generated, which is passed to the initial node encoder $N$ to generate initial node embeddings $h_{v_{2}}^{0}$ . Similarly, for each adjacent node $u$ of $v_{2}$ in path $\mathbb{P}$ , the edge descriptor $e_{v_{2},u}$ is computed as $e_{v_{2},u} = h_{v_{2}}^{0} - h_{u}^{0}$ which is passed to the $N_{e}^{0}$ to produce initial edge embeddings $e_{v_{2},u}^{0}$ . The traversal path, initial node, and edge embeddings are then used to formulate $\mathbb{A}$ . + +![](images/14c82b4d1316757e2c78c69e1613a120d9c24f8014675ec93fc6726a732227c4.jpg) +(C) +(D) + +Algorithm 1: Construction of $\mathbb{A}$ +Input: Superpixels S, Path P +Output: A +A $\leftarrow$ GenerateGraph(S) +foreach node v in P do + $s_p^v\gets$ GetSuperPixelPatch(v) + $s_p^v\gets$ ResizeSuperPixelPatch(sp) + $h_v^0\gets N(s_v^0)$ +A $\leftarrow$ UpdateNodeFeatures(A,v,hv) + $\Psi_v\gets$ GetAdjacentNodes(A,v) +foreach node u in $\Psi_v$ do + $s_p^u\gets$ GetSuperPixelPatch(u) + $s_p^u\gets$ ResizeSuperPixelPatch(su) + $h_u^0\gets N(s_u^0)$ $e_{v,u}\gets h_v^0 -h_u^0$ $e_{v,u}^{0}\gets N_{0}^{e}(e_{v,u})$ +A $\leftarrow$ UpdateEdgeFeatures(v,u,e0v,u,A) +end +end +return A + +stream nodes precede $v$ in the traversal order. This Hamiltonian path $\mathbb{P}$ serves as a structured framework for the orderly propagation of node and edge features across the graph $\mathbb{G}$ , facilitating the efficient flow of contextual information between superpixels. + +# 4.2.3. Adjacency Network Construction + +To systematically capture and organize node and edge features, we introduce the concept of an adjacency network (A). This network integrates the traversal path $\mathbb{P}$ with an augmented node structure that incorporates both node and edge descriptors at each node $v$ . + +The node descriptor is initialized with the feature vector $h_v^0$ , extracted using the pre-trained encoder $N$ . Notably, both node and edge features undergo iterative refinement in subsequent processing stages. This refinement is performed + +using six pre-trained feature extractor models (reported in Table 1) to enhance the representation of structural and relational features. + +The edge descriptor $e_{v,u}$ captures the relational characteristics between nodes $v$ and $u$ . Specifically, it is computed as the difference between their initial embeddings, expressed as: $e_{v,u} = h_v^0 - h_u^0$ . This edge descriptor is non-commutative, meaning that $e_{v,u} \neq e_{u,v}$ . To obtain the initial edge embeddings $(e_{v,u}^0)$ , the edge descriptor $(e_{v,u})$ is processed through the initial feature extractor model $N_e^0$ . The algorithm for the adjacency network construction is discussed in Algorithm 1. + +# 4.3. Neural Graph Featurization + +To model the relationships between lesion and non-lesion features, we employ an iterative neural graph featurization process. At each iteration, feature representations are propagated across the graph, facilitating the refinement of both node descriptors and edge connections. This iterative updating enhances the contextual representation of each node by integrating information from its neighboring nodes. The algorithm for the neural graph featurization routine is encapsulated in Algorithm 2, and we provide a detailed description in the subsequent sections. + +# 4.3.1. Edge Update + +At each iteration, the edge embeddings $e_{v,u}^{k}$ , which capture the relationship between node $v$ and its neighboring nodes $u$ , are updated using the neural network model $N_{e}^{p}$ . These updates in the edge embeddings are achieved by integrating the node embeddings from the previous iteration, $h_{v}^{k-1}$ and $h_{u}^{k-1}$ , along with the prior edge embeddings $e_{v,u}^{k-1}$ , as formally expressed in Equation 1. + +$$ +e _ {v, u} ^ {k} = N _ {e} ^ {p} \left(h _ {v} ^ {k - 1}, h _ {u} ^ {k - 1}, e _ {v, u} ^ {k - 1}\right). \tag {1} +$$ + +# 4.3.2. Node Update + +To refine the feature representation of each node $v$ , its neighboring nodes are categorized into two distinct groups: + +Algorithm 2: NGF Scheme +Input: Path $\mathbb{P}$ Adjacency Network A +Output: Updated A + $k\gets 1$ $k_{m}\gets$ maxIteration // propagation iterations +converge $\leftarrow$ False +while converge is False do if $k\geq k_m$ then break//the algorithm ran for max iterations end foreach node v in $\mathbb{P}$ do $\Psi_v\gets$ GetAdjacentNodes(A,v) $h^{k}_{Ds_v}\gets \phi$ $h^{k}_{Us_v}\gets \phi$ foreach node u in $\Psi_v$ do $e^{k}_{v,u}\gets N^{p}_{e}(h^{k - 1}_{v},h^{k - 1}_{u},e^{k - 1}_{v,u})$ $\mathbb{A}\gets$ UpdateEdgeFeatures(v,u, $e^{k}_{v,u}$ ,A) if $d\mathbb{P}_u < d\mathbb{P}_v$ then $h^{k}_{Ds_v}\gets$ $h^{k}_{Ds_v} + N_{Ds}(h^0_v,h^k_v,e^k_u)$ end if $d\mathbb{P}_u > d\mathbb{P}_v$ then $h^{k}_{Us_v}\gets h^{k}_{Us_v} + N_{Us}(h^{0}_v,h^{k - 1}_v,e^{k}_v)$ end +end + $h^k_v\gets N_n^p ([h^{k}_D s_v],h^{k}_U s_v])$ $\mathbb{A}\gets$ UpdateNodeFeatures(A,v, $h^k_v)$ +end + $k\gets k + 1$ +converge $\leftarrow$ HasConverged() // See Eq. (5) +end +return A + +upstream nodes $(Us)$ and downstream nodes $(Ds)$ . The upstream nodes are those encountered after node $v$ along the traversal path $\mathbb{P}$ , while the downstream nodes are those visited before $v$ . This division facilitates the incorporation of both appearance and geometric information as the traversal progresses through $\mathbb{P}$ . + +The embeddings for upstream $(h_{Us_v}^k)$ and downstream $(h_{Ds_v}^k)$ nodes of $v$ are computed as follows: + +$$ +h _ {U s _ {v}} ^ {k} = \sum_ {u \in U s} ^ {n _ {U s}} N _ {U s} \left(h _ {v} ^ {0}, h _ {v} ^ {k - 1}, e _ {v, u} ^ {k}\right), \tag {2} +$$ + +$$ +h _ {D s _ {v}} ^ {k} = \sum_ {u \in D s} ^ {n _ {D s}} N _ {D s} \left(h _ {v} ^ {0}, h _ {v} ^ {k - 1}, e _ {v, u} ^ {k}\right), \tag {3} +$$ + +where $n_{Us}$ and $n_{Ds}$ denote the number of upstream and downstream nodes, respectively. $N_{Us}$ and $N_{Ds}$ are distinct feature extraction models which generates the upstream em + +beddings $(h_{Us_v}^k)$ and downstream embeddings $(h_{Ds_v}^k)$ , respectively, for node $v$ in each iteration $k$ using the current edge embeddings $e_{v,u}^k$ , the node embeddings from the previous iteration $h_v^{k-1}$ , and the initial node descriptor $h_v^0$ . Incorporating $h_v^0$ ensures that the initial feature representation is preserved during iterative refinement. + +At each iteration $k$ , the node embeddings $h_{Ds_v}^k$ and $h_{Us_v}^k$ are concatenated and processed by the neural network model $N_n^p$ , yielding refined node embeddings $h_v^k$ . This operation, mathematically defined in Equation 4, enhances the contextual representation of each node, facilitating improved differentiation between dissimilar nodes. + +$$ +h _ {v} ^ {k} = N _ {n} ^ {p} \left(\left[ h _ {D s _ {v}} ^ {k}, h _ {U s _ {v}} ^ {k} \right]\right). \tag {4} +$$ + +The convergence of the algorithm is evaluated by monitoring changes in the adjacency matrix $(A^{m})$ over a span of $K$ iterations. Specifically, + +$$ +\sigma = \frac {\sum_ {j = 2} ^ {k _ {m}} A _ {j} ^ {m} \oplus A _ {j - 1} ^ {m}}{n _ {e}}, \tag {5} +$$ + +where $n_e$ is the normalization term representing the number of edges in the graph $\mathbb{G}$ , and $\oplus$ is an XOR operator. The metric $\sigma$ quantifies the cumulative change in the adjacency matrix $A^m$ over a window of $k_m$ iterations. It serves as a convergence criterion, indicating the point when the graph structure stabilizes during the NGF process. + +Graph Partitioning: After achieving the convergence of NGF process, we apply a graph-cut algorithm to partition the graph $\mathbb{G}$ into two disjoint subgraphs, $\mathbb{G}_a(V_a,E_a)$ and $\mathbb{G}_b(V_b,E_b)$ , with the objective of distinguishing between lesion and non-lesion patterns. The partitioning is determined using the normalized cut criterion, $N_{c}(V_{a},V_{b})$ , as introduced by Shi et al. [28]. + +To compute the cut cost, we first define the node association $h_{as}^{(a,b)}$ between arbitrary node sets $a$ and $b$ as the sum of the Euclidean distances between their respective node feature vectors $(h_m^k, h_n^k)$ , as expressed in Equation 6. Here, $n_f$ denotes the feature dimensionality obtained from the featurization step. + +$$ +h _ {a s} ^ {(a, b)} = \sum_ {m \in a, n \in b} \sqrt {\sum_ {i = 0} ^ {n _ {f} - 1} \left(h _ {m} ^ {k} (i) - h _ {n} ^ {k} (i)\right) ^ {2}}, \tag {6} +$$ + +Similarly, the edge association $e_{as}^{(a,b)}$ is defined as the normalized sum of edge feature values $(e_{a,b}^k)$ between nodes in $a$ and $b$ , as given in Equation 7. Here, $e_f$ represents the edge feature dimensionality. + +$$ +e _ {a s} ^ {(a, b)} = \frac {\sum_ {p \in a , q \in b | p \neq q} \sum_ {i = 0} ^ {e _ {f} - 1} e _ {p , q} ^ {k} (i)}{\sum_ {m \in V , n \in V | m \neq n} \sum_ {j = 0} ^ {e _ {f} - 1} e _ {m , n} ^ {k} (j)}, \tag {7} +$$ + +![](images/ad95ab851413d0e89607e8d387c036e2942ac968c75bdd97ba3496ffbdcbd947.jpg) +Figure 4. Visual representation of the outcome of neural graph featurization and the graph partitioning in the latent feature space. The top row (left to right) shows the outcomes after 2, 4, and 6 iterations, respectively. Bottom row (left to middle) shows the outcomes after 8 and 10 iterations, and the right shows the output after performing graph partitioning, which results in the formation of lesion and non-lesion clusters. + +![](images/54104501a2f49471991fbeebe9fdb6d3bc882792796d9ee9627e20e02143542f.jpg) +Figure 5. The proposed vision-language model-based approach for lesion classification. The output of the proposed neural graph cut is fed into the pre-trained vision encoder (PVT [31]), while the text prompt is provided to the pre-trained text encoder (BERT [7]), generating visual and textual embeddings. Cosine similarity between the visual and textual embeddings is computed, and the lesion class is assigned to the lesion patch based on the highest similarity. + +The cut cost $c(a, b)$ is then computed as the sum of the node and edge associations between $a$ and $b$ , as formulated in Equation 8. + +$$ +c (a, b) = \sum_ {m \in a, n \in b | m \neq n} \left(h _ {a s} ^ {(m, n)} + e _ {a s} ^ {(m, n)}\right). \tag {8} +$$ + +By substituting $V_{a}$ and $V_{b}$ for $a$ and $b$ , respectively, in Equation 8, we obtain the cut cost $c(V_{a}, V_{b})$ , which is subsequently used to compute the normalized cut $N_{c}(V_{a}, V_{b})$ , as defined in Equation 9. + +$$ +N _ {c} \left(V _ {a}, V _ {b}\right) = \frac {c \left(V _ {a} , V _ {b}\right)}{c \left(V _ {a} , V\right)} + \frac {c \left(V _ {a} , V _ {b}\right)}{c \left(V _ {b} , V\right)}. \tag {9} +$$ + +Figure 4 showcases the output of the NGF iterations and the graph partitioning in the latent space, which leads to the formation of lesion and non-lesion clusters. + +Table 1. The architectural details of the six models that are used in the proposed scheme. The abbreviation 'LL' stands for 'Linear Layer'. + +
NetworksInput DimensionOutput DimensionTotal Model Layers
N224 × 224 × 31 × 2048174 without LL
N01 × 20481 × 164 without LL
NEp1 × 41121 × 164 without LL
NUs1 × 41121 × 20484 without LL
NDs1 × 41121 × 20484 without LL
NNp1 × 40961 × 20484 without LL
+ +# 4.4. VLM-Based Lesion Classification: + +After obtaining the clusters $\mathbb{G}_a$ and $\mathbb{G}_b$ , which correspond to lesion and non-lesion superpixels within the input scan, respectively, we employed a zero-shot vision-language model for further classification of the extracted lesion superpixels into hard exudates, soft exudates, microaneurysms, and haemorrhages. Figure 5 depicts the employed vision-language-based classification scheme. To reduce semantic mismatch, we aligned the image and word embeddings within a shared dimensional space and computed their similarity using the cosine similarity function. The similarities are calculated between the visual and textual features computed, respectively, from each lesion superpixel within the input scan and the prompt template: "The fundus patch contains {lesion}", in which the keyword {lesion} is taken from the lesion vocabulary. + +For visual feature extraction, we utilized a Pyramid Vision Transformer (PVT) [31] as a feature extractor using its pretrained weights. In the text branch, we employed BERT [7] as the text encoder to generate textual features from the prompt template containing lesion labels, such as hard exudates, soft exudates, microaneurysms, and haemorrhages. Moreover, to compute cosine similarity, the BERT feature vector is first resized to length $n$ via sub-sampling, where $n$ is the size of the PVT feature vector. + +Afterward, the final label assigned to each superpixel patch was determined by the prompt whose text embedding exhibited the highest cosine similarity with the patch's visual embeddings, as extracted by the vision encoder. + +# 5. Experimental Setup + +This section presents the description of the datasets, the training and implementation protocols, and the evaluation metrics we used to evaluate the performance of the proposed scheme and compare it with state-of-the-art methods. + +# 5.1. Datasets + +We evaluated the proposed framework against state-of-the-art approaches using publicly available datasets, including ODIR [3], FIVES [17], BIOMISA [11], and the IDRiD [25]. Due to space constraints, we refer the reader to each dataset paper for its details. + +Table 2. Performance comparison with varying segment values for SLIC [1]. 'segment' represents the number of superpixels. Bold indicates the best scores. + +
DatasetSegmentsPrecisionRecallF1IoU
ODIR1000.880.910.890.80
500.840.870.850.74
100.770.810.790.65
BIOMISA1000.920.940.930.86
500.850.900.870.77
100.790.820.800.67
IDRiD1000.910.930.920.85
500.850.890.870.76
100.760.790.770.63
+ +Table 3. Analysis of the optimal balance between segmentation performance and computational complexity of different encoders on ODIR, BIOMISA, and IDRiD datasets. Bold indicates the best scores. + +
DatasetModelPrecisionRecallF1IoUGMACsTime (ms)
ODIRViT0.860.870.860.7619.60384
ResNet-500.840.860.850.7317.10462
PVT0.880.910.890.8010.40323
BIOMISAViT0.880.900.890.8017.40341
ResNet-500.890.910.900.8114.90427
PVT0.920.940.930.868.63306
IDRiDViT0.890.910.900.8116.20356
ResNet-500.870.900.880.7912.90392
PVT0.910.930.920.858.04284
+ +Table 4. Determining the convergence criteria across the ODIR, BIOMISA, and IDRiD datasets using different values of $K / \sigma$ . Bold indicates the best combination. $1\%$ , $2\%$ , and $3\%$ represent the acceptable percentage of change for NGF convergence. + +
K/σODIRBIOMISAIDRiD
1%2%3%1%2%3%1%2%3%
5118596413106
101612913107171411
1521171316129242119
+ +# 5.2. Training Details + +Six learnable models are employed in the proposed framework to extract and iteratively update node and edge features 1. In the zero-shot configuration, we utilized the node and edge models with their pre-trained weights directly, without undergoing any additional training or fine-tuning routines. For the few-shot experiments, we first fine-tuned all the NGF models offline using the ODIR dataset. The fine-tuning process involved training on 20 examples per class for 40 epochs, employing the ADAM optimizer and cross-entropy loss. After fine-tuning, we removed the classification heads from the models and used them as feature extractors in the neural graph featurization process. + +# 5.3. Evaluation Metrics + +To rigorously assess the effectiveness of the proposed framework, we employed a comprehensive set of standard performance metrics computed at the pixel level, including precision, recall, F1, and intersection-over-union (IoU). + +Table 5. Performance comparison of zero-shot and few-shot variants of the proposed scheme. Bold indicates the best scores. + +
DatasetMethodPrecisionRecallF1IoU
ODIRZero-shot0.880.910.890.80
Few-shot0.930.950.940.88
BIOMISAZero-shot0.920.940.930.86
Few-shot0.950.970.960.92
IDRiDZero-shot0.910.930.920.85
Few-shot0.940.960.950.90
+ +# 6. Results and Discussion + +This section provides a comprehensive analysis of the proposed framework's performance, benchmarking it against state-of-the-art methods across four public datasets. Additionally, we present ablation studies to assess the impact of key hyperparameters and determine the optimal configuration for the proposed framework. + +# 6.1. Ablation Studies + +We conducted a series of experiments to evaluate the various ablation aspects of the proposed method. + +# 6.1.1. Segmentation Analysis: + +The performance of the SLIC [1] to generate small superpixels was evaluated across various ODIR, BIOMISA, and IDRiD datasets using different numbers of segments: 10, 50, and 100. Table 2 presents the quantitative results, indicating that the best performance was achieved for segment '100'. + +# 6.1.2. Performance vs Computational Cost: + +To assess the trade-off between performance, cost, and efficiency, we conducted an ablation study by plugging different pre-trained models as $N$ within the proposed method. Specifically, we evaluated ViT [8], ResNet-50 [14], and PVT [31]. As shown in Table 3, the best trade-off between recognition performance, computational cost, and run-time efficiency is achieved with PVT [31] across the ODIR, BIOMISA, and IDRiD datasets. Given its capacity to provide an optimal trade-off between performance, cost, and efficiency, PVT [31] was chosen as the optimal node encoder for further experiments. + +# 6.1.3. Convergence Criteria: + +Table 4 summarizes the convergence results for different $\sigma$ thresholds and window lengths $(K)$ across all datasets. The results indicate that optimal convergence was consistently achieved with $K = 5$ and $\sigma = 1$ . + +# 6.1.4. Zero-shot vs Few-shot Analysis: + +In the zero-shot variant, pre-trained models were incorporated into the proposed method without undergoing additional training or fine-tuning rounds. However, fine-tuning these models on pre-task data with a small training set of 20 + +Table 6. Performance comparison of the proposed framework with SOTA methods across the four datasets. Bold indicates the best scores. + +
DatasetMethodApproachPrecisionRecallF1IoU
ODIRProposedUnsupervised0.880.910.890.80
DMSF [10]Supervised0.830.860.840.73
MSAG [34]Supervised0.810.840.820.70
BIOMISAProposedUnsupervised0.920.940.930.86
MSAG [34]Supervised0.890.910.900.81
DMSF [10]Supervised0.900.920.910.83
DRUNET [2]Supervised0.870.890.880.78
IDRiDProposedUnsupervised0.910.930.920.85
MSAG [34]Supervised0.860.890.870.77
DMSF [10]Supervised0.870.900.880.79
DRUNET [2]Supervised0.850.860.850.74
FIVESProposedUnsupervised0.920.930.920.86
SegDRoWS [21]Supervised0.870.890.880.79
MSAG [34]Supervised0.820.850.830.72
+ +![](images/26d781295fa0b1932ff0c5d32fd0f2f98dfc6fea639fd7034cfb7760f476391f.jpg) +Figure 6. Exemplar cases that showcase how accurately the graphs are constructed from the superpixels of different sizes by the proposed scheme. + +![](images/1d34f01af7b58ea89a07dd29f74d54d43fa7aef667ef7c49c5fbe9aa77d5a315.jpg) + +![](images/51a5e1b0233430f9239a35a0f450279aa926a8bd9fd10f9e505cfeddd37378a1.jpg) + +![](images/9144c55714c25be8bc0a6f9839a5b0121de4fe12b9ce5e6cc8a4e99cb69ab13b.jpg) + +![](images/ad2aa1a55fa170c86f4cb19f09cd7b777b1a9247e7b273bcb6e94c45c8da3536.jpg) + +![](images/df8725d6760f5c74f19f1abc840a72b3476aa730813ffce3404854bcf202a1b7.jpg) + +![](images/9f72afc36625d5595a2dc8a56c4586310f6388bb4395bff8ced5f153ecd5984f.jpg) +(A) + +![](images/cf2b331f10f5e3758f23d1be3db6b50f7f8f9eebef7c39dc839eab29eca1596c.jpg) +(B) + +![](images/f197d417a0ec3044e40df3b15e204cc75f651ad960e65778bd20d2361b0db3ea.jpg) +(C) + +![](images/441a0cea336b1b2f58823819edca32e0a0d4523f3eddba4798e30102d7ee72b7.jpg) +(D) +Figure 7. Qualitative analysis of retinal lesion classification using zero-shot and few-shot variants of the proposed method. (A) Fundus scans, (B) Ground truth, (C) Few-shot variant, and (D) Zero-shot variant. Blue represents hard exudates, cyan represents hemorrhages, and yellow represents soft exudates. Zoom in for the best visualization. + +samples significantly enhanced the overall performance of the proposed model across ODIR, BIOMISA, and IDRiD datasets (see Table 5). Nevertheless, the performance of the zero-shot variant is also remarkable compared to the SOTA works (see Table 6). + +# 6.2. Comparative Analysis with SOTA + +Table 6 presents a quantitative comparison of the proposed method with state-of-the-art works across all datasets. The + +results clearly indicate that the proposed method substantially surpasses the fully supervised SOTA models. These improvements are achieved by exploiting the homogeneity and heterogeneity between lesion and non-lesion nodes during the neural graph featurization process, which ultimately leads to the robust discrimination of lesion and non-lesion superpixels within the input scans. The lesion regions are further classified into their appropriate classes using the vision language model. + +# 6.3. Qualitative Analysis + +Figure 6 presents some exemplar cases that reveal how accurately the graphs are constructed from the superpixels of different sizes by the proposed scheme, where these graphs are utilized in the NGF scheme to extract the retinal lesions accurately. Similarly, Figure 7 reports exemplar cases that demonstrate the qualitative evaluation of the zero-shot and few-shot variants of the proposed method. We can notice from Figure 7 that the few-shot variant produces remarkable performance toward extracting different lesion types. Nevertheless, the zero-shot variant of the proposed method also produces promising lesion extraction performance, especially considering the fact that it does not require any training or fine-tuning routines. + +# 7. Conclusion + +This paper presents an unsupervised vision-language neural graph featurization method that extracts retinal lesions from the fundus scans irrespective of the scanner specifications, the dataset variations, or the vendor artifacts. Furthermore, the proposed method does not require any exhaustive training or fine-tuning rounds, as well as ground truth labels, to extract the retinal lesions. The proposed method is rigorously evaluated on four public datasets, dubbed ODIR, FIVES, BIOMISA, and IDRiD, where across each dataset, it outperforms state-of-the-art supervised learning methods in terms of F1 and IoU scores. The intrinsic capability of the proposed method to extract retinal lesions from the scans acquired with different scanners highlights its potential to be deployed in the real-world clinical setting to facilitate doctors, unlike the competitive supervised learning approaches. In the future, we envisage testing the proposed method in broader medical and non-medical domains to perform fine-grained segmentation and classification of small-sized objects. + +# Acknowledgment + +This research is funded by Abu Dhabi University (Ref: 19300950, 19300910, and 19300979), and the Center for Cyber-physical Systems (C2PS), Khalifa University (Ref: CIRA-2021-052). + +# References + +[1] Radhakrishna Achanta, Appu Shaji, Kevin Smith, Aurelien Lucchi, Pascal Fua, and Sabine Susstrunk. SLIC Superpixels Compared to State-of-the-Art Superpixel Methods. IEEE Transactions on Pattern Analysis and Machine Intelligence, November 2012. 1, 2, 3, 7 +[2] Meshal Alharbi and Deepak Gupta. Segmentation of diabetic retinopathy images using deep feature fused residual with U-Net. *Alexandria Engineering Journal*, November 2023. 8 +[3] N Balakrishna, MB Mukesh Krishnan, E Venkata Ram Sai, S Vinish Ranganath, K Sonika, and L Gouri Priyanka. Ocular disease recognition using efficientnet. In 2024 3rd International Conference on Applied Artificial Intelligence and Computing (ICAAIC), pages 1069-1074. IEEE, 2024. 6 +[4] Jesse Beisegel, Fabienne Ratajczak, and Robert Scheffler. Computing Hamiltonian Paths with Partial Order Restrictions. ACM Transactions on Computation Theory, Volume 17, Issue 1, February 2025. 3 +[5] J Jasper Gnana Chandran, J Jabez, and Senduru Srinivasulu. Auto-metric graph neural network optimized with capuchin search optimization algorithm for coinciding diabetic retinopathy and diabetic macular edema grading. Biomedical Signal Processing and Control, 80:104386, 2023. 1, 2 +[6] Yinlin Cheng, Mengnan Ma, Xingyu Li, and Yi Zhou. Multi-label classification of fundus images based on graph convolutional network. BMC Medical Informatics and Decision Making, 21:1-9, 2021. 2 +[7] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. ACL: Human Language Technologies, 2019. 6 +[8] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale. International Conference on Learning Representations (ICLR), 2021. 2, 7 +[9] Anjing Guo, Leyuan Fang, Min Qi, and Shutao Li. Unsupervised denoising of optical coherence tomography images with nonlocal-generative adversarial network. IEEE Transactions on Instrumentation and Measurement, 70:1-12, 2020. 1 +[10] Tianjiao Guo, Jie Yang, and Qi Yu. Diabetic Retinopathy Lesion Segmentation using Deep Multi-scale Framework. Biomedical Signal Processing and Control, February 2024. 1, 8 +[11] Taimur Hassan, M Usman Akram, M Furqan Masood, and Ubaidullah Yasin. Biomisa retinal image database for macular and ocular syndromes. In Image Analysis and Recognition: 15th International Conference, ICIAR 2018, Póvoa de Varzim, Portugal, June 27–29, 2018, Proceedings 15, pages 695–705. Springer, 2018. 6 +[12] Taimur Hassan, Sajid Javed, Arif Mahmood, Talha Qaiser, Naoufel Werghi, and Nasir Rajpoot. Nucleus classification in histology images using message passing network. Medical Image Analysis, 79:102480, 2022. 2 + +[13] Taimur Hassan, Zhu Li, Sajid Javed, Jorge Dias, and Naoufel Werghi. Neural graph refinement for robust recognition of nuclei communities in histopathological landscape. IEEE Transactions on Image Processing, 33:241-256, 2023. 2 +[14] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 7 +[15] Shiqi Huang, Jianan Li, Yuze Xiao, Ning Shen, and Tingfa Xu. Rtnet: relation transformer network for diabetic retinopathy multi-lesion segmentation. IEEE Transactions on Medical Imaging, 41(6):1596–1607, 2022. 2 +[16] Yijin Huang, Li Lin, Pujin Cheng, Junyan Lyu, and Xiaoying Tang. Lesion-based contrastive learning for diabetic retinopathy grading from fundus images. In Medical Image Computing and Computer Assisted Intervention—MICCAI 2021: 24th International Conference, Strasbourg, France, September 27–October 1, 2021, Proceedings, Part II 24, pages 113–123. Springer, 2021. 2 +[17] Kai Jin, Xingru Huang, Jingxing Zhou, Yunxiang Li, Yan Yan, Yibao Sun, Qianni Zhang, Yaqi Wang, and Juan Ye. Fives: A fundus image dataset for artificial intelligence based vessel segmentation. Scientific data, 9(1):475, 2022. 6 +[18] Yogesh Kumar and Bharat Gupta. Retinal image blood vessel classification using hybrid deep learning in cataract diseased fundus images. Biomedical Signal Processing and Control, 84:104776, 2023. 1, 2 +[19] Yongjia Lei, Shuyuan Lin, Zhiying Li, Yachao Zhang, and Taotao Lai. Gnn-fused capsnet with multi-head prediction for diabetic retinopathy grading. Engineering Applications of Artificial Intelligence, 133:107994, 2024. 2 +[20] Yiyue Li, Qicheng Lao, Qingbo Kang, Zekun Jiang, Shiyi Du, Shaoting Zhang, and Kang Li. Self-supervised anomaly detection, staging and segmentation for retinal images. Medical Image Analysis, 87:102805, 2023. 2 +[21] Jian Liu, Haiying Che, Aidi Zhao, Na Li, Xiao Huang, Hui Li, and Zhihong Jiang. SegDRoWS: Segmentation of diabetic retinopathy lesions by a whole-stage multi-scale feature fusion network. Biomedical Signal Processing and Control, July 2025. 8 +[22] Yanbei Liu, Xinwen Peng, Xin Wei, Lei Geng, Fang Zhang, Zhitao Xiao, and Jerry Chun-Wei Lin. Label-aware dual graph neural networks for multi-label fundus image classification. IEEE Journal of Biomedical and Health Informatics, 2024. 2 +[23] Sandeep Madarapu, Samit Ari, and KK Mahapatra. A deep integrative approach for diabetic retinopathy classification with synergistic channel-spatial and self-attention mechanism. Expert Systems with Applications, 249:123523, 2024. 2 +[24] Saif Khalid Musluh, Ammar M Okran, Saddam Abdulwahab, Domenec Puig, and Hatem A Rashwan. Advanced diabetic retinopathy classification: Integrating pathological indicators segmentation and morphological feature analysis. In International Workshop on Ophthalmic Medical Image Analysis, pages 104-114. Springer, 2024. 2 +[25] Prasanna Porwal, Samiksha Pachade, Manesh Kokare, Girish Deshmukh, Jaemin Son, Woong Bae, Lihong Liu, + +Jianzong Wang, Xinhui Liu, Liangxin Gao, et al. Idrid: Diabetic retinopathy-segmentation and grading challenge. Medical image analysis, 59:101561, 2020. 6 +[26] Amritha Abdul Salam, Manjunatha Mahadevappa, Asha Das, and Madhu S Nair. Drg-net: A graph neural network for computer-aided grading of diabetic retinopathy. Signal, Image and Video Processing, 16(7):1869–1875, 2022. 2 +[27] P Saranya, S Prabakaran, Rahul Kumar, and Eshani Das. Blood vessel segmentation in retinal fundus images for proliferative diabetic retinopathy screening using deep learning. The visual computer, pages 1-16, 2022. 2 +[28] Jianbo Shi and Jitendra Malik. Normalized cuts and image segmentation. IEEE Transactions on pattern analysis and machine intelligence, 22(8):888-905, 2000. 5 +[29] Julio Silva-Rodriguez, Hadi Chakor, Riadh Kobbi, Jose Dolz, and Ismail Ben Ayed. A foundation language-image model of the retina (flair): Encoding expert knowledge in text supervision. Medical Image Analysis, 99:103357, 2025. 2 +[30] Muhammad Zeeshan Tahir, Muhammad Nasir, and Sanyuan Zhang. Advances in retinal microaneurysms detection, segmentation and datasets for the diagnosis of diabetic retinopathy: a systematic literature review. Multimedia Tools and Applications, 83(30):74897-74935, 2024. 1 +[31] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pyramid vision transformer: A versatile backbone for dense prediction without convolutions. In Proceedings of the IEEE/CVF international conference on computer vision, pages 568-578, 2021. 6, 7 +[32] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pyramid vision transformer: A versatile backbone for dense prediction without convolutions. In Proceedings of the IEEE/CVF international conference on computer vision, pages 568-578, 2021. 2 +[33] Xin Wei, Yanbei Liu, Fang Zhang, Lei Geng, Chunyan Shan, Xiangyu Cao, and Zhitao Xiao. Mstnet: Multi-scale spatial-aware transformer with multi-instance learning for diabetic retinopathy classification. Medical Image Analysis, page 103511, 2025. 2 +[34] Changzhuan Xu, Song He, and Hailin Li. An attentional mechanism model for segmenting multiple lesion regions in the diabetic retina. Scientific Reports, 2024. 8 +[35] Hao Xu and Yun Wu. G2vit: Graph neural network-guided vision transformer enhanced network for retinal vessel and coronary angiograph segmentation. *Neural Networks*, 176: 106356, 2024. 2 +[36] Guanghua Zhang, Jing Pan, Zhaoxia Zhang, Heng Zhang, Changyuan Xing, Bin Sun, and Ming Li. Hybrid graph convolutional network for semi-supervised retinal image classification. IEEE Access, 9:35778-35789, 2021. 2 +[37] Theodore Zhao, Yu Gu, Jianwei Yang, Naoto Usuyama, Ho Hin Lee, Tristan Naumann, Jianfeng Gao, Angela Crabtree, Jacob Abel, Christine Moung-Wen, et al. Biomed-parse: a biomedical foundation model for image parsing of everything everywhere all at once. arXiv preprint arXiv:2405.12971, 2024. 2 + +[38] Yi Zhou, Boyang Wang, Lei Huang, Shanshan Cui, and Ling Shao. A benchmark for studying diabetic retinopathy: segmentation, grading, and transferability. IEEE transactions on medical imaging, 40(3):818-828, 2020. 1 +[39] Muhammad Zubair, Muhammad Umair, Rizwan Ali Naqvi, Dildar Hussain, Muhammad Owais, and Naoufel Werghi. A comprehensive computer-aided system for an early-stage diagnosis and classification of diabetic macular edema. Journal of King Saud University-Computer and Information Sciences, 35(8):101719, 2023. 1 \ No newline at end of file diff --git a/ICCV/2025/Vision-Language Neural Graph Featurization for Extracting Retinal Lesions/images.zip b/ICCV/2025/Vision-Language Neural Graph Featurization for Extracting Retinal Lesions/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..c6a6eb3803c9faf80939376401e31c17c0e1bfc2 --- /dev/null +++ b/ICCV/2025/Vision-Language Neural Graph Featurization for Extracting Retinal Lesions/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:81a3384f57588f9fb4c71501ca8c10867c51743b609d872a6fbe3a797e17e814 +size 391146 diff --git a/ICCV/2025/Vision-Language Neural Graph Featurization for Extracting Retinal Lesions/layout.json b/ICCV/2025/Vision-Language Neural Graph Featurization for Extracting Retinal Lesions/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..982c5729910cd60af83bdf31ecd67f1ca05a4637 --- /dev/null +++ b/ICCV/2025/Vision-Language Neural Graph Featurization for Extracting Retinal Lesions/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:60accdf9ec1e4851bc3f8150ddc24435aa982028ffacfdbfc1f8cc8154a97340 +size 484731 diff --git a/ICCV/2025/VisionMath_ Vision-Form Mathematical Problem-Solving/e1365b02-1cb8-4330-95ba-78f25112df6d_content_list.json b/ICCV/2025/VisionMath_ Vision-Form Mathematical Problem-Solving/e1365b02-1cb8-4330-95ba-78f25112df6d_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..fca193cde363c2a0a6edb39626a9179d3db1fc9b --- /dev/null +++ b/ICCV/2025/VisionMath_ Vision-Form Mathematical Problem-Solving/e1365b02-1cb8-4330-95ba-78f25112df6d_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a147eb89ea57f034d9ab85abc2149e8a858becd0063315221b402e136a35f66e +size 88966 diff --git a/ICCV/2025/VisionMath_ Vision-Form Mathematical Problem-Solving/e1365b02-1cb8-4330-95ba-78f25112df6d_model.json b/ICCV/2025/VisionMath_ Vision-Form Mathematical Problem-Solving/e1365b02-1cb8-4330-95ba-78f25112df6d_model.json new file mode 100644 index 0000000000000000000000000000000000000000..d7912a25205fff2c4e05e75a1ae57369127593d2 --- /dev/null +++ b/ICCV/2025/VisionMath_ Vision-Form Mathematical Problem-Solving/e1365b02-1cb8-4330-95ba-78f25112df6d_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6bf3b4d0a37d48d3461726121f9dfe6c2e10101de2b47e312ae19b9ebd60c8dd +size 111021 diff --git a/ICCV/2025/VisionMath_ Vision-Form Mathematical Problem-Solving/e1365b02-1cb8-4330-95ba-78f25112df6d_origin.pdf b/ICCV/2025/VisionMath_ Vision-Form Mathematical Problem-Solving/e1365b02-1cb8-4330-95ba-78f25112df6d_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..dc8d15e1bff4b79499d7b7313a316f5911db1569 --- /dev/null +++ b/ICCV/2025/VisionMath_ Vision-Form Mathematical Problem-Solving/e1365b02-1cb8-4330-95ba-78f25112df6d_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ec50cefd517b1dfdaea220fe44dac904fdcda8908587e47141e9e4e80467d07a +size 1081849 diff --git a/ICCV/2025/VisionMath_ Vision-Form Mathematical Problem-Solving/full.md b/ICCV/2025/VisionMath_ Vision-Form Mathematical Problem-Solving/full.md new file mode 100644 index 0000000000000000000000000000000000000000..7d8849139c6313afc66c9cb94b3d7ee3aa4f0ec9 --- /dev/null +++ b/ICCV/2025/VisionMath_ Vision-Form Mathematical Problem-Solving/full.md @@ -0,0 +1,347 @@ +# VisionMath: Vision-Form Mathematical Problem-Solving + +Zongyang Ma $^{1,2,4,5*}$ , Yuxin Chen $^{2*†}$ , Ziqi Zhang $^{1,5}$ , Zhongang Qi $^{6}$ , Chunfeng Yuan $^{1,5}$ , Shaojie Zhu $^{3}$ , Chengxiang Zhuo $^{3}$ , Bing Li $^{1,5}$ , Ye Liu $^{2,7}$ , Zang Li $^{3}$ , Ying Shan $^{2}$ , Weiming Hu $^{1,4,5,8}$ $^{1}$ State Key Laboratory of Multimodal Artificial Intelligence Systems, Institute of Automation, Chinese Academy of Sciences ${}^{2}$ ARC Lab, ${}^{3}$ Tencent PCG ${}^{4}$ School of Artificial Intelligence, University of Chinese Academy of Sciences ${}^{5}$ Beijing Key Laboratory of Super Intelligent Security of Multi-Modal Information ${}^{6}$ vivo Mobile Communication Co. ${}^{7}$ The Hong Kong Polytechnic University ${}^{8}$ School of Information Science and Technology, ShanghaiTech University $\{\mathrm{mazongyang2020,zhangziqi2017}\} @ia.ac.cn$ , $\{\mathrm{cfyuan,bli,wmhu}\} @nlpr.ia.ac.cn$ $\{\mathrm{uasonchen,alisyzhu,felixzhuo,gavinzli,yingsshan}\} @tencent.com$ , zhongangqi@gmail.com, coco.ye.liu@connect.polyu.hk + +# Abstract + +Mathematical problems in real-world scenarios are often presented in a purely vision-form, where textual problem statement and accompanying math figures, e.g., geometry figures and functional graphs, are integrated into a single image. This vision-form problem-solving task requires precise comprehension and reasoning on both textual and graphical elements in the images, posing significant challenge to current Multimodal Large Language Models (MLLMs), which process text and math figures in isolation. In this work, we propose VisionMath, the first exploration for vision-form mathematical problem-solving model, which employs a three-stage progressive multimodal reasoning alignment strategy to systematically enhance task-specific capabilities. Building upon a LLM proficient in unimodal mathematical reasoning, VisionMath first establishes foundational OCR capabilities through capturing rendered mathematical problem images. Subsequently, the model develops comprehensive understanding of figure structures and properties via learning from figure descriptions and mathematical educational videos. Finally, the model's reasoning capacity is activated using carefully constructed visual-form problem-solving datasets VisionMath-IT with chain-of-thought annotations. For comprehensive evaluation, we construct multilingual benchmarks covering diverse problem types, including geometry, algebra, function problems in both English and Chinese. Experimental results demonstrate that VisionMath significantly outperforms existing general-purpose + +![](images/f182661044a71f3f41a8fe839c8c0d5f3e4d0a223fd1c798c7492c2801b6f080.jpg) + +(a) +Images: +![](images/d77cae28c529a6fdc4faec70e0dcb564ee3697370d6e18fa2148d9ef867b27fd.jpg) +By cutting off two triangular pyramids from the square shape (as shown in Figure 1), we obtain the geometry shown in Figure 2. The left view of this geometric shape is(). + +![](images/b6f78fc89f7d667231199ca3b32ddfe3400d95eb4e7ed5b2eec3e0f7743fc73b.jpg) +(b) +Figure 1. Demonstration of mathematical problem examples of image-text-separation-form (a) and vision-form (b). + +and math-specialized models, showcasing superior multilingual and multi-type problem-solving capabilities. Our model weights, data and code will be made available at https://github.com/mengqiDyangge/VisionMath. + +# 1. Introduction + +Recent advancements in reasoning capabilities have significantly enhanced the performance of both general-purpose [3, 7-9, 22, 27-29, 50, 52] and math-specialized MLLMs [10, 14, 32, 42], enabling them to demonstrate re + +markable proficiency in solving various types of mathematical problems. Current MLLMs primarily address "imagetext-separation-form" mathematical problems, where the textual problem statement (including formulas) and math figures, such as geometric figures and functional graphs, are input separately into the model, as shown in Figure 1(a). However, in real-world scenarios, problem statement and math figures are often captured within a single image, as the demonstration in Figure 1(b). The lack of exploration has led to the inability of existing MLLMs to effectively process such "vision-form" problems. For instance, on the MathVista benchmark [32], when switching the input samples from image-text-separation-form to vision-form, the accuracies of GPT-4o mini for Geometry Problem Solving (GPS) and Math Word Problems (MWP) decreases from $58.1\%$ to $53.4\%$ and from $59.6\%$ to $53.2\%$ , respectively. Such performance gap limits the model's real-world application, especially when the problem involves complex formulas and figures. + +The above issue drives us to take the first step in exploring a MLLM tailored for vision-form math problem-solving. To enhance the model's understanding and reasoning capabilities for both text and figures, we develop our MLLM, named VisionMath, focusing on three critical aspects: text recognition, figure comprehension, and integrated text-figure reasoning. Specifically, we design a three-stage progressive multimodal reasoning alignment strategy. In the first stage, we establish the basic OCR capabilities, enabling accurate perception of problem statements, especially formulas. The training combines document image OCR datasets [6, 16, 21, 37, 47, 51] with rendered math problem image datasets, requiring the model to transcribe the text present in the images. The second stage strengthens figure comprehension with a dual-path enhancement strategy. On one hand, math figure captioning datasets [14, 60] are employed to explicitly improve the structural visual parsing of geometric figures and functional graphs. Note that these captions generated by LLMs or rule-based systems often suffer from limited diversity and domain gap with real-world math figures. Thus, on the other hand, we foster a more in-depth math property understanding of real-world figures via learning from mathematical educational videos [61], which offer diverse and real-world math figures across complexity levels accompanied by teachers' profound analysis. In the final stage, we employ vision-form problem-solving datasets to instruct the model to perform step-by-step reasoning based on the joint text-figures understanding, thus improving the integrated text-figure reasoning capabilities. + +To deal with the lack of vision-form mathematical problem-solving datasets, we introduce VisionMath-IT, a large-scale, high-quality bilingual (English/Chinese) instruction tuning dataset. This dataset, developed through a + +![](images/6e706938a202876a42748a1c8fe9d3f996c9c7d62c72b705c7d5b1eae21f72ef.jpg) +Figure 2. Performance comparisons between VisionMath and other representative MLLMs. + +systematic pipeline involving data collection, filtering, rendering, and re-formating, offers a rich collection of visionform math problems. Furthermore, we develop specialized evaluation benchmarks to thoroughly assess the performance of MLLMs in solving various types of vision-form math problems. The first is VisionMath-Bench, which is a Chinese benchmark encompassing geometry, algebra, and function problems. Others are English benchmarks derived from commonly used image-text-separation-form datasets, including Vision-MathVista [32] and Vision-MathVerse [59]. Extensive experimental results demonstrate that our VisionMath significantly outperforms all open-source general-purpose and math-specialized MLLMs across these benchmarks. Simultaneously, VisionMath shows substantial improvements over GPT-4o on VisionMath-Bench and Vision-MathVista, while achieving overall performance gains compared to GPT-4o on Vision-MathVerse. Intuitive comparisons of representative models are visualized in Figure 2. Our contributions are summarized as follows: + +- We present the first exploration of practical vision-form math problem-solving task, proposing a progressive multimodal alignment reasoning strategy to systematically enhance the model's visual-form mathematical comprehension and reasoning capabilities. +- Through the design of a rigorous data construction pipeline, we meticulously curated a large-scale, high-quality dataset for vision-form mathematical problem-solving, named VisionMath-IT. +- We develop comprehensive benchmarks covering bilingual algebraic, geometric, and functional problems, and evaluate multiple state-of-the-art MLLMs to benchmark their performance on vision-form math problem-solving tasks. Extensive results highlight the superiority of our proposed approach. + +
SamplesDataset
VisionMath-OCRDocument image OCR and understanding1.36MChartQA [35], Deepform [46], DocVQA [36], InfographicVQA [38], Kleister Charity [43], Tabfact [6], WikiTableQuestions [40], Visualmrc [47], Ccpdf [51], RVL-CDIP [15] synthdog [20], Vary [54], pubtabnet [62], Turl [11], OCRCC [5], Dvqa [18], FigureQA [19], PlotQA [39], MMC [26] VisionMath-IT-Chinese, MMathCoT-1M [33]
Mathematical Problem OCR1.34 M
VisionMath-FigureFigure Caption0.55MGeo170k-Caption [14], MAVIS-Caption [60]
Mathematical Educational Video0.60MMultimodal-textbook [61]
VisionMath-ITEnglish Vision-Form Problem1.96MGeo170k [14], MathV360 [42], MultiMath-300k-English [41], MMathCoT-1M [33], Numbia [4]
Chinese Vision-Form Problem0.62MVisionMath-IT-Chinese, MultiMath-300k-Chinese [41]
+ +Table 1. Detailed subsets and data volumes of different training datasets. + +# 2. Related Work + +General-purpose Multimodal Large Language Model. Recent years have witnessed remarkable advancements in general-purpose MLLMs. LLaVA series [22, 27-29] has pioneered systematic explorations of effective training strategies, including multi-stage alignment, multi-task data rationing, reinforcement learning, and synthetic data training, significantly improving the capabilities. Meanwhile, QwenVL series [3, 50, 52] demonstrate SOTA performance across diverse tasks such as high-resolution image understanding, document understanding, and long-form video understanding. InternVL series [7-9] use larger-scale visual encoder for multimodal pre-training to bridge the performance gap between commercial [2, 48, 49] and open-source models. EMU series [44, 45, 53] aims to unify multimodal understanding and generation. MiniCPM [57] substantially advances edge-device multimodal intelligence through architectural and efficiency innovations. While these advancements have collectively propelled MLLM capabilities across diverse vision tasks [13, 23-25, 30, 34-36, 58], their multimodal reasoning and mathematical problem-solving proficiency remains suboptimal due to the inherent trade-offs in general-purpose design. + +Math-specialized Multi-Modal Language Model. To improve mathematical reasoning, several works have made multifaceted attempts. G-LLaVA [14] first collects geometry data to improve geometry problem solving. Math-LLaVA [42] and R-COT [10] employ rule-based synthetic data generation, significantly improving performance on geometry and algebra problems. MAVIS [60], MathGLM-Vision [56], and MultiMath [41] develop large-scale multisource multimodal math datasets to strengthen mathematical reasoning through domain-adaptive fine-tuning. MathPUMA [63] constructs paired samples conveying identical mathematical concepts to bridge the performance discrepancies between text-rich and image-rich problems. URSA [33] introduces error localization supervision to minimize logical errors and visual misinterpretations during reasoning. Despite these developments significantly expanding the capabilities frontier, previous works overlook fully visionform mathematical problem-solving, leading to suboptimal performance in this practical task. This has inspired us to conduct an in-depth analysis on the characteristics of this task and make targeted improvements. + +# 3. Method + +# 3.1. Data Engine of VisionMath + +High-quality data is crucial for effective model training. To meet the capability requirements for solving vision-based mathematical problems, ranging from basic perception to complex reasoning, we meticulously develop three specialized datasets: VisionMath-OCR, VisionMath-Figure, and VisionMath-IT. The detailed composition of each dataset is shown in Table 1. + +# 3.1.1. VisionMath-OCR + +VisionMath-OCR is designed to enable accurate recognition and parsing of general text, formulas, and symbolic notations within problem statements. Specifically, this dataset is partitioned into two components. + +Document Image OCR. This component includes a collection of document image OCR and understanding datasets (as listed in Table 1), aimed at establishing general-purpose OCR capabilities with a focus on recognizing standard text, charts, and tables. + +Mathematical Problem OCR. This component consists of rendered images with multilingual mathematical problems from the datasets in Table 1. It requires the model to accurately restate the complete problem statement, ensuring precise OCR of formulaic structures and symbolic notations. + +# 3.1.2. VisionMath-Figure + +VisionMath-Figure is introduced to enhance math figure perception and comprehension, particularly the properties and structural relationships of points, lines, and surfaces in geometry figures and function graphs. Based on the sources of data, the dataset can be divided into two parts: + +Figure Captions. This portion of the data is derived from the Geo170k [14] and MAVIS [60] datasets, consisting of mathematical figures and their corresponding descriptions. These descriptions contain basic component information about the figures, which helps enhance the model's figure structure perception and understanding. It is important to note that the figures in the Geo170k dataset are primarily planar geometric shapes, which limits their diversity. On the other hand, the figures in the MAVIS dataset are mainly constructed using rule-based methods, often resulting in a domain gap compared to real-world mathematical figures. To address this, we also utilize mathematical educational + +![](images/3eecf347c61337a1d803fc2772bef319a488d5af57a9a955849ecaab2d14f449.jpg) +Figure 3. Pipeline of Chinese Vision-Form Problem Synthesis. + +videos as an additional data source to further improve the understanding of mathematical figures. + +Mathematical Educational Videos. We observe that mathematical educational videos contain a rich variety of real-world mathematical figures, along with in-depth analysis and extensive knowledge about these figures, such as the properties of geometric shapes. This information is beneficial for the understanding and reasoning of mathematical figures. Therefore, we collect mathematical videos from the Multimodal Textbook [61] and construct training samples with the following steps: + +Step1: Subject Filtering. The videos in the Multimodal Textbook encompass multiple subjects. Therefore, we first utilize a Large Language Model (LLM) to filter the videos based on their titles and select those that are related to mathematics education. + +Step2: ASR Text Filtering. Multimodal Textbook has already extracted frames from the videos and used ASR tools to transcribe the corresponding teacher explanations. However, some of the transcriptions contain non-informative filler phrases devoid of mathematical knowledge. Therefore, we employ an LLM to filter out text that does not include mathematical explanations. + +# 3.1.3. VisionMath-IT + +VisionMath-IT is a bilingual (English/Chinese) instruction-tuning dataset aimed at empowering end-to-end vision-form mathematical problem-solving, and English-Chinese examples of different problem types are shown in Sec. C. of the supplementary materials. Its construction adheres to two principles: (1) Diversity. Curating visually-presented mathematical problems spans algebra, geometry (including plane geometry, analytic geometry, solid geometry) and function problems with accompanying figures. (2) Structured Solution Formatting. The responses are unifying into (chain-of-thought, final answer) pairs. This not only covers a wide range of problem types to improve generalization but also eliminates style variations in solution formats, allowing to concentrate on learning the reasoning process. + +To ensure VisionMath-IT meets rigorous diversity requirements and enables models to effectively learn heterogeneous reasoning patterns from various solution structures, we implement a multi-source data curation strategy. + +# English Vision-Form Problem Synthesis. + +The publicly available English image-text-separation-form datasets already encompass diverse question types and substantial samples. Therefore, the focus is on selecting high-quality data for vision-form problem transformation and solution formalizing. The specific steps are as follows: + +Step 1: Data Filtering. To filter out high-quality data from existing datasets, we first sample a portion of instances from different datasets and manually inspect the clarity of the mathematical images. We then retain the datasets with overall acceptable visual quality to serve as the source for training data. The retained dataset is shown in the "English Vision-Form Problem" section of Table 1. + +Step 2: Image-form Problem Rendering. For each sample in the retained image-text-separation-form datasets, the text-based problem statement and the associated math figures are jointly rendered into a single image, resulting in vision-form mathematical datasets. + +Step 3: Solution Formatting. For each problem, we design rules to split the original solution's chain-of-thought process and final answer, and unify the response format using the template “{CoT process}. Therefore, the answer is \boxed{\text{final answer}}”. + +# Chinese Vision-Form Problem Synthesis. + +Due to the lack of specialized high-quality and diversity in existing datasets for Chinese mathematical problems, we conduct extensive data collection on the Internet and perform automatic data filtering and curation strategies. As shown in Figure 3, the specific pipeline are as follows: + +Step 1: Source HTML Decomposition. The HTML collected from the web often contains much irrelevant information. Therefore, we first extract the HTML parts corresponding to the question, solution, and answer based on predefined rules. And the question images can then be obtained by rendering the html code corresponding to question. + +![](images/fd676b87749a8ea0013338584312089c2a8c9aa241d74e20a5bd624c736581b7.jpg) + +![](images/7b2f811b23c50086d1accb3e429a26b14043dd68235bb2e57618f4456583bb86.jpg) + +![](images/6458ee7dd83af60d00fd7c908227424af6c86f34618f984f136e12d3198708cf.jpg) +Stage2: Figure Perception & Understanding +Figure 4. The architecture of VisionMath, and the accompanying three stage progressive multimodal reasoning alignment training strategy for enhancing Mathematical OCR, deepening math figure Understanding, and fine-tuning problem-solving capabilities. + +![](images/e35cb5521eeecbeddcc64c7f1195b866427b8fbfc7642020163ac1ad509d050b.jpg) +Stage3: Vision-Form Mathematical Instruction Tuning + +Step 2: Formula Standardization. In the HTML code, mathematical formulas may appear as image paths or in style code formats. To obtain readable question and answer text, these image paths or style codes must be extracted, rendered into images, and subsequently converted into LaTeX code using OCR tools. By replacing the original mathematical formula codes with the LaTeX codes, we can ultimately obtain the raw question, solution, and answer text. + +Step 3: Text Cleaning and Formatting. The raw texts obtained from stage 2 often encounter some quality issues, including incorrectly recognizing mathematical images as formulas and converting them into garbled LaTeX, along with improperly mixed presentation of solution steps and final answers. To clean and standardize the raw texts, we leverage the powerful text understanding capabilities of LLM to filter out any garbled LaTeX. Additionally, we also use LLMs to explicit separate the CoT process and answer text, fill in any missing answers, and organize them into the format “{CoT process}. Therefore, the answer is \boxed{final answer}”. + +The final dataset, VisionMath-IT-Chinese, consists of 332K math problems related to K-12 education. This corpus encompasses multiple formats including multiple-choice, fill-in-the-blank, and free-form problems, spanning core mathematical subjects such as geometry, algebra, function, and word problem-solving. Furthermore, we have also processed Multimath-300k-Chinese using the same pipeline as English Vision-Form Problem Synthesis. + +# 3.2. Architecture of VisionMath + +The architecture of the Vision-Math model is illustrated in Figure 4, which follows the structure that has been widely + +verified to work well in general MLLMs. This comprises three core components: a pre-trained visual encoder InternViT [9], which is good at encoding semantically rich visual features; a lightweight projector, i.e., a two-layer MLP [27], that transforms visual representations into the semantic space of LLMs; and an open-source LLM Qwen2.5-Math [55], which has advanced reasoning capabilities for text-form mathematical problems. + +# 3.3. Progressive Multimodal Reasoning Alignment + +To harness the synergistic integration of diverse datasets in Table 1 for building strong vision-form mathematical problem-solving model, we propose a three-stage multimodal reasoning alignment strategy that incrementally endows the model with capabilities ranging from foundational visual perception to advanced complex reasoning, as illustrated in Figure 4. The underlying design logic follows a hierarchical progression: first achieving accurate perception of problem statements, then advancing to comprehensive understanding of math figures, and finally enabling integrated text-figure reasoning. + +# 3.3.1. Mathematical OCR Capability Establishment + +During the first stage, we focus on equipping the model with the capabilities to accurately perceive fundamental problem-statement elements, including text, mathematical formulas, charts, and tables. To achieve this, we leverage the curated VisionMath-OCR dataset for training, updating all parameters in the visual encoder, projection layer, and language model. To prevent the LLM's catastrophic forgetting of reasoning capabilities during multimodal training, we also strategically interleave a small proportion (25%) of + +text-form mathematical dataset Numbia [4] into the training batches, ensuring balanced language skill retention. + +# 3.3.2. Math Figure Perception & Understanding + +In the second stage, we employ a dual-path enhancement strategy to improve the model's comprehension of mathematical figures. On the first path, leveraging figure caption data, we instruct the model to generate descriptive captions for input figures, thereby enhancing its ability to accurately perceive and understand figure primitives (such as points, lines, planes) and their spatial relationships, e.g., parallelism, intersections and containment. On the complementary path, utilizing the constructed educational video dataset, we task the model with predicting corresponding lecture content (i.e., refined ASR text) based on input video frames. The diverse mathematical figures present in these videos further strengthen the model's figure understanding capabilities. Additionally, the lecture content provided by educators contains rich knowledge, including in-depth figure analysis and geometric properties of figures, which further deepens the model's understanding and reasoning capabilities regarding visual content. Similarly, a small amount $(25\%)$ of Numbia is added at this stage to prevent the language model from degrading. + +# 3.3.3. Vision-Form Mathematical Instruction Tuning + +After completing the above two-phase training curriculum, the model has demonstrated robust foundational comprehension of vision-form mathematical problems. Building upon the acquired capabilities, the final stage orchestrates a seamless transition toward advanced mathematical reasoning through end-to-end supervised instruction tuning using the complete VisionMath-IT dataset. + +# 4. Evaluation Benchmarks + +To comprehensively evaluate and compare model performance on mathematical problems across multiple languages and subjects, our evaluation benchmarks are designed to include the following parts. + +# 4.1. Chinese Mathematical Proficiency Benchmark + +VisionMath-Bench is a Chinese evaluation benchmark comprising 100 Plane Geometry (Pla Geo), 90 Analytic Geometry (Ana Geo), 103 Solid Geometry (Sol Geo), 72 Algebra (Alg) and 135 Function (Func) problems, with test samples systematically collected according to standardized mathematics curricula spanning primary to senior secondary education in China. Test accuracies for different problem types are reported separately during evaluation. + +# 4.2. English Mathematical Proficiency Benchmarks + +English mathematical evaluation benchmarks are derived through the conversion of commonly used image-text-separation-form evaluation datasets, i.e., MathVista and + +MathVerse. Detailed dataset processing and statistics are shown in the Sec. B. of the supplementary materials. + +# 5. Experiment + +# 5.1. Implementation Details + +VisionMath employs InternViT-300M-448px as the visual encoder and Qwen2.5-Math-7B-Instruct as the base language model. The three-stage alignment training generally adopts the same training settings, except for using different datasets. The maximum number of image tiles for visual encoding and the maximum input sequence length for LLM are limited to 9 and 4096, respectively. The learning rate for the visual encoder is 2e-6, while that for the projection layer and LLM is 1e-5. Each stage undergoes 1 epoch of training with a batch size set at 384. + +# 5.2. Comparisons with SOTAs + +To make thoroughly performance comparisons of existing methods on vision-form mathematical problem-solving task, we extensively evaluate 8 open-source general-purpose MLLMs, 5 open-source math-specialized MLLMs, and 4 closed-source commercial general-purpose MLLMs, on the constructed evaluation benchmarks. General-purpose MLLMs are evaluated using VLMEvalKit [12], while math-specialized MLLMs are evaluated based on the original codebases. The detailed results are listed in Table 2, and the key findings are summarized below. + +VisionMath substantially outperforms open-source MLLMs. As can be seen from Table 2, leading open-source general-purpose MLLMs (e.g., Qwen2-VL) surpass math-specialized MLLMs (e.g., URSA) across multiple benchmarks, despite URSA exhibiting superior performance than Qwen2-VL on image-text-separation-form problems. This divergence stems from math-specialized models' excessive overfitting to figure-text separated problems, which induces catastrophic degradation of essential OCR for recognizing problem statements, making it difficult to generalize to the vision-form problem-solving tasks. However, both categories of current open-source models exhibit significant performance gaps compared to VisionMath across different languages and problem types. On the Chinese VisionMath-Bench, VisionMath outperforms the top general-purpose model Qwen2-VL by $16.0\%$ , $24.4\%$ , and $9.7\%$ on plane geometry, analytic geometry, and solid geometry problems, while exceeds the dedicated math-specialized model MultiMath by $44.0\%$ , $36.6\%$ , and $34.3\%$ . In Chinese algebra and function problem-solving, VisionMath achieves $18.0\%$ and $9.6\%$ advantages over Qwen2-VL and $48.6\%$ and $31.8\%$ leads against MultiMath. Similar superiority is also observed on English benchmarks Vision-MathVista and Vision-MathVerse, where VisionMath substantially outperforms + +
ModelParamVisionMath-BenchVision-MathVistaVision-MathVerse
Pla GeoAna GeoSol GeoAlgFuncGPSMWPTDTLVIVDVO
Closed-Source General-purpose MLLMs
Gemini-1.5-flash [49]-23.012.620.426.431.162.364.453.944.344.240.044.2
Gemini-1.5-pro [49]-33.023.231.140.336.373.666.761.756.252.852.253.3
GPT-4o mini [17]-22.013.718.433.327.453.453.247.338.734.034.534.4
GPT-4o [17]-31.028.437.844.431.059.667.253.947.145.142.142.4
Open-Source General-purpose MLLMs
Phi3V [1]4B4.03.23.94.25.918.821.016.417.015.515.414.0
Phi3.5V [1]4B0.00.04.94.21.519.737.618.815.914.314.117.5
DeepSeek-VL [31]7B2.01.12.91.44.418.326.98.59.810.410.212.8
MiniCPM-V2.6 [57]8B6.08.411.716.718.537.539.217.618.117.417.416.4
LLaVA-OneVision [22]7B11.08.48.711.121.529.855.924.521.621.121.116.7
InternVL2 [9]8B17.011.615.525.027.443.352.737.329.825.826.525.9
InternVL2.5 [7]8B18.015.816.525.026.745.252.236.729.127.027.724.7
Qwen2-VL [52]7B33.018.935.038.934.839.452.135.031.529.729.228.3
Open-Source Math-specilized MLLMs
G-LLaVA [14]7B0.01.11.01.40.023.12.117.015.514.112.25.3
Math-LLaVA [42]13B1.02.20.02.84.431.327.417.417.516.416.012.6
Math-PUMA-Qwen2 [63]7B1.01.11.02.84.425.030.818.318.919.218.515.5
Math-PUMA-DeepSeek2 [63]7B1.02.21.14.25.919.729.319.719.520.017.912.2
MultiMath [41]7B5.06.710.48.312.631.826.317.118.217.618.111.7
URSA [33]8B2.00.06.84.210.422.129.015.021.321.720.513.3
VisionMath (Ours)7B49.043.344.756.944.470.369.458.048.045.344.239.6
OCR tool + Open-Source MLLMs
OCR tool + InternVL2.5 [7]8B22.016.813.727.827.448.154.938.329.427.325.429.3
OCR tool + URSA [33]8B22.014.721.426.426.734.642.528.723.721.621.431.2
+ +Table 2. Comparisons of representative MLLMs on vision-form mathematical problem-solving benchmarks. The best results of open-sourced and closed-sourced MLLMs are marked in orange and blue, respectively. + +all open-source MLLMs across all problem types. This demonstrates that well-designed datasets combined with rationally progressive training strategies can methodically build the model's capabilities—from foundational math element perception to complex reasoning, thereby achieving excellent vision-form problem-solving proficiency. + +VisionMath is superior to the tool-integrated methods. A natural question is whether the end-to-end VisionMath model is better than pipelines that employ OCR tools to extract textual problem statements, and then feed visionform problem image and extracted statements into MLLMs for solving. The results of InternVL2.5 and URSA integrated with GPT-4o as OCR tools show that our method still achieves all-around superiority across benchmarks. This demonstrates that current MLLMs' performance bottleneck in solving vision-form math problems stems not merely from inaccurate mathematical OCR, but more from the need for enhancements in figure perception and reasoning. + +VisionMath achieves competitive results compared to closed-source MLLMs. Compared to advanced small-scale closed-source MLLMs, such as GPT-4o mini and Gemini-1.5-flash, VisionMath maintains comprehensive performance superiority across all evaluation benchmarks. When benchmarked against large-scale closed-source MLLMs, e.g., GPT-4o and Gemini-1.5-pro, VisionMath's benchmark performance varies. It leads GPT-4o and Geimi-1.5-pro on VisionMath-Bench, surpasses GPT + +4o while matching performance with Gemini-1.5-pro on Vision-MathVista, and outperforms GPT-4o's but trails behind Gemini-1.5-pro on Vision-MathVerse. + +# 5.3. Ablations + +In this section, we conduct ablation studies on dataset components and training strategies employed in our method, evaluating their contributions on Chinese VisionMath-Bench and English Vision-MathVerse (overall accuracy is reported due to space limitations) to validate the efficacy in enhancing problem-solving capabilities. Note that, Due to source constraints, 1M subsets of VisionMath-OCR and VisionMath-IT are sampled for training. + +Is a dedicated OCR capability enhancement stage necessary? Yes. As demonstrated in Table 3, omitting the foundational mathematical OCR capability acquisition during initial training leads to significant performance degradation across both evaluation benchmarks. This verifies that precise OCR, specifically optimized for text, symbol, formulas of problem statement, serves as a critical enabler for subsequent correct reasoning execution. + +Is the math figure comprehension strengthen stage beneficial? Yes. Table 4 presents results of different methods: (1) baseline without enhanced figure comprehension, (2) figure structural understanding enhancement using captions, (3) figure property understanding enhancement leveraging mathematical educational videos, and (4) a combined + +
VisionMath-BencMathVerse
Pla GeoAna GeoSol GeoAlgFunc
wo OCR-stage37.231.134.331.335.237.5
Ours39.033.236.934.337.841.1
+ +Table 3. Effectiveness of Mathematical OCR enhancement. + +
Fig. Cap.M.E. Vid.VisionMath-BencMathVerse
Pla GeoAna GeoSol GeoAlgFunc
37.030.332.231.132.436.7
✓38.031.333.033.633.338.2
✓38.533.834.631.934.039.8
✓✓39.033.236.934.337.841.1
+ +Table 4. Contribution of Figure Captioning ("Fig. Cap.") and Mathematical Education Video Learning ("M.E. Vid.") to enhancing the mathematical figure comprehension. + +
Eng. Pro.Chi. Pro.VisionMath-BencMathVerse
Pla GeoAna GeoSol GeoAlgFunc
✓24.023.825.230.135.239.2
✓38.232.035.234.937.326.1
✓✓39.033.236.934.337.841.1
+ +Table 5. Ablations on the effect of the collected vision-form English Problem ("Eng. Pro.") and Chinese Problem ("Chi. Pro.")-Solving datasets. + +strategy integrating the above enhancement. Compared to the (1) baseline, both (2) figure structural enhancement and (3) property understanding enhancement strategies improve performance, and (4) combining them yields further gains. This underscores the importance of understanding and utilizing the component structures and properties of math figures for correct reasoning, making the specialized figure understanding strengthen stage essential. + +Are collected English and Chinese vision-form problem-solving datasets effective and complementary? Yes. Table 5 demonstrates that instruction tuning with exclusively English (or Chinese) data achieves only acceptable performance on their language-specific benchmarks-MathVerse for English and VisionMath-Bench for Chinese. In contrast, simultaneous utilization of bilingual data yields performance improvements on both benchmarks compared to monolingual training. This demonstrates that mathematical reasoning capabilities transcend language barriers, thus aggregating multilingual datasets synergistically enhances reasoning to develop better problem-solving models. + +# 5.4. Qualitative Analysis + +Figure 5 presents visualization examples of visionform problem-solving from advanced open-source model Qwen2-VL, closed-source model GPT-4o, and our VisionMath. Analysis of the geometry problem in (a) reveals that both Qwen2-VL and GPT-4o exhibit misinterpretations of geometric properties: Qwen2-VL incorrectly states "... AB is parallel to CD, angle A is equal to angle BOC ..., while GPT-4o asserts "... AB is parallel to CD... angle C $(50^{\circ})$ and angle A are corresponding angles". Erroneous geo + +# Problem Image (a) + +As shown in the figure, AB parallel CD, CP intersects AB at O, AO = PO, if angle C = 50.0°, then the degree of angle A is () + +![](images/7ff5b9e3f37974e7652e190bcf3d466b7603e4c59dcac3f5a1511398416d752e.jpg) +Figure 5. Quantitative comparisons of Qwen2-VL, GPT-4o, and VisionMath on the vision-form mathematical problem-solving task. Red and green text in answers mark incorrect and correct math understanding and reasoning, respectively. + +Qwen2-VL Answer: Since AB is parallel to CD, angle A is equal to angle BOC (error 1). Since AO is equal to PO, angle A is equal to angle BOC, which is equal to angle C. (error 2) Therefore, the answer is C. (error 3) + +# GPT-4o Answer: + +1. Since AB is parallel to CD and CP intersects them, angle C $(50^{\circ})$ and angle A are corresponding angles (error 1) Corresponding angles are equal when two lines are parallel and are intersected by a transversal. + +2. Therefore, angle A is also $50^{\circ}$ . So, the degree of angle A is $50^{\circ}$ . + +The correct option is D. (error 2) + +# VisionMath Answer: + +Since AB is parallel to CD, we have angle $C =$ angle POB $= 50^{\circ}$ + +Since $AO = PO$ , we have angle $A =$ angle $P$ . + +Therefore, angle A = 1/2 angle POB = 2° Therefore, the answer is S |boxed{A5}. + +# Problem Image (b) + +What is the size of the shaded area under the curve? Round the answer to 2 decimal places + +![](images/f69f238d04e2e6b2e76428d904f568cd3b19d04f2be41d5756e8e619db8b1971.jpg) + +# Qwen2-VL Answer: + +area under the curve is half of the full circle's area (error 1): + +[ \backslash[\text{frac}\{1\}\{2\} \backslash \text{times} 9\backslash \text{pi} = \backslash \text{frac}\{9\backslash \text{pi}\} \backslash \{2\} \backslash ] + +# (4) App19.10:Answers 3.14)(2) + +To find the area under this curve from $\backslash (x = 0)$ to $\backslash (x = 6$ + +we need to calculate the area of the semicircle. (error 1) + +A.{_tset[semicircle]}={\frac[9}(2)}\backslash \pi \}approx{\frac[9](2)}\cdot{3.14159} + +[A_ \{text\semicircle\} ]\approxprox 14.13717 (error2) + +# VisionMath Answer: + +Step 1: The given function is $\backslash \backslash (f(x) = \backslash \backslash \operatorname{sqrt}(9 - (x - 3)^2)\backslash)$ . + +Step 2: The shaded area is a quarter of the circle. + +shaded area is $\backslash \backslash \backslash \backslash \text{frac}\{9\backslash \text{pi}\} \{4\} \backslash \backslash$ + +Therefore, the answer is $ \backslash \boxed{7.07} $. + +metric propositions that propagate through their reasoning chains, resulting in incorrect answers. In contrast, our VisionMath precisely interprets the spatial relationships with the correct proposition "... AB is parallel to CD, we have angle $C =$ angle $POB \ldots$ ", demonstrating rigorous geometric reasoning that ultimately arrives at the correct answer. This further substantiates the efficacy of the specialized enhanced figure comprehension training. In case (b), where the problem requires calculating the shaded area under the curve, both Qwen2-VL and GPT-4o fail to capture the constraint, erroneously computing the entire semicircle area beneath the curve. In contrast, VisionMath accurately parse the condition, correctly computing the quarter-circle area to obtain at the validated solution. More qualitative examples can be found in the Sec. D. of the supplementary materials. + +# 6. Conclusion + +In this work, we pioneer the exploration of a practical vision-form mathematical problem-solving task. We systematically deconstruct three core capabilities required for vision-form problem resolution, i.e., precise problem statement OCR, advanced comprehension of math figure structures and properties, and excellent multimodal reasoning capacity. Correspondingly, we design a three-phase progressive multimodal reasoning alignment strategy to incrementally develop these competencies. To comprehensively evaluate model performance, we construct multilingual benchmarks spanning various problem types. Experimental results demonstrate that our method significantly outperforms open-source general-purpose, math-specialized and tool-integrated MLLMs, while achieve comparable performance to closed-source models with significantly larger parameter scales. + +Acknowledgments This work is supported by the Natural Science Foundation of China (Grants No. 62192782, No. 62302501, No. 62403462), the Key Research and Development Program of Xinjiang Uyghur Autonomous Region, Grant No. 2023B03024. + +# References + +[1] Marah Abdin, Jyoti Aneja, Hany Awadalla, Ahmed Awadallah, Ammar Ahmad Awan, Nguyen Bach, Amit Bahree, Arash Bakhtiari, Jianmin Bao, Harkirat Behl, et al. Phi-3 technical report: A highly capable language model locally on your phone. arXiv preprint arXiv:2404.14219, 2024. 7 +[2] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 3 +[3] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond. arXiv preprint arXiv:2308.12966, 2023. 1, 3 +[4] Edward Beeching, Shengyi Costa Huang, Albert Jiang, Jia Li, Benjamin Lipkin, Zihan Qina, Kashif Rasul, Ziju Shen, Roman Soletskyi, and Lewis Tunstall. Numinamath 7b cot. https://huggingface.co/AI-MO/NuminaMath-7B-CoT, 2024.3,6 +[5] Siow-Wee Chang, Sameem Abdul-Kareem, Amir Feisal Merican, and Rosnah Binti Zain. Oral cancer prognosis based on clinicopathologic and genomic markers using a hybrid of feature selection and machine learning methods. BMC bioinformatics, 14:1-15, 2013. 3 +[6] Wenhu Chen, Hongmin Wang, Jianshu Chen, Yunkai Zhang, Hong Wang, Shiyang Li, Xiyou Zhou, and William Yang Wang. Tabfact: A large-scale dataset for table-based fact verification. arXiv preprint arXiv:1909.02164, 2019. 2, 3 +[7] Zhe Chen, Weiyun Wang, Yue Cao, Yangzhou Liu, Zhangwei Gao, Erfei Cui, Jinguo Zhu, Shenglong Ye, Hao Tian, Zhaoyang Liu, et al. Expanding performance boundaries of open-source multimodal models with model, data, and test-time scaling. arXiv preprint arXiv:2412.05271, 2024. 1, 3, 7 +[8] Zhe Chen, Weiyun Wang, Hao Tian, Shenglong Ye, Zhangwei Gao, Erfei Cui, Wenwen Tong, Kongzhi Hu, Jiapeng Luo, Zheng Ma, et al. How far are we to gpt-4v? closing the gap to commercial multimodal models with open-source suites. arXiv preprint arXiv:2404.16821, 2024. +[9] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24185–24198, 2024. 1, 3, 5, 7 +[10] Linger Deng, Yuliang Liu, Bohan Li, Dongliang Luo, Liang Wu, Chengquan Zhang, Pengyuan Lyu, Ziyang Zhang, Gang + +Zhang, Errui Ding, et al. R-cot: Reverse chain-of-thought problem generation for geometric reasoning in large multimodal models. arXiv preprint arXiv:2410.17885, 2024. 1, 3 +[11] Xiang Deng, Huan Sun, Alyssa Lees, You Wu, and Cong Yu. Turl: Table understanding through representation learning. ACM SIGMOD Record, 51(1):33-40, 2022. 3 +[12] Haodong Duan, Junming Yang, Yuxuan Qiao, Xinyu Fang, Lin Chen, Yuan Liu, Xiaoyi Dong, Yuhang Zang, Pan Zhang, Jiaqi Wang, et al. Vlmevalkit: An open-source toolkit for evaluating large multi-modality models. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 11198-11201, 2024. 6 +[13] Chaoyou Fu, Peixian Chen, Yunhang Shen, Yulei Qin, Mengdan Zhang, Xu Lin, Jinrui Yang, Xiawu Zheng, Ke Li, Xing Sun, et al. Mme: A comprehensive evaluation benchmark for multimodal large language models. arXiv preprint arXiv:2306.13394, 2023. 3 +[14] Jiahui Gao, Renjie Pi, Jipeng Zhang, Jiacheng Ye, Wanjun Zhong, Yufei Wang, Lanqing Hong, Jianhua Han, Hang Xu, Zhenguo Li, and Lingpeng Kong. G-llava: Solving geometric problem with multi-modal large language model, 2023. 1, 2, 3, 7 +[15] Adam W Harley, Alex Ufkes, and Konstantinos G Derpanis. Evaluation of deep convolutional nets for document image classification and retrieval. In International Conference on Document Analysis and Recognition (ICDAR). 3 +[16] Adam W Harley, Alex Ufkes, and Konstantinos G Derpanis. Evaluation of deep convolutional nets for document image classification and retrieval. In 2015 13th International Conference on Document Analysis and Recognition (ICDAR), pages 991-995. IEEE, 2015. 2 +[17] Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Weli-hinda, Alan Hayes, Alec Radford, et al. Gpt-4o system card. arXiv preprint arXiv:2410.21276, 2024. 7 +[18] Kushal Kafle, Brian Price, Scott Cohen, and Christopher Kanan. Dvqa: Understanding data visualizations via question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5648-5656, 2018. 3 +[19] Samira Ebrahimi Kahou, Vincent Michalski, Adam Atkinson, Ákos Kádár, Adam Trischler, and Yoshua Bengio. Figurea: An annotated figure dataset for visual reasoning. arXiv preprint arXiv:1710.07300, 2017. 3 +[20] Geewook Kim, Teakgyu Hong, Moonbin Yim, Jeong Yeon Nam, Jinyoung Park, Jinyeong Yim, Wonseok Hwang, Sang-doo Yun, Dongyoon Han, and Seunghyun Park. Ocr-free document understanding transformer. In European Conference on Computer Vision (ECCV), 2022. 3 +[21] Sunjun Kweon, Yeonsu Kwon, Seonhee Cho, Yohan Jo, and Edward Choi. Open-wikitable: Dataset for open domain question answering with complex reasoning over table. arXiv preprint arXiv:2305.07288, 2023. 2 +[22] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 1, 3, 7 + +[23] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Yi Liu, Zun Wang, Jilan Xu, Guo Chen, Ping Luo, et al. Mybench: A comprehensive multi-modal video understanding benchmark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22195-22206, 2024. 3 +[24] Zongzhao Li, Jiacheng Cen, Bing Su, Wenbing Huang, Tingyang Xu, Yu Rong, and Deli Zhao. Large language-geometry model: When llm meets equivariance. arXiv preprint arXiv:2502.11149, 2025. +[25] Zongzhao Li, Zongyang Ma, Mingze Li, Songyou Li, Yu Rong, Tingyang Xu, Ziqi Zhang, Deli Zhao, and Wenbing Huang. Star-r1: Spatial transformation reasoning by reinforcing multimodal llms. arXiv preprint arXiv:2505.15804, 2025.3 +[26] Fuxiao Liu, Xiaoyang Wang, Wenlin Yao, Jianshu Chen, Kaiqiang Song, Sangwoo Cho, Yaser Yacoob, and Dong Yu. Mmc: Advancing multimodal chart understanding with large-scale instruction tuning. arXiv preprint arXiv:2311.10774, 2023. 3 +[27] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning, 2023. 1, 3, 5 +[28] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning, 2023. +[29] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava next: Improved reasoning,OCR, and world knowledge, 2024.1,3 +[30] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, et al. Mmbench: Is your multi-modal model an all-around player? In European conference on computer vision, pages 216-233. Springer, 2024. 3 +[31] Haoyu Lu, Wen Liu, Bo Zhang, Bingxuan Wang, Kai Dong, Bo Liu, Jingxiang Sun, Tongzheng Ren, Zhuoshu Li, Hao Yang, et al. Deepseek-vl: towards real-world vision-language understanding. arXiv preprint arXiv:2403.05525, 2024.7 +[32] Pan Lu, Hritik Bansal, Tony Xia, Jiacheng Liu, Chunyuan Li, Hannaneh Hajishirzi, Hao Cheng, Kai-Wei Chang, Michel Galley, and Jianfeng Gao. Mathvista: Evaluating mathematical reasoning of foundation models in visual contexts. In International Conference on Learning Representations (ICLR), 2024. 1, 2 +[33] Ruilin Luo, Zhuofan Zheng, Yifan Wang, Yiyao Yu, Xinzhe Ni, Zicheng Lin, Jin Zeng, and Yujiu Yang. Ursa: Understanding and verifying chain-of-thought reasoning in multimodal mathematics. arXiv preprint arXiv:2501.04686, 2025. 3, 7 +[34] Karttikeya Mangalam, Raiymbek Akshulakov, and Jitendra Malik. Egoschema: A diagnostic benchmark for very long-form video language understanding. Advances in Neural Information Processing Systems, 36:46212-46244, 2023. 3 +[35] Ahmed Masry, Do Xuan Long, Jia Qing Tan, Shafiq Joty, and Enamul Hoque. Chartqa: A benchmark for question answering about charts with visual and logical reasoning. arXiv preprint arXiv:2203.10244, 2022. 3 + +[36] Minesh Mathew, Dimosthenis Karatzas, and CV Jawahar. Docvqa: A dataset for vqa on document images. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 2200-2209, 2021. 3 +[37] Minesh Mathew, Viraj Bagal, Ruben Tito, Dimosthenis Karatzas, Ernest Valveny, and C.V. Jawahar. Infographicvqa. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 1697-1706, 2022. 2 +[38] Minesh Mathew, Viraj Bagal, Ruben Tito, Dimosthenis Karatzas, Ernest Valveny, and CV Jawahar. Infographicvqa. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1697-1706, 2022. 3 +[39] Nitesh Methani, Pritha Ganguly, Mitesh M Khapra, and Pratyush Kumar. Plotqa: Reasoning over scientific plots. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1527-1536, 2020. 3 +[40] Panupong Pasupat and Percy Liang. Compositional semantic parsing on semi-structured tables. arXiv preprint arXiv:1508.00305, 2015. 3 +[41] Shuai Peng, Di Fu, Liangcai Gao, Xiuqin Zhong, Hongguang Fu, and Zhi Tang. Multimath: Bridging visual and mathematical reasoning for large language models. arXiv preprint arXiv:2409.00147, 2024. 3, 7 +[42] Wenhao Shi, Zhiqiang Hu, Yi Bin, Junhua Liu, Yang Yang, See-Kiong Ng, Lidong Bing, and Roy Ka-Wei Lee. Math-llava: Bootstrapping mathematical reasoning for multimodal large language models, 2024. 1, 3, 7 +[43] Tomasz Stanisławek, Filip Graliński, Anna Wróblewska, Dawid Lipinski, Agnieszka Kaliska, Paulina Rosalska, Bartosz Topolski, and Przemysław Biecek. Kleister: key information extraction datasets involving long documents with complex layouts. In International Conference on Document Analysis and Recognition, pages 564-579. Springer, 2021. 3 +[44] Quan Sun, Yufeng Cui, Xiaosong Zhang, Fan Zhang, Qiying Yu, Zhengxiong Luo, Yueze Wang, Yongming Rao, Jingjing Liu, Tiejun Huang, and Xinlong Wang. Generative multimodal models are in-context learners. 2023. 3 +[45] Quan Sun, Qiying Yu, Yufeng Cui, Fan Zhang, Xiaosong Zhang, Yueze Wang, Hongcheng Gao, Jingjing Liu, Tiejun Huang, and Xinlong Wang. Generative pretraining in multimodality. 2023. 3 +[46] Stacey Svetlichnaya. Deepform: Understand structured documents at scale. 2020. 3 +[47] Ryota Tanaka, Kyosuke Nishida, and Sen Yoshida. Visualmrc: Machine reading comprehension on document images. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 13878-13888, 2021. 2, 3 +[48] Gemini Team, Rohan Anil, Sebastian Borgeaud, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, Katie Millican, et al. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023. 3 +[49] Gemini Team, Petko Georgiev, Ving Ian Lei, Ryan Burnell, Libin Bai, Anmol Gulati, Garrett Tanzer, Damien Vincent, Zhufeng Pan, Shibo Wang, et al. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context. arXiv preprint arXiv:2403.05530, 2024. 3, 7 + +[50] Qwen Team. Qwen2.5-v1, 2025. 1, 3 +[51] Michal Turski, Tomasz Stanisławek, Karol Kaczmarek, Paweł Dyda, and Filip Graliński. Ccpdf: Building a high quality corpus for visually rich documents from web crawl data, 2023. 2, 3 +[52] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Yang Fan, Kai Dang, Mengfei Du, Xuancheng Ren, Rui Men, Dayiheng Liu, Chang Zhou, Jingren Zhou, and Junyang Lin. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 1, 3, 7 +[53] Xinlong Wang, Xiaosong Zhang, Zhengxiong Luo, Quan Sun, Yufeng Cui, Jinsheng Wang, Fan Zhang, Yueze Wang, Zhen Li, Qiying Yu, et al. Emu3: Next-token prediction is all you need. arXiv preprint arXiv:2409.18869, 2024. 3 +[54] Haoran Wei, Lingyu Kong, Jinyue Chen, Liang Zhao, Zheng Ge, Jinrong Yang, Jianjian Sun, Chunrui Han, and Xiangyu Zhang. Vary: Scaling up the vision vocabulary for large vision-language models. arXiv preprint arXiv:2312.06109, 2023. 3 +[55] An Yang, Beichen Zhang, Binyuan Hui, Bofei Gao, Bowen Yu, Chengpeng Li, Dayiheng Liu, Jianhong Tu, Jingren Zhou, Junyang Lin, Keming Lu, Mingfeng Xue, Runji Lin, Tianyu Liu, Xingzhang Ren, and Zhenru Zhang. Qwen2.5-math technical report: Toward mathematical expert model via self-improvement. arXiv preprint arXiv:2409.12122, 2024.5 +[56] Zhen Yang, Jinhao Chen, Zhengxiao Du, Wenmeng Yu, Wei-han Wang, Wenyi Hong, Zhihuan Jiang, Bin Xu, Yuxiao Dong, and Jie Tang. Mathglm- vision: Solving mathematical problems with multi-modal large language model. arXiv preprint arXiv:2409.13729, 2024. 3 +[57] Yuan Yao, Tianyu Yu, Ao Zhang, Chongyi Wang, Junbo Cui, Hongji Zhu, Tianchi Cai, Haoyu Li, Weilin Zhao, Zhihui He, et al. Minicpm-v: A gpt-4v level mllm on your phone. arXiv preprint arXiv:2408.01800, 2024. 3, 7 +[58] Xiang Yue, Yuansheng Ni, Kai Zhang, Tianyu Zheng, Ruoqi Liu, Ge Zhang, Samuel Stevens, Dongfu Jiang, Weiming Ren, Yuxuan Sun, et al. Mmmu: A massive multi-discipline multimodal understanding and reasoning benchmark for expert agi. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9556-9567, 2024. 3 +[59] Renrui Zhang, Dongzhi Jiang, Yichi Zhang, Haokun Lin, Ziyu Guo, Pengshuo Qiu, Aojun Zhou, Pan Lu, Kai-Wei Chang, Yu Qiao, et al. Mathverse: Does your multi-modal llm truly see the diagrams in visual math problems? In European Conference on Computer Vision, pages 169–186. Springer, 2024. 2 +[60] Renrui Zhang, Xinyu Wei, Dongzhi Jiang, Yichi Zhang, Ziyu Guo, Chengzhuo Tong, Jiaming Liu, Aojun Zhou, Bin Wei, Shanghang Zhang, Peng Gao, and Hongsheng Li. Mavis: Mathematical visual instruction tuning, 2024. 2, 3 +[61] Wenqi Zhang, Hang Zhang, Xin Li, Jiashuo Sun, Yongliang Shen, Weiming Lu, Deli Zhao, Yueting Zhuang, and Lidong Bing. 2.5 years in class: A multimodal textbook for vision- + +language pretraining. arXiv preprint arXiv:2501.00958, 2025. 2, 3, 4 +[62] Xu Zhong, Elaheh ShafieiBavani, and Antonio Jimeno Yepes. Image-based table recognition: data, model, and evaluation. In European conference on computer vision, pages 564-580. Springer, 2020. 3 +[63] Wenwen Zhuang, Xin Huang, Xiantao Zhang, and Jin Zeng. Math-puma: Progressive upward multimodal alignment to enhance mathematical reasoning. arXiv preprint arXiv:2408.08640, 2024. 3, 7 \ No newline at end of file diff --git a/ICCV/2025/VisionMath_ Vision-Form Mathematical Problem-Solving/images.zip b/ICCV/2025/VisionMath_ Vision-Form Mathematical Problem-Solving/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..c130d4ebbbed1896e039a7fa6aa7911835600e45 --- /dev/null +++ b/ICCV/2025/VisionMath_ Vision-Form Mathematical Problem-Solving/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:93f9eb9e2e0bbd55d38f6df8235c58f77b204d828bd8947a48edea657ac4482c +size 573581 diff --git a/ICCV/2025/VisionMath_ Vision-Form Mathematical Problem-Solving/layout.json b/ICCV/2025/VisionMath_ Vision-Form Mathematical Problem-Solving/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..800ee119b995a2e887183b644d9e6cd75672484c --- /dev/null +++ b/ICCV/2025/VisionMath_ Vision-Form Mathematical Problem-Solving/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f4e76b016b4353802ac7594b99eb2d5d70d6dd6018a72a7bfddb6607aeaf7e16 +size 392873 diff --git a/ICCV/2025/VistaDream_ Sampling multiview consistent images for single-view scene reconstruction/7ae1ef6e-97d6-4f46-866a-e139e1269a78_content_list.json b/ICCV/2025/VistaDream_ Sampling multiview consistent images for single-view scene reconstruction/7ae1ef6e-97d6-4f46-866a-e139e1269a78_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..765c8590ca6a47740979dde8fe006396bb1b0192 --- /dev/null +++ b/ICCV/2025/VistaDream_ Sampling multiview consistent images for single-view scene reconstruction/7ae1ef6e-97d6-4f46-866a-e139e1269a78_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:16f8802e5c9051814bae4558bfb2ce5bc57d02ee01ec82d5b32c43418928390a +size 83186 diff --git a/ICCV/2025/VistaDream_ Sampling multiview consistent images for single-view scene reconstruction/7ae1ef6e-97d6-4f46-866a-e139e1269a78_model.json b/ICCV/2025/VistaDream_ Sampling multiview consistent images for single-view scene reconstruction/7ae1ef6e-97d6-4f46-866a-e139e1269a78_model.json new file mode 100644 index 0000000000000000000000000000000000000000..52055baa0f37d724efef87bd3b96549d233b7a35 --- /dev/null +++ b/ICCV/2025/VistaDream_ Sampling multiview consistent images for single-view scene reconstruction/7ae1ef6e-97d6-4f46-866a-e139e1269a78_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7c635d8d233b5b9cfccc05b5da4a1225f5cdaa8a943b4ff81336455425623f40 +size 104498 diff --git a/ICCV/2025/VistaDream_ Sampling multiview consistent images for single-view scene reconstruction/7ae1ef6e-97d6-4f46-866a-e139e1269a78_origin.pdf b/ICCV/2025/VistaDream_ Sampling multiview consistent images for single-view scene reconstruction/7ae1ef6e-97d6-4f46-866a-e139e1269a78_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..de6438c49b650c96a44eff80ebfcd00e94d1986f --- /dev/null +++ b/ICCV/2025/VistaDream_ Sampling multiview consistent images for single-view scene reconstruction/7ae1ef6e-97d6-4f46-866a-e139e1269a78_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:da1a9c482f2c9950cd04c1ce1fefd58de17659b70e5b75b20cbda3c9b4ef566d +size 8471794 diff --git a/ICCV/2025/VistaDream_ Sampling multiview consistent images for single-view scene reconstruction/full.md b/ICCV/2025/VistaDream_ Sampling multiview consistent images for single-view scene reconstruction/full.md new file mode 100644 index 0000000000000000000000000000000000000000..1459477bd01c893ccfb2c987cf091c43c3471ad0 --- /dev/null +++ b/ICCV/2025/VistaDream_ Sampling multiview consistent images for single-view scene reconstruction/full.md @@ -0,0 +1,366 @@ +# VistaDream: Sampling multiview consistent images for single-view scene reconstruction + +Haiping Wang $^{1}$ Yuan Liu $^{2,3,\dagger}$ Ziwei Liu $^{3}$ Wenping Wang $^{4}$ Zhen Dong $^{1,\dagger}$ Bisheng Yang $^{1}$ $^{1}$ Wuhan University $^{2}$ Hong Kong University of Science and Technology + $^{3}$ Nanyang Technological University $^{4}$ Texas A&M University +{hpwang, dongzhenwhu, bshyang}@whu.edu.cn +yuanly@ust.hk ziwei.liu@ntu.edu.sg wenping@tamu.edu + +![](images/05f4c3fcdff4741e91311243b2a7b5da403840fc8e5ba56ccb20286ecd2e3bd3.jpg) +Figure 1. Overview. (Top) Given a single-view image of a scene, VistaDream reconstructs a 3D scene represented by 3D Gaussian Splitting (3DGS) [22] for novel view synthesis. (Bottom) The proposed Multiview Consistency Sampling (MCS) significantly improves scene quality and achieves better results compared to the commonly used Score Distillation Sampling (SDS) [44]. + +# Abstract + +In this paper, we propose VistaDream, a novel framework to reconstruct a 3D scene from a single-view image. Recent diffusion models enable generating high-quality novel-view images from a single-view input image. Most existing methods only concentrate on building the consistency between the input image and the generated images while losing the consistency between the generated images. VistaDream addresses this problem by a two-stage pipeline. In the first stage, VistaDream builds a global coarse 3D scaffold by zooming out a little step with inpainted boundaries and an estimated depth map. Then, on this global scaffold, we use iterative diffusion-based RGB-D inpaint + +ing to generate novel-view images to inpaint the holes of the scaffold. In the second stage, we further enhance the consistency between the generated novel-view images by a novel training-free Multiview Consistency Sampling (MCS) that introduces multi-view consistency constraints in the reverse sampling process of diffusion models. Experimental results demonstrate that without training or fine-tuning existing diffusion models, VistaDream achieves high-quality scene reconstruction and novel view synthesis using a single-view image and outperforms baseline methods by a large margin. The code, videos, and demos are available at https://vistadream-project-page.github.io/. + +# 1. Introduction + +Reconstructing 3D scenes is a critical task in computer vision, robotics, and graphics. Traditionally, this has required multiple images from different viewpoints [49] or specialized hardware like RGBD scanners [39, 61] to capture both geometry and appearance. However, in real-world applications like AR/VR and robotics, we often only have access to a single-view image. Single-view 3D scene reconstruction is a highly challenging, ill-posed problem. Recent advances in diffusion models [16] demonstrate strong capabilities in generating realistic images, offering promise for creating novel views from single images to aid in 3D reconstruction. The key challenge, however, is ensuring consistency across the generated views to produce coherent 3D scenes. + +To tackle this challenge, prior works [38, 56, 66, 67, 69] have primarily focused on enforcing consistency between the input single-view image and the generated novel views but struggle with ensuring consistency between the generated views themselves. Early approaches [56, 69] introduced techniques like epipolar line attention to aligning the input and generated views. More recent methods [50, 51, 66, 67] adopt a warp-and-inpaint approach, where they estimate depth, warp the image to a new viewpoint, inpaint it, and repeat this process iteratively to reconstruct the 3D scene. While promising, these methods still suffer from inconsistencies in the depth maps of the novel views, as monocular depth estimators [21, 62] fail to maintain a consistent scale across viewpoints. Multiview diffusion models [15, 38] attempt to address this by generating all novel views simultaneously for improved consistency but are limited by the number of views they can produce and demand extensive datasets and computational resources for training. In short, achieving multiview consistency in the generated images of a scene from a single-view input remains an unresolved and significant challenge. + +In this paper, we propose VistaDream, a framework for 3D scene reconstruction from single-view images without the requirement of fine-tuning diffusion models. Given the single-view images as inputs, VistaDream reconstructs the scene of the given single-view image as a set of 3D Gaussian kernels [22], which enables us to render arbitrary novel-view images in the scene by the splatting technique. VistaDream is built upon the existing image diffusion models [37, 73] and maintains the multiview consistency of generated images by a two-stage pipeline as follows. + +In the first stage, VistaDream begins by constructing a coarse 3D scaffold, achieved by zooming out the camera from the input view while applying inpainting and depth estimation. This process establishes a rough yet valuable global geometry constraint for the 3D reconstruction. By zooming out, we generate expanded views and utilize the Fooocus model [73] to inpaint the black borders created by the zooming out. We further enhance this step by leveraging + +detailed text descriptions provided by a Visual-Language Model [33], which helps produce high-quality, well-defined zoomed-out images. Next, we estimate a depth map on the zoomed-out image, providing a coarse 3D geometry of the entire scene to serve as a constraint for subsequent generation. Building on this global scaffold, we then apply a warp-and-inpaint approach [50, 51, 66, 67] to fill gaps in the 3D scene. This step produces a rough 3D reconstruction with some inconsistencies among the generated views. + +In the second stage, we introduce a novel Multiview Consistency Sampling (MCS) algorithm to resample multiview-consistent images from a pre-trained diffusion model [37] to refine the reconstructed 3D scene. In contrast to SDS [43] which only considers one view in the regeneration process for refinement, our MCS simultaneously utilizes multiple rendered images to explicitly enforce the consistency among all images, which greatly improves the ability to model fine details, avoids averaging issues, and leads to stable convergence. This is formulated as a constrained sampling process, where multiview consistency is enforced during the reverse diffusion process. We begin by rendering multiple views from the current 3D scene and introducing noise to these renderings. The MCS algorithm then denoises these images to regenerate multiview-consistent outputs. At each denoising step, we utilize the predicted $x_0$ to train a new 3DGS representation, replacing the predicted $x_0$ with a corrected $\hat{x}_0$ rendered from this 3DGS representation to enhance consistency for denoising. Our results demonstrate that this consistency rectification significantly improves the multiview consistency of the generated images, leading to higher-quality 3D scene reconstructions. + +We conduct experiments on single-view images in both indoor and outdoor datasets of diverse styles. The results demonstrate that VistaDream, requiring no training or finetuning, surpasses state-of-the-art scene generation methods both qualitatively and quantitatively. Comprehensive ablation studies also validate the effectiveness of our global scaffold initialization and Multiview Consistency Sampling in enhancing scene consistency and quality. + +# 2. Related work + +Diffusion Models. Diffusion models have recently shown remarkable generation capabilities [16, 52]. These models gradually corrupt data into noise via a predefined Markov chain in the forward process and learn to reverse this process by progressive denoising, mapping noise distributions to data distributions. This enables effective novel data sampling or generation. Models such as Stable Diffusion [11, 12, 41, 42] leverage this framework for remarkable text-based image generation by scaling the model size and training data. Recent advancements fine-tune these models for tasks like depth estimation [14, 21] and image inpainting [50, 60, 73], achieving impressive perfor + +![](images/c5f1bb3667e901a794c197b8d656d6c7a6bb35a8dc8744585de3f8cc86f50e5b.jpg) +(b) Construct 3D Global Scaffold +Figure 2. Coarse Gaussian field reconstruction. (a) Given an image, VistaDream initializes a 3D global scaffold by enlarging FoV and inpainting, then iteratively inpaints the warped RGB-D images to complete a coarse Gaussian field. (b) Existing move-based scaffold construction [17, 51] struggle to accurately connect the multiview extending regions, yielding distorted scaffold causing the restriction failure. The proposed zoom-out strategy is simple yet effective in constructing a accurate and consistent global scaffold. + +mances. Additionally, methods like Latent Consistency Models [37, 48, 64] distill pre-trained diffusion models in the latent space to enable faster one- or few-step inference. Our method is built upon existing text-to-image diffusion models [37, 73]. Some existing works [3, 23, 24] also pose constraints on the predicted $x_0$ in diffusion models for editing while our work focuses on scene generation. Concurrent works Cycle3D [54] and 3D-Adaptor [8] also build a sampling and reconstruction pipeline in diffusion models, which target for single-view object reconstruction and require additional diffusion model training. + +Large Vision-Language Model. In contrast to task-specific visual models, such as those used for segmentation [20] or image captioning [27, 28, 45], Large Vision-Language Models (VLMs) [1, 33] align visual embeddings with the latent space of Large Language Models (LLMs) [40, 55]. By leveraging the strong knowledge priors of LLMs, VLMs enable advanced image understanding [71], supporting tasks like visual question answering, image descriptions, and task decomposition. In our pipeline, we adopt the LLaVA [33] to generate captions. + +Single-view reconstruction. Single-view reconstruction aims to generate a 3D distribution from a single image for novel view rendering [35, 51]. Some approaches learn to convert monocular photos into 3D objects via end-to-end training, producing 3D representations like multiview consistent images [29, 35, 36, 57], meshes [34], neural fields [53, 76], and tri-planes [18]. Alternatively, DreamFusion [43] proposes Score Distillation Sampling (SDS) that iteratively optimizes 3D scenes through single-step sampling from noisy images rendered at various viewpoints. SDS or its variants [30, 31, 58, 74] achieve lightweight 3D object generation from pre-trained 2D diffusion models. + +However, randomness in SDS denoise introduces inconsistencies among iterations, yielding averaged results [74]. + +Training end-to-end reconstruction models remains challenging for scene-level distributions due to their complexity, yielding limited fields of view, diversity, and multiview inconsistency [6, 15, 47, 75]. Recent methods use inpainting models [10, 25, 51, 72] or video generation models [59, 63, 65, 70] to iteratively complete missing regions to expand the scene scope while suffering from instability, noise, and distortion. SDS is then introduced for optimization at the cost of blurriness [51]. VistaDream addresses these limitations by leveraging large Vision-Language Models to enhance the reliability and diversity of scene expansion. Additionally, we propose Multi-view Consistent Sampling (MCS) to generate high-quality, consistent multi-view images directly from pre-trained diffusion models, significantly improving scene quality. + +# 3. Method + +Given a single-view image of a scene, the target of VistaDream is to reconstruct the 3D Gaussian field of the scene and enable novel view synthesis in the scene. VistaDream achieves this with a two-stage pipeline. The first stage builds a coarse 3D Gaussian field while the second stage refines the 3D Gaussian field with Multiview Consistency Sampling of a diffusion model. + +# 3.1. Coarse Gaussian field reconstruction + +In this stage, our target is to build a coarse Gaussian field from the single-view input image. First, similar to existing methods [51], we construct a global 3D scaffold to constrain the overall appearances and geometry for most regions. Then we apply the warp-and-inpaint scheme built on + +![](images/de1046efa2e8d4e465816536c67cbcc02cd7a2c8fcb6bbdd0753d4ffaac782e8.jpg) + +![](images/9038c4f9a687e94f794f8d2197f8f0b2854fcef17d0904f8888357d6853188fe.jpg) +Figure 3. Multiview Consistency Sampling for Scene Refinement. (a) We optimize the Gaussian field using high-quality, multi-view images regenerated by diffusion models. (b) The key component is the MCS algorithm, which enforces consistency during multi-view optimization. (c) A real case demonstrates that the MCS optimization process can progressively enhance the quality (red box) and consistency (yellow box) of multi-view images. Utilizing multiview images from MCS to optimize the Gaussian field can significantly enhance its quality. + +the global 3D scaffold to generate novel-view images and depth maps. Finally, we reconstruct a 3D Gaussian field from generated images and depth maps. + +Motivation. Direct warp-and-inpaint for scene generation struggles to maintain loop consistency when revisiting generated regions [10]. Recent works [17, 51] propose to first roughly but largely expand the scene to create a scaffold, serving as a global constraint preventing the warp-and-inpaint from deviating significantly. This is done by moving the camera within a small neighbor but with large rotations, extending the scene scope by outpainting in 3D from multiple viewpoints, called moving-based scaffold [46, 51]. However, aligning multiview outpainted regions is challenging as in Fig. 2 (b), causing geometric distortions and reconstruction failure. To enhance scaffold consistency, we propose a simple yet effective zoom-out strategy. + +Building scaffold by zoom-out and VLM. We first zoom out from the input view by changing the Field of View (FoV) of the camera as shown in Fig. 2 (b), which leaves a quite large regular region on all directions for the inpainting model Fooocus [73] to inpaint once, ensuring coherent new contents on these large regions as shown in Fig. 2 (a/b). After generating the zoomed-out images, we apply a depth estimator [5, 19] to estimate the metric depth values for all pixels on the zoomed-out image. The zoomed-out image along with the estimated depth map ensures a consistent and accurate 3D scaffold to constrain the 3D scene. However, consistent scaffold by zoom-out is non-trivial. The inpainting models usually require a text prompt to generate new content, which is vital especially for a large missing region + +from zoom-out. The short and simple descriptions from BLIP [28] leads to incorrect and distorted scaffold while the detailed descriptions from LLaVA [33] will greatly improve the inpainting quality, where examples are shown in the supplementary material. + +Warp-and-inpaint. Then, based on the 3D scaffold, we apply warp-and-inpaint scheme to generate a set of novel-view images. For a new viewpoint, we warp the global zoomed-out image with its depth map to this viewpoint and then fill the empty regions with Fooocus [73], guided by the LLaVA prompt. After that, we estimate the depth map on this new viewpoint with a depth estimator [5, 14] and align it with the scaffold. The estimated depth map is further optimized to align with the global depth map [10]. We repeat this process for several new viewpoints to generate a set of RGBD images. Finally, we train a 3D Gaussian field using these generated RGBD images and the global 3D scaffold as our coarse 3D Gaussian field. + +# 3.2. Multiview Consistency Sampling Refinement + +In the above stage, we enforce the consistency between the 3D scaffold and novel-view images generated by warp-and-inpaint scheme. However, there still remains inconsistency among the generated images causing the resulting 3D Gaussian to be noisy. In Stage II, we will improve the reconstructed 3D Gaussian field by a Multiview Consistency Sampling algorithm. Specifically, as shown in Fig. 3 (a), we first render $N$ images on $N$ predefined viewpoints from the coarse Gaussian field, denoted by $\mathbf{x}^{(1:N)} \coloneqq \{\mathbf{x}^{(n)}| = 1,\dots,N\}$ . Then, we use the diffusion models to refine all + +these renderings into new images $\hat{\mathbf{x}}^{(1:N)}$ which have enhanced quality and multiview consistency. Finally, we refine the 3D Gaussian field by training on these refined renderings. The key problem here is how to refine these renderings using a diffusion model while maintaining multiview consistency because simply regenerating these images leads to inconsistent results. We address this problem with a Multiview Consistency Sampling (MCS) algorithm. + +# 3.2.1. Multiview Consistency Sampling + +Given the renderings $\mathbf{x}^{(1:N)}$ , we first follow the forward process of the diffusion model to add noises to them to get a set of noisy renderings $\hat{\mathbf{x}}_T^{(1:N)}$ where $T$ is a predefined time step. Then, to regenerate these images, we sample the Markov Chain of reverse process $\prod_{n}^{N}\prod_{t}^{T}p_{\theta}(\hat{\mathbf{x}}_{t - 1}^{(n)}|\hat{\mathbf{x}}_{t}^{(n)})$ while enforcing the multiview consistency between all the images $\hat{\mathbf{x}}_0^{(1:N)}$ . To achieve this, we enforce the multiview consistency on every timestep by training a 3D Gaussian field to rectify the denoising direction. + +$\mathbf{x}_0$ -prediction of DDPM. Specifically, recall that on one denoising step $t$ of the DDPM [16] model, the predicted noise $\epsilon_{\theta}(\hat{\mathbf{x}}_t^{(n)}, t)$ gives an estimation of the final denoising result $\hat{\mathbf{x}}_0^{(n)}$ by + +$$ +\hat {\mu} \left(\mathbf {x} _ {t} ^ {(n)}, t\right) = \frac {1}{\bar {\alpha} _ {t}} \left(\hat {\mathbf {x}} _ {t} ^ {(n)} - \bar {\beta} _ {t} \epsilon_ {\theta} \left(\hat {\mathbf {x}} _ {t} ^ {(n)}, t\right)\right), \tag {1} +$$ + +where $\epsilon_{\theta}(\hat{\mathbf{x}}_t^{(n)},t)$ denotes the predicted noises on the $n$ -th rendered view on the timestep $t,\hat{\mu}_t^{(n)}\coloneqq \hat{\mu} (\mathbf{x}_t^{(n)},t)$ is the estimated $\hat{\mathbf{x}}_0^{(n)}$ from the current noisy version $\hat{\mathbf{x}}_t^{(n)}$ $\bar{\alpha}_{t}$ and $\bar{\beta}_t$ are predefined constants. Thus, one denoising step can also be written in the form of $\hat{\mu}_t^{(n)}$ instead of $\epsilon_{\theta}(\hat{\mathbf{x}}_t^{(n)},t)$ by + +$$ +\hat {\mathbf {x}} _ {t - 1} ^ {(n)} = s _ {t} \hat {\mathbf {x}} _ {t} ^ {(n)} + d _ {t} \hat {\mu} _ {t} ^ {(n)} + \sigma_ {t} \epsilon , \epsilon \sim \mathcal {N} (0, I), \tag {2} +$$ + +where $s_t, d_t, \sigma_t$ all are predefined constants and $\epsilon$ is a noise sampled from the standard Gaussian distribution. Eq. (2) indicates that the denoising direction is determined by the $\hat{\mu}_t^{(n)}$ . Thus, the key idea of MCS is to rectify $\hat{\mu}_t^{(1:N)} \coloneqq \{\hat{\mu}_t^{(n)}\}$ to new $\tilde{\mu}_t^{(1:N)}$ and then use the rectified $\tilde{\mu}_t^{(1:N)}$ for denoising in Eq. (2). + +Enforcing consistency. Since $\hat{\mu}_t^{(1:N)}$ is an estimation of the noisy free $\hat{\mathbf{x}}_0^{(1:N)}$ , we enforce the multiview consistency between them by training a 3D Gaussian field on the noisy-free $\hat{\mu}_t^{(1:N)}$ . Then, we render the images on this temporal 3D Gaussian field, which are denoted by $\bar{\mu}_t^{(n)}$ . Then, the rectified $\tilde{\mu}_t^{(1:N)}$ are computed by + +$$ +\tilde {\mu} _ {t} ^ {(n)} = w _ {t} \gamma_ {t} ^ {(n)} \bar {\mu} _ {t} ^ {(n)} + (1 - w _ {t}) \hat {\mu} _ {t} ^ {(n)}, \tag {3} +$$ + +where $\gamma_{t}$ stands for $std(\hat{\mu}_t) / std(\bar{\mu}_t)$ to avoid overexposure [32], $w_{t}$ is a predefined weight to balance between the denoising results $\hat{\mu}_{t}^{(1:N)}$ and the rendered multiview consistent $\bar{\mu}_t^{(1:N)}$ . $w_{t}$ determines how much multiview + +
MethodTrain(a) Geo. TSED(b) LVM-IQA(c) Fid. PSNR(d) Time T.(min) / GPU
LLaVAInterVLQwen
Realm [51]✓0.7060.5130.6950.60115.2~150 / 1×A10 (24G)
CAT3D [15]✓0.8220.6840.8060.870-~1 / 16×A100 (80G)
MVGM [6]✓0.8010.6550.7740.84924.5~4 / 1×4090 (24G)
SEVA [75]✓0.7890.6320.7660.77923.9~5 / 1×4090 (24G)
Ours0.8460.7180.7980.87426.6~7 / 1×4090 (24G)
+ +Table 1. Quantitative comparisons on (a) Geometric accuracy, (b) Rendering Image Quality Assessment by LVMs, (c) Fidelity to the input image, and (d) Time consumption. "Train" means requiring training on the single-view reconstruction task. + +consistency is imposed on the denoising process. Learning a 3D Gaussian field forces the multiview consistency to get $\bar{\mu}_t^{(1:N)}$ but may oversmooth some regions. Directly utilizing the denoising directions from $\hat{\mu}_t^{(1:N)}$ produces images with more details but less multiview consistency. Therefore, we set the $w_t$ to balance the denoising directions between $\bar{\mu}_t^{(1:N)}$ and $\hat{\mu}_t^{(1:N)}$ . We repeat this process for every denoising step to get the refined renderings $\hat{\mathbf{x}}_0^{(1:N)}$ . Then, these refined renderings are used for the refinement of the coarse 3D Gaussian field to improve the rendering quality as shown in Fig. 3 (c). + +Discussion. Previous methods [50, 51, 67] mainly focus on enforcing the consistency between the input image and a single generated image in a sequential manner, which struggles to maintain consistency on a long trajectory. Our MCS allows the simultaneous generation of multiple novel-view images and enforcement of consistency among all generated images, which does not suffer from consistency lost in sequential modeling. Thus, MCS generates more high-quality and consistent images than baseline methods and improves the quality of single-view 3D reconstruction. An alternative way is to adopt the SDS-based refinement method [43, 51]. However, the SDS method only considers one rendered view for one denoising step in the optimization, which often tends to average the generated contents. In comparison, our MCS simultaneously considers multiple rendered views to maintain multiview consistency and thus improves the consistency and generation quality. + +# 4. Experiments + +# 4.1. Experimental protocol + +Datasets. We adopt 32 single-view images from baseline methods [51, 65], copyright-free online photos, and generated models [12] for evaluation. The images cover both indoor and outdoor, real and simulated scenes. Among these, 11 images from RealmDreamer [51] were used in the quantitative comparisons. Evaluations on more data are provided in the Supplementary Material. + +Baselines. We use RealmDreamer [51], GenWarp [50], CAT3D [15], MVGenMaster [6], and Stable Virtual Camera (SEVA) [75] as the baseline methods. Given an input image, RealmDreamer [51] trains inpainting networks for extending the scene, which is optimized a Score Distilla + +![](images/ea3cdae7dd7853f0cd51ce14a21e200d7b7f240163234941ba5387f3357e59fd.jpg) +(a) Input + +![](images/fdb30664b2c13274b7be615a9175e6fef581d142cb954b83110ce2a61f0a49c2.jpg) +(b) RealmDreamer + +![](images/f5f01b7ff2cae2d430c485d9ffc5a70a2a906725424d8de5852d992692125551.jpg) + +![](images/f9a5370af429f8578d34f50cee6d496e985c203de93902b54293fea34f8be111.jpg) + +![](images/7751b629b29b3caf566165c7abd787199ce74dbb25c9d501574e90effa35b4fe.jpg) + +![](images/1c4f00584ad782b1d0af54338aaf7fec25146dc22ffd846aed8564984a8951aa.jpg) +(c) Ours + +![](images/0caff5b86fb1964fb389a29679fbc044c4b2a7c61f8fe172cd7160b20c414a11.jpg) + +![](images/0334656bb01e447b4896b64cdd1069150f0618dcf9b7840b2c6ab98b7ec57528.jpg) + +![](images/245c9585f0a2657fe37b47d4bcb58a556f1693046add879fdbecd975ba75737b.jpg) +(b) GenWarp + +![](images/219fac4c6d490f2219fead0ecb9c803d4b0ead46120276d51426e083a16d0d13.jpg) + +![](images/6658727ccc7b78d9abb0e5c6d8a3685603498f6752df0bfb47fef45c2e1f707a.jpg) + +![](images/3ff23ddafaac2f78e3a7d63ed82f15a00d524359756229f88a28d07ef7f5a9d4.jpg) + +![](images/c674ef0cea9c69b27b60a95a42eb783a7c7e2b6a462cda762f2d12301da6e860.jpg) + +![](images/28cf9fa7468a46fc896f05e35a5c084cbc0cee726dea9ec6b46918c709654cd5.jpg) + +![](images/f1d4137f79a81209581966dd17151b9598e78f1622d202b35e45f06206f6a589.jpg) +(a) Input +(a) Input + +![](images/7462e9fd28f4fc0861cbdc4a95971bf06eabf5f6951b07273500ffb16481ed52.jpg) + +![](images/a05ca89b4d9d6711f9e05f7e7f7dbb8eaf5d8f315a7687bc3693a05b9f7da41f.jpg) +(b) CAT3D + +![](images/adf5a2b6f32712d574aa37e869d97e3d7b7cb2928d3a6b528b69e2c069f6ddff.jpg) + +![](images/6359cdb3bcf18acb2641954846933fb9ea7a82f419e5a0a7ac47e9f49f8de076.jpg) + +![](images/cc5675255b6e0b735e54e344666ff457c5ec7c95fdd2227ee075cbe7857e1c30.jpg) +(c) Ours +(c) Ours + +![](images/ad9f9d56baf93fadd2ed196ca796f78abebd1cda6b042c597033a41ea22c898f.jpg) + +![](images/0c168a6fd3e405b30a08865f57db77107f3f6cdc313ffc392fae7a84d9f6bbc9.jpg) + +![](images/4c852f590ec007a60110d2bcdafbe00a27e6ce38c17833cc43c620f2fd5629e5.jpg) + +![](images/e1384ae7c0ec1e7f488b48208cb004e8fae38804b7ba05e70e098cf44e1415ff.jpg) + +![](images/8b33c14aaf4ebfcb40b8fc3c3648f3bc41feb4660f89fe6e21013dcafb328ff4.jpg) + +![](images/43120a35606c34ae370e38461d9a8e80f12541ecb3f11efb78ca974c76629530.jpg) + +![](images/5daeb3e2faad4c9138706362b44e598371a1c5a77a87d63def900770d3bf64c8.jpg) +(c) Ours + +![](images/00ec36bd3e6bdd6fe07c925ae06e507f66e428e9a2aefe0f1c809d776db21cab.jpg) + +![](images/7491ba432da54835a847a971bb82ebd17c58aff9b0cbe8bebd10d508549191a0.jpg) +(a) Input +(a) Input +Figure 4. Qualitative comparisons between the baselines and our method. RealmDreamer reconstruct the corresponding 3D Gaussian Field through a coarse-to-fine strategy; GenWarp generates a Gaussian field by applying InstantSplat [13] to the images estimated view-by-view; CAT3D conducts multi-view diffusion by conditioning on the input image and reconstructs the scene with these images by Zip-NeRF [4]; MVGenMaster and SEVA conduct multi-view diffusion and recover the scene with Mast3r [26] and InstantSplat [13]. For RealmDreamer, MVGenMaster, and SEVA, we visualize a mixture of rendered images and depth maps of the scene in the third column. + +![](images/19f0b0f6561091753679cb79e10edb8216f6088259fde894b7a7d5fa4fb8349a.jpg) +(b) MVGenMaster + +![](images/0b7786c425250f900ba5c7b4e3f166551f987af14cdbcd7b098651abf703320f.jpg) +(b) SEVA + +![](images/56b92d971d66766a9b162845a34f787f573c5e246536dc41474c46e0a983b37a.jpg) + +![](images/c5bfc16473209a646db110f3479feb149c201ef50dcd3dd8ad939dfe34a43a18.jpg) + +![](images/c7dd849fff295adca62f1a347f48fee957a8e0e87d5f1563f534cc09dd720a03.jpg) +(c) Ours + +![](images/e8462abbd3a65b856fdc99373b1f225729c0441c8328816c9dfb4e7762b7cf2b.jpg) + +![](images/fdb652b73600a54716c1e9f280efb8b9cbb2da20b41aac13d4a841a09791230c.jpg) +Figure 5. Effectiveness of MCS refinement. (a) The coarse Gaussian field contains some noisy and messy objects as marked by red boxes. (b) After our MCS refinement, the rendering results demonstrate improved quality. + +tion Sampling (SDS) [44] loss. GenWarp [50] learns to generate novel-view images of the same scene and leverages InstantSplat [13] to reconstruct the scene with a Gaussian field. CAT3D [15] trains a multi-view diffusion model to simultaneously sample the multi-view images by conditioning on the input image and then reconstructing the scene with Zip-NeRF [4]. Both MVGenMaster [6] and SEVA [75] also learns to synthesize novel views. We then adopt Mast3R [26] to recover their geometry and convert them to a Gaussian scene by PixelSplat [7]. For GenWarp [50], MVGenMaster [6] and SEVA [75], we test with their official implementation. For CAT3D [15] and RealmDreamer [51], we use the results in their project pages. + +Metrics. We adopt TSED [68] for geometry assessment of the generated scenes. Moreover, inspired by WonderJourney [67], we propose VLM-IQA metric to utilize VLMs, specifically LLaVA [33], InternVL [9], and Qwen2.5-VL [2], to evaluate the quality of the multiview images rendered from the reconstructed scenes, averaged on five aspects: noise free level (NF), sharp edge (Edge), structure (Struc.), detail clarity (Detail), and overall quality (Quality). LVM-IQA metric details are given in Sec.A.3 of the Supplementary Material. We also use PSNR to evaluate fidelity to the input reference image and further report the required VRAM and time to generate one scene. More implementation details are included in Sec.A.2 of the supplementary material. + +# 4.2. Comparisons with baselines + +The quantitative results in Table 1 show that our VistaDream, without any finetuning on the single-view scene reconstruction task, demonstrates significant improvements + +over other baselines on the geometry accuracy, the visual quality, and the fidelity to the input image. For LVM-IQA, VistaDream achieves comparable qualities as CAT3D which has been extensively trained on the single-view scene reconstruction task. + +In Fig. 4, we show qualitative results of the proposed method and the baseline methods across various scenes. More image/text-to-scene results and rendering videos are provided in the supplementary material. RealmDreamer exhibits significant distortion and noise due to the inconsistency introduced by warp-and-inpaint and shows blurry renderings due to the SDS refinement. The multi-view images generated by GenWarp exhibit noticeable inconsistencies, leading to noise and distortion in the reconstructed scenes. The multi-view images generated by CAT3D exhibit high quality and strong consistency by training on a large-scale multiview dataset, enabling the reconstruction of reliable and clear scenes. The inconsistent multi-view images from MVGenMaster and SEVA yields blurry or distorted results. In comparison to existing methods, VistaDream enlarges the initial scene scope with VLM-assisted inpainting, which improves the consistency and stability of subsequent inpainting. Furthermore, we ensured better multi-view consistency during scene optimization, ultimately yielding consistently higher visual and geometric quality. + +VistaDream achieves $20 \times$ speedup than optimization-based RealmDreamer, while consuming much less VRAM than CAT3D, which requires much GPU memory to interact multiple views in a single diffusion forward pass. VistaDream shows comparable efficiency to MVGenMaster and SEVA with no training needed. More runtime details are provided in Supplementary Material. + +![](images/631c937edcc48c02a997fd27b6c0ab11bc6f4d245a06907dec3f93b1090bbe3f.jpg) +Figure 6. Scene refinement via Score Distillation Sampling (SDS) or our Multiview Consistency Sampling (MCS). (a) The SDS refinement yields an overly smooth 3D scene, yielding blurry and inconsistent artifacts. (b) In contrast, our method produces enhanced qualities and better realism by considering multiview consistency. + +![](images/5d24e80bb80d40085a3ec23881ce5589217682b479729f9974fda928b68c72ee.jpg) + +![](images/d28d6a5693cbbce089169c51db5545b2b8170ba47a1ed1366b0d0acd9f0f533b.jpg) + +![](images/cb171fd2d80f6ce2bf9558c914efe5992fe7e4cde3526cd4086b25e7b6a0a7a6.jpg) + +![](images/2fec98ed3f91c5941e8fd17f9861d6ae55482bfe3a1dbdb95602bc317bf26462.jpg) + +![](images/2ef90df839ef9349d6e60825a4d12a23c047b97d5fa445ede70045c9278d2528.jpg) + +![](images/05ca12c22d89e9ae48fa89382d3f9e0d19127cb4499e809bbb8c953c24ef0989.jpg) + +
MethodNF↑Edge↑Struc.↑Detail↑Quality↑
Realm-DP [51]0.810.160.470.860.56
Realm-DP+MCS0.84+0.030.21+0.050.52+0.050.90+0.040.64+0.08
InstantSplat [13]0.390.090.340.830.41
InstantSplat+MCS0.57+0.180.30+0.210.59+0.250.88+0.050.59+0.18
+ +Table 2. Improve other methods with MCS. LLaVA-IQA results by applying MCS on single-view reconstruction method RealmDP and sparse-view reconstruction method InstantSplat. + +# 4.3. More analysis + +More analysis about $w_{t}$ setting in Eq. 3, results with sparse input views, time-consuming details, more ablations of our scaffold and MCS, more qualitative results, and limitation analysis are provided in the supplementary material. + +Effectiveness of MCS refinement. As shown in Fig. 5, the rendering of coarse Gaussian fields exhibits noticeable noise and artifacts, including distorted object boundaries and chaotic structures in complex regions. After MCS refinement, the accuracy and overall coherence of the scene have improved, allowing for the rendering of high-quality novel views, though minor detail blurring may occur to enforce multi-view consistency. The quantitative results in Table 1 further support these analyses. + +Compare MCS with SDS refinement. Score Distillation Sampling (SDS) [44] is a commonly used scene optimization technique that iteratively refines the single-view renderings by one-step diffusion. However, SDS only considers one input view, it tends to average the generation results for consistency, yielding blurry results [74] as shown in Fig. 6 (a). The proposed Multi-view Consistency Sampling simultaneously samples multi-view images by explicitly enforcing consistency, yielding high-quality and coherent multi-view images. These images achieve accurate and realistic scene optimization as shown in Fig. 6 (b). + +Improve other reconstruction methods with MCS. As in Table 2, we apply Multiview Consistent Sampling for Gaussian field refinement of sparse-view reconstruction method InstantSplat [13] and single-view reconstruc + +![](images/af78fbd4d92d7fa589d5f417759d56d009dd9a998fb4a0d72e076cd502277460.jpg) + +![](images/a0fd651178473a86c19d547259e7087f1fba8710f277c54532449592fbb96095.jpg) + +![](images/22355496c216c2327b122d77f06dbb43879d359400785ae94637b44ddd0a5dbd.jpg) + +![](images/6c66fc0b826a5d8b81434f0ad5a160b5b12c10eaff236275078dc0f7b7a6fb7e.jpg) + +![](images/f86fd3b4a7187f6c74450f4d0e16d46bfc6eb3ce0fe5e6931f1b2936abd72ed3.jpg) + +![](images/1ce9689aa338031a212e7579dbd4cd2d4807aa2b47cab977e8fbb096067c7c76.jpg) +(a) InstantSplat +Figure 7. Improving InstantSplat [13] with MCS. (a) Novel views from the sparse-view reconstruction method, InstantSplat [13] reconstruction suffers from distortion artifacts. (b) MCS effectively improves the quality of InstantSplat, yielding less noise, artifacts, and distortions. + +![](images/33f2e67e8aad54ba3ee9926fdc6279ea736b0ec6f7fb6790b7d4dc01919190da.jpg) +(b) InstantSplat + MCS + +![](images/d331c88959507fceffbb4db9d795b62bec60b725be95d514229c7fe17412c88b.jpg) + +tion method Realm-DP [51]. It can be seen that MCS effectively enhances Realm-DP by a $5\%$ average improvement and InstantSplat by a $17\%$ average improvement. Realm-DP means RealmDreamer with advanced Depth-Pro for depth estimation. More details are given in supplementary material. Qualitative results in Fig. 7 show that MCS greatly benefit the visual quality of InstantSplat reconstructions. + +# 5. Conclusion + +We propose VistaDream, a two-stage framework for 3D scene reconstruction from a single image. In the first stage, we enhance the stability of scene reconstruction by introduction a VLM-assisted global scaffold. In the second stage, we introduce Multi-view Consistency Sampling to sample high-quality and consistent multi-view images for scene optimization. Experimental results demonstrate that our method requires no fine-tuning on the single-view scene reconstruction task but achieves superior qualitative and quantitative results compared to the baseline methods. + +# 6. Acknowledgments + +This research is jointly supported by NSFC Basic Research Project for Ph.D. Candidate (No.424B2012) and the National Natural Science Foundation Project (No.42130105). + +# References + +[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 3 +[2] Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, et al. Qwen2. 5-vl technical report. arXiv preprint arXiv:2502.13923, 2025. 7 +[3] Arpit Bansal, Hong-Min Chu, Avi Schwarzschild, Soumyadip Sengupta, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Universal guidance for diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 843-852, 2023. 3 +[4] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Zip-nerf: Anti-aliased grid-based neural radiance fields. In ICCV, pages 19697-19705, 2023. 6, 7 +[5] Aleksei Bochkovskii, Amaël Delaunoy, Hugo Germain, Marcel Santos, Yichao Zhou, Stephan R Richter, and Vladlen Koltun. Depth pro: Sharp monocular metric depth in less than a second. arXiv:2410.02073, 2024. 4 +[6] Chenjie Cao, Chaohui Yu, Shang Liu, Fan Wang, Xiangyang Xue, and Yanwei Fu. Mvgenmaster: Scaling multi-view generation from any image via 3d priors enhanced diffusion model. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 6045-6056, 2025. 3, 5, 7 +[7] David Charatan, Sizhe Lester Li, Andrea Tagliasacchi, and Vincent Sitzmann. pixelsplat: 3d gaussian splats from image pairs for scalable generalizable 3d reconstruction. In CVPR, pages 19457-19467, 2024. 7 +[8] Hansheng Chen, Bokui Shen, Yulin Liu, Ruoxi Shi, Linqi Zhou, Connor Z Lin, Jiayuan Gu, Hao Su, Gordon Wetzstein, and Leonidas Guibas. 3d-adapter: Geometry-consistent multi-view diffusion for high-quality 3d generation. arXiv preprint arXiv:2410.18974, 2024. 3 +[9] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 24185–24198, 2024. 7 +[10] Jaeyoung Chung, Suyoung Lee, Hyeongjin Nam, Jaerin Lee, and Kyoung Mu Lee. Luciddreamer: Domain-free generation of 3d gaussian splatting scenes. arXiv preprint arXiv:2311.13384, 2023. 3, 4 +[11] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. NeurIPS, 34:8780-8794, 2021. 2 +[12] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In ICML, 2024. 2, 5 + +[13] Zhiwen Fan, Wenyan Cong, Kairun Wen, Kevin Wang, Jian Zhang, Xinghao Ding, Danfei Xu, Boris Ivanovic, Marco Pavone, Georgios Pavlakos, et al. Instantsplat: Unbounded sparse-view pose-free gaussian splatting in 40 seconds. arXiv preprint arXiv:2403.20309, 2024. 6, 7, 8 +[14] Xiao Fu, Wei Yin, Mu Hu, Kaixuan Wang, Yuexin Ma, Ping Tan, Shaojie Shen, Dahua Lin, and Xiaoxiao Long. Geowizard: Unleashing the diffusion priors for 3d geometry estimation from a single image. arXiv preprint arXiv:2403.12013, 2024. 2, 4 +[15] Ruiqi Gao, Aleksander Holynski, Philipp Henzler, Arthur Brussee, Ricardo Martin-Brualla, Pratul Srinivasan, Jonathan T Barron, and Ben Poole. Cat3d: Create anything in 3d with multi-view diffusion models. arXiv preprint arXiv:2405.10314, 2024. 2, 3, 5, 7 +[16] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. NeurIPS, 2020. 2, 5 +[17] Lukas Hollein, Ang Cao, Andrew Owens, Justin Johnson, and Matthias Nießner. Text2room: Extracting textured 3d meshes from 2d text-to-image models. In ICCV, pages 7909-7920, 2023. 3, 4 +[18] Yicong Hong, Kai Zhang, Juixiang Gu, Sai Bi, Yang Zhou, Difan Liu, Feng Liu, Kalyan Sunkavalli, Trung Bui, and Hao Tan. Lrm: Large reconstruction model for single image to 3d. arXiv preprint arXiv:2311.04400, 2023. 3 +[19] Mu Hu, Wei Yin, Chi Zhang, Zhipeng Cai, Xiaoxiao Long, Hao Chen, Kaixuan Wang, Gang Yu, Chunhua Shen, and Shaojie Shen. Metric3d v2: A versatile monocular geometric foundation model for zero-shot metric depth and surface normal estimation. arXiv:2404.15506, 2024. 4 +[20] Jitesh Jain, Jiachen Li, Mang Tik Chiu, Ali Hassani, Nikita Orlov, and Humphrey Shi. Oneformer: One transformer to rule universal image segmentation. In CVPR, pages 2989-2998, 2023. 3 +[21] Bingxin Ke, Anton Obukhov, Shengyu Huang, Nando Metzger, Rodrigo Caye Daudt, and Konrad Schindler. Repurposing diffusion-based image generators for monocular depth estimation. In CVPR, pages 9492-9502, 2024. 2 +[22] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139-1, 2023. 1, 2 +[23] Jeongsol Kim, Geon Yeong Park, and Jong Chul Ye. Dreamsampler: Unifying diffusion sampling and score distillation for image manipulation. In ECCV, pages 398-414. Springer, 2024. 3 +[24] Juil Koo, Chanho Park, and Minhyuk Sung. Posterior distillation sampling. In CVPR, pages 13352-13361, 2024. 3 +[25] Jiabao Lei, Jiapeng Tang, and Kui Jia. Rgbd2: Generative scene synthesis via incremental view inpainting using rgbd diffusion models. In CVPR, pages 8422-8434, 2023. 3 +[26] Vincent Leroy, Yohann Cabon, and Jérôme Revaud. Grounding image matching in 3d with mast3r. In European Conference on Computer Vision, pages 71-91. Springer, 2024. 6, 7 +[27] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified + +vision-language understanding and generation. In ICML, pages 12888-12900. PMLR, 2022. 3 +[28] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, pages 19730–19742. PMLR, 2023. 3, 4 +[29] Peng Li, Yuan Liu, Xiaoxiao Long, Feihu Zhang, Cheng Lin, Mengfei Li, Xingqun Qi, Shanghang Zhang, Wenhan Luo, Ping Tan, et al. Era3d: High-resolution multiview diffusion using efficient row-wise attention. arXiv preprint arXiv:2405.11616, 2024. 3 +[30] Xinyang Li, Zhangyu Lai, Linning Xu, Yansong Qu, Liujuan Cao, Shengchuan Zhang, Bo Dai, and Rongrong Ji. Director3d: Real-world camera trajectory and 3d scene generation from text. arXiv preprint arXiv:2406.17601, 2024. 3 +[31] Yixun Liang, Xin Yang, Jiantao Lin, Haodong Li, Xiaogang Xu, and Yingcong Chen. Luciddreamer: Towards high-fidelity text-to-3d generation via interval score matching. In CVPR, pages 6517–6526, 2024. 3 +[32] Shanchuan Lin, Bingchen Liu, Jiashi Li, and Xiao Yang. Common diffusion noise schedules and sample steps are flawed. In WACV, pages 5404-5411, 2024. 5 +[33] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. NeurIPS, 2024. 2, 3, 4, 7 +[34] Minghua Liu, Chao Xu, Haian Jin, Linghao Chen, Mukund Varma T, Zexiang Xu, and Hao Su. One-2-3-45: Any single image to 3d mesh in 45 seconds without per-shape optimization. NeurIPS, 36, 2024. 3 +[35] Yuan Liu, Cheng Lin, Zijiao Zeng, Xiaoxiao Long, Lingjie Liu, Taku Komura, and Wenping Wang. Syncdreamer: Generating multiview-consistent images from a single-view image. arXiv preprint arXiv:2309.03453, 2023. 3 +[36] Xiaoxiao Long, Yuan-Chen Guo, Cheng Lin, Yuan Liu, Zhiyang Dou, Lingjie Liu, Yuexin Ma, Song-Hai Zhang, Marc Habermann, Christian Theobalt, et al. Wonder3d: Single image to 3d using cross-domain diffusion. In CVPR, pages 9970-9980, 2024. 3 +[37] Simian Luo, Yiqin Tan, Longbo Huang, Jian Li, and Hang Zhao. Latent consistency models: Synthesizing high-resolution images with few-step inference. arXiv preprint arXiv:2310.04378, 2023. 2, 3 +[38] Norman Müller, Katja Schwarz, Barbara Rössle, Lorenzo Porzi, Samuel Rota Bulò, Matthias Nießner, and Peter Kontschieder. Multidiff: Consistent novel view synthesis from a single image. In CVPR, 2024. 2 +[39] Richard A Newcombe, Shahram Izadi, Otmar Hilliges, David Molyneaux, David Kim, Andrew J Davison, Pushmeet Kohi, Jamie Shotton, Steve Hodges, and Andrew Fitzgibbon. Kinectfusion: Real-time dense surface mapping and tracking. In ISMAR, 2011. 2 +[40] OpenAI. Chatgpt. https://chatgpt.com/, 2023.3 +[41] OpenAI. Dall-e3. https://openai.com/index/dall-e-3/, 2023.2 +[42] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 2 + +[43] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 2, 3, 5 +[44] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 1, 7, 8 +[45] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PMLR, 2021. 3 +[46] Chris Rockwell, David F Fouhey, and Justin Johnson. Pixelsynth: Generating a 3d-consistent experience from a single image. In ICCV, pages 14104-14113, 2021. 4 +[47] Kyle Sargent, Zizhang Li, Tanmay Shah, Charles Herrmann, Hong-Xing Yu, Yunzhi Zhang, Eric Ryan Chan, Dmitry Lagun, Li Fei-Fei, Deqing Sun, et al. Zeronvs: Zero-shot 360-degree view synthesis from a single real image. arXiv preprint arXiv:2310.17994, 2023. 3 +[48] Axel Sauer, Dominik Lorenz, Andreas Blattmann, and Robin Rombach. Adversarial diffusion distillation. arXiv preprint arXiv:2311.17042, 2023. 3 +[49] Johannes Lutz Schonberger, Enliang Zheng, Marc Pollefeys, and Jan-Michael Frahm. Pixelwise view selection for unstructured multi-view stereo. In ECCV, 2016. 2 +[50] Junyoung Seo, Kazumi Fukuda, Takashi Shibuya, Takuya Narihira, Naoki Murata, Shoukang Hu, Chieh-Hsin Lai, Seunglyong Kim, and Yuki Mitsufuji. Genwarp: Single image to novel views with semantic-preserving generative warping. arXiv preprint arXiv:2405.17251, 2024. 2, 5, 7 +[51] Jaidev Shriram, Alex Trevithick, Lingjie Liu, and Ravi Ramamoorthi. Realmdreamer: Text-driven 3d scene generation with inpainting and depth diffusion. arXiv preprint arXiv:2404.07199, 2024. 2, 3, 4, 5, 7, 8 +[52] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 2 +[53] Jiaxiang Tang, Zhaoxi Chen, Xiaokang Chen, Tengfei Wang, Gang Zeng, and Ziwei Liu. Lgm: Large multi-view gaussian model for high-resolution 3d content creation. arXiv preprint arXiv:2402.05054, 2024. 3 +[54] Zhenyu Tang, Junwu Zhang, Xinhua Cheng, Wangbo Yu, Chaoran Feng, Yatian Pang, Bin Lin, and Li Yuan. Cycle3d: High-quality and consistent image-to-3d generation via generation-reconstruction cycle. arXiv preprint arXiv:2407.19548, 2024. 3 +[55] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 3 +[56] Hung-Yu Tseng, Qinbo Li, Changil Kim, Suhib Alsisan, Jia-Bin Huang, and Johannes Kopf. Consistent view synthesis with pose-guided diffusion models. In CVPR, 2023. 2 +[57] Peng Wang and Yichun Shi. Imagedream: Image-prompt multi-view diffusion for 3d generation. arXiv preprint arXiv:2312.02201, 2023. 3 + +[58] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. NeurIPS, 36, 2024. 3 +[59] Zhouxia Wang, Ziyang Yuan, Xintao Wang, Yaowei Li, Tianshui Chen, Menghan Xia, Ping Luo, and Ying Shan. Motionctrl: A unified and flexible motion controller for video generation. In SIGGRAPH, pages 1-11, 2024. 3 +[60] Shaoan Xie, Zhifei Zhang, Zhe Lin, Tobias Hinz, and Kun Zhang. Smartbrush: Text and shape guided object inpainting with diffusion model. In CVPR, 2023. 2 +[61] Bisheng Yang, Zhen Dong, Fuxun Liang, and Xiaoxin Mi. Ubiquitous Point Cloud: Theory, Model, and Applications. CRC Press, 2024. 2 +[62] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In CVPR, 2024. 2 +[63] Yiying Yang, Fukun Yin, Jiayuan Fan, Xin Chen, Wanzhang Li, and Gang Yu. Scene123: One prompt to 3d scene generation via video-assisted and consistency-enhanced mae. arXiv preprint arXiv:2408.05477, 2024. 3 +[64] Tianwei Yin, Michael Gharbi, Richard Zhang, Eli Shechtman, Fredo Durand, William T Freeman, and Taesung Park. One-step diffusion with distribution matching distillation. In CVPR, pages 6613-6623, 2024. 3 +[65] Meng You, Zhiyu Zhu, Hui Liu, and Junhui Hou. Nvs-solver: Video diffusion model as zero-shot novel view synthesizer. arXiv preprint arXiv:2405.15364, 2024. 3, 5 +[66] Hong-Xing Yu, Haoyi Duan, Charles Herrmann, William T Freeman, and Jiajun Wu. Wonderworld: Interactive 3d scene generation from a single image. arXiv preprint arXiv:2406.09394, 2024. 2 +[67] Hong-Xing Yu, Haoyi Duan, Junhwa Hur, Kyle Sargent, Michael Rubinstein, William T Freeman, Forrester Cole, Deqing Sun, Noah Snavely, Jiajun Wu, et al. Wonderjourney: Going from anywhere to everywhere. In CVPR, pages 6658-6667, 2024. 2, 5, 7 +[68] Jason J Yu, Fereshteh Forghani, Konstantinos G Derpanis, and Marcus A Brubaker. Long-term photometric consistent novel view synthesis with diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7094-7104, 2023. 7 +[69] Jason J Yu, Fereshteh Forghani, Konstantinos G Derpanis, and Marcus A Brubaker. Long-term photometric consistent novel view synthesis with diffusion models. In CVPR, 2023. 2 +[70] Wangbo Yu, Jinbo Xing, Li Yuan, Wenbo Hu, Xiaoyu Li, Zhipeng Huang, Xiangjun Gao, Tien-Tsin Wong, Ying Shan, and Yonghong Tian. Viewcrafter: Taming video diffusion models for high-fidelity novel view synthesis. arXiv preprint arXiv:2409.02048, 2024. 3 +[71] Jingyi Zhang, Jiaxing Huang, Sheng Jin, and Shijian Lu. Vision-language models for vision tasks: A survey. T-PAMI, 2024. 3 +[72] Jingbo Zhang, Xiaoyu Li, Ziyu Wan, Can Wang, and Jing Liao. Text2nerf: Text-driven 3d scene generation with neural radiance fields. TVCG, 2024. 3 + +[73] Lvming Zhang. Fooocus. https://github.com/1llyasviel/Fooocus, 2023.2, 3, 4 +[74] Yiming Zhong, Xiaolin Zhang, Yao Zhao, and Yunchao Wei. Dreamlcm: Towards high-quality text-to-3d generation via latent consistency model. arXiv preprint arXiv:2408.02993, 2024. 3, 8 +[75] Jensen Zhou, Hang Gao, Vikram Voleti, Aaryaman Vasishta, Chun-Han Yao, Mark Boss, Philip Torr, Christian Rupprecht, and Varun Jampani. Stable virtual camera: Generative view synthesis with diffusion models. arXiv preprint arXiv:2503.14489, 2025. 3, 5, 7 +[76] Zi-Xin Zou, Zhipeng Yu, Yuan-Chen Guo, Yangguang Li, Ding Liang, Yan-Pei Cao, and Song-Hai Zhang. Triplane meets gaussian splatting: Fast and generalizable single-view 3d reconstruction with transformers. In CVPR, pages 10324-10335, 2024. 3 \ No newline at end of file diff --git a/ICCV/2025/VistaDream_ Sampling multiview consistent images for single-view scene reconstruction/images.zip b/ICCV/2025/VistaDream_ Sampling multiview consistent images for single-view scene reconstruction/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..b5ab027e4571ec58fd22c71ec584850729f3d45b --- /dev/null +++ b/ICCV/2025/VistaDream_ Sampling multiview consistent images for single-view scene reconstruction/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b6ee9064fe69b5b510318c6221a69319ae23e2d6d2d0934f371bbdad0165590b +size 1176451 diff --git a/ICCV/2025/VistaDream_ Sampling multiview consistent images for single-view scene reconstruction/layout.json b/ICCV/2025/VistaDream_ Sampling multiview consistent images for single-view scene reconstruction/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..cc37378dfa50d5411ac241b7c2bf0df20594343c --- /dev/null +++ b/ICCV/2025/VistaDream_ Sampling multiview consistent images for single-view scene reconstruction/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5f24e65e582cc9417236d02d0740d0c8f946130162f5ee4b67d113ba43a16d00 +size 444507 diff --git a/ICCV/2025/Visual Chronicles_ Using Multimodal LLMs to Analyze Massive Collections of Images/faf5d8e8-4101-4ccc-91eb-41873d923f20_content_list.json b/ICCV/2025/Visual Chronicles_ Using Multimodal LLMs to Analyze Massive Collections of Images/faf5d8e8-4101-4ccc-91eb-41873d923f20_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..8fbd53f7fbbeb6df3484d8d19f4410dfb8640e9e --- /dev/null +++ b/ICCV/2025/Visual Chronicles_ Using Multimodal LLMs to Analyze Massive Collections of Images/faf5d8e8-4101-4ccc-91eb-41873d923f20_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:da76b2dc2a93c6cbf424178170d019ac5723c60cc2ca2653c4833f3af2967f51 +size 81446 diff --git a/ICCV/2025/Visual Chronicles_ Using Multimodal LLMs to Analyze Massive Collections of Images/faf5d8e8-4101-4ccc-91eb-41873d923f20_model.json b/ICCV/2025/Visual Chronicles_ Using Multimodal LLMs to Analyze Massive Collections of Images/faf5d8e8-4101-4ccc-91eb-41873d923f20_model.json new file mode 100644 index 0000000000000000000000000000000000000000..e4f892f416a9c31f6fcdbe32ed08c4bdd8235bb1 --- /dev/null +++ b/ICCV/2025/Visual Chronicles_ Using Multimodal LLMs to Analyze Massive Collections of Images/faf5d8e8-4101-4ccc-91eb-41873d923f20_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0179226c25f78b26a5f9eee54fd81f03fcdeddc8939e600759cbe685dbb509cd +size 103281 diff --git a/ICCV/2025/Visual Chronicles_ Using Multimodal LLMs to Analyze Massive Collections of Images/faf5d8e8-4101-4ccc-91eb-41873d923f20_origin.pdf b/ICCV/2025/Visual Chronicles_ Using Multimodal LLMs to Analyze Massive Collections of Images/faf5d8e8-4101-4ccc-91eb-41873d923f20_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..8fb229204b17287178dbbb62248f853a86ab1c3d --- /dev/null +++ b/ICCV/2025/Visual Chronicles_ Using Multimodal LLMs to Analyze Massive Collections of Images/faf5d8e8-4101-4ccc-91eb-41873d923f20_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a113d8b5836a17243c24165f49c7c1df522c34ea23eb44a7ec4a4bcad1ece84c +size 47876791 diff --git a/ICCV/2025/Visual Chronicles_ Using Multimodal LLMs to Analyze Massive Collections of Images/full.md b/ICCV/2025/Visual Chronicles_ Using Multimodal LLMs to Analyze Massive Collections of Images/full.md new file mode 100644 index 0000000000000000000000000000000000000000..3612bc3435f50e68199c8458d75bf453601db96c --- /dev/null +++ b/ICCV/2025/Visual Chronicles_ Using Multimodal LLMs to Analyze Massive Collections of Images/full.md @@ -0,0 +1,344 @@ +# Visual Chronicles: Using Multimodal LLMs to Analyze Massive Collections of Images + +Boyang Deng $^{1}$ Songyou Peng $^{2,*}$ Kyle Genova $^{2,*}$ +Gordon Wetzstein $^{1}$ Noah Snavely $^{2}$ Leonidas Guibas $^{1,2}$ Thomas Funkhouser $^{2}$ $^{1}$ Stanford University $^{2}$ Google DeepMind + +![](images/5e66abb8ad4d039c7b21be9f459b68425384d020eb2164aae78d725e6457ff13.jpg) + +![](images/6c93f007965d8ba0e0cb2e112a40c1bbf09ed4ab12853905447122af3b2a3dce.jpg) +Figure 1. Visual Trend Discovery from Massive-Scale Collections of Images. We introduce Visual Chronicles, a novel approach that uses Multimodal LLMs (MLLMs) to make discoveries from massive collections of images over time. We analyze millions of images collected over more than 10 years by Google Street View. As seen above, our method discovers both expected and surprising visual trends in San Francisco—e.g. outdoor dining areas were added to many storefronts, and miles of overpass was painted blue, both with visual evidence. + +![](images/1202ee3aab979429e1f774ea76992a19ec1b9d344e940cb564cfe4f834061869.jpg) + +![](images/ab1da394b8daec46bc0b30acdde84199fb02c9819a98ad7e122c4e168d26c8d6.jpg) + +![](images/342dc82d1c147620de5138e13d1314c6ee09e0be7c49b1b6cf6a1a54e21778ea.jpg) + +![](images/e33177e7af39fa538f528bc6097bcfa0e45068fe30d354202870066581f88001.jpg) + +# Abstract + +We present a system using Multimodal LLMs (MLLMs) to analyze a large database with tens of millions of images captured at different times, with the aim of discovering patterns in temporal changes. Specifically, we aim to capture frequent co-occurring changes ("trends") across a city over a certain period. Unlike previous visual analyses, our analysis answers open-ended queries (e.g., "what are the frequent types of changes in the city?") without any predetermined target subjects or training labels. These properties cast prior learning-based or unsupervised visual analysis tools unsuitable. We identify MLLMs as a novel tool for their open-ended semantic understanding capabilities. Yet, our datasets are four orders of magnitude too large for an MLLM to ingest as context. So we introduce a bottom-up procedure that decomposes the massive visual analysis problem into more tractable sub-problems. We carefully design MLLM-based solutions to each sub-problem. During experiments and ablation studies with our system, we find it significantly outperforms baselines and is able to discover interesting + +trends from images captured in large cities (e.g., "addition of outdoor dining," "overpass was painted blue," etc.). + +# 1. Introduction + +Searching for semantic patterns and correlations in image databases is a long-standing problem in computer vision with applications in urban planning, visual analytics, and other fields. Classical research on this problem includes tracking changes in images within specific domains over time (e.g., "A Century of Portraits" [15]) and finding salient patterns of objects in images (e.g., estimating demographics from vehicle sightings [14]). However, these methods have only been demonstrated for predetermined specific target subjects (e.g., faces [15] or cars [14]); and often require labeled training data (e.g., labeled car attributes [14]). + +In this paper, we study how to make open-ended discoveries about frequently occurring changes ("trends") across + +a city within a given time period, using visual analysis of a massive collection of images without required labeled training data. We provide a study using Google Street View as a data source [3]. Street View provides tens of millions of posed and timestamped images covering large cities with repeated captures spanning a decade. Our goal is to answer open-vocabulary queries about how a city has changed over time. In contrast to prior work, our queries are open-ended, and not limited to a predetermined list of subject categories. Additionally, the data doesn't come with labeled answers to our queries, and manual labeling at scale is infeasible because the nature of the queries are unknown a priori. + +The new properties of our analysis make it a difficult problem. Any pretrained subject-specific recognition model, e.g., a car detector, is insufficient for open-ended queries. Meanwhile, the lack of labels prevents us from training a tailored learning-based model as in prior work [14]. Training-free image analysis techniques, from image features [11] to unsupervised remote sensing [38], have also been used in prior work. However, our goal is to find semantically interesting insights. This demands a holistic semantic understanding of images, e.g., understanding that a shop is closed from visual evidence, which these prior techniques fail to offer. + +Therefore, we develop a new, scalable, training-free solution to open-ended visual analysis, leveraging the considerable scene understanding capabilities of Multimodal Large Language Models (MLLMs). Our solution is motivated by a key observation—readymade MLLMs are very effective at open-ended semantic analyses of a small collection of images. They can detect semantically salient temporal changes with high accuracy in 20-40 images. However, a straightforward application of MLLMs does not solve our problem due to the massive scale of the data. For instance, simply asking an MLLM to "name the trends" without using images would fail, because we empirically find they mainly produce abstract answers such as economical growth. Alternatively, we could feed all images as context to an MLLM, and ask it to "find the trends". However, this does not work due to the inability of MLLMs to handle massive context. Thus, we design a novel bottom-up hierarchical system to decompose the problem into smaller subtasks: detecting and summarizing local visual changes followed by finding global trends among these local changes. For each subtask, we propose an effective MLLM-based solution. To the best of our knowledge, our system marks the first use of MLLMs for the analysis of a database of more than a million images. + +Experiments with our proposed system reveal several interesting insights. According to results of tests with manually labeled data produced for this paper, we find that our system discovers visual trends more accurately than prior solutions or trivial use of MLLMs. In experiments with 20 million Street View images from each of two major cities (NYC and SF), it finds several interesting trends that would be difficult + +to discover without detailed visual analyses (see Fig. 1). Overall, our research contributions are as follows: + +- We study a novel open-ended visual analysis problem on a massive-scale dataset without labels. +- We identify MLLMs as the critical tool for such unprecedented analyses, and design a novel system for massive-scale analyses using MLLMs, bypassing the context and efficiency limitations. +- We present experiments using our system in two major cities yielding unexpected and informative findings. + +These contributions together signal a new paradigm of computer vision studies—advancing from previous task-specific supervised training or supervised fine-tuning, to carefully designed use of MLLMs in a system like ours, for complicated large-scale tasks without labels. + +# 2. Related Work + +# 2.1. Large-Scale Image Analysis + +General Analysis. Analyzing large-scale visual data—such as datasets composed of millions of images—can unveil important characteristics of the data that cannot be observed from just a few examples. Prior work has applied such analytical study to various visual data collections, leading to interesting discoveries, including distinctive visual styles [11, 21, 45], recurring visual patterns in art [17, 44], dataset biases and irregularities [4, 24, 49], and socioeconomic insights such as economic outcomes [19], urban neighborhood perceptions [12], and demographic compositions [14]. Among these works, imagery linked to GIS [3] has been a popular data source [11, 12, 14, 19, 36], closely related to the data in our study. While these methods have made significant contributions, they largely rely on image-driven analysis, e.g., clustering of visual elements [11, 36], interconnecting images through shared visual structure [16], learning visual prototypes [51], average images [15, 59], color diagrams [49], image recognition models [12, 14, 19], or analysis-by-synthesis approaches with image generation models [6, 9, 10, 45]. These methods primarily analyze visual information without considering temporal change. + +Temporal Analysis. Building upon the foundation of large-scale image analysis, other methods examine visual data over time to uncover changes, trends, and evolving patterns [15, 18, 26, 34-36, 39, 40, 53, 59]. For example, Martin-Brualla et al. [34, 35] mined time-lapse videos from internet photos to visualize landmark transformations, revealing patterns of construction, decay, and seasonal variations. Naik et al. [40] predicted urban change by analyzing sequential street view imagery, identifying factors that influence physical transformations in cities. Ginosar et al. [15] analyzed a century of American high school yearbook photos to study the evolution of facial expressions and fashion trends, demonstrating how temporal analysis can provide sociolog + +Image A (Mar. 2011) +![](images/0e98b07ec5c9dbd72ec9c887340f1b64b04bff0ab5201a1c40bc8e2d2f378f47.jpg) +"These images were taken in the same place over multiple years. Please find all changes and tell me when they are seen." + +![](images/c931a10c5aa64f61caf9660ac9b99d512efbb2b1c6e4ae02c722705c384447ed.jpg) +Image B (Feb. 2014) + +![](images/fe02c823f061770273c9f4fdb9fc92b0061a858b2838eb29a20c26dee9e9996e.jpg) +Image C (Oct. 2015) + +![](images/dce20c288f4a9cbc55b1ee65ab28bba5880b8a11ae39070840ece748838418f2.jpg) +MLLM +Figure 2. Using MLLMs for Visual Change Detection. Given a set of images captured from nearby views at different timestamps (top row), we use an MLLM as a visual analyst to detect changes. + +Here's what changed: + +Change 1 (Image $A \to$ Image $B$ ): Outdoor seats + +were added in front of the restaurant. + +Change 2 (Image $B \to$ Image $C$ ): An advertisement + +board was put up in front of the store. + +·· + +ical insights. These methods either require specific, predetermined target categories [15] or labelled training data [40], whereas our method leverages MLLMs for open-ended analyses without the need for labeled data. Our work can capture temporal and contextual changes more effectively, defining a new direction for future image analysis. Note that a separate line of research focuses on analyzing video data, targeting much smaller time scales and relying on the continuity between frames [4, 23, 27, 28, 32, 50, 55]. In contrast, we focus on temporal analysis of non-continuous temporal image collections. + +# 2.2. Vision-Language Models in Visual Analysis + +The evolution of VLMs has significantly advanced visual analysis. Early models, such as image captioning systems [22, 52], were limited to generating descriptions for individual images, restricting analyses to small-scale tasks like captioning differences between pairs of images [5, 25, 41, 54]. Models like CLIP [42], ALIGN [20], and their variants [8, 56, 60] have also been used in various visual analysis tasks. However, as VLM encoders without generative capabilities, they fall short for tasks requiring detailed image descriptions and reasoning. + +The advent of large-scale VLMs, from Flamingo [2], PaLI [7], and BLIP-2 [30], to powerful new models like Gemini [47, 48] and ChatGPT [1], affords new opportunities for visual analysis. These models integrate the expansive understanding of large language models with visual inputs, enabling sophisticated visual reasoning and generation. While LLMs have been employed in large-scale text analyses [31, 46, 57], large-scale visual analyses using large VLMs remain scarce. A recent method, VisDiff [13], uses MLLMs to find differences between datasets of a few thousand images, but does not scale to our problem due to dataset size and differing objectives. Also, their verification step is largely based on CLIP embeddings that prove to be unsat + +isfactory in our task. One concurrent work, TeDeSC [33], applies VisDiff's framework to semantic clustering tasks in relatively small datasets, organizing images based on text descriptions to find groups. Another recent work [58] introduces a general analytical framework based on VLMs, utilizing creative language predicates as interpretable hypotheses. This approach, however, does not directly apply to our task due to the intractably large hypothesis space and the inability of LLMs to effectively predict them. Additionally, completing their analysis requires labels for regression, equivalent to annotating "is an image in a trend?", which is extremely challenging in our context. + +# 3. Massive-Scale Image Analysis with MLLMs + +# 3.1. Task: Open-Ended Urban Imagery Analysis + +Our work uses visual analysis to make open-ended discoveries from massive collections of images without labels. We use two image collections (New York City and San Francisco) provided by Google Street View with permission, each with 20M posed and timestamped images captured at street level from 2011 to 2023. + +Our objective is to identify and describe trends of visual changes over time in these image collections, with interpretable text descriptions accompanied by visually grounded evidence. By visual change, we refer to a semantic change event that occurred at a specific imaged location between two discrete times (corresponding to times when that location was imaged). This query is open-ended because the event can be of any nature and happen to any kind of subject. A visual trend is a group of similar changes that have occurred at least $N$ times across different locations within a certain period, where $N$ is a hyper-parameter determining the minimum frequency for a trend to be considered significant. Moreover, we desire the most detailed and specific insights we can find that still occur at least $N$ times, to gain a deeper understanding of subtle urban transformations. + +# 3.2. Using MLLMs for Massive Image Analysis + +We aim to leverage the powerful analytical capabilities of MLLMs at large scales. However, the naive approach of feeding all the data to MLLMs does not work; it is not possible to feed tens of millions of images into an MLLM due to context size limitations (e.g., Gemini-1.5 Pro [43]'s limit is 2 million tokens, which corresponds to at most 8k images). Instead, we decompose massive dataset queries into a set of local queries that can be answered independently by an MLLM. Then, we combine the answers to those local queries into an answer about the whole dataset. This bottom-up approach is appropriate because there are natural ways to decompose an image dataset into small sets of images related to individual visual changes. During early experiments, we discovered that MLLMs are very effective (see Sec. 4.2) at + +# Algorithm 1 Using MLLMs for Trend Discovery + +Require: Trend T, set of changes $\mathcal{C}$ , parameters $k$ , $N$ +Ensure: Is T a real trend or not? + +1: Encode $\mathbf{T}$ and each $c_{i} \in \mathcal{C}$ using TextEmb +2: Compute distances $d_{i}$ between $\mathbf{T}$ and each $c_{i}$ +3: Select $k$ nearest $c_{i}$ based on $d_{i}$ as set $C_k$ +4: Use MLLM to determine if each $c_{j} \in C_{k}$ belongs to $\mathbf{T}$ +5: Let $N_{\mathrm{pos}}$ be the number of positive responses +6: if $N_{\mathrm{pos}} \geq N$ then +7: return Trend T is positive +8: else +9: return Trend T is negative + +describing subtle changes within a temporal image sequence captured from nearly the same perspective (e.g., those shown in Fig. 2). MLLMs not only detect small changes within such sequences, but also effectively ignore many common trivial changes (variations in time of day, season, etc.) and focus instead more on consistent semantic content that might be of interest to a person. They also identify where in the sequence the visual changes occur in addition to producing detailed text descriptions of them (e.g., "the store got a new 'for sale' sign on the window between the 3rd and 4th image"). + +With this key insight regarding the capabilities of MLLMs, we design a two-step analysis system consisting of local change detection followed by trend discovery. In both steps, we harness the analytical power of MLLMs, and design careful solutions to effectively process and analyze the data. The following sections describe each step in detail. + +# 3.3. Using MLLMs for Visual Change Detection + +In the first step of our system, we detect local visual changes by analyzing small sets of images captured from approximately the same viewpoint at different times (Fig. 2 top row, see data processing details in Supplementary). Since the images are time-stamped, we can sort them into chronological order at each location. We feed every sorted image sequence into an MLLM and craft a query that, in essence, asks: 1) what changes have occurred, and 2) from which images these changes were detected. The output for the first question provides a text description for each detected visual change, while the second grounds these detections to specific visual evidence and timestamps. Please refer to the supplement for more details about image sequence preparation, MLLM prompting, and more example model outputs. + +# 3.4. Using MLLMs for Visual Trend Discovery + +In the second step, we aim to analyze the set of local visual changes produced by local change detection to discover frequent visual trends occurring throughout a city. Specifically, we look for trends where similar changes occur at least $N$ times at different locations in the same time period. + +At first glance, one may consider a simple solution of presenting all local change descriptions to an MLLM and asking for trends. However, this is impractical for massive city-scale datasets. For instance, from a collection of 20 million images for one city, the first step of the system produces over 3 million local changes. Analyzing them all at once would exceed the context window size of current MLLMs. + +To address this issue, we designed a two-step process that 1) produces visual trend proposals, and then 2) verifies which proposed trends are supported by at least $N$ visual changes. The first proposal step is relatively simple. We embed the text description for every visual change into a vector space and use Canopy Clustering [37] with a tolerant loose threshold to find potential cluster proposals. + +The second verification step is more challenging. An easy and efficient approach would be to assign each visual change to a trend proposal based on its proximity in an embedding space. However, we find that this approach performs poorly because distances in embedding spaces are not distinctive enough to capture the subtle similarities and differences between visual changes required for our application. Image embeddings may encode both change-relevant and change-irrelevant content, and text embeddings may fail to capture subtle wording differences (e.g., adding "not") or different conceptual granularities (e.g., "a storefront changed its business" and "a pizza shop changed into a café" are difficult to cluster because they represent different levels of specificity). Moreover, there is no universal embedding distance threshold to classify whether a change belongs to a trend, making it difficult to achieve consistent accuracy. + +An alternative, albeit impractical, approach would be to feed every detected local change and every proposed trend to an MLLM and ask it whether the change fits within the trend (yes or no). This approach is compelling because we find that MLLMs provide remarkably high accuracy for this small task. However, the downside of this approach is efficiency. Verifying 200 trend proposals exhaustively against 3 million changes using MLLMs would take more than one year. + +To address this issue, we propose a hybrid algorithm which combines the efficiency of text embeddings with the precision of MLLMs, as shown in Alg. 1. Since we only need to find at least $N$ changes that belong to a trend to confirm it, we first use text embeddings to sort all changes based on their similarity to the trend proposal. Next, we use the MLLM to verify only the top $k > N$ nearest neighbors. If $N$ positive matches are found among these $k$ , the trend is confirmed. We set $k = 1500$ for $N = 500$ in our experiments, reducing the MLLM inference cost by $2000 \times$ and making the process computationally feasible. + +# 4. Evaluations + +In this section, we evaluate Visual Chronicles against alternatives as well as design choices made in Visual Chronicles. + +We use Gemini-1.5 Pro [43] as the MLLM throughout the paper, for its impressive long-context multimodal analytical capability. We collect a small set of data for evaluation. + +# 4.1. Comparisons to Trivial Use of MLLMs + +We first investigate how well our system works in comparison to naive solutions using MLLMs trivially. + +Can an MLLM predict similar visual trends without looking at any images? In the first experiment, we investigate how the visual trends produced by our proposed method compare to those produced directly by an MLLM when prompted only with text (i.e., without providing any images as context). We acknowledge that this baseline is not apples-to-apples since our method can access lots of extra data and use significantly more compute. However, it addresses an important question: "Are MLLMs already knowledgeable enough to list visual trends on their own?" To perform this comparison, we asked Gemini-1.5 Pro and GPT-4o to predict visual trends over the last decade using various prompt strategies. We found that MLLMs' answers are usually abstract, e.g. "Increased focus on sustainability", and general without intriguing details, e.g. "store closures". Please see complete answers in Supplementary. In contrast, our method identifies specific and subtle trends, e.g. addition of solar panels in Fig. 4, more nuanced, e.g. closures of grocery stories in Fig. 5, and sometimes unexpected trends, e.g. an overpass painted blue in Fig. 1, all supported by visual evidence. These visual trends provide a more detailed and verifiable understanding of changes in the city than the broad abstract ones provided by the MLLM on its own. + +Can an MLLM process all the images in a massive collection in one query? In the second experiment, we investigate the limits of using MLLMs to predict visual trends with a single query using a set of images as context. For this test, ideally, we'd provide the entire massive image collection of millions of images as context to the MLLM, and then ask "what are the visual trends?" However, this is not possible, since MLLMs have limits on the number of tokens that can be provided as context (8K images for the MLLM with largest limit [43]). So, instead we performed tests with randomly sampled subsets of 8K images out of 20M in NYC as context, and take the union of answers from all subsets as the final result. The results however are not satisfactory—due to the severe under-sampling of images in each batch, only the extremely frequent trends can be found, e.g. the addition and removal of scaffoldings in NYC which our system also identifies. Additionally, this solution only finds less than 10 trends throughout a whole city even after union of multiple sub-samples. See complete answers in Supplementary. + +# 4.2. Local Change Detection Evaluation + +Next, we quantitatively evaluate the design choices made in Visual Chronicles, starting from local change detection. To + +
HoG [11]C-Hist [49]R-Sensing [38]CLIP [42]NV-Emb [29]Gemini [43]
AP16.44%16.76%18.51%26.52%23.75%76.56%
+ +Table 1. Evaluating Visual Change Detection. Gemini detects visual changes with significantly higher average precision (AP) than baselines based on image features (HoG, C-Hist), remote sensing (R-Sensing), and semantic embeddings (CLIP, NV-Emb). + +understand how good Visual Chronicles is at local change detection, we collect a dataset of 200 locations with 3036 images in total. For each location, we manually label the changes, including between which 2 consecutive images each change happens and the description of changes. + +We compare our method with prior unsupervised image analysis tools, including image features such as HoG features [11] and Color Histograms (C-Hist) [49], and semantic features such as CLIP [42] and NV-Emb [29]. These methods first featurize every image in the image sequence for each location and measures the feature distance between consecutive images as an indicator for potential change. For NV-Emb, we use Gemini [43] to caption each image and embed the caption text. Additionally, we compare ours with an unsupervised remote sensing method (R-Sensing) [38]. Note that other learning-based approaches [14] requires labelled training data, hence are inapplicable. + +To have a fair comparison between MLLMs and baselines that only predict the change likelihood between two images, we use a classification metric, Average Precision (AP). For any consecutive image pairs in our test set, our human labels give us binary ground truth on whether there is a change. We then compute the AP on this binary classification task for all methods. Note that if there are multiple changes between a pair of images, we only count the pair once. Additionally, to further measure MLLMs' performance, we also evaluate the precision of its prediction by manually labelling if each detected change is true, that is if the text is actually describing a change and the change align with the corresponding evidence images, as well as its recall by manually labelling if each human labelled change is covered in MLLMs' answers. + +The results in Tab. 1 show that the MLLM (Gemini) is significantly better than the two methods based on embedding distances - i.e., by more than 50 AP percentage points. Overall, we find that the MLLM is remarkably good at this task - its precision is $81.34\%$ and its recall is $89.87\%$ . + +# 4.3. Trend Discovery Evaluation + +Trend discovery is not trivial for MLLMs. We first evaluate a trivial solution of feeding all change descriptions to MLLMs and asking for trends, to the extent this is possible. Current context limits can support at most 50K changes at once, out of 3M changes per city (less than $2\%$ ). We therefore subsample sets of 50K changes and take the union over multiple sets, but the answers are not satisfactory. Similar to feeding subsampled image collections to MLLMs, only + +
RepresentationTypeScalableAverage Precision
Random——47.70%
CLIPImgEmb✓54.78%
NV-EmbTxtEmb✓73.13%
GeminiMLLM×86.63%
+ +Table 2. Evaluating Trend Discovery. The MLLM discovers trends with significantly higher average precision (AP) than baselines based on embedding distances, but is not scalable due to expensive inference. + +
MethodAllTrueNV-EmbRandMLLMHybrid (Ours)
Acc.@5072.7%77.9%31.8%93.9%
Acc.@10054.1%69.6%49.9%94.6%
Acc.@20028.9%81.8%74.9%98.3%
+ +Table 3. Evaluating Change-Trend Classifiers. AllTrue is predicting positive for all. Our hybrid algorithm is significantly more accurate at verifying change-trend pairs than scalable alternatives. + +very few trends can be identified and they are dominated by extremely frequent trends, e.g. scaffolding and construction. See the supplementary for more details. + +MLLMs are good at verifying trends, but are not scalable. Next, we test MLLM trend verification capabilities. We sample 50 cluster proposals each with 40 change candidates to verify. This makes 2000 trend-change pairs for the binary classification task "does this change belong to this trend?" We use human labellers to manually label these pairs. To ensure more balanced label distribution, we over-label more randomly sampled pairs and subsample 2000 pairs to balance positives and negatives. We compare MLLMs with SOTA image embedding (CLIP) and text embedding (NV-Emb), as well as a trivial solution that produces uniformly random binary predictions. As in the previous test, we use average precision as the metric. Our results in Tab. 2 show that the MLLM (Gemini) significantly outperforms the methods based on embeddings, by a margin more than $13\%$ . However, MLLMs are expensive in inference time. Even with 64 MLLMs running in parallel, it would take 380 days to verify 200 trend proposals against 3M changes (0.6B queries at 3.5s/query/MLLM), while embedding approaches can finish in less than a day with 64 A100 GPUs. + +Our hybrid solution is both accurate and scalable. At last, we verify performance of our hybrid solution. We know that our solution reduces the required MLLM inference cost by $2000 \times$ , making it sufficiently fast to verify trends, e.g. 4.6 hours to verify 200 trends with the same 64 parallel MLLMs. So, the remaining question is whether it can discover trends accurately. To test this, we sample 1000 trend proposals, each with 5000 changes to verify against. We use an MLLM to exhaustively verify every trend-change pair, 5M in total, and label each trend proposal as true or false, based + +on whether the MLLM confirmed at least $N$ changes for it (shown as $@N$ in Tab. 3). We then test our hybrid method against with this labeled set of trends to measure how accurately it classifies them as true or false, and report the binary classification accuracy. We compare our hybrid solution with: (1) A trivial approach that always predicts true, (2) A pure text embedding approach (NV-Emb) where we use a threshold to binarize trend-change similarity and use the binarised label to decide if a trend proposal is true. We report the best accuracy among 1000 grid-searched thresholds, and (3) Randomly selecting changes to verify (RandMLLM). Unlike our text embedding sorted verification, this solution randomly picks the same amount of changes as our solution to verify for each trend proposal. In Tab. 3, we find that, compared to other scalable alternatives, our hybrid solution is significantly more accurate. It does not degrade much the accuracy from exhaustive MLLM inference while being much more efficient. For the hyper-parameter $k$ in our solution, we use $k = 3N$ and ablate the choice of $k$ in Supplementary. + +# 5. Applications + +In this section, we apply our system to study visual trends spanning a decade in two major US cities, San Francisco (SF) and New York City (NYC). For each city, we collect 20M Street View images, captured from 2011 to 2023. For NYC, the images are sampled from 8.8M different locations, with an average of 23 images per location. For SF, we sample from 9.5M locations, with an average of 21 images per location. Local change detection yields 2.9M change entries in NYC and 3.6M in SF. For each city, we verify the most detailed 500 trend proposals. + +# 5.1. What are the trends of the decade? + +We first show the results of an open-ended attempt to discover "what changes were frequently observed over a decade?" in each city. Unlike most prior work, these results are produced without any preconditioning. That is, no external input was provided to guide the system to find specific types of changes, to avoid finding commonly occurring trivial changes (time of day, seasons, etc.), and to avoid semantically irrelevant changes (e.g., presence of distractors). The notion of what types of trends are worth reporting is determined by the MLLM during our local change detection and trend discovery processes. + +Fig. 3 and Fig. 4 show a sampling of the visual trends found in NYC and SF, respectively (more appear in the Supplementary). For each trend, we can provide hundreds of image pairs as supporting visual evidence. + +Yet, we still must be careful when drawing conclusions. For example, in Fig. 4, it appears many new solar panels were added near the raised freeway. In reality, it may be that roof-mounted solar panels are more likely to be visible in an image from a raised freeway than from ground level. We + +![](images/33e05d8c856f019079980e7fce7ce7636257f1174c47feeb79e054aa7b485cff.jpg) +"Security cameras added" (×745) + +![](images/25a79ded2ed75b4db201a16236e19be8492b6a9626b28351a35948b0cf3e891a.jpg) + +![](images/94d88f4cd19303c76a7ea423dc503e2734c3c0ae04cac20839583868560d4085.jpg) + +![](images/24d20983b834b2b076afb8362bd5b2f78f505352ee44807fbf9521ff80e14f9a.jpg) + +![](images/fa2f75b5f245ab4614d3250d3da629076fe87180a4c3719f52373528cbcc3a25.jpg) +Figure 3. Trends of the Decade in New York City. We discover interesting trends of visual change in NYC, including 745 new security cameras, 509 new fences around parking lots, and 519 red ADA warning pads on sidewalks. For each example, the top row shows images before a change, bottom row shows after, and markers appear on the map where the changes occurred. + +![](images/05cf2ba2b434eaf3f5403151e0c3a90b84780742285becb25020d65bcbaabd86.jpg) +"Parking lots fenced" (×509) + +![](images/d6cd1bd3fd5e575347f14939bf8a4b52afc6cfdaa0ab936e3383f9d7e7baaae3.jpg) +"New marking" (A×519) + +![](images/8b2738a3d1a328e665898b179491044c7c62d567fe7fbf13adf8538d41abb2f2.jpg) +Figure 4. Trends of the Decade in San Francisco. In San Francisco, we discover 1504 new solar panels (mostly in residential areas), 751 bus lanes conversions (mostly along a few major roads), and 1799 new bike racks (mostly near downtown). + +![](images/dafcb18a5f393c95e35e17377e4a8425cb858a5d21ce0ad0fe9503d1825aeea6.jpg) + +![](images/9c941126d8eaf2ebb30c2319afd1aee36b2fb05bb9baf34b06e74790c93cc8fd.jpg) + +![](images/0ffdf5376cdb05fc2a3f5582a56d90227762ed9a8e55592d2977f130ac642103.jpg) + +![](images/5f69be4c7fc534e87e1068260daa41377e57a477a54deb777d9f3f8a154c1151.jpg) + +![](images/aef857bfceb251a5b865cb0f8011eb2b64c2677d878f52a729d69580163fb439.jpg) + +![](images/9aa1f58be705fa06c21d79e904bb224d886e1fb2120ba6ca4fd8d2e01a07f94d.jpg) + +must be careful to account for spatial and temporal image sampling biases when reviewing the results. Despite this, we believe our system can be useful for discovering what changes happened, and where they happened, in an evidence-driven way that previously would not have been possible without an enormous amount of human effort. + +# 5.2. What happened in SF since 2020? + +Our system also supports searching for visual trends regarding a specific sub-period of interest. To demonstrate, we showcase a time-conditioned search for the range 2020-01-01 to 2022-12-31. Here, we filter to changes that happened strictly within this period, i.e., both images (before and after) are captured within this period, and we rerun only the trend discovery step. To encourage the system to find trends that might be unique to post-2020, we use a simple heuristic to sort trend proposals by their potential frequency differences + +between post-2020 years (2020–2022) and pre-2020 years (2017–2019). We conduct this study on data from SF. + +Outdoor dining. One notable trend is "The storefront added tables and chairs outside", indicating the addition of outdoor dining, as shown in Fig. 1. We observe this occurrence 1,482 times in 2020-2022 and 668 times in 2017-2019. + +An Overpass Painted Blue. To our surprise, Fig. 1 shows the visual trend "The support of the overpass was painted blue," observed 481 times since 2020 (5 times before 2020). An internet search reveals the likely explanation—San Francisco settled on 'Coronado Blue' for its $31 million paint job of the Central Freeway (SF Standard reporting). + +# 5.3. Trends related to retail stores in NYC + +Visual Chronicles also supports searches for trends relevant to a specific subject. We provide an example study of re + +![](images/a95c2e1c46180f5f8fd239af93e54dab1f8779be61e9bddf3ec8ed4a91144b63.jpg) +Figure 5. NYC Retail Visual Trends. By conditioning on "a retail store has changed," we are able to discover visual trends related to the opening and closing of types of retail stores in New York. + +tail stores in NYC. Similar to Sec. 5.2, we can re-use the local changes detected in Sec. 5.1, and this time filter them by semantic relevance. To compute semantic relevance to retail stores, we first embed the sentence, "A retail store has changed". Then, similar to our hybrid trend verification solution, we pick the nearest $500\mathrm{K}$ changes in NYC and use an MLLM to check that they are indeed changes relevant to retail stores. After filtering, we run the trend discovery stage, except we verify 150 trends uniformly across trends of varying word count (rather than the lengthiest trends), hoping to find a variety of trends about retail stores. + +Bakeries and Juice Shops. Seen in Fig. 5, one type of trend that arises within this search is the opening of various shops, including "There is a bakery newly opened at the storefront" and "A juice shop opened at the storefront". + +Groceries and Bank Branches. Opposite to the opening of certain stores, the closing of certain stores also happened in 2011-2023 (Fig. 5). Among our trends, we find that many grocery stores and bank branches closed in NYC. + +# 5.4. A Non-Temporal Query: "Unusual Things" + +Our system also supports non-temporal open-ended queries. As one example, we investigated searching for "unusual things" that actually happen a lot in a city. To do so, we simply changed the first step of our system to send just a single image to the MLLM and ask it "what are the unusual things in the image?." The MLLM then produces text describing the local unusual things, which will be processed by the second trend discovery step without change. Applying this + +![](images/39407a9bef146f6cc96aad352e01ed357bb192c39bd589d9c83937a2d8b46792.jpg) +Figure 6. Unusual Things on the Streets of NYC. Beyond temporal queries, our system can also analyze individual images, for non-temporal open-ended queries. Here we show findings from the query "What are the unusual things in the image?" We find there are many unusual "large, abstract" sculptures in NYC. + +![](images/417dccc5311fb2b2e7e33d5eaa2a7f0e2aed2af8bf8fd31ac3ec5c0be4dbce15.jpg) + +process to a large set of images in NYC yields a cluster of images showing "a large, abstract sculpture" (Fig. 6), among other interesting clusters. + +# 6. Conclusions, Limitations, and Future Work + +This paper explores how to use MLLMs to analyze a massive image collection without labels to find trends with open-ended semantic queries. In addition to introducing this novel task, we carefully design a novel system effectively leveraging the strong capabilities of MLLMs for large scale analyses. Experiments show the proposed system is able to discover fascinating trends automatically. + +This paper is the first to analyze image datasets with millions of images using MLLMs, and so it should be considered an initial proof-of-concept, rather than an exhaustive analysis of all possible design choices. Accordingly, there are many limitations. One key avenue for future work is to integrate MLLM analysis over large-scale, distributed imagery into a robust statistical framework that takes into account biases, e.g., due to uneven sampling of imagery and inaccuracies or bias in MLLM results — our work is a scalable way to generate hypotheses from vast amounts of visual data, but the methods we devise could also be developed into tools that allow for more rigorous conclusions, grounded in evidence. Another fruitful direction would be to apply our methods to other kinds of large visual datasets, including video (e.g., archival news footage), and to other kinds of queries, like “find interesting styles.” In any case, we are likely near the beginning of a long line of research related to using MLLMs for analyzing massive image collections, and this paper takes a small step in that direction. + +Acknowledgements. Thanks to Jiahui Lei, Anh Thai, Jiapeng Tang, Linyi Jin, Luming Tang, Rundi Wu, Ian Huang, Colton Stearns, Francis Engelman, Manu Gopakumar, Suyeon Choi, Haley So, Richard Tucker, Abhijit Kundu, Jonathan Barron, Glenn Entis, and David Salesin, for their comments and constructive discussions; to Abhijit Kundu, William Freeman, and John Quintero for helping review our draft; G.W. was in part supported by Google, Samsung, and Stanford HAI. B.D. was in part supported by a Qualcomm Innovation Fellowship. + +# References + +[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 3 +[2] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. NeurIPS, 2022. 3 +[3] Dragomir Anguelov, Carole Dulong, Daniel Filip, Christian Frueh, Stephane Lafon, Richard Lyon, Abhijit Ogale, Luc Vincent, and Josh Weaver. Google street view: Capturing the world at street level. Computer, 2010. 2 +[4] Oren Boiman and Michal Irani. Detecting irregularities in images and in video. *IJCV*, 2007. 2, 3 +[5] Shizhen Chang and Pedram Ghamisi. Changes to captions: An attentive network for remote sensing change captioning. IEEE TIP, 2023. 3 +[6] Eric Ming Chen, Jin Sun, Apoorv Khandelwal, Dani Lischinski, Noah Snively, and Hadar Averbuch-Elor. What's in a decade? transforming faces through time. In Computer Graphics Forum, 2023. 2 +[7] Xi Chen, Xiao Wang, Soravit Changpinyo, AJ Piergiovanni, Piotr Padlewski, Daniel Salz, Sebastian Goodman, Adam Grycner, Basil Mustafa, Lucas Beyer, et al. Pali: A jointly-scaled multilingual language-image model. In ICLR, 2023. 3 +[8] Gabriela Csurka, Tyler L Hayes, Diane Larlus, and Ricardo Volpi. What could go wrong? discovering and describing failure modes in computer vision. arXiv preprint arXiv:2408.04471, 2024. 3 +[9] Théophile Dalens, Mathieu Aubry, and Josef Sivic. Bilinear image translation for temporal analysis of photo collections. IEEE TPAMI, 2019. 2 +[10] Boyang Deng, Richard Tucker, Zhengqi Li, Leonidas Guibas, Noah Snavely, and Gordon Wetzstein. Streetscapes: Large-scale consistent street view generation using autoregressive video diffusion. In SIGGRAPH 2024 Conference Papers, 2024. 2 +[11] Carl Doersch, Saurabh Singh, Abhinav Gupta, Josef Sivic, and Alexei A Efros. What makes paris look like paris? Communications of the ACM, 2015. 2, 5 +[12] Abhimanyu Dubey, Nikhil Naik, Devi Parikh, Ramesh Raskar, + +and César A Hidalgo. Deep learning the city: Quantifying urban perception at a global scale. In ECCV, 2016. 2 +[13] Lisa Dunlap, Yuhui Zhang, Xiaohan Wang, Ruiqi Zhong, Trevor Darrell, Jacob Steinhardt, Joseph E Gonzalez, and Serena Yeung-Levy. Describing differences in image sets with natural language. In CVPR, 2024. 3 +[14] Timnit Gebru, Jonathan Krause, Yilun Wang, Duyun Chen, Jia Deng, Erez Lieberman Aiden, and Li Fei-Fei. Using deep learning and google street view to estimate the demographic makeup of neighborhoods across the united states. Proceedings of the National Academy of Sciences, 2017. 1, 2, 5 +[15] Shiry Ginosar, Kate Rakelly, Sarah Sachs, Brian Yin, and Alexei A Efros. A century of portraits: A visual historical record of american high school yearbooks. In CVPR Workshops, 2015. 1, 2, 3 +[16] Kyle Heath, Natasha Gelfand, Maks Ovsjanikov, Mridul Aanjaneya, and Leonidas J Guibas. Image webs: Computing and exploiting connectivity in image collections. In CVPR, 2010. 2 +[17] Phillip Isola, Devi Parikh, Antonio Torralba, and Aude Oliva. Understanding the intrinsic memorability of images. NeurIPS, 2011. 2 +[18] Phillip Isola, Jianxiong Xiao, Antonio Torralba, and Aude Oliva. What makes an image memorable? In CVPR, 2011. 2 +[19] Neal Jean, Marshall Burke, Michael Xie, W Matthew Davis, David B Lobell, and Stefano Ermon. Combining satellite imagery and machine learning to predict poverty. Science, 2016. 2 +[20] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, 2021. 3 +[21] Sergey Karayev, Matthew Trentacoste, Helen Han, Aseem Agarwala, Trevor Darrell, Aaron Hertzmann, and Holger Winnemoeller. Recognizing image style. arXiv preprint arXiv:1311.3715, 2013. 2 +[22] Andrej Karpathy and Li Fei-Fei. Deep visual-semantic alignments for generating image descriptions. In CVPR, 2015. 3 +[23] Andrej Karpathy, George Toderici, Sanketh Shetty, Thomas Leung, Rahul Sukthankar, and Li Fei-Fei. Large-scale video classification with convolutional neural networks. In CVPR, 2014. 3 +[24] Aditya Khosla, Tinghui Zhou, Tomasz Malisiewicz, Alexei A Efros, and Antonio Torralba. Undoing the damage of dataset bias. In ECCV, 2012. 2 +[25] Hoeseong Kim, Jongseok Kim, Hyungseok Lee, Hyunsung Park, and Gunhee Kim. Agnostic change captioning with cycle consistency. In ICCV, 2021. 3 +[26] Pierre-Yves Laffont, Zhile Ren, Xiaofeng Tao, Chao Qian, and James Hays. Transient attributes for high-level understanding and editing of outdoor scenes. ACM TOG, 2014. 2 +[27] Ivan Laptev. On space-time interest points. IJCV, 2005. 3 +[28] Ivan Laptev and Tony Lindeberg. Interest point detection and scale selection in space-time. In International Conference on Scale-Space Theories in Computer Vision, 2003. 3 + +[29] Chankyu Lee, Rajarshi Roy, Mengyao Xu, Jonathan Raiman, Mohammad Shoeybi, Bryan Catanzaro, and Wei Ping. Nv-embed: Improved techniques for training llms as generalist embedding models. arXiv preprint arXiv:2405.17428, 2024. 5 +[30] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, 2023. 3 +[31] Weixin Liang, Yuhui Zhang, Hancheng Cao, Binglu Wang, Daisy Yi Ding, Xinyu Yang, Kailas Vodrahalli, Siyu He, Daniel Scott Smith, Yian Yin, et al. Can large language models provide useful feedback on research papers? a large-scale empirical analysis. NEJM AI, 2024. 3 +[32] Tony Lindeberg. On automatic selection of temporal scales in time-causal scale-space. In International workshop on algebraic frames for the perception-action cycle, 1997. 3 +[33] Mingxuan Liu, Zhun Zhong, Jun Li, Gianni Franchi, Subhankar Roy, and Elisa Ricci. Organizing unstructured image collections using natural language. arXiv preprint arXiv:2410.05217, 2024. 3 +[34] Ricardo Martin-Brualla, David Gallup, and Steven M Seitz. 3d time-lapse reconstruction from internet photos. In ICCV, 2015. 2 +[35] Ricardo Martin-Brualla, David Gallup, and Steven M Seitz. Time-lapse mining from internet photos. ACM TOG, 2015. 2 +[36] Kevin Matzen, Kavita Bala, and Noah Snavely. StreetStyle: Exploring world-wide clothing styles from millions of photos. arXiv preprint arXiv:1706.01869, 2017. 2 +[37] Andrew McCallum, Kamal Nigam, and Lyle H Ungar. Efficient clustering of high-dimensional data sets with application to reference matching. In ACM SIGKDD international conference on Knowledge discovery and data mining, 2000. 4 +[38] Farid Melgani, Gabriele Moser, and Sebastiano B Serpico. Unsupervised change-detection methods for remote-sensing images. Optical Engineering, 41(12):3288-3297, 2002. 2, 5 +[39] Nikhil Naik, Jade Philipoom, Ramesh Raskar, and Cesar Hidalgo. Streetscore-predicting the perceived safety of one million streetscapes. In CVPR workshops, 2014. 2 +[40] Nikhil Naik, Scott Duke Kominers, Ramesh Raskar, Edward L Glaeser, and Cesar A Hidalgo. Computer vision uncovers predictors of physical urban change. Proceedings of the National Academy of Sciences, 2017. 2, 3 +[41] Dong Huk Park, Trevor Darrell, and Anna Rohrbach. Robust change captioning. In ICCV, 2019. 3 +[42] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 3, 5 +[43] Machel Reid, Nikolay Savinov, Denis Teptyashin, Dmitry Lepikhin, Timothy Lillicrap, Jean-baptiste Alayrac, Radu Soricut, Angeliki Lazaridou, Orhan Firat, Julian Schrittwieser, et al. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context. arXiv preprint arXiv:2403.05530, 2024. 3, 5 + +[44] Xi Shen, Alexei A Efros, and Mathieu Aubry. Discovering visual patterns in art collections with spatially-consistent feature learning. In CVPR, 2019. 2 +[45] Ioannis Siglidis, Aleksander Holynski, A. Alexei Efros, Mathieu Aubry, and Shiry Ginosar. Diffusion models as data mining tools. In ECCV, 2024. 2 +[46] Chandan Singh, Jeevana Priya Inala, Michel Galley, Rich Caruana, and Jianfeng Gao. Rethinking interpretability in the era of large language models. arXiv preprint arXiv:2402.01761, 2024. 3 +[47] Gemini Team, Rohan Anil, Sebastian Borgeaud, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, Katie Millican, et al. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023. 3 +[48] Gemini Team, Petko Georgiev, Ving Ian Lei, Ryan Burnell, Libin Bai, Anmol Gulati, Garrett Tanzer, Damien Vincent, Zhufeng Pan, Shibo Wang, et al. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context. arXiv preprint arXiv:2403.05530, 2024. 3 +[49] Antonio Torralba and Alexei A Efros. Unbiased look at dataset bias. In CVPR, 2011. 2, 5 +[50] Du Tran, Heng Wang, Lorenzo Torresani, Jamie Ray, Yann LeCun, and Manohar Paluri. A closer look at spatiotemporal convolutions for action recognition. In CVPR, 2018. 3 +[51] Nanne Van Noord. Prototype-based dataset comparison. In ICCV, 2023. 2 +[52] Oriol Vinyals, Alexander Toshev, Samy Bengio, and Dumitru Erhan. Show and tell: A neural image caption generator. In CVPR, 2015. 3 +[53] Michael Xie, Neal Jean, Marshall Burke, David Lobell, and Stefano Ermon. Transfer learning from deep features for remote sensing and poverty mapping. In AAAI, 2016. 2 +[54] Linli Yao, Weiying Wang, and Qin Jin. Image difference captioning with pre-training and contrastive learning. In AAAI, 2022. 3 +[55] Lihi Zelnik-Manor and Michal Irani. Event-based analysis of video. In CVPR, 2001. 3 +[56] Zeliang Zhang, Mingqian Feng, Zhiheng Li, and Chenliang Xu. Discover and mitigate multiple biased subgroups in image classifiers. In CVPR, 2024. 3 +[57] Ruiqi Zhong, Peter Zhang, Steve Li, Jinwoo Ahn, Dan Klein, and Jacob Steinhardt. Goal driven discovery of distributional differences via language descriptions. NeurIPS, 2023. 3 +[58] Ruiqi Zhong, Heng Wang, Dan Klein, and Jacob Steinhardt. Explaining datasets in words: Statistical models with natural language parameters. arXiv preprint arXiv:2409.08466, 2024. 3 +[59] Jun-Yan Zhu, Yong Jae Lee, and Alexei A Efros. Averageexplorer: Interactive exploration and alignment of visual data collections. ACM TOG, 2014. 2 +[60] Zhiying Zhu, Weixin Liang, and James Zou. Gsclip: A framework for explaining distribution shifts in natural language. arXiv preprint arXiv:2206.15007, 2022. 3 \ No newline at end of file diff --git a/ICCV/2025/Visual Chronicles_ Using Multimodal LLMs to Analyze Massive Collections of Images/images.zip b/ICCV/2025/Visual Chronicles_ Using Multimodal LLMs to Analyze Massive Collections of Images/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..7598a4f43a7add2635b3840a2d9a75553715c6f3 --- /dev/null +++ b/ICCV/2025/Visual Chronicles_ Using Multimodal LLMs to Analyze Massive Collections of Images/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0d58fd4a4f8891c3a56397bc1a12f9c7995ad8549ad6737df08a168315761e02 +size 609186 diff --git a/ICCV/2025/Visual Chronicles_ Using Multimodal LLMs to Analyze Massive Collections of Images/layout.json b/ICCV/2025/Visual Chronicles_ Using Multimodal LLMs to Analyze Massive Collections of Images/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..783c50bbdd4befc7a5d5df99170e01b5fe72ec32 --- /dev/null +++ b/ICCV/2025/Visual Chronicles_ Using Multimodal LLMs to Analyze Massive Collections of Images/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:104daf24a1e4c0c20544b85d2ae5e6485f47f44e31e53dc64d19ee332cf306ac +size 398744 diff --git a/ICCV/2025/Visual Intention Grounding for Egocentric Assistants/eedc0963-0f97-46ed-a0a7-79c0f04d3f23_content_list.json b/ICCV/2025/Visual Intention Grounding for Egocentric Assistants/eedc0963-0f97-46ed-a0a7-79c0f04d3f23_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..98896c428c0e326bcb77d56eea1b3d124fac8627 --- /dev/null +++ b/ICCV/2025/Visual Intention Grounding for Egocentric Assistants/eedc0963-0f97-46ed-a0a7-79c0f04d3f23_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8100e5e5079c9269512e7397e8e57e6fa3ca96e432aae6823dcc672483dfa1aa +size 80567 diff --git a/ICCV/2025/Visual Intention Grounding for Egocentric Assistants/eedc0963-0f97-46ed-a0a7-79c0f04d3f23_model.json b/ICCV/2025/Visual Intention Grounding for Egocentric Assistants/eedc0963-0f97-46ed-a0a7-79c0f04d3f23_model.json new file mode 100644 index 0000000000000000000000000000000000000000..5b76dc61c368f2b2dfab1260928256a8fb59c2bc --- /dev/null +++ b/ICCV/2025/Visual Intention Grounding for Egocentric Assistants/eedc0963-0f97-46ed-a0a7-79c0f04d3f23_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:26267d42643a29ea2c9064e86231cc1b1bea6c9cad0140b6c62524f120f0b6a3 +size 103873 diff --git a/ICCV/2025/Visual Intention Grounding for Egocentric Assistants/eedc0963-0f97-46ed-a0a7-79c0f04d3f23_origin.pdf b/ICCV/2025/Visual Intention Grounding for Egocentric Assistants/eedc0963-0f97-46ed-a0a7-79c0f04d3f23_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..d03d665805a3fbcf46f73c4ea143057065f31f14 --- /dev/null +++ b/ICCV/2025/Visual Intention Grounding for Egocentric Assistants/eedc0963-0f97-46ed-a0a7-79c0f04d3f23_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d0fa53f8e7c81a1907871ba7649698d2c09b6cbbc5a3681313dc8f7c425ef413 +size 715908 diff --git a/ICCV/2025/Visual Intention Grounding for Egocentric Assistants/full.md b/ICCV/2025/Visual Intention Grounding for Egocentric Assistants/full.md new file mode 100644 index 0000000000000000000000000000000000000000..78b88f4312df25f859e5347230f367efe52e78a4 --- /dev/null +++ b/ICCV/2025/Visual Intention Grounding for Egocentric Assistants/full.md @@ -0,0 +1,318 @@ +# Visual Intention Grounding for Egocentric Assistants + +Pengzhan Sun $^{1}$ Junbin Xiao $^{1}$ * Tze Ho Elden Tse $^{1}$ Yicong Li $^{1}$ Arjun Akula $^{2}$ Angela Yao $^{1}$ $^{1}$ National University of Singapore $^{2}$ Google DeepMind {pengzhan, junbin, eldentse, ayao}@comp.nus.edu.sg, liyicong@u.nus.edu, arjunakula@gmail.com + +![](images/55018a6d595e1a7d4cc5790eca572e466bde20d72fff6039e39a903314bf0a5f.jpg) +Conventional Visual Grounding +Query: white chair +Figure 1. Traditional visual grounding (left) vs. egocentric visual intention understanding (center and right). Traditional grounding identifies the "white chair" by detecting specific objects from third-person perspectives. Egocentric visual intention understanding must infer user needs in complex, first-person scenarios, e.g., seating in a workshop (center) or using a chair to reach the sink (right). + +![](images/3c408381b0729ad48b80921b33f63f61fa0f2a7883058de32867a8870293cfe1.jpg) +Egocentric Visual Intention Grounding +Query: It would be helpful to have somewhere to sit while I organize all these tools and supplies. + +![](images/79594228998240f971f03a6317453a5e54d176a1c67a499a6fc133f4bfa385aa.jpg) +Query: I need a little boost to reach the sink. + +# Abstract + +Visual grounding associates textual descriptions with objects in an image. Conventional methods target third-person image inputs and named object queries. In applications such as AI assistants, the perspective shifts – inputs are egocentric, and objects may be referred to implicitly through needs and intentions. To bridge this gap, we introduce EgoIntention, the first dataset for egocentric visual intention grounding. EgoIntention challenges multimodal LLMs to 1) understand and ignore unintended contextual objects and 2) reason about uncommon object functionalities. Benchmark results show that current models misidentify context objects and lack affordance understanding in egocentric views. We also propose Reason-to-Ground (RoG) instruction tuning; it enables hybrid training with normal descriptions and egocentric intentions with a chained intention reasoning and object grounding mechanism. RoG significantly outperforms naive finetuning and hybrid training on EgoIntention, while maintaining or slightly improving + +naive description grounding. This advancement enables unified visual grounding for egocentric and exocentric visual inputs while handling explicit object queries and implicit human intentions. Our code and model are available at https://github.com/pengzhansun/EgoIntention. + +# 1. Introduction + +Consider the following scenarios: a person is looking for a place to sit down for organizing tools in a messy workshop, or a kid is trying to reach the sink in a kitchen (refer to Figure 1). A wearable artificial intelligent (AI) assistant could enhance these tasks by identifying contextually relevant objects (e.g., a chair) without explicit object references. To achieve this, such an assistant must possess strong egocentric visual perception capabilities [11, 37, 43, 51]. This would significantly improve task efficiency, reduce cognitive load, and support hands-free, context-aware interaction in dynamic environments. + +Building on this vision, we introduce the egocentric visual intention grounding task. Given an egocentric visual + +Figure 2. Challenge of Visual Intention Grounding. The model must infer the intended object from the full intention sentence, rather than simply detecting explicitly mentioned objects. In this example, "gather my phone and belongings" explicitly mentions "phone" (highlighted in red), which often misleads existing visual grounding models to identify the wrong object (red box). The correct target, a handbag (green box), is only implied. +![](images/4d35320ccf2435c06bb734521d7e61e455eb169556dd8685872d4363dfa47e72.jpg) +Please locate an object from the image that match with this intention: I need to gather my phone and belongings before heading out. + +input and a human intention query, a model must accurately localize the intended object within the scene. This task supports real-world applications where users locate objects based on their needs. The object may not be directly named but can be inferred from the user's intention. Additionally, egocentric AI assistants operate from a first-person perspective, introducing challenges such as occlusions and dynamic viewpoints not present in conventional third-person vision systems. Unlike conventional visual grounding tasks [27, 30, 47], such as referring expression comprehension [9, 10, 16, 28, 29, 38, 39], this task requires reasoning about object affordance beyond explicit mentions of the object. As shown in Figure 2, existing models often misinterpret explicit mentions and fail to infer the actual intended object, highlighting the need for contextual reasoning in visual intention grounding. + +Despite recent advancements in vision-language models [4, 6, 8, 17, 19, 31, 35, 36, 46, 48, 50], existing methods [1, 3, 5, 26, 40, 41] struggle to solve visual intention grounding in diverse real-world scenarios. First, solving the visual intention grounding task using two separate models, such as a language model (e.g., GPT-4 [1]) for reasoning and an object detector (e.g., GroundingDINO [23]) for localization, yields suboptimal results. These off-the-shelf models process visual and language information independently, often resulting in inconsistent associations between the two modalities. Consequently, a model may hallucinate objects by incorrectly identifying items that are not present in the scene. Secondly, while multimodal large language models (MLLMs) offer a unified approach, existing methods are primarily designed for third-person visual grounding tasks. Without datasets connecting egocentric visual data to inten + +tion sentences, these models struggle to adapt to and perform well on first-person intention grounding. + +To bridge the gap in egocentric visual intention grounding, we introduce EgoIntention. EgoIntention is a comprehensive dataset built upon Ego4D [11], the largest real-world egocentric vision dataset. We inherit object bounding boxes annotation from PACO-Ego4D [33], a dataset that annotates object parts and attributes, and augment them with carefully curated human intention descriptions. Additionally, we address the inherent subjectivity in intention-object relationships by incorporating supplementary bounding box annotations for alternative objects that could reasonably fulfill each stated intention. Our EgoIntention dataset addresses two main gaps in visual grounding research: 1) the lack of egocentric data, and 2) the limited coverage of complex intentions arising from first-person viewpoints. Comprising 26,384 images, 52,768 human intention descriptions, and 89,841 annotated target object bounding boxes, this comprehensive dataset establishes a robust foundation for advancing visual intention understanding in real-world egocentric applications. + +Despite providing a rich benchmark for egocentric intention grounding, EgoIntention also exposes significant challenges for existing models. We identify two major limitations. First, multimodal models rely on explicit object-centric prompts (e.g., "Where is the white chair?"), directly mapping object names to their locations. However, intention-based queries require implicit reasoning. For example, when assisting in a workshop, the model must infer that the user needs a chair for sitting, rather than simply detecting a chair. Second, models must reason beyond direct object matches, recognizing alternative objects based on affordances. In the case of a child trying to reach the sink, the model should identify the chair as a suitable support, even though the query does not explicitly mention "chair." This requires a deeper understanding of object functionality and context, which is a challenge for current MLLMs. + +To address the above challenges, we propose a model-agnostic Reason-to-Ground (RoG) instruction tuning approach. RoG disentangles intention understanding from object grounding to enhance multimodal models' reasoning capabilities. By doing so, RoG reduces spurious correlations between unintended object mentions and the actual intended object locations. On EgoIntention, our RoG instruction tuning improves MiniGPTv2's performance by 3.9 Precision@0.5 compared to naive finetuning and significantly outperforms the off-the-shelf GPT-4 [1] + GroundingDINO pipeline by 12.2 Precision@0.5. + +Our contributions can be summarized as follows: + +1. We construct the EgoIntention dataset for visual intention grounding. The dataset is the first egocentric visual grounding dataset with multiple intention queries. +2. We benchmark and reveal that existing MLLMs struggle with intention reasoning and egocentric visual grounding. + +Table 1. Comparison of intention-related visual grounding datasets. + +
Dataset#ImagesLanguage QueryEgo-viewMulti-intention annotations
ADE-Aff [7]10,000Verb××
PAD [25]4,002Verb××
COCO-Tasks [34]39,724Phrase××
RIO [32]40,214Template language×✓
IntentionVG [42]101,648Free-form language✓×
EgoIntention26,384Free-form language✓✓
+ +These models misinterpret explicitly mentioned objects as targets or hallucinate objects not present in the scene. + +3. We propose Reason-to-Grounding (RoG) instruction tuning, a model-agnostic training approach to enhance MLLMs for egocentric intention grounding while retaining their performances for normal visual grounding. + +# 2. Related Work + +# 2.1. Visual Grounding Datasets + +Visual grounding [27, 30, 47] is a multimodal task that locates a target object in an image based on a given language query. Early works focused on referring expression comprehension [9, 10, 16, 28, 29, 38, 39], which matches descriptive phrases to objects within an image. Datasets such as RefCOCO [16], RefCOCO+ [47], and RefCOCOg [27] have played key roles in advancing this field. More recently, the scope of language input has been widened to encompass descriptions of object affordance [2, 18, 45]. This evolution led to datasets using verbs or phrases, (e.g., ADE-Aff [7], PAD [25], COCO-Tasks [34]) and full sentences in datasets like RIO [32] and IntentionVG [42]. + +Compared to IntentionVG, where egocentric images are captured with a fixed viewpoint centered on objects, our dataset is constructed from Ego4D [11], introducing greater visual challenges such as motion blur, small object sizes, and perspective distortions inherent to first-person vision. For language queries, our dataset provides multiple intention sentences per object, reflecting the diverse ways an object can be used to fulfill different needs. Similar to the RIO dataset, we annotate each sample with both a context sentence, describing an object's typical use in its expected environment, and an uncommon sentence, which represents a less conventional use case requiring creative object substitution. + +# 2.2. Visual Grounding Models + +Traditional visual grounding methods [20, 21, 24, 27, 44] are specialized models explicitly trained to map language queries to object locations. Models such as MDETR [14], SeqTR [52], and Polyformer [22] leverage Transformer + +based architectures to enhance this association. GroundingDINO [23], a recent advancement in open-set object detection, extends DINO [49] with grounded pre-training, allowing it to detect arbitrary objects given category names or referring expressions. Recently, multimodal large language models (MLLMs) have emerged as the dominant paradigm for vision-language tasks [1, 3, 5, 15, 26, 40, 41]. By leveraging vast amounts of image-text data and instruction tuning, these models achieve impressive generalization across various multimodal benchmarks. While effective in conventional visual grounding tasks, both specialist models and MLLMs primarily perform word-by-word detection on the input language query, rather than truly understanding the underlying human intention. To address these challenges, we propose Reason-to-Ground (RoG), a method that explicitly disentangles intention reasoning from object localization, as detailed in method Section 4. + +# 3. Intention Grounding & EgoIntention Dataset + +# 3.1. Task Description + +We introduce visual intention grounding, a novel paradigm that establishes a direct mapping between human intentions and target objects in visual scenes. This approach bridges the gap between natural language understanding and visual object grounding. Formally, given an egocentric image $I$ and a human intention query $Q$ , the task requires the model to comprehend the underlying intention and localize the target object $O$ that satisfies the user's need by predicting its bounding box coordinates $(x_{1},y_{1},x_{2},y_{2})$ . + +# 3.2. Dataset Collection + +Our dataset, EgoIntention, sources its images from the Ego4D dataset [11] and its associated object bounding box annotations from PACO [33]. The key contribution of EgoIntention is the collection of multiple intention queries per object, reflecting the diverse ways objects are used in real-world scenarios. Our systematic data collection pipeline consists of three stages, as illustrated in Figure 3. + +The initial stage generates intention descriptions. We use GPT-4 [1]'s multimodal capabilities to analyze egocentric visual inputs and generate contextually relevant human intention sentences. To capture diverse real-world scenarios, we design two types of intention descriptions: + +- Context-aware intentions: These sentences reflect complex object relationships and environmental cues from a first-person perspective. For instance, a user might think, "I noticed a wobbly table leg that needs fixing," expressing the need for a hammer. + +- Uncommon intentions: These describe atypical object uses, where users repurpose objects based on necessity. For example, a backpack might be used as an improvised umbrella during unexpected rain. + +![](images/6ef7298f857c421c721561e46426ceb6dc8033215b4adbb2fc2c13c7eaf1baed.jpg) +Ground truth object and bounding boxes + +![](images/bb39d3b23e3051245dcf2c7f6fc28a88d718e7dbfdd680ab93d8cc0a315c24b1.jpg) +Step 1. +ChatGPT +Prompt + +![](images/fa7fdf0e354315bfeb0ccbf593690e812b5193b5a9ac850d590eb4625b07e251.jpg) +Sentence +candidate set +Human check +instructions +Step 2. +MTurk + +Qualified sentence: I need to cut some herbs for the recipe + +![](images/c999f4c983bdd6de1cbcd4989894456842cb5b3464b0a1da03b076e04d69273b.jpg) +Step 3. +MTurk + +![](images/322d761e0c396f5ba37d6ef4bf012421d27c2cefeeeb04b639ef5bfdc4f9d305.jpg) +Alternative bounding box +collection instructions +Collected objects and bounding boxes +Figure 3. Dataset collection pipeline for EgoIntention. Our data collection process consists of three key stages. (1) Intention Sentence Generation: We use GPT-4 to generate egocentric human intention sentences based on visual input, covering both context-aware and uncommon intention scenarios. (2) Human Validation via MTurk: Annotators assess the semantic validity and real-world applicability of generated sentences, filtering out low-quality or ambiguous descriptions. (3) Alternative Object and Bounding Box Collection: Given the inherent subjectivity of human intentions, additional valid object candidates are identified by human annotators, supplementing the original ground truth annotations with alternative bounding boxes. + +These sentences are generated using carefully crafted prompts through an in-context learning approach as shown in Appendix. + +The second stage involves human validation via Amazon Mechanical Turk (MTurk). Annotators assess each generated sentence using detailed guidelines to ensure semantic correctness and real-world plausibility. Since multiple valid intentions may exist for a given scenario, we do not use inter-annotator agreement as a filtering criterion. Instead, one annotator selects and, if necessary, refines a suitable sentence. This selection is then verified by a second annotator and a GPT-4-based checker, both of which independently assess whether the sentence expresses a genuine need for the target object. Only those passing both checks are retained. + +To account for the subjectivity of human intentions, we include additional object annotations. A single intention may be satisfied by various objects (e.g., flower pots, bottles, or cups for desktop decoration). Annotators identify such alternatives, which are added as supplementary bounding boxes. As before, only those passing both human and GPT-4 verification are included. + +We observed that GPT-4 performs more reliably as a verifier than as a generator. While context-aware sentences passed human checks at $97.2\%$ , uncommon intentions had a lower pass rate of $74.1\%$ , often due to generic outputs or object name leakage. However, as a verifier, GPT-4 achieved $92\%$ agreement with human judgments in a 500-sample study, supporting its role as an auxiliary checker. To further validate our pipeline, we conducted an inter-annotator study on 500 verified samples, yielding agreement rates of $98.6\%$ for context-aware and $91.2\%$ for uncommon intentions. The maintained or slightly improved RefCOCO performance after joint training further supports the quality of our annotations. + +# 3.3. Dataset Statistics + +
ImageContext BBoxUncommon BBox
Train15,66725,77225,933
Val8251,4021,366
Test9,89217,69917,669
+ +Table 2. Number of images and bounding boxes (BBox) in the EgoIntention dataset. Context BBox refers to bounding boxes associated with objects commonly used in the given scene, aligning with expected human intentions based on environmental cues. Uncommon BBox represents bounding boxes for objects used in unconventional ways, where the intended action requires creative or atypical object usage. + +Our EgoIntention dataset builds upon PACO's image splits, comprising 15,667 training, 825 validation, and 9,892 test samples. For each image, we annotate two types of intention queries: context-aware intentions that leverage environmental cues, and uncommon intentions that capture alternative object uses. Following our multi-stage annotation pipeline, we enrich the dataset with supplementary bounding box annotations to accommodate the inherent diversity of object choices for each intention. The distribution of bounding box annotations across different splits and intention types is summarized in Table 2. + +# 4. Method + +This section begins with an observational study of off-the-shelf hybrid models in Section 4.1, highlighting key limitations in their reasoning and detection pipeline. Based on these findings, we propose Reason-to-Ground (RoG) to enhance multimodal large language models for visual intention + +grounding in Section 4.2. We then detail our supervised fine-tuning approach in Section 4.3. + +# 4.1. Observation of Off-the-Shelf Hybrid Models + +A straightforward approach to visual intention grounding is to leverage two off-the-shelf models separately for reasoning and detection. Specifically, we use ChatGPT for intention reasoning and GroundingDINO for object detection. From our observation study (detailed in Section 5), we identify two key findings: Performing reasoning first improves accuracy. Narrowing the search space before detection leads to more precise object localization. Hybrid off-the-shelf models suffer from inconsistent representations. These models operate in distinct visual and language spaces, causing discrepancies that hinder effective grounding. Motivated by these findings, our main method introduces Reason-to-Ground, an instruction tuning approach for multimodal large language models. + +# 4.2. Reason-to-Ground Instruction Tuning + +We propose a novel training strategy, Reason-to-Ground Instruction Tuning (RoG). Our method decomposes the task into two essential components: human intention understanding and visual grounding. Existing approaches [3, 5] directly feed implicit intention sentences with a task-specific token , which prompts the MLLM to output the bounding box corresponding to the input language query. In contrast, RoG employs a two-stage process. In the first stage, we facilitate human intention understanding by querying the MLLM with a token followed by the implicit intention sentence. This prompts the model to output the target object category. In the second stage, we perform token and the explicit object description derived from the first stage. These two stages are illustrated with an example in Figure 4. + +By disentangling intention reasoning and object grounding, RoG prevents the model from naively associating explicitly mentioned objects with target bounding boxes, leading to more accurate intention-driven visual grounding. + +# 4.3. Supervised Fine-tuning + +We propose a unified grounding framework that can process exo- and egocentric visual inputs while handling explicit object queries and implicit human intentions. To achieve this comprehensive capability, we leverage traditional exocentric datasets (RefCOCO [16], RefCOCOg [27], and RefCOCO+ [47]) alongside our proposed EgoIntention dataset during training. Our method RoG effectively extends MLLMs' visual grounding capabilities to accommodate egocentric image inputs and implicit human intention queries through LoRA [12] supervised fine-tuning. + +![](images/010788b787d8ba5a5017004b897536d49b6aac89ac9c9bf1c8d04b7956af5edd.jpg) +Figure 4. Overview of Reason-to-Ground Instruction Tuning (RoG): The model first infers an explicit object category from an implicit intention sentence (intention reasoning), then localizes the object in the image (object grounding). + +# 5. Experiments + +# 5.1. Implementation Details + +We conducted supervised finetuning using a comprehensive dataset that combines traditional referring expression comprehension datasets (RefCOCO, RefCOCO+, and RefCOCOg) with our proposed visual intention grounding dataset, EgoIntention. For our experiments, we evaluated five state-of-the-art multimodal large language models (MLLMs), all with 7B parameters: MiniGPTv2, Groma, CogVLM-grounding, and Qwen-VL. All model training was performed on 4 NVIDIA A40 GPUs. To maintain computational efficiency while preserving model performance, we employed Low-Rank Adaptation (LoRA) [12] for parameter-efficient finetuning. + +# 5.2. Models + +# 5.2.1. Hybrid Model Baselines + +A straightforward approach to visual intention grounding is to use off-the-shelf state-of-the-art models to separately address the vision and language components of the task. For vision, we adopt GroundingDINO, a popular open-set object detector that accepts a set of object category queries and returns the corresponding bounding boxes. For language understanding, we use ChatGPT, a large-scale language model + +developed by OpenAI, to reason about the user's intention and infer the target object category. + +As detailed in Section 4, we implement this two-stage framework in two variants: (1) first GPT-4 then GroundingDINO (Reasoning-Detection baseline, R-D), and (2) first GroundingDINO then GPT-4 (Detection-Reasoning baseline, D-R). + +D-R baseline. We begin by applying GroundingDINO to detect a comprehensive set of candidate object bounding boxes in the egocentric image. To ensure broad coverage, we use all object categories provided by the PACO dataset as text queries. GroundingDINO returns bounding boxes along with associated logits and predicted phrases. We then pass these predicted phrases, along with the human intention sentence, to GPT-4. GPT-4 performs reasoning over the textual descriptions to identify the object category most aligned with the expressed intention. + +R-D baseline. In contrast, this variant begins with GPT-4. Given the egocentric image and a prompt listing all candidate object categories, GPT-4 reasons about the intention and outputs the predicted object category. This output is then used to filter the query space for GroundingDINO, which subsequently predicts the bounding box for the inferred object by focusing only on the categories identified through GPT-4's reasoning. + +# 5.2.2. Multimodal Large Language Models + +We evaluate our approach on several state-of-the-art MLLMs, categorized into two groups based on their specialized capabilities: + +Grounding Specialist MLLMs excel at visual grounding tasks including grounded captioning, referring expression generation, referring expression comprehension, and grounded visual question answering: + +- CogVLM-grounding incorporates a trainable visual expert module within the attention and FFN layers, effectively bridging the gap between the frozen pretrained language model and image encoder. +- Groma demonstrates exceptional region-level grounding capability by integrating region tokens into both user instructions and model responses, enabling precise localization of described language queries. + +Compared to grounding specialists, Generalist MLLMs exhibit enhanced reasoning abilities—a critical capability for solving visual intention grounding tasks: + +- MiniGPT-v2 provides a unified interface for numerous vision-language tasks, employing unique task identifiers during model training to facilitate multi-task learning. +- Qwen-VL, built upon the Qwen-LM foundation, transcends conventional image description and question-answering capabilities by incorporating robust visual grounding functionality. + +# 5.3. Evaluation Metrics + +Our evaluation accounts for multiple alternative ground-truth boxes. We calculate each sample's score by computing the IoU between the predicted box and all ground truth boxes and then take the maximum IoU. A prediction is correct if its IoU with any ground truth box exceeds a threshold of 0.3 or 0.5 (Precision@0.3 or @0.5). + +# 5.4. Zero-shot Evaluation + +We first test all the 6 above mentioned methods in a zero-shot setting in Table 3. Our experiments reveal that the order of using the reasoner and detector significantly impacts results, despite the D-R and R-D baselines using the same models. For visual context-aware intention understanding, the D-R baseline correctly understands and accurately detects only $21.1\%$ of samples. In contrast, the R-D baseline improves $\mathrm{P@0.5}$ accuracy to $46.6\%$ . We attribute this difference to GroundingDINO's limitations when processing numerous language queries. When used first, it must handle all the object categories from the EgoIntention dataset and omits many object candidates. However, by using GPT-4 for initial reasoning, we narrow the target objects to one or two categories. This focused input allows GroundingDINO to perform $25\%$ better than in the D-R baseline. For uncommon intention understanding, we observe a similar trend. The R-D baseline $(23.6\%)$ outperforms the D-R baseline $(14.6\%)$ , but $\mathrm{P@0.5}$ improves by only $9\%$ . This smaller improvement stems from GPT-4's lower reasoning accuracy for uncommon intentions, failing to provide GroundingDINO with the correct object category for detection. + +Grounding-specialist MLLMs, such as CogVLM-grounding and Groma, perform poorly on EgoIntentions. This can be attributed to human intention reasoning being crucial for solving the visual intention grounding task. However, these specialist models are primarily aligned with grounding-related tasks and lack the necessary reasoning capability to infer which object should be grounded. Consequently, they fail to identify the intended object, leading to suboptimal results. In contrast, generalist MLLMs such as MiniGPT-v2 and Qwen-VL demonstrate more reasonable grounding performance when given an intention query. Despite reasoning capabilities learned from pretraining and alignment, generalist MLLMs' overall performance remains limited. MiniGPT-v2 achieves an overall Precision@0.5 score of $19.9\%$ , while Qwen-VL reaches $21.7\%$ ; both are significantly lower than the R-D GPT4-GroundingDINO baseline $(35.1\%)$ . This performance gap highlights the importance of integrating strong intention reasoning with visual grounding for improved results. + +# 5.5. RoG Supervised Finetuning Results + +In addition to the EgoIntention training set, we incorporate RefCOCO, RefCOCO+, and RefCOCOg (RefCOCO+/g) to + +Table 3. Benchmark comparison on EgoIntention across various methods, including two-stage pipeline approaches, grounding-specialist MLLMs, and generalist MLLMs. Performance is evaluated using Precision@0.3, Precision@0.5, and mIoU, reported for both the Context Split and Uncommon Split. The Overall P@0.5 metric summarizes the general performance across splits. + +
MethodContext SplitUncommon SplitOverall P@0.5
P@0.3P@0.5mIoUP@0.3P@0.5mIoU
D-R GroundingDINO-GPT430.121.10.21720.614.60.15017.8
R-D GPT4-GroundingDINO54.346.60.40231.723.60.24235.1
CogVLM-grounding8.05.90.0576.34.90.0425.4
Groma9.67.40.0748.96.90.0707.2
MiniGPT-v231.321.70.22425.518.00.18619.9
Qwen-VL27.921.40.22527.322.00.22821.7
+ +Table 4. Comparison of instruction tuning methods on RefCOCO, RefCOCO+, RefCOCOg, and EgoIntention datasets. Naive SFT refers to LoRA-based supervised fine-tuning, while RoG SFT represents LoRA-based fine-tuning with our Reason-to-Ground Instruction Tuning (RoG) instruction tuning approach. We report Precision@0.5 as the evaluation metric across different validation and test splits. + +
ModelRefCOCORefCOCO+RefCOCOgEgoIntention
valtestAtestBvaltestAtestBvaltestcontextuncommonoverall
Zero-shot MiniGPTv287.491.383.779.085.172.883.584.121.718.019.9
Naive SFT86.691.083.079.084.972.082.684.246.040.943.4
RoG SFT87.891.484.079.885.473.884.385.249.944.747.3
Zero-shot Qwen-VL89.392.485.483.288.277.285.385.621.422.021.7
Naive SFT89.592.885.783.488.877.885.986.332.126.129.1
RoG SFT89.392.585.383.388.877.486.286.435.531.733.6
+ +maintain model performance on referring expression comprehension while also leveraging these datasets to enhance intention grounding performance. As a result, our training data consists of three components: RefCOCO+/g, RefCOCOIntention+/g, and EgoIntention. The RefCOCOIntention+/g dataset is generated automatically using GPT-4, applying the same prompt used for collecting EgoIntention, to create human intention queries. + +While Naive SFT improves EgoIntention performance, it slightly degrades the model's capability in referring expression comprehension as shown in Table 4. In contrast, our RoG instruction tuning strategy not only further enhances performance on EgoIntention but also leads to improved results on the RefCOCO series datasets. After fine-tuning with our RoG strategy, the generalist MLLM MiniGPT-v2 surpasses the best off-the-shelf two-stage method (R-D GPT4-GroundingDINO) by $12.2\%$ according to Precision@0.5, demonstrating the effectiveness of our approach in bridging intention reasoning and grounding. Unlike R-D GPT4-GroundingDINO, which treats these as two independent sub-tasks, our method jointly models the reasoning and grounding processes, leading to more coherent and accurate results. Additionally, we observe that RoG SFT + +performs better on MiniGPT-v2 than on Qwen-VL. This discrepancy arises because MiniGPT-v2 exhibits stronger adaptability to new task-specific tokens, enabling it to integrate hierarchical reasoning instructions more effectively. In contrast, Qwen-VL demonstrates weaker instruction-following capabilities during supervised fine-tuning, limiting its performance gains under the same strategy. See visualization examples in Appendix. + +# 5.6. Ablation Study + +# 5.6.1. Impact of Supervised Finetuning Datasets + +We analyze the performance gains from different datasets used during the SFT stage in Table 5. Finetuning exclusively on EgoIntention hurts generalization, dropping RefCOCO validation performance from $87.4\%$ to $66.5\%$ . Combining RefCOCO/+/g with EgoIntention maintains REC performance while improving EgoIntention metrics. Specifically, context intention performance improves from $42.8\%$ to $45.9\%$ , and uncommon intention performance increases from $39.2\%$ to $40.8\%$ (see Table 5). This suggests that conventional REC datasets contribute additional gains, particularly in object grounding. + +We explore whether adding human intention annotations + +Table 5. Ablation study of training datasets used for MiniGPT-v2 fine-tuning and their impact on visual grounding datasets. + +
SFT DatasetsMethodRefCOCORefCOCO+RefCOCOgEgoIntention
RC+/gRCInt.+/gEgoInt.ValTestATestBValTestATestBValTestConUncoAve.Obj.
-0-shot87.491.383.779.085.172.883.584.121.718.019.940.8
✓Naive SFT87.691.384.480.085.373.884.885.223.719.421.538.1
✓Naive SFT66.571.560.460.266.651.864.665.842.839.241.046.2
✓✓Naive SFT87.591.584.679.985.673.584.785.445.940.843.348.6
✓✓✓Naive SFT86.691.083.079.085.072.082.684.246.040.943.451.3
✓✓✓RoG SFT87.891.484.079.885.473.884.385.249.944.747.352.2
+ +to REC datasets improves EgoIntention performance. We create RefCOCOInt $+ / \mathrm{g}$ by collecting intention sentences for RefCOCO $+ / \mathrm{g}$ training sets. This augmentation yields only slight improvements. Without human verification, data quality remains a bottleneck. These limited gains suggest high-quality annotations are crucial for advancing intention grounding. + +The best overall performance comes from applying our RoG strategy across all training datasets, demonstrating the effectiveness of our approach in jointly improving REC task and visual intention grounding task. + +# 5.6.2. RoG Improves Explicit Object Grounding + +As shown in Table 5, we further evaluate the model's performance on EgoIntention with explicit object queries, as indicated in the last column labeled "object". Compared to Naive SFT, RoG SFT disentangles visual intention understanding from object grounding, preventing the model from being misled by explicitly mentioned but unintended objects in the intention query. Our approach also leads to a significant improvement in egocentric visual grounding with explicit object queries, boosting the accuracy from $51.3\%$ with Naive SFT to $52.2\%$ with RoG SFT. + +# 5.6.3. Hallucination and Misleading Errors. + +Failure cases are due to hallucination (object is neither mentioned in the query nor present in the image) and object reasoning. The latter can be divided into misleading language (object mentioned in the query but not intended) and misleading vision (object appears in the image but is not intended). We use RAM++ [13] to extract all object tags in the image to check for misleading vision errors. RoG fine-tuned MiniGPTv2 achieves a lower error rate across three categories as shown in the table 6. + +Table 6. Error rates $(\%)$ for hallucination and reasoning failures across context and uncommon splits. + +
ModelContext SplitUncommon Split
Language MisleadVision MisleadHallucinationLanguage MisleadVision MisleadHallucination
Zero-shot3.48.430.16.97.558.2
Naive SFT1.46.423.43.97.544.0
RoG SFT2.13.812.02.46.624.3
+ +# 5.6.4. Uncommon Intention Coverage + +Uncommon intentions are inherently subjective and underexplored. Prior work like RIO [32] offers limited uncommon cases with 4,826 test samples only for testing and none for training. We curated five diverse, uncommon samples per object category as in-context learning GPT-4 prompts to construct the first large-scale training set for this setting. Table 7 shows that removing our uncommon split from training leads to a clear drop in Precision@0.5 on EgoIntention uncommon test set. Training with our uncommon intention data also improves zero-shot performance on RIO uncommon set for both naive and RoG SFT models. + +Table 7. Impact of training with uncommon intentions on EgoIntention and RIO benchmarks. + +
MethodsTraining w/ EgoInt. UncommonEgoInt. UncommonZero-shot on RIO Uncommon
Naive SFT-33.021.4
✓40.922.1
RoG SFT-33.723.8
✓44.726.0
+ +# 6. Conclusion + +We introduce egocentric visual intention grounding, where AI assistants infer and localize intended objects based on implicit human intentions. To support this research, we construct EgoIntention and benchmark state-of-the-art large vision-language models (LVLMs). Our results reveal that LVLMs struggle with implicit intention inference and egocentric visual grounding. For improvements, we propose Reason-to-Ground Instruction Tuning (RoG), a model-agnostic approach that disentangles intention reasoning from visual grounding, reducing spurious correlations and improving alignment with human intent. By applying RoG in supervised fine-tuning with hybrid data from normal visual grounding and intention grounding tasks, LVLMs retain strong performance on conventional visual grounding while achieving significant improvements in egocentric visual intention grounding, offering a promising approach for both object and intention queries in exo- and ego-centric visual environments. + +# References + +[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 2, 3 +[2] Shikhar Bahl, Russell Mendonca, Lili Chen, Unnat Jain, and Deepak Pathak. Affordances from human videos as a versatile representation for robotics. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13778-13790, 2023. 3 +[3] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond. arXiv preprint arXiv:2308.12966, 1(2):3, 2023. 2, 3, 5 +[4] Gongwei Chen, Leyang Shen, Rui Shao, Xiang Deng, and Liqiang Nie. Lion: Empowering multimodal large language model with dual-level visual knowledge. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26540-26550, 2024. 2 +[5] Jun Chen, Deyao Zhu, Xiaogian Shen, Xiang Li, Zechun Liu, Pengchuan Zhang, Raghuraman Krishnamoorthi, Vikas Chandra, Yunyang Xiong, and Mohamed Elhoseiny. Minigpt-v2: large language model as a unified interface for vision-language multi-task learning. arXiv preprint arXiv:2310.09478, 2023. 2, 3, 5 +[6] Keqin Chen, Zhao Zhang, Weili Zeng, Richong Zhang, Feng Zhu, and Rui Zhao. Shikra: Unleashing multimodal llm's referential dialogue magic. arXiv preprint arXiv:2306.15195, 2023. 2 +[7] Ching-Yao Chuang, Jiaman Li, Antonio Torralba, and Sanja Fidler. Learning to act properly: Predicting and explaining affordances from images. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 975-983, 2018. 3 +[8] Ziqing Fan, Cheng Liang, Chaoyi Wu, Ya Zhang, Yanfeng Wang, and Weidi Xie. Chestx-reasoner: Advancing radiology foundation models with reasoning through step-by-step verification. arXiv preprint arXiv:2504.20930, 2025. 2 +[9] Nicholas FitzGerald, Yoav Artzi, and Luke Zettlemoyer. Learning distributions over logical forms for referring expression generation. In Proceedings of the 2013 conference on empirical methods in natural language processing, pages 1914-1925, 2013. 2, 3 +[10] Dave Golland, Percy Liang, and Dan Klein. A game-theoretic approach to generating spatial descriptions. In Proceedings of the 2010 conference on empirical methods in natural language processing, pages 410-419, 2010. 2, 3 +[11] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, et al. Ego4d: Around the world in 3,000 hours of egocentric video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18995-19012, 2022. 1, 2, 3 +[12] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: + +Low-rank adaptation of large language models. ICLR, 1(2):3, 2022. 5 +[13] Xinyu Huang, Yi-Jie Huang, Youcai Zhang, Weiwei Tian, Rui Feng, Yuejie Zhang, Yanchun Xie, Yaqian Li, and Lei Zhang. Open-set image tagging with multi-grained text supervision. arXiv preprint arXiv:2310.15200, 2023. 8 +[14] Aishwarya Kamath, Mannat Singh, Yann LeCun, Gabriel Synnaeve, Ishan Misra, and Nicolas Carion. Mdetr-modulated detection for end-to-end multi-modal understanding. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1780-1790, 2021. 3 +[15] Liu Kangcheng, Xiao Junbin, Zhang Rui, Lv Hanqi, and Du Zidong. Bottom-up and top-down thoughts for visual intention grounding. In Proceedings of the 2025 International Conference on Multimedia Retrieval, page 889-898. Association for Computing Machinery, 2025. 3 +[16] Sahar Kazemzadeh, Vicente Ordonez, Mark Matten, and Tamara Berg. Referitgame: Referring to objects in photographs of natural scenes. In Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP), pages 787-798, 2014. 2, 3, 5 +[17] Yicong Li, Xiang Wang, Junbin Xiao, Wei Ji, and Tat-Seng Chua. Invariant grounding for video question answering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2928-2937, 2022. 2 +[18] Yicong Li, Na Zhao, Junbin Xiao, Chun Feng, Xiang Wang, and Tat-seng Chua. Laso: Language-guided affordance segmentation on 3d object. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14251-14260, 2024. 3 +[19] Zhaowei Li, Qi Xu, Dong Zhang, Hang Song, Yiqing Cai, Qi Qi, Ran Zhou, Junting Pan, Zefeng Li, Van Tu Vu, et al. Groundinggpt: Language enhanced multi-modal grounding model. arXiv preprint arXiv:2401.06071, 2024. 2 +[20] Yue Liao, Si Liu, Guanbin Li, Fei Wang, Yanjie Chen, Chen Qian, and Bo Li. A real-time cross-modality correlation filtering method for referring expression comprehension. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10880-10889, 2020. 3 +[21] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer vision-ECCV 2014: 13th European conference, zurich, Switzerland, September 6-12, 2014, proceedings, part v 13, pages 740-755. Springer, 2014. 3 +[22] Jiang Liu, Hui Ding, Zhaowei Cai, Yuting Zhang, Ravi Kumar Satzoda, Vijay Mahadevan, and R Manmatha. Polyformer: Referring image segmentation as sequential polygon generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18653-18663, 2023. 3 +[23] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 2, 3 +[24] Gen Luo, Yiyi Zhou, Xiaoshuai Sun, Liujuan Cao, Chenglin Wu, Cheng Deng, and Rongrong Ji. Multi-task collaborative + +network for joint referring expression comprehension and segmentation. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 10034-10043, 2020. 3 +[25] Hongchen Luo, Wei Zhai, Jing Zhang, Yang Cao, and Dacheng Tao. One-shot affordance detection. arXiv preprint arXiv:2106.14747, 2021. 3 +[26] Chuofan Ma, Yi Jiang, Jiannan Wu, Zehuan Yuan, and Xiaojuan Qi. Groma: Localized visual tokenization for grounding multimodal large language models. In European Conference on Computer Vision, pages 417-435. Springer, 2025. 2, 3 +[27] Junhua Mao, Jonathan Huang, Alexander Toshev, Oana Camburu, Alan L Yuille, and Kevin Murphy. Generation and comprehension of unambiguous object descriptions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 11-20, 2016. 2, 3, 5 +[28] Margaret Mitchell, Kees van Deemter, and Ehud Reiter. Natural reference to objects in a visual domain. In Proceedings of the 6th international natural language generation conference, 2010. 2, 3 +[29] Margaret Mitchell, Kees Van Deemter, and Ehud Reiter. Generating expressions that refer to visible objects. In Proceedings of the 2013 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 1174-1184, 2013. 2, 3 +[30] Varun K Nagaraja, Vlad I Morariu, and Larry S Davis. Modeling context between objects for referring expression understanding. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part IV 14, pages 792-807. Springer, 2016. 2, 3 +[31] Shraman Pramanick, Guangxing Han, Rui Hou, Sayan Nag, Ser-Nam Lim, Nicolas Ballas, Qifan Wang, Rama Chellappa, and Amjad Almahairi. Jack of all tasks master of many: Designing general-purpose coarse-to-fine vision-language model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14076-14088, 2024. 2 +[32] Mengxue Qu, Yu Wu, Wu Liu, Xiaodan Liang, Jingkuan Song, Yao Zhao, and Yunchao Wei. Rio: A benchmark for reasoning intention-oriented objects in open environments. Advances in Neural Information Processing Systems, 36, 2024. 3, 8 +[33] Vignesh Ramanathan, Anmol Kalia, Vladan Petrovic, Yi Wen, Baixue Zheng, Baishan Guo, Rui Wang, Aaron Marquez, Rama Kovvuri, Abhishek Kadian, et al. Paco: Parts and attributes of common objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7141-7151, 2023. 2, 3 +[34] Johann Sawatzky, Yaser Souri, Christian Grund, and Jurgen Gall. What object should i use?-task driven object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7605-7614, 2019. 3 +[35] Hao Shao, Shengju Qian, Han Xiao, Guanglu Song, Zhuofan Zong, Letian Wang, Yu Liu, and Hongsheng Li. Visual cot: Unleashing chain-of-thought reasoning in multi-modal language models. arXiv e-prints, pages arXiv-2403, 2024. 2 +[36] Wenhao Shi, Zhiqiang Hu, Yi Bin, Junhua Liu, Yang Yang, See-Kiong Ng, Lidong Bing, and Roy Ka-Wei Lee. Math + +llava: Bootstrapping mathematical reasoning for multimodal large language models. arXiv preprint arXiv:2406.17294, 2024. 2 +[37] Pengzhan Sun, Bo Wu, Xunsong Li, Wen Li, Lixin Duan, and Chuang Gan. Counterfactual debiasing inference for compositional action recognition. In Proceedings of the 29th ACM International Conference on Multimedia, pages 3220-3228, 2021. 1 +[38] Kees van Deemter, Ielka van der Sluis, and Albert Gatt. Building a semantically transparent corpus for the generation of referring expressions. In Proceedings of the fourth international natural language generation conference, pages 130-132, 2006. 2, 3 +[39] Jette Viethen and Robert Dale. The use of spatial relations in referring expression generation. In Proceedings of the Fifth International Natural Language Generation Conference, pages 59-67, 2008. 2, 3 +[40] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 2, 3 +[41] Weihan Wang, Qingsong Lv, Wenmeng Yu, Wenyi Hong, Ji Qi, Yan Wang, Junhui Ji, Zhuoyi Yang, Lei Zhao, Xixuan Song, et al. Cogvlm: Visual expert for pretrained language models. arXiv preprint arXiv:2311.03079, 2023. 2, 3 +[42] Wenxuan Wang, Yisi Zhang, Xingjian He, Yichen Yan, Zijia Zhao, Xinlong Wang, and Jing Liu. Beyond literal descriptions: Understanding and locating open-world objects aligned with human intentions. arXiv preprint arXiv:2402.11265, 2024.3 +[43] Junbin Xiao, Nanxin Huang, Hao Qiu, Zhulin Tao, Xun Yang, Richang Hong, Meng Wang, and Angela Yao. Egoblind: Towards egocentric visual assistance for the blind people. arXiv preprint arXiv:2503.08221, 2025. 1 +[44] Sibei Yang, Guanbin Li, and Yizhou Yu. Dynamic graph attention for referring expression comprehension. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4644-4653, 2019. 3 +[45] Tomoya Yoshida, Shuhei Kurita, Taichi Nishimura, and Shin-suke Mori. Text-driven affordance learning from egocentric vision. arXiv preprint arXiv:2404.02523, 2024. 3 +[46] Haoxuan You, Haotian Zhang, Zhe Gan, Xianzhi Du, Bowen Zhang, Zirui Wang, Liangliang Cao, Shih-Fu Chang, and Yinfei Yang. Ferret: Refer and ground anything anywhere at any granularity. arXiv preprint arXiv:2310.07704, 2023. 2 +[47] Licheng Yu, Patrick Poirson, Shan Yang, Alexander C Berg, and Tamara L Berg. Modeling context in referring expressions. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, pages 69-85. Springer, 2016. 2, 3, 5 +[48] Ao Zhang, Yuan Yao, Wei Ji, Zhiyuan Liu, and Tat-Seng Chua. Next-chat: An Imm for chat, detection and segmentation. arXiv preprint arXiv:2311.04498, 2023. 2 +[49] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel M Ni, and Heung-Yeung Shum. Dino: Detr + +with improved denoising anchor boxes for end-to-end object detection. arXiv preprint arXiv:2203.03605, 2022. 3 +[50] Hao Zhang, Hongyang Li, Feng Li, Tianhe Ren, Xueyan Zou, Shilong Liu, Shijia Huang, Jianfeng Gao, Leizhang, Chunyuan Li, et al. Llava-grounding: Grounded visual chat with large multimodal models. In European Conference on Computer Vision, pages 19-35. Springer, 2024. 2 +[51] Sheng Zhou, Junbin Xiao, Qingyun Li, Yicong Li, Xun Yang, Dan Guo, Meng Wang, Tat-Seng Chua, and Angela Yao. Ego-textvqa: Towards egocentric scene-text aware video question answering. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 3363-3373, 2025. 1 +[52] Chaoyang Zhu, Yiyi Zhou, Yunhang Shen, Gen Luo, Xingjia Pan, Mingbao Lin, Chao Chen, Liujuan Cao, Xiaoshuai Sun, and Rongrong Ji. Seqtr: A simple yet universal network for visual grounding. In European Conference on Computer Vision, pages 598-615. Springer, 2022. 3 \ No newline at end of file diff --git a/ICCV/2025/Visual Intention Grounding for Egocentric Assistants/images.zip b/ICCV/2025/Visual Intention Grounding for Egocentric Assistants/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..65bd40b118a5a9c2d98d49bf0e85bd908749625c --- /dev/null +++ b/ICCV/2025/Visual Intention Grounding for Egocentric Assistants/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7a24e4fad03f642691bd8512df0fdee6658440785957c580271021a2da5440a5 +size 440484 diff --git a/ICCV/2025/Visual Intention Grounding for Egocentric Assistants/layout.json b/ICCV/2025/Visual Intention Grounding for Egocentric Assistants/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..82d9525f3f60e33eab3dc532126304583a7b3418 --- /dev/null +++ b/ICCV/2025/Visual Intention Grounding for Egocentric Assistants/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:08d9586dffab80f2ff322f0256532eea29e1a3676a989152aa33cf128aa820e8 +size 370482 diff --git a/ICCV/2025/Visual Interestingness Decoded_ How GPT-4o Mirrors Human Interests/38264d21-1016-40ad-892a-9a9170d58abd_content_list.json b/ICCV/2025/Visual Interestingness Decoded_ How GPT-4o Mirrors Human Interests/38264d21-1016-40ad-892a-9a9170d58abd_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..0cd9f4dc93265697b3c29a9bf3a8f33efbd36780 --- /dev/null +++ b/ICCV/2025/Visual Interestingness Decoded_ How GPT-4o Mirrors Human Interests/38264d21-1016-40ad-892a-9a9170d58abd_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:37c867bfdfceae0c7fa2f0948125408cd58517c3964947a5ce4f400eb1fba526 +size 119793 diff --git a/ICCV/2025/Visual Interestingness Decoded_ How GPT-4o Mirrors Human Interests/38264d21-1016-40ad-892a-9a9170d58abd_model.json b/ICCV/2025/Visual Interestingness Decoded_ How GPT-4o Mirrors Human Interests/38264d21-1016-40ad-892a-9a9170d58abd_model.json new file mode 100644 index 0000000000000000000000000000000000000000..97265c33bb9d4dfe900adc5da745b0f4c9da33f2 --- /dev/null +++ b/ICCV/2025/Visual Interestingness Decoded_ How GPT-4o Mirrors Human Interests/38264d21-1016-40ad-892a-9a9170d58abd_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ee3d83cdf67bb6a6c6b40505d213f09e64ad5f701f25567c6241d2a529d5f245 +size 136976 diff --git a/ICCV/2025/Visual Interestingness Decoded_ How GPT-4o Mirrors Human Interests/38264d21-1016-40ad-892a-9a9170d58abd_origin.pdf b/ICCV/2025/Visual Interestingness Decoded_ How GPT-4o Mirrors Human Interests/38264d21-1016-40ad-892a-9a9170d58abd_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..88866e084a7eddd17c7286805cb6f7f8a7d0704e --- /dev/null +++ b/ICCV/2025/Visual Interestingness Decoded_ How GPT-4o Mirrors Human Interests/38264d21-1016-40ad-892a-9a9170d58abd_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dd5d583813119272729298339d9ef6e9e210bc3b4a59f06ba7ca75ef2ddfcebf +size 2521737 diff --git a/ICCV/2025/Visual Interestingness Decoded_ How GPT-4o Mirrors Human Interests/full.md b/ICCV/2025/Visual Interestingness Decoded_ How GPT-4o Mirrors Human Interests/full.md new file mode 100644 index 0000000000000000000000000000000000000000..f6d4c0e0022f45f5f29ac11447180352a8567222 --- /dev/null +++ b/ICCV/2025/Visual Interestingness Decoded_ How GPT-4o Mirrors Human Interests/full.md @@ -0,0 +1,369 @@ +# Visual Interestingness Decoded: How GPT-4o Mirrors Human Interests + +Fitim Abdullahu and Helmut Grabner +Zurich University of Applied Sciences, Switzerland + +{fitim.abdullahu, helmut.grabner}@zhaw.ch + +# Abstract + +Our daily life is highly influenced by what we consume and see. Attracting and holding one's attention – the definition of (visual) interestingness – is essential. The rise of Large Multimodal Models (LMMs) trained on large-scale visual and textual data has demonstrated impressive capabilities. We explore these models' potential to understand to what extent the concepts of visual interestingness are captured and examine the alignment between human assessments and GPT-4o's, a leading LMM, predictions through comparative analysis. Our studies reveal partial alignment between humans and GPT-4o. It already captures the concept as best compared to state-of-the-art methods. Hence, this allows for the effective labeling of image pairs according to their (commonly) interestingness, which are used as training data to distill the knowledge into a learning-to-rank model. The insights pave the way for a deeper understanding of human interest. Code and materials: https://github.com/fiabdu/Visual-Interestingness-Decoded + +# 1. Introduction + +Online media data continues to expand, making it increasingly challenging to deliver relevant and engaging content to users. A key aspect of this challenge is the concept of (visual) interestingness - capturing attention and influencing behavior, which dates back to Berlyne's work in 1949 [2]. On the other hand, a vast amount of online accessible media is scraped to empower the training of foundation models in a self-supervised manner. Large Multimodal Models (LMMs), especially Language-Vision Models like GPT-40 [32], encode human-like knowledge and perform impressively across tasks. While they can reliably categorize images or answer visual questions, their ability to recognize subjective concepts remains uncertain. + +This work investigates how well state-of-the-art LMMs capture the fuzzy concept of visual interestingness. We explore whether these models can identify features associated with interestingness and compare their assessments to human judgments through user studies. As illustrated in + +![](images/e07209e2f896531b027f01d6f8e4189dc71ab9e99c86383e35d645adf3fc45df.jpg) +Figure 1. LMMs, such as GPT-4o, encode human-like knowledge and perform well across various tasks. We explore image interestingness, a highly subjective concept, by examining consistent labeling between humans and LMMs and their level of agreement. Pairwise labeling accesses relative measures, which are used to train a rank model to finally assess an image's interestingness. + +Fig. 1, we focus on both (i) the alignment and divergence between human- and model-based evaluations and (ii) assessing LMMs' potential to reduce manual labeling effort by distilling knowledge from their internal representations. + +We focus on everyday images to make the task traceable, consistent with prior work [11, 12, 18]. By achieving high consensus among humans, we explicitly limit subjectivity in the task, aiming to identify images that resonate with a broader audience [1, 5, 16]. We balanced the dataset size with experimental feasibility to ensure traceability and manage computational costs. Despite this trade-off, our results remain significant, and the insights are valid. + +
ApproachDataDefinition of InterestingnessLabels & Computational Model
IJCV'21 [9] et al.images, image features and meta-dataexplicit human annotations, e.g., AMT“direct” supervised training using the human annotations
ECCV'24 [1]images (large scale from Flickr)user interaction, i.e., user favoritesunsupervised estimation of user agreement to identify commonly interesting images; used to train a regressor
proposedimages (large scale)LMM trained on a vast amount of (human-generated) datalabel pairwise images to capture local preferences, which are then combined into a global learning-to-rank model
+ +Table 1. Most related work has focused on explicit human annotation. However, obtaining reliable annotations at scale is challenging and costly. Recent work explores implicit information to obtain "real" data from photo-sharing platforms at scale. However, this data may be biased toward a particular application, and only positive intentions (i.e., favorites) are available. To overcome these limitations, our approach leverages knowledge represented in LMMs to distill a computational model for visual interestingness. + +To this, our main contributions are: + +- We present a dataset of 1,000 images (2,500 image pairs) with metadata and multiple verified labels from human annotators and three state-of-the-art LMMs. +- We propose a novel approach to estimating image interestingness without direct user input, which outperforms state-of-the-art models for everyday images. +- We analyze and compare human and LMMs annotations, highlighting areas of agreement and divergence, paving the way for better understanding human and machine interests. + +# 2. Related Work + +Interestingness is inherently ambiguous1. It is subjective, varying by context, observer, and individual background [2, 9, 18]. Yet, recent work suggests that certain image features appeal broadly, regardless of individual differences [1, 5, 16]. Constantin et al. [9] provides a comprehensive overview of computational approaches to visual interestingness. For instance, Flickr calculates an "interestingness" score to help users discover engaging content on its platform [4, 14]. Early research in this area primarily relied on classical machine learning and computer vision methods, often constrained by limited datasets [12, 17, 18]. As the field has grown, the advent of deep learning and more powerful computational resources enabled larger-scale studies and more complex models, e.g., [8]. + +However, unbiased data collection and consistent annotations at a large scale are challenging and costly. Abdul-lahu and Grabner [1] recently proposed a more progressive approach that explores indirectly labeled data from multiple Flickr users. "Interestingness" is defined by analyzing user engagement, and their approach aims to identify commonly appealing image characteristics directly from users' favorites. In our work, we aim to extend this idea and get an + +notations from the knowledge encoded into LMMs trained on a vast amount of (human-generated) data - see Tab. 1. + +Large Multimodal Models. The emergence of multimodal foundation models has transformed AI by combining computer vision and natural language processing within a unified framework [7, 13, 32, 40, 43]. This paradigm shift changes training approaches and enables broader, more flexible model applications, moving from discrete classification tasks to prompt-based interactions [15]. This flexibility is advantageous in question-answering tasks, where users may ask about unfamiliar or abstract concepts. GPT-4o achieves state-of-the-art performance on various visual perception benchmarks [31, 32]. + +Recent models like GPT-4o have been trained on vast amounts of web data, encoding extensive knowledge from different fields. Fine-tuning and distilling knowledge into smaller, more specific models is widely used; see [44] for a recent survey. Leveraging the general knowledge of LMMs, specific models are used to design automatic evaluators that mirror human performance. Applications include evaluating text and images (e.g., 3D models) [42], assisting graphic design [19], assessing fashion aesthetics [20], or measuring content appeal [6]. + +To align with human values, models are fine-tuned with supervision [34] or automatically [36]. However, their latent knowledge can go beyond what LMMs are explicitly taught [25, 28], potentially revealing both overlaps and gaps between human and machine understanding [23, 39]. Understanding these overlaps and gaps could help uncover new concepts and insights2, especially in subjective areas. + +Outline of the Paper. We aim to explore how the implicit knowledge encoded in LMMs relates to the concept of visual interestingness. Sec. 3 and Sec. 4 present consensus and agreement for single and relative image interestingness assessment, respectively. Sec. 5 uses automatically annotated image pairs from GPT-4o to train a computational model for predicting image interestingness within a learning-to-rank framework. Finally, Sec. 6 discusses the + +similarities and differences between humans and GPT-40, providing insights into what makes an image interesting. + +# 3. Single Image Interestingness Assessment + +# 3.1. Experimental Design + +Dataset. We chose images from the photo-sharing platform Flickr as these images capture a wide variety of content from diverse communities, including professionals and amateur users [1, 24]. To select representative and diverse images, we selected 1,000 images from the Flickr-User dataset [1], equally sampled according to their proposed commonly interestingness score. + +Human Annotations. We conducted our user study using Amazon Mechanical Turk. Human workers were instructed on evaluating image interestingness and describing their choice, with no right or wrong answers – it was their decision if an image matched their interest (intentionally kept very open). More specifically, a randomly selected image was shown along with the question “Is this image interesting to you?” with two response options: yes or no. Additionally, they were prompted to provide a brief explanation for their choice. Five Human Intelligence Tasks (HITs) were created for each of the 1,000 images. Each HIT contained one image, and workers were compensated $0.01 per completed task. A total of 258 unique workers participated, with each worker labeling images without seeing the same image more than once. + +For further analysis, we split the dataset $S = \mathcal{C}_H \cup \mathcal{D}_H$ . An image belongs to $\mathcal{C}_H$ if four or five participants $((\cdot)_H$ , for human) agree on their response, indicating a high level of consensus. Otherwise, it belongs to $\mathcal{D}_H$ , reflecting dissent. Additionally, we define the interestingness label $y_H$ as 1 if the majority finds the image interesting and 0 otherwise. + +LMM Annotations. The rapid development of foundation models is remarkable, especially considering how quickly new models are released. In our work, we use state-of-the-art LMMs, specifically OpenAI's GPT-4o [32] ('gpt-4o-2024-11-20'), Meta's Llama 3.2 [13] ('Llama-3.2-11B-Vision-Instruct'), , and DeepSeeks Vision-Language VL2 [43] ('DeepSeek-VL2-tiny'). All of which can take text-image inputs and produce text outputs. The models are given the same prompt as the human annotators (the exact prompt is provided at the beginning of the subsequent subsections), with each image being evaluated five times. Analog to human annotations, consistently labeled images belong to sets $\mathcal{C}_G$ (for GPT-4o), $\mathcal{C}_L$ (for Llama 3.2), and $\mathcal{C}_D$ (for DeepSeek) others to the corresponding dissent sets $\mathcal{D}_{\{G,L,D\}}$ . Annotation labels $y_{\{G,L,D\}} \in \{0,1\}$ are defined by majority vote. + +Asking for Explanations. Asking "why" is used twofold, (i) to ensure quality for the judgments (see discussion for Llama 3.2 and DeepSeek annotations in Sec. 4.1) + +![](images/30d02f5f736fd3f00110b6041477f661f583b866e81bc01e08bce31039c96ff3.jpg) +(a) Human: 5/5 yes; Explanations: "ILIKE", "It is good photo and interesting", "It is amazing", ... +GPT-4o: 5/5 yes; Explanations: "The vibrant sunset and scenic landscape create a captivating visual appeal," "The sunset and landscape create a visually stunning scene," "The sunset with radiant clouds over a vast field creates a captivating and serene scene," ... +Figure 2. Examples of images and corresponding responses from human annotators and GPT-4o. Almost all images are consistently labeled, most of them as interesting. + +![](images/5ec6b1a6474be4f15d7b7ef155cb944fa0da84d6b86f5d8e65c3f02421380084.jpg) +(b) Human: 5/5 yes; Explanations: "NEAT", "Thinking", ... +GPT-4o: 4/5 yes; Explanations: "The cluttered workshop with a person lounging hints at a story," "It depicts a unique juxtaposition of a living space and a workshop," "The image depicts a person relaxing in a cluttered garage, which might not appeal to everyone,"... + +
|Cx|yx=1A(H,x)(S)A(H,x)(Ch)
Human91.9 %99.9 %--
GPT-4o93.9 %95.3 %92.9 %93.6 %
Llama93.1 %99.8 %97.1 %98.3 %
DeepSeek76.2 %81.4 %75.3 %77.3 %
+ +Table 2. Consistency and agreement with human annotation for single image interestingness assessment. Unfortunately almost all images are consistently labeled as interesting. + +and (ii) to gain deeper insights about how humans and LMMs come to the particular conclusion whether the image is interesting (see Sec. 6). Examples are shown in Fig. 2. + +# 3.2. "Is This Image Interesting?" + +Prompt: Is this image interesting? Answer with one word (yes or no) without punctuation and in lowercase. Add a semicolon without space. Explain why in one sentence without going into detail. + +As summarized in Tab. 2, the annotations from humans and LMMs show high consistency (almost all images are in the respective consistent set $\mathcal{C}_x$ ). This indicates that humans and LMMs generally agree on the interestingness of images. Furthermore, almost all images in the respective sets are considered interesting ( $y_x = 1$ ), indicating that humans and LMMs find almost all images on which they agree to be interesting. These suggest that humans and LMMs actively look for something interesting when explicitly asked for it, leading to a predominantly positive response. + +The agreement $A^{(M,N)}(\mathcal{S}) \coloneqq \frac{1}{|S|} \sum_{i=1}^{|S|} \mathbb{I}(y_M^{(i)} = y_N^{(i)})$ between annotation $M$ and $N$ on set $\mathcal{S}$ measures how well annotations of humans and LMMs are aligned. Not surprisingly, as almost all images are interesting, the results indicate a high level of agreement between humans and LMMs. Furthermore, the agreement between human and LMM increases slightly when focusing on consistently labeled images in $\mathcal{C}_H$ . It seems that it is somewhat easier for the LMM to distinguish between interesting and uninteresting if the humans agree on this question. + +Notably, the DeepSeek model has significantly less consistency (among itself) and less agreement with humans. However, for humans and the other LMMs, some images fall into the dissenting sets. This inconsistency may arise because an image may not interest a broad audience. For example, in Fig. 2b, GPT-4o stated: "The image depicts a person relaxing in a crowded garage, which might not appeal to everyone." This response suggests that the model subjectively evaluates the content to determine its interestingness, as discussed in [1]. + +Key Insight. Responses from humans, GPT-4o, and Llama 3.2 are very consistent and aligned. Almost all images were deemed interesting, and a story was made up to support the decision. + +# 4. Relative Image Interestingness Assessment + +As demonstrated in the last sections, whether an image is interesting is hard to answer generally. Results on single images are rendered meaningless, as almost all images are consistently labeled as interesting by humans and state-of-the-art LMMs. Relative comparisons are more affordable and often used for similar judgments, e.g., [18, 41]. + +# 4.1. Experimental Design + +Dataset. We created image pairs based on the 1,000 images used previously. Each image was used in five different (random) pairs, resulting in 2,500 image pairs. + +Human Annotations. As in the previous experiment, Amazon Mechanical Turk was used to obtain human annotations. A randomly selected image pair was shown to a worker, who was asked: "Which image is more interesting to you?" The worker could click on their preferred image and was asked to briefly explain their choice. Consistent with the previous experiment, five HITs were created for each of the 2,500 image pairs. Each HIT contained one image pair, and workers were compensated \(0.01 per completed task. A total of 553 unique workers participated, with each worker labeling image pairs without seeing the same pair more than once. + +As above, the consistency of the answers is defined if four or more humans agree on the labeling $(\mathcal{C}_H)$ . Furthermore, let $y_{H}$ represent the human labels, where $y_{H} = 1$ + +![](images/f1988977201274566f5d1e250ace44844ec6f17c82b52b71441df3271f69573f.jpg) + +![](images/f28e9f794348a12494ff02a9637b13de6c6fcfc230fdd3a531333e8c12861e16.jpg) + +Human: 5/5 second; Explanations: "Love Ferraris!", "Very nice", "It looks good", ... + +GPT-40: 5/5 second; Explanations: "The vibrant color and modern design stand out more," "The modern, sleek design of the vehicle coupled with the vibrant color captures attention more effectively," "It's visually striking due to its modern design and vivid color," ... + +![](images/2d5c8a3b1455ef8aa221c7de4384e80bb1ad95de3ce4518c2443c778e26e3222.jpg) + +![](images/e5c79b4ec57f28c6aad52b35d2d3d9fac91d6c2a23d09ab4e2eceedc4e82d545.jpg) + +Human: 5/5 second; Explanations: "I like birds", "This bird is cute", "LOOKING NATURAL", ... + +GPT-40: 5/5 first; Explanations: "The intricate design and craftsmanship make it more visually engaging," "It showcases a unique and artistic representation," "The first image depicts a unique and intricate metallic insect sculpture, making it more visually striking,"... + +Figure 3. Image pairs illustrating instances where humans and GPT-4o agree and disagree. For example, humans and GPT-4o have differing opinions regarding images of insects and birds. At first glance, it may not be immediately evident that the insect image is a metallic sculpture, which could explain why people did not find it as interesting—humans may not give the image the same level of attention as a machine. + +if the majority prefers the first image and $y_{H} = 0$ if the majority prefers the second image. + +GPT-4o Annotations. In this study, two image inputs are used, which is supported by GPT-4o. + +Systematic Error. Even though the model allows multiple images as input, we have discovered a systematic error. Image pairs were presented twice to GPT-4o, and the images were swapped on the second run. For $36\%$ of the cases, GPT-4o always reported the second image as more interesting. Only $64\%$ (1,599 out of 2,500) of the GPT-4o annotation remained the same, independent of the image order. For subsequent experiments, only these image pairs were kept. Please note that this systematic error does not seem to be much correlated to the human consensus ( $56.3\%$ ) + +of image pairs in $\mathcal{C}_H$ and $47.6\%$ in $\mathcal{D}_H$ are error-free). + +GPT-4o Demographics. As interestingness is subjective, it would be nice to test different user groups based on their demographics automatically. We used the system prompt of GPT-4o for that purpose: "You are a [gender] from [continent] and between [age] and [age] years old." If one uses prominent images that have become naturalized in society for men or women, such as a blue car or a pink flower, GPT-4o responds differently. A car is more interesting for men and a pink flower for women, somehow capturing prejudices. However, this vanishes when using everyday images where this distinction is no longer so prominent. In a more extensive study involving 500 random image pairs, we used male or female for gender and North America or Africa for the continent, and a range of 25 to 34 and 45 to 54 years for age, respectively. Running all eight combinations, filtering out pairs with a systematic error, and combining the remaining pairs, we ended up with 116 image pairs. Unfortunately, the results were identical for all image pairs, regardless of gender, continent, or age specified. + +Llama 3.2 and DeepSeeks-VL2 Annotations. As Llama 3.2 [13] currently does not support multiple image inputs, we combined the two images into a single input for the model. However, this workaround did not yield reliable results. For instance, the model often selected the first image as more interesting while providing an explanation that referred to the second image. Similarly, DeepSeek's recent Janus Pro Model [7] does not support multiple image inputs. When combining images, the model consistently selected the second image as more interesting. While DeepSeek's Vision-Language V2 model do allow for multiple image inputs [43], the selection and descriptions often do not align with the actual content, exhibiting issues similar to those observed with Llama 3.2. + +Due to these inconsistencies, which result in an unfair comparison between these LMMs and GPT-4o, we limited our further analyses to GPT-4o. + +Key Insight: GPT-4o responds independently of the demographic tested; however, it has a significant systematic bias in favoring the second image over the first. + +# 4.2. "Which image is More interesting?" + +Prompt: Which of the two images is more interesting? Answer with one word (first or second) without punctuation and in lowercase. Add a semicolon without space. Explain in one sentence why you have chosen this image without going into detail. + +Unlike in the single-image study above, people's responses are less consistent in the paired-image experiment. Set $\mathcal{C}_H$ contains $56.3\%$ of the image pairs, indicating consensus in about half of them. GPT-4o exhibits much higher consistency, with $95.5\%$ of all image pairs in $\mathcal{C}_G$ . The overall agreement between GPT-4o and human annotations is + +$A^{(H,G)}(\mathcal{S}) = 66.2\%$ In case of human consensus, the agreement increases to $A^{(H,G)}(\mathcal{C}_H) = 73.8\%$ while decreases on the dissent set to $A^{(H,G)}(\mathcal{D}_H) = 56.5\%$ - as one would expect. Examples are depicted in Fig. 3. + +Key Insight. GPT-4o's annotations are aligned with human judgments, especially when there is consensus among people. + +# 4.3. Comparisons and Relation to other Approaches + +Our annotations are compared to other models and concepts related to visual interestingness prediction (c.f. [9]). All approaches provide a measurement, score, or probability per image, which (is claimed) to be related to the image's interestingness. An image pair is labeled according to which image yields the higher response. + +Aesthetics [22]. The VILA (Vision-Language Aesthetics) model learns image aesthetics by analyzing user comments alongside images. It models subjective aesthetic judgments by aligning visual features with language and categorizing images according to these learned aesthetics. + +Memorability [21]. A predictive model assigns memorability scores to images. Through large-scale experiments with memory recall tasks, quantified memorability is established as a stable metric across viewers. We use the recent re-implementation from [30]. + +Commonly Interesting Images (CI) [1]. Interests is subjective. However, some images appeal to a broader audience and are, therefore, of common interest. A predictive model was trained by analyzing how many unique Flickr users "favored" images from a certain category (i.e., visually similar images). + +Social Interestingness [11]. Whereas being related to visual interestingness, factors beyond image features are relevant to make an image go viral. Social interestingness metrics use the number of views, favorites, and comments of a post on a social media platform. Using Flickr images in our study, we directly sourced these values for every image. + +Zero shot learning [37]. We use Customized Prompts via Language (CuPL) to generate prompts to determine an image's interestingness. Using a Large Language Model, in our case GPT-40, to "Describe what an interesting image looks like", we get various prompts such as "An interesting image features vibrant colors, unexpected elements, and a captivating composition that draws the viewer's eye". Overall, 500 prompts are created following [37]. After removing duplicates and highly similar prompts, we ended up with 250 unique prompts. Text embeddings for these prompts using CLIP [38] are calculated and averaged. The final score is the cosine similarity between text and image CLIP embeddings. + +Results and Discussion. Results can be seen in Tab. 3 (left). Every approach is compared to human $(y_{H})$ and GPT-4o annotations $(y_{G})$ , in terms of agreement on the set + +
GroupModelAnnotations (Sec. 4)Learning to Rank (Sec. 5)
A(H,x)A(G,x)Acc.(H)rS(H)Acc.(G)rS(G)
HumanHuman-73.8 %77.5 ± 2.5 %-72.0 ± 3.4 %0.59 ± 0.06
LMMsGPT-4o73.8 %-73.4 ± 3.4 %0.59 ± 0.0684.8 ± 2.5 %-
CuPL60.3 %60.9 %61.5 ± 3.5 %0.34 ± 0.0763.2 ± 3.1 %0.42 ± 0.08
Computational +ModelsCI69.6 %67.6 %69.6 ± 3.6 %0.54 ± 0.0669.1 ± 3.3 %0.52 ± 0.06
Memorability35.5 %39.1 %34.7 ± 4.0 %-0.42 ± 0.0838.3 ± 3.6 %-0.34 ± 0.07
Aesthetic68.3 %75.1 %69.0 ± 3.7 %0.50 ± 0.0773.6 ± 3.7 %0.67 ± 0.06
Social Inter- +estingness#Views61.7 %63.9 %63.4 ± 3.4 %0.39 ± 0.0866.3 ± 3.4 %0.48 ± 0.08
#Favorites66.4 %74.0 %66.3 ± 3.2 %0.47 ± 0.0769.4 ± 3.1 %0.57 ± 0.07
#Comments68.0 %74.8 %66.6 ± 3.1 %0.46 ± 0.0770.2 ± 3.2 %0.58 ± 0.07
+ +Table 3. GPT-4o achieves the highest agreement $A^{(\cdot ,x)}$ among all models using human responses as the ground truth. It also outperforms existing models focused on visual interestingness, related concepts, and social metrics. On the right, we show the model's accuracy Acc. (·) on the image pairs. The global ranking (measured by the Spearman rank correlation $r_S^{(\cdot)}$ ) of the test dataset remains consistent, indicating that the learning-to-rank model generalizes beyond pairwise relationships for both human and GPT-4o annotations. + +$\mathcal{C}_H$ . GPT-4o is superior to previous models in this context, followed by models using aesthetic or common interestingness. Social interestingness scores reveal that images with more comments tend to be considered more interesting than those without, which aligns with research in that regard [10]. Memorability score has the weakest link to interestingness, consistent with prior findings [18]. When GPT-4o is used as the ground truth, the VILA aesthetic model performs the best, followed by the GPT-4o model, while the human model ranks third. It also appears to have a stronger agreement with social interestingness, possibly due to their pre-training. + +Key Insight. GPT-4o's annotations are superior to previously proposed approaches to predict human interest. + +# 5. Learning a Computational Model + +So far, image pairs have been annotated by humans, GPT-40, and various computational models. In this section, we distill this knowledge into a simple computational model. + +Learning-To-Rank. A simple learning-to-rank model can be implemented using a Siamese network architecture with shared weights [3]. As we are using images $\mathbf{I}_0$ and $\mathbf{I}_1$ as input, they are first embedded using $\mathrm{CLIP}^3$ , passing through a linear layer with a single neuron. The scoring function is the difference between them, passed through a sigmoid function:, i.e., $S(\mathbf{I_0},\mathbf{I_1})\coloneqq \sigma (\mathbf{w}^\top \mathrm{CLIP}(\mathbf{I_0}) - \mathbf{w}^\top \mathrm{CLIP}(\mathbf{I_1}))$ . Learning is done to maximize the score differences between pairs. Binary cross-entropy loss on the + +target $y \in \{0,1\}$ is used; $y = 1$ if $\mathbf{I_0}$ ranks higher than $\mathbf{I_1}$ and $y = 0$ otherwise. As the weights are shared, after training, a score can also be obtained using a single input $S(\mathbf{I}) = \sigma (\mathbf{w}^{\top}\mathrm{CLIP}(\mathbf{I}))$ . For multiple images, the individual scores are used to rank them. Besides its simplicity, this approach has been used successfully many times, also for distilling information from LLM or LMM, e.g., [6, 41] + +Training/ Testing. The dataset was split in half for training and testing. We train learning-to-rank models for all annotations (human and GPT-4o and approaches from Sec. 4.3). Each model was trained for 25 epochs, and no overfitting was observed. Each experiment was repeated 50 times with different training/test splits. The results are depicted in Tab. 3 (right). Acc. denotes the model accuracy on the individual image pairs (as trained) and $r_S$ the Spearman rank correlation on the global ranking based on the scores $S(\cdot)$ for each image. + +Results and Discussion. The best-performing model is obtained when training and test data are from the same source (human or GPT-4o). This serves as the baseline for comparing the other models. All the results match nicely with those from the previous sections (left side of the table) for the individual performance of labeling image pairs. + +According to human annotations, the model generalizes to unseen data, although the average accuracy of approximately $77.5\%$ may reflect the subjectivity of the task. GPT-4o achieves the best performance among the models, although a gap remains compared to the baseline. Other computational models, such as CI or aesthetics, perform well but still fall short of GPT-4o's results. Examining the Spearman correlations, we find that human responses positively correlate with all models except memorability, which aligns with current research findings. Notably, the correlation between humans and GPT-4o is 0.59, indicating a moderate positive + +relationship between these two variables. + +Based on GPT-4o annotations, the aesthetic model ranks highest in accuracy after GPT-4o itself, followed closely by the human model. The global ranking of the test dataset is inline, meaning that the learning-to-rank model can generalize beyond pairwise relationships for both human and GPT-4o annotations. + +Int10k Dataset. The Int10k dataset [9] focuses on video summarization, and there is a significant domain gap between single everyday images (our focus in this work) and cinematic image sequences. Nevertheless, we applied our approach. Overall, the results show a significant drop, with accuracy for the human model and human-provided annotations being around $59.3\% \pm 2.5\%$ . Even for this dataset, GPT-4o outperformed all other approaches, achieving a comparable accuracy of $59.2\% \pm 2.1\%$ . For more details, see the supplementary material. + +Key Insight. Computational models obtained from distilling the information from the annotations match them very well. GPT-4o is superior to previous models. + +# 6. Similarities and Differences in Assessment + +This section discusses the agreement and disagreement between humans and GPT-4o in more detail. + +Embeddings. We analyzed the responses from both GPT-4o and human annotators to create a semantic embedding using OpenAI's 'text-embedding-3-small' model (1536-dimensional vector for each text input). Since the human responses were predominantly short and repetitive – often consisting of simple terms like "nice", "beautiful", or "good" – they were less suitable for in-depth analysis (cf. Fig. 4). Therefore, we focused our analysis on the text embeddings from GPT-4o responses. + +Exploiting the "Why" responses. We perform hierarchical clustering on the embeddings. Several clusters emerged from the analysis, including the cute and emotional clusters shown in Fig. 4a, which are of interest to both humans and GPT-4o. While humans respond to certain images as cute, GPT-4o tends to associate them with emotions. Additionally, humans and GPT-4o demonstrate consensus regarding uniqueness. Nevertheless, there are clusters that only GPT-4o finds interesting, such as images featuring vibrant color or depicting action. These images are not necessarily captivating for humans, see Fig. 4b. Please note that these clusters represent semantically similar texts, not semantically similar images. + +Exploiting Image Appearance. To gain deeper insight into what makes an image interesting or uninteresting, we conducted an additional experiment focusing on the visual characteristics of the images. GPT-4o was asked to describe each image, and embeddings were calculated for those descriptions. Fig. 5a- 5b illustrates a hierarchical clustering of the description embeddings in the original space and a 2d + +![](images/0d78be4ea393e730904da0228370be3ff5de1b4b0c3a0dd9ea06eaf6f6c1a3d2.jpg) + +![](images/608389261e3887d4249a173817841cd2fb6afbd7d3c7c1941e510a96643db3c8.jpg) + +![](images/41d6c3147dd0eeae76b5266d30c594e7d8894e15dd7c114723ac89190e713699.jpg) + +![](images/90f8bf7a2c7b48e442a8ec7305d4ae646efd9181ee099a72d11340c59d0cf61a.jpg) + +Human: "Cute Girl", "This side is cute", "THIS IMAGE IS VERY BEAUTIFULL", "It looks interesting" + +GPT-40: "It captures a joyful moment," "It shows human interaction and activity," "The image of the baby animals captures a unique and lively moment," "The emotional connection between the two beings adds an engaging element." + +![](images/5149ff05d789ec8c6b65e9e2ae9b6edae556b024b51c11a7272dc12efc829dd9.jpg) + +![](images/97fb798098f880fcbe0a6c09ded7cb4825cccd0117c907852ed58411438e4a88.jpg) + +![](images/31ad3f560a8d93545327e5d2274697e9844d3e5b2a73b7d5197837775c2f3800.jpg) + +![](images/b2c739a053e96feceb55dbca1f2776806ea1849c2f5036a1bcfba11e95ffe91f.jpg) + +Human: "Right is better than left", "This building's design is very interesting", "Appreciate nature more", "What a two wheeler excited to ride immediately" + +GPT-40: "Features unique graffiti on a monument.", "The building's unique architecture makes it stand out.", "The image depicts a unique and intriguing structure situated in a picturesque and seemingly remote location.", "The vintage motorcycle has a unique and nostalgic appeal." + +(a) Clusters considered as "interesting" for humans and GPT-4o. + +![](images/132584ca363973c5c2478d52fe35dd14c8f3e218ba267a5214923f9f8ab28ed2.jpg) + +![](images/b38bbf3e4234bd48f7c770e0a7473bebb8561e4840fe8f85f2ba55d5e8a161ae.jpg) + +![](images/8d5ca3bba49843322e78042aef7fa3103fb04de1d7c7e612a47296293538d629.jpg) + +![](images/0ed915b9179561d1553e04bf3a30ef159a97dc0962d452b847d3948fa7fd1824.jpg) + +GPT-40: "The action and dynamic scene with soldiers makes it more interesting," "It has more vibrant colors and dynamic elements", "The arrangement and colors are visually appealing.", "The image shows a person performing music on the street which adds more dynamic and action." + +(b) Only "interesting" for GPT-4o (no human descriptions are available, as these images were never considered interesting in pairwise comparisons by human annotators). + +Figure 4. Explanations: The clusters are derived from text embeddings of responses "Why" an image is interesting. Clusters of common interest (a) include "cute", "joyful" moments as well as "uniqueness". A minority of clusters (b) shows disagreement between humans and GPT-4o. + +projection using UMAP [27]. The clustering reveals groups of semantically similar images consistently found interesting, uninteresting, or divisive by humans and GPT-4o, see Fig. 5c-5e. There is agreement on images depicting flowers, birds, or nature scenes as interesting. In contrast, humans and GPT-4o find images of mundane scenes or people in ordinary situations uninteresting. Images depicting people at events or performing art elicit mixed responses. + +Tab. 4 reports quantitative information per cluster, including agreement between humans and GPT-4o and if the + +![](images/00b6c1e316e8d3d1da8e8e04f1ed149e99ccb2add79397e51abdd51932442dd0.jpg) +(a) Hierarchical clustering reveals groups of semantically similar descriptions ( $d < 2$ ), each containing either exclusively interesting (red), uninteresting (blue), or a mix of differing images (green). + +![](images/b56799d7b62f8371fb2c93d3f279b0f78403ab03c7eb04985b77160a6271d6fe.jpg) +(b) 2d visualization of the description text, embeddings using UMAP. + +![](images/f796b66c78ba1a24e7decb70eb756c0369408757f8ce11a742b120573a9ccc29.jpg) +(c) Agreement on interestingness (red) + +![](images/61a1e56334ea6d3c3ba191911ea6317b77df56a51714ca0b8eca1dd4a2c169af.jpg) +Figure 5. Image Content Descriptions: Text embeddings of the image descriptions are divided into semantically similar groups using hierarchical clustering (a, b). Most clusters indicate agreement between humans and GPT-4 (c, d), while some indicate disagreement (e). Compare with Tab. 4. + +![](images/725f6524b5c063cace7967a2ed9fe644c733f5f94d2f15e43c2fd57271193664.jpg) + +![](images/6fc4d3c4d9b9973b3cb047507895523982989c4b36cee7182e45b427107e31d1.jpg) +(d) Agreement on uninterestingness (blue) + +![](images/a24fca339c8a429f0a23656dc100442346643f574bc46a54aeaeaacbdfcf41de.jpg) + +![](images/1e7a4e8824916c13719c0a95c3622176450a9eaf7ac49666853a40cd724fc221.jpg) + +![](images/0f1828a78b8a93e8f958e04a0f3392f9b3dfa68fdf1328bef3ef0b4b64edb4ac.jpg) +(e) Disagreement (green) + +![](images/78802ef1073d3d28e13824f1f136a65f1282efbe42fdad6efb869cbaa4243222.jpg) + +![](images/d661ba813007a0cc466b36660c212da804c545dfadcd05b6aa95e9e8c1d342c9.jpg) + +cluster is mainly interesting or uninteresting. Additionally, we calculated the mean ranks of images in each cluster based on both human $(\overline{R}^{(H)})$ and GPT-4o $(\overline{R}^{(G)})$ annotations. The mean ranks align closely, particularly when agreement is high. The top four words describe the cluster obtained after removing stop words and lemmatizing the automatically generated appearance descriptions. These words align well with examples from Fig. 5 (marked). + +Key Insight. Humans and GPT-4o generally agree well on what is interesting for topics and scenes. + +# 7. Conclusion + +GPT-4o cannot be used directly to assess image interestingness due to uninformative (almost always positive) responses on single images and systematic errors. However, we demonstrated that GPT-4o outperforms existing models in a comparative annotation setting. Its alignment with human assessments highlights its potential to support (i) large-scale studies and (ii) knowledge distillation. Further investigation into the gap between GPT-4o and human annotation, including demographic factors, would be insightful. Extending the study to datasets beyond Flickr images could provide a broader perspective. Together, these efforts will help refine our understanding of visual interestingness and its contributing factors. + +
A#Apos +#AR(H)R(G)Frequent Words (Appearance)
94%60%143184train, track, station, railway
86%23%246236people, group, front, standing
84%86%86107water, sky, background, body
82%97%5380perched, bird, branch, flower
81%24%217207person, sitting, room, window
81%59%165142building, modern, street, large
80%46%204193people, two, smiling, together
80%78%125131dog, lying, person, cat
78%88%7279tree, sky, landscape, water
77%18%223221various, small, featuring, ...
77%41%202191person, wearing, standing, red
72%44%212158person, people, background, ...
72%89%7491water, bird, swimming, white
71%61%197158people, group, person, flag
74%54%154185car, parked, red, background
65%72%95160flower, green, yellow, pink
47%71%207163playing, stage, performing, guitar
40%50%207163stage, guitar, person, group
+ +Table 4. Agreement of humans and GPT-4o concerning the image content; agreement $(A)$ on un- and interestingness, average ranks form models trained using human and GPT-4o annotations as well as disagreement along its clusters text descriptions. Color markings match Fig. 5. + +Acknowledgements. This research was funded by the Swiss National Science Foundation (SNSF) under grant number 206319 "Visual Interestingness - All images are equal but some images are more equal than others". + +# References + +[1] Fitim Abdullahu and Helmut Grabner. Commonly interesting images. In Proceedings of the European Conference on Computer Vision (ECCV), 2024. 1, 2, 3, 4, 5 +[2] Daniel E Berlyne. Interest as a psychological concept. British Journal of Psychology, 39(4):184, 1949. 1, 2 +[3] Chris Burges, Tal Shaked, Erin Renshaw, Ari Lazier, Matt Deeds, Nicole Hamilton, and Greg Hullender. Learning to rank using gradient descent. In Proceedings of the International Conference on Machine learning (ICML), pages 89-96, 2005. 6 +[4] Daniel S Butterfield, Caterine Fake, Callum J Henderson-Begg, and Serguei Mourachov. *Interestingness ranking of media objects*, USPTO #US8732175B2. 2 +[5] Fabienne Bünzli, Wibke Weber, Fitim Abdullahu, and Helmut Grabner. Depicting humans, animals, and objects in motion: The effect of implied motion on engagement and persuasion in advertising. Journal of Advertising, 0(0):1-21, 2024. 1, 2 +[6] Sherry X. Chen, Yaron Vaxman, Elad Ben Baruch, David Asulin, Aviad Moreshet, Misha Sra, and Pradeep Sen. Aidappeal: Automatic image dataset and algorithm for content appeal enhancement and assessment labeling. In Proceedings of the European Conference on Computer Vision (ECCV), 2024. 2, 6 +[7] Xiaokang Chen, Zhiyu Wu, Xingchao Liu, Zizheng Pan, Wen Liu, Zhenda Xie, Xingkai Yu, and Chong Ruan. Januspro: Unified multimodal understanding and generation with data and model scaling. arXiv preprint arXiv:2501.17811, 2025.2.5 +[8] Mihai Gabriel Constantin, Miriam Redi, Gloria Zen, and Bogdan Ionescu. Computational understanding of visual interestingness beyond semantics: Literature survey and analysis of covariates. In ACM Computing Surveys, volume 52. Association for Computing Machinery, 5 2019. 2 +[9] Mihai Gabriel Constantin, Liviu Daniel Stefan, Bogdan Ionescu, Ngoc Q.K. Duong, Claire Hélène Demarty, and Mats Sjöberg. Visual Interestingness Prediction: A Benchmark Framework and Literature Review. International Journal of Computer Vision (IJCV), 129(5):1526-1550, 5 2021. 2, 5, 7 +[10] Lisette de Vries, Sonja Gensler, and Peter S.H. Leeflang. Popularity of brand posts on brand fan pages: An investigation of the effects of social media marketing. Journal of Interactive Marketing, 26(2):83-91, 2012. 6 +[11] Arturo Deza and Devi Parikh. Understanding image vitality. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1818-1826, 2015. 1, 5 +[12] Sagnik Dhar, Vicente Ordonez, and Tamara L Berg. High level describable attributes for predicting aesthetics and interestingness. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages + +1657-1664,2011.1,2 +[13] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, Anirudh Goyal, Anthony Hartshorn, Aobo Yang, Archi Mitra, Archie Sravankumar, Artem Korenev, Arthur Hinsvark, Arun Rao, Aston Zhang, Aurelien Rodriguez, Austen Gregerson, Ava Spataru, Baptiste Roziere, Bethany Biron, Binh Tang, Bobbie Chern, Charlotte Caucheteux, Chaya Nayak, Chloe Bi, Chris Marra, Chris McConnell, Christian Keller, Christophe Touret, Chunyang Wu, Corinne Wong, Cristian Canton Ferrer, Cyrus Nikolaidis, Damien Allonsius, Daniel Song, Danielle Pintz, Danny Livshits, David Esiobu, Dhruv Choudhary, Dhruv Mahajan, Diego Garcia-Olano, Diego Perino, Dieuwke Hupkes, Egor Lakomkin, Ehab AlBadawy, Elina Lobanova, Emily Dinan, Eric Michael Smith, Filip Radenovic, Frank Zhang, Gabriel Synnaeve, Gabrielle Lee, Georgia Lewis Anderson, Graeme Nail, Gregoire Mialon, Guan Pang, Guillem Cucurell, Hailey Nguyen, Hannah Korevaar, Hu Xu, Hugo Touvron, Iliyan Zarov, Imanol Arrieta Ibarra, Isabel Kloumann, Ishan Misra, Ivan Evtimov, Jade Copet, Jaewon Lee, Jan Geffert, Jana Vranes, Jason Park, Jay Mahadeokar, Jeet Shah, Jelmer van der Linde, Jennifer Billock, Jenny Hong, Jenya Lee, Jeremy Fu, Jianfeng Chi, Jianyu Huang, Jiawen Liu, Jie Wang, Jiecao Yu, Joanna Bitton, Joe Spisak, Jongsoo Park, Joseph Rocca, Joshua Johnstun, Joshua Saxe, Junteng Jia, Kalyan Vasuden Alwala, Kartikeya Upasani, Kate Plawiak, Ke Li, Kenneth Heafield, Kevin Stone, Khalid El-Arini, Krithika Iyer, Kshitiz Malik, Kuenley Chiu, Kunal Bhalla, Lauren Rantala-Yeary, Laurens van der Maaten, Lawrence Chen, Liang Tan, Liz Jenkins, Louis Martin Lovish Madaan Labo Malo Lukas Blecher Lukas Landzaat Luke de Oliveira Madeline Muzzi Mahesh Pasupuleti, Mannat Singh Manohar Paluri Marcin Kardas Mathew Oldham Mathieu Rita Maya Pavlova Melanie Kambadur Mike Lewis Min Si,Mitesh Kumar Singh Mona Hassan Naman Goyal Narjes Torabi Nikolay Bashlykov Nikolay Bogoychev,Niladri Chatterji Olivier Duchenne Onur Celebi Patrick Alrassy Pengchuan Zhang Pengwei Li,Petar Vasic Peter Weng Prajjwal Bhargava Pratik Dubal Praveen Krishnan,Punit Singh Koura Puxin Xu Qing He,Qingxiao Dong Ragavan Srinivasan Raj Ganapathy,Ramon Calderer Ricardo Silveira Cabral Robert Stojnic,Roterta Raileanu Rohit Girdhar Rohit Patel Romain SauvestreRonnie Polidoro Roshan Sumbaly Ross Taylor Ruan Silva Rui Hou Rui Wang Saghar Hosseini,Sahana Chennabasappa Sanjay Singh Sean Bell Seohyun Sonia KimSergey Edunov Shaoliang Nie Sharan Narang. Sharath Rarparthy Sheng Shen Shengye Wan Shruti BhosaleShun Zhang Simon Vandenhende Soumya Batra Spencer Whitman Sten Sootla Stephane Collot Suchin Gururangan Sydney Borodinsky Tamar Herman Tara Fowler Tarek Sheasha Thomas Georgiou Thomas Scialom Tobias Speckbacher,Todor MihaylovTong Xiao Ujjwal Karn. Vedanuj Goswami,Vibhor Gupta,Vignesh Ramanathan. Viktor Kerkez,Vincent Gonguet,Virginie Do,Vish Vogeti Vladan Petrovic Weiwei Chu Wenhan Xiong,Wenyin Fu Whitney Meers Xavier Martinet Xiaodong Wang Xia + +qing Ellen Tan, Xinfeng Xie, Xuchao Jia, Xuewei Wang, Yaelle Goldschlag, Yashesh Gaur, Yasmine Babaei, Yi Wen, Yiwen Song, Yuchen Zhang, Yue Li, Yuning Mao, Zacharie Delpierre Coudert, Zheng Yan, Zhengxing Chen, Zoe Papakipos, Aaditya Singh, Aaron Grattafori, Abha Jain, Adam Kelsey, Adam Shajnfeld, Adithya Gangidi, Adolfo Victoria, Ahuva Goldstand, Ajay Menon, Ajay Sharma, Alex Boesenberg, Alex Vaughan, Alexei Baevski, Allie Feinstein, Amanda Kallet, Amit Sangani, Anam Yunus, Andrei Lupu, Andres Alvarado, Andrew Caples, Andrew Gu, Andrew Ho, Andrew Poulton, Andrew Ryan, Ankit Ramchandani, Annie Franco, Aparajita Saraf, Arkabandhu Chowdhury, Ashley Gabriel, Ashwin Bharambe, Assaf Eisenman, Azadeh Yazdan, Beau James, Ben Maurer, Benjamin Leonhardi, Bernie Huang, Beth Loyd, Beto De Paola, Bhargavi Paranjape, Bing Liu, Bo Wu, Boyu Ni, Braden Hancock, Bram Wasti, Brandon Spence, Brani Stojkovic, Brian Gamido, Britt Montalvo, Carl Parker, Carly Burton, Catalina Mejia, Changhan Wang, Changkyu Kim, Chao Zhou, Chester Hu, ChingHsiang Chu, Chris Cai, Chris Tindal, Christoph Feichtenhofer, Damon Civin, Dana Beaty, Daniel Kreymer, Daniel Li, Danny Wyatt, David Adkins, David Xu, Davide Testuggine, Delia David, Devi Parikh, Diana Liskovich, Didem Foss, Dingkang Wang, Duc Le, Dustin Holland, Edward Dowling,Eissa Jamil, Elaine Montgomery,Eleonora Presani, Emily Hahn, Emily Wood, Erik Brinkman Esteban Arcaute, Evan Dunbar, Evan Smothers, Fei Sun Felix Kreuk,Feng Tian,Firat Ozgenel, Francesco Caggioni,Francisco Guzmán Frank Kanayet,Frank Seide,Gabriela Medina Florez,Gabriella Schwarz,Gada Badeer, Georgia Swee Gil Halpern,Govind Thattai Grant HermanGrigory Sizov Guangyi,ZhangGuna Lakshminarayanan Hamid Shojanazeri,Han Zou,Hannah Wang,Hanwen Zha,Haroun Habeeb,Harrison Rudolph,Helen Suk,Henry Aspegren Hunter Goldman,Ibrahim Damlaj,Igor Molybog,Igor Tufanov,Irina-Elena Veliche Itai Gat,Jake WeissmanJames Geboski,James Kohli,Japhet AsherJean-Baptiste Gaya Jeff MarcusJeff TangJennifer ChanJenny ZhenJeremy Reizenstein Jeremy Teboul Jessica Zhong Jian Jin Jingyi YangJoe CummingsJon Carvill Jon Shepard Jonathan McPhie,Jonathan Torres Josh Ginsburg Junjie Wang Kai Wu,Kam Hou U,Karan Saxena,Karthik Prasad,Kartikay Khandelwal,Katayoun Zand,Kathy Matosich,Kaushik Veeraraghavan,Kelly Michelena Keqian LiKun Huang. Kunal ChawlaKushal LakhotiaKyle HuangLailin Chen Lakshya Garg,Lavender A,Leandro SilvaLee Bell Lei Zhang,Liangpeng Guo,Licheng Yu,Liron Moshkovich Luca Wehrstedt,Madian Khabsa Manav Avalani Manish BhattMaria Tsimpoukelli,Martynas MankusMatan Hasson Matthew LennieMatthias Reso Maxim Groshev Maxim Naumov Maya Lathi Meghan Keneally Michael L. SeltzerMichal Valko,Michelle Restrepo,Mihir PatelMik Vyatskov,Mikayel Samvelyan Mike Clark Mike Macey Mike WangMiquel Jubert Hermoso,Mo Metanat Mohammad Rastegari Munish Bansal Nandhini Santhanam. Natascha ParksNatasha White,Navyata BawaNayan Singhal Nick Egebo Nicolas Usunier Nikolay Pavlovich Laptev Ning Dong Ning ZhangNorman Cheng Oleg + +Chernoguz, Olivia Hart, Omkar Salpekar, Ozlem Kalinli, Parkin Kent, Parth Parekh, Paul Saab, Pavan Balaji, Pedro Rittner, Philip Bontrager, Pierre Roux, Piotr Dollar, Polina Zvyagina, Prashant Ratanchandani, British Yuvraj, Qian Liang, Rachad Alao, Rachel Rodriguez, Rafi Ayub, Raghotham Murthy, Raghu Nayani, Rahul Mitra, Raymond Li, Rebekkah Hogan, Robin Battey, Rocky Wang, Rohan Maheswari, Russ Howes, Rudy Rinott, Sai Jayesh Bondu, Samyak Datta, Sara Chugh, Sara Hunt, Sargun Dhillon, Sasha Sidorov, Satadru Pan, Saurabh Verma, Seiji Yamamoto, Sharadh Ramaswamy, Shaun Lindsay, Shaun Lindsay, Sheng Feng, Shenghao Lin, Shengxin Cindy Zha, Shiva Shankar, Shuqiang Zhang, Shuqiang Zhang, Sinong Wang, Sneha Agarwal, Soji Sajuyigbe, Soumith Chintala, Stephanie Max, Stephen Chen, Steve Kehoe, Steve Satterfield, Sudarshan Govindaprasad, Sumit Gupta, Sungmin Cho, Sunny Virk, Suraj Subramanian, Sy Choudhury, Sydney Goldman, Tal Remez, Tamar Glaser, Tamara Best, Thilo Kohler, Thomas Robinson, Tianhe Li, Tianjun Zhang, Tim Matthews, Timothy Chou, Tzook Shaked, Varun Vontimitta, Victoria Ajayi, Victoria Montanez, Vijai Mohan, Vinay Satish Kumar, Vishal Mangla, Vitor Albiero, Vlad Ionescu, Vlad Poenaru, Vlad Tiberiu Mihalescu, Vladimir Ivanov, Wei Li, Wenchen Wang, Wenwen Jiang, Wes Bouaziz, Will Constable, Xiaocheng Tang, Xiaofang Wang, Xiaojian Wu, Xiaolan Wang, Xide Xia, Xilun Wu, Xinbo Gao, Yanjun Chen, Ye Hu, Ye Jia, Ye Qi, Yenda Li, Yilin Zhang, Ying Zhang, Yossi Adi, Youngjin Nam, Yu, Wang. Yuchen Hao, Yundi Qian, Yuzi He, Zach Rait, Zachary Devito, Zef Rosnbrick. Zhaoduo Wen. Zhenyu Yang and Zhiwei Zhao. The llama 3 herd of models. arXiv preprint arXiv:2407.217832024.235 + +[14] Flickr. About interestingness, https://www.flickr.com/explore/interesting/, 2024. 2024-02-23. 2 +[15] Andrea Frome, Greg S Corrado, Jon Shlens, Samy Bengio, Jeff Dean, Marc' Aurelio Ranzato, and Tomas Mikolov. Devise: A deep visual-semantic embedding model. In C.J. Burges, L. Bottou, M. Welling, Z. Ghahramani, and K.Q. Weinberger, editors, Advances in Neural Information Processing Systems (NeurIPS), volume 26. Curran Associates, Inc., 2013. 2 +[16] Maham Gardezi, King Hei Fung, Usman Mirza Baig, Mariam Ismail, Oren Kadosh, Yoram S Bonneh, and Bhavin R Sheth. What makes an image interesting and how can we explain it. Frontiers in Psychology, 12, 2021. 1, 2 +[17] Helmut Grabner, Fabian Nater, Michel Druey, and Luc Van Gool. Visual interestingness in image sequences. In Proceedings of the ACM International Conference on Multimedia, pages 1017-1026, 2013. 2 +[18] Michael Gygli, Helmut Grabner, Hayko Riemenschneider, Fabian Nater, and Luc Van Gool. The interestingness of images. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 1633-1640, 2013. 1, 2, 4, 6 +[19] Daichi Haraguchi, Naoto Inoue, Wataru Shimoda, Hayato Mitani, Seiichi Uchida, and Kota Yamaguchi. Can gpts evaluate graphic design based on design principles? In SIG- + +GRAPH Asia 2024 Technical Communications, 2024. 2 +[20] Yuki Hirakawa, Takashi Wada, Kazuya Morishita, Ryotaro Shimizu, Takuya Furusawa, Sai Htaung Kham, and Yuki Saito. An empirical analysis of gpt-4v's performance on fashion aesthetic evaluation. In SIGGRAPH Asia 2024 Technical Communications, 2024. 2 +[21] Phillip Isola, Jianxiong Xiao, Antonio Torralba, and Aude Oliva. What makes an image memorable? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 145-152, 2011. 5 +[22] Junjie Ke, Keren Ye, Jiahui Yu, Yonghui Wu, Peyman Milanfar, and Feng Yang. Vila: Learning image aesthetics from user comments with vision-language pretraining. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 5 +[23] Been Kim. Beyond interpretability: developing a language to shape our relationships with ai, Apr 2022. 2 +[24] Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan Popov, Matteo Malloci, Alexander Kolesnikov, Tom Duerig, and Vittorio Ferrari. The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale. International Journal of Computer Vision (IJCV), 2020. 3 +[25] Ziming Liu, Ouail Kitouni, Niklas S Nolte, Eric Michaud, Max Tegmark, and Mike Williams. Towards understanding grokking: An effective theory of representation learning. In S. Koyejo, S. Mohamed, A. Agarwal, D. Belgrave, K. Cho, and A. Oh, editors, Advances in Neural Information Processing Systems (NeurIPS), volume 35, pages 34651-34663. Curran Associates, Inc., 2022. 2 +[26] René Magritte. The treachery of images (this is not a pipe) (la trahison des images [ceci n'est pas une pipe]), 1929. 2024-10-24. 2 +[27] Leland McInnes, John Healy, and James Melville. Umap: Uniform manifold approximation and projection for dimension reduction. arXiv preprint arXiv:1802.03426, 2018. 7 +[28] Kevin Meng, David Bau, Alex Andonian, and Yonatan Belinkov. Locating and editing factual associations in gpt. In S. Koyejo, S. Mohamed, A. Agarwal, D. Belgrave, K. Cho, and A. Oh, editors, Advances in Neural Information Processing Systems (NeurIPS), volume 35, pages 17359-17372. Curran Associates, Inc., 2022. 2 +[29] Cade Metz. In two moves, alphago and lee sedol redefined the future. WIRED.com, 16, 2016. 2 +[30] Coen D Needell and Wilma A Bainbridge. Embracing new techniques in deep learning for estimating image memorability. Computational Brain & Behavior, 5(2):168-184, 2022. 5 +[31] OpenAI. Hello gpt-4o, https://openai.com/index/hello-gpt-4o/, 2024. 2024-07-01. 2 +[32] OpenAI, Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, Red Avila, Igor Babuschkin, Suchir Balaji, Valerie Balcom, Paul Baltescu, Haiming Bao, Mohammad Bavarian, Jeff Belgium, Irwan Bello, Jake Berdine, Gabriel Bernadett-Shapiro, Christopher Berner, Lenny Bogdonoff, Oleg Boiko, Madelaine Boyd, Anna-Luisa Brakman, Greg Brockman, + +Tim Brooks, Miles Brundage, Kevin Button, Trevor Cai, Rosie Campbell, Andrew Cann, Brittany Carey, Chelsea Carlson, Rory Carmichael, Brooke Chan, Che Chang, Fotis Chantzis, Derek Chen, Sully Chen, Ruby Chen, Jason Chen, Mark Chen, Ben Chess, Chester Cho, Casey Chu, Hyung Won Chung, Dave Cummings, Jeremiah Currier, Yunxing Dai, Cory Decareaux, Thomas Degry, Noah Deutsch, Damien Deville, Arka Dhar, David Dohan, Steve Dowling, Sheila Dunning, Adrien Ecoffet, Atty Eleti, Tyna Eloundou, David Farhi, Liam Fedus, Niko Felix, Simón Posada Fishman, Juston Forte, Isabella Fulford, Leo Gao, Elie Georges, Christian Gibson, Vik Goel, Tarun Gogineni, Gabriel Goh, Rapha Gontijo-Lopes, Jonathan Gordon, Morgan Grafstein, Scott Gray, Ryan Greene, Joshua Gross, Shixiang Shane Gu, Yufei Guo, Chris Hallacy, Jesse Han, Jeff Harris, Yuchen He, Mike Heaton, Johannes Heidecke, Chris Hesse, Alan Hickey, Wade Hickey, Peter Hoeschele, Brandon Houghton, Kenny Hsu, Shengli Hu, Xin Hu, Joost Huizinga, Shantanu Jain, Shawn Jain, Joanne Jang, Angela Jiang, Roger Jiang, Haozhun Jin, Denny Jin, Shino Jomoto, Billie Jonn, Heewoo Jun, Tomer Kaftan, Lukasz Kaiser, Ali Kamali, Ingmar Kanitscheider, Nitish Shirish Keskar, Tabarak Khan, Logan Kilpatrick, Jong Wook Kim, Christina Kim, Yongjik Kim, Jan Hendrik Kirchner, Jamie Kiros, Matt Knight, Daniel Kokotajlo, Lukasz Kondraciuk, Andrew Kondrich, Aris Konstantinidis, Kyle Kosic, Gretchen Krueger, Vishal Kuo, Michael Lampe, Ikai Lan, Teddy Lee, Jan Leike, Jade Leung, Daniel Levy, Chak Ming Li, Rachel Lim, Molly Lin, Stephanie Lin, Mateusz Litwin, Theresa Lopez, Ryan Lowe, Patricia Lue, Anna Makanju, Kim Malfacini, Sam Manning, Todor Markov, Yaniv Markovsky, Bianca Martin, Katie Mayer, Andrew Mayne, Bob McGrew, Scott Mayer McKinney, Christine McLeavey, Paul McMillan, Jake McNeil, David Medina, Aalok Mehta, Jacob Menick, Luke Metz, Andrey Mishchenko, Pamela Mishkin, Vinnie Monaco, Evan Morikawa, Daniel Mossing, Tong Mu. Mira Murati Oleg Murk David Mély Ashvin Nair Reichiro Nakano Rajeev Nayak Arvind Neelakantan Richard Ngo,Hyeonwoo Noh Long OuyangCullen O'Keefe,Jakub Pachocki Alex Paino Joe Palermo Ashley Pantuliano Giambattista Parascandolo Joel Parish Emy Parparita Alex Passos,Mikhail PavlovAndrew PengAdam Perelman Filipe de Avila Belbute Peres Michael Petrov Henrique Ponde de Oliveira Pinto Michael Pokorny Michelle Pokrass Vitchyr H. Pong Tolly Powell Alethea Power Boris Power Elizabeth Proehl Raul Puri,Alec Radford Jack Rae Aditya Ramesh,Cameron Raymond Francis Real Kendra Rimbach Carl RossBob RotstedHenri Roussez Nick Ryder Mario Saltarelli Ted Sanders Shibani Santurkar Girish Sastry. Heather Schmidt David Schnurr John Schulman Daniel Selsam Kyla Sheppard Toki Sherbakov Jessica Shieh Sarah Shoker Pranav Shyam Szymon Sidor Eric Sigler Maddie Simens Jordan Sitkin Katarina Slama Ian Sohl Benjamin Sokolowsky Yang Song Natalie Staudacher Felipe Petroski Such Natalie Summers Ilya Sutskever Jie Tang,Nikolas Tezak Madeleine B.Thompson Phil Tillet Amin Tootoonchian Elizabeth Tseng Preston Tuggle Nick Turley Jerry Tworek Juan Felipe Cerón Uribe Andrea Val + +lone, Arun Vijayvergiya, Chelsea Voss, Carroll Wainwright, Justin Jay Wang, Alvin Wang, Ben Wang, Jonathan Ward, Jason Wei, CJ Weinmann, Akila Welihinda, Peter Welinder, Jiayi Weng, Lilian Weng, Matt Wiethoff, Dave Willner, Clemens Winter, Samuel Wolrich, Hannah Wong, Lauren Workman, Sherwin Wu, Jeff Wu, Michael Wu, Kai Xiao, Tao Xu, Sarah Yoo, Kevin Yu, Qiming Yuan, Wojciech Zaremba, Rowan Zellers, Chong Zhang, Marvin Zhang, Shengjia Zhao, Tianhao Zheng, Juntang Zhuang, William Zhuk, and Barret Zoph. GPT-4 Technical Report. arXiv preprint arXiv:2303.08774, 2024. 1, 2, 3 +[33] Maxime Oquab, Timothée Darcet, Theo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Russell Howes, Po-Yao Huang, Hu Xu, Vasu Sharma, Shang-Wen Li, Wojciech Galuba, Mike Rabbat, Mido Assran, Nicolas Ballas, Gabriel Synnaeve, Ishan Misra, Herve Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 6 +[34] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, John Schulman, Jacob Hilton, Fraser Kelton, Luke Miller, Maddie Simens, Amanda Askell, Peter Welinder, Paul F Christiano, Jan Leike, and Ryan Lowe. Training language models to follow instructions with human feedback. In S. Koyejo, S. Mohamed, A. Agarwal, D. Belgrave, K. Cho, and A. Oh, editors, Advances in Neural Information Processing Systems (NeurIPS), volume 35, pages 27730-27744. Curran Associates, Inc., 2022. 2 +[35] Alicia Parrish, Susan Hao, Sarah Laszlo, and Lora Aroyo. Is a picture of a bird a bird? a mixed-methods approach to understanding diverse human perspectives and ambiguity in machine vision models. In Gavin Abercrombie, Valerio Basile, Davide Bernadi, Shiran Dudy, Simona Frenda, Lucy Havens, and Sara Tonelli, editors, Proceedings of the 3rd Workshop on Perspectivist Approaches to NLP (NLPPerspectives) @ LREC-COLING 2024, pages 1-18, Torino, Italia, May 2024. ELRA and ICCL. 2 +[36] Yuang Peng, Yuxin Cui, Haomiao Tang, Zekun Qi, Runpei Dong, Jing Bai, Chunrui Han, Zheng Ge, Xiangyu Zhang, and Shu-Tao Xia. Dreambench++: A human-aligned benchmark for personalized image generation. arXiv preprint arXiv:2406.16855, abs/2406.16855, 2024. 2 +[37] Sarah Pratt, Ian Covert, Rosanne Liu, and Ali Farhadi. What does a platypus look like? generating customized prompts for zero-shot image classification. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 5 +[38] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In Proceedings of the International Conference on Machine Learning (ICML), pages 8748-8763, 2021. 5 +[39] Lisa Schut, Nenad Tomasev, Tom McGrath, Demis Hassabis, Ulrich Paquet, and Been Kim. Bridging the human-ai knowledge gap: Concept discovery and transfer in alphazero, 2023. + +2 + +[40] Gemini Team, Rohan Anil, Sebastian Borgeaud, JeanBaptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M. Dai, Anja Hauth, Katie Millican, David Silver, Melvin Johnson, Ioannis Antonoglou, Julian Schrittwieser, Amelia Glaese, Jilin Chen, Emily Pitler, Timothy Lillicrap, Angeliki Lazaridou, Orhan First, James Molloy, Michael Isard, Paul R. Barham, Tom Hennigan, Benjamin Lee, Fabio Viola, Malcolm Reynolds, Yuanzhong Xu, Ryan Doherty, Eli Collins, Clemens Meyer, Eliza Rutherford, Erica Moreira, Kareem Ayoub, Megha Goel, Jack Krawczyk, Cosmo Du, Ed Chi, Heng-Tze Cheng, Eric Ni, Purvi Shah, Patrick Kane, Betty Chan, Manaal Faruqui, Aliaksei Severyn, Hanzhao Lin, YaGuang Li, Yong Cheng, Abe Ittycheriah, Mahdis Mahdieh, Mia Chen, Pei Sun, Dustin Tran Sumit Bagri, Balaji Lakshminarayanan, Jeremiah Liu, Andras Orban, Fabian Gura, Hao Zhou, Xinying Song, Aurelien Boffy, Harish Ganapathy, Steven Zheng, HyunJeong Choe, Agoston Weisz, Tao Zhu, Yifeng Lu, Siddharth Gopal, Jarrod Kahn, Maciej Kula, Jeff Pitman, Rushin Shah, Emanuel Taropa, Majd Al Merey, Martin Baeuml, Zhifeng Chen, Laurent El Shafey, Yujing Zhang, Olcan Sercinoglu, George Tucker, Enrique Piqueras, Maxim Krikun, Iain Barr, Nikolay Savinov, Ivo Danihelka, Becca Roelofs, Anais White, Anders Andreassen, Tamara von Glehn, Lakshman Yagati, Mehran Kazemi, Lucas Gonzalez, Misha Khalman, Jakub Sygnowski, Alexandre Frechette, Charlotte Smith Laura Culp, Lev Proleev, Yi Luan, Xi Chen, James Lottes, Nathan Schucher Federico Lebron Alban Rrustemi,Natalie Clay Phil CroneTomas KociskyJeffrey Zhao Bartek PerzDian YuHeidi HowardAdam BloniarzJack W. Rae,Han LuLaurent Sifre Marcello Maggioni Fred AlcoberDan GarretteMegan BarnesShantanu ThakoorJacob Austin Gabriel Barth-MaronWilliam WongRishabh Joshi,Rahma ChaabouniDeeni FatihaArun Ahuja,Gaurav Singh Tomar Evan SenterMartin ChadwickIlya KornakovNithya Attaluri Inaki IturrateRuibo LiuYunxuan Li Sarah Cogan Jeremy ChenChao JiaChenjie Gu Qiao Zhang Jordan Grimstad Ale Jakse Hartman Xavier Garcia Thanumalayan Sankaranarayana Pillai Jacob Devlin Michael Laskin Diego de Las Casas,Dasha Valter Connie TaoLorenzo BlancoAdria Puigdomenech Badia David ReitterMianna ChenJenny Brennan Clara Rivera,Sergey Brin, Shariq Iqbal,Gabriela SuritaJane Labanowski Abhi Rao Stephanie Winkler Emilio Parisotto,Yiming Gu Kate Olszewska Ravi Addanki Antoine Miech Annie Louis Denis Teptyashin Geoff Brown Elliot Catt Jan BalaguerJackie XiangPidong WangZoe Ashwood Anton BriukhovAlbert WebsonSanjay GanapathySmit Sanghavi Ajay Kannan Ming-Wei Chang Axel Stjerngren Josip Djolonga Yuting Sun Ankur Bapna Matthew Aitchison Pedram Pejman Henryk Michalewski Tianhe Yu Cindy WangJuliette LoveJunwhan Ahn Dawn Bloxwich Kehang HanPeter HumphreysThibault SellamJames Bradbury Varun GodboleSina Samangooei Bogdan Damoc Alex KaskasoliSebastien M.R.ArnoldVijay Vasudevan Shubham Agrawal Jason Riesa Dmitry Lepikhin Richard Tanburn Srivatsan Srinivasan Hyeontaek Lim Sarah Hod + +Kinson, Pranav Shyam, Johan Ferret, Steven Hand, Ankush Garg, Tom Le Paine, Jian Li, Yujia Li, Minh Giang, Alexander Neitz, Zaheer Abbas, Sarah York, Michel Reid, Elizabeth Cole, Aakanksha Chowdhery, Dipanjan Das, Dominika Rogozinska, Vitaliy Nikolaev, Pablo Sprechmann, Zachary Nado, Lukas Zilka, Flavien Prost, Luheng He, Marianne Monteiro, Gaurav Mishra, Chris Welty, Josh Newlan, Dawei Jia, Miltiadis Allamanis, Clara Huiyi Hu, Raoul de Liedekerke, Justin Gilmer, Carl Saroufim, Shruti Rijhwani, Shaobo Hou, Disha Shrivastava, Anirudh Baddepudi, Alex Goldin, Adnan Ozturel, Albin Cassirer, Yunhan Xu, Daniel Sohn, Devendra Sachan, Reynald Kim Amplayo, Craig Swanson, Dessie Petrova, Shashi Narayan, Arthur Guez, Siddhartha Brahma, Jessica Landon, Miteyan Patel, Ruizhe Zhao, Kevin Villela, Luyu Wang, Wenhao Jia, Matthew Rahtz, Mai Giménez, Legg Yeung, James Keeling, Petko Georgiev, Diana Mincu, Boxi Wu, Salem Haykal, Rachel Saputro, Kiran Vodrahalli, James Qin, Zeynep Cankara, Abhanshu Sharma, Nick Fernando, Will Hawkins, Behnam Neyshabur, Solomon Kim, Adrian Hutter, Priyanka Agrawal, Alex Castro-Ros, George van den Driessche, Tao Wang, Fan Yang, Shuo yiin Chang, Paul Komarek, Ross McIlroy, Mario Lucic, Guodong Zhang, Wael Farhan, Michael Sharman, Paul Natsev, Paul Michel, Yamini Bansal, Siyuan Qiao, Kris Cao, Siamak Shakeri, Christina Butterfield, Justin Chung Paul Kishan Rubenstein, Shivani Agrawal, Arthur Mensch. Kedar Soparkar, Karel Lenc, Timothy Chung, Aedan Pope. Loren Maggiore, Jackie Kay, Priya Jhakra, Shibo Wang. Joshua Maynez, Mary Phuong, Taylor Tobin, Andrea Tacchetti, Maja Trebacz, Kevin Robinson, Yash Katariya. Sebastian Riedel, Paige Bailey, Kefan Xiao, Nimesh Ghelani. Lora Aroyo, Ambrose Slone, Neil Houlsby. Xuehan Xiong. Zhen Yang. Elena Gribovskaya. Jonas Adler. Mateo Wirth. Lisa Lee. Music Li. Thais Kagohara. Jay Pavagadhi. Sophie Bridgers. Anna Bortsova. Sanjay Ghemawat. Zafarali Ahmed. Tianqi Liu. Richard Powell. Vijay Bolina. Mariko Inuma. Polina Zablotskaia. James Besley. Da-Woon Chung. Timothy Dozat. Ramona Comanescu. Xiance Si. Jeremy Greer. Guolong Su. Martin Polacek. Raphael Lopez Kaufman. Simon Tokumine. Hexiang Hu. Elena Buchatskaya. Yingjie Miao. Mohamed Elhawaty. Aditya Siddhant. Nealad Tomasev. Jinwei Xing. Christina Greer. Helen Miller. Shereen Ashraf. Aurko Roy. Zizhao Zhang. Ada Ma. Angelos Filos. Milos Besta. Rory Blevins. Ted Klimenko. Chih-Kuan Yeh. Soravit Changpinyo. Jiaqi Mu. Oscar Chang. Mantas Pajarskas. Carrie Muir. Vered Cohen. Charline Le Lan. Krishna Haridasan. Amit Marathe. Steven Hansen. Sholto Douglas. Rajkumar Samuel. Mingqiu Wang. Sophia Austin. Chang Lan. Jiepu Jiang. Justin Chiu. Jaime Alonso Lorenzo. Lars Lowe Sjösund. Sébastien Cevey. Zach Gleicher. Thi Avrahami. Anudhyan Boral. Hansa Srinivasan. Vittorio Selo. Rhys May. Konstantinos Aisopoulos. Léonard Hussenot. Livio Baldini Soares. Kate Baumli. Michael B. Chang. Adrià Recasens. Ben Caine. Alexander Pritzel. Filip Pavetic. Fabio Pardo. Anita Gergely. Justin Frye. Vinay Ramasesh. Dan Horgan. Kartikeya Badola. Nora Kassner Subhrajit Roy. Ethan Dyer. Víctor Campos Campos. Alex Tomala. Yunhao Tang. Dalia El Badawy. Elspeth White + +Basil Mustafa, Oran Lang, Abhishek Jindal, Sharad Vikram, Zhitao Gong, Sergi Caelles, Ross Hemsley, Gregory Thornton, Fangxiaoyu Feng, Wojciech Stokowiec, Ce Zheng, Phoebe Thacker, Caglar Unlu, Zhishuai Zhang, Mohammad Saleh, James Svensson, Max Bileschi, Piyush Patil, Ankesh Anand, Roman Ring, Katerina Tsihlas, Arpi Vezer, Marco Selvi, Toby Shevlane, Mikel Rodriguez, Tom Kwiatkowski, Samira Daruki, Keran Rong, Allan Dafoe, Nicholas FitzGerald, Keren Gu-Lemberg, Mina Khan, Lisa Anne Hendricks, Marie Pellat, Vladimir Feinberg, James Cobon-Kerr, Tara Sainath, Maribeth Rauh, Sayed Hadi Hashemi, Richard Ives, Yana Hasson, Eric Noland, Yuan Cao, Nathan Byrd, Le Hou, Qingze Wang, Thibault Sottiaux, Michela Paganini, Jean-Baptiste Lespiau, Alexandre Moufarek, Samer Hassan, Kaushik Shivakumar, Joost van Amersfoort, Amol Mandhane, Pratik Joshi, Anirudh Goyal, Matthew Tung, Andrew Brock, Hannah Sheahan, Vedant Misra, Cheng Li, Nemanja Rakicevic, Mostafa Dehghani, Fangyu Liu, Sid Mittal, Junhyuk Oh, Seb Noury, Eren Sezener, Fantine Huot, Matthew Lamm, Nicola De Cao, Charlie Chen, Sidharth Mudgal, Romina Stella, Kevin Brooks, Gautam Vasudevan, Chenxi Liu, Mainak Chain, Nivedita Melinkeri, Aaron Cohen, Venus Wang, Kristie Seymour, Sergey Zubkov, Rahul Goel, Summer Yue, Sai Krishnakumaran, Brian Albert, Nate Hurley, Motoki Sano, Anhad Mohananey, Jonah Joughin, Egor Filonov, Tomasz Kepa, Yomna Eldawy, Jiawern Lim, Rahul Rishi, Shirin Badiezadegan, Taylor Bos, Jerry Chang. Sanil Jain, Sri Gayatri Sundara Padmanabhan, Subha Puttagunta Kalpesh Krishna Leslie Baker Norbert Kalb Vamsi Bedapudi Adam Kurzrok Shuntong Lei Anthony Yu Oren Litvin Xiang Zhou Zhichun WuSam Sobell Andrea SicilianoAlan Papir Robby NealeJonas Bragagnolo Tej Toor Tina Chen Valentin Anklin Feiran Wang,Richie Feng Milad Gholami Kevin LingLijuan Liu Jules Walter Hamid Moghaddam Arun Kishore Jakub Adamek Tyler Mercado Jonathan Mallinson Siddhinita Wandekar Stephen CagleEran Ofek Guillermo Garrido Clemens Lombriser. Maksim Mukha Botu Sun Hafeezul Rahman Mohammad Josip Matak Yadi Qian Vikas Peswani Pawel Janus Quan Yuan Leif Schelin Oana David Ankur Garg Yifan He Oleksii Duzhyi Anton Algmyr Timothee Lottaz Qi Li Vikas Yadav Luyao Xu Alex Chinien Rakesh Shivanna Aleksandr Chuklin Josie LiCarrie Spadine Travis Wolfe Kareem Mohamed Subhabrata Das Zihang Dai Kyle He Daniel von Dincklage Shyam Upadhyay Akanksha Maurya Luyan Chi Sebastian Krause Khalid Salama Pam G Rabinovitch,Pavan Kumar Reddy M Aarush Selvan,Mikhail Dektiarev,GolNaz Ghiasi,Erdem Guven,Himanshu Gupta Boyi Liu Deepak Sharma Idan Heimlich Shtacher Shachi Paul Oscar Akerlund Francois-Xavier Aubet Terry HuangChen Zhu Eric Zhu Elico Teixeira Matthew Fritze Francesco Bertolini Liana-Eleonora Marinescu Martin Bolle Dominik Paulus Khyatti Gupta Tejasi Latkar Max Chang Jason Sanders Roopa Wilson Xuewei Wu Yi-Xuan Tan Lam Nguyen Thiet Tulsee Doshi Sid Lall Swaroop Mishra Wanming Chen Thang Luong Seth Benjamin Jasmine Lee Ewa Andrejczuk Dominik Rabiej. Vipul Ranjan Krzysztof Styrc Pengcheng Yin Jon Simon + +Malcolm Rose Harriott, Mudit Bansal, Alexei Robsky, Geoff Bacon, David Greene, Daniil Mirylenka, Chen Zhou, Obaid Sarvana, Abhimanyu Goyal, Samuel Andermatt, Patrick Siegler, Ben Horn, Assaf Israel, Francesco Pongetti, ChihWei "Louis" Chen, Marco Selvatici, Pedro Silva, Kathie Wang, Jackson Tolins, Kelvin Guu, Roey Yogev, Xiaochen Cai, Alessandro Agostini, Maulik Shah, Hung Nguyen, Noah O Donnaile, Sebastien Pereira, Linda Friso, Adam Stambler, Adam Kurzrok, Chenkai Kuang, Yan Romanikhin, Mark Geller, ZJ Yan, Kane Jang, Cheng-Chun Lee, Wojciech Fica, Eric Malmi, Qijun Tan, Dan Banica, Daniel Balle, Ryan Pham, Yanping Huang, Diana Avram, Hongzhi Shi, Jasot Singh, Chris Hidey, Niharika Ahuja, Pranab Saxena, Dan Dooley, Srividya Pranavi Potharaju, Eileen O'Neill, Anand Gokulchandran, Ryan Foley, Kai Zhao, Mike Dusenberry, Yuan Liu, Pulkit Mehta, Ragha Kotikalapudi, Chalance Safranek-Shrader, Andrew Goodman, Joshua Kessinger, Eran Globen, Prateek Kolhar, Chris Gorgolewski, Ali Ibrahim, Yang Song, Ali Eichenbaum, Thomas Brovelli, Sahitya Potluri, Preethi Lahoti, Cip Baetu, Ali Ghorbani, Charles Chen, Andy Crawford, Shalini Pal, Mukund Sridhar, Petru Gurita, Asier Mujika, Igor Petrovski, PierreLouis Cedoz, Chenmei Li, Shiyuan Chen, Niccolò Dal Santo, Siddharth Goyal, Jitesh Punjabi, Karthik Kappaganthu, Chester Kwak, Pallavi LV, Sarmishta Velury, Himadri Choudhury, Jamie Hall, Premal Shah, Ricardo Figueira, Matt Thomas, Minjie Lu Ting Zhou Chintu Kumar Thomas Jurdi, Sharat Chikkerur, Yenai Ma Adams Yu Soo Kwak Victor Ahdel,Sujeevan RajayogamTravis ChomaFei Liu Aditya Barua Colin Ji,Ji Ho ParkVincent HellendoornAlex Bailey Taylan Bilal,Huanjie Zhou,Mehrdad Khatir Charles Sutton,Wojciech Rzadkowski,Fiona Macintosh Konstantin Shagin Paul Medina Chen Liang Jinjing ZhouPararth Shah,Yingying Bi Attila Dankovics Shipra Banga Sabine Lehmann Marissa Bredesen Zifan Lin John Eric Hoffmann Jonathan Lai Raynald Chung Kai Yang,Nihal Balani Arthur Brazinskas Andrei Sozanschi Matthew Hayes Hector Fernandez Alcalde Peter Makarov Will Chen Antonio Stella Liselotte Snijders Michael Mandl Ante Karrman Pawe Nowak Xinyi Wu Alex Dyck Krishnan Vaidyanathan Raghavender R Jessica Mallet Mitch Rudominer Eric Johnston Sushil Mittal Akhil Udathu Janara Christensen Vishal Verma Zach Irving Andreas Santucci,Gamageldin Elsayed Elnaz Davoodi Marin Georgiev Ian Tenney Nan Hua Geoffrey Cideron Edouard Leurent Mahmoud Alnahlawi Ionut Georgescu Nan Wei Ivy ZhengDylan Scandinavian Heinrich Jiang Jasper Snoek,Mukund SundararajanXuezhi WangZack Ontiveros Itay Karo Jeremy Cole,Vinu Rajashekhar Lara TumehEyal Ben-DavidRishub JainJonathan Uesato Romina Datta Oskar BunyanShimu Wu John Zhang Piotr Stanczyk Ye Zhang David Steiner Subhajit Naskar Michael Azzam Matthew Johnson Adam Paszke ChungCheng ChiuJaume Sanchez EliasAfroz Mohiuddin Faizan MuhammadJin MiaoAndrew LeeNino VieillardJane Park Jiageng ZhangJeff Stanway Drew Garmon Abhijit Karmarkar,Zhe Dong,Jong LeeAviral Kumar Luowei Zhou Jonathan Evens William IsaacGeoffrey Irv + +ing, Edward Loper, Michael Fink, Isha Arkatkar, Nanxin Chen, Izhak Shafran, Ivan Petrychenko, Zhe Chen, Johnson Jia, Anselm Levskaya, Zhenkai Zhu, Peter Grabowski, Yu Mao, Alberto Magni, Kaisheng Yao, Javier Snaider, Norman Casagrande, Evan Palmer, Paul Suganthan, Alfonso Castano, Irene Giannoumis, Wooyel Kim, Mikolaj Rybinski, Ashwin Sreevatsa, Jennifer Prendki, David Soergel, Adrian Goedeckemeyer, Willi Gierke, Mohsen Jafari, Meenu Gaba, Jeremy Wiesner, Diana Gage Wright, Yawen Wei, Harsha Vashisht, Yana Kulizhskaya, Jay Hoover, Maigo Le, Lu Li, Chimezie Iwuanyanwu, Lu Liu, Kevin Ramirez, Andrey Khorlin, Albert Cui, Tian LIN, Marcus Wu, Ricardo Aguilar, Keith Pallo, Abhishek Chakladar, Ginger Perng, Elena Allica Abellan, Mingyang Zhang, Ishita Dasgupta, Nate Kushman, Ivo Penchev, Alena Repina, Xihui Wu, Tom van der Weide, Priya Ponnapalli, Caroline Kaplan, Jiri Simsa, Shuangfeng Li, Olivier Dousse, Fan Yang, Jeff Piper, Nathan IE, Rama Pasumarthi, Nathan Lintz, Anitha Vijayakumar, Daniel Andor, Pedro Valenzuela, Minnie Lui, Cosmin Paduraru, Daiyi Peng, Katherine Lee, Shuyuan Zhang, Somer Greene, Duc Dung Nguyen, Paula Kurylowicz, Cassidy Hardin, Lucas Dixon, Lili Janzer, Kiam Choo, Ziqiang Feng, Biao Zhang, Achintya Singhal, Dayou Du, Dan McKinnon, Natasha Antropova, Tolga Bolukbasi, Orgad Keller, David Reid, Daniel Finchelstein, Maria Abi Raad, Remi Crocker, Peter Hawkins, Robert Dadashi, Colin Gaffney, Ken Franko, Anna Bulanova, Remi Leblond, Shirley Chung Harry Askham, Luis C. Cobo, Kelvin Xu, Felix Fischer, Jun Xu, Christina Sorokin, Chris Alberti, Chu-Cheng Lin, Colin Evans, Alek Dimitriev, Hannah Forbes, Dylan Banarsez Zora Tung, Mark Omernick Colton Bishop, Rachel Sterneck Rohan Jain Jiawei Xia Ehsan Amid Francesco Piccinno Xingyu Wang Praseem Banzal Daniel J. Mankowitz Alex Polozov Victoria Krakovna Sasha Brown Mohammad Hossein Bateni Dennis Duan Vlad Firoiu Meghana Thotakuri Tom Natan Matthieu Geist Ser tan Girgin Hui Li Jiayu Ye Ofir Roval Reiko Tojo Michael Kwong James Lee-Thorp Christopher Yew Danila Sinopalnikov Sabela Ramos John Mellor Abhishek Sharma Kathy Wu David Miller Nicolas Sonnerat Denis Vnukov Rory Greig Jennifer Beattie Emily Caveness Libin Bai Julian Eisenschlos Alex Korchemniy Tom Tsai Mimi Jasarevic Weize Kong Phuong Dao Zeyu Zheng Frederick Liu Fan Yang Rui Zhu Tian Huey Teh Jason Sanmiya Evgeny Gladchenko Nejc Trdin Daniel Toyama Evan Rosen,Sasan Tavakkol Linting XueChen Elkind Oliver Woodman John Carpenter George PapamakariosRupert KempSushant Kafle Tanya Grunina,Rishika Sinha Alice Talbert Diane WuDenese Owusu-Afriyie Cosmo Du Chloe Thornton Jordi PontTuset Pradyumna Narayana Jing Li Saaber Fatehi John WietingOmar Ajmeri Benigno Uria Yeongil Ko Laura KnightAmelie Heliou Ning Niu Shane Gu Chenxi Pang Yeqing Li,Nir Levine,Ariel StolovichRebeca SantamariaFernandez Sonam Goenka Wenny Yustalim Robin Strudel Ali Elquresh Charlie Deck Hyo Lee Zonglin Li Kyle Levin Raphael Hoffmann Dan Holtmann-Rice Olivier Bachem Sho Arora Christy Koh Soheil Hassas Yeganeh Siim Poder,Mukarram Tariq,Yanhua SunLucian Ionita Mo + +jtaba Seyedhosseini, Pouya Tafti, Zhiyu Liu, Anmol Gulati, Jasmine Liu, Xinyu Ye, Bart Chrzaszcz, Lily Wang, Nikhil Sethi, Tianrun Li, Ben Brown, Shreya Singh, Wei Fan, Aaron Parisi, Joe Stanton, Vinod Koverkathu, Christopher A. Choquette-Choo, Yunjie Li, TJ Lu, Abe Ittycheriah, Prakash Shroff, Mani Varadarajan, Sanaz Bahargam, Rob Willoughby, David Gaddy, Guillaume Desjardins, Marco Cornero, Brona Robenek, Bhavishya Mittal, Ben Albrecht, Ashish Shenoy, Fedor Moiseev, Henrik Jacobsson, Alireza Ghaffarkhah, Morgane Riviere, Alanna Walton, Clément Crepy, Alicia Parrish, Zongwei Zhou, Clement Farabet, Carey Radebaugh, Praveen Srinivasan, Claudia van der Salm, Andreas Fidjeland, Salvatore Scellato, Eri Latorre-Chimoto, Hanna Klimczak-Plucinska, David Bridson, Dario de Cesare, Tom Hudson, Piermaria Mendolicchio, Lexi Walker, Alex Morris, Matthew Mauger, Alexey Guseynov, Alison Reid, Seth Odoom, Lucia Loher, Victor Cotruta, Madhavi Yenugula, Dominik Grewe, Anastasia Petrushkina, Tom Duerig, Antonio Sanchez, Steve Yadlowsky, Amy Shen, Amir Globerson, Lynette Webb, Sahil Dua, Dong Li, Surya Bhupatiraju, Dan Hurt, Haroon Qureshi, Ananth Agarwal, Tomer Shani, Matan Eyal, Anuj Khare, Shreyas Rammohan Belle, Lei Wang, Chetan Tekur, Mihir Sanjay Kale, Jinliang Wei, Ruoxin Sang, Brennan Saeta, Tyler Liechty, Yi Sun, Yao Zhao, Stephan Lee, Pandu Nayak, Doug Fritz, Manish Reddy Vuyyuru, John Aslanides, Nidhi Vyas, Martin Wicke, Xiao Ma, Evgenii Eltsyhev, Nina Martin Hardie Cate, James Manyika, Keyvan Amiri, Yelin Kim, Xi Xiong, Kai Kang, Florian Luisier, Nilesh Tripuraneni, David Madras, Mandy Guo, Austin Waters Oliver Wang Joshua Ainslie Jason Baldridge Han Zhang Garima Pruthi Jakob Bauer Feng Yang Riham Mansour Jason Gelman Yang XuGeorge Polovets,Ji Liu Honglong CaiWarren ChenXiangHai ShengEmily XueSherjil Ozair Christof Angermueller Xiaowei Li Anoop Sinha Weiren Wang Julia Wiesinger Emmanouil Koukoumidis Yuan Tian Anand Iyer Madhu Gurumurthy Mark Goldenson Parashar Shah MK Blake Hongkun Yu Anthony Urbanowicz Jennimaria Palomaki Chrisantha Fernando Ken Durden Harsh Mehta Nikola Momchev,Elahe Rahimtoroghi Maria Georgaki Amit Raul Sebastian Ruder Morgan Redshaw Jinhyuk Lee Denny Zhou Komal Jalan Dinghua Li Blake Hechtman Parker Schuh Milad Nasr Kieran Milan Vladimir Mikulik Juliana Franco Tim Green Nam Nguyen Joe Kelley Aroma Mahendru Andrea Hu Joshua Howland Ben Vargas Jeffrey Hui Kshitij Bansal,Vikram Rao Rakesh Ghiya Emma Wang Ke Ye Jean Michel Sarr Melanie Moranski Preston Madeleine Elish Steve Li Aakash Kaku Jigar Gupta Ice Pasupat Da-Cheng Juan Milan Someswar Tejvi M., Xinyun Chen,Aida Amini Alex Fabrikant Eric Chu Xuanyi Dong Amruta Muthal Senaka Buthpitiya Sarthak Jauhari Nan Hua Urvashi Khandelwal Ayal Hitron Jie Ren Larissa Rinaldi Shahar Death Avigail Dabush NanJiang Jiang Harshal Godhia Uli Sachs Anthony Chen Yicheng Fan Hagai Taitelbaum Hila Noga Zhuyun Dai James WangChen Liang Jenny Hamer Chun-Sung Ferng Chenel Elkind Aviel Atias Paulina Lee,Vit Listik Mathias Carlen Jan van de Kerkhof Marcin Pikus Krunoslav Za + +her, Paul Müller, Sasha Zykova, Richard Stefanec, Vitaly Gatsko, Christoph Hirnschall, Ashwin Sethi, Xingyu Federico Xu, Chetan Ahuja, Beth Tsai, Anca Stefanoiu, Bo Feng, Keshav Dhandhania, Manish Katyal, Akshay Gupta, Atharva Parulekar, Divya Pitta, Jing Zhao, Vivaan Bhatia, Yashodha Bhavnani, Omar Alhadlaq, Xiaolin Li, Peter Danenberg, Dennis Tu, Alex Pine, Vera Filippova, Abhipso Ghosh, Ben Limonchik, Bhargava Urala, Chaitanya Krishna Lanka, Derik Clive, Yi Sun, Edward Li, Hao Wu, Kevin Hongtongsak, Ianna Li, Kalind Thakkar, Kuanysh Omarov, Kushal Majmundar, Michael Alverson, Michael Kucharski, Mohak Patel, Mudit Jain, Maksim Zabelin, Paolo Pelagatti, Rohan Kohli, Saurabh Kumar, Joseph Kim, Swetha Sankar, Vineet Shah, Lakshmi Ramachandruni, Xiangkai Zeng, Ben Bariach, Laura Weidinger, Tu Vu, Alek Andreev, Antoine He, Kevin Hui, Sheleem Kashem, Amar Subramanya, Sissie Hsiao, Demis Hassabis, Koray Kavukcuoglu, Adam Sadovsky, Quoc Le, Trevor Strohman, Yonghui Wu, Slav Petrov, Jeffrey Dean, and Oriol Vinyls. Gemini: A family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2024. 2 + +[41] Haoning Wu, Hanwei Zhu, Zicheng Zhang, Erli Zhang, Chaofeng Chen, Liang Liao, Chunyi Li, Annan Wang, Wenxiu Sun, Qiong Yan, Xiaohong Liu, Guangtao Zhai, Shiqi Wang, and Weisi Lin. Towards open-ended visual quality comparison. arXiv preprint arXiv:2402.16641, 2024. 4, 6 +[42] Tong Wu, Guandao Yang, Zhibing Li, Kai Zhang, Ziwei Liu, Leonidas Guibas, Dahua Lin, and Gordon Wetzstein. Gpt-4v(ision) is a human-aligned evaluator for text-to-3d generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2 +[43] Zhiyu Wu, Xiaokang Chen, Zizheng Pan, Xingchao Liu, Wen Liu, Damai Dai, Huazuo Gao, Yiyang Ma, Chengyue Wu, Bingxuan Wang, Zhenda Xie, Yu Wu, Kai Hu, Jiawei Wang, Yaofeng Sun, Yukun Li, Yishi Piao, Kang Guan, Aixin Liu, Xin Xie, Yuxiang You, Kai Dong, Xingkai Yu, Haowei Zhang, Liang Zhao, Yisong Wang, and Chong Ruan. Deepseek-vl2: Mixture-of-experts vision-language models for advanced multimodal understanding. arXiv preprint arXiv:2412.10302, 2024. 2, 3, 5 +[44] Xiaohan Xu, Ming Li, Chongyang Tao, Tao Shen, Reynold Cheng, Jinyang Li, Can Xu, Dacheng Tao, and Tianyi Zhou. A survey on knowledge distillation of large language models. arXiv preprint arXiv:2402.13116, 2024. 2 \ No newline at end of file diff --git a/ICCV/2025/Visual Interestingness Decoded_ How GPT-4o Mirrors Human Interests/images.zip b/ICCV/2025/Visual Interestingness Decoded_ How GPT-4o Mirrors Human Interests/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..5761e50f72a3a86a34addda5a643d7fa230075e7 --- /dev/null +++ b/ICCV/2025/Visual Interestingness Decoded_ How GPT-4o Mirrors Human Interests/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fad375e253e6f267062bb5a021d1a8b156d15a8a63265db47bb4ae692d5f4343 +size 596751 diff --git a/ICCV/2025/Visual Interestingness Decoded_ How GPT-4o Mirrors Human Interests/layout.json b/ICCV/2025/Visual Interestingness Decoded_ How GPT-4o Mirrors Human Interests/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..e5b323b2edde5c35773f9e0d2938499757e61860 --- /dev/null +++ b/ICCV/2025/Visual Interestingness Decoded_ How GPT-4o Mirrors Human Interests/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:62050644733cdf242541919a7bba3a7013dd275c5d7bbcdef23a16a677a471ec +size 437009 diff --git a/ICCV/2025/Visual Modality Prompt for Adapting Vision-Language Object Detectors/fcab79ff-b3e2-457c-a329-0dcc0ab0c321_content_list.json b/ICCV/2025/Visual Modality Prompt for Adapting Vision-Language Object Detectors/fcab79ff-b3e2-457c-a329-0dcc0ab0c321_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..79970f9b518e926fa1c30968ff940f53899a9b0b --- /dev/null +++ b/ICCV/2025/Visual Modality Prompt for Adapting Vision-Language Object Detectors/fcab79ff-b3e2-457c-a329-0dcc0ab0c321_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b2af76358004cbec157afa29b06a0d27179c7bff0efc7a3d6f600182516af3ab +size 76419 diff --git a/ICCV/2025/Visual Modality Prompt for Adapting Vision-Language Object Detectors/fcab79ff-b3e2-457c-a329-0dcc0ab0c321_model.json b/ICCV/2025/Visual Modality Prompt for Adapting Vision-Language Object Detectors/fcab79ff-b3e2-457c-a329-0dcc0ab0c321_model.json new file mode 100644 index 0000000000000000000000000000000000000000..d043f8feae43e2ccb336479b3f057b3299868111 --- /dev/null +++ b/ICCV/2025/Visual Modality Prompt for Adapting Vision-Language Object Detectors/fcab79ff-b3e2-457c-a329-0dcc0ab0c321_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8984edd0e68a3dcf0fe3e76084aea742858d5790489a9ca8b7030836d8ce8ff8 +size 94679 diff --git a/ICCV/2025/Visual Modality Prompt for Adapting Vision-Language Object Detectors/fcab79ff-b3e2-457c-a329-0dcc0ab0c321_origin.pdf b/ICCV/2025/Visual Modality Prompt for Adapting Vision-Language Object Detectors/fcab79ff-b3e2-457c-a329-0dcc0ab0c321_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..948cb10a41cd32f6fbf5689b557b89207e90fc4d --- /dev/null +++ b/ICCV/2025/Visual Modality Prompt for Adapting Vision-Language Object Detectors/fcab79ff-b3e2-457c-a329-0dcc0ab0c321_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d27e96d2b3dd16c978b3b2218d4f3b793acee1b6506ccaac4d745e577bf035d3 +size 28166411 diff --git a/ICCV/2025/Visual Modality Prompt for Adapting Vision-Language Object Detectors/full.md b/ICCV/2025/Visual Modality Prompt for Adapting Vision-Language Object Detectors/full.md new file mode 100644 index 0000000000000000000000000000000000000000..06dde3f3f3cba10d283af30ceb66ec01918ae554 --- /dev/null +++ b/ICCV/2025/Visual Modality Prompt for Adapting Vision-Language Object Detectors/full.md @@ -0,0 +1,252 @@ +# Visual Modality Prompt for Adapting Vision-Language Object Detectors + +Heitor R. Medeiros* Atif Belal Srikanth Muralidharan Eric Granger Marco Pedersoli + +LIVIA, Dept. of Systems Engineering, ETS Montreal, Canada International Laboratory on Learning Systems (ILLS) + +![](images/e7f8d3c16f2b50c156d86f3894b6ccd24a64ba18c6a0e45f2df64387ef25d735.jpg) +Figure 1. YOLO-World detections of different approaches across modalities: LLVIP and FLIR datasets (infrared) and $\mathrm{NYU}_{v2}$ (depth). Each column corresponds to a different approach: (a) Ground Truth (GT): Shows in yellow the GT bounding boxes for objects. (b) Zero-Shot (ZS): Displays detections (in red) from a zero-shot model. It has missed several detections and some inaccurate boxes without tuning. (c) Visual Prompt: Illustrates detections with a visual prompt added to the image. It shows improvements over ZS, with more accurate detection in certain areas, but still misses some objects. (d) ModPrompt (Ours): Detections from our model. ModPrompt generates artifacts on the image to enhance objects and suppress the background, facilitating detection. + +# Abstract + +The zero-shot performance of object detectors degrades when tested on different modalities, such as infrared and depth. While recent work has explored image translation techniques to adapt detectors to new modalities, these methods are limited to a single modality and traditional detectors. Recently, vision-language detectors (VLDs), such asYOLO-World and Grounding DINO, have shown promising zero-shot capabilities; however, they have not yet been adapted for other visual modalities. Traditional fine-tuning approaches compromise the zero-shot capabilities of the detectors. The visual prompt strategies commonly used for classification with vision-language models apply the same linear prompt translation to each image, making them less effective. To address these limitations, we propose Mod + +Prompt, a visual prompt strategy to adapt VLDs to new modalities without degrading zero-shot performance. In particular, an encoder-decoder visual prompt strategy is proposed, further enhanced by the integration of inference-friendly modality prompt decoupled residual, facilitating a more robust adaptation. Empirical benchmarking results show our method for modality adaptation on YOLO-World and Grounding DINO for challenging infrared (LLVIP, FLIR) and depth $(NYU_{v2})$ datasets, achieving performance comparable to full fine-tuning while preserving the model's zero-shot capability. Our code is available at https://github.com/heitorrapela/ModPrompt. + +# 1. Introduction + +Object detection (OD) is a key challenge in computer vision, aiming to localize and classify objects in images [57]. + +![](images/6e87076e0aa77ae833943eb41b3f26cb08b2edc79535068ef6d916df1aa0ed15.jpg) +Figure 2. Strategies to adapt object detectors to new modalities: (a) Full Fine-tuning: Both the backbone (the part of the model responsible for feature extraction) and the head (responsible for the final output, like object detection) are updated with new training data. (b) Head Fine-tuning: Only the head is fine-tuned while the backbone remains frozen. (c) Visual Prompt: Uses a visual prompt added to the input. The backbone and head remain unchanged, but the visual prompt guides the model to better interpret the new modality. (d) Our Modality Prompt. Similarly to a visual prompt, the input image is added to a visual prompt. The main difference is that here the prompt is not static; it is a transformation of the input image. + +Recent OD advancements, driven by applications like autonomous driving [35, 44], surveillance [10, 40], and robotics [11, 19, 38], have improved real-world systems significantly. Although traditional detectors like YOLO [21] or DINO [52] perform well, they are limited to fixed vocabularies, such as the 80 categories in the COCO [29] dataset, and exhibit poor zero-shot performance [2]. These detectors are trained on predefined categories, restricting their adaptability to varying scenarios. In contrast, Vision-Language Models (VLMs) combine visual representations and semantic text embeddings, allowing them to dynamically understand pixel-level features in context [39]. VLMs enhance accuracy and flexibility by recognizing unseen objects through the integration of language, paving the way for open-vocabulary detection [49], such as Grounding DINO [30] and YOLO-World [7]. + +Generally, these open-vocabulary detectors are pretrained on large-scale RGB datasets, making them capable of zero-shot detection [30, 51]. However, their performance degrades if the domain shift is large, as in a different modality, such as infrared or depth. A common approach to mitigate this is by fine-tuning the VLMs on the downstream modality dataset. However, it makes the VLMs lose their zero-shot capabilities due to catastrophic forgetting [24]. Some techniques have been explored to adapt the VLMs to downstream tasks for image classification. These include text prompt tuning (TPT) [54, 55] and visual prompt tuning (VPT) [1, 20]. Unlike traditional fine-tuning, prompt tuning involves adding learnable prompts to the model's in + +put, allowing it to remain unchanged and preserve its zero-shot capability. VPT methods for classification learn visual prompt tokens to adapt the classifier of transformer-based models to the downstream tasks. Visual prompts are effective when adapting VLMs to a downstream task of the same modality (normally RGB). However, these methods are also less effective if the downstream data is from another visual modality with a large modality shift between pre-trained and downstream data, as shown in Table 1. + +In addition to incorporating language into detectors, another major driver of advancements in real-world applications is the integration of diverse visual modalities, such as infrared [26, 34, 45] and depth [27, 46, 47]. These modalities provide additional spatial information, enhancing visibility in low-light and obstructed conditions, resulting in more robust and accurate object detectors [4]. Adapting detectors for different modalities unlocks their potential in specialized contexts by combining the efficiency of transfer learning with modality-specific enhancements. This approach bridges the gap between general-purpose detectors capable of zero-shot detection and the nuanced demands of novel modalities, where labels are still scarce compared to RGB data. + +Therefore, to overcome limitations in modality adaptation, image translation [16, 36] methods have been explored for traditional detector-based models. However, these approaches primarily optimize the image translation loss rather than directly focusing on improving detection performance. More recent approaches, such as Hallu + +ciDet [33] and ModTr [32], have demonstrated improvement by emphasizing detection loss optimization during adaptation. However, these methods are limited to traditional detectors without language mechanisms, and they did not explore more challenging modalities such as depth. For HalluciDet, the prior pre-trained knowledge is lost during the adaptation phase. In ModTr, the COCO pre-trained knowledge is preserved, but it lacks the zero-shot powerful capability of VLMs. In contrast, our work leverages recent vision-language detectors, incorporating textual information and allowing for adaptation to different modalities, exploring the benefits of both visual and text information for improving modality adaptation. + +In this paper, we propose ModPrompt, a visual prompt strategy to adapt vision-language ODs in the input space to new visual modalities without sacrificing zero-shot knowledge of the vision-language detector. ModPrompt is a detector-agnostic encoder-decoder visual prompt strategy that can easily be integrated with any VLM detector, regardless of the detector's backbone type. We also introduce an inference-friendly residual mechanism named Modality Prompt Decoupled Residual (MPDR) capable of adapting the learnable text embeddings with the ModPrompt loss, while preserving full knowledge of the model due to the decoupled embedding parameters. We demonstrate that the combination of MPDR with ModPrompt improves the performance of the detector across different modalities while retaining the original zero-shot knowledge, both in the vision and language components. To the best of our knowledge, this is the first work that focuses on adapting state-of-the-art VLM-based OD models to new modalities. Furthermore, ModPrompt is validated on two state-of-the-art VLM ODs, Grounding DINO and YOLO-World, and across four widely used visual modalities, infrared, depth, event-based, and LiDAR. + +# Our main contributions can be summarized as follows. + +(1) We introduce ModPrompt, a novel pixel-level prompt adaptation strategy based on input translation. Our method is backbone-agnostic, leading to improved adaptability of vision-language detectors across different modalities and preserving the vision encoder knowledge. Additionally, we introduce MPDR, a text-embedding modality prompt decoupled residual mechanism, which preserves full knowledge of the text-embedding while increasing the performance of the model for the novel modality. +(2) We conduct a detailed investigation of various pixel-level visual prompt techniques, offering insights into why traditional prompt methods struggle with adapting object detectors to new modalities. Our analysis demonstrates how ModPrompt overcomes these limitations to significantly enhance detection performance. +(3) We provide a comprehensive benchmark on the vi- + +sual prompts for modality adaptation across two primary open-vocabulary OD models, YOLO-World and Grounding DINO. We evaluate our method across infrared, depth, even-based, and LiDAR datasets, achieving in some cases a level of performance comparable to full fine-tuning while preserving the model's zero-shot capability. + +# 2. Related Works + +Open-Vocabulary Object Detection. Recent advances in open-vocabulary OD exploit vision-language models (VLMs) with region-level multimodal alignment to recognize novel concepts. RegionCLIP [53] uses pseudo labels for contrastive pretraining, while Gu et al. [14] employ knowledge distillation with CLIP-based region embeddings. Lin et al. [28] leverage bipartite matching for image-text alignment. GLIP [25] and GLIPv2 [51] enhance zero-shot detection via multimodal fusion and region-grounded pretraining. Grounding DINO [30] introduces multiphase image-text fusion, and YOLO-World [7] enables efficient open-vocabulary detection with reparametrizable fusion. + +Downstream Adaptation of Large Vision Models. Recent work has focused on parameter-efficient adaptation of large vision models without erasing prior knowledge. Co-op [55] learns continuous text prompts for downstream tasks. Visual Prompt Tuning [20] adds tunable parameters at various network stages. Bahng et al. [1] propose image-space linear probing. CLIP-Adapter [12] introduces feature-space adaptation via bottlenecks and residuals. Yu et al. [48] use task residuals to decouple new parameters for classification. While these methods seem effective, they are not suitable for detection. + +Image Translation for Detection. Image translation methods adapt images across domains by modifying domain-specific characteristics while preserving content [37], often using generative models [13, 23]. Pix2Pix [18] leverages paired supervision, while CycleGAN [56] enables unpaired translation. For OD, HalluciDet [33] performs a two-step adaptation using RGB fine-tuning and translation, whereas ModTr [32] employs zero-shot COCO-trained detectors with fusion strategies to preserve prior knowledge. In contrast, we adapt large VLM-based ODs (e.g., YOLO-World, Grounding DINO) to novel modalities (e.g., IR, depth) while retaining their zero-shot capabilities. + +# 3. Proposed Method + +# 3.1. Preliminary Definitions + +Vision Language Object Detection. The training dataset for traditional object detectors can be represented as $\mathcal{D} =$ + +
DatasetMethodYOLO-WorldGrounding DINO
AP50AP75APAP50AP75AP
LLVIP - IRZero-Shot (ZS)81.00 ± 0.0057.80 ± 0.0053.20 ± 0.0085.50 ± 0.0062.70 ± 0.0056.50 ± 0.00
Head Finetuning (HFT)93.57 ± 0.0573.83 ± 0.1964.80 ± 0.0887.53 ± 0.0665.57 ± 0.2358.10 ± 0.20
Full Finetuning (FT)97.43 ± 0.0577.93 ± 0.2167.73 ± 0.0997.17 ± 0.3179.93 ± 0.8367.83 ± 0.96
Visual Prompt (Fixed)70.30 ± 7.8945.67 ± 6.9743.53 ± 5.7983.83 ± 0.0661.53 ± 0.2355.13 ± 0.15
Visual Prompt (Random)60.13 ± 0.2938.73 ± 0.1736.87 ± 0.1283.87 ± 0.0661.37 ± 0.0655.03 ± 0.06
Visual Prompt (Padding)79.87 ± 1.0051.77 ± 0.9049.30 ± 0.8382.73 ± 0.3160.00 ± 0.3555.13 ± 0.15
Visual Prompt (WM)82.00 ± 1.5953.90 ± 1.0650.90 ± 0.9469.57 ± 0.9341.37 ± 1.2740.77 ± 0.87
Visual Prompt (WMv2)74.10 ± 0.4346.47 ± 0.6244.70 ± 0.2269.87 ± 1.1241.77 ± 1.3041.13 ± 0.96
ModPrompt (Ours)92.80 ± 0.2970.73 ± 1.0262.87 ± 0.6393.13 ± 0.1567.17 ± 0.7860.10 ± 0.50
NYUv2 - DepthZero-Shot (ZS)04.80 ± 0.0003.10 ± 0.0003.00 ± 0.0008.30 ± 0.0005.60 ± 0.0005.30 ± 0.00
Head Finetuning (HFT)21.03 ± 0.1212.03 ± 0.3712.00 ± 0.1418.24 ± 0.8513.23 ± 0.4612.10 ± 0.21
Full Finetuning (FT)49.90 ± 0.0836.40 ± 0.2933.57 ± 0.2651.60 ± 2.0939.17 ± 1.8835.77 ± 1.70
Visual Prompt (Fixed)04.67 ± 0.0503.07 ± 0.0502.90 ± 0.0008.27 ± 0.0605.57 ± 0.0605.27 ± 0.06
Visual Prompt (Random)04.23 ± 0.1202.63 ± 0.0502.53 ± 0.0508.33 ± 0.0605.53 ± 0.0605.27 ± 0.06
Visual Prompt (Padding)03.97 ± 0.0502.50 ± 0.0002.43 ± 0.0507.63 ± 0.0605.17 ± 0.0604.63 ± 0.06
Visual Prompt (WM)10.67 ± 0.1206.90 ± 0.2206.57 ± 0.0504.87 ± 0.0602.97 ± 0.0602.97 ± 0.06
Visual Prompt (WMv2)10.63 ± 0.1206.67 ± 0.1206.50 ± 0.0805.00 ± 0.1003.07 ± 0.0603.03 ± 0.06
ModPrompt (Ours)37.17 ± 0.5727.50 ± 0.6424.93 ± 0.5021.70 ± 0.2015.03 ± 0.2914.13 ± 0.21
+ +Table 1. Detection performance (APs) for YOLO-World and Grounding DINO for the two main datasets evaluated: LLVIP-IR and NYUv2-Depth. The different visual prompt adaptation techniques are compared with our ModPrompt, and the zero-shot (ZS), head finetuning (HFT), and full finetuning (FT) are also reported, where the full finetuning is the upper bound. + +$\{(X,Y)\}_{j = 1}^{M}$ .Here, $X\in \mathbb{R}^{W\times H\times C}$ represents an image with dimensions $W\times H$ and $C$ channels, and $Y = \{(b_{i},c_{i})\}_{i = 1}^{N}$ consist of bounding boxes $b_{i}$ and object category $c_{i}$ .For the visual language model, the annotations are reformulated as a region-text pair $Y = \{(b_i,t_i)\}_{i = 1}^N$ ,where $t_i$ corresponds to the text for the region $b_{i}$ .Specifically, $t_i$ is the name of the object category in $b_{i}$ . The objective of the object detection task is to accurately identify all objects in a given image. The average precision (AP) metric across classes is used as the standard evaluation protocol. An object detector is formally represented as the mapping $f_{\theta}:\mathbb{R}^{W\times H\times C}\to \hat{Y}$ , where $\theta$ denotes the parameter vector. The detection cost function $\mathcal{C}_{det}(\theta)$ , a differentiable proxy for the AP metric, is computed as an average loss of the detection loss $\mathcal{L}_{det}$ over the entire dataset $\mathcal{D}$ ,described mathematically as: + +$$ +\mathcal {C} _ {d e t} (\theta) = \frac {1}{| \mathcal {D} |} \sum_ {(x, Y) \in \mathcal {D}} \mathcal {L} _ {d e t} (f _ {\theta} (x), Y). \tag {1} +$$ + +Visual Prompt for OD. Given a frozen pre-trained detector, the objective of the Visual Prompt method is to learn a task-specific visual prompt $v_{\phi}$ , parametrized by $\phi$ , which can be combined with the input $x$ to improve the final detection performance. During test time, this visual prompt is added to the test images to adapt them for the desired detection task. As described in the following equation: + +$$ +\mathcal {C} _ {\mathrm {v p}} (\phi) = \frac {1}{| \mathcal {D} |} \sum_ {(x, Y) \in \mathcal {D}} \mathcal {L} _ {\text {d e t}} \left(f _ {\theta} \left(x + v _ {\phi}\right), Y\right). \tag {2} +$$ + +# 3.2. ModPrompt + +The visual prompt methods discussed above learn simple linear transformations, like adding a fixed patch to each image. These transformations are learned during the training process and are not conditioned on the input image at inference, making them less effective. In ModPrompt, we incorporate a function $h_{\vartheta}$ , an encoder-decoder inspired by U-Net [41], dependent on the input image $x$ , which is trained conditioned on labels $Y$ . Our encoder-decoder is flexible in terms of the encoder backbone initialized from pre-trained weights. We adapt the last layer of the decoder to constrain it to be 3-channels with values ranging between 0 and 1 to simulate a pseudo-RGB image. Note that this mechanism is completely different from traditional U-Nets, which are trained to reconstruct and generate $c$ -channels for semantic segmentation tasks, requiring both input and ground truth segmentation map during training. However, ModPrompt requires access only to the pre-trained RGB detector and final target modality data, so it is based on guidance detection loss instead of reconstruction. The ModPrompt training cost is defined by the following equation: + +$$ +\mathcal {C} _ {\mathrm {m p}} (\vartheta) = \frac {1}{| \mathcal {D} |} \sum_ {(x, Y) \in \mathcal {D}} \mathcal {L} _ {\text {d e t}} \left(f _ {\theta} \left(x + h _ {\vartheta} (x)\right), Y\right). \tag {3} +$$ + +This translation function $(f_{\theta})$ is responsible for adapting the input image to a modality representation optimally suited to address detection in the target modality. Our + +idea of a visual ModPrompt is inspired by text conditional prompt [54] work, which incorporates a small network for the text encoder conditional adaptation. In our approach, the visual prompt depends on the input image. + +# 3.3. Modality Prompt Decoupled Residual with Knowledge Preservation (MPDR) + +The vision-language models take both text and image as input. ModPrompt aims to transform the input image to adapt to the new modality. For the adaptation of the textual part, text prompt tuning methods [54, 55] have been explored for classification. But these methods require backpropagating through the whole text backbone, making them computationally expensive. Recently, task residual [48] was proposed that tunes the precomputed text embedding of the vision-language classifiers. Inspired by this, we design a strategy to efficiently tune the embedding of the vision-language detectors. As illustrated in Figure 3, offline text embeddings are generated for each object category. Then, the MPDR, which are learnable embedding tokens, are tuned to adapt to a specific modality end-to-end with ModPrompt loss. The MPDR is crucial for decoupling the knowledge from the original embedding, which preserves prior text-embedding knowledge and is able to incorporate new knowledge during the adaptation stage. During the training, the loss of ModPrompt and MPDR work in synergy to adapt to any new modality. In the test phase, the MPDR can be deactivated easily by zero-masking its parameters, which is the same as full zero-shot embedding knowledge, or enabled, which is the adapted embedding for this modality (that's why decoupling the knowledge is important) and doesn't introduce any inference overhead. The following equation describes the final training cost: + +$$ +\mathcal {C} _ {\mathrm {m p - t p}} (\vartheta , \phi) = \mathcal {C} _ {\mathrm {m p}} (\vartheta) + \mathcal {C} _ {\mathrm {t p}} (\phi), \tag {4} +$$ + +where $\mathcal{C}_{\mathrm{mp}}(\vartheta)$ is the optimization of parameters $\vartheta$ of the encoder-decoder function, and $\mathcal{C}_{\mathrm{tp}}(\phi)$ is the text-prompt adaptation on the embedding space by training new parameters $\phi$ for the class embeddings in the offline vocabulary. + +# 3.4. Training Summary + +For training the ModPrompt with MPDR, we first precompute the text embeddings for the target classes, using the text encoder. Then, we sum the learnable embedding residuals with the frozen embeddings and train them together with the ModPrompt. Note that, different from other embedding adaptation approaches, our MPDR can be used with ModPrompt for full zero-shot knowledge preservation due to the residual decoupled embedding learning strategy. The cost function is based on Equation 4, but instead of the online cost function for text $\mathcal{C}_{\mathrm{tp}}(\phi)$ , we work on the text em + +![](images/711f3ce147a17d75ba8fd0729ad4198d85816fd51f8c113c7a2ebcf95a1c54d4.jpg) +Figure 3. Our proposed strategy for text-prompt tuning: an inference-friendly and knowledge-preserving decoupled embedding tuning method. An offline embedding is generated for each object category, and then a novel decoupled residual trainable parameters and the ModPrompt are integrated into the detector to adapt it to new modalities. + +bedding space, which is better suited for efficient training and inference. + +# 4. Results and Discussion + +# 4.1. Experimental Methodology + +(a) Datasets: LLVIP: LLVIP is a surveillance dataset composed of 12,025 IR images in the training set and 3,463 IR images in the test set with only pedestrians annotated. FLIR ALIGNED: We used the sanitized and aligned paired sets provided by Zhang et al. [50]. It has 4,129 training IR images and 1,013 IR test images. FLIR images are captured from the perspective of a camera in the front of a car, with a resolution of 640 by 512. It contains the bicycles, dogs, cars, and people classes. Following Cao et al. [5], we decided to remove "dog" objects because the number of annotations is inadequate for training with the FLIR dataset; results of FLIR can be seen in the supp. material due to space constraints. $\mathbf{NYU}_{v2}$ : The NYU-Depth V2 [43] dataset provides a valuable collection of indoor video sequences captured with a Microsoft Kinect camera. In this dataset, we used the depth information, which is composed of 795 training images and 654 test images with a resolution of 640 by 480. The dataset comprised 19 different classes, such as bathtubs, beds, bookshelves, boxes, and chairs. In all our experiments, only IR and depth images were used. + +(b) Implementation Details: The YOLO-World models were trained on an A100 NVIDIA GPU, and Grounding DINO on V100, and were implemented using PyTorch. For YOLO-World, we used the original work [7], which is based on MMDET [6], and for Grounding DINO, we used the open-source version from the MMDET library. For the YOLO-World, we use both YOLO-World-S-v1 and YOLO- + +
DetectorMethodLLVIP-IRNYUv2-DEPTH
AP50AP75APAP50AP75AP
YOLO-WorldFixed86.60 ± 0.00 (+16.3)63.53 ± 0.05 (+17.8)57.67 ± 0.05 (+14.1)12.47 ± 0.05 (+7.80)07.50 ± 0.00 (+4.43)07.37 ± 0.05 (+4.47)
Random86.43 ± 0.05 (+26.3)63.40 ± 0.08 (+24.6)57.50 ± 0.00 (+20.6)12.23 ± 0.05 (+8.00)07.30 ± 0.08 (+4.67)07.20 ± 0.00 (+4.67)
Padding83.57 ± 0.12 (+3.70)59.03 ± 0.05 (+7.26)54.23 ± 0.05 (+4.93)10.63 ± 0.09 (+6.66)05.60 ± 0.08 (+3.10)05.97 ± 0.05 (+3.54)
WeightMap87.47 ± 0.17 (+5.47)62.23 ± 0.31 (+8.33)57.43 ± 0.05 (+6.53)13.43 ± 0.09 (+2.76)08.07 ± 0.12 (+1.17)07.90 ± 0.08 (+1.33)
ModPrompt96.60 ± 0.37 (+3.80)77.27 ± 0.33 (+6.54)67.37 ± 0.05 (+4.50)44.67 ± 0.17 (+7.50)32.53 ± 0.66 (+5.03)29.93 ± 0.12 (+5.00)
Grounding DINOFixed88.70 ± 0.66 (+4.87)67.73 ± 0.50 (+6.20)60.13 ± 0.45 (+5.00)10.00 ± 0.00 (+1.73)06.40 ± 0.28 (+0.83)06.40 ± 0.28 (+1.13)
Random88.80 ± 0.95 (+4.93)68.00 ± 1.21 (+6.63)60.37 ± 0.90 (+5.33)09.90 ± 0.14 (+1.57)06.50 ± 0.00 (+0.97)06.15 ± 0.07 (+0.88)
Padding86.93 ± 0.55 (+4.20)65.23 ± 1.24 (+5.23)58.07 ± 1.10 (+4.17)09.27 ± 0.15 (+1.64)06.00 ± 0.17 (+0.83)05.57 ± 0.15 (+0.87)
WeightMap74.90 ± 0.92 (+5.33)46.80 ± 0.70 (+5.43)45.77 ± 0.46 (+5.00)06.17 ± 0.12 (+1.30)03.53 ± 0.15 (+0.56)03.60 ± 0.10 (+0.63)
ModPrompt93.73 ± 0.25 (+0.60)68.27 ± 0.40 (+1.10)60.87 ± 0.50 (+0.77)26.63 ± 0.38 (+4.93)18.53 ± 0.32 (+3.50)17.27 ± 0.46 (+3.14)
+ +World-S-v2, with the YOLOv8 backbone and CLIP [39] as the language model. For Grounding DINO, we use Swin-Transformer Tiny [31] as the detector backbone and BERT [9] for the language encoder. Our models are evaluated in terms of AP performance using COCO API [29]. The batch size for YOLO-World models was 8, and the models were trained with a single GPU. For Grounding DINO, the batch size was 16, and 2 GPUs were used for training using DDP. Yolo-World and Grounding Dino were trained with the maximum training epoch of 80 and 60, respectively. For offline embedding extraction, we used CLIP-ViT-base-patch32 and Bert-base-uncased for YOLO-World and Grounding DINO, respectively. Additional details are provided in the supp. materials. + +# (c) Baseline Methods: + +- Zero-shot Model. The zero-shot model refers to performing the inference with the original pre-trained open-vocabulary detector model, but on our final detection modality. Normally, these models are trained with large RGB datasets, such as Object365 [42] or GoldG [22]. + +- Finetuning. The finetuning strategy refers to adapting a pre-trained model by changing its parameters directly for the final task, but it destroys the pre-training knowledge of the model. The most common strategies for adapting detectors are the head finetuning, which adapts only the head of the detector, and full finetuning, which adapts all the weights of the detector. + +- Visual Prompt. For this baseline, we define a family of different visual prompt strategies explored in our work. Specifically, we have fixed, random, padding, and weight map. In fixed, we add a fixed patch in a fixed position for all training. In random, we have a random padding that is sampled per image. In padding, we add a padding prompt and combine it with the image. In weight map, we consider the learnable weights to be a mask with the dimensions of the input image. + +- Modality Prompt. Here, we refer to our proposed encoder-decoder visual prompt, in which we learn how to translate the input, guided by the final detection loss, and then we combine its output with the original input. + +Table 2. Detection performance (APs) for YOLO-World and Grounding DINO on LLVIP-IR and $\mathbf{NYU}_{v2}$ -Depth datasets. Each visual prompt adaptation strategy is compared with the learnable MPDR (results in parentheses are the gain with the MPDR module), which is responsible for updating the new modality embeddings and not changing the original embedding knowledge. + +
MethodVariationLLVIP - IR
AP50AP75AP
Fixed3061.60 ± 0.7539.93 ± 0.5237.97 ± 0.56
30070.30 ± 7.8945.67 ± 6.9743.53 ± 5.79
Random3060.13 ± 0.2938.73 ± 0.1736.87 ± 0.12
30056.27 ± 0.4633.73 ± 0.6233.13 ± 0.42
Padding3079.87 ± 1.0051.77 ± 0.9049.30 ± 0.83
30039.53 ± 2.3615.90 ± 1.0219.07 ± 1.18
ModPromptMB92.80 ± 0.2970.73 ± 1.0262.87 ± 0.63
RES91.03 ± 0.1268.40 ± 1.1061.43 ± 0.58
MethodVariationNYUv2 - Depth
AP50AP75AP
Fixed3004.67 ± 0.0503.07 ± 0.0502.90 ± 0.00
30003.43 ± 0.0502.00 ± 0.0802.10 ± 0.00
Random3004.23 ± 0.1202.63 ± 0.0502.53 ± 0.05
30001.53 ± 0.1700.77 ± 0.1200.87 ± 0.12
Padding3003.97 ± 0.0502.50 ± 0.0002.43 ± 0.05
20000.37 ± 0.1200.10 ± 0.0800.17 ± 0.05
ModPromptMB35.37 ± 0.1225.20 ± 0.2423.27 ± 0.17
RES37.17 ± 0.5727.50 ± 0.6424.93 ± 0.50
+ +Table 3. Detection performance (APs) for YOLO-World under the two main datasets evaluated: LLVIP-IR and $\mathbf{NYU}_{v2}$ -Depth. We compared the main visual prompt strategies fixed, random, padding, and ModPrompt. The variations consist of the number of prompt pixels ( $p_s = 30$ , 200, or 300) and for ModPrompt, the MobileNet (MB) or ResNet (RES). + +![](images/4067800effe26d7a347e21240215c57555f6d1d139fb38b1705692c169b60746.jpg) +Figure 4. Detection performance on LLVIP dataset of different SOTA Modality Translation OD methods in terms of APs. + +# 4.2. Visual Modality Adaptation + +This section explores two main open-vocabulary detectors, YOLO-World and Grounding DINO, for the input space visual prompt adaptation on IR and depth modalities. In the Table 1, we tabulate the results using YOLO-World and Grounding DINO. We compare zero-shot (ZS), full-finetuning (FT) of the visual backbone, visual prompt with fixed patch, random patch, padding patch, weight map (WM) on all pixels, and weight map with scale shift $(\mathrm{WM}_{v2})$ , and our ModPrompt. Every visual prompt strategy adapts at the input level, so we keep all other parameters frozen to avoid catastrophic forgetting. In this setup, the YOLO-World with LLVIP-IR dataset, we observe that ModPrompt outperforms the other prompt strategies with $\mathrm{AP}_{50}$ 92.80 and the second best was WM with 82.00, similar trends for the $\mathrm{AP}_{75}$ with ModPrompt reaching 70.73 and WM with 53.90. Additionally, in terms of AP, our method has 62.87 AP, while the second-best approach has 50.90 AP. For Grounding DINO, in LLVIP-IR, our ModPrompt had an 93.13 $\mathrm{AP}_{50}$ , 67.17 $\mathrm{AP}_{75}$ , and 60.10 AP, and the second best (fixed patch) had an $\mathrm{AP}_{50}$ of 83.87, but close to random and padding prompts. We observe that ModPrompt performs better when objects are well-defined in the image and when objects are not too small, otherwise, like all other input-level pixel strategies it struggles, especially on refined bounding-box localization, which can be seen with $\mathrm{AP}_{75}$ and AP, whereas in $\mathrm{AP}_{50}$ it shows good results. Performance on the FLIR dataset is reported in the supp. material. + +Comparison with SOTA Modality Translation OD methods. Our ModPrompt technique is compared with recent works on modality translators for ODs: HalluciDet [33] and ModTr [32]. In Fig. 4, we observe that our results are better in all APs for the LLVIP dataset. Additional results for the FLIR-IR dataset are presented in the supplementary material. + +Comparison of the number of Train Parameters and Knowledge Preservation. In Table 4 we compare the + +![](images/b33c1848001749a052a3b1bc6535a5a42f4e999d965e9c72611a429b61ee3c96.jpg) +Figure 5. Detections for YOLO-World for the different approaches: First two rows for LLVIP (infrared), and last two rows for $\mathrm{NYU}_{v2}$ (depth). Each column corresponds to a different approach: (a) GT (Ground Truth): Shows in yellow the ground-truth bounding boxes for objects. (b) Zero-Shot (ZS): Displays detections (in red) from a zero-shot model. (c) Visual Prompt (VP): Illustrates detections with a visual prompt added to the image. (d) ModPrompt (MP) (Ours): Detections from our proposed model. + +number of training parameters and show the catastrophic forgetting in the compared baseline methods. As we can see, ModPrompt preserves the zero-shot performance on the COCO dataset, while the performance drops significantly in HFT and FT. Other visual prompt strategies, such as WM and $\mathrm{WM}_2$ , can preserve the zero-shot performance, but our average performance is much better than theirs while requiring fewer trainable parameters. + +
MethodParams (M)LLVIPCOCOAvg.
ZS0.0081.00 ± 0.0051.90 ± 0.0066.45 (00.00)
HFT2.3193.57 ± 0.0500.66 ± 0.0447.12 (-19.33)
FT76.8197.43 ± 0.0500.10 ± 0.0048.77 (-17.68)
WM3.9387.47 ± 0.1751.90 ± 0.0069.69 (+3.24)
WMv27.8685.00 ± 0.0851.90 ± 0.0068.45 (+2.00)
ModPrompt3.0895.63 ± 0.0451.90 ± 0.0073.77 (+7.32)
+ +Table 4. ${\mathbf{{AP}}}_{50}$ of YOLO-World on LLVIP-IR and COCO data. + +We compare the number of trainable parameters and show the catastrophic forgetting in HFT and FT baselines. + +# 4.3. Ablation Studies + +Visual Prompts. We evaluate different variations of the visual prompt adaptation methods introduced in our paper. Specifically, we compare the performance when different input patch sizes are used; for instance, $p_s = 30$ refers to a patch size of 30 pixels. In this study, we tested various patch sizes for each of the visual prompt methods and reported the performance in Table 3. We + +evaluate ModPrompt using two different translators with U-Net based backbones- MobileNet (MB) [17] and ResNet (RES) [15]. Empirically, we observe minimal variation in results when increasing the prompt patch size. In some cases, a larger prompt could even degrade the performance of the detector on the new modality, as seen with $\mathrm{p}_s = 200$ on $\mathrm{NYU}_{v2}$ datasets. A similar trend can be observed with ModPrompt, where sometimes the MB version, with far fewer parameters, can outperform the RES version. This finding suggests that a lighter MB translator can be a viable choice over a heavier RES translator, making it better suited for real-time performance, with minimal degradation in detection metrics. + +MPDR Knowledge Preservation mechanism. Table 2 shows that MDPR can be beneficial for improving the majority of the visual prompt strategies while preserving the original knowledge of the detector. For the text adaptation setting, we pre-compute the embeddings for the target classes using our language model, avoiding forward/backpropagation on it. In this setting, we also tried performing adaptations directly on the embeddings without using MPDR parameters, similar to text/embedding prompt tuning; in this case, the performance on the final target modality was close to MDPR. However, the zero-shot knowledge of the model was lost; thus, we opted to report only the MDPR in the main manuscript. + +Broader Modality Generalization. To further demonstrate the effectiveness of ModPrompt, we extended our evaluation to include Event-based (PEDRo) [3] and LiDAR (STCrowd) [8] modalities. As shown in Table 5, ModPrompt achieves strong performance, outperforming HFT and the visual prompt baseline. These results demonstrate that our method generalizes effectively across four distinct modalities (IR, Depth, Event-based, and LiDAR), confirming its robustness beyond the original experimental scope. + +
MethodEvent-based (PEDRo)LiDAR (STCrowd)
AP50AP75APAP50AP75AP
ZS23.908.911.400.000.000.0
HFT62.127.731.456.821.827.0
FFT94.178.766.988.561.255.7
WM49.921.324.734.810.014.6
ModPrompt92.974.263.885.754.851.2
+ +Table 5. Results on PEDRo (event-based) and STCrowd (Li-DAR) datasets on YOLO-World (YW). + +Study on Computational Cost. Our MPDR module is explicitly designed for inference-friendliness. Since text embeddings are precomputed, no forward pass through the text encoder is required either during inference or training. As shown in Figure 6, the MPDR module reduces the train- + +![](images/7dac91325eb3757aa6da5b32fc61125bdb864611d096d0421d5f2600edb3dbf4.jpg) +Figure 6. Comparison of Training Time and Inference Speed vs. Detection Performance on $\mathbf{NYU}_{v2}$ . Left: shows training time per epoch (in seconds) vs. $\mathrm{AP}_{50}$ . Right: presents inference time (FPS) vs. $\mathrm{AP}_{50}$ . Each point is a method, with HFT, FFT, WM, Modality Prompt (MP), and MP with Decoupled Residuals (MPDR) variants shown in distinct colors. Higher values in both FPS and $\mathrm{AP}_{50}$ indicate better runtime and accuracy trade-offs. + +ing time and increases the inference FPS. Although MPDR and MPDR WM require less computation, our MPDR ModPrompt achieves substantially higher detection performance while maintaining reasonable efficiency. + +# 4.4. Qualitative Results + +Figure 5 shows the visualization of ModPrompt on YOLO-World compared with zero-shot and the best visual prompt baseline. Thus, we can see that visual prompt-based methods, in general, struggle a bit to provide attached bounding boxes, but they do a good job of adapting. The VP method could not detect people in the LLVIP image well, while ModPrompt did a good job. Additional visualizations are provided in the supplementary material. + +# 5. Conclusion + +In this work, we presented ModPrompt, a visual prompt strategy that adapts open-vocabulary object detectors to new modalities (e.g., infrared, depth, event-based, and LiDAR) while preserving zero-shot capabilities. Our encoder-decoder design yields strong performance across challenging benchmarks. We also introduced MPDR, a decoupled text-prompt mechanism that retains language knowledge during adaptation and can match full fine-tuning performance. ModPrompt outperforms prior pixel-level prompts and generalizes across both CNN and transformer backbones, offering a robust and efficient solution for modality-adaptive object detection. By expanding the potential of vision-language models in different applications, ModPrompt offers a compelling solution for future advancements in adaptable object detection frameworks, especially in settings that require full knowledge preservation. + +Acknowledgments. This work was supported by Distech Controls Inc., the Natural Sciences and Engineering Research Council of Canada, the Digital Research Alliance of Canada, and MITACS. + +# References + +[1] Hyojin Bahng, Ali Jahanian, Swami Sankaranarayanan, and Phillip Isola. Exploring visual prompts for adapting large-scale models. arXiv preprint arXiv:2203.17274, 2022. 2, 3 +[2] Ankan Bansal, Karan Sikka, Gaurav Sharma, Rama Chellappa, and Ajay Divakaran. Zero-shot object detection. In Proceedings of the European conference on computer vision (ECCV), pages 384-400, 2018. 2 +[3] Chiara Boretti, Philippe Bich, Fabio Pareschi, Luciano Prono, Riccardo Rovatti, and Gianluca Setti. Pedro: an event-based dataset for person detection in robotics. In IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), 2023. 8 +[4] Nicolas Bustos, Mehrsa Mashhadi, Susana K Lai-Yuen, Sudeep Sarkar, and Tapas K Das. A systematic literature review on object detection using near infrared and thermal images. Neurocomputing, page 126804, 2023. 2 +[5] Yue Cao, Junchi Bin, Jozsef Hamari, Erik Blasch, and Zheng Liu. Multimodal object detection by channel switching and spatial attention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 403-411, 2023. 5 +[6] Kai Chen, Jiaqi Wang, Jiangmiao Pang, Yuhang Cao, Yu Xiong, Xiaoxiao Li, Shuyang Sun, Wansen Feng, Ziwei Liu, Jiarui Xu, Zheng Zhang, Dazhi Cheng, Chenchen Zhu, Tianheng Cheng, Qijie Zhao, Buyu Li, Xin Lu, Rui Zhu, Yue Wu, Jifeng Dai, Jingdong Wang, Jianping Shi, Wanli Ouyang, Chen Change Loy, and Dahua Lin. MMDetection: Open mmlab detection toolbox and benchmark. arXiv preprint arXiv:1906.07155, 2019. 5 +[7] Tianheng Cheng, Lin Song, Yixiao Ge, Wenyu Liu, Xinggang Wang, and Ying Shan. Yolo-world: Real-time open-vocabulary object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16901-16911, 2024. 2, 3, 5 +[8] Peishan Cong, Xinge Zhu, Feng Qiao, Yiming Ren, Xidong Peng, Yuenan Hou, Lan Xu, Ruigang Yang, Dinesh Manocha, and Yuexin Ma. Stcrowd: A multimodal dataset for pedestrian perception in crowded scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19608-19617, 2022. 8 +[9] Jacob Devlin. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 6 +[10] Thomas Dubail, Fidel Alejandro Guerrero Peña, Heitor Rapela Medeiros, Masih Aminbeidokhti, Eric Granger, and Marco Pedersoli. Privacy-preserving person detection using low-resolution infrared cameras. In European Conference on Computer Vision, pages 689-702. Springer, 2022. 2 +[11] Andreas Eitel, Jost Tobias Springenberg, Luciano Spinello, Martin Riedmiller, and Wolfram Burgard. Multimodal deep learning for robust rgb-d object recognition. In 2015 + +IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 681-687. IEEE, 2015. 2 +[12] Peng Gao, Shijie Geng, Renrui Zhang, Teli Ma, Rongyao Fang, Yongfeng Zhang, Hongsheng Li, and Yu Qiao. Clip-adapter: Better vision-language models with feature adapters. International Journal of Computer Vision, 132(2): 581-595, 2024. 3 +[13] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 3 +[14] Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, and Yin Cui. Open-vocabulary object detection via vision and language knowledge distillation. In International Conference on Learning Representations, 2022. 3 +[15] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 8 +[16] Christian Herrmann, Miriam Ruf, and Jürgen Beyerer. Cnn-based thermal infrared person detection by domain adaptation. In Autonomous Systems: Sensors, Vehicles, Security, and the Internet of Everything, page 1064308. International Society for Optics and Photonics, 2018. 2 +[17] Andrew G Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco Andretto, and Hartwig Adam. Mobilenets: Efficient convolutional neural networks for mobile vision applications. arXiv preprint arXiv:1704.04861, 2017. 8 +[18] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In Computer Vision and Pattern Recognition (CVPR), 2017 IEEE Conference on, 2017. 3 +[19] Eugenio Ivorra, Mario Ortega, Mariano Alcaniz, and Nicolas Garcia-Aracil. Multimodal computer vision framework for human assistive robotics. In 2018 Workshop on Metrology for Industry 4.0 and IoT, pages 1-5. IEEE, 2018. 2 +[20] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In European Conference on Computer Vision, pages 709-727. Springer, 2022. 2, 3 +[21] Glenn Jocher, Ayush Chaurasia, and Jing Qiu. Ultralytics yolov8. https://github.com/ultralytics/ultralytics, 2023.2 +[22] Aishwarya Kamath, Mannat Singh, Yann LeCun, Gabriel Synnaeve, Ishan Misra, and Nicolas Carion. Mdetr-modulated detection for end-to-end multi-modal understanding. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1780-1790, 2021. 6 +[23] Diederik P Kingma and Max Welling. Auto-encoding variational bayes, 2022. 3 +[24] Brian Lester, Rami Al-Rfou, and Noah Constant. The power of scale for parameter-efficient prompt tuning. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 3045-3059, 2021. 2 +[25] Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu + +Yuan, Lei Zhang, Jenq-Neng Hwang, et al. Grounded language-image pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10965-10975, 2022. 3 +[26] Shasha Li, Yongjun Li, Yao Li, Mengjun Li, and Xiaorong Xu. Yolo-firi: Improved yolov5 for infrared image object detection. IEEE access, 9:141861-141875, 2021. 2 +[27] Yinhao Li, Zheng Ge, Guanyi Yu, Jinrong Yang, Zengran Wang, Yukang Shi, Jianjian Sun, and Zeming Li. Bevdepth: Acquisition of reliable depth for multi-view 3d object detection. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1477-1485, 2023. 2 +[28] Chuang Lin, Peize Sun, Yi Jiang, Ping Luo, Lizhen Qu, Gholamreza Haffari, Zehuan Yuan, and Jianfei Cai. Learning object-language alignments for open-vocabulary object detection. In The Eleventh International Conference on Learning Representations, 2023. 3 +[29] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 2, 6 +[30] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Qing Jiang, Chunyuan Li, Jianwei Yang, Hang Su, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. In European conference on computer vision, pages 38-55. Springer, 2024. 2, 3 +[31] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 6 +[32] Heitor Rapela Medeiros, Masih Aminbeidokhti, Fidel Alejandro Guerrero Peña, David Latortue, Eric Granger, and Marco Pedersoli. Modality translation for object detection adaptation without forgetting prior knowledge. In European Conference on Computer Vision, pages 51-68. Springer, 2024. 3, 7 +[33] Heitor Rapela Medeiros, Fidel A Guerrero Pena, Masih Aminbeidokhti, Thomas Dubail, Eric Granger, and Marco Pedersoli. Hallucidet: Hallucinating rgb modality for person detection through privileged information. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1444-1453, 2024. 3, 7 +[34] Heitor R Medeiros, David Latortue, Eric Granger, and Marco Pedersoli. Mixed patch visible-infrared modality agnostic object detection. In 2025 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 9023-9032. IEEE Computer Society, 2025. 2 +[35] Claudio Michaelis, Benjamin Mitzkus, Robert Geirhos, Evgenia Rusak, Oliver Bringmann, Alexander S Ecker, Matthias Bethge, and Wieland Brendel. Benchmarking robustness in object detection: Autonomous driving when winter is coming. arXiv preprint arXiv:1907.07484, 2019. 2 +[36] Mehmet Akif Özkanoglu and Sedat Ozer. Infragan: A gan architecture to transfer visible images to infrared domain. Pattern Recognition Letters, 155:69-76, 2022. 2 + +[37] Yingxue Pang, Jianxin Lin, Tao Qin, and Zhibo Chen. Image-to-image translation: Methods and applications. IEEE Transactions on Multimedia, 24:3859-3881, 2021. 3 +[38] Harry A Pierson and Michael S Gashler. Deep learning in robotics: a review of recent research. Advanced Robotics, 31 (16):821-835, 2017. 2 +[39] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 6 +[40] Anitha Ramachandran and Arun Kumar Sangaiah. A review on object detection in unmanned aerial vehicle surveillance. International Journal of Cognitive Computing in Engineering, 2:215-228, 2021. 2 +[41] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical image computing and computer-assisted intervention, pages 234-241. Springer, 2015. 4 +[42] Shuai Shao, Zeming Li, Tianyuan Zhang, Chao Peng, Gang Yu, Xiangyu Zhang, Jing Li, and Jian Sun. Objects365: A large-scale, high-quality dataset for object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2019. 6 +[43] Nathan Silberman, Derek Hoiem, Pushmeet Kohli, and Rob Fergus. Indoor segmentation and support inference from rgbd images. In Computer Vision-ECCV 2012: 12th European Conference on Computer Vision, Florence, Italy, October 7-13, 2012, Proceedings, Part V 12, pages 746-760. Springer, 2012. 5 +[44] Jack Stilgoe. Machine learning, social learning and the governance of self-driving cars. Social studies of science, 48(1): 25-56, 2018. 2 +[45] Qingwang Wang, Yongke Chi, Tao Shen, Jian Song, Zifeng Zhang, and Yan Zhu. Improving rgb-infrared object detection by reducing cross-modality redundancy. Remote Sensing, 14(9):2020, 2022. 2 +[46] Xiangyang Xu, Yuncheng Li, Gangshan Wu, and Jiebo Luo. Multi-modal deep feature learning for rgb-d object detection. Pattern Recognition, 72:300-313, 2017. 2 +[47] Zeyu Yang, Jiaqi Chen, Zhenwei Miao, Wei Li, Xiatian Zhu, and Li Zhang. Deepinteraction: 3d object detection via modality interaction. Advances in Neural Information Processing Systems, 35:1992-2005, 2022. 2 +[48] Tao Yu, Zhihe Lu, Xin Jin, Zhibo Chen, and Xinchao Wang. Task residual for tuning vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10899-10909, 2023. 3, 5 +[49] Yuhang Zang, Wei Li, Kaiyang Zhou, Chen Huang, and Chen Change Loy. Open-vocabulary detr with conditional matching. In European Conference on Computer Vision, pages 106-122. Springer, 2022. 2 +[50] Heng Zhang, Elisa Fromont, Sébastien Lefèvre, and Bruno Avignon. Multispectral fusion for object detection with cyclic fuse-and-refine blocks. In 2020 IEEE International + +Conference on Image Processing (ICIP), pages 276-280. IEEE, 2020.5 +[51] Haotian Zhang, Pengchuan Zhang, Xiaowei Hu, Yen-Chun Chen, Liunian Li, Xiyang Dai, Lijuan Wang, Lu Yuan, Jenq-Neng Hwang, and Jianfeng Gao. Glipv2: Unifying localization and vision-language understanding. Advances in Neural Information Processing Systems, 35:36067-36080, 2022. 2, 3 +[52] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel Ni, and Heung-Yeung Shum. Dino: Detr with improved denoising anchor boxes for end-to-end object detection. In The Eleventh International Conference on Learning Representations, 2023. 2 +[53] Yiwu Zhong, Jianwei Yang, Pengchuan Zhang, Chunyuan Li, Noel Codella, Liunian Harold Li, Luowei Zhou, Xiyang Dai, Lu Yuan, Yin Li, et al. Regionclip: Region-based language-image pretraining. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16793-16803, 2022. 3 +[54] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Conditional prompt learning for vision-language models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16816-16825, 2022. 2, 5 +[55] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, 130(9):2337-2348, 2022. 2, 3, 5 +[56] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Computer Vision (ICCV), 2017 IEEE International Conference on, 2017. 3 +[57] Zhengxia Zou, Keyan Chen, Zhenwei Shi, Yuhong Guo, and Jieping Ye. Object detection in 20 years: A survey. Proceedings of the IEEE, 111(3):257-276, 2023. 1 \ No newline at end of file diff --git a/ICCV/2025/Visual Modality Prompt for Adapting Vision-Language Object Detectors/images.zip b/ICCV/2025/Visual Modality Prompt for Adapting Vision-Language Object Detectors/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..978dfcced8cd21d9731dcf34b9fd280886aa7f2a --- /dev/null +++ b/ICCV/2025/Visual Modality Prompt for Adapting Vision-Language Object Detectors/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c62ac000a7368d78ac93d4d9477cfaa5cd50e66c22b1ce35baa5587237ada3d5 +size 743851 diff --git a/ICCV/2025/Visual Modality Prompt for Adapting Vision-Language Object Detectors/layout.json b/ICCV/2025/Visual Modality Prompt for Adapting Vision-Language Object Detectors/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..2af86724973e069d3e4c238dbfc84c12faaae1a8 --- /dev/null +++ b/ICCV/2025/Visual Modality Prompt for Adapting Vision-Language Object Detectors/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e9cf80caca9b0e3fbda0f635b915a0761ba5d077364356454728b033a670e297 +size 324300 diff --git a/ICCV/2025/Visual Relation Diffusion for Human-Object Interaction Detection/dc9e94e3-6027-418f-9af5-8cec6e4a6997_content_list.json b/ICCV/2025/Visual Relation Diffusion for Human-Object Interaction Detection/dc9e94e3-6027-418f-9af5-8cec6e4a6997_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..114a1a8382e14fab84d1b1aebba1d743a00e6874 --- /dev/null +++ b/ICCV/2025/Visual Relation Diffusion for Human-Object Interaction Detection/dc9e94e3-6027-418f-9af5-8cec6e4a6997_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7958b258b7ae96791f3ffebe1338308df87a6e2a0667c852d747c36b8c68a5b2 +size 81117 diff --git a/ICCV/2025/Visual Relation Diffusion for Human-Object Interaction Detection/dc9e94e3-6027-418f-9af5-8cec6e4a6997_model.json b/ICCV/2025/Visual Relation Diffusion for Human-Object Interaction Detection/dc9e94e3-6027-418f-9af5-8cec6e4a6997_model.json new file mode 100644 index 0000000000000000000000000000000000000000..3cdddf554f274f3f8aec8def27f8fa04235a9335 --- /dev/null +++ b/ICCV/2025/Visual Relation Diffusion for Human-Object Interaction Detection/dc9e94e3-6027-418f-9af5-8cec6e4a6997_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5ccb3bdecf079198d24646f8c83bf83d7c33c0a372ade59be04e8fe4384a852c +size 99640 diff --git a/ICCV/2025/Visual Relation Diffusion for Human-Object Interaction Detection/dc9e94e3-6027-418f-9af5-8cec6e4a6997_origin.pdf b/ICCV/2025/Visual Relation Diffusion for Human-Object Interaction Detection/dc9e94e3-6027-418f-9af5-8cec6e4a6997_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..c5814cd0f5d77a5fb256599241585ff58a61921d --- /dev/null +++ b/ICCV/2025/Visual Relation Diffusion for Human-Object Interaction Detection/dc9e94e3-6027-418f-9af5-8cec6e4a6997_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f315487010873c97f8e265c679a65dd1e0febbb8887d8fbf98ae0799dfde8cfc +size 10590040 diff --git a/ICCV/2025/Visual Relation Diffusion for Human-Object Interaction Detection/full.md b/ICCV/2025/Visual Relation Diffusion for Human-Object Interaction Detection/full.md new file mode 100644 index 0000000000000000000000000000000000000000..774adda929bc05707360e15e7dff1887f5438d13 --- /dev/null +++ b/ICCV/2025/Visual Relation Diffusion for Human-Object Interaction Detection/full.md @@ -0,0 +1,306 @@ +# Visual Relation Diffusion for Human-Object Interaction Detection + +Ping Cao $^{1,2}$ , Yepeng Tang $^{1,2}$ , Chunjie Zhang $^{1,2,*}$ , Xiaolong Zheng $^{3,4,5}$ , Chao Liang $^{6}$ , Yunchao Wei $^{1,2}$ , Yao Zhao $^{1,2}$ + +$^{1}$ Institute of Information Science, School of Computer Science and Technology, BJTU + +$^{2}$ Visual Intelligence + X International Cooperation Joint Laboratory of MOE, BJTU + +$^{3}$ State Key Laboratory of Multimodal Artificial Intelligence Systems, CASIA + +$^{4}$ State Key Laboratory of Management and Control for Complex Systems, CASIA + +$^{5}$ School of Artificial Intelligence, UCAS $^{6}$ NERCMS, School of Computer Science, Wuhan University + +{pingcao, yepengtang, cjzhang, yunchao.wei, yzhao}@bjtu.edu.cn, + +xiaolong.zheng@ia.ac.cn, cliang@whu.edu.cn + +# Abstract + +Human-object interaction (HOI) detection relies on fine-grained visual understanding to distinguish complex relationships between humans and objects. While recent generative diffusion models have demonstrated remarkable capability in learning detailed visual concepts through pixel-level generation, their potential for interaction-level relationship modeling remains largely unexplored. To bridge this gap, we propose a Visual Relation Diffusion model (VRDiff), which introduces dense visual relation conditions to guide interaction understanding. Specifically, we encode interaction-aware condition representations that capture both spatial responsiveness and contextual semantics of human-object pairs, conditioning the diffusion process purely on visual features rather than text-based inputs. Furthermore, we refine these relation representations through generative feedback from the diffusion model, enhancing HOI detection without requiring image synthesis. Extensive experiments on the HICO-DET benchmark demonstrate that VRDiff achieves competitive results under both standard and zero-shot HOI detection settings. + +# 1. Introduction + +Human-object interaction (HOI) detection [7, 13, 14, 41, 51] aims to identify interacting human-object pairs within an image, as well as classify their corresponding interaction categories. As a fundamental task in scene understanding [9, 47-50, 53], HOI detection plays a crucial role in various real-world applications, including robotics intelligence, image captioning, and action recognition [18, 19, 24, + +![](images/94cdb46f2043b0678226665a6173e0d30768c0b0921e5d4ab382110a466532a3.jpg) +Figure 1. Comparison of HOI detection paradigms. (a) Traditional HOI detection methods. (b) VLM-based HOI detection leveraging multimodal knowledge. (c) Our approach, which transfers fine-grained perception capabilities from diffusion models to HOI detection through visual conditions. + +43, 45]. Traditional HOI detection methods, as illustrated in Fig. 1 (a) primarily focus on object detection and interaction classification, with optimizations centered around network design or feature enhancement. However, accurately detecting the $<$ human, verb, object $>$ triplet from images remains highly challenging due to the complexity of real-world scenes and the diversity of interactions, especially when predicting unseen HOI categories. + +A set of studies [27-29, 35] has attempted to enhance the generalization capability of HOI detectors by incorporating large-scale vision-language models (VLMs), such as CLIP. As shown in Fig. 1(b), unlike traditional HOI detection methods, VLM-based HOI approaches obtain extra linguistic knowledge by encoding HOI descriptions through the text encoder. While CLIP excels in learning high-level semantics, it exhibits limitations in perceiving fine-grained visual details due to its contrastive learning paradigm and the noisy image-text pairs used during training. Consequently, VLM-based HOI detection models tend to focus on salient human-object interactions, significantly hinder + +ing their capacity to discern spatial relationships and fine-grained interaction details. + +Recently, text-to-image diffusion models have shown remarkable potential in capturing complex visual concepts [1, 12, 32, 37], which is attributed to their well-designed architectures and large-scale paired image-text training data. These models generate high-quality images with rich details, demonstrating their strong capability to perceive fine-grained visual information, which is also crucial for HOI detection. An intuitive way to incorporate diffusion models into HOI detection is to expand the interaction dataset through text-to-image generation paradigm. However, generating additional HOI images via diffusion introduces substantial computational overhead, making it inefficient for applications. Moreover, while text-based conditioning in conventional diffusion models provides strong semantic guidance, directly leveraging it in HOI detection presents two key challenges. First, HOI detection primarily relies on visual relationship representation, and the presence of strong textual priors may unintentionally weaken the model's ability to refine visual features. Second, the instance-level design of traditional diffusion models, which emphasizes pixel-level image synthesis, limits their ability to capture interaction-level relationships that depend on contextual and relational understanding. Bridging this gap and transferring the fine-grained perception capabilities of diffusion models to human-object interaction modeling remains a significant and underexplored challenge. + +In this paper, we propose VRDiff, a visual-based diffusion framework that fully exploits the potential of diffusion models for HOI detection. To address the limitations of conventional diffusion models in relation modeling, we introduce dense visual conditions. Unlike previous works that transfer knowledge through text-based conditions, our framework adopts a purely visual strategy, eliminating the need for synthetic image generation, as shown in Fig. 1 (c). Our visual conditioning approach enables the diffusion model to provide fine-grained feedback directly on visual features, offering more precise guidance for HOI detection. To bridge the gap between instance-level pixel generation and interaction-level relation modeling, we incorporate spatial information and interaction context to enhance the relation modeling capability of condition features. Specifically, we propose an interaction-aware condition representation that captures spatial cues from human-object bounding boxes and integrates contextual information across interaction pairs, resulting in relational conditions enriched with spatial sensitivity and semantic cues. By using these representations as conditions for the diffusion model and applying a reconstruction loss to dynamically refine them, our VRDiff effectively enhances HOI detection. Experimental results demonstrate that our method achieves competitive performance across various zero-shot and standard + +settings. In summary, our main contributions are as follows: + +- We propose a visual condition diffusion feedback mechanism that transfers fine-grained perception capabilities from diffusion models to HOI detection. +- We introduce an interaction-aware condition that captures spatial and contextual interaction cues, guiding the diffusion model to focus on interaction understanding. +- Our VRDiff framework achieves competitive results in both standard and zero-shot settings, demonstrating strong generalization for HOI detection. + +# 2. Related Work + +# 2.1. Human-Object Interaction Detection + +As human-object interaction detection involves both detection and classification tasks, early HOI detection methods typically employ offline instance detectors to identify all humans and objects in an image, then associate them into pairs and classify interactions [8, 13, 14, 24, 51]. Two-stage methods relying on Convolutional Neural Network (CNN)-based detectors often suffer from suboptimal performance due to their dependence on instance detection quality. With the advent of Transformer-based detectors [5], two-stage human-object interaction detection methods have achieved significant advancements in recent years [4, 30, 34, 39, 52]. Meanwhile, one-stage methods formulate HOI detection as an end-to-end set prediction problem [7, 33, 46, 56], jointly predicting human-object bounding boxes and interaction categories. Many one-stage approaches focus on designing interaction-aware decoders [16, 18, 35, 42], enabling the query-based object detection paradigm to simultaneously output detection results and capture interaction information. In this work, we adopt a two-stage HOI detection approach, focusing on the representation of interaction relationships. + +Moreover, with the advancement of large-scale vision-language models (VLMs) [22, 31], an increasing number of HOI detection studies have incorporated pre-trained models into their frameworks. Some approaches focus on visual knowledge extraction [29, 36], leveraging VLM-based vision encoder to enhance visual features. Another strategy integrates language priors into HOI detection by integrating text encoder [35]. Additionally, several approaches utilize both vision and text encoders to obtain more discriminative interaction cues [19, 26-28, 40]. VLM-based methods have significantly enhanced the zero-shot capability of HOI detectors by distilling vision-text alignment or directly utilizing VLM-extracted features. However, as noted in recent works [17, 38], CLIP-based methods struggle to perceive fine-grained visual concepts beyond high-level semantics. To address this limitation, we incorporate diffusion models into HOI detection, strengthening the model's ability to capture fine-grained interaction perception. + +![](images/48d98028c1cadd8b054b9d5f911a9c4cdcef89053bab0075ebc3bb3fce095d58.jpg) +Figure 2. The overall pipeline of VRDiff. We propose a diffusion-based framework for HOI detection that relies solely on visual conditions. We construct visual relation conditions with spatial priors and contextual interaction cues to guide the diffusion model. Through reconstruction loss, diffusion feedback dynamically refines visual relation representations. During inference, we remove the diffusion branch and directly classify interactions using the refined visual relation features. + +# 2.2. Diffusion Models for Visual Representation + +In recent years, text-to-image diffusion models [15, 54] have achieved significant advancements in image generation, leveraging an iterative process of forward noise addition and reverse denoising to produce high-fidelity images with rich fine-grained details. These advancements stem from improvements in model architecture, training techniques, and the integration of large-scale datasets. Unlike traditional generative models, diffusion models refine noisy inputs step by step through a progressive denoising process, enabling them to capture intricate textures, structural details, and complex spatial relationships. Due to their superior ability to perceive multi-level visual concepts, diffusion models have been widely applied to downstream tasks such as image classification, image segmentation, and depth estimation [1, 2, 37]. Given the critical role of fine-grained visual perception in HOI detection, we integrate diffusion-based refinement into interaction modeling to enhance the representation of human-object relationships. + +In contrast to methods that rely on VLMs, recent diffusion-based approaches [23, 38, 44] largely follow the conventional text-to-image generation paradigm, either leveraging synthetic data generated by diffusion models to enhance generalization or extracting static visual features from diffusion encoders for knowledge transfer. Among these, CycleHOI [38] introduces a cycle consistency framework to guide HOI detector training through a detection-generation cycle. While effective, such methods rely on + +textual descriptions for the generation task, which may limit its ability to capture fine-grained interaction details, particularly when text conditions include annotation errors or omissions. Building upon this direction, our VRDiff introduces a purely visual conditioning mechanism that leverages dense visual relations to unlock the fine-grained perception capabilities of diffusion models. This enables effective HOI detection through direct visual relation refinement, without requiring image synthesis. + +# 3. VRDiff + +In this section, we present the Visual Relation Diffusion network for human-object interaction Detection (VRDiff), as illustrated in Fig. 2. + +# 3.1. Preliminary Detection and Encoding + +Instance Detector. In this work, we formulate HOI detection as a two-stage framework, where the first stage detects all instances, and the second stage performs interaction recognition for each human-object pair. For the instance detection stage, we employ an off-the-shelf object detector $D_{ins}$ , e.g. DETR [5], which is pre-trained on COCO to ensure robust performance in localizing instances. Given an image $I \in \mathbb{R}^{3 \times H \times W}$ with height $H$ and width $W$ , the instance detector outputs $(b_h, b_o, c_o)$ . Here, $b_h, b_o \in \mathbb{R}^4$ denote the detected bounding boxes of human and object instances, and $c_o \in \{1, \dots, O\}$ indicates the object category among $O$ total classes. These detection results provide prior + +information for the next stage of interaction recognition. + +Global Feature Encoding. In the second stage, we first extract global visual features from the input image $I$ using a pre-trained image encoder $\mathcal{F}_{enc}$ . Specifically, we leverage the CLIP image encoder to obtain a high-dimensional feature map denoted as $f_{I} \in \mathbb{R}^{C \times H' \times W'}$ , where $H'$ and $W'$ represent the height and width of the feature map, and $C$ denotes the number of channels. In order to optimize the utilization of knowledge acquired during pre-training, we follow [20] by integrating a series of learnable semantic adapter layers into each block of the visual encoder, while keeping the remaining parameters frozen. Subsequently, we apply ROAlign [11] to extract human-specific and object-specific features from the global feature map, using the bounding boxes of instances from the first stage: + +$$ +f _ {h} = \operatorname {R O I A l i g} \left(\mathcal {F} _ {\text {e n c}} (I), b _ {h}\right) \tag {1} +$$ + +$$ +f _ {o} = \operatorname {R O I A l i g n} \left(\mathcal {F} _ {\text {e n c}} (I), b _ {o}\right) \tag {2} +$$ + +where $f_{h}$ and $f_{o}$ represent the features extracted from the human and object regions, respectively. + +# 3.2. Interaction-aware Condition Representation + +Spatial Interaction Embedding. Based on the detection results of instance detector, we generate candidate human-object pairs by filtering and permuting the detected instances, producing a set of potential interactions. Unlike methods that rely solely on visual features, our approach explicitly encodes the spatial interactions between human-object pairs by extracting geometric features from their bounding boxes. Specially, for each human-object pair $(b_h, b_o)$ , we compute a set of geometric features $f_s \in \mathbb{R}^k$ , including relative position, scale ratio, and intersection-over-union (IoU). These features are then fed into a multilayer perceptron (MLP)-based spatial adapter, which learns a mapping $\mathcal{F}_{adapter}: \mathbb{R}^k \to \mathbb{R}^d$ to generate the spatial-aware interaction embedding $f_s \in \mathbb{R}^d$ . This embedding captures the geometric context of each pair, providing a robust spatial prior for subsequent visual relation modeling and interaction refinement in the diffusion process. + +Visual Relation Conditioning. To enable precise interaction refinement in the diffusion process, we develop a visual relation conditioning module that integrates spatial and visual interaction-aware contextual cues into a robust condition for human-object interaction detection, as shown in Fig. 3. For each candidate human-object pair identified in Section 3.1, we start with the human feature $f_{h}$ and object feature $f_{o}$ , extracted via ROAlign from the global feature $\mathcal{F}_{\mathrm{enc}}(I)$ , together with the spatially-aware embedding $f_{s}$ from the spatial interaction module. To leverage the complementary strengths of appearance and geometry, we concatenate these features into a unified vector: + +$$ +f _ {\text {c a t}} = \left[ f _ {h}, f _ {o}, f _ {s} \right] \tag {3} +$$ + +![](images/989d8de702556e7392e11dac107081b428200899abe74d51846b10b89fcc2b4a.jpg) +Figure 3. Overview of visual relation conditioning. We construct the visual condition for the diffusion model by integrating spatial features from bounding boxes with interaction features while incorporating contextual information across interaction pairs. + +This concatenated feature is then processed by an MLP with two hidden layers and ReLU activations, which learns a non-linear mapping $\mathcal{F}_{\mathrm{MLP}}: \mathbb{R}^j \to \mathbb{R}^l$ to produce an initial fused interaction feature: + +$$ +f _ {\text {f u s e d}} = \mathrm {M L P} \left(f _ {\text {c a t}}\right) \tag {4} +$$ + +To further enhance the representation, we apply a multihead self-attention (MHSA) adapter across the fused features of all $m$ candidate pairs within the image, denoted as $\{f_{\mathrm{fused}}^1, f_{\mathrm{fused}}^2, \ldots, f_{\mathrm{fused}}^m\}$ . The MHSA mechanism computes attention weights based on pairwise interaction similarities, enabling the model to emphasize relevant contextual interactions and mitigate noise from irrelevant pairs. Formally, the refined interaction condition is obtained as: + +$$ +f _ {\mathrm {r e l}} = \operatorname {M H S A} \left(f _ {\text {f u s e d}} ^ {1}, f _ {\text {f u s e d}} ^ {2}, \dots , f _ {\text {f u s e d}} ^ {m}\right) \tag {5} +$$ + +where $f_{\mathrm{rel}}$ integrates local pair-wise semantics with cross-pair relational context, providing a comprehensive representation that enhances the perception of interaction details. This enriched representation serves as the visual condition, guiding the refinement of interaction features through generative feedback, as detailed in Section 3.3. + +# 3.3. Relation-Driven Diffusion Model + +Relation-Conditioned Denoising. Diffusion models have demonstrated remarkable capabilities in capturing fine-grained visual details in image generation tasks with its unique stepwise denoising mechanism. Beyond producing high-quality images, it effectively learns and models mid- and low-level visual concepts, such as shape, texture, and local structure. Building on this strength, we design a relation-driven diffusion model that leverages the denoising process to refine interaction features for HOI detection, relying solely on visual conditions. Formally, given a clean image $x_0$ , a typical conditional diffusion model follows a Markovian denoising process. In the forward process, Gaussian noise $\epsilon \sim \mathcal{N}(0,I)$ is progressively added + +across $T$ steps, producing a sequence of increasingly noisy features $x_{1},x_{2},\ldots ,x_{T}$ .. + +$$ +x _ {t} = \sqrt {\bar {\alpha} _ {t}} x _ {0} + \sqrt {1 - \bar {\alpha} _ {t}} \epsilon \tag {6} +$$ + +where $x_{t}$ represents the noisy state at timestep $t$ , $\alpha_{t}$ is a predefined noise schedule parameter that decreases with $t$ , and $\bar{\alpha}_{t} = \prod_{i=1}^{t} \alpha_{i}$ , causing $x_{t}$ to approach a standard Gaussian at $t = T$ . The reverse process iteratively reconstructs the clean image $x_{0}$ by refining $x_{t}$ into $x_{t-1}$ at each step $t$ : + +$$ +x _ {t - 1} = \frac {1}{\sqrt {\alpha_ {t}}} \left(x _ {t} - \frac {1 - \alpha_ {t}}{\sqrt {1 - \bar {\alpha} _ {t}}} \epsilon_ {\phi} (x _ {t}, t, c)\right) + \sigma_ {t} z \tag {7} +$$ + +where $\epsilon_{\phi}$ is a denoising neural network predicting the noise component conditioned on a signal $c$ in the forward diffusion process, and $\sigma_t$ controls noise variance. + +Unlike previous methods that condition diffusion models on text conditions for pixel-level generation, our approach directly optimizes interaction-level visual features. Specifically, in our proposed VRDiff, we adapt the denoising process for HOI detection by conditioning the denoising steps with the visual relation embeddings $f_{\mathrm{rel}}$ from Section 3.2. At each timestep $t$ , the denoising network $\epsilon_{\theta}$ predicts the noise component guided by $f_{\mathrm{rel}}$ : + +$$ +x _ {t - 1} = \frac {1}{\sqrt {\alpha_ {t}}} \left(x _ {t} - \frac {1 - \alpha_ {t}}{\sqrt {1 - \bar {\alpha} _ {t}}} \epsilon_ {\phi} \left(x _ {t}, t, f _ {\mathrm {r e l}}\right)\right) + \sigma_ {t} z \tag {8} +$$ + +By incorporating the visual relation embedding $f_{\mathrm{rel}}$ , which encodes spatial and visual contextual cues for $m$ pairs, the diffusion process is effectively aligned with interaction-specific details. This alignment not only enables more effective refinement of interaction representations but also enhances the model's capacity to capture fine-grained relationships, particularly in challenging scenarios. + +Diffusion-based Interaction Refinement. Unlike prior methods that rely on static interaction features from diffusion model [44], our VRDiff leverages diffusion model feedback to dynamically refine these features, ensuring a more context-aware and adaptive representation of human-object interactions. Considering the incompleteness and ambiguity in HOI annotations, we avoid using text-based HOI labels as conditions in the diffusion process. Instead, as mentioned in the above relation-driven denoising process, we condition the diffusion model on interaction-aware embeddings $f_{\mathrm{rel}}$ , which encode both spatial and visual relational cues. At each timestep $t$ , the denoising network $\epsilon_{\phi}$ integrates $f_{\mathrm{rel}}$ via cross-attention: + +$$ +\operatorname {C r o s s A t t n} \left(x _ {t}, f _ {\text {r e l}}\right) = \operatorname {s o f t m a x} \left(\frac {Q K ^ {\top}}{\sqrt {d _ {k}}}\right) V \tag {9} +$$ + +where $Q = W_{q}x_{t}$ , $K = W_{k}f_{\mathrm{rel}}$ , and $V = W_{v}f_{\mathrm{rel}}$ . In this way, our approach ensures that the refinement process + +is guided by structured visual interaction priors, rather than relying on high-level text descriptions that may not fully capture the fine-grained details of human-object relationships. Notably, $f_{\mathrm{rel}}$ evolves dynamically via the diffusion model's feedback, rather than remaining a fixed representation, enabling step-wise enhancement of interaction representations throughout the process. When the denoising network $\epsilon_{\phi}$ processes the noisy feature $x_{t}$ under the guidance of visual relation condition $c = f_{\mathrm{rel}}$ , it generates a refined feature representation $x_{t-1}$ , which ensures a more interaction-aware and discriminative embedding. The reconstruction loss for this refinement process is defined as: + +$$ +L _ {d i f f} = \mathbb {E} _ {t, x _ {0}, \epsilon , c} \left[ \left\| \epsilon - \epsilon_ {\phi} \left(\sqrt {\bar {\alpha} _ {t}} x _ {0} + \sqrt {1 - \bar {\alpha} _ {t}} \epsilon , t, c\right) \right\| ^ {2} \right] \tag {10} +$$ + +where $\epsilon$ is the noise added to the clean feature and $\epsilon_{\phi}\left(\sqrt{\bar{\alpha}_t} x_0 + \sqrt{1 - \bar{\alpha}_t}\epsilon ,t,c\right)$ is the noise predicted by the denoising network $\epsilon_{\phi}$ . + +After obtaining the diffusion-refined interaction features, we perform action classification on these features. The action prediction score is computed as follows: + +$$ +s _ {h, o} = \left(\alpha_ {r e l} f _ {r e l} + \alpha_ {u} f _ {u}\right) C _ {L} ^ {T} \tag {11} +$$ + +where $s_{h,o} \in \mathbb{R}^K$ represents the score vector across $K$ interaction classes for a given human-object pair, $f_{rel}, f_u \in \mathbb{R}^d$ denote the visual relation, and unified interaction features, respectively. The coefficients $\alpha_{rel}, \alpha_u$ balance the contributions of each feature, weighting their influence based on spatial and visual relevance. Following [20], $C_L$ serves as a concept-guided prior integrating contributions from distinct concept branches. The entire model is optimized using a focal loss $L_{\mathrm{cls}}$ to address class imbalance in HOI detection. Consequently, the overall loss for our VRDiff framework combines this classification loss with the diffusion denoising objective: + +$$ +L _ {\text {t o t a l}} = \lambda_ {d i f f} L _ {d i f f} + \lambda_ {c l s} L _ {c l s} \tag {12} +$$ + +# 4. Experiments + +# 4.1. Datasets + +We conduct experiments on widely used HICO-DET [6] dataset to verify the effectiveness of our proposed VRDiff model. HICO-DET contains 47,776 images, with 38,118 for training and 9,658 for testing, respectively. The whole dataset contains 80 object categories, 117 interaction categories, and $600 < \text{human}, \text{verb}, \text{object} >$ categories. Among the 600 HOI categories, 138 categories with less than 10 training samples are labeled as "Rare" categories, while the remaining categories are defined as "Non-Rare" categories. For zero-shot HOI detection, our method follows the standard setting from [19, 23], which requires predicting HOI categories that are unseen during training. + +
MethodsBackboneDefaultKnown Object
FullRareNon-RareFullRareNon-Rare
CDN [46] (NeurIPS21)ResNet-10132.0727.1933.5334.7929.4836.38
HOI-Trans [56] (CVPR21)ResNet-10126.6119.1528.8429.1320.9831.57
QPIC [33] (CVPR21)ResNet-10129.9023.9231.6932.3826.0634.27
AS-Net [7] (CVPR21)ResNet-5028.8724.2530.2531.7427.0733.14
SSRT [16] (CVPR22)ResNet-10131.3424.3133.32---
GEN-VLKT [26] (CVPR22)ResNet-10134.9631.1836.0838.2234.3639.37
MUREN [18] (CVPR23)ResNet-5032.8728.6734.1235.5230.8836.91
HOICLIP [29] (CVPR23)ResNet-5034.6931.1235.7437.6134.4738.54
AGER [35] (ICCV23)ResNet-5036.7533.5337.7139.8435.5840.2
Xie et al. [42] (CVPR23)ResNet-10136.0333.1636.8938.8235.5139.81
DIFFUSIONHOI [23] (NeurIPS24)VQGAN38.1238.9337.8440.9342.8740.04
DRG [8] (ECCV20)ResNet-50-FPN24.5319.4726.0427.9823.1129.43
IDN [24] (NeurIPS20)ResNet-5024.5820.3325.8627.8923.6429.16
FCL [14] (CVPR21)ResNet-5025.2720.5726.6727.7122.3428.93
ATL [13] (CVPR21)ResNet-5023.8117.4327.4227.3822.0928.96
SCG [51] (ICCV21)ResNet-50-FPN29.2624.6130.6532.8727.8934.35
STIP [55] (CVPR22)ResNet-5032.2228.1533.4335.2931.4336.45
UPT [52] (CVPR22)ResNet-10132.6228.6233.8136.0831.4137.47
Iwin [34] (ECCV22)ResNet-10132.7927.8435.4035.8428.7436.09
MCPC [41] (ECCV22)ResNet-5035.1533.7135.5837.5635.8738.06
ViPLO [30] (CVPR23)ResNet-50+ViT-B37.2235.4537.7540.6138.8241.15
ADA-CM [20] (ICCV23)ResNet-50+ViT-L38.4037.5238.66---
RmLR [4] (ICCV23)ResNet-10137.4128.8139.9738.6931.2740.91
PVic [53] (ICCV23)ResNet-5034.6932.1435.4538.1435.3838.97
BCOM [36] (CVPR24)ResNet-50+ViT-L39.3439.9039.1742.2442.8642.05
Wu et al. [39] (CVPR24)ResNet-10136.8233.9937.6640.5637.0241.69
CMMP [21] (ECCV24)ResNet-50+ViT-L38.1437.7538.25---
VRDiff (Ours)ResNet-50+ViT-B36.7735.6637.1139.9838.6940.36
VRDiff (Ours)ResNet-50+ViT-L41.4041.6941.3144.2544.6544.13
+ +Table 1. Comparison of HOI Detection Performance Under Standard Settings on HICO-DET. + +# 4.2. Implementation Details + +We follow the standard protocol of two-stage HOI detection methods [20, 21, 25], where the instance detector DETR [5] with a ResNet-50 [10] backbone is fine-tuned before training VRDiff. During training, we freeze the instance detector, image encoder (CLIP [31]) and diffusion model (SD-2-1-base [32]), and only train the adapter within the image encoder along with the remaining components of the VRDiff model. During the inference phase, we remove the SD branch and classify interactions directly using visual relation features. For all experiments, we use AdamW as the optimizer with a learning rate of $1e^{-3}$ . The model is trained for a maximum of 12 epochs, with loss weights $\lambda_{diff}$ and $\lambda_{cls}$ both set to 1. VRDiff is conducted on NVIDIA GeForce RTX 3090 GPUs with a batch size of 8. + +# 4.3. Comparisons with the State-of-the-Arts + +We compare VRDiff with existing methods on HOI detection under the standard setting. As shown in Tab. 1, VRDiff + +achieves the best performance in both default and known-object settings. Our diffusion feedback mechanism refines visual relation features, minimizing the performance gap between Rare and Non-Rare categories, even with a smaller image encoder (ViT-B). With a stronger encoder (ViT-L), VRDiff surpasses DIFFUSIONHOI [23] by 2.76 mAP on Rare categories and further reduces the Rare vs. Non-Rare gap to 0.38. Unlike methods relying on diffusion-generated synthetic data, VRDiff consistently improves performance without additional data augmentation, highlighting the effectiveness of visual relation diffusion. + +To assess the generalization of our method, we compare it with state-of-the-art approaches under zero-shot settings, where models are required to predict unseen HOI categories. Further details are provided in the supplementary materials. As shown in Tab. 2 and Tab. 3, VRDiff consistently outperforms existing approaches across all zero-shot scenarios. The methods above and below the dashed line are implemented with smaller and larger backbones, respec + +
MethodsTypeUnseenSeenFull
GEN-VLKT [26] (CVPR22)RF-UC21.3632.9130.56
EoID [40] (AAAI23)RF-UC22.0431.3929.52
HOICLIP [29] (CVPR23)RF-UC25.5328.4732.99
ADA-CM [20] (ICCV23)RF-UC27.6334.3533.01
CLIP4HOI [28] (NeurIPS23)RF-UC28.4735.4834.08
CMMP [21] (ECCV24)RF-UC29.4532.8732.18
DIFFUSIONHOI [23] (NeurIPS24)RF-UC32.0636.7735.89
EZ-HOI [19] (NeurIPS24)RF-UC29.0234.1533.13
VRDiff (Ours)RF-UC31.8036.8335.82
UniHOI [3] (NeurIPS23)RF-UC28.6833.1632.27
BCOM [36] (CVPR24)RF-UC28.5235.0433.74
CMMP [21] (ECCV24)RF-UC35.9837.4237.13
EZ-HOI [19] (NeurIPS24)RF-UC34.2437.3536.73
VRDiff (Ours)RF-UC38.9240.8340.45
GEN-VLKT [26] (CVPR22)NF-UC25.0523.3823.71
EoID [40] (AAAI23)NF-UC26.7726.6626.69
HOICLIP [29] (CVPR23)NF-UC26.3928.1027.75
ADA-CM [20] (ICCV23)NF-UC32.4131.1331.39
CLIP4HOI [28] (NeurIPS23)NF-UC28.9031.4428.26
CMMP [21] (ECCV24)NF-UC32.0929.7130.18
DIFFUSIONHOI [23] (NeurIPS24)NF-UC30.0430.2930.25
EZ-HOI [19] (NeurIPS24)NF-UC31.1733.6630.55
VRDiff (Ours)NF-UC36.4833.8434.37
UniHOI [3] (NeurIPS23)NF-UC28.4532.6331.79
BCOM [36] (CVPR24)NF-UC33.1231.7632.03
CMMP [21] (ECCV24)NF-UC33.5235.5335.13
EZ-HOI [19] (NeurIPS24)NF-UC36.3334.4734.84
VRDiff (Ours)NF-UC39.3339.0339.09
+ +tively. With a smaller backbone, VRDiff achieves the best performance on Seen categories, while with a larger backbone, it outperforms all existing methods across all categories. By leveraging visual relation encoding, VRDiff effectively transfers the fine-grained perception capabilities of diffusion models into interaction feature learning. + +# 4.4. Ablation Studies + +We conduct ablation studies on HICO-DET under UV Setting, exploring the effect of key components, condition design for VRDiff, interaction relation modeling strategy, and relation condition composition. + +Key Components. As shown in Tab. 4, each component of VRDiff contributes meaningfully to HOI detection performance. The baseline model without visual relation modules yields the lowest accuracy, highlighting the limitations of relying solely on detection-based features. Incorporating the Visual Relation Encoding (VRE) module enhances the discriminative capacity of interaction features by explicitly modeling human-object relationships. Building upon this, the Spatial Interaction Embedding (SIE) module further improves spatial reasoning capabilities, with notable gains in complex interaction understanding. Finally, integrating Stable Diffusion (SD) yields the full VRDiff model, which achieves the highest overall performance and demonstrates + +Table 2. Comparison of Rare-First Unseen Composition (RF-UC) and Non-Rare First Unseen Composition (NF-UC) on HICO-DET with State-of-the-Art Methods. + +
MethodsTypeUnseenSeenFull
FCL [14] (CVPR21)UO15.5420.7419.87
ATL [13] (CVPR21)UO15.1121.5420.47
GEN-VLKT [26] (CVPR22)UO10.5128.9225.63
HOICLIP [29] (CVPR23)UO16.2030.9928.53
CLIP4HOI [28] (NeurIPS23)UO31.7932.7332.58
CMMP [21] (ECCV24)UO33.7631.1531.59
DIFFUSIONHOI [23] (NeurIPS24)UO22.3732.0331.12
EZ-HOI [19] (NeurIPS24)UO33.2832.0632.27
VRDiff (Ours)UO37.2934.4934.96
UniHOI [3] (NeurIPS23)UO19.7234.7631.56
CMMP [21] (ECCV24)UO39.6736.1536.74
EZ-HOI [19] (NeurIPS24)UO38.1736.0236.38
VRDiff (Ours)UO41.5239.6639.97
GEN-VLKT [26] (CVPR22)UV19.4726.0424.53
EoID [40] (AAAI23)UV22.7130.7329.61
HOICLIP [29] (CVPR23)UV24.3032.1931.09
CLIP4HOI [28] (NeurIPS23)UV26.0231.1430.42
CMMP [21] (ECCV24)UV26.2332.7531.84
DIFFUSIONHOI [23] (NeurIPS24)UV28.0533.2432.67
EZ-HOI [19] (NeurIPS24)UV25.1033.4932.32
VRDiff (Ours)UV23.6236.8434.99
UniHOI [3] (NeurIPS23)UV26.0536.7834.68
CMMP [21] (ECCV24)UV30.8437.2836.38
EZ-HOI [19] (NeurIPS24)UV28.8238.1536.84
VRDiff (Ours)UV32.1440.6139.43
+ +Table 3. Comparison of Unseen Object (UO) and Unseen Verb (UV) on HICO-DET with State-of-the-Art Methods. + +particularly strong benefits for generalizing to unseen interactions. To further illustrate these quantitative gains, Fig. 4 presents attention maps extracted from the image encoder across different model variants, from the baseline to the full VRDiff. As the VRE, SIE, and SD modules are progressively integrated, the model's attention increasingly focuses on regions corresponding to human-object interactions. This indicates that VRDiff enhances fine-grained visual perception by conditioning the diffusion model with dense interaction-aware visual relations. + +Condition Design for VRDiff. Traditional stable diffusion relies on text conditions for image generation, and prior SD-based HOI detection methods have largely followed this approach. In contrast, our work focuses on conditioning the diffusion model solely with visual features, combining empty text and visual relations (EmptyText+Visual) as the final condition. To validate the effectiveness of our condition design, we compare it against two alternative settings: using real interaction annotations as text conditions (RealText) and combining real interaction text with visual relations (RealText+Visual) in Tab. 5. The results indicate that real interaction annotations constrain the diffusion model's ability to enhance visual representations of the HOI detection model, leading to a $1.24\mathrm{mAP}$ drop compared to our approach. Adding visual features mitigates + +
SDSIEVREUnseenSeenFull
×××22.2638.8336.51
××✓24.8139.6337.56
×✓✓25.9940.1338.15
✓✓✓32.1440.6139.43
+ +![](images/7df7a0af9242688199f5ed065eeade03062653506b239a2ef4ec99eefd63e559.jpg) +Figure 4. Visualization of attention maps across different model variants. From left to right: Base model, Base+VRE, Base+VRE+SIE, and VRDiff (full model). + +this impact but still underperforms our approach by 0.62 mAP. This demonstrates that VRDiff more effectively transfers the fine-grained perceptual capabilities of the diffusion model to the HOI detection model. + +Interaction Relation Modeling. As shown in Tab. 6, we explore different approaches to modeling interaction relations by extracting various types of interaction-aware features from image-level representations. Using only object features (O) without explicit interaction modeling leads to suboptimal performance, highlighting the importance of relational modeling. Incorporating both human and object features (HO) achieves the best performance, demonstrating that explicitly capturing pairwise interactions provides a more discriminative and informative representation. While extracting features from the union of human and object regions (Uni) also improves performance, it remains slightly inferior to HO, suggesting that a broader spatial region does not necessarily provide richer interaction cues compared to explicitly modeling human-object relationships. + +Relation Condition Composition. To evaluate the effectiveness of our visual relation condition, we present HOI detection results with different relation condition compositions in Tab. 7. The results show that using only spatial (S) relation features from bounding boxes leads to a decline in HOI detection performance due to the absence + +Table 4. Ablation study on the Key Components of VRDiff. + +
Condition DesignUnseenSeenFull
RealText32.0939.1838.19
RealText+Visual32.1339.9038.81
EmptyText+Visual32.1440.6139.43
+ +Table 5. Ablation study on the Condition Design of VRDiff. + +
Relation ModelingUnseenSeenFull
O31.5639.8938.72
HO32.1440.6139.43
Uni31.5739.9438.77
+ +Table 6. Ablation study on the Relation Modeling Strategy. + +
Condition CompositionUnseenSeenFull
S30.0839.3038.01
V29.6839.1337.80
S + V32.1440.6139.43
+ +Table 7. Ablation study on the Relation Condition Composition. + +of fine-grained dense visual guidance. Relying solely on interaction-level visual (V) features results in a $0.31\mathrm{mAP}$ drop compared to the S-based setting, highlighting the importance of spatial priors in refining interaction relationships. In contrast, our final relation condition $(S + V)$ , which integrates spatial enhancement within interactions and relation augmentation across interactions, achieves the best performance by providing a more comprehensive and structured relation representation. + +# 5. Conclusion + +In this paper, we propose the Visual Relation Diffusion model (VRDiff), a novel diffusion feedback mechanism for HOI detection. Unlike traditional text-conditioned approaches, VRDiff introduces dense visual relation conditions that enable the diffusion model to focus on fine-grained interaction details. Through interaction-aware condition representations and generative feedback from text-to-image diffusion models, our model effectively learns spatially responsive features and enhances relational modeling capabilities. Extensive experiments on HICO-DET demonstrate that VRDiff achieves competitive performance under both standard and zero-shot settings. + +# 6. Acknowledgements + +This work was supported by National Natural Science Foundation of China (62476021, 72225011, 72434005, 62372339, 92470203), Beijing Natural Science Foundation (No. L242022). + +# References + +[1] Omri Avrahami, Dani Lischinski, and Ohad Fried. Blended diffusion for text-driven editing of natural images. In CVPR, pages 18208-18218, 2022. 2, 3 +[2] Dmitry Baranchuk, Ivan Rubachev, Andrey Voynov, Valentin Khrulkov, and Artem Babenko. Label-efficient semantic segmentation with diffusion models. arXiv preprint arXiv:2112.03126, 2021. 3 +[3] Yichao Cao, Qingfei Tang, Xiu Su, Song Chen, Shan You, Xiaobo Lu, and Chang Xu. Detecting any human-object interaction relationship: Universal hoi detector with spatial prompt learning on foundation models. Advances in Neural Information Processing Systems, 36:739-751, 2023. 7 +[4] Yichao Cao, Qingfei Tang, Feng Yang, Xiu Su, Shan You, Xiaobo Lu, and Chang Xu. Re-min, learn and reason: Exploring the cross-modal semantic correlations for language-guided hoi detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 23492-23503, 2023. 2, 6 +[5] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European Conference on Computer Vision, pages 213-229. Springer, 2020. 2, 3, 6 +[6] Yu-Wei Chao, Yunfan Liu, Xieyang Liu, Huayi Zeng, and Jia Deng. Learning to detect human-object interactions. In 2018 IEEE winter conference on applications of computer vision (wacv), pages 381-389. IEEE, 2018. 5 +[7] Mingfei Chen, Yue Liao, Si Liu, Zhiyuan Chen, Fei Wang, and Chen Qian. Reformulating hoi detection as adaptive set prediction. In CVPR, pages 9004-9013, 2021. 1, 2, 6 +[8] Chen Gao, Jiarui Xu, Yuliang Zou, and Jia-Bin Huang. Drg: Dual relation graph for human-object interaction detection. In European Conference on Computer Vision, pages 696-712. Springer, 2020. 2, 6 +[9] Saurabh Gupta and Jitendra Malik. Visual semantic role labeling. arXiv preprint arXiv:1505.04474, 2015. 1 +[10] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6 +[11] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2961-2969, 2017. 4 +[12] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2 +[13] Zhi Hou, Baosheng Yu, Yu Qiao, Xiaojiang Peng, and Dacheng Tao. Affordance transfer learning for human-object interaction detection. In CVPR, pages 495-504, 2021. 1, 2, 6, 7 +[14] Zhi Hou, Baosheng Yu, Yu Qiao, Xiaojiang Peng, and Dacheng Tao. Detecting human-object interaction via fabricated compositional learning. In CVPR, pages 14646-14655, 2021. 1, 2, 6, 7 + +[15] Hexiang Hu, Kelvin CK Chan, Yu-Chuan Su, Wenhu Chen, Yandong Li, Kihyuk Sohn, Yang Zhao, Xue Ben, Boqing Gong, William Cohen, et al. Instruct-imagen: Image generation with multi-modal instruction. In CVPR, pages 4754-4763, 2024. 3 +[16] ASM Iftekhar, Hao Chen, Kaustav Kundu, Xinyu Li, Joseph Tighe, and Davide Modolo. What to look at and where: Semantic and spatial refined transformer for detecting human-object interactions. In CVPR, pages 5353-5363, 2022. 2, 6 +[17] Dahun Kim, Anelia Angelova, and Weicheng Kuo. Region-aware pretraining for open-vocabulary object detection with vision transformers. In CVPR, pages 11144-11154, 2023. 2 +[18] Sanghyun Kim, Deunsol Jung, and Minsu Cho. Relational context learning for human-object interaction detection. In CVPR, pages 2925-2934, 2023. 1, 2, 6 +[19] Qinqian Lei, Bo Wang, and Robby Tan. Ez-hoi: VIm adaptation via guided prompt learning for zero-shot hoi detection. Advances in Neural Information Processing Systems, 37:55831-55857, 2024. 1, 2, 5, 7 +[20] Ting Lei, Fabian Caba, Qingchao Chen, Hailin Jin, Yuxin Peng, and Yang Liu. Efficient adaptive human-object interaction detection with concept-guided memory. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6480-6490, 2023. 4, 5, 6, 7 +[21] Ting Lei, Shaofeng Yin, Yuxin Peng, and Yang Liu. Exploring conditional multi-modal prompts for zero-shot hoi detection. In European Conference on Computer Vision, pages 1-19. Springer, 2024. 6, 7 +[22] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, pages 12888-12900. PMLR, 2022. 2 +[23] Liulei Li, Wenguan Wang, and Yi Yang. Human-object interaction detection collaborated with large relation-driven diffusion models. arXiv preprint arXiv:2410.20155, 2024. 3, 5, 6, 7 +[24] Yong-Lu Li, Xinpeng Liu, Xiaogian Wu, Yizhuo Li, and Cewu Lu. Hoi analysis: Integrating and decomposing human-object interaction. Advances in Neural Information Processing Systems, 33:5011-5022, 2020. 1, 2, 6 +[25] Zhuolong Li, Xingao Li, Changxing Ding, and Xiangmin Xu. Disentangled pre-training for human-object interaction detection. In CVPR, pages 28191-28201, 2024. 6 +[26] Yue Liao, Aixi Zhang, Miao Lu, Yongliang Wang, Xiaobo Li, and Si Liu. Gen-vlkt: Simplify association and enhance interaction understanding for hoi detection. In CVPR, pages 20123-20132, 2022. 2, 6, 7 +[27] Jinguo Luo, Weihong Ren, Weibo Jiang, Xi'ai Chen, Qiang Wang, Zhi Han, and Honghai Liu. Discovering syntactic interaction clues for human-object interaction detection. In CVPR, pages 28212-28222, 2024. 1 +[28] Yunyao Mao, Jiajun Deng, Wengang Zhou, Li Li, Yao Fang, and Houqiang Li. Clip4hoi: Towards adapting clip for practical zero-shot hoi detection. Advances in Neural Information Processing Systems, 36:45895-45906, 2023. 2, 7 +[29] Shan Ning, Longtian Qiu, Yongfei Liu, and Xuming He. Hoiclip: Efficient knowledge transfer for hoi detection with + +vision-language models. In CVPR, pages 23507-23517, 2023. 1, 2, 6, 7 +[30] Jeeseung Park, Jin-Woo Park, and Jong-Seok Lee. Viplo: Vision transformer based pose-conditioned self-loop graph for human-object interaction detection. In CVPR, pages 17152-17162, 2023. 2, 6 +[31] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PmLR, 2021. 2, 6 +[32] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 2, 6 +[33] Masato Tamura, Hiroki Ohashi, and Tomoaki Yoshinaga. Qpic: Query-based pairwise human-object interaction detection with image-wide contextual information. In CVPR, pages 10410-10419, 2021. 2, 6 +[34] Danyang Tu, Xiongkuo Min, Huiyu Duan, Guodong Guo, Guangtao Zhai, and Wei Shen. Iwin: Human-object interaction detection via transformer with irregular windows. In European Conference on Computer Vision, pages 87-103. Springer, 2022. 2, 6 +[35] Danyang Tu, Wei Sun, Guangtao Zhai, and Wei Shen. Agglomerative transformer for human-object interaction detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 21614-21624, 2023. 1, 2, 6 +[36] Guangzhi Wang, Yangyang Guo, Ziwei Xu, and Mohan Kankanhalli. Bilateral adaptation for human-object interaction detection with occlusion-robustness. In CVPR, pages 27970–27980, 2024. 2, 6, 7 +[37] Wenxuan Wang, Quan Sun, Fan Zhang, Yepeng Tang, Jing Liu, and Xinlong Wang. Diffusion feedback helps clip see better. arXiv preprint arXiv:2407.20171, 2024. 2, 3 +[38] Yisen Wang, Yao Teng, and Limin Wang. Cyclehoi: Improving human-object interaction detection with cycle consistency of detection and generation. arXiv preprint arXiv:2407.11433, 2024. 2, 3 +[39] Eastman ZY Wu, Yali Li, Yuan Wang, and Shengjin Wang. Exploring pose-aware human-object interaction via hybrid learning. In CVPR, pages 17815-17825, 2024. 2, 6 +[40] Mingrui Wu, Jiaxin Gu, Yunhang Shen, Mingbao Lin, Chao Chen, and Xiaoshuai Sun. End-to-end zero-shot hoi detection via vision and language knowledge distillation. In Proceedings of the AAAI Conference on artificial intelligence, pages 2839-2846, 2023. 2, 7 +[41] Xiaoqian Wu, Yong-Lu Li, Xinpeng Liu, Junyi Zhang, Yuzhe Wu, and Cewu Lu. Mining cross-person cues for body-part interactiveness learning in hoi detection. In European Conference on Computer Vision, pages 121-136. Springer, 2022. 1, 6 +[42] Chi Xie, Fangao Zeng, Yue Hu, Shuang Liang, and Yichen Wei. Category query learning for human-object interaction classification. In CVPR, pages 15275-15284, 2023. 2, 6 + +[43] Xianghui Xie, Bharat Lal Bhatnagar, and Gerard Pons-Moll. Visibility aware human-object interaction tracking from single rgb camera. In CVPR, pages 4757-4768, 2023. 1 +[44] Jie Yang, Bingliang Li, Ailing Zeng, Lei Zhang, and Ruimao Zhang. Open-world human-object interaction detection via multi-modal prompts. In CVPR, pages 16954-16964, 2024. 3, 5 +[45] Ting Yao, Yingwei Pan, Yehao Li, and Tao Mei. Exploring visual relationship for image captioning. In European Conference on Computer Vision, pages 684-699, 2018. 1 +[46] Aixi Zhang, Yue Liao, Si Liu, Miao Lu, Yongliang Wang, Chen Gao, and Xiaobo Li. Mining the benefits of two-stage and one-stage hoi detection. Advances in Neural Information Processing Systems, 34:17209-17220, 2021. 2, 6 +[47] Chunjie Zhang, Huihui Bai, and Yao Zhao. Fine-grained image classification by class and image-specific decomposition with multiple views. IEEE Transactions on Multimedia, 25: 6756-6766, 2022. 1 +[48] Chunjie Zhang, Chao Liang, and Yao Zhao. Exemplar-based, semantic guided zero-shot visual recognition. IEEE Transactions on Image Processing, 31:3056-3065, 2022. +[49] Chunjie Zhang, Ping Cao, Huihui Bai, and Xiaolong Zheng. Semantic coherence guided multi-view similarity for image classification with varied supervision. IEEE Transactions on Artificial Intelligence, 5(6):3072-3082, 2023. +[50] Chunjie Zhang, Chuankai Liu, Shaohua Duan, Xiaolong Zheng, Tianyi Yu, and JITAO Zhang. Embodied cognitive intelligence guided moon sample collection. The Innovation, 6(7):100939, 2025. 1 +[51] Frederic Z Zhang, Dylan Campbell, and Stephen Gould. Spatially conditioned graphs for detecting human-object interactions. In Proceedings of the IEEE/CVF international conference on computer vision, pages 13319-13327, 2021. 1, 2, 6 +[52] Frederic Z Zhang, Dylan Campbell, and Stephen Gould. Efficient two-stage detection of human-object interactions with a novel unary-pairwise transformer. In CVPR, pages 20104-20112, 2022. 2, 6 +[53] Frederic Z Zhang, Yuhui Yuan, Dylan Campbell, Zhuoyao Zhong, and Stephen Gould. Exploring predicate visual context in detecting of human-object interactions. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10411-10421, 2023. 1, 6 +[54] Qinsheng Zhang and Yongxin Chen. Fast sampling of diffusion models with exponential integrator. arXiv preprint arXiv:2204.13902, 2022. 3 +[55] Yong Zhang, Yingwei Pan, Ting Yao, Rui Huang, Tao Mei, and Chang-Wen Chen. Exploring structure-aware transformer over interaction proposals for human-object interaction detection. In CVPR, pages 19548-19557, 2022. 6 +[56] Cheng Zou, Bohan Wang, Yue Hu, Junqi Liu, Qian Wu, Yu Zhao, Boxun Li, Chenguang Zhang, Chi Zhang, Yichen Wei, et al. End-to-end human object interaction detection with hoi transformer. In CVPR, pages 11825-11834, 2021. 2, 6 \ No newline at end of file diff --git a/ICCV/2025/Visual Relation Diffusion for Human-Object Interaction Detection/images.zip b/ICCV/2025/Visual Relation Diffusion for Human-Object Interaction Detection/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..4fade12f879121cbd79a84ca44174cf18678cea2 --- /dev/null +++ b/ICCV/2025/Visual Relation Diffusion for Human-Object Interaction Detection/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f67137394c4297e9cfccf9ae791344bbd21e2a10eed07b3e34830b30a2a08ea7 +size 759819 diff --git a/ICCV/2025/Visual Relation Diffusion for Human-Object Interaction Detection/layout.json b/ICCV/2025/Visual Relation Diffusion for Human-Object Interaction Detection/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..a96c1e577eef5433e394b70076e286e579271a66 --- /dev/null +++ b/ICCV/2025/Visual Relation Diffusion for Human-Object Interaction Detection/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9e16e302a9c44df8a76807a45eaa822ebe977ba9b93652e1b12a60db403c435e +size 376063 diff --git a/ICCV/2025/Visual Surface Wave Elastography_ Revealing Subsurface Physical Properties via Visible Surface Waves/f44e2841-6bf1-41ce-9953-fc74116e88aa_content_list.json b/ICCV/2025/Visual Surface Wave Elastography_ Revealing Subsurface Physical Properties via Visible Surface Waves/f44e2841-6bf1-41ce-9953-fc74116e88aa_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..69603897c8c0a257d7da99700eeb2f770cb71659 --- /dev/null +++ b/ICCV/2025/Visual Surface Wave Elastography_ Revealing Subsurface Physical Properties via Visible Surface Waves/f44e2841-6bf1-41ce-9953-fc74116e88aa_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9f6993717fe91687d70e4df418b031a8e273b71c43b34d7944dfc116621e823c +size 74329 diff --git a/ICCV/2025/Visual Surface Wave Elastography_ Revealing Subsurface Physical Properties via Visible Surface Waves/f44e2841-6bf1-41ce-9953-fc74116e88aa_model.json b/ICCV/2025/Visual Surface Wave Elastography_ Revealing Subsurface Physical Properties via Visible Surface Waves/f44e2841-6bf1-41ce-9953-fc74116e88aa_model.json new file mode 100644 index 0000000000000000000000000000000000000000..2c6100f1662507f33016e0a8880e50d43de1605a --- /dev/null +++ b/ICCV/2025/Visual Surface Wave Elastography_ Revealing Subsurface Physical Properties via Visible Surface Waves/f44e2841-6bf1-41ce-9953-fc74116e88aa_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1b0a0ff0b3af27ba9e48c9d4a06dc7e0d95b4816dfd6dd3f419dc3e3bbc1bee7 +size 94067 diff --git a/ICCV/2025/Visual Surface Wave Elastography_ Revealing Subsurface Physical Properties via Visible Surface Waves/f44e2841-6bf1-41ce-9953-fc74116e88aa_origin.pdf b/ICCV/2025/Visual Surface Wave Elastography_ Revealing Subsurface Physical Properties via Visible Surface Waves/f44e2841-6bf1-41ce-9953-fc74116e88aa_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..ba1440ef835a8c4fc82f3444a66c294f93eb3ed8 --- /dev/null +++ b/ICCV/2025/Visual Surface Wave Elastography_ Revealing Subsurface Physical Properties via Visible Surface Waves/f44e2841-6bf1-41ce-9953-fc74116e88aa_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fe444e01a9d6c1b0437e5d54d20d13e2e8283865334f497b242b802f45a6d8b4 +size 17272156 diff --git a/ICCV/2025/Visual Surface Wave Elastography_ Revealing Subsurface Physical Properties via Visible Surface Waves/full.md b/ICCV/2025/Visual Surface Wave Elastography_ Revealing Subsurface Physical Properties via Visible Surface Waves/full.md new file mode 100644 index 0000000000000000000000000000000000000000..47c3a4ebd3f76e14af0ca3fff335d748f2bdeee9 --- /dev/null +++ b/ICCV/2025/Visual Surface Wave Elastography_ Revealing Subsurface Physical Properties via Visible Surface Waves/full.md @@ -0,0 +1,297 @@ +# Visual Surface Wave Elastography: Revealing Subsurface Physical Properties via Visible Surface Waves + +Alexander C. Ogren* Berthy T. Feng* Jihoon Ahn Katherine L. Bouman Chiara Daraio California Institute of Technology + +# Abstract + +Wave propagation on the surface of a material contains information about physical properties beneath its surface. We propose a method for inferring the thickness and stiffness of a structure from just a video of waves on its surface. Our method works by extracting a dispersion relation from the video and then solving a physics-based optimization problem to find the best-fitting thickness and stiffness parameters. We validate our method on both simulated and real data, in both cases showing strong agreement with ground-truth measurements. Our technique provides a proof-of-concept for at-home health monitoring of medically-informative tissue properties, and it is further applicable to fields such as human-computer interaction. + +# 1. Introduction + +How waves propagate on the surface of a medium reveals information about its subsurface properties. For example, by watching how ocean waves evolve and break as they near the shore, one can infer the rises and dips of the seafloor below [49]. Waves can also be observed on the surfaces of biological systems. Imagine applying a massage gun to your calf. The ripples on your skin convey information about the underlying layers of fat, muscle, and bone. In fact there is a well-defined relationship linking the thickness and stiffness of each layer to the wave propagation behavior. We propose Visual Surface Wave Elastography (VSWE), a physics-based method to estimate the thickness and stiffness of a medium from a video of waves on its surface. + +We are primarily motivated by the task of biological tissue characterization, which has broad applications including at-home health monitoring. For example, tumors [26, 30], musculoskeletal degeneration [31, 44, 45], and liver disease [23, 42] often lead to changes in tissue thickness or stiffness. However, existing techniques for elastography (i.e., measuring tissue stiffness) require specialized + +![](images/15443269b5732f1a80e1f8faf03874eaabcc5b3558ababa950a6fa6d4ce36f74.jpg) +Figure 1. Estimating subsurface tissue properties of a human leg with VSWE. From a video of surface wave motion, VSWE estimates the thickness and stiffness of the soft tissue layer. In this example, we are able to recover the three different thicknesses of the three regions highlighted on the leg. We also recover the stiffness of the leg tissue, which does not change across the leg. + +ultrasound [43, 50] and sometimes magnetic resonance [34] devices, along with medical experts to operate the equipment [7, 12]. We show that it is possible to obtain coarse estimates of thickness and stiffness given surface waves measured with just a video camera. Another application is human computer interaction (HCI), where visually inferring subsurface changes in muscle stiffness may unlock new modes of gesture recognition. In biomechanics, many regions of the body are often modeled as a series of layers of biological tissue, such as skin, muscle, or bone [4, 22]. Similarly, in this work, we model the medium as a layer of soft tissue with unknown thickness and stiffness atop a layer of bone. We leverage the mathematical relationship between the geometric and mechanical properties of the soft layer and the wave propagation behavior of the medium. + +The wave propagation behavior of a material is compactly described by a mathematical object known as a dispersion relation. Under some common biomechanical as + +sumptions in our layer model (discussed further in Sec. 3.1), the thickness and stiffness of the soft layer fully determine the dispersion relation. The main idea of our method is to find the thickness and stiffness values that lead to a dispersion relation that matches the dispersion relation extracted from surface waves observed in the video. + +In this paper, we first discuss related work on video-based and wave-based material characterization and provide background on wave mechanics. We then describe our method, which consists of taking a video of surface waves, extracting the dispersion relation from the video, and then solving an optimization problem to find the best-fitting thickness and stiffness parameters. We validate our method on real and simulated data, including real videos of gelatin-based phantoms that mimic biological tissue and realistic simulations of a human leg. + +# 2. Related work + +# 2.1. Video-based material characterization + +Videos encode rich information about the environment and physical objects. For example, previous work has leveraged surface vibrations [14] or sound [13, 19] in video to recover unknown information. For material characterization tasks, the way an object moves in a video provides useful information about its physical properties. Videos have been used to estimate material parameters of specific object classes, such as fabrics [8, 11, 35], rods [15, 16], and trees [47], whose dynamics are entirely visible. Going a step further, Visual Vibration Tomography (VVT) [18] infers subsurface material properties. Specifically, it recovers the spatially-varying stiffness and density throughout a 3D object with known geometry by analyzing its global vibrational modes. + +In contrast to vibrational modes that exist at discrete frequencies, surface waves contain information over a continuous range of frequencies and wavevectors. Surface waves have been under-utilized for video-based material characterization, even though they can also be observed in video and contain useful information about underlying physical properties [51]. Our work shows how to leverage surface waves in video to estimate geometric and mechanical properties. A surface-wave-based approach is especially beneficial when one wishes to analyze local structure without having to model the global structure (as would be required by VVT), which may be complicated or have unknown geometry. In the case of biological tissue characterization, for example, it would be impractical to model the 3D vibrational modes of the entire human anatomy in order to estimate local tissue properties. Our surface-wave-based approach circumvents the need to model a complex geometry and solve for its global modes by targeting local regions where we can analyze wave modes of a simpler geometry. + +# 2.2. Wave-based material characterization + +In general, wave-based imaging relies on an understanding of how physical characteristics of interest affect wave propagation. For example, magnetic resonance imaging [25] uses knowledge of how radio-frequency waves are absorbed and re-emitted by different types of tissues in the body. Optical tomography methods leverage near-infrared light measurements at multiple positions to recover a 3D map of tissue absorption [5, 33]. Ultrasound techniques, which are pervasive in non-destructive testing [29], medical imaging [6], and wearable technology [27], use high-frequency mechanical bulk waves to characterize tissue and locate features. Among tissue characterization techniques, transient elastography [50], shear wave elastography [43], and magnetic resonance elastography [34] leverage knowledge of bulk wave physics to estimate the elasticity of tissues and organs. However, such methods often require not only high-end, expensive equipment, but also trained medical specialists [7, 12], making regular screening infeasible. In contrast, our work leverages surface waves, which are generally less expensive to observe than their bulk counterparts. + +Although very different fields from our application of interest, geophysics and seismic imaging provide insights into the possibility of using surface waves to infer subsurface features [24, 32, 36, 37, 40, 52]. Inspired by the success of surface-wave methods in seismology, we propose to harness surface waves observed in video to infer subsurface tissue properties. Whereas previous work [9] suggested an approach using sensors sparsely placed on the skin, we leverage dense visual data. Vision-based tissue characterization would enable the next generation of health monitoring systems that take advantage of the ubiquity of visual sensors. + +# 3. Background + +# 3.1. Dispersion relations + +Just as sound waves can be expressed as the combination of simple harmonic modes, waves traveling through any medium can be expressed as the combination of wave modes of different spatial and temporal frequencies. A dispersion relation compactly describes wave propagation by defining all the possible wave modes of a medium. Specifically, it defines the wavevector and frequency of each possible wave mode, where the wavevector and frequency indicate the spatial and temporal rates of oscillation, respectively. + +Mathematically, the dispersion relation can be determined by solving the harmonic elastic wave equation (a PDE) subject to phase-shifted periodic boundary conditions. Let $\mathbf{x} = [x,y,z]^{\mathrm{T}}$ denote the spatial location and $\mathbf{u}(\mathbf{x}) = [u(\mathbf{x}),v(\mathbf{x}),w(\mathbf{x})]^{\mathrm{T}}$ denote the displacement at $\mathbf{x}$ . Assuming an isotropic linear-elastic material, the harmonic + +elastic wave equation can be written as + +$$ +- \omega^ {2} \mathcal {M} (\mathbf {u}) (\mathbf {x}) = \mathcal {K} (\mathbf {u}) (\mathbf {x}) \quad \forall \mathbf {x} \in \Omega , \tag {1} +$$ + +where $\mathcal{M}$ and $\kappa$ denote linear operators representing the mass and stiffness of the medium. The mass operator $\mathcal{M}$ depends on the density field $\rho (\mathbf{x})$ . The stiffness operator $\kappa$ depends on the elastic modulus $E(\mathbf{x})$ and Poisson's ratio $\nu (\mathbf{x})$ . The eigenvector solution $\mathbf{u}(\mathbf{x})$ describes the shape of the wave mode, and the eigenfrequency $\omega$ defines the temporal frequency of the wave. Note that the solutions also depend on the domain $\Omega$ , which in our case is parameterized by the thickness and length of the tissue layer. + +When solving Eq. (1), boundary conditions are necessary to impose real-life assumptions of the physical system. In our setting, we assume that $\Omega$ is a finite subregion of the full wave medium (e.g., a section of a leg). Furthermore, to simplify our analysis, we target waves traveling in one direction (i.e., along the $x$ direction). This leads us to apply 1D periodic boundary conditions on the domain $x \in [0, a]$ to impose the assumption that the wave medium continues past the boundary of $\Omega$ . + +Specifically, we apply the Bloch-Floquet [10, 21] (a.k.a. phase-shifted) periodic boundary conditions, which define a boundary condition for every wavenumber $\gamma \in [0,\pi /a]$ . We note that in general $\gamma$ is a wavevector, but in the case of 1D wave analysis, it is a wavenumber that indicates the number of wavelengths per unit length. For a given $\gamma$ , the boundary condition is defined as + +$$ +\mathbf {u} (x = a, y, z) = \mathbf {u} (x = 0, y, z) e ^ {\mathbf {i} \gamma a}. \tag {2} +$$ + +In words, the solution at one end $(x = 0)$ only differs from the solution at the other end $(x = a)$ by the phase shift $\gamma a$ . + +To compute a dispersion relation, we solve Eq. (1) subject to Eq. (2) for $\gamma \in [0,\pi /a]$ . For each $\gamma$ , we solve a generalized eigenvalue problem and obtain multiple $(\omega ,\mathbf{u}(\mathbf{x}))$ solution pairs. The dispersion relation is exactly the set of all $\omega$ solutions for every $\gamma$ . Supplementary details on solid mechanics may be found in Appendix D. + +# 3.2. Assumptions for tissue characterization + +For biological tissue characterization, we can model the medium as a layer of soft tissue atop a hard bone layer, where the top layer has uniform thickness $T$ . We note that throughout this work, we refer to the elastic modulus of the tissue as its stiffness. We make several assumptions to reduce the complexity of the problem: + +1. The soft layer has a uniform stiffness $E(\mathbf{x}) = E$ . +2. The density and Poisson's ratio of the soft tissue are known. We set $\rho = 1\mathrm{gcm}^{-3}$ [39] and $\nu = 0.45$ [28]. +3. The bone is much stiffer than the soft tissue. This means neither the thickness nor the exact stiffness of the bone layer matters, and we can model it as motionless. + +![](images/920f58025f0469205a87b8208e0f61a010a730a0744d09c1ec345137acac034a.jpg) +Figure 2. The dispersion relation depends on both the thickness and the stiffness of the soft tissue. We show how the dispersion relation changes when the soft tissue is made slightly thicker or slightly stiffer, with examples of the wave modes that are embedded in each dispersion relation. Wave mode dynamics occurring beneath the surface affect their expression on the surface. + +These assumptions are common in biomechanical analysis [4, 22]. Under these assumptions, the stiffness operator $\mathcal{K}$ in Eq. (1) only depends on the elastic modulus $E$ (a.k.a. stiffness) of the soft tissue, and the spatial domain $\Omega$ on which we solve Eq. (1) only depends on the thickness $T$ . This mean we can fully determine the dispersion relation based on $T$ and $E$ , so we denote the dispersion relation as + +$$ +\mathfrak {D} (T, E) = \left\{\omega_ {i} (\gamma) \right\} _ {i = 1} ^ {N} \quad \forall \gamma \in [ 0, \pi / a ], \tag {3} +$$ + +where $N$ is the number of branches, or the number of eigenvalue solutions computed for Eq. (1). + +The dispersion relation is only concerned with the spatial $(\gamma)$ and temporal $(\omega)$ frequencies of the mode shapes rather than the mode shapes $\mathbf{u}(\mathbf{x})$ themselves. From waves that express themselves, even in part, on the surface, we can extract spatial and temporal frequency information to obtain a dispersion relation. Fig. 2 shows how the dispersion relation changes with the thickness and stiffness of the soft tissue layer. Our method leverages the sensitivity of the dispersion relation to small changes in these parameters. + +# 4. Methods + +This section describes the computational method of VSWE. The input is a video of motion on the surface of the medium of interest, and the output is the estimated thickness and stiffness of the medium. In between there are two broad stages: (1) extracting a dispersion relation from the video and (2) solving for the tissue properties that best agree with the observed dispersion relation. Fig. 3 presents an overview of the pipeline. + +# 4.1. Extracting a dispersion relation from video 4.1.1. Motion extraction + +The first step is to quantify the image-space displacements in the video. Since surface wave motion tends to be small, we use phase-based motion processing [46], which is sensitive to sub-pixel displacements. Specifically, we compute + +![](images/134b65feffaab59d72cf3833bf76228b9af81ada687776cea3ce5798fdd4dff0.jpg) +Figure 3. Method overview. Given a video of the surface of the medium of interest, we first extract motion fields in image-space. From these we extract the observed dispersion relation, which dictates the spatial and temporal frequencies of waves that travel through the medium. We solve a physics-based optimization problem to estimate the thickness and stiffness values that lead to a theoretical dispersion relation that best agrees with the observed dispersion relation. + +local phase shifts in a complex steerable pyramid [41] and then convert these to pixel displacements [20]. The result is the image-space horizontal displacement $\tilde{u} (\tilde{x},\tilde{y},t)$ and vertical displacement $\tilde{v} (\tilde{x},\tilde{y},t)$ at each pixel $(\tilde{x},\tilde{y})$ and video frame $t$ , where the displacements are relative to the first frame. (Image-space coordinates are notated with a tilde to avoid confusion with world-space coordinates.) + +# 4.1.2. Estimating the dispersion relation + +We process the image-space surface displacements into a dispersion relation via the fast Fourier transform (FFT). We assume that the waves are traveling in the horizontal direction in image-space. For each row of pixels in the $\tilde{u} (\tilde{x},\tilde{y},t)$ video, we take a 2D FFT, transforming the space dimension $\tilde{x}$ to the wavenumber dimension $\gamma$ and the time dimension $t$ to the frequency dimension $\omega$ . That is, for the row where $\tilde{y} = \tilde{y}_i$ , we convert the $\tilde{u} (\tilde{x},\tilde{y} = \tilde{y}_i,t)$ signal to the complex-valued signal $\widehat{\tilde{u}} (i)(\gamma ,\omega)$ , whose magnitude gives the dispersion relation. To improve the signal-to-noise ratio, we average across all the rows and both displacement directions, resulting in the observed dispersion relation $\mathbf{D}_{\mathrm{obs}}$ : + +$$ +\mathbf {D} _ {\mathrm {o b s}} := \frac {1}{2 H} \sum_ {i = 1} ^ {H} \left(\left| \widehat {\bar {u}} (i) \right| + \left| \widehat {\bar {v}} (i) \right|\right), \tag {4} +$$ + +where $H$ is the number of rows in the image. Fig. 4 illustrates this process, showing how the 2D FFT pulls out spatial and temporal frequency content. + +The main challenge is that the observed dispersion relation likely does not fully agree with the true theoretical dispersion relation of the medium. First of all, it is incomplete: wave modes that do not prominently express on the surface or require large amounts of energy to propagate may not appear strongly in the video, and frequencies above the Nyquist sampling rate of the video cannot be captured. Second of all, the extracted motion may be noisy due to camera noise or spurious motion (from, e.g., lights flickering or rigid-body motion). VSWE has some margin of error due to the discrepancy between the observed and true dispersion relations, although we show in our experiments in Sec. 5.1 it is sensitive to $5\%$ changes in the true parameters. + +# 4.2. Estimating the thickness and stiffness + +The next stage is to find the geometric and mechanical parameters that best explain the observed dispersion relation. We do this via physical simulation and optimization. For a given hypothesized thickness $T$ and stiffness $E$ , we employ the finite element method (FEM) to numerically compute the dispersion relation. We wrote specialized FEM code to do so efficiently, which can be accessed at github/aco8ogren/tissue-dispersion. The goal is to maximize an objective function that rewards similarity between $\mathfrak{D}(T,E)$ and $\mathbf{D}_{\mathrm{obs}}$ . + +After testing various objective functions, we found that treating dispersion relations as images and using the structural similarity (SSIM) index [48] to work best. That is, we aim to maximize the SSIM between the images of the observed and proposed dispersion relations. The observed dispersion relation $\mathbf{D}_{\mathrm{obs}}$ , since it is derived via a 2D FFT, is already represented as an image. However, the physics-based dispersion relation $\mathfrak{D}(T,E)$ is computed as a set of curves. We transform $\mathfrak{D}(T,E)$ into an image $\mathbf{D}_{\mathrm{hyp}}(T,E)$ by assigning intensities with a Gaussian kernel based on the distance from each $(\gamma ,\omega)$ pixel to the curves. Then we solve the following optimization problem: + +$$ +T ^ {*}, E ^ {*} = \underset {T, E} {\arg \max } \operatorname {S S I M} \left(\mathbf {D} _ {\text {h y p}} (T, E), \mathbf {D} _ {\text {o b s}}\right). \tag {5} +$$ + +We solve Eq. (5) with a grid search over possible thickness and stiffness values, although more efficient approaches can be taken for computationally-constrained settings. + +# 5. Results + +We validate our approach on real and simulated data. We demonstrate remarkable accuracy on real gelatin samples, and we demonstrate recovering spatially-varying thickness and stiffness across a simulated 3D human leg. + +# 5.1. Simulated plane strain + +We created a two-layer tissue model in COMSOL [1] that would allow us to simulate samples with ground-truth geometric and mechanical parameters. We set stiffness values + +![](images/d97223050881ff5a51ab33d5ddb4dde589a548198c1e06a57ec3c9298c4422ab.jpg) +Figure 4. Obtaining a dispersion relation from image-space motion. Here we demonstrate with the vertical displacements taken from a real video of a gelatin sample. Taking a 1D FFT across time for every pixel allows us to separate wave modes by frequency $(\omega)$ . Taking another 1D FFT across space (in the $\tilde{x}$ direction) decomposes the wave modes into spatial frequencies $(\gamma)$ . The result is a 2D FFT representing the dispersion relation. We average the dispersion relations across image rows and both displacement directions $(\tilde{u}$ and $\tilde{v})$ . + +on the order of $10\mathrm{kPa}$ , which is a reasonable range for soft tissue [17]. We simulated the model's response to a chirp excitation signal applied to the leftmost side of the surface of the sample. The simulations were done with the plane-strain assumption, which allows for modeling 3D dynamics with only two dimensions by assuming that strain is zero in the $z$ direction (more details in Appendix D). We used the simulated horizontal and vertical displacements, sampled at $600\mathrm{Hz}$ and $\sim 4300\mathrm{samp / m}$ over 1 s and $29.5\mathrm{cm}$ , to obtain dispersion relations. + +We tested the sensitivity of our method by applying slight changes to the true thickness and stiffness of the simulated sample. That is, for a certain thickness and stiffness, we can simulate samples assuming those parameters as well as perturbations of those parameters and assess whether VSWE picks up on these slight changes. As an experiment, we simulated samples at 5 distinct thicknesses and 4 distinct moduli. For each thickness-stiffness pair, we looked at $5\%$ and $10\%$ perturbations of either parameter. This led to 9 clusters with 9 simulated samples each. As Fig. 5 shows, within each cluster, the estimated parameters track with changes in the true parameters, even when the true parameters have changed by only $5\%$ . + +# 5.2. Real videos of gelatin-based phantoms + +To test our method on real data, we created gelatin-based phantoms (i.e., samples mimicking biological tissue) of varying thicknesses. We poured varying amounts of gelatin (1000, 1100, and $1500~\mathrm{mL}$ ) into the same-sized container, leading to three different thicknesses. We set the samples in the refrigerator for about 24 hours. Once they were set, we sprinkled garlic powder onto the samples to create texture for motion extraction. + +For each set sample, we measured the thickness with calipers, and we used rheometry to obtain ground-truth stiffness values. To obtain videos, we applied a shaker at one end of the sample to excite waves with a chirp signal and recorded the surface of the sample with a high-speed camera at 600 FPS at a resolution of $96 \times 1280$ pixels (each video was about four seconds long, with an observation ex + +![](images/a50a70801d75623f0f27463608a78f9db9b6d98d0609ec9e0b74776238a161ad.jpg) +Figure 5. Sensitivity analysis. We applied VSWE to plane-strain simulations with a range of true thickness and stiffness values, along with slight perturbations of those values. In each cluster, we perturbed the true parameters by $\pm 5\%$ and $\pm 10\%$ from the central values and found that VSWE was sensitive to these changes. + +![](images/5fda282aa8f8f4b0d83e941ad730fd67d74dcf2c582b858f21deeb59e71ba339.jpg) +Figure 6. Experiment setup. A shaker was applied on one side of the gelatin-based phantom to excite waves in the medium. A high-speed camera, zoomed in on the phantom's surface, captured videos of the response to the shaker. As the left picture shows, we took temperature recordings with a thermometer. + +tent of about $30\mathrm{cm}$ ). Fig. 6 shows pictures of the experiment setup, and more details can be found in Appendix F. + +For each of the three samples, we took many rheometry measurements and videos over the course of about one hour after removing the sample from the refrigerator. In total we obtained about 60 videos per sample over time, each with corresponding ground-truth parameters. Fig. 7 shows the inferred parameters for each sample and each point in time. Fig. 7(a) shows that our method clearly identifies the three + +![](images/3d6be07da545250938739aef1ddf283736279bfb84554ef66395f44ce160da6f.jpg) +Figure 7. (a) Estimated thickness and stiffness for three phantoms over a range of temperatures. The phantoms were produced with three different volumes (1000, 1100, $1500~\mathrm{mL}$ ) of gelatin but set in the same-sized container, thus generating three different thicknesses. The thickness of each sample was measured with calipers. We produced a confidence interval of the ground-truth thickness by taking the 0.2 and 0.8 quantiles of multiple calipers measurements. The VSWE-estimated thickness consistently falls within the confidence interval. Note that while each phantom was polymerized according to the same approximate recipe, slight variations in the recipe, along with slightly different polymerization durations, led to some variation in the stiffness across samples. The temperature changes occurred as the samples spent more time at room temperature after being removed from the fridge. (b) Estimated stiffness values for the thinnest sample over the same range of temperatures. The ground-truth elastic modulus was measured with rheometry over a similar temperature range. Because the modulus of hydrogels can depend on both temperature and frequency [53], the rheometry measurements show a different stiffness range at each temperature. Note that the rheometry measurements were only taken between $10 - 100\mathrm{Hz}$ due to instrument limitations, while the excitation signal ranged from $40 - 200\mathrm{Hz}$ . This may explain some of the discrepancy between VSWE and rheometry, especially for the colder (and thus stiffer) samples which exhibit a stronger expression of higher-frequency wave dynamics. Regardless, in each case, VSWE estimates the stiffness extremely well, within $1.2\%$ error of the rheometry range. + +![](images/762c38a53582eb66b7f0e6e4f9bf8de3c9a7ddbc9935c1e9c974582067f2e148.jpg) + +different thicknesses of the samples. As a sample spends more time out of the refrigerator, its temperature increases and its stiffness decreases. Fig. 7(b) shows that the estimated stiffness decreases accordingly with temperature. + +# 5.3. 3D human leg with spatially-varying thickness + +As another step towards realism, we tested VSWE on a simulated female human leg. The anatomical geometry of the leg was obtained as an STL file from the dataset of Andreassen et al. [3], who created 3D models from the National Library of Medicine's Visible Human Project [2]. We simulated the leg's response to a chirp excitation applied on the leg in COMSOL, running a full 3D physics simulation without any assumptions besides linear elasticity. For the sake of computational feasibility, we ran the simulation on the lower half of the leg, which we refer to as the calf region. To obtain dispersion relations, we considered the simulated displacements in two directions: one tangent to the leg's surface and one normal to the leg's surface. + +Displacements were sampled at $800\mathrm{Hz}$ for 1s. The observation windows varied per region of the leg, with sizes between $\sim 3\mathrm{cm}$ and $\sim 11.5\mathrm{cm}$ . See Appendix E for more details. + +We applied a sweeping window across the upper calf region, estimating the thickness and modulus at each window location. Fig. 8 shows how, as the window sweeps around the leg, the inferred thickness changes. We computed the true thickness by taking the distance from the point on the skin to the nearest point on the bone. This results in a thickness distribution within the window, since the thickness changes slightly in the lengthwise direction of the leg, as well. Fig. 8 shows the inferred thickness and the distribution of ground-truth thicknesses for each location of the sweeping window. We find that the estimated thickness roughly agrees with the true thickness distribution. There appears to be slightly more error near the edges of the simulated domain, which may be due to boundary effects. While the physics-based simulation in COMSOL is well-modeled in the interior of the domain, there may be artifacts near the boundary of the domain, making it harder to infer the correct tissue parameters near the boundary. + +We also applied VSWE to three different windows in the lower calf region. Rather than perform a sweep in this region, we chose to focus on three distinct windows since there are three subregions in the lower calf area that have notably different thicknesses, due to the tissue structure changing a lot near the ankle. As Fig. 8 shows, the inferred thickness agrees with the true thickness distribution for each subregion. Our method also recovered the constant stiffness throughout the calf, as shown in Fig. 1. + +# 5.3.1. Objective function ablation + +As mentioned in Sec. 4.2, we find that SSIM performs best for the optimization objective function. This choice was crucial for obtaining acceptable results with the 3D leg. Fig. 9 shows the optimization landscape for different choices of the objective function, demonstrated on both the upper calf simulation and a real video of gelatin. We compare SSIM to the following objective functions: a curve-based objective function, the negative mean squared error (MSE) between the images, and the peak signal-to-noise ratio (PSNR) between the images. The curve-based loss function goes through all the curves (i.e., $\{\omega_i(\gamma)\}_{i=1}^N \forall \gamma \in [0, \pi/a])$ in $\mathfrak{D}(T,E)$ and integrates the values of the observed dispersion relation $\mathbf{D}_{\mathrm{obs}}$ along the curves. Fig. 9 shows that SSIM leads to the sharpest loss landscape and the most accurate estimated parameters. More details can be found in Appendix B. + +# 6. Discussion + +# 6.1. Practical considerations + +VSWE's performance depends on many factors, including the types of waves supported by the tissue (e.g., surface + +![](images/c6b42685fefa299d1724eb330c547ca2021fde544537b55084800014eae8cf83.jpg) +(a) anatomical geometry & VSWE on upper and lower calf + +![](images/e55698f87742d5241602e273e70bb39187b4d5ee6afc1f36537f82d4424d1fe3.jpg) +(b) sweep across upper calf region + +![](images/4b0df91ac869ad340699389e7b21f6f59dfc03382d60319f8d01e918451655c1.jpg) +(c) regions of drastically different thickness near ankle + +![](images/3a9280712bdc38d70fa548b162d1093db1686e8816755afd0951230b7a6f5b03.jpg) +Figure 8. 3D anatomical inference. Panel (a) shows the realistic 3D leg anatomy (taken from the Visible Human Project [2]). We ran a full 3D simulation of the leg's response to a chirp excitation applied on the skin. The observed dispersion relation, along with the optimization landscape and hypothesized dispersion relation of the optimal parameters, is shown for two regions: one on the upper calf and one on the lower calf near the ankle. Panel (b) shows results of sweeping an observation window across a region of the upper calf. A ground-truth thickness distribution was obtained by computing the distance from the skin to the nearest point on the bone across the observation window. The VSWE-estimated thickness as the observation window slides from left to right tracks with the changing thickness of the leg. Panel (c) shows results for three distinct regions near the ankle. The VSWE estimations reflect the drastically different thicknesses of these regions. +Figure 9. Ablation of optimization objective function. The optimization landscape is shown for various objective functions, applied in two different scenarios. Each landscape is visualized with its own color map limits, but bright yellow and dark blue indicate high (good) and low (bad) objective values, respectively. SSIM gives the sharpest optimization landscape. For the simulated upper calf, all the objective functions besides SSIM lead to the wrong optimal parameters (cross hairs indicate target parameters). Additionally, note how the simulated upper calf leads to a blurrier dispersion relation and hence blurrier optimization landscapes than the real gelatin sample, which has a simple geometry. + +waves must be observable), the quality and extent of the observation (e.g., sufficient texture for motion extraction), and the quality of the physics model. + +One practical consideration is the resolution of the FEM mesh used to model physics, as the element size determines the modeling accuracy and the smallest modeled wavelength. Fig. 10 shows that as the element size decreases, the estimated parameters approach the true parameters. + +![](images/44df775864fa3a9271fe56ffa6bd0d248ed3dd43586507b017ae880b9a7f81cd.jpg) +Figure 10. Ablation of FEM mesh resolution. As the mesh becomes finer, the VSWE-estimated parameters approach the true parameters used for the COMSOL plane-strain simulation. + +![](images/bc716d18d02626608e66a9e94323b7861eb5d1f2fde86916eacfb1b8ac21c8a5.jpg) +Figure 11. Ablation of spatial extent of observation. Keeping all else fixed, decreasing the spatial length $L$ of the observation window degrades the quality of the observed dispersion relation. A degraded dispersion relation in turn leads to a fuzzy optimization landscape (see Fig. 9 for an example). + +Another factor is that the spatial length $L$ of the obser + +![](images/aa611bc2de453877ce44856dbfc694450d44ddcdf6e8e51592f3101d05e98790.jpg) +Figure 12. Leveraging similitude of characteristic numbers enables the application of VSWE to extremely different parameter ranges. In this simulated example, we show that by preserving the values of $\pi_{1-6}$ , we see a similar observed signal from a system with parameters $10\times$ different than the primary ranges studied in this paper. In particular, we rescaled the size of the observation window, the thickness of the tissue, the spatial sampling rate, the total observation time, the frequency range of the excitation, and the temporal sampling rate. Note that the observation window on the right is smaller than that of the most degraded FFT in Fig. 11, yet the FFT here has not degraded. + +vation window determines the largest wavelength and the number of wavelengths that can be observed. Fig. 11 shows how, keeping all else fixed, the FFT degrades as the observed domain shrinks, which makes inference more challenging. + +# 6.2. Characteristic numbers + +Characteristic numbers (a.k.a. dimensionless numbers or $\pi$ -groups) allow us to roughly quantify limiting effects like those in the ablation studies of Fig. 10 and Fig. 11. They can be used to understand working limits for VSWE, and they give guidance for applying VSWE to scenarios beyond those studied in this work. + +Some characteristic numbers reflect the observability of wave modes, such as $\pi_1 \coloneqq \gamma L$ , $\pi_2 \coloneqq \mathrm{PPM} / \gamma$ , $\pi_3 \coloneqq \omega \tau$ , and $\pi_4 \coloneqq \mathrm{FPS} / \omega$ , where PPM (pixels per meter) is the spatial sampling rate, FPS (frames per second) is the temporal sampling rate, and $\tau$ is the total observation time. Increasing $\pi_{1-4}$ improves VSWE's performance by improving the quality of the FFT dispersion. The effect of varying $\pi_1$ can be observed in Fig. 11. + +Other characteristic numbers reflect the numerical accuracy of the FEM model used to fit the observation, such as $\pi_5\coloneqq \frac{1}{\gamma e}$ , where $e$ is the FEM element size. Increasing $\pi_5$ improves performance by ensuring that physics is wellmodeled in the parameter-fitting process (Fig. 10). + +Yet other characteristic numbers reflect the actual physics that are possible in the system. For example, $\pi_6 \coloneqq \gamma T$ conveys the "shallowness" of the wave. If $\pi_6$ is too + +large, then waves behave similarly to how they would in an infinitely-thick medium, making it difficult to pinpoint the thickness. If $\pi_6$ is too small, waves above a certain wavelength are not permitted to exist, making it necessary to capture smaller wavelengths. $^2$ + +Across our experiments, we found that $\pi_1$ , $\pi_4$ , and $\pi_6$ were often the most relevant to practical working limits and helped us to identify and tune the minimum $L$ , minimum FPS, and valid $T$ ranges, respectively. See Appendix C for examples of $\pi$ values in our experiments. + +Fig. 12 provides an example of how to take advantage of characteristic numbers. Specifically, we overcome the challenge of a smaller observed domain (as illustrated in Fig. 11) by rescaling other parameters to preserve the relevant characteristic numbers and thus maintain the quality of the observed dispersion relation. In general, one should look to characteristic numbers as guidelines for adjusting system and/or inference parameters in order to maintain the performance of VSWE across a variety of scenarios. + +# 7. Conclusion + +We have introduced Visual Surface Wave Elastography for estimating subsurface geometric and mechanical properties from visible surface waves. Our method works by extracting spatial and temporal frequency information from surface waves observed in video. By representing this information in the form of a dispersion relation, we can use a fully physics-based approach to solve for the properties that explain the wave behavior observed in the video. We validated our method on both simulated and real data, showing that our method is sensitive to small changes in the properties of tissue-like media. We demonstrated the promise of applying VSWE in the real world by testing it on real videos of gelatin-based phantoms and on a simulated realistic 3D human leg. Our work shows the potential of everyday visual data to reveal critical information deep beneath the skin. + +# Acknowledgments + +ACO and CD are supported by DOE award no. DE-SC0021253 and the New Frontiers of Sound (NewFoS) NSF center. BTF and KLB acknowledge support from the NSF GRFP, the Amazon AI4Science Partnership Discovery Grant, and the Heritage Medical Research Institute (HMRI) Fellowship. JA and CD are supported by the HMRI and the NSF Center to Stream Healthcare in Place (C2SHIP) award no. 2052827. + +# References + +[1] COMSOL AB. Comsol multiphysics® v5.5, 2019. 4 +[2] M.J. Ackerman. The visible human project. Proceedings of the IEEE, 86(3):504-511, 1998. 6, 7 +[3] Thor E Andreassen, Donald R Hume, Landon D Hamilton, Karen E Walker, Sean E Higinbotham, and Kevin B Shelburne. Three dimensional lower extremity musculoskeletal geometry of the visible human female and male. Scientific Data, 10(1):34, 2023. 6 +[4] J. W. Andrews, M. J. Adams, and T. D. Montenegro-Johnson. A universal scaling law of mammalian touch. Science Advances, 6(41):eabb6912, 2020. 1, 3 +[5] S R Arridge. Optical tomography in medical imaging. Inverse Problems, 15(2):R41, 1999. 2 +[6] Danilo Avola, Luigi Cinque, Alessio Fagioli, Gianluca Foresti, and Alessio Mecca. Ultrasound medical imaging techniques: a survey. ACM Computing Surveys (CSUR), 54 (3):1-38, 2021. 2 +[7] R. A. Bell. Economics of MRI technology. Journal of magnetic resonance imaging: JMR1, 6(1):10-25, 1996. 1, 2 +[8] Kiran S. Bhat, Christopher D. Twigg, Jessica K. Hodgins, Pradeep K. Khosla, Zoran Popovic, and Steven M. Seitz. Estimating cloth simulation parameters from video. In Proceedings of the 2003 ACM SIGGRAPH/Eurographics Symposium on Computer Animation, page 37-51, Goslar, DEU, 2003. Eurographics Association. 2 +[9] Marcin Bialas and Bojan B. Guzina. On the viscoelastic characterization of thin tissues via surface-wave sensing. International Journal of Solids and Structures, 48(14):2209-2217, 2011. 2 +[10] Felix Bloch. Über die quantenmechanik der elektronen in kristallgittern. Zeitschrift für physik, 52(7):555-600, 1929. 3 +[11] Katherine L Bouman, Bei Xiao, Peter Battaglia, and William T Freeman. Estimating the material properties of fabric from video. In Proceedings of the IEEE International Conference on Computer Vision, pages 1984-1991, Sydney, Australia, 2013. IEEE. 2 +[12] Nathalie J Bureau and Daniela Ziegler. Economics of musculoskeletal ultrasound. Current radiology reports, 4:1-7, 2016. 1, 2 +[13] Ziyang Chen, Shengyi Qian, and Andrew Owens. Sound localization from motion: Jointly learning sound direction and camera rotation. International Conference on Computer Vision (ICCV), 2023. 2 +[14] Abe Davis, Michael Rubinstein, Neal Wadhwa, Gautham Mysore, Fredo Durand, and William T. Freeman. The visual microphone: Passive recovery of sound from video. ACM Transactions on Graphics (Proc. SIGGRAPH), 33(4):79:1-79:10, 2014. 2 +[15] Abe Davis, Katherine L Bouman, Justin G Chen, Michael Rubinstein, Fredo Durand, and William T Freeman. Visual vibrometry: Estimating material properties from small motion in video. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5335-5343, Boston, MA, USA, 2015. IEEE. 2 + +[16] Abe Davis, Katherine L. Bouman, Justin G. Chen, Michael Rubinstein, Oral Buyukozturk, Frédo Durand, and William T. Freeman. Visual vibrometry: Estimating material properties from small motions in video. IEEE Trans. Pattern Anal. Mach. Intell., 39(4):732-745, 2017. 2 +[17] V Egorov, S Tsyuryupa, S Kanilo, M Kogit, and AJME Sarvazyan. Soft tissue elastometer. Medical engineering & physics, 30(2):206-212, 2008. 5 +[18] Berthy T Feng, Alexander C Ogren, Chiara Daraio, and Katherine L Bouman. Visual vibration tomography: Estimating interior material properties from monocular video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16231-16240, 2022. 2 +[19] Chao Feng, Ziyang Chen, and Andrew Owens. Self-supervised video forensics by audio-visual anomaly detection. Computer Vision and Pattern Recognition (CVPR), 2023. 2 +[20] David J Fleet and Allan D Jepson. Computation of component image velocity from local phase information. International journal of computer vision, 5(1):77-104, 1990. 4 +[21] Gaston Floquet. Sur les équations différentielles linéaires à coefficients périodiques. In Annales scientifiques de l'École normale supérieure, pages 47-88, 1883. 3 +[22] Freija Geldof, Behdad Dashtbozorg, Benno HW Hendriks, Henricus JCM Sterenborg, and Theo JM Ruers. Layer thickness prediction and tissue classification in two-layered tissue structures using diffuse reflectance spectroscopy. Scientific reports, 12(1):1698, 2022. 1, 3 +[23] Penelope C. Georges, Jia-Ji Hui, Zoltan Gombos, Margaret E. McCormick, Andrew Y. Wang, Masayuki Uemura, Rosemarie Mick, Paul A. Janmey, Emma E. Furth, and Rebecca G. Wells. Increased stiffness of the rat liver precedes matrix deposition: Implications for fibrosis. American Journal of Physiology-Gastrointestinal and Liver Physiology, 293(6):G1147-G1154, 2007. 1 +[24] N. Gucunski and R. D. Woods. Numerical simulation of the SASW test. Soil Dynamics and Earthquake Engineering, 11 (4):213-227, 1992. 2 +[25] Ray Hashman Hashemi, William G Bradley, and Christopher J Lisanti. MRI: the basics: The Basics. Lippincott Williams & Wilkins, 2012. 2 +[26] Kenneth Hoyt, Benjamin Castaneda, Man Zhang, Priya Nigwekar, P Anthony di Sant'Agnese, Jean V Joseph, John Strang, Deborah J Rubens, and Kevin J Parker. Tissue elasticity properties as biomarkers for prostate cancer. Cancer Biomarkers, 4(4-5):213-225, 2008. 1 +[27] Hao Huang, Ray S Wu, Muyang Lin, and Sheng Xu. Emerging wearable ultrasound technology. IEEE Transactions on Ultrasonics, Ferroelectrics, and Frequency Control, 71(7): 713-729, 2023. 2 +[28] Yousef Javanmardi, Huw Colin-York, Nicolas Szita, Marco Fritzsche, and Emad Moeendarbary. Quantifying cell-generated forces: Poisson's ratio matters. Communications physics, 4(1):237, 2021. 3 +[29] Jitesh Jodhani, Abhay Handa, Anandita Gautam, Ramakant Rana, et al. Ultrasonic non-destructive evaluation of compos- + +ites: A review. Materials Today: Proceedings, 78:627-632, 2023. 2 +[30] Sangwoo Kwon, Woochul Yang, Donggerami Moon, and Kyung Sook Kim. Comparison of Cancer Cell Elasticity by Cell Type. Journal of Cancer, 11(18):5403-5412, 2020. 1 +[31] Lilian Lacourpaille, François Hug, Arnaud Guével, Yann Pérez, Armelle Magot, Jean-Yves Hogrel, and Antoine Nordez. Non-invasive assessment of muscle stiffness in patients with duchenne muscular dystrophy. Muscle & Nerve, 51(2):284-286, 2015. 1 +[32] Chih-Ping Lin, Chun-Hung Lin, and Chih-Jung Chien. Dispersion analysis of surface wave testing - SASW vs. MASW. Journal of Applied Geophysics, 143:223-230, 2017. 2 +[33] Chao Liu, Akash K. Maity, Artur W. Dubrawski, Ashutosh Sabharwal, and Srinivasa G. Narasimhan. High resolution diffuse optical tomography using short range indirect subsurface imaging. In 2020 IEEE International Conference on Computational Photography (ICCP), pages 1-12, 2020. 2 +[34] Yogesh K. Mariappan, Kevin J. Glaser, and Richard L. Ehman. Magnetic resonance elastography: A review. *Clinical Anatomy*, 23(5):497-511, 2010. 1, 2 +[35] Eder Miguel, Derek Bradley, Bernhard Thomaszewski, Bernd Bickel, Wojciech Matusik, Miguel A Otaduy, and Steve Marschner. Data-driven estimation of cloth simulation models. In Computer Graphics Forum, pages 519-528, Hoboken, NJ, USA, 2012. Wiley Online Library, Wiley Online Library. 2 +[36] Soheil Nazarian and Kenneth Stokoe. Evaluation of Moduli and Thicknesses of Pavement Systems by Spectral-Analysis of-Surface-Waves Method. The Study of New Technologies for Pavement Evaluation Research Project 3-8-80-256, as conducted for Texas State Department of Highways and Public Transportation, by the Center for Transportation Research Bureau of Engineering Research - The University of Texas at Austin, 1983. 2 +[37] Choon Park. MASW for geotechnical site investigation. The Leading Edge, 32(6):656-662, 2013. 2 +[38] Lord Rayleigh. On Waves Propagated along the Plane Surface of an Elastic Solid. s1-17(1):4-11. 12 +[39] Babette S. Zemel. Chapter 18 - body composition during growth and development. In Human Growth and Development (Second Edition), pages 461-486. Academic Press, Boston, second edition edition, 2012. 3 +[40] Nikolai M. Shapiro, Michel Campillo, Laurent Stehly, and Michael H. Ritzwoller. High-Resolution Surface-Wave Tomography from Ambient Seismic Noise. Science, 307 (5715):1615-1618, 2005. 2 +[41] Eero P Simoncelli and William T Freeman. The steerable pyramid: A flexible architecture for multi-scale derivative computation. In Proceedings., International Conference on Image Processing, pages 444-447, Washington, D.C., USA, 1995. IEEE, IEEE. 4 +[42] Shan Sun, Zhenyuan Song, Scott J. Cotler, and Michael Cho. Biomechanics and functionality of hepatocytes in liver cirrhosis. Journal of Biomechanics, 47(9):2205-2210, 2014. 1 +[43] Mihra S. Taljanovic, Lana H. Gimber, Giles W. Becker, L. Daniel Latt, Andrea S. Klauser, David M. Melville, Liang + +Gao, and Russell S. Witte. Shear-Wave Elastography: Basic Physics and Musculoskeletal Applications. *RadioGraphics*, 37(3):855-870, 2017. 1, 2 +[44] N. H. Varady and A. J. Grodzinsky. Osteoarthritis year in review 2015: Mechanics. Osteoarthritis and Cartilage, 24 (1):27-35, 2016. 1 +[45] Kelley M. Virgilio, Kyle S. Martin, Shayn M. Peirce, and Silvia S. Blemker. Multiscale models of skeletal muscle reveal the complex effects of muscular dystrophy on tissue mechanics and damage susceptibility. Interface Focus, 5(2): 20140080, 2015. 1 +[46] Neal Wadhwa, Michael Rubinstein, Frédo Durand, and William T Freeman. Phase-based video motion processing. ACM Transactions on Graphics (TOG), 32(4):1-10, 2013. 3 +[47] Bohan Wang, Yili Zhao, and Jernej Barbič. Botanical materials based on biomechanics. ACM Transactions on Graphics (TOG), 36(4):1-13, 2017. 2 +[48] Zhou Wang, A.C. Bovik, H.R. Sheikh, and E.P. Simoncelli. Image quality assessment: From error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4): 600-612, 2004. 4, 11 +[49] Paul Webb. Introduction to oceanography. Roger Williams University, 2021. 1, 8, 12 +[50] Vincent Wai-Sun Wong and Henry Lik-Yuen Chan. Transient elastography. Journal of Gastroenterology and Hepatology, 25(11):1726-1731, 2010. 1, 2 +[51] Tianyuan Zhang, Mark Sheinin, Dorian Chan, Mark Rau, Matthew O'Toole, and Srinivasa G. Narasimhan. Analyzing physical impacts using transient surface wave imaging. In Proc. IEEE CVPR, 2023. 2 +[52] Zhendong Zhang, Tariq Alkhalifah, and Yike Liu. Full dispersion-spectrum inversion of surface waves. JGR Solid Earth, 2024. 2 +[53] S Zörner, M Kaltenbacher, R Lerch, A Sutor, and M Döllinger. Measurement of the elasticity modulus of soft tissues. Journal of Biomechanics, 43(8):1540-1545, 2010. 6 \ No newline at end of file diff --git a/ICCV/2025/Visual Surface Wave Elastography_ Revealing Subsurface Physical Properties via Visible Surface Waves/images.zip b/ICCV/2025/Visual Surface Wave Elastography_ Revealing Subsurface Physical Properties via Visible Surface Waves/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..e5d5b4cf81eae25e3b228c4d9c4e86bee1253166 --- /dev/null +++ b/ICCV/2025/Visual Surface Wave Elastography_ Revealing Subsurface Physical Properties via Visible Surface Waves/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8ea7a22f69b60a78b82da906923de6268c8d499be5822757ea084429101b3e80 +size 533323 diff --git a/ICCV/2025/Visual Surface Wave Elastography_ Revealing Subsurface Physical Properties via Visible Surface Waves/layout.json b/ICCV/2025/Visual Surface Wave Elastography_ Revealing Subsurface Physical Properties via Visible Surface Waves/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..70e455fb3a1ea14627b1a11daeaefede8e3ba68c --- /dev/null +++ b/ICCV/2025/Visual Surface Wave Elastography_ Revealing Subsurface Physical Properties via Visible Surface Waves/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a2b671898a369062f40426ac8df3ccefe4777fdd586a763182d2e480672f876a +size 407778 diff --git a/ICCV/2025/Visual Test-time Scaling for GUI Agent Grounding/b0a9cda2-be4a-426f-a5b8-2a6b64ffb5fe_content_list.json b/ICCV/2025/Visual Test-time Scaling for GUI Agent Grounding/b0a9cda2-be4a-426f-a5b8-2a6b64ffb5fe_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..da184a2a15df6cf4267662c5a97ad9d1967d9688 --- /dev/null +++ b/ICCV/2025/Visual Test-time Scaling for GUI Agent Grounding/b0a9cda2-be4a-426f-a5b8-2a6b64ffb5fe_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c99401e07bb9de14cdd8d033ade604e498e84b4f59692ea0d9d334a7fc992370 +size 78059 diff --git a/ICCV/2025/Visual Test-time Scaling for GUI Agent Grounding/b0a9cda2-be4a-426f-a5b8-2a6b64ffb5fe_model.json b/ICCV/2025/Visual Test-time Scaling for GUI Agent Grounding/b0a9cda2-be4a-426f-a5b8-2a6b64ffb5fe_model.json new file mode 100644 index 0000000000000000000000000000000000000000..f7f5634d347f6960de804fbb500d2bac571b30bb --- /dev/null +++ b/ICCV/2025/Visual Test-time Scaling for GUI Agent Grounding/b0a9cda2-be4a-426f-a5b8-2a6b64ffb5fe_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bbeb9a90080a1434526b33c8a93d7e6e8d2bd623fa5cb384bcafe3584536a945 +size 95947 diff --git a/ICCV/2025/Visual Test-time Scaling for GUI Agent Grounding/b0a9cda2-be4a-426f-a5b8-2a6b64ffb5fe_origin.pdf b/ICCV/2025/Visual Test-time Scaling for GUI Agent Grounding/b0a9cda2-be4a-426f-a5b8-2a6b64ffb5fe_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..06806f61f54780cc247cc322d91112ead691e653 --- /dev/null +++ b/ICCV/2025/Visual Test-time Scaling for GUI Agent Grounding/b0a9cda2-be4a-426f-a5b8-2a6b64ffb5fe_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:80d329c80a224604d41e257a4d90c15d56bb81f544b232f945084c99b6a5847c +size 4214433 diff --git a/ICCV/2025/Visual Test-time Scaling for GUI Agent Grounding/full.md b/ICCV/2025/Visual Test-time Scaling for GUI Agent Grounding/full.md new file mode 100644 index 0000000000000000000000000000000000000000..7ba3fd39d55a52741508162e441ecf39f481d082 --- /dev/null +++ b/ICCV/2025/Visual Test-time Scaling for GUI Agent Grounding/full.md @@ -0,0 +1,251 @@ +# Visual Test-time Scaling for GUI Agent Grounding + +Tiang Luo $^{1,2}$ Lajanugen Logeswaran $^{2,\dagger}$ Justin Johnson $^{1,\dagger}$ Honglak Lee $^{1,2,\dagger}$ University of Michigan $^{1}$ LG AI Research $^{2}$ equal advising + +# Abstract + +We introduce RegionFocus, a visual test-time scaling approach for Vision Language Model Agents. Understanding webpages is challenging due to the visual complexity of GUI images and the large number of interface elements, making accurate action selection difficult. Our approach dynamically zooms in on relevant regions, reducing background clutter and improving grounding accuracy. To support this process, we propose an image-as-map mechanism that visualizes key landmarks at each step, providing a transparent action record and enables the agent to effectively choose among action candidates. Even with a simple region selection strategy, we observe significant performance gains of $28+\%$ on ScreenSpot-Pro and $24+\%$ on WebVoyager benchmarks on top of two state-of-the-art open vision language model agents, UI-TARS-72B and Qwen2.5-VL-72B, highlighting the effectiveness of visual test-time scaling in interactive settings. We achieve a new state-of-the-art grounding performance of $61.6\%$ on the ScreenSpot-Pro benchmark by applying RegionFocus to a Qwen2.5-VL-72B model. Our code is publicly available at https://github.com/tiangeluo/RegionFocus. + +# 1. Introduction + +Graphical user interface (GUI) agents have become increasingly pivotal in modern computing, powering applications ranging from automated web browsing to intuitive operating system navigation [1, 21]. With the proliferation of large-scale vision-language models (VLMs), researchers have sought to harness both textual and visual information to build more capable interactive systems [2]. While many existing frameworks rely heavily on text-based reasoning [39, 46] or simple visual grounding [11, 22], real-world GUIs often contain a substantial number of irrelevant elements—such as menu bars, ads, and extraneous buttons—that can overwhelm purely textual or naive visual approaches. This mismatch between text-heavy inference and the visual complexity of GUIs leads to frequent errors (e.g., clicking the wrong button or navigating to an unintended section). Since these tasks are typically high-level, + +![](images/07eea9e08ac7bad51b050308dba592450c2504bc1a4ebccb4fb3355092edcb0b.jpg) +Figure 1. Overview. When GUI agents encounter execution errors, we instruct the model to focus on a specific point of interest and extract multiple sub-regions around this focal point (1). The agent then independently generates candidate actions for each sub-region. Actions that interact with a specific coordinate are marked with pink-star landmarks (e.g., "click") to visually indicate the relevant location, along with action numbers (2). We retain the pink-star landmarks annotated with history indices to track RegionFocus's past interactions for diverse exploration (3). + +such low-level mistakes accumulate and ultimately result in higher failure rates and poorer overall performance. + +Recent research on GUI agents typically falls into two main categories: those relying on textual cues for planning and reasoning, and those incorporating visual information through VLMs. Text-based approaches often generate text labels or bounding boxes for each visual element to guide agent actions [22, 46]. However, they can struggle when textual descriptions are ambiguous, incomplete, or fail to capture crucial visual features (e.g., floating windows), even when using accessibility trees. On the other hand, vision-based pipelines [11, 28] often rely heavily on the VLM's ability to ground visual elements. We observe that many errors arise from inadvertently clicking empty or incorrect interface components, underscoring the limitations of existing single-inference visual grounding methods using VLMs. Once an error occurs, there is no feedback loop to correct it, causing mistakes to compound throughout the process. + +Motivated by these shortcomings, we propose a visual test-time scaling framework, RegionFocus, designed to narrow the GUI model's attention to salient interface regions when execution errors occur or other conditions are trig- + +gered (e.g., VLM self-judgment). Specifically, as illustrated in Figure 1(1), we leverage the VLM's capability to identify points of interest and combine this with bounding-box proposals generated either from fixed-ratio masks or segmentation models such as SAM [16]. For each sub-region, the agent independently predicts actions based solely on the local context (Figure 1(2)), subsequently aggregating the top candidate actions to form a refined, single-step response. Furthermore, interactions with web or OS interfaces allow our method to zoom into targeted areas, enhancing the resolution of selected regions for more careful examination. RegionFocus works as a modular plug-in for GUI agents without affecting the original workflow. + +In order to keep track of regions visited by RegionFocus, we introduce an image-as-map mechanism to record temporal information. In this approach, elements previously considered by the agent are annotated in the UI screenshot with visual landmarks (e.g., the pink-stars in Figure 1 (3)). These landmarks prevent the agent from revisiting regions it has already examined and guide the agent towards unexplored areas. These markers do not interfere with the model's regular inference process which leverages unaltered webpage screenshots, and are only used in the RegionFocus component. Once the agent navigates to a new page, all landmarks are refreshed to have a new RegionFocus history. + +In addition to representing RegionFocus history, we also leverage image-as-map in the action aggregation process (Figure 1 (2)), where candidate actions are represented using landmarks for selecting an optimal action. We find image-as-map to be highly effective in representing both temporal information (e.g., previously visited regions) and spatial information (e.g., multiple action candidates) compared to alternative representations such as element coordinates represented in the form of text alone. This is particularly crucial for distinguishing between screen elements in close proximity, which is challenging to reason about based on a text representation of the coordinates alone. + +With our proposed visual test-time scaling framework, we help existing models—UI-TARS [28] and QWen2.5-VL [3]—achieve better performance in both web-based and desktop interfaces. In particular, we demonstrate substantial performance gains on benchmarks including ScreenSpot-Pro [18] for OS-level GUI navigation, as well as WebVoyager [13] for browser automation. Through our experiments, we show that even a simple fixed-ratio bounding-box generation approach yields pronounced improvements over baseline systems, underscoring the efficacy of focusing the model's attention on visually relevant regions. Our empirical studies further indicate that image-as-map consistently outperforms text-based representations for VLM agents. Overall, our findings highlight the value of visual test-time scaling as a simple yet powerful extension to existing VLM-based GUI agents. + +# 2. Related Work + +# 2.1. GUI Agents + +Recent advancements in Large Language Models (LLMs) and Vision Language Models (VLMs) have significantly enhanced GUI automation, enabling agents to interact with diverse graphical environments through textual and visual modalities [8, 11, 15, 21, 24, 27, 31, 33, 34, 38, 40]. Prior studies generally adopt two distinct approaches: (1) text-based reasoning, which leverages structured representations such as HTML or accessibility trees [5, 17, 47], extracting structured interface information [22, 43] and supplementary textual details for input into LLMs/VLMs [46]; and (2) vision-based inference, relying on VLMs to directly interpret GUI elements [11, 28]. While text-based techniques efficiently handle structured information, they often struggle with visually complex or ambiguous interfaces [13, 18, 36], resulting in inaccuracies and reduced reliability. Similar observations have been reported in [39]. Conversely, visual grounding approaches may inadvertently interact with irrelevant or empty regions due to overly broad visual attention. In contrast to prior methods using entire interface screens as input [6, 14, 29], our approach explicitly separates planning from visual grounding via a novel visual test-time scaling framework. This framework selectively targets salient GUI regions through precise bounding-box proposals and integrates an innovative "image-as-map" strategy, maintaining contextual coherence throughout interactions. + +# 2.2. Test-Time Scaling in AI Agents + +Test-time scaling involves dynamically adjusting computational resources during inference to enhance model performance [32, 35, 41, 42]. This approach allows AI agents to allocate additional processing power to challenging tasks, thereby improving decision-making and accuracy. Inspired by advancements in test-time scaling for LLMs, several studies have extended similar principles to GUI agents. For example, prior work leverage intermediate action histories [45], collect external information during inference [25] and incorporate reflection mechanisms into AI agents [44] during inference. Despite their success in improving performance, these methods do not utilize the unique advantages of visual information. We propose an approach toward visual test-time scaling, dynamically adjusting the image focus region and employ an "image-as-map" technique to encode history information for effective GUI agent inference. + +# 2.3. Visual Image Attention + +Visual image attention mechanisms have a rich history of enabling AI models to selectively focus on pertinent regions within visual inputs [9, 12, 20, 37]. Such mechanisms are especially critical for GUI-based agents, where accurately identifying and interacting with interface elements amidst + +![](images/f351d336c42dba320d41f785475b7bad5b0622a1bedddd5d7ef25193e7f96495.jpg) +1) Agent makes initial prediction on full screenshot + +![](images/4d0328b33c1025ec7fd32b48fc90aab5e6771100ae97cb270409f2664e16b98c.jpg) +4) Examine region-based predictions and determine final action + +![](images/bba1ff39be1e5b7ead75e5b22a3c5dbe0f9b852c7e8b940629301c82cbcb91e4.jpg) +Figure 2. Overview of integrating RegionFocus into GUI agent pipelines. The standard inference process (blue arrow) takes input information and continuously predicts subsequent actions. When the GUI agent encounters errors, RegionFocus (green arrow) activates, proposing focal points to extract targeted sub-regions. Actions are then predicted individually for these sub-regions and aggregated into a single refined action for standard inference. + +visually cluttered environments is essential. Our proposed approach introduces a region-focused mechanism utilizing predefined bounding boxes that progressively refine their attention through history recording, incrementally concentrating focus more precisely on target elements. This iterative refinement shares conceptual similarities with prior recurrent models [4, 23]; however, our method leverages VLMs to achieve refinement [26], uniquely integrating this process directly into GUI agents' test-time scaling. Accurately generating these attention regions directly via VLMs remains an open direction for future research. + +# 3. Method + +To address the limitations of current GUI agent frameworks, we propose a visual test-time scaling approach that enhances the robustness and accuracy of VLM agents interacting with complex graphical user interfaces. Unlike traditional methods, which uniformly treat all interface elements, our method dynamically adjusts the model's focus by selectively emphasizing visually salient regions whenever potential errors are detected. This approach significantly reduces misclicks and navigation mistakes. + +Crucially, our framework operates entirely at inference time, enabling straightforward integration into existing VLM agents without requiring retraining or architectural modifications. In the following sections, we first de + +scribe the integration process for our pipeline with existing GUI agents, followed by a detailed explanation of the design principles and implementation of each component. + +# 3.1. Overview + +Figure 2 illustrates our proposed pipeline. On top of the standard interaction pipeline, our approach introduces an error-triggered refinement mechanism. Specifically, when the agent encounters a trigger condition—such as clicking on non-interactive elements (e.g., selecting empty space instead of the intended date option) or repeating unsuccessful actions (e.g., repeatedly typing “Los Angeles” without successfully clicking the correct button)—the RegionFocus component is activated. + +During this refinement stage, the agent initially predicts a focal point near the intended target element. Based on this approximate location, it generates a bounding box likely to encapsulate the target element (Section 3.2). For each region defined by the bounding boxes, the agent independently predicts candidate actions. Finally, the agent aggregates a single action to be executed based on the predictions for each region. + +Additionally, we maintain an annotated history of previously examined focal points on the same UI image using an image-as-map representation (Section 3.3). This visual history guides the agent in avoiding redundant searches and + +helps it progressively focus on the correct target. Interactive predictions involving specific coordinates are visually marked (e.g., using pink-star landmarks) to help the model verify the correctness of its selections. + +# 3.2. Visual Region Focus + +The core of our pipeline is a dynamic visual zoom-in mechanism activated during inference. When initial action predictions result in errors—such as clicks on non-interactive or empty regions—the model dynamically adjusts its visual attention. Specifically, it refines attention by generating bounding-box proposals around visually salient regions, leading to more precise single-step actions. + +Trigger Condition We define two primary types of triggering conditions. The first relies on environmental feedback obtained from direct interaction with dynamic GUI environments, such as interactive webpages. Errors, like clicks on non-interactive elements, can be easily detected based on environment feedback (e.g., webpage change). The second type of trigger occurs in cases where environmental interaction isn't possible and we are dealing with static screenshots (e.g., ScreenSpot-Pro [19] scenario). Here, the VLM evaluates predicted actions, shifting its role from merely making predictions to explicitly evaluating the correctness of actions. Although the VLM's judgments aren't always perfectly accurate, they significantly help identify and mitigate errors shown in our experiments. + +Bounding-box Proposal Empirically, VLM agents reliably produce focal points near target elements but struggle to directly predict accurate bounding boxes [10]. Instead, we derive bounding boxes from these focal points rather than predicting them directly. Although advanced segmentation models (e.g., SAM [16]) could provide more precise bounding boxes, we currently focus exclusively on leveraging the GUI agent itself. This ensures that future enhancements in GUI agents directly benefit our approach. + +We employ a heuristic approach, defining bounding boxes with fixed dimensions and aspect ratios (e.g., 0.5 width by 0.5 height) centered around the focal point. If a bounding box exceeds the image boundary, it is adjusted to remain fully within the screenshot. This simple yet effective strategy strikes a good balance between accuracy and computational efficiency. + +Action Candidate Prediction Given the regions identified by the bounding boxes extracted from the previous stage, the agent predicts an action for each region. If the agent can interact with the environment (e.g., in the case of a webpage), a zoomed-in, high resolution view of the region is provided to the agent. In this case, at least one side of the region can be made to match the original full image + +![](images/137217ede51f9f386c78acae20c50f0b5af39d1fd5404b3697cf6d750637c07f.jpg) +Figure 3. Image-as-Map records temporal information. We place numbered pink stars in the image as visual landmarks to indicate previously focused points. Each time a RegionFocus attempt fails (i.e., action takes no effect), we add a new pink star at the attempted location. Once an action successfully takes effect, we refresh the history and remove any existing landmarks. + +resolution via zooming in. If environment interaction is not possible, we simply crop the region from the initial image and upsample it for prediction. + +Action Aggregation After the GUI agent independently analyzes each bounding-box proposal and generates candidate actions, we prompt the model to select a single action from all candidates based on the overall task objective, serving as the next step in the inference pipeline. For coordinate-based actions (e.g., "Click (x, y)" or "Scroll (x, y) down"), we visually mark the candidate action coordinates on the snapshot along with their action numbers, as illustrated in Figure 1 (2). This process significantly reduces the model's workload by simplifying the mapping and interpretation of textual coordinates onto the image. Empirically, we observe that incorporating these visual markers leads the model to select action candidates more accurately. + +# 3.3. Image-as-Map + +When RegionFocus is triggered multiple times for a given UI image, it is important for the agent to generate diverse focal points and avoid revisiting previously explored regions. Initially, we attempted to represent the region focus history using textual coordinates but found this approach ineffective, as the agent often revisited similar focal points despite explicit prompts to avoid them. + +To address this, we propose an image-as-map representation, where we visually encode previously examined focal points as pink-star landmarks annotated with historical indices directly onto UI snapshots, as shown in Figure 3 & Figure 1 (3), to reflect temporal information. Unlike text-based histories, this visual representation more effectively conveys temporal information, allowing the agent to reason directly over the image and avoid revisiting the same areas. Note that the visual landmark annotation process used for action aggregation (Figure 1 (2)) is also a form of the + +
Agent ModelDevelopmentCreativeCADScientificOfficeOSAvg
texticonavgtexticonavgtexticonavgtexticonavgtexticonavgtexticonavgtexticonavg
QwenVL-7B0.00.00.00.00.00.00.00.00.00.70.00.40.00.00.00.00.00.00.10.00.1
GPT-4o1.30.00.71.00.00.62.00.01.52.10.01.21.10.00.60.00.01.30.00.8
SeeClick0.60.00.31.00.00.62.50.01.93.50.02.01.10.00.52.80.01.51.80.01.1
Qwen2-VL-7B2.60.01.31.50.00.90.50.00.46.30.03.53.41.93.00.90.00.52.50.21.6
OS-Atlas-4B7.10.03.73.01.42.32.00.01.59.05.57.55.13.84.45.60.03.15.01.73.7
ShowUI-2B16.91.49.49.10.05.32.50.01.913.27.310.615.37.513.510.32.26.610.82.67.7
CogAgent-18B14.90.78.09.60.05.67.13.16.122.21.813.413.00.06.55.60.03.112.00.87.7
Aria-UI16.20.08.423.72.114.77.61.66.127.16.418.120.31.916.14.70.02.617.12.011.3
UGround-7B26.62.114.727.32.817.014.21.611.131.92.719.331.611.327.917.80.09.725.02.816.5
Claude Comp.Use22.03.912.625.93.416.814.53.711.933.915.825.830.116.326.211.04.58.123.47.117.1
OS-Atlas-7B33.11.417.728.82.817.912.24.710.337.57.324.433.95.727.427.14.516.828.14.018.9
UGround-V1-7B--35.5--27.8--13.5--38.8--48.8--26.1--31.1
UI-TARS-7B58.412.436.150.09.132.820.89.418.063.931.850.063.320.853.530.816.924.547.816.235.7
+RegionFocus59.715.938.559.611.939.630.57.824.967.430.051.269.530.260.445.821.334.755.218.741.2
UI-TARS-72B63.017.340.857.115.439.618.812.517.264.620.945.763.326.454.842.115.730.150.917.538.1
+RegionFocus72.126.950.268.722.449.335.525.033.077.130.957.172.945.366.563.627.046.964.028.050.2
Qwen2.5-VL-7B45.51.424.132.86.321.722.36.218.450.77.331.952.515.143.936.410.124.539.36.626.8
+RegionFocus53.23.429.142.94.927.028.43.122.256.910.937.059.924.551.741.115.729.646.68.832.1
Qwen2.5-VL-72B66.213.840.864.615.444.047.712.539.178.529.157.174.637.766.160.722.543.464.920.247.8
+RegionFocus75.325.551.276.330.857.271.628.160.987.539.166.587.060.480.974.836.057.178.634.161.6
+ +Table 1. Comparison of various models on ScreenSpot-Pro. [19]. Our proposed RegionFocus helps the UI-TARS-72B [28] model achieve a $31.8\%$ improvement, while Qwen2.5-VL-72B [3] sees a $28.9\%$ gain, thereby achieving state-of-the-art performance. Additionally, integrating RegionFocus into UI-TARS-7B allows it to surpass the performance of the substantially larger UI-TARS-72B model. + +image-as-map strategy, but capturing spatial information. Landmark-annotated snapshots are used only in the Region-Focus process, whereas the original inference pipeline (e.g., for the initial action prediction) receives unaltered UI images. We maintain these highlighted landmarks on the page until an action takes effect (i.e., causes a meaningful state change), at which point the history is refreshed. Our empirical results show that this image-as-map strategy consistently outperforms text-based methods, especially in more complex, multi-step GUI tasks. We also observe that it helps the agent distinguish between two GUI elements that are very close to each other, as shown in Figure 7 (2). + +# 4. Experiments + +In the experimental sections below, we integrate our proposed pipeline with UI-TARS [28] and Qwen2.5-VL [3], two recently proposed GUI agent models that autonomously interacts with GUI screenshots, which have demonstrated exceptional performance in various GUI tasks. We evaluate our pipeline across both OS operation tasks and Web Interaction. For our main experiments, we adopt a fixed bounding box approach (Section 3.2), using only the agent model itself. We also include ablation studies using SAM [16], which provides more accurate bounding boxes. Additional experimental details are provided in Appendix B. + +# 4.1. ScreenSpot-Pro + +ScreenSpot-Pro [19] is a benchmark designed to evaluate GUI grounding in complex, high-resolution desktop envi + +ronments (typically larger than $3k \times 2k$ ). It includes expert-annotated tasks across 23 applications spanning five functional domains (Development, Creative, CAD, Scientific, Office) and three operating systems. Tasks are categorized into text-based and icon-based grounding challenges. Grounding accuracy is measured by verifying if the model-predicted location falls within the bounding box of the target element. + +Because ScreenSpot-Pro only provides static screenshots—without an OS environment for interaction—we employ a VLM-based judge to trigger RegionFocus. Specifically, when the agent predicts a point, we will highlight that point in the input screenshot with pink-star landmarks and ask the model itself to assess the correctness of that point. If deemed incorrect, we initiate the RegionFocus process. + +Results In Table 1, we summarize the reported grounding accuracy of various methods evaluated on ScreenSpot-Pro. For fair comparisons, we employ the official test code released by ScreenSpot-Pro for evaluation. We report the original numbers from the UI-TARS paper in Table 1. + +Table 1 shows that RegionFocus consistently improves performance across all categories for both text and icon grounding compared to the base model. UI-TARS-72B + RegionFocus achieves a $31.7\%$ improvement over the base UI-TARS-72B model. Moreover, the UI-TARS-7B + RegionFocus variant outperforms the UI-TARS-72B model overall, demonstrating the effectiveness of our approach. Furthermore, RegionFocus further helps Qwen2.5-VL-72B + +
Agent ModelAllrecipesAmazonAppleArXivGitHubBookingESPNCoursera
Claude45.9%±3.4%58.6%±4.2%58.1%±4.0%55.0%±7.0%56.9%±1.4%19.0%±1.3%46.2%±1.3%68.2%±1.3%
GPT-4o56.3%±1.3%53.7%±2.5%56.6%±1.3%60.5%±0.0%57.7%±3.7%43.9%±3.5%44.0%±2.7%65.1%±2.8%
WebVoyager51.1%±2.2%52.9%±1.4%62.8%±2.3%52.0%±1.3%59.3%±3.7%32.6%±2.7%47.0%±1.3%57.9%±2.7%
Qwen2.5-VL-7B47.2%±4.4%49.1%±7.5%47.3%±2.5%14.9%±2.1%23.9%±4.8%10.0%±0.9%39.2%±7.2%46.4%±0.3%
+ RegionFocus49.2%±7.2%53.6%±4.4%67.1%±3.6%51.7%±1.3%35.0%±0.3%30.0%±0.0%39.6%±6.4%71.1%±4.8%
Qwen2.5-VL-72B28.6%±1.5%56.2%±4.3%53.0%±0.9%32.6%±2.4%61.5%±4.3%24.8%±2.9%49.8%±3.8%72.9%±1.9%
+ RegionFocus43.4%±1.3%60.6%±1.5%69.7%±4.0%45.4%±2.0%67.3%±3.8%37.3%±2.9%59.8%±1.4%78.2%±1.4%
UI-TARS-7B17.8%±1.3%30.9%±1.4%17.1%±1.3%20.9%±2.3%32.5%±2.5%7.6%±1.3%45.0%±1.3%70.7%±1.2%
+ RegionFocus35.6%±2.2%39.0%±2.4%31.8%±2.7%37.2%±0.0%58.1%±1.5%15.2%±1.3%62.8%±0.0%74.2%±1.4%
UI-TARS-72B16.2%±2.6%38.6%±2.0%45.9%±1.2%64.9%±4.1%43.2%±2.0%40.2%±2.3%47.2%±3.8%58.9%±1.6%
+ RegionFocus57.5%±1.1%55.8%±1.4%52.6%±1.3%50.3%±3.6%58.9%±3.7%44.7%±2.6%69.7%±1.3%82.2%±1.5%
Agent ModelCambridge DictionaryBBC NewsGoogle FlightsGoogle MapGoogle SearchHuggingfaceWolfram AlphaOverall
Claude71.3%±3.6%66.7%±4.8%15.1%±5.5%55.3%±1.4%72.9%±1.3%53.5%±4.7%51.5%±5.4%52.8%±1.4%
GPT-4o82.2%±1.3%54.8%±2.4%28.6%±0.0%56.9%±2.8%63.6%±1.3%42.6%±3.6%65.2%±2.2%55.5%±0.8%
WebVoyager71.3%±1.3%60.3%±2.8%51.6%±1.4%64.3%±2.8%77.5%±2.7%55.8%±2.3%60.9%±2.2%57.1%±0.2%
Qwen2.5-VL-7B21.1%±4.8%45.1%±3.0%10.0%±0.9%30.2%±1.5%10.0%±1.3%41.4%±0.3%51.3%±3.8%32.5%±1.3%
+ RegionFocus17.3%±0.3%52.9%±0.9%12.8%±4.8%17.1%±1.2%18.3%±2.4%60.0%±2.8%40.8%±1.2%41.1%±1.2%
Qwen2.5-VL-72B63.7%±1.3%45.9%±1.6%17.7%±1.5%31.2%±1.5%11.5%±0.0%38.3%±2.9%48.9%±2.5%42.4%±0.5%
+ RegionFocus68.9%±1.3%54.4%±2.9%34.6%±4.4%42.2%±1.5%20.3%±0.0%51.5%±4.9%56.0%±1.1%52.7%±1.1%
UI-TARS-7B57.1%±1.3%41.3%±2.7%10.3%±1.4%17.5%±0.0%49.6%±1.3%40.7%±1.4%38.4%±1.3%33.2%±0.5%
+ RegionFocus55.8%±0.0%52.4%±2.4%28.6%±4.8%29.3%±2.4%60.5%±2.3%45.7%±1.3%44.2%±1.3%44.7%±0.5%
UI-TARS-72B72.9%±1.4%39.3%±3.4%33.9%±2.1%27.2%±5.2%60.6%±1.2%24.9%±1.6%48.3%±2.3%44.1%±0.5%
+ RegionFocus75.9%±1.3%49.8%±3.6%49.9%±2.1%59.9%±1.4%65.8%±1.2%59.6%±1.3%60.2%±1.3%59.5%±0.1%
+ +Table 2. Comparison of various models on WebVoyager [13]. For each automatic evaluation, we run GPT evaluator three times to calculate the performance mean and standard deviation. We evaluated our method on UI-TARS [28] and Qwen2.5-VL [3], consistently observing performance improvements. The performance on Google Search was significantly affected by bot blocking - See text for details. + +![](images/ac5c8cd3431b1cbea85eab59534ec24696da7a855726c0625be34740eb429170.jpg) +Figure 4. Qualitative results - Screenshot-Pro. In this example, the agent successfully rejects the initial action via self-VLM-judge and proposes a correct grounding point based on the zoomed-in view. More qualitative results are listed in Appendix B. + +achieve the state-of-the-art performance of $61.6\%$ . + +Figure 4 illustrates our pipeline for ScreenSpot-Pro. We first ask the agent to judge the initial action prediction. If it is incorrect, we initiate RegionFocus by zooming into the regions predicted by the agent, predicting actions within each region, and finally aggregating all actions into a sin + +gle outcome. We used the same VLM to both predict the action and evaluate its correctness, and found that although it may initially generate incorrect coordinates, it can still reliably assess whether the click points are correct with the help of image-as-map—enabling enabling those predictions to be corrected by RegionFocus through prediction over the zoomed-in images. This generation-verification gap has also been noted in recent literature [7, 30]. + +# 4.2. WebVoyager + +WebVoyager [13] is a benchmark designed to evaluate autonomous web agents' capabilities in performing open-ended tasks through multimodal interactions with real-world websites. Distinct from previous web agent benchmarks, WebVoyager comprises 643 tasks across 15 popular, real-world websites such as Amazon, Apple, ArXiv, and Google Maps. In this scenario, the agent actively interacts with the web environment. We employ a Playwright-controlled Chrome browser to navigate webpages, with the VLM agent determining the appropriate action based on each webpage screenshot. After task execution, we use the official evaluation setting, where a GPT-based judge reviews the last 15 screenshots along with an optional textual response to determine whether the task has been successfully accomplished. + +![](images/ad159f6cb41916fcc8f5bb68b79e285045778044b5c414723855c79b88605345.jpg) +Task: Check the Dataset Viewer for ai2lumos/lumos_complex_qa_plan_onetime on Hugging face. What is the content corresponding to user in the first message? + +![](images/089d3199b5678427e8142adfa19de40b00c6790ba5844d5a8e53d015c9919952.jpg) +Task: Search the latest article about space exploration on BBC News and summarize its key points. + +Results Table 2 presents the task success rates of various web agents evaluated on the WebVoyager benchmark. RegionFocus consistently improves performance across all types of websites—including “Booking” and “Search”—highlighting the effectiveness of integrating RegionFocus into the GUI agent for web browsing. It also brings consistent improvements over two open-source model, UI-TARS and Qwen2.5-VL. We note that our model performance was impacted by online interaction constraints—such as bot blocking and intermittent VPN issues. For example, “Google-Search” performance dropped noticeably after Google recently tightened its bot detection systems and access controls; our UI-TARS results were obtained prior to this, whereas Qwen2.5-VL results were obtained afterward. While addressing these constraints can further improve the model's overall performance, bot-blocking and non-stationarity remain inherent challenges when evaluating agents on real-world websites. + +We present qualitative examples from WebVoyager in Figures 5 and 6. Figure 5 shows RegionFocus zooming in on the sub-region of interest, enlarging key content and simplifying target localization. Figure 6 demonstrates a correction scenario: after initially clicking an unrelated element, our pipeline proposes two closely positioned buttons, and the image-as-map mechanism helps distinguish between these nearly identical elements with slight coordinate differences. More analysis are included in Appendix A. + +# 4.3. Ablation Studies & Analysis + +We conduct ablation studies on the entire pipeline, including our "image-as-map" design choice and the use of a predefined bounding box based on the point predicted by the agent. We also demonstrate that by leveraging SAM [16] and increasing the number of trajectory steps, performance + +![](images/46ed4f49891e966729ced60f2bf2443f44b428938aa9c552490d888ba13c9f5f.jpg) +Figure 5. RegionFocus. We illustrate how RegionFocus reduces background noise by emphasizing salient regions of an image. The mouse pointer indicates the agent's initial action prediction, which is suboptimal. The second image is zoomed in, reducing distracting details and allowing the agent to focus on the relevant information. + +![](images/8069ee00be2687cbfe7c5522716b2b5f20822130357bcc9f4a4b657630b6eedd.jpg) +Figure 6. Image-as-Map. We show that action aggregation, enhanced by the proposed image-as-map, effectively distinguishes subtle coordinate differences between elements. Star landmarks are generated during RegionFocus prior to aggregation; the two landmarks at the top left correspond to the home and search buttons. + +can be further improved. Due to limited computation, we employed the UI-TARS-7B-DPO model for these ablations on a subset of the WebVoyager benchmark. The results are shown in Table 3, where "image-as-map" and "Fixed-BBox" refers to our same 7B model configuration with RegionFocus enabled, a maximum limit of 100 action steps, and predefined bounding box with fixed aspect ratios. + +Text-based RegionFocus History Representation By using image-as-map, we can directly provide visual location information to the agent, helping it distinguish even minor differences and thereby enhancing its perception. For instance, as shown in Figure 7 (2), we color-code click points in the image to denote the image-as-map mechanism, while the corresponding coordinates are listed in the text box on the right. Notably, although the textual coordinate difference is within only five pixels, the resulting action can vary significantly. Using a purely text-based representation for both history tracking and action aggregation leads to notable degradation compared to our image-as-map approach (Text-as-History vs. Image-as-Map in Table 3). + +Proposing regions directly In our pipeline (Section 3), instead of directly predicting bounding boxes, we prompt the agent model to first identify a point of interest and then generate a predefined bounding box around it. This design choice addresses the agent's difficulty in accurately predicting bounding boxes. To validate this, we conducted an experiment requiring the model to predict both upper-left and bottom-right coordinates for cropping the UI image. As indicated by the "Predict-Region" results, this approach significantly reduced performance. + +![](images/a878f91d699ffab4ef95844b673bf06887aa9359015709e175b451837d69d7cf.jpg) +1) SAM provides Bounding Box + +![](images/7a7e2146429fde3090381fea1fa5591fe473cb547f0711f9dd641e852007b72c.jpg) +2) Image-as-Map tells small coordinates difference +If recoding in text solely: +{Instruction} +Select one action point +from the blew options +- Click-Point 1: (671, 291) +- Click-Point 2: (671, 295) + +SAM Our pipeline can naturally leverage segmentation models that take point inputs as indicator, such as SAM [16]. As shown in Figure 7 (1), even when the model provides a point in a non-interactive empty area, SAM successfully generates a bounding box covering the correct region—demonstrating its suitability. The effectiveness of incorporating point-based segmentation models into Region-Focus is further validated by the "SAM" results in Table 3. + +Test-time Thinking Budget In our main experiments, we limited the maximum number of inference steps to 100. We investigate how performance changes when varying the inference-step limit above and below the default of 100 steps. Here, inference steps refer to actual browser-interactive actions, excluding RegionFocus overhead. We evaluated limits of 10, 50, and 300 steps, using the UI-TARS-7B model. As shown in Figure 8, raising the limit to 300 steps improved performance from 43.2 to 45.3, although most trajectories terminated before reaching this threshold. However, incremental benefits gradually diminish with higher inference-step limits. + +Runtime Analysis For each trigger, the VLM performs one inference for focal-point proposal, four inferences for region action prediction (which can be executed in parallel), and one inference for action aggregation. The four region action prediction inferences result from adopting four predefined bounding boxes around each predicted focal point, a hyper-parameter. For the interactive environment (WebVoyager), RegionFocus occurs in $61.7\%$ of all trajectories; among these, it is triggered an average of 5.84 times, incurring an average overhead of $66.8\%$ per trajectory. Notably, $32.3\%$ of triggered cases occur only once, with this single trigger yielding an $83.7\%$ improvement in success. For the static environment (ScreenSpot-Pro), RegionFocus occurs in $60.2\%$ and $33\%$ of cases for the 72B and 7B models, respectively. The 72B model incurs overheads of $180\%$ when region predictions are executed in parallel, and $360\%$ if executed sequentially. Regarding memory usage, RegionFocus requires no additional GPU memory if region predictions are performed sequentially without employing SAM [16]. + +![](images/922ccb757897767d2e901c630498fa0f2918350cc9dec26a5ffb1a12bf85f0ae.jpg) +Figure 7. Ablation studies. (1) RegionFocus can natively use the SAM to generate bounding boxes. (2) how image-as-map helps highlight subtle differences of different GUI elements. + +![](images/a1465c0bad41da16f8ddc424a64ee5fe01f9d19811514b339aa002e4c83ea5b4.jpg) +Table 3. Ablation results on WebVoyager. Higher scores indicate better performance. +Figure 8. Inference-step limits. We observe improvements as the inference-step limit increases. + +Whether region action predictions should be executed sequentially or in parallel, the number of regions to zoom into, and whether to utilize SAM depend on the specific use case. + +Failure Case Analysis According to our observations, we summarize the main causes of failures as follows: (1) *Element not visible*: In cases where the correct element is not visible on the current page (e.g., requires scrolling actions), zooming in on the current page is not beneficial. (2) *Non-interactable elements*: Some webpages contain extensive text that appears relevant but is not clickable, causing *RegionFocus* to zoom into these regions and subsequently fail to proceed due to the lack of interactable elements. For instance, on the Hugging Face homepage (https://huggingface.co/), there is abundant relevant text such as "Image-to-Text" and "Sentence Similarity," which relates to agent tasks but is non-clickable. (3) *Wrong focal points*: In some rare cases, *RegionFocus* proposes the wrong focal point, directing attention to irrelevant regions. (4) *Action prediction failure*: Occasionally, despite identifying the correct focal point, all regional action predictions fail; our observations suggest that employing SAM [14] could alleviate this issue as it further reduces the background clutter. (5) *Action aggregation failure*: Among the regional actions, the selected one may not be optimal. (6) *Others*: Other errors arise from general reasoning failures, bot detection, or reaching the maximum step limit. + +# 5. Conclusion + +We introduced RegionFocus, a visual test-time scaling approach that dynamically zooms in on relevant interface regions to address the clutter and ambiguity of modern GUIs. By integrating an image-as-map mechanism that marks key landmarks, our method provides transparent action records and improves coordinate-based action predictions. Experiments on Screenspot-pro and WebVoyager show substantial performance gains—even with a simple fixed-ratio bounding box strategy—highlighting the power of visual test-time scaling in enhancing interactive AI systems. + +# 6. Acknowledgment + +# References + +[1] Peter Anderson, Qi Wu, Damien Teney, Jeffrey Bruce, Mark Johnson, Stephen Gould, and Anton van den Hengel. Vision-and-language navigation: Interpreting visually-grounded navigation instructions in real environments. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 3674-3683, 2018. 1 +[2] Anthropic. Developing a computer use model. https://www.anthropic.com/news/developing-computer-use, 2024. Accessed: 2025-03-02.1 +[3] Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, Humen Zhong, Yuanzhi Zhu, Mingkun Yang, Zhaohai Li, Jianqiang Wan, Pengfei Wang, Wei Ding, Zheren Fu, Yiheng Xu, Jiabo Ye, Xi Zhang, Tianbao Xie, Zesen Cheng, Hang Zhang, Zhibo Yang, Haiyang Xu, and Junyang Lin. Qwen2.5-v1 technical report. arXiv preprint arXiv:2502.13923, 2025. 2, 5, 6 +[4] Zhaowei Cai and Nuno Vasconcelos. Cascade r-cnn: Delving into high quality object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6154-6162, 2018. 3 +[5] Ruisheng Cao, Fangyu Lei, Haoyuan Wu, Jixuan Chen, Yeqiao Fu, Hongcheng Gao, Xinzhuang Xiong, Hanchong Zhang, Wenjing Hu, Yuchen Mao, et al. Spider2-v: How far are multimodal agents from automating data science and engineering workflows? Advances in Neural Information Processing Systems, 37:107703-107744, 2025. 2 +[6] Kanzhi Cheng, Qiushi Sun, Yougang Chu, Fangzhi Xu, Yantao Li, Jianbing Zhang, and Zhiyong Wu. Seeclick: Harnessing gui grounding for advanced visual gui agents. arXiv preprint arXiv:2401.10935, 2024. 2 +[7] Stephen A Cook. The complexity of theorem-proving procedures. In Logic, automata, and computational complexity: The works of Stephen A. Cook, pages 143-152. 2023. 6 +[8] Xiang Deng, Yu Gu, Boyuan Zheng, Shijie Chen, Sam Stevens, Boshi Wang, Huan Sun, and Yu Su. Mind2web: Towards a generalist agent for the web. Advances in Neural Information Processing Systems, 36:28091-28114, 2023. 2 +[9] António Farinhas, André F. T. Martins, and Pedro M. Q. Aguiar. Multimodal continuous visual attention mechanisms, 2021. 2 +[10] Yongchao Feng, Yajie Liu, Shuai Yang, Wenrui Cai, Jinqing Zhang, Qiqi Zhan, Ziyue Huang, Hongxi Yan, Qiao Wan, Chenguang Liu, et al. Vision-language model for object detection and segmentation: A review and evaluation. arXiv preprint arXiv:2504.09480, 2025. 4 + +The work was completed during Tiange's part-time internship at LG AI Research and was supported by an LG AI Research grant awarded through University of Michigan. We thank Jaekyeom Kim and Sungryull Sohn for their discussions and contributions to the codebase. +[11] Boyu Gou, Ruohan Wang, Boyuan Zheng, Yanan Xie, Cheng Chang, Yiheng Shu, Huan Sun, and Yu Su. Navigating the digital world as humans do: Universal visual grounding for gui agents. arXiv preprint arXiv:2410.05243, 2024. 1, 2 +[12] Meng-Hao Guo, Tian-Xing Xu, Jiang-Jiang Liu, Zheng-Ning Liu, Peng-Tao Jiang, Tai-Jiang Mu, Song-Hai Zhang, Ralph R Martin, Ming-Ming Cheng, and Shi-Min Hu. Attention mechanisms in computer vision: A survey. Computational visual media, 8(3):331-368, 2022. 2 +[13] Hongliang He, Wenlin Yao, Kaixin Ma, Wenhao Yu, Yong Dai, Hongming Zhang, Zhenzhong Lan, and Dong Yu. Webvoyager: Building an end-to-end web agent with large multimodal models. arXiv preprint arXiv:2401.13919, 2024. 2, 6 +[14] Wenyi Hong, Weihan Wang, Qingsong Lv, Jiazheng Xu, Wenmeng Yu, Junhui Ji, Yan Wang, Zihan Wang, Yuxiao Dong, Ming Ding, et al. Cogagent: A visual language model for gui agents. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14281-14290, 2024. 2 +[15] Geunwoo Kim, Pierre Baldi, and Stephen McAleer. Language models can solve computer tasks. Advances in Neural Information Processing Systems, 36:39648-39677, 2023. 2 +[16] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hans Marx, Carl Rolland, Laura Gustafson, Tsung-Yi Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dólar, and Ross Girshick. Segment anything. arXiv preprint arXiv:2304.02643, 2023. 2, 4, 5, 7, 8 +[17] Jing Yu Koh, Robert Lo, Lawrence Jang, Vikram Duvvur, Ming Chong Lim, Po-Yu Huang, Graham Neubig, Shuyan Zhou, Ruslan Salakhutdinov, and Daniel Fried. Visualwebarena: Evaluating multimodal agents on realistic visual web tasks. arXiv preprint arXiv:2401.13649, 2024. 2 +[18] Kaixin Li, Ziyang Meng, Hongzhan Lin, Ziyang Luo, Yuchen Tian, Jing Ma, Zhiyong Huang, and Tat-Seng Chua. Screenshot-pro: Gui grounding for professional high-resolution computer use, 2025. 2 +[19] Kaixin Li, Ziyang Meng, Hongzhan Lin, Ziyang Luo, Yuchen Tian, Jing Ma, Zhiyong Huang, and Tat-Seng Chua. Screenshot-pro: Gui grounding for professional high-resolution computer use. arXiv preprint arXiv:2504.07981, 2025.4.5 +[20] Drew Linsley, Dan Shiebler, Sven Eberhardt, and Thomas Serre. Learning what and where to attend. arXiv preprint arXiv:1805.08819, 2018. 2 +[21] Xiao Liu, Tianjie Zhang, Yu Gu, Iat Long Iong, Yifan Xu, Xixuan Song, Shudan Zhang, Hanyu Lai, Xinyi Liu, Hanlin Zhao, et al. Visualagentbench: Towards large multimodal models as visual foundation agents. arXiv preprint arXiv:2408.06327, 2024. 1, 2 +[22] Yadong Lu, Jianwei Yang, Yelong Shen, and Ahmed Awadallah. Omniparser for pure vision based gui agent, 2024. 1, 2 +[23] Volodymyr Mnih, Nicolas Heess, Alex Graves, et al. Recurrent models of visual attention. Advances in neural information processing systems, 27, 2014. 3 +[24] Shikhar Murty, Hao Zhu, Dzmitry Bahdanau, and Christopher D. Manning. Nnetnav: Unsupervised learning of + +browser agents through environment interaction in the wild, 2025. 2 +[25] Reiichiro Nakano, Jacob Hilton, Suchir Balaji, Jeff Wu, Long Ouyang, Christina Kim, Christopher Hesse, Shantanu Jain, Vineet Kosaraju, William Saunders, et al. Webgpt: Browser-assisted question-answering with human feedback. arXiv preprint arXiv:2112.09332, 2021. 2 +[26] Soroush Nasiriany, Fei Xia, Wenhao Yu, Ted Xiao, Jacky Liang, Ishita Dasgupta, Annie Xie, Danny Driess, Ayzaan Wahid, Zhuo Xu, et al. Pivot: Iterative visual prompting elicits actionable knowledge for vlms. arXiv preprint arXiv:2402.07872, 2024. 3 +[27] Vardaan Pahuja, Yadong Lu, Corby Rosset, Boyu Gou, Arindam Mitra, Spencer Whitehead, Yu Su, and Ahmed Awadallah. Explorer: Scaling exploration-driven web trajectory synthesis for multimodal web agents, 2025. 2 +[28] Yujia Qin, Yining Ye, Junjie Fang, Haoming Wang, Shihao Liang, Shizuo Tian, Junda Zhang, Jiahao Li, Yunxin Li, Shijue Huang, et al. Ui-tars: Pioneering automated gui interaction with native agents. arXiv preprint arXiv:2501.12326, 2025. 1, 2, 5, 6 +[29] Peter Shaw, Mandar Joshi, James Cohan, Jonathan Berant, Panupong Pasupat, Hexiang Hu, Urvashi Khandelwal, Kenton Lee, and Kristina N Toutanova. From pixels to ui actions: Learning to follow instructions via graphical user interfaces. Advances in Neural Information Processing Systems, 36:34354-34370, 2023. 2 +[30]Gokul Swamy, Sanjiban Choudhury, Wen Sun, Zhiwei Steven Wu, and J Andrew Bagnell. All roads lead to likelihood: The value of reinforcement learning in fine-tuning. arXiv preprint arXiv:2503.01067, 2025. 6 +[31] Brandon Trabucco, Gunnar Sigurdsson, Robinson Piramuthu, and Ruslan Salakhutdinov. Towards internet-scale training for agents, 2025. 2 +[32] Hieu Tran Bao, Nguyen Cong Dat, Nguyen Duc Anh, and Hoang Thanh-Tung. Learning to stop overthinking at test time. arXiv e-prints, pages arXiv-2502, 2025. 2 +[33] Junyang Wang, Haiyang Xu, Haitao Jia, Xi Zhang, Ming Yan, Weizhou Shen, Ji Zhang, Fei Huang, and Jitao Sang. Mobile-agent-v2: Mobile device operation assistant with effective navigation via multi-agent collaboration. In Advances in Neural Information Processing Systems (NeurIPS), 2024. 2 +[34] Junyang Wang, Haiyang Xu, Jiabo Ye, Ming Yan, Weizhou Shen, Ji Zhang, Fei Huang, and Jitao Sang. Mobile-agent: Autonomous multi-modal mobile device agent with visual perception. arXiv preprint arXiv:2401.16158, 2024. 2 +[35] Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc Le, Ed Chi, Sharan Narang, Aakanksha Chowdhery, and Denny Zhou. Self-consistency improves chain of thought reasoning in language models. arXiv preprint arXiv:2203.11171, 2022. 2 +[36] Tianbao Xie, Danyang Zhang, Jixuan Chen, Xiaochuan Li, Siheng Zhao, Ruisheng Cao, Jing Hua Toh, Zhoujun Cheng, Dongchan Shin, Fangyu Lei, et al. Osworld: Benchmarking multimodal agents for open-ended tasks in real computer environments. Advances in Neural Information Processing Systems, 37:52040-52094, 2025. 2 + +[37] Kelvin Xu, Jimmy Ba, Ryan Kiros, Kyunghyun Cho, Aaron Courville, Ruslan Salakhudinov, Rich Zemel, and Yoshua Bengio. Show, attend and tell: Neural image caption generation with visual attention. In International conference on machine learning, pages 2048-2057. PMLR, 2015. 2 +[38] An Yan, Zhengyuan Yang, Wanrong Zhu, Kevin Lin, Linjie Li, Jianfeng Wang, Jianwei Yang, Yiwu Zhong, Julian McAuley, Jianfeng Gao, et al. Gpt-4v in wonderland: Large multimodal models for zero-shot smartphone gui navigation. arXiv preprint arXiv:2311.07562, 2023. 2 +[39] Ke Yang, Yao Liu, Sapana Chaudhary, Rasool Fakoor, Pratik Chaudhari, George Karypis, and Huzefa Rangwala. Agentoccam: A simple yet strong baseline for llm-based web agents, 2024. 1, 2 +[40] Shunyu Yao, Howard Chen, John Yang, and Karthik Narasimhan. Webshop: Towards scalable real-world web interaction with grounded language agents. Advances in Neural Information Processing Systems, 35:20744-20757, 2022. 2 +[41] Shunyu Yao, Dian Yu, Jeffrey Zhao, Izhak Shafran, Tom Griffiths, Yuan Cao, and Karthik Narasimhan. Tree of thoughts: Deliberate problem solving with large language models. Advances in neural information processing systems, 36:11809-11822, 2023. 2 +[42] Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao. React: Synergizing reasoning and acting in language models. In International Conference on Learning Representations (ICLR), 2023. 2 +[43] Wenwen Yu, Zhibo Yang, Jianqiang Wan, Sibo Song, Jun Tang, Wenqing Cheng, Yuliang Liu, and Xiang Bai. Omniparser v2: Structured-points-of-thought for unified visual text parsing and its generality to multimodal large language models. arXiv preprint arXiv:2502.16161, 2025. 2 +[44] Xiao Yu, Baolin Peng, Vineeth Vajipey, Hao Cheng, Michel Galley, Jianfeng Gao, and Zhou Yu. Exact: Teaching ai agents to explore with reflective-mcts and exploratory learning. arXiv preprint arXiv:2410.02052, 2024. 2 +[45] Zhuosheng Zhang and Aston Zhang. You only look at screens: Multimodal chain-of-action agents. arXiv preprint arXiv:2309.11436, 2023. 2 +[46] Boyuan Zheng, Boyu Gou, Jihyung Kil, Huan Sun, and Yu Su. Gpt-4v (isdiction) is a generalist web agent, if grounded. arXiv preprint arXiv:2401.01614, 2024. 1, 2 +[47] Shuyan Zhou, Frank F Xu, Hao Zhu, Xuhui Zhou, Robert Lo, Abishek Sridhar, Xianyi Cheng, Yonatan Bisk, Daniel Fried, Uri Alon, et al. Webarena: A realistic web environment for building autonomous agents. ICLR, 2024. 2 \ No newline at end of file diff --git a/ICCV/2025/Visual Test-time Scaling for GUI Agent Grounding/images.zip b/ICCV/2025/Visual Test-time Scaling for GUI Agent Grounding/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..14bcfdbb8b1c005c259b0eb8b27dc4ef2e1b58aa --- /dev/null +++ b/ICCV/2025/Visual Test-time Scaling for GUI Agent Grounding/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c868887fae2d2e5c11be7ec03c3866d454083fb92ddd3b5bbbf263c542546dfe +size 828006 diff --git a/ICCV/2025/Visual Test-time Scaling for GUI Agent Grounding/layout.json b/ICCV/2025/Visual Test-time Scaling for GUI Agent Grounding/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..66f3dd84b56e2001a8676f7213ab7e18a475c875 --- /dev/null +++ b/ICCV/2025/Visual Test-time Scaling for GUI Agent Grounding/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fe8e7843b024c034a44128361f7f85ce5171757bd55a5c2a37136094072aea77 +size 302168 diff --git a/ICCV/2025/Visual Textualization for Image Prompted Object Detection/5ef81f8a-58e2-436d-920c-1c45226041f0_content_list.json b/ICCV/2025/Visual Textualization for Image Prompted Object Detection/5ef81f8a-58e2-436d-920c-1c45226041f0_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..7bf74bc647edd79a575f6353a9618e8fefb5cc50 --- /dev/null +++ b/ICCV/2025/Visual Textualization for Image Prompted Object Detection/5ef81f8a-58e2-436d-920c-1c45226041f0_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:93813f0c542edc9c9f1b07a4deeb46e66b377a91691e2140a5564a9aaebcab5b +size 89680 diff --git a/ICCV/2025/Visual Textualization for Image Prompted Object Detection/5ef81f8a-58e2-436d-920c-1c45226041f0_model.json b/ICCV/2025/Visual Textualization for Image Prompted Object Detection/5ef81f8a-58e2-436d-920c-1c45226041f0_model.json new file mode 100644 index 0000000000000000000000000000000000000000..8fca262a5a318f20b99ca18d206e60990e323e6c --- /dev/null +++ b/ICCV/2025/Visual Textualization for Image Prompted Object Detection/5ef81f8a-58e2-436d-920c-1c45226041f0_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f85f6675f28a0f30b127ae7fce0261e849aec874f80a5050c9feb640eee80ffa +size 109310 diff --git a/ICCV/2025/Visual Textualization for Image Prompted Object Detection/5ef81f8a-58e2-436d-920c-1c45226041f0_origin.pdf b/ICCV/2025/Visual Textualization for Image Prompted Object Detection/5ef81f8a-58e2-436d-920c-1c45226041f0_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..9680f0c1f321ae6fe9af78200e5564187e295535 --- /dev/null +++ b/ICCV/2025/Visual Textualization for Image Prompted Object Detection/5ef81f8a-58e2-436d-920c-1c45226041f0_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:47590530df0846d156f235531fbf18eef9408e4a466986017d9e1a4570d3ecf5 +size 13236984 diff --git a/ICCV/2025/Visual Textualization for Image Prompted Object Detection/full.md b/ICCV/2025/Visual Textualization for Image Prompted Object Detection/full.md new file mode 100644 index 0000000000000000000000000000000000000000..2c520843a850b1f1ba451e05f76fdb41424a699c --- /dev/null +++ b/ICCV/2025/Visual Textualization for Image Prompted Object Detection/full.md @@ -0,0 +1,293 @@ +# Visual Textualization for Image Prompted Object Detection + +Yongjian Wu $^{1*}$ Yang Zhou $^{1*}$ Jiya Saiyin $^{1}$ Bingzheng Wei $^{2}$ Yan Xu $^{1\dagger}$ $^{1}$ School of Biological Science and Medical Engineering, Beihang University + $^{2}$ ByteDance Inc. + +{wuyongjian,zhouyangbme}@buaa.edu.cn xuyan04@gmail.com + +# Abstract + +We propose VisTex-OVLM, a novel image prompted object detection method that introduces visual textualization — a process that projects a few visual exemplars into the text feature space to enhance Object-level Vision-Language Models' (OVLMs) capability in detecting rare categories that are difficult to describe textually and nearly absent from their pretraining data, while preserving their pre-trained object-text alignment. Specifically, VisTex-OVLM leverages multi-scale textualizing blocks and a multi-stage fusion strategy to integrate visual information from visual exemplars, generating textualized visual tokens that effectively guide OVLMs alongside text prompts. Unlike previous methods, our method maintains the original architecture of OVLM, maintaining its generalization capabilities while enhancing performance in few-shot settings. VisTex-OVLM demonstrates superior performance across open-set datasets which have minimal overlap with OVLM's pre-training data and achieves state-of-the-art results on few-shot benchmarks PASCAL VOC and MSCOCO. The code will be released at VisTex-OVLM. + +# 1. Introduction + +Recently, vision-language models (VLMs) have demonstrated exceptional generalization in CV by pre-training on extensive image-text pairs, allowing zero-shot transfer via natural language prompts and inspiring new approaches for object detection [15, 17, 19, 30, 52, 58]. Among VLMs, object-level VLMs (OVLMs) [5, 31, 33], represented by GLIP [31], achieve better object-text alignment. OVLMs are pre-trained with extensive object detection and phrase grounding data, featuring multi-stage and multi-scale encoders to enhance object-text alignment beyond general VLMs. Furthermore, OVLMs commonly employ a multistage encoding structure based on cross-attention, allowing text prompts to actively guide object feature represen + +tation and location regression. These characteristics enable OVLMs to surpass conventional VLMs like CLIP [42] in zero-shot transfer for downstream object detection tasks. + +However, OVLM's zero-shot object detection (ZSOD) has intrinsic limitations. First, many objects in downstream tasks are underrepresented in the pre-training data, leading to suboptimal zero-shot transfer performance. Second, text prompts often lack sufficient description granularity, resulting in semantic bias and omissions. Third, fine-grained objects share similar descriptions in the text space, making it difficult to distinguish them using text alone [7, 53]. These challenges make introducing new visual information from downstream tasks essential for improving OVLM's transfer ability. One solution is to leverage visual exemplars for fine-tuning, which also refers to few-shot object detection (FSOD) [3, 17, 22, 40, 45]. For example, some VLM-based object detection methods incorporate few-shot visual exemplars information by fine-tuning original weights or adding processing structures. However, these approaches risk disrupting the OVLM's original object-text alignment and harm its generalization capabilities. + +To illustrate this, without loss of generality, we conducted an empirical analysis by transferring GLIP, pre-trained on Object365 [44], to MSCOCO [32] using various common transfer methods. Post-transfer, we computed the cosine similarity between paired text and object features from Object365 (Fig. 1). Results revealed that any modification in model structure or weights (methods 2-4) with limited transfer data skewed the object-text similarity distribution in the source domain, partially degrading alignment. This raises the question: can we introduce the semantics of a few visual exemplars while preventing the limited target training data from disrupting OVLM's object-text alignment? + +Using images as prompts, i.e., image prompting [36, 53], to provide semantic guidance alongside text prompts may address the question above. However, most OVLMs [5, 31, 33] are text-prompt-exclusive, lacking native support for image prompting due to architectural incompatibilities. MQ-Det [53] is the first to integrate visual exemplars into OVLM's prompting process by adding extra trainable cross-attention + +![](images/4d4b3869605874fe3577876d14771461fd6eb494290f68f82f15ad0c1af26295.jpg) +Figure 1. Frequency distribution of feature cosine similarity between object-text pairs from Object365 dataset after transferring the Object365-pretrained GLIP to MSCOCO with different methods. + +modules within the text encoder, enabling exemplar images to modulate the text prompts token-wisely and introduce the visual information of novel classes during inference. However, MQ-Det only re-weights tokens of existing text prompts without directly introducing new visual information, which may result in the loss of key clues. Moreover, the newly added cross-attention structures also bias OVLM's pre-trained object-text alignment, as shown in method 5, Fig. 1. How to enable image prompting for text-prompt-exclusive OVLMs without disrupting their pre-trained object-text alignment remains underexplored. + +This paper proposes VisTex-OVLM, a novel image prompting method for expanding OVLM's detection capacity beyond pre-trained categories. To preserve OVLM's object-text alignment, we introduce visual textualization, which projects support objects into the text feature space. This process allows limited exemplars from novel categories to participate in prompting, guiding OVLM to perform open-set detection alongside semantic text prompts. The visual textualized support images, combined with the text prompt, are directly input into the unmodified pre-trained OVLM, maintaining its established object-text alignment. Specifically, we first leverage OVLM's visual encoder to extract representations from visual exemplars and design lightweight multiscale textualizing blocks (MSTBs) for projection. MSTBs process the multi-scale intermediate visual features of the given exemplars produced by each layer of OVLM's visual encoder and uniformly project them into OVLM's text feature space as textualized visual features. Furthermore, we employ a non-parametric multi-stage fusion strategy (MSF) to fuse the textualized visual features from different stages of the visual encoder, leveraging OVLM's multi-stage object-text alignment without introducing additional parameters. Each visual exemplar is projected into a textualized visual + +token, guiding OVLM in inferring the target image alongside the text prompts. MSTB and MSF fully leverage an OVLM's original structure, robust object-text alignment, and its capability for object-level feature extraction, enabling the textualized visual tokens to achieve optimal semantic representation. During training, MSTB is the only trainable structure. Without fine-tuning on novel classes, MSTB can directly project the visual exemplars from novel categories into semantic-rich textualized visual tokens, which is suitable for practical applications such as open-vocabulary detection on personal mobile devices using image prompts. Previous methods break OVLMs' pre-trained object-text alignment when introducing image prompts. VisTex-OVLM's novelty lies in effectively leveraging the original OVLM structure to incorporate visual information that complements underspecified semantics missing from the text prompts, without altering the model's pre-trained alignment. + +Unlike MQ-Det which modulates text features, VisTex-OVLM directly incorporates visual information, preserving OVLM's inference structure and object-text alignment to the greatest extent (method 6, Fig. 1). When applied to GLIP (VisTex-GLIP) and GroundingDINO [33] (VisTex-DINO), VisTex-OVLM demonstrates superior performance in open-set scenarios, including LVIS [16] and 16 datasets with minimal overlap with OVLM pre-training data. Furthermore, VisTex-OVLM also achieves state-of-the-art (SOTA) results on standard few-shot benchmarks, including PASCAL VOC [9] and MSCOCO [32], showing broad applicability. Extensive ablation studies further validate the effectiveness of using images as semantic complements to text prompts. The contributions are summarized as follows: + +- We propose VisTex-OVLM, introducing visual textualization, which projects visual exemplars into the text feature space for image prompting to expand OVLM's practical applicability in detecting categories absent from pre-training data without disrupting their object-text alignment. +- Multi-scale textualizing blocks (MSTBs) and a multi-stage fusion (MSF) strategy are designed to preserve the original structure and robust alignment of OVLMs while enabling efficient utilization of their pre-trained knowledge. +- VisTex-OVLM excels with limited visual exemplars in open-set scenarios and achieves SOTA performance on standard FSOD benchmarks. + +# 2. Related works + +# 2.1. Few-shot object detection + +Few-shot object detection aims at detecting novel objects utilizing only few-shot annotated instances from novel classes. Mainstream methods mostly rely on complex episode-based meta-learning training paradigms [10, 18, 20, 21, 56] or intricate inter-class relationship modeling [40, 47, 48, 50]. Despite the progress made, FSOD remains a challenging + +![](images/0b500fdd078db16e000ad3a516dd6d1765b9513ac153bee2a6badd64491ced1c.jpg) +Figure 2. Overview of VisTex-OVLM. VisTex-OVLM performs visual textualization of support images through parameter-shared multi-scale textualizing blocks (MSTB) and a non-parametric multi-stage fusion strategy, mapping support images to text feature space for direct prompting of unmodified OVLMs, such as GLIP [31] and GroundingDINO [33]. + +task. Recently, VLMs have shown remarkable generalization capabilities, which led to their growing application in FSOD [15, 17, 19, 30, 52, 58]. However, most approaches demonstrate limited performance. The current SOTA model, MTL-FSOD [43], leverages VLMs solely for enhancing classification in FSOD through knowledge distillation, falling short of fully exploiting the potential of VLMs in FSOD. Our VisTex-OVLM probes deeper, aiming to fully leverage the potential of VLM in FSOD. + +# 2.2. VLM-based detection and grounding + +VLMs leverage vast text-image pairs obtained from the internet to train its text encoder and image encoder with contrastive learning, allowing the model to learn highly generalizable features. Among them, CLIP [42] is widely recognized for its image-level alignment, supporting effective transfer to image classification and text-image retrieval tasks. However, due to its lack of object-level knowledge, CLIP requires fine-tuning or adding detector modules for detection and grounding tasks [14, 28, 51, 59], increasing model complexity and training costs. Object-level VLMs (OVLMs) designed for detection tasks have emerged to address this [5, 31, 33]. A representative work, GLIP [31] combines object detection and phrase grounding data, utilizing grounded pre-training and a multi-scale, multi-stage cross-attention architecture to establish strong object-text alignment, lead + +ing to superior downstream detection and grounding performance over image-level open-vocabulary detectors like regionCLIP [60]. Therefore, our VisTex-OVLM is built upon OVLMs, which are inherently well-suited for detection and grounding tasks, to further explore the potential of VLM on FSOD. + +# 2.3. Prompting methods + +Prompting provides task-specific hints to guide large pretrained models, enhancing their understanding and results [13]. Originating in NLP, it now extends to VLMs. + +Design and tuning Prompt design refers to manually constructing prompts [8, 41], which can be complex, time-consuming, and prone to subjective biases. Prompt tuning adds learnable tokens to VLMs and fine-tunes [2, 4, 24, 26, 61, 62]. However, tuning may disrupt VLMs' cross-modal alignment, defecting its generalization capabilities. + +Image prompting Image prompting is a novel method where images are regarded as prompts to provide semantic guidance alongside text prompts. CLIPSeg [36] applies this for one-shot segmentation but requires an extra segmentation module, which can disrupt cross-modal alignment. In detection tasks, OWL-ViT [38] allows image prompting, but it relies on image-object pre-training and a predefined architecture and thus cannot be generalized to other OVLMs. MQ-Det [53] integrates a trainable cross-attention + +module within the text encoder of OVLM, allowing the image prompt to modulate the text prompt in a token-wise manner and thereby introducing the information of prompt images. However, MQ-Det only re-weights tokens without fully incorporating new visual information. Moreover, the additional cross-attention structure may also compromise OVLM's object-text alignment and generalization. + +In contrast to the aforementioned approaches, our VisTex-OVLM leverages few-shot support images as prompts through visual textualization, inserting their semantics directly into OVLM's text features. This allows the visual information of image prompts to be directly incorporated into the text prompt. More importantly, our approach does not alter the original structure of OVLM, fully maintaining its excellent object-text alignment, which results in excellent generalization ability and performance. + +# 3. Method + +Without loss of generality, we introduce our approach using notations of few-shot object detection (FSOD). + +# 3.1. Task definition + +In the classical few-shot object detection (FSOD) setting, there are two distinct datasets: a base set $D_{b}$ with category set $C_{b}$ and a novel set $D_{n}$ with category set $C_{n}$ , where $C_{b} \cap C_{n} = \emptyset$ . The base set $D_{b}$ contains abundant annotated objects that can be utilized for training the FSOD system, while in the novel set $D_{n}$ , each category $C \in C_{n}$ has at most $K$ bounding box annotations for $K$ -shot (e.g., $K = 1,2,3\dots 10$ ) object detection learning. In the inference phase, the images providing $K$ -shot annotations are referred to as support images, while the images to be predicted, termed query images, belong to the category set $C_{n}$ . + +In the more practical generalized few-shot object detection (GFSOD) setting, the target category set during inference becomes $C_b \cup C_n$ , requiring the model to maintain robust performance on base classes as well. + +# 3.2. Preliminaries and notations for OVLMs + +Commonly, OVLMs extend conventional VLMs by incorporating specialized designs for object-level visual tasks [5, 31, 33]. First, they use a hierarchical visual encoder to perform multi-scale encoding and generate region features at different spatial scales—aligned with text features—to capture objects of varying sizes. Second, they expand the text encoder's vocabulary using web-sourced phrase grounding data. Third, they adopt a multi-stage architecture with cross-modal attention to enhance region-text alignment. These specialized designs enable OVLMs to achieve superior object-text alignment, making them particularly suitable for object detection tasks. + +To facilitate subsequent discussions, we now define the notation for the OVLM encoding process. Let the inter + +mediate visual features and text features in OVLMs be denoted as $R^i$ and $P^i$ , respectively, where $i = 0,1,\dots,L$ and $L$ is the total number of stages. $R^0$ denotes the visual tokens outputted from a pre-trained visual tokenizer, typically a Swin [34], $P^0$ denotes the text prompt tokens tokenized by a pre-trained text tokenizer, usually BERT [25]. Suppose the text prompt $t$ for target objects contains $|C|$ classes, and for simplicity, assume each class is described using $N$ words ( $N = 1$ if only the class name is used as the prompt). Throughout the encoding process, $P \in \mathbb{R}^{|C| \cdot N \times d_T}$ , where $d_T$ is the dimension of the text feature space. Due to the multi-scale encoding, $R^i = \left[R^{i,(0)},\ldots,R^{i,(j)},\ldots,R^{i,(M-1)}\right] = \left[R^{i,(j)}\right]_{j=0}^{M-1}$ , where $R^{i,(j)} \in \mathbb{R}_{2^j}^{\frac{H}{2^j} \cdot \frac{W}{2^j} \times d_I}$ , $H$ and $W$ are the height and width of the largest scale visual features, $M$ is the total number of scales, and $d_I$ is the visual feature space dimension. The notation $[\cdot]$ represents concatenation. + +# 3.3. VisTex-OVLM + +As revealed by the empirical analysis in the introduction, the fine-tuning of weights or additional processing structures tends to compromise an OVLM's well-established object-text alignment, adversely affecting the generalization. To address these challenges, we propose VisTex-OVLM, a novel approach that textualizes support images as prompts during the inference phase to guide a non-finetuned pre-trained OVLM in detecting novel objects, as illustrated in Fig. 2. VisTex-OVLM introduces multi-scale textualizing blocks (MSTB) and a multi-stage fusion strategy to achieve the visual textualization of support images. The MSTB maps the visual features of support images into the text feature space, enabling direct support image prompting of an unmodified OVLM model while preserving its excellent pre-trained object-text alignment. + +# 3.3.1. Visual textualization + +An effective approach to modifying VLM's prediction preferences without altering its weights and architecture is to adjust the text prompts in the language branch. Thus, we propose mapping support images to the text feature space, i.e. visual textualization, to incorporate information about novel classes while preserving OVLM's object-text alignment. + +Image prompt engineering Since object detection is a dense prediction task, it is crucial to maximize the utilization of few-shot annotations for distinguishing the target object in the support image from irrelevant objects and the background. Following Luddecke et al.'s [36] experience with image prompting in semantic segmentation, we implement image prompt engineering by combining the support images with the target bounding boxes. Specifically, we apply background blur to regions outside the target bounding box. + +Multi-scale textualizing Let $x_{S}$ denote the prompt-engineered one-shot support image. We utilize OVLM's + +frozen visual encoder to extract multi-scale features containing information about objects of different sizes. For stage $i$ , we denote the extracted visual features of $x_{S}$ as: $R_{S}^{i} = \left[R_{S}^{i,(j)}\right]_{j=0}^{M-1}$ , where $R_{S}^{i,(j)} \in \mathbb{R}_{2^j}^{\frac{H}{2^j} \cdot \frac{W}{2^j} \times d_I}$ . We design a multi-scale textualizing block (MSTB) to process visual features at all scales uniformly and map them to the text feature space, as illustrated in Fig. 2. Specifically, we employ $3 \times 3$ convolutions with stride 2 to progressively downsample and extract the most effective visual information from larger-scale features, followed by an MLP that maps the visual features to the text feature dimension $d_T$ . + +To reduce the cost of training MSTB on the base set $D_{b}$ and effectively utilize shared knowledge across features of different scales, we adopt a parameter-sharing strategy. Specifically, we apply a shared set of convolutions to process all features except those at the smallest scale, followed by an MLP mapping to obtain the textualized visual feature $\widetilde{P_S^i}$ : + +$$ +\widetilde {R _ {S} ^ {i , (j)}} = \left\{ \begin{array}{l l} \left(\prod_ {j} ^ {M - 2} \operatorname {C o n v} _ {\text {d o w n}} ^ {i, (j)}\right) \left(R _ {S} ^ {i, (j)}\right), & \text {i f} j \neq M - 1 \\ R _ {S} ^ {i, (j)}, & \text {i f} j = M - 1 \end{array} \right. \tag {1} +$$ + +$$ +\widetilde {R _ {S} ^ {i}} = \left[ \widetilde {R _ {S} ^ {i , (j)}} \right] _ {j = 0} ^ {M - 1} \in \mathbb {R} ^ {M \cdot \frac {H}{2 ^ {M - 1}} \cdot \frac {W}{2 ^ {M - 1}} \times d _ {I}}, \tag {2} +$$ + +$$ +\widetilde {P _ {S} ^ {i}} = \mathrm {M L P} ^ {i} \left(\widetilde {R _ {S} ^ {i}}\right) \in \mathbb {R} ^ {1 \times d _ {T}}. \tag {3} +$$ + +This strategy facilitates learning the mapping functions that project different-scale features into a unified text feature space. For FSOD, multi-scale textualizing enables comprehensive feature extraction from support samples across different scales, capturing both fine-grained local details and global contextual relationships. This multi-scale approach significantly enhances the representational capability of $\widetilde{P_S^i}$ for target objects in query images. + +Multi-stage fusion We further implement a multi-stage fusion (MSF) strategy to integrate textualized visual features $\widetilde{P_S^i}$ from different visual encoder stages of support images. This fusion process consolidates each support image into a single textualized visual token $\widetilde{P_S}$ : + +$$ +\widetilde {P _ {S}} = \operatorname {M S F} ^ {i} \left(\{\widetilde {R _ {S} ^ {i}} \} _ {i = 0} ^ {L}\right). \tag {4} +$$ + +This fusion process can be implemented through non-parametric operations, conserving computational resources while achieving effective results. In fact, the synergistic effect of multi-stage fusion and multi-scale textualizing fully leverages OVLM's well-established object-text alignment and object-level visual feature extraction capabilities, enabling the textualized visual token $\widetilde{R_S^i}$ to achieve optimal semantic representation. + +# 3.3.2. Direct support image prompting + +Previous methods that use images for prompting either add extra components [36, 53] that risk disrupting OVLM's object-text alignment or modulate text features [53] without directly incorporating visual information. In contrast, VisTex-OVLM achieves direct integration of support image visual information through visual textualization. + +In VisTex-OVLM, we concatenate the textualized visual token $\widetilde{P_S}$ with text prompt tokens processed by a pre-trained BERT and input them directly into an unmodified pre-trained OVLM: + +$$ +P ^ {0} = \left[ \operatorname {B E R T} (t), \widetilde {P _ {S}} \right]. \tag {5} +$$ + +For $K$ -shot tasks, we maintain each shot's independence — essential for OVLM to capture support sample distribution for query prediction — by directly concatenating textualized tokens from different support images, avoiding information-losing fusion: + +$$ +P ^ {0} = \left[ \operatorname {B E R T} (t), \widetilde {P _ {S _ {1}}}, \dots , \widetilde {P _ {S _ {K}}} \right]. \tag {6} +$$ + +For a setting with $|C|$ classes, $P^0 \in \mathbb{R}^{|C| \cdot (N + K) \times d_T}$ . During the training phase, we construct corresponding text prompts $t$ using class labels from the base set $D_b$ . Following previous FSOD work, we randomly select $K$ -shot samples as support images for each base class, utilizing the remaining data as queries, and train MSTB end-to-end using OVLM's pre-training loss functions. After training, MSTB can directly perform textualization on novel class data. + +Unlike previous image prompting methods, we directly embed the semantics of few-shot support images into OVLM's text prompt $P^0$ in the form of textualized visual tokens, enabling direct incorporation of visual information. Our approach maintains the OVLM's original architecture, fully preserving its excellent pre-trained object-text alignment. Implementations of VisTex-OVLM on GLIP and GroundingDINO, namely VisTex-GLIP and VisTex-DINO, demonstrate superior generalization ability and performance. + +# 4. Experiment + +# 4.1. Datasets + +Open-set scenarios: To rigorously validate VisTex-OVLM's effectiveness on genuinely novel categories, we evaluated its transferability in open-set scenarios, including LVIS Mini-Val [16] and 16 unseen datasets with minimal overlap with OVLM pre-training data. LVIS challenges long-tail object detection, while 11 datasets from ODinW35 [29] (mAP $\leq 2$ for GLIP-L) and 5 medical datasets (MoNu [27], CCRCC [11], ConSep [12], LIDC [1], and Deeplesion [54]) test robustness in non-natural imaging contexts. + +Standard FSOD benchmarks Following previous FSOD works [46, 56], we evaluated our method on PASCAL VOC + +
MethodLVIS MiniValUnseen subsets from ODinW35Unseen medical datasets
APAPrABCDEFGHIJKMoNuCCRCCConSePLIDCDeeplesion
Meta-DETR [56]24.620.122.821.530.523.628.929.423.632.412.223.17.315.518.219.91.20.6
DiGeo [37]24.120.535.922.621.425.131.538.729.643.123.121.811.618.220.422.86.12.3
DeFRCN [40]32.328.742.341.531.133.836.841.132.640.233.925.219.316.315.613.15.22.8
MFD [50]28.926.436.233.532.324.827.238.528.43431.742.222.89.413.512.17.53.5
MQ-Det [53]36.231.144.842.938.133.541.840.122.632.122.446.326.17.17.68.26.33.1
OWL-ViT [38]--53.642.242.743.640.238.421.938.530.536.123.50.216.715.90.20.0
OWL-ViTv2 [39]47.237.865.452.145.847.842.350.638.245.138.549.830.74.423.422.56.51.2
GroundingDINO-ZS36.130.80.10.10.30.30.20.80.40.71.31.30.32.016.20.70.10.0
GroundingDINO-FF*38.333.417.53.66.98.219.77.28.120.94.266.924.425.844.6222.80.5
VisTex-DINO (Ours)42.837.270.268.458.566.562.073.156.163.339.871.632.527.738.529.69.610.8
GLIP-ZS37.328.20.00.00.10.40.91.00.70.11.30.40.44.28.81.30.30.0
GLIP-FF*49.540.575.869.861.470.567.475.455.363.144.568.741.228.530.635.111.311.9
VisTex-GLIP (Ours)50.742.977.570.263.168.069.575.758.465.746.973.538.928.931.535.712.211.7
+ +Table 1. Performance in open-set scenarios. Subsets A~K from ODinW35 [29]: PKLot_640, openPoetryV, boggleBoards, dicemedCol, OxfPetsbybreed, UnoCards, plantdoc, EgoHands_s, webScreenshots, OxfPetsbyspecies, MaskWearing. Best results are marked in bold. + +
Method/ShotSplit 1Split 2Split 3Mean
123510123510123510
non-VLM-based
MetaDet [45]ICCV 1918.920.630.236.849.621.823.127.831.743.020.623.929.443.944.131.0
MPSR [49]ECCV 2041.7-51.455.261.824.4-39.239.947.835.6-42.348.049.744.8
DeFRCN [40]ICCV 2153.657.561.564.160.830.138.147.053.347.948.450.952.354.957.451.9
Meta-DETR [56]TPAMI 2240.651.458.059.263.637.036.643.749.154.641.645.952.758.960.650.2
MFD [50]ECCV 2263.466.367.769.468.142.146.553.455.353.856.158.359.062.263.759.0
ICPE [35]AAAI 2354.359.562.465.766.233.540.148.751.752.550.963.155.360.660.155.0
VFA [23]AAAI 2357.764.664.767.267.441.446.251.151.851.648.954.856.659.058.956.1
Du et al. [6]ICCV 2352.355.563.165.966.742.745.848.754.856.347.851.856.860.362.455.4
SNIDA-MFD [48]CVPR 2464.967.969.771.470.542.247.854.556.654.958.161.360.763.666.060.7
VLM-based
D&R [30]AAAI 2341.051.755.761.865.430.739.042.546.651.737.947.151.756.859.549.3
Norm-VAE [52]CVPR 2362.164.967.869.267.539.946.854.454.253.658.260.361.064.065.559.1
FM-FSOD-L [17]CVPR 2440.153.557.068.672.033.136.348.854.864.739.250.255.763.468.153.7
DP-DDCL [15]KBS 2449.460.163.967.569.137.543.448.452.456.245.456.359.063.065.755.8
VEIC [58]ESWA 2467.769.570.570.469.149.551.555.056.254.763.364.763.665.465.462.4
MTL-FSOD [43]ECCV 2468.971.572.174.572.265.569.873.574.473.168.869.870.071.671.971.2
OWL-ViT [38]ECCV 2245.348.251.452.351.742.843.148.251.250.343.945.647.349.149.648.0
OWL-ViTv2 [39]NeurlPS 2348.649.351.252.653.145.747.850.252.151.442.646.249.651.051.849.5
MQ-Det [53]NeurlPS 2353.958.461.464.765.149.251.355.658.460.249.953.557.460.263.157.5
GroundingDINO-ZSECCV 2449.643.347.446.8
GroundingDINO-FFECCV 2449.550.451.152.753.142.843.644.846.547.847.548.950.852.452.849.0
VisTex-DINOOurs51.853.155.85760.243.546.748.551.153.248.650.653.95658.552.6
GLIP-ZSCVPR 2250.543.049.547.7
GLIP-FFCVPR 2251.355.259.461.261.844.346.248.649.351.250.152.658.361.461.954.2
VisTex-GLIPOurs69.471.873.474.674.866.570.173.174.274.468.970.271.172.372.671.8
+ +Table 2. Comparison of different FSOD methods in terms of AP50 on three PASCAL VOC Novel Split sets. Best results are marked in bold. + +[9] and MSCOCO [32]. PASCAL VOC: We used three partitions of base and novel categories (SPLIT1, SPLIT2, and SPLIT3). Each partition divides the 20 PASCAL VOC categories into 15 base classes $(C_b)$ and 5 novel classes $(C_n)$ , reporting AP50 results on the novel set $(D_n)$ under 1, 2, 3, 5, and 10 shots. MSCOCO: For MSCOCO's 80 classes, the 20 classes overlapping with PASCAL VOC are designated as novel $(C_n)$ , and the remaining 60 as base $(C_b)$ , with mAP results reported under 1, 2, 3, 5, 10, and 30 shots. + +# 4.2. Implementation details + +Without loss of generality, we conducted main experiments using the pre-trained GLIP-L [31], referred to as VisTex-GLIP. We also implemented VisTex-OVLM on + +GroundingDINO-T [33], denoted as VisTex-DINO. Unless specified otherwise, we used the following experimental settings for VisTex-GLIP. + +The MSTB's MLP consists of two "fully connected (fc) + ReLU" layers. The two fc layers are responsible for spatial aggregation $\left(M\cdot \frac{H}{2^{M - 1}}\cdot \frac{W}{2^{M - 1}}\rightarrow 1\right)$ and channel dimension transformation $(d_I\to d_T)$ , respectively. According to the feature size and scale extracted by the vision encoder of the pre-trained GLIP-L, both $H$ and $W$ are set to 100, and $M$ is 5. By default, MSTB is applied to all 8 stages of GLIP-L, and max pooling across stages is used for non-parametric multi-stage fusion. + +For the training of MSTB, we utilized 2 Nvidia RTX 3090 GPUs, each contains 24 GB of memory. For each epoch on + +
MethodShot Number
12351030
non-VLM-based
Meta R-CNN [55]ICCV 191.01.82.84.06.511.1
MPSR [49]ECCV 205.16.77.48.79.814.1
DeFRCN [40]ICCV219.312.914.816.118.522.6
Meta-DETR [56]TPAMI 227.5-13.515.419.022.2
MFD [50]ECCV 2210.813.915.016.419.422.7
Du et al. [6]ICCV23----20.322.8
SNIDA-MFD [48]CVPR 2412.015.416.417.820.723.8
VLM-based
D&R [30]AAAI238.312.714.316.418.721.8
Norm-VAE [52]CVPR239.513.714.315.918.722.5
FM-FSOD-L [17]CVPR 245.711.015.721.927.737.0
DP-DDCL [15]KBS 249.012.714.617.219.923.0
VEIC [58]ESWA 2411.514.415.316.919.721.9
MTL-FSOD [43]ECCV 2412.816.917.519.522.725.2
OWL-ViT [38]ECCV 2222.226.131.432.431.532.8
OWL-ViTv2 [39]NeurlPS 2326.528.330.532.933.533.1
MQ-Det [53]NeurlPS 2342.246.347.547.747.848.1
GroundingDINO-ZSECCV 2448.5
GroundingDINO-FFECCV 2448.748.949.350.851.653.0
VisTex-DINOOurs48.949.250.252.554.355.4
GLIP-ZSCVPR 2242.9
GLIP-FFCVPR 2241.244.747.548.849.349.4
VisTex-GLIPOurs47.951.852.252.652.753.6
+ +$(D_b)$ , $K$ -shot support images were randomly selected for each class $C \in C_b$ , while the rest served as query images. Based on query image labels, the corresponding support images were used to generate textualized image tokens $\widetilde{P_S}$ which were then concatenated with text prompts. The training spanned 30 epochs using AdamW with an initial learning rate of $10^{-4}$ and weight decay of 0.01. + +Detailed implementation for VisTex-DINO and comparison methods are provided in the supplementary materials. + +# 4.3. Open-set scenarios + +We evaluated VisTex-OVLM's performance in open-set scenarios, including LVIS MiniVal and 16 unseen datasets with categories exhibiting minimal overlap with the OVLM pretraining data. We compared VisTex-OVLM with methods that have publicly available code and weights, reporting mAP unless otherwise specified in Tab. 1. We established two baselines: OVLM-ZS and OVLM-FF, e.g., GLIP-ZS for GLIP. In OVLM-ZS, we used pre-trained OVLM weights for zero-shot testing. In OVLM-FF, we fine-tuned all pretrained parameters on random 2-shot support images from the novel dataset, using class names as text prompts for both training and test evaluation, and thus is denoted by the "*'" marker in the results. Notably, OVLM-ZS cannot incorporate support images' visual information during testing. Other methods requiring training, including ours, were trained on MSCOCO's base set, treating open-set task sets as novel and testing with 2-shot support images, while OWL-ViT (v2) and OVLM-ZS were evaluated using their frozen pre-trained weights. The results represent the average of five experimental runs. The results show that the zero-shot performance of OVLM significantly drops for novel classes or domains + +Table 3. FSOD performance on MSCOCO. Best results are marked in bold. + +
Method10-shot30-shot
APbAPnAPAPbAPnAP
non-VLM-based
DiGeo [37]32.039.210.333.139.414.2
DE-VIT-L [57]30.629.434.030.629.534.0
VLM-based
FM-FSOD-L [17]40.044.227.743.145.237.0
MQ-Det [53]46.946.547.847.346.848.1
VisTex-GLIP48.146.252.748.947.253.6
+ +Table 4. GFSOD performance on MSCOCO. Best results are marked in bold. + +rarely encountered during pre-training compared to its performance on PASCAL VOC and MSCOCO. In contrast, our method outperforms or matches other approaches in most cases, demonstrating strong transferability. It is worth noting that OVLM-FF may achieve slightly higher performance in some cases due to the introduction of new textual knowledge through full fine-tuning. However, VisTex-OVLM, without any training on these datasets, still achieves comparable results. These findings underscore VisTex-OVLM's robust performance in open-set detection and its broad applicability across diverse domains. + +# 4.4. Comparison on FSOD benchmarks + +We compared our method with other recent FSOD methods, including both non-VLM-based methods and VLM-based methods, with results presented in Tabs. 2 and 3. In OVLM-FF, we fine-tuned all pre-trained parameters on $K$ -shot support images from the novel dataset. Additionally, we directly compared our method with MQ-Det, which is also based on OVLM and uses support images to modulate text prompts for image prompting. Tabs. 2 and 3 demonstrate the superiority of our approach over common transfer methods in utilizing a limited number of support images by textualizing them and directly prompting OVLM, for both GLIP and GroundingDINO. + +Our method achieves SOTA performance in classic FSOD while retaining strong recognition for base categories, meeting generalized FSOD (GFSOD) requirements. Tab. 4 shows comparisons with other GFSOD methods on MSCOCO in 10-shot and 30-shot settings. By preserving and leveraging VLMs' generalization, our method avoids overfitting and knowledge forgetting on base classes, achieving high performance across both base and novel categories. + +We also conducted compatibility experiments on Region-CLIP and FIBER in the supplementary materials to further validate the generalization ability of VisTex-OVLM. + +# 4.5. Ablation study + +We conducted extensive ablation studies on VisTex-OVLM in a 2-shot setting on MSCOCO using GLIP-L, reporting mAP on both the base and the novel classes as in Tab. 4. All other settings were kept at optimal configurations. More + +![](images/fc9c24a63d7459cf6beafc7990622189dd339634bc717a685b74a06b0699163e.jpg) +Figure 3. Comparison output visualizations on COCO. (a) MetaDETR, (b) DiGeo, (c) DeFRCN, (d) MFD, (e) MQ-Det, (f) VisTex-GLIP. Green, red, and yellow boxes denote true positives, false positives, and false negatives. + +ablation results and analysis regarding VisTex-OVLM's computational overhead are in the supplementary materials. + +We examined the effects of the prompting mode and crucial components, as shown in Tab. 5. Text prompts were generated by concatenating class words following the GLIP paper [31]. We compared GLIP-ZS, GLIP-FF, and GLIP-MaPLe (integrating MaPLe [26]) as a prompt-tuning example. Multi-scale textualization is fundamental to our method, so the "MSTB sharing" column shows only whether parameter-sharing is applied in MSTB. "x" in the "MSF" column for "VisTex-GLIP" indicates that textualization is applied only in stage 1. As detailed in Tab. 5, combining components achieves the best mAP for VisTex-GLIP. Utilizing visual support images addresses GLIP-ZS's limitation of relying solely on text prompts. VisTex-GLIP outperforms GLIP-FF, GLIP-MaPLe, and MQ-Det, which disrupt the object-text alignment by modifying model weights or structure. In contrast, MSTB avoids fine-tuning on novel classes, directly mapping novel support samples into textualized image tokens for prompting, which is user-friendly and avoids disrupting the object-text alignment of GLIP. Notably, image prompts alone without text prompts reduce performance, as text prompts offer essential semantic guidance, impacting VLMs' image-text alignment. + +We present attention heatmaps (Fig. 4) showing the effects of different prompting modes and key components, aligned with methods in Tab. 5. These heatmaps visualize object feature attention relative to text features. For "VisTex-GLIP without text prompts", the heatmap represents attention exclusively to textualized visual tokens. The heatmaps show that GLIP-FF and GLIP-MAPLE exhibit significant background noise, while MQ-Det suppresses noise but shows less focused object attention. In contrast, VisTex-GLIP achieves stronger target focus and reduced noise. Notably, VisTex-GLIP without text prompts shows weaker target attention than with text prompts. This observation underscores the importance of text prompts in providing essential semantic guidance, which crucially influences OVLMs' object-text + +
MethodText PromptImage PromptMSTB sharingMSF#Par(M)2-shot
APbAPnAP
GLIP-ZS✓×××0.0042.944.940.5
GLIP-FF✓✓*××397.5944.344.144.7
GLIP-MaPLe✓✓*××5.9248.547.849.7
MQ-Det✓✓××53.1044.743.546.3
VisTex-GLIP×✓××38.5230.632.525.9
✓✓××38.5237.129.741.2
✓✓✓×29.2048.345.749.0
✓✓✓✓63.0650.348.651.8
+ +Table 5. Effectiveness of the prompting mode and crucial components. "√*" represents directly using support images for fine-tuning, #Par(M) denotes the number of trainable parameters. Best results are marked in bold. + +![](images/c22f825a53e98bd1e0c66a7ffc99dff86e96a83b0272770ce9b68d66ae801bb5.jpg) +Figure 4. Attention heatmaps comparison. (a) Ground truth, (b) GLIP-ZS (zero-shot), (c) GLIP-FF (fully fine-tuning), (d) GLIP-MaPLe, (e) MQ-Det, (f) VisTex-GLIP without text prompts, (g) VisTex-GLIP. + +alignment and overall performance. + +# 5. Conclusion + +In this paper, we introduce VisTex-OVLM, a novel object detection approach that uses visual textualization to project visual exemplars into the text feature space to prompt an OVLM without disrupting its pre-trained object-text alignment. To implement visual textualization, we design MSTBs and an MSF strategy that fully leverage OVLM's strong object-text alignment and object-level feature extraction capabilities, allowing the textualized visual tokens to achieve optimal semantic representation. VisTex-OVLM achieves state-of-the-art performance on both open-set detection (LVIS and 16 unseen datasets) and few-shot benchmarks (PASCAL VOC and MSCOCO). VisTex-OVLM is applicable to common OVLMs including GLIP and GroundingDINO, underscoring the benefits of using image prompts to complement text prompts and validating its effectiveness in addressing zero-shot object detection limitations. + +Acknowledgement Supported by the National Natural Science Foundation in China under Grant 62371016 and U23B2063, the Beijing Natural Science Foundation Haidian District Joint Fund in China under Grant L222032, the Academic Excellence Foundation of BUAA for PhD Students. + +# References + +[1] Samuel G Armato III, Geoffrey McLennan, Michael F McNitt-Gray, Charles R Meyer, David Yankelevitz, Denise R Aberle, Claudia I Henschke, Eric A Hoffman, Ella A Kazerooni, Heber MacMahon, et al. Lung image database consortium: developing a resource for the medical imaging research community. Radiology, 232(3):739-748, 2004. 5 +[2] Hyojin Bahng, Ali Jahanian, Swami Sankaranarayanan, and Phillip Isola. Exploring visual prompts for adapting large-scale models. arXiv preprint arXiv:2203.17274, 2022. 3 +[3] Adrian Bulat, Ricardo Guerrero, Brais Martinez, and Georgios Tzimiropoulos. Fs-detr: Few-shot detection transformer with prompting and without re-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 11793-11802, 2023. 1 +[4] Sanjoy Chowdhury, Sayan Nag, and Dinesh Manocha. Apollo: Unified adapter and prompt learning for vision language models. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 10173-10187, 2023. 3 +[5] Zi-Yi Dou et al. Coarse-to-fine vision-language pre-training with fusion in the backbone. NeurIPS, 35:32942-32956, 2022. 1, 3, 4 +[6] Jinhao Du, Shan Zhang, Qiang Chen, Haifeng Le, Yanpeng Sun, Yao Ni, Jian Wang, Bin He, and Jingdong Wang. s-adaptive decoupled prototype for few-shot object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 18950-18960, 2023. 6, 7 +[7] Yu Du, Fangyun Wei, Zihe Zhang, Miaojing Shi, Yue Gao, and Guoqi Li. Learning to prompt for open-vocabulary object detection with vision-language model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14084-14093, 2022. 1 +[8] Avia Efrat and Omer Levy. The turking test: Can language models understand instructions? arXiv preprint arXiv:2010.11982, 2020. 3 +[9] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International journal of computer vision, 88:303-338, 2010. 2, 6 +[10] Qi Fan, Wei Zhuo, Chi-Keung Tang, and Yu-Wing Tai. Few-shot object detection with attention-rpn and multi-relation detector. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4013–4022, 2020. 2 +[11] Zeyu Gao, Jiangbo Shi, Xianli Zhang, Yang Li, Haichuan Zhang, Jialun Wu, Chunbao Wang, Deyu Meng, and Chen Li. Nuclei grading of clear cell renal cell carcinoma in histopathological image by composite high-resolution network. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 132-142. Springer, 2021. 5 +[12] Simon Graham, Quoc Dang Vu, Shan E Ahmed Raza, Ayesha Azam, Yee Wah Tsang, Jin Tae Kwak, and Nasir Rajpoot. Hover-net: Simultaneous segmentation and classification of nuclei in multi-tissue histology images. Medical Image Analysis, 58:101563, 2019. 5 + +[13] Jindong Gu, Zhen Han, Shuo Chen, Ahmad Beirami, Bailan He, Gengyuan Zhang, Ruotong Liao, Yao Qin, Volker Tresp, and Philip Torr. A systematic survey of prompt engineering on vision-language foundation models. arXiv preprint arXiv:2307.12980, 2023. 3 +[14] Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, and Yin Cui. Open-vocabulary object detection via vision and language knowledge distillation. arXiv preprint arXiv:2104.13921, 2021. 3 +[15] Yinsai Guo, Liyan Ma, Xiangfeng Luo, and Shaorong Xie. Dp-ddcl: A discriminative prototype with dual decoupled contrast learning method for few-shot object detection. Knowledge-Based Systems, 297:111964, 2024. 1, 3, 6, 7 +[16] Agrim Gupta, Piotr Dollar, and Ross Girshick. Lvis: A dataset for large vocabulary instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5356-5364, 2019. 2, 5 +[17] Guangxing Han and Ser-Nam Lim. Few-shot object detection with foundation models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 28608-28618, 2024. 1, 3, 6, 7 +[18] Guangxing Han, Yicheng He, Shiyuan Huang, Jiawei Ma, and Shih-Fu Chang. Query adaptive few-shot object detection with heterogeneous graph convolutional networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3263-3272, 2021. 2 +[19] Guangxing Han, Long Chen, Jiawei Ma, Shiyuan Huang, Rama Chellappa, and Shih-Fu Chang. Multi-modal few-shot object detection with meta-learning-based cross-modal prompting. arXiv preprint arXiv:2204.07841, 2022. 1, 3 +[20] Guangxing Han, Shiyuan Huang, Jiawei Ma, Yicheng He, and Shih-Fu Chang. Meta faster r-cnn: Towards accurate few-shot object detection with attentive feature alignment. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 780-789, 2022. 2 +[21] Guangxing Han, Jiawei Ma, Shiyuan Huang, Long Chen, and Shih-Fu Chang. Few-shot object detection with fully cross-transformer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5321-5330, 2022. 2 +[22] Guangxing Han, Jiawei Ma, Shiyuan Huang, Long Chen, and Shih-Fu Chang. Few-shot object detection with fully cross-transformer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5321-5330, 2022. 1 +[23] Jiaming Han, Yuqiang Ren, Jian Ding, Ke Yan, and GuiSong Xia. Few-shot object detection via variational feature aggregation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 755-763, 2023. 6 +[24] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In European Conference on Computer Vision, pages 709-727. Springer, 2022. 3 +[25] Jacob Devlin Ming-Wei Chang Kenton and Lee Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of naacL-HLT, page 2. Minneapolis, Minnesota, 2019. 4 + +[26] Muhammad Uzair Khattak, Hanoona Rasheed, Muhammad Maaz, Salman Khan, and Fahad Shahbaz Khan. Maple: Multimodal prompt learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19113-19122, 2023. 3, 8 +[27] Neeraj Kumar, Ruchika Verma, Sanuj Sharma, Surabhi Bhargava, Abhishek Vahadane, and Amit Sethi. A dataset and a technique for generalized nuclear segmentation for computational pathology. IEEE transactions on medical imaging, 36 (7):1550-1560, 2017. 5 +[28] Weicheng Kuo, Yin Cui, Xiuye Gu, AJ Piergiovanni, and Anelia Angelova. F-vlm: Open-vocabulary object detection upon frozen vision and language models. arXiv preprint arXiv:2209.15639, 2022. 3 +[29] Chunyuan Li, Haotian Liu, Liunian Li, Pengchuan Zhang, Jyoti Aneja, Jianwei Yang, Ping Jin, Houdong Hu, Zicheng Liu, Yong Jae Lee, et al. Elevater: A benchmark and toolkit for evaluating language-augmented visual models. Advances in Neural Information Processing Systems, 35:9287-9301, 2022. 5, 6, 1 +[30] Jiangmeng Li, Yanan Zhang, Wenwen Qiang, Lingyu Si, Chengbo Jiao, Xiaohui Hu, Changwen Zheng, and Fuchun Sun. Disentangle and remerge: interventional knowledge distillation for few-shot object detection from a conditional causal perspective. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1323-1333, 2023. 1, 3, 6, 7 +[31] Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, et al. Grounded language-image pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10965-10975, 2022. 1, 3, 4, 6, 8 +[32] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 1, 2, 6 +[33] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Qing Jiang, Chunyuan Li, Jianwei Yang, Hang Su, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 1, 2, 3, 4, 6 +[34] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 4 +[35] Xiaonan Lu, Wenhui Diao, Yongqiang Mao, Junxi Li, Peijin Wang, Xian Sun, and Kun Fu. Breaking immutable: Information-coupled prototype elaboration for few-shot object detection. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1844-1852, 2023. 6 +[36] Timo Lüddecke and Alexander Ecker. Image segmentation using text and image prompts. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7086-7096, 2022. 1, 3, 4, 5, 2 + +[37] Jiawei Ma, Yulei Niu, Jincheng Xu, Shiyuan Huang, Guangxing Han, and Shih-Fu Chang. Digeo: Discriminative geometry-aware learning for generalized few-shot object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3208-3218, 2023. 6, 7 +[38] Matthias Minderer, Alexey Gritsenko, Austin Stone, Maxim Neumann, Dirk Weissenborn, Alexey Dosovitskiy, Aravindh Mahendran, Anurag Arnab, Mostafa Dehghani, Zhuoran Shen, et al. Simple open-vocabulary object detection. In European conference on computer vision, pages 728-755. Springer, 2022. 3, 6, 7 +[39] Matthias Minderer, Alexey Gritsenko, and Neil Houlsby. Scaling open-vocabulary object detection. Advances in Neural Information Processing Systems, 36:72983-73007, 2023. 6, 7 +[40] Limeng Qiao, Yuxuan Zhao, Zhiyuan Li, Xi Qiu, Jianan Wu, and Chi Zhang. Defrcn: Decoupled faster r-cnn for few-shot object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 8681-8690, 2021. 1, 2, 6, 7 +[41] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 3 +[42] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 3 +[43] Yan Ren, Yanling Li, and Adams Wai-Kin Kong. Adaptive multi-task learning for few-shot object detection. In European Conference on Computer Vision, pages 297-314. Springer, 2024. 3, 6, 7 +[44] Shuai Shao, Zeming Li, Tianyuan Zhang, Chao Peng, Gang Yu, Xiangyu Zhang, Jing Li, and Jian Sun. Objects365: A large-scale, high-quality dataset for object detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 8430-8439, 2019. 1 +[45] Tao Wang, Xiaopeng Zhang, Li Yuan, and Jiashi Feng. Few-shot adaptive faster r-cnn. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7173-7182, 2019. 1, 6 +[46] Xin Wang, Thomas E Huang, Trevor Darrell, Joseph E Gonzalez, and Fisher Yu. Frustratingly simple few-shot object detection. arXiv preprint arXiv:2003.06957, 2020. 5 +[47] Xin Wang, Thomas E Huang, Trevor Darrell, Joseph E Gonzalez, and Fisher Yu. Frustratingly simple few-shot object detection. arXiv preprint arXiv:2003.06957, 2020. 2 +[48] Yanjie Wang, Xu Zou, Luxin Yan, Sheng Zhong, and Jiahuan Zhou. Snida: Unlocking few-shot object detection with nonlinear semantic decoupling augmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12544–12553, 2024. 2, 6, 7 +[49] Jiaxi Wu, Songtao Liu, Di Huang, and Yunhong Wang. Multiscale positive sample refinement for few-shot object detection. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XVI 16, pages 456-472. Springer, 2020. 6, 7 + +[50] Shuang Wu, Wenjie Pei, Dianwen Mei, Fanglin Chen, Jiandong Tian, and Guangming Lu. Multi-faceted distillation of base-novel commonality for few-shot object detection. In European Conference on Computer Vision, pages 578-594. Springer, 2022. 2, 6, 7 +[51] Xiaoshi Wu, Feng Zhu, Rui Zhao, and Hongsheng Li. Cora: Adapting clip for open-vocabulary detection with region prompting and anchor pre-matching. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7031-7040, 2023. 3 +[52] Jingyi Xu, Hieu Le, and Dimitris Samaras. Generating features with increased crop-related diversity for few-shot object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19713-19722, 2023. 1, 3, 6, 7 +[53] Yifan Xu, Mengdan Zhang, Chaoyou Fu, Peixian Chen, Xiaoshan Yang, Ke Li, and Changsheng Xu. Multi-modal queried object detection in the wild. In Proceedings of the 37th International Conference on Neural Information Processing Systems, pages 4452-4469, 2023. 1, 3, 5, 6, 7 +[54] Ke Yan, Xiaosong Wang, Le Lu, and Ronald M Summers. Deepleson: Automated deep mining, categorization and detection of significant radiology image findings using large-scale clinical lesion annotations. arXiv preprint arXiv:1710.01766, 2017. 5 +[55] Xiaopeng Yan, Ziliang Chen, Anni Xu, Xiaoxi Wang, Xiaodan Liang, and Liang Lin. Meta r-cnn: Towards general solver for instance-level low-shot learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9577-9586, 2019. 7 +[56] Gongjie Zhang, Zhipeng Luo, Kaiwen Cui, Shijian Lu, and Eric P Xing. Meta-detr: Image-level few-shot detection with inter-class correlation exploitation. IEEE transactions on pattern analysis and machine intelligence, 45(11):12832-12843, 2022. 2, 5, 6, 7 +[57] Xinyu Zhang, Yuting Wang, and Abdeslam Boularias. Detect every thing with few examples. arXiv preprint arXiv:2309.12969, 2023. 7 +[58] Taijin Zhao, Heqian Qiu, Yu Dai, Lanxiao Wang, Hefei Mei, Fanman Meng, Qingbo Wu, and Hongliang Li. Vlm-guided explicit-implicit complementary novel class semantic learning for few-shot object detection. Expert Systems with Applications, 256:124926, 2024. 1, 3, 6, 7 +[59] Xiaowei Zhao, Xianglong Liu, Duorui Wang, Yajun Gao, and Zhide Liu. Scene-adaptive and region-aware multi-modal prompt for open vocabulary object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16741-16750, 2024. 3 +[60] Yiwu Zhong et al. Regionclip: Region-based language-image pretraining. In CVPR, pages 16793-16803, 2022. 3, 1 +[61] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Conditional prompt learning for vision-language models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16816-16825, 2022. 3 +[62] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, 130(9):2337-2348, 2022. 3 \ No newline at end of file diff --git a/ICCV/2025/Visual Textualization for Image Prompted Object Detection/images.zip b/ICCV/2025/Visual Textualization for Image Prompted Object Detection/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..bc7f05d32445af849d87a128fcf82370ed0c0d44 --- /dev/null +++ b/ICCV/2025/Visual Textualization for Image Prompted Object Detection/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3724a5d2f57573233d7324ab5448feca40ca794c43ddd2a6f40048c55042253e +size 808193 diff --git a/ICCV/2025/Visual Textualization for Image Prompted Object Detection/layout.json b/ICCV/2025/Visual Textualization for Image Prompted Object Detection/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..91a5ac71d2481119fca0e5b497365eb5b83e554e --- /dev/null +++ b/ICCV/2025/Visual Textualization for Image Prompted Object Detection/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:81f72ecbcd5d753b73230cbdcf9b4dd4c8c52b50f5718da5192d7de6f3125b0b +size 374189 diff --git a/ICCV/2025/Visual-Oriented Fine-Grained Knowledge Editing for MultiModal Large Language Models/a7f3a1fd-fa58-4e7e-b24e-304e05b363a7_content_list.json b/ICCV/2025/Visual-Oriented Fine-Grained Knowledge Editing for MultiModal Large Language Models/a7f3a1fd-fa58-4e7e-b24e-304e05b363a7_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..5a2700c0cb0286dd20feee65a428f97cc44d8b64 --- /dev/null +++ b/ICCV/2025/Visual-Oriented Fine-Grained Knowledge Editing for MultiModal Large Language Models/a7f3a1fd-fa58-4e7e-b24e-304e05b363a7_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e016cc8310a616e7c309a37492cd3cae935c48475c10a309e14a504e4f4fce2c +size 76723 diff --git a/ICCV/2025/Visual-Oriented Fine-Grained Knowledge Editing for MultiModal Large Language Models/a7f3a1fd-fa58-4e7e-b24e-304e05b363a7_model.json b/ICCV/2025/Visual-Oriented Fine-Grained Knowledge Editing for MultiModal Large Language Models/a7f3a1fd-fa58-4e7e-b24e-304e05b363a7_model.json new file mode 100644 index 0000000000000000000000000000000000000000..5d86e7e6e2938a2da6d0774afdd39c35f86bb441 --- /dev/null +++ b/ICCV/2025/Visual-Oriented Fine-Grained Knowledge Editing for MultiModal Large Language Models/a7f3a1fd-fa58-4e7e-b24e-304e05b363a7_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2aff000c71049cf1a2c3749406ff7dbc1e3185dc780fdda4bf2540a27dfb8015 +size 91705 diff --git a/ICCV/2025/Visual-Oriented Fine-Grained Knowledge Editing for MultiModal Large Language Models/a7f3a1fd-fa58-4e7e-b24e-304e05b363a7_origin.pdf b/ICCV/2025/Visual-Oriented Fine-Grained Knowledge Editing for MultiModal Large Language Models/a7f3a1fd-fa58-4e7e-b24e-304e05b363a7_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..c7bc2ec2a4ae5f12bc70e99ed57dd50b0038380b --- /dev/null +++ b/ICCV/2025/Visual-Oriented Fine-Grained Knowledge Editing for MultiModal Large Language Models/a7f3a1fd-fa58-4e7e-b24e-304e05b363a7_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:50d6c5ccb2be592e17df444069b2ce0664c5ef9dd34213d7b8fcc26923b03205 +size 969849 diff --git a/ICCV/2025/Visual-Oriented Fine-Grained Knowledge Editing for MultiModal Large Language Models/full.md b/ICCV/2025/Visual-Oriented Fine-Grained Knowledge Editing for MultiModal Large Language Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..e1c638ae1f21d3dfb7a6298ef1ab147b14c2fc41 --- /dev/null +++ b/ICCV/2025/Visual-Oriented Fine-Grained Knowledge Editing for MultiModal Large Language Models/full.md @@ -0,0 +1,339 @@ +# Visual-Oriented Fine-Grained Knowledge Editing for MultiModal Large Language Models + +Zhen Zeng $^{1,*}$ Leijiang Gu $^{1,*}$ Xun Yang $^{2}$ Zhangling Duan $^{3,\dagger}$ Zenglin Shi $^{1,\dagger}$ Meng Wang $^{1}$ $^{1}$ Hefei University of Technology + $^{2}$ University of Science and Technology of China + $^{3}$ Hefei Comprehensive National Science Center + +{zengzhen, 2024170839}@mail.hfut.edu.cn {zenglin.shi, wangmeng}@hfut.edu.cn xyang21@ustc.edu.cn duanzl1024@ahu.edu.cn + +# Abstract + +Existing knowledge editing works for MultiModal Large Language Models primarily focus on text-oriented, coarse-grained scenarios, where modifying textual content alone is sufficient. As a result, they fail to capture the unique challenges of multimodal editing, particularly when visual information is central to knowledge representation. In this paper, we introduce a visual-oriented, fine-grained multimodal knowledge editing task that targets precise modifications in images containing multiple interacting entities. To support this, we propose the Fine-Grained Visual Knowledge Editing (FGVEdit) benchmark, designed to evaluate the accuracy and effectiveness of multimodal editing at a granular level. To address this challenge, we present the Multimodal Scope Classifier-based Knowledge Editor (MSCKE), a new framework that leverages a multimodal scope classifier to integrate both textual and visual information. By accurately identifying and updating knowledge localized within images, MSCKE ensures precise editing while preserving unrelated content. Extensive experiments on the FGVEdit benchmark highlight the complexity of this new task and demonstrate that existing methods struggle with fine-grained multimodal editing. Our results highlight MSCKE as a scalable and promising framework for advancing multimodal knowledge editing. Code is available at https://github.com/zeng-zhen/FGVEdit. + +# 1. Introduction + +Traditional knowledge editing primarily strives to update and correct knowledge in Large Language Models (LLMs) to ensure their accuracy and reliability [22, 24, 29, 32]. Many works, e.g., [6, 16, 19, 20, 33, 34], have pro + +![](images/0951b10cd2a4606ae188c9dfca588fd78e9409acb2d8edad46beca5138c1e10a.jpg) +Figure 1. Comparison of fine-grained and coarse-grained knowledge editing in multimodal models. + +posed knowledge editing methods for LLMs, which have shown strong performance in refining model knowledge. Recently, there has been a growing interest in extending knowledge editing to Multimodal Large Language Models (MLLMs) [4, 30]. MLLMs enhance the capabilities of traditional LLMs by incorporating multiple modalities, such as text and images, allowing them to process and generate content that integrates both textual and visual information [2, 13, 15, 26, 28, 35]. In MLLMs, errors can arise not only from the language module but also from the visual components or the interactions between modalities. As a result, knowledge editing in MLLMs presents unique challenges that go beyond the complexities found in LLMs. + +To advance the field, some pioneering works [4, 9, 14, 31] have introduced multiple multimodal knowledge editing benchmarks. However, these benchmarks still primarily focus on text-oriented knowledge editing involving coarse-grained visual understanding, only considering treating the entire image as the target entity. In such settings, editing methods do not require access to the visual module. As illustrated in Fig. 1, simply editing the language model suf + +fices (e.g., by replacing the word "bird" with "kite"). As a result, knowledge editing methods for LLMs tend to perform well within these settings. In contrast, this work addresses the more complex task of editing knowledge associated with multiple interacting entities within a single image. To this end, we introduce a new FGVEdit benchmark, designed for visual-oriented fine-grained knowledge editing. Our problem setting requires editing methods to accurately identify and localize the target entity within the image, updating the content related to it (e.g., the kites and their associated relationships). In this context, traditional knowledge editing methods for LLMs are less effective. + +In this work, we introduce the Multimodal Scope Classifier-based Knowledge Editor (MSCKE) framework, specifically designed for visual-oriented, fine-grained knowledge editing. Building upon the text-based editing method SERAC [20], we make a key modification by replacing the text-only scope classifier with a multimodal scope classifier. Our proposed classifier combines visual and textual modalities to identify fine-grained target entities and extract relevant knowledge from images (e.g., the position relationship between "kites" and "water" shown in Fig. 1), enabling the capture of subtle entity similarities. This multimodal approach ensures that edited knowledge is applied only when the new input is meaningfully related to the target entity being edited. Such integration is crucial, as relying solely on textual semantics may fail to capture the nuanced relationships between entities in the same image. Our contributions are summarized as follows: + +- We introduce a visual-oriented, fine-grained multimodal knowledge editing task. To advance research in this direction, we propose MSCKE, a scalable multimodal knowledge editing framework. Together, this task and MSCKE enable precise, fine-grained modifications, setting a new standard for controllable knowledge updates in MLLMs. +- Building on the proposed task, we introduce FGVEdit, a benchmark designed to comprehensively evaluate the visual-oriented, fine-grained editing capabilities of multimodal knowledge editing methods. Featuring images with multiple entities, FGVEdit challenges existing approaches to achieve higher levels of editing performance. +- Our extensive experiments on the FGVEdit benchmark reveal that existing methods perform poorly in addressing the challenges of the proposed task. In contrast, our MSCKE framework demonstrates strong performance, highlighting its effectiveness and potential as a robust solution for advancing multimodal knowledge editing. + +# 2. Related Works + +# 2.1. Knowledge Editing for LLMs + +The existing knowledge editing methods for LLMs can be categorized into two main paradigms: parameter-preserving + +methods and parameter-modifying methods [1, 3, 12, 25]. + +Parameter-preserving methods. This approach modifies the output during inference of LLMs by utilizing additional components, while keeping the original model parameters unchanged. These additional components can include memory that stores edited examples and supplementary models that process the output. Representative methods include SERAC [20], IKE [33], and MemPrompt [16]. Additionally, T-Patcher [10] and CaliNET [7] incorporate extra trainable parameters into LLMs. + +Parameter-modifying methods. This approach involves updating parameters within LLMs to modify the model. One intuitive strategy is to identify and modify the parameters corresponding to specific knowledge, as demonstrated by KN [5], ROME [17], and MEMIT [18]. Another modification method employs meta-learning to train a hypernetwork that updates the parameters of LLMs, exemplified by MEND [19] and KE [6]. + +Building upon the parameter-preserving text editing method SERAC [20], we introduce the Multimodal Scope Classifier-based Knowledge Editor (MSCKE) framework by replacing the text-only scope classifier with a multimodal scope classifier. MSCKE establishes a new framework to extend knowledge editing from LLMs to MLLMs. + +# 2.2. Knowledge Editing for MLLMs + +MLLMs are capable of processing information from different modalities, such as text and images [2, 21, 23, 30]. Typically, they consist of a visual encoder, a language model, and an alignment module that integrates the visual and textual spaces. When performing knowledge editing on MLLMs, it is crucial to account for various types of errors, including visual comprehension errors, textual knowledge errors, and multimodal coordination errors. + +MMEdit [4] is the first benchmark for multimodal knowledge editing, evaluating editing methods for modifying visual knowledge. KEBench [9] introduces a portability metric to assess how well models leverage edited knowledge in relation to associated content. MIKE [14] and MC-MKE [31] propose fine-grained multimodal knowledge editing benchmarks at different levels. Specifically, MIKE [14] updates entity knowledge from coarse-grained to fine-grained (e.g., changing "old man" to "Trump"), while MC-MKE [31] distinguishes between error types. + +Existing approaches typically treat the entire image as a single target entity for editing, relying primarily on textual knowledge. Consequently, knowledge editing methods designed for LLMs can be directly applied in these scenarios. In contrast, this paper addresses the intricate interactions among multiple entities within a single image and introduces the FGVEdit benchmark to advance fine-grained visual knowledge editing. Under this problem setting, traditional LLM-based knowledge editing methods prove less + +![](images/815ee5ea469532a108ed3afc1df06dea35ccf7e3b556564e57477e7e90affc9d.jpg) +Figure 2. Architecture of the MSCKE method, illustrating the multimodal scope classifier, multimodal edit memory, base model with frozen parameters, counterfactual model with trainable parameters. During the editing phase, the editing samples are stored in the multimodal edit memory. In the inference phase, the multimodal classifier is employed to compute the similarity between the input and the editing samples in multimodal edit memory. Inputs with a similarity score less than 0.5 are classified as out-of-scope inputs and are processed by the base model; conversely, inputs are handled by the counterfactual model. + +![](images/ab1e469f62329f0af101e11ce920f0de48b8bd3fe3c2ae0e915d07e631c347e4.jpg) + +effective. To bridge this gap, we propose MSCKE, a multimodal knowledge editing framework equipped with a multimodal scope classifier, which achieves strong performance on the FGVEdit benchmark. + +# 3. Fine-Grained Visual Knowledge Editing + +# 3.1. Problem Formulation + +Fine-grained visual knowledge editing (FGVEdit) task focuses on updating specific and detailed knowledge within MLLMs that process both visual and textual information. Unlike other multimodal knowledge editing tasks, which treat the entire image as the target entity and perform coarse-grained editing that may overlook complex visual details, FGVEdit aims to modify knowledge related to specific content within complex images. + +Formally, let $f_{\theta}$ denote a pre-trained MLLM with parameters $\theta$ that can generate outputs based on combined image-text inputs. An editing sample is represented as $(i_e, t_e, y_e)$ , where $i_e$ denotes an image containing multiple entities, $t_e$ is the textual prompt, and $y_e$ represents the expected output for the input $(i_e, t_e)$ . The original output is given by $y_o = f_{\theta}(i_e, t_e)$ and $y_o \neq y_e$ . + +The objective of the FGVEdit task is to update the knowledge so that it accurately reflects $y_{e}$ when queried for specific fine-grained details, without influencing its performance on unrelated content. We define a knowledge editing operation $\mathcal{E}$ to update the model parameters as $\theta_{e} = \mathcal{E}(\theta, i_{e}, t_{e}, y_{e})$ . After applying $\mathcal{E}$ , the updated model $f_{\theta_{e}}$ should be capable of generating the corrected output for the specific input as $y_{e} = f_{\theta_{e}}(i_{e}, t_{e})$ . + +Fine-grained visual editing scope. The editing scope refers to the set of inputs that should be affected by knowl- + +edge editing. For in-scope inputs, which pertain to the same knowledge as the editing sample, the model's outputs should be updated accordingly. Conversely, for out-of-scope inputs, the model's behavior must remain unchanged to preserve its broader capabilities. For the fine-grained knowledge defined by the editing sample $(i_e, t_e, y_e)$ , there exists a fine-grained editing scope $S(i_e, t_e, y_e)$ . This editing scope is determined by both visual and textual information. For in-scope inputs on the same image, represented as $(i_e, t_{in}) \in S(i_e, t_e, y_e)$ , the original output is $y_{in} = f_\theta(i_e, t_{in})$ , and the edited model is required to produce a correction as $y_{in}' = f_{\theta_e}(i_e, t_{in})$ . For out-of-scope inputs on the same image, represented as $(i_e, t_{out}) \notin S(i_e, t_e, y_e)$ , the original output is $y_{out} = f_\theta(i_e, t_{out})$ , and the edited model maintains the original output as $y_{out} = f_{\theta_e}(i_e, t_{out})$ . + +# 3.2. MSCKE Framework + +To address the challenges of the proposed task, we introduce a Multimodal Scope Classifier-based Knowledge Editor (MSCKE) framework, as illustrated in Fig. 2. Structurally, our MSCKE framework comprises the following four major components: the Multimodal Edit Memory, the Multimodal Scope Classifier, the Base Model ( $f_{\mathrm{base}}$ ), and the Counterfactual Model ( $f_{\mathrm{cfr}}$ ). + +During the editing process, MSCKE keeps the basic model parameters while storing the updated knowledge in a model-agnostic multimodal editing memory as editing samples. These samples specify precise modifications or corrections to the model's knowledge. By maintaining these edit samples separately, MSCKE can reference them when necessary without altering the base model directly. + +When receiving a new input, the multimodal scope classifier evaluates the relevance of the input to the stored edit- + +ing samples based on visual and textual information. It then decides whether to invoke the edited information or rely on the response of the base model. This ensures that the editing is applied appropriately to relevant inputs, while unrelated knowledge remains stable. The multimodal scope classifier is defined as $f_{\mathrm{cls}}(i_{\mathrm{test}}, t_{\mathrm{test}}, i_e, t_e, y_e)$ . + +The index $k^*$ of the most similar editing sample and its corresponding similarity $\rho$ are determined as follows: + +$$ +k ^ {*} = \underset {k} {\arg \max } f _ {\mathrm {c l s}} \left(i _ {\text {t e s t}}, t _ {\text {t e s t}}, i _ {e} ^ {k}, t _ {e} ^ {k}, y _ {e} ^ {k}\right), \tag {1} +$$ + +$$ +\rho = f _ {\mathrm {c l s}} \left(i _ {\text {t e s t}}, t _ {\text {t e s t}}, i _ {e} ^ {k ^ {*}}, t _ {e} ^ {k ^ {*}}, y _ {e} ^ {k ^ {*}}\right), \tag {2} +$$ + +where $\rho \in [0,1]$ . Inputs with similarity less than 0.5 are considered out-of-scope samples, while those with similarity equal to or greater than 0.5 are considered in-scope samples. The final output of the model is: + +$$ +y _ {\text {t e s t}} = \left\{ \begin{array}{l l} f _ {\text {b a s e}} \left(i _ {\text {t e s t}}, t _ {\text {t e s t}}\right), & \rho < 0. 5, \\ f _ {\text {c f r}} \left(t _ {e} ^ {k ^ {*}}, y _ {e} ^ {k ^ {*}}, i _ {\text {t e s t}}, t _ {\text {t e s t}}\right), & \rho \geq 0. 5, \end{array} \right. \tag {3} +$$ + +where $f_{\mathrm{base}}$ denotes the base model, and $f_{\mathrm{cfr}}$ denotes the counterfactual model. We omit $i_e^{k^*}$ as an input to $f_{\mathrm{cfr}}$ because the test image $i_{\mathrm{test}}$ already provides the complete visual context needed for counterfactual reasoning. The role of the edit sample is primarily to guide this reasoning process through its textual components $(t_e^{k^*}, y_e^{k^*})$ . + +The base model is a well-trained multimodal large language model (e.g., MiniGPT-4 [35]), designed to process inputs that are unrelated to the editing samples in the multimodal edit memory. By freezing the parameters of the base model, we ensure that its general knowledge remains stable and uninfluenced by the editing samples. In contrast, the counterfactual model responds to inputs related to editing samples. Specifically, the counterfactual model accepts the input $(i_{\mathrm{test}}, t_{\mathrm{test}})$ while incorporating the corresponding editing samples from the multimodal edit memory. Guided by the prompts of the editing samples in the multimodal edit memory, the counterfactual model generates outputs that reflect the knowledge updates or corrections. The counterfactual model can either be a trainable model that shares the same output space as the base model or employ alternative editing methods that accurately reflect the new facts. + +# 3.3. Multimodal Scope Classifier + +To integrate both visual and textual modalities, we develop a new multimodal scope classifier. This allows the classifier to capture fine-grained similarities between input and editing samples. Such integration is essential, as relying solely on textual semantics may fail to capture true similarity when $t_e$ and $t_{\mathrm{test}}$ refer to different entities related to $i_e$ and $i_{\mathrm{test}}$ . By combining image and text information, our classifier enables a more comprehensive evaluation of the relevance between the query and editing samples. For both + +images and texts, we first map them into a unified feature space: + +$$ +h _ {e} ^ {i} = A _ {i} \left(E _ {i} \left(i _ {e}\right)\right), \quad h _ {e} ^ {t} = A _ {t} \left(E _ {t} \left(t _ {e}, y _ {e}\right)\right), \tag {4} +$$ + +$$ +h _ {\text {t e s t}} ^ {i} = A _ {i} \left(E _ {i} \left(i _ {\text {t e s t}}\right)\right), \quad h _ {\text {t e s t}} ^ {t} = A _ {t} \left(E _ {t} \left(t _ {\text {t e s t}}\right)\right), \tag {5} +$$ + +in which $E_{i}(\cdot)$ and $E_{t}(\cdot)$ are encoders for images and texts, respectively, which extract high-level features from the inputs. The alignment modules $A_{i}(\cdot)$ and $A_{t}(\cdot)$ project the encoded features into a shared multimodal space, ensuring that the image and text representations are compatible for fusion. To simplify the process, we utilize a pretrained CLIP [23] model for fine-tuning, which is capable of extracting aligned features from both images and text. We then compute the image-text fused features: + +$$ +z _ {e} = \operatorname {F u s i o n} \left(h _ {e} ^ {i}, h _ {e} ^ {t}\right), \tag {6} +$$ + +$$ +z _ {\text {t e s t}} = \operatorname {F u s i o n} \left(h _ {\text {t e s t}} ^ {i}, h _ {\text {t e s t}} ^ {t}\right), \tag {7} +$$ + +in which the Fusion $(\cdot, \cdot)$ function combines the visual and textual features to capture the cross-modal interactions. To simplify the training process, we employ the dot-product attention mechanism to extract the content of the image that is relevant to the text. Finally, the similarity is calculated from the fused features: + +$$ +\rho = \operatorname {S i m} \left(z _ {e}, z _ {\text {t e s t}}\right), \tag {8} +$$ + +where $\operatorname{Sim}(\cdot, \cdot)$ is a similarity function, such as cosine similarity or a learned distance metric, which quantifies the relevance between the query and the editing sample in the fused feature space. Note that our multimodal scope classifier can also accept text-only input by setting the image input to null, ensuring compatibility with text editing tasks. + +Training. The multimodal scope classifier is treated as a binary classifier to determine whether a given query is related to the editing samples stored in the multimodal edit memory. Thus, we train the classifier with a binary cross-entropy loss on a multimodal dataset, denoted as $D_{e} = \{(i_{e}^{k},t_{e}^{k},y_{e}^{k})\}$ . For each editing sample, the dataset provides both in-scope inputs $(i_{in}^{k},t_{in}^{k})$ and out-of-scope inputs $(i_{out}^{k},t_{out}^{k})$ . The loss function is defined as: + +$$ +\rho_ {\mathrm {i n}} ^ {k} = f _ {\mathrm {c l s}} \left(i _ {\mathrm {i n}} ^ {k}, t _ {\mathrm {i n}} ^ {k}, i _ {e} ^ {k}, t _ {e} ^ {k}, y _ {e} ^ {k}\right), \tag {9} +$$ + +$$ +\rho_ {\text {o u t}} ^ {k} = f _ {\text {c l s}} \left(i _ {\text {o u t}} ^ {k}, t _ {\text {o u t}} ^ {k}, i _ {e} ^ {k}, t _ {e} ^ {k}, y _ {e} ^ {k}\right), \tag {10} +$$ + +$$ +\mathcal {L} _ {c l s} = - \frac {1}{N} \sum_ {k = 1} ^ {N} \left[ \log \left(\rho_ {\text {i n}} ^ {k}\right) + \log \left(1 - \rho_ {\text {o u t}} ^ {k}\right) \right], \tag {11} +$$ + +where $N$ is the total number of training samples. + +# 4. FGVEdit Benchmark + +# 4.1. Evaluation Metrics + +In FGVEdit, we use four metrics to evaluate the performance of the editing method: specificity, reliability, locality, and generality. Among them, specificity is a new metric + +![](images/5bfe889ae1c5bdc26a40fc1a9aaac6f3cf91e74813486fa10cefb204b4fddd3c.jpg) +Figure 3. The construction pipeline of specificity dataset. For all questions related to a given image, the first question is selected as the classification criterion. Subsequent questions are assessed using ChatGPT. This classification process occurs in two stages: the first stage categorizes the data into in-scope and out-of-scope based on the image, while the second stage employs a text-based classification to identify and retain difficult data. Difficult data refers to instances where the category cannot be determined solely through textual analysis. + +tailored to our fine-grained visual editing task. The other three are common metrics that have been widely used in other multimodal knowledge editing benchmarks [4, 9]. + +# 4.1.1. Fine-Grained Visual Editing Metric + +Specificity. Previous multimodal knowledge editing metrics treat an image as a single entity and fail to account for the multiple entities present in the same image. These metrics primarily assess the impact of the editing method on the language model of MLLMs, rather than the nuanced interactions within the image itself. To this end, we introduce a new metric Specificity, denoted as $M_{\mathrm{specificity}}$ . Specificity evaluates whether the editing method can only influence the knowledge related to the editing target entity while keeping unrelated knowledge unchanged when editing complex images with multiple entities. + +To capture the different relationships between entities in an image, $M_{\mathrm{specificity}}$ is computed by two components: $M_{\mathrm{in}}^{\mathrm{v}}$ and $M_{\mathrm{out}}^{\mathrm{v}}$ . $M_{\mathrm{in}}^{\mathrm{v}}$ evaluates the ability to respond correctly to questions within the visual scope of the editing entity, while $M_{\mathrm{out}}^{\mathrm{v}}$ assesses the performance on questions outside the visual scope of the editing entity. These components are defined as follows: + +$$ +M _ {\text {s p e c i f i c i t y}} = \frac {1}{2} \left[ M _ {\text {i n}} ^ {v} + M _ {\text {o u t}} ^ {v} \right] \tag {12} +$$ + +$$ +M _ {i n} ^ {v} = \mathbb {E} _ {\left(t _ {i n}, y _ {i n}\right) \sim D _ {i n} ^ {v}} \left[ \mathbb {I} _ {f _ {\theta_ {e}} \left(i _ {e}, t _ {i n}\right) = y _ {i n}} \right] \tag {13} +$$ + +$$ +M _ {o u t} ^ {v} = \mathbb {E} _ {\left(t _ {o u t}, y _ {o u t}\right) \sim D _ {o u t} ^ {v}} \left[ \mathbb {I} _ {f _ {\theta_ {e}} \left(i _ {e}, t _ {o u t}\right)} = f _ {\theta} \left(i _ {e}, t _ {o u t}\right) \right] \tag {14} +$$ + +where $D_{in}^{v}$ refers to the dataset of in-visual-scope data, $D_{out}^{v}$ represents dataset of out-of-visual-scope data. Fig. 3 + +shows an example of the data used to evaluate specificity. + +# 4.1.2. Common Editing Metrics + +Reliability. Reliability refers to the ability of the editing methods to revise the predictions of MLLM from the original output $y_{o}$ to the expected output $y_{e}$ when provided with image and text input. This is a key objective in knowledge editing. We measure reliability as follows: + +$$ +M _ {r e l} = \mathbb {E} _ {\left(i _ {e}, t _ {e}, y _ {e}\right) \sim D _ {e d i t}} \left[ \mathbb {I} _ {f _ {\theta_ {e}} \left(i _ {e}, t _ {e}\right) = y _ {e}} \right] \tag {15} +$$ + +where $D_{edit}$ denotes the dataset of Reliability. + +Locality. Locality measures the stability of editing methods, ensuring that MLLM produces consistent outputs for data outside the editing scope before and after knowledge editing. We measure locality as follows: + +$$ +M _ {l o c} = \mathbb {E} _ {\left(t _ {l}, y _ {l}\right) \sim D _ {l o c}} \left[ \mathbb {I} _ {f _ {\theta_ {e}} \left(t _ {l}\right) = f _ {\theta} \left(t _ {l}\right)} \right] \tag {16} +$$ + +where $t_l$ represents a question out of editing scope, $y_l$ is the corresponding label, $D_{loc}$ refers to the dataset of Locality. + +Generality. In addition to accurately changing the prediction from the original $y_{o}$ to the desired $y_{e}$ , the edited multimodal model should also adapt its predictions within the editing scope, such as for rephrased questions. We define generality as follows: + +$$ +M _ {g e n} = \mathbb {E} _ {(t r) \sim D _ {g e n}} \left[ \mathbb {I} _ {f _ {\theta_ {e}} (i e, t r) = f _ {\theta_ {e}} (i e, t e)} \right] \tag {17} +$$ + +where $t_r$ represents the rephrased question of $t_e$ . $D_{gen}$ refers to the dataset of Generality. + +# 4.2. Dataset Construction + +We select Visual Question Answering (VQA) as the evaluation task and use VQAv2 [8] as the foundation for constructing our FGVEdit dataset. VQAv2 was originally designed to assess the ability to answer open-ended questions about images. It contains over 1 million question-answer pairs, each associated with an image, spanning a broad array of topics and requiring diverse reasoning capabilities. + +Specificity Dataset Construction. We select the first question of each image as the classification criterion. Additionally, it also serves as the editing sample for the calculation of reliability. To categorize the remaining questions, we manually construct two prompts and apply them twice. First, using the classification criterion and the corresponding image, we request GPT-4o-mini with a prompt "Using the provided image, analyze whether classification criterion logically entails this question, specifically, whether a change in Answer 1 would impact Answer 2. Please make a simple judgment (Yes, No, Maybe), and your response should not contain any other characters.", to classify the remaining questions, assessing whether the criterion logically entails the other questions. The results of this first classification yield two lists: In-scope samples, representing questions logically entailed by the first question, and Out-of-scope samples, representing questions that are not logically entailed by it. Next, based solely on the classification criterion, we request GPT-4o-mini again with a prompt "analyze whether classification criterion logically entails this question, specifically, whether a change in Answer 1 would impact Answer 2. Please make a simple judgment (Yes, No, Maybe), and your response should not contain any other characters.", to classify the questions in In-scope samples and Out-of-scope samples, generating the final results: "hard in-visual-scope" for questions within the visual context of the image and "hard out-of-visual-scope" for questions outside the visual context. + +Locality and Generality Dataset Construction. For the Locality dataset, we follow the approach in MMEdit [4] and use the NQ dataset [11], a benchmark for model stability, which includes question-answer pairs that fall outside the editing scope. For the Generality dataset, we request GPT-4o-mini with a prompt like "Please rewrite the following question differently. Do not include the original question" and assess whether the model can accurately respond to the generated questions within the editing scope. + +# 4.3. Data Statistics + +For each sample, we generate instances of specificity, locality, and generality data to ensure dataset completeness. Due to the limited availability of specificity data that includes both hard in-visual-scope and hard out-of-visual-scope data, we retain specificity data that contains either one of these categories. This approach maintains the task's chal + +lenge while ensuring sufficient data coverage. Non-hard data is used as a substitute for the other category when necessary. The final dataset consists of 11,112 samples, which are then split into training and testing sets in a 3:1 ratio. + +# 5. Experiments and Results + +# 5.1. Experimental Setup + +Base MLLMs. We utilize BLIP-2 OPT [13] and MiniGPT-4 [35] as the base MLLMs for Editing. BLIP-2 OPT is a vision-language model that combines a pre-trained Vision Transformer (ViT) with a lightweight Querying Transformer (Q-Former). It uses a two-stage pre-training strategy: the first stage focuses on learning vision-language representations, while the second emphasizes generative learning, allowing for efficient zero-shot image-to-text generation. MiniGPT-4 integrates a pre-trained ViT with a Q-Former to effectively align visual features with the Vicuna LLM. This architecture enables it to process and generate text based on visual inputs, enhancing its performance on complex vision-language tasks. + +Baselines. Since there are currently no methods specifically designed for FGVEdit, we apply the LLM editing methods to the LLM within MLLMs as baselines for evaluation. Fine-tuning adjusts a pre-trained model parameters to incorporate new information. While effective, it is computationally intensive and may unintentionally degrade unrelated behaviors. IKE [33] is a paradigm that utilizes contextual learning to modify specific factual knowledge in LLMs without altering their parameters. By providing carefully crafted demonstrations as input prompts, IKE guides the model to update its responses to reflect new or corrected information, avoiding the computational costs of fine-tuning. SERAC [20] stores new facts in an additional memory, and employs a scope classifier to determine whether the input is related to the facts stored in memory. By utilizing a counterfactual model to answer questions related to the new facts while preserving the original model, SERAC enables targeted modifications without accessing the underlying model parameters, thereby enhancing computational efficiency. MEND [19] leverages an auxiliary neural network to convert fine-tuning gradients into targeted updates, facilitating rapid edits while maintaining performance on unrelated inputs. It serves as a scalable solution for model improvement. + +Implementation details. For SERAC, MEND, and IKE, we leverage the code provided by the EasyEdit [27] toolkit, while fine-tuning is from KEBench [9]. By default, the implementation of our method is built on the SERAC [20] framework, utilizing the EasyEdit [27] toolkit. For the multimodal scope classifier, we employ the CLIP-ViT-B/32 model to encode and align image-text pairs. During the fusion stage, we utilize a parameter-free dot product attention + +
BLIP-2 OPTMiniGPT-4
SpecificitySpecificity
FT-LLM24.2135.02
FT-Visual16.5630.53
IKE20.0725.26
SERAC31.9237.85
MSCKE61.60 (↑29.68)57.20 (↑19.35)
MEND65.8567.39
MSCKE-MEND68.38 (↑2.53)71.98 (↑4.59)
+ +Table 1. Fine-grained performance comparison of various knowledge editing methods on BLIP-2 OPT and MiniGPT-4. Results are presented as percentages. $\uparrow$ indicates the improvement of our method over the baseline. + +
BLIP-2 OPT
ReliabilityLocalityGenerality
FT-LLM100.076.9399.96
FT-Visual99.68100.099.21
IKE99.8948.4998.02
SERAC93.0899.9096.83
MSCKE99.13 (↑6.05)100.0 (↑0.10)98.56 (↑1.73)
MEND97.0098.6096.40
MSCKE-MEND97.40 (↑0.40)100.0 (↑1.40)96.50 (↑0.10)
MiniGPT-4
ReliabilityLocalityGenerality
FT-LLM93.3986.2693.35
FT-Visual93.39100.091.87
IKE100.052.4598.88
SERAC99.50100.092.90
MSCKE99.50 (↑0.00)100.0 (↑0.00)93.00 (↑0.10)
MEND94.8598.5894.82
MSCKE-MEND97.05 (↑2.20)100.0 (↑1.42)96.70 (↑1.88)
+ +Table 2. Performance of various knowledge editing methods across common metrics. Results are presented as percentages. $\uparrow$ indicates the improvement of our method over the baseline. + +mechanism to integrate features. To highlight the extensibility of our approach, we replace the original counterfactual model with MEND [19] and introduce an enhanced version, called MSCKE-MEND. This demonstrates how easily our method can be adapted to incorporate advanced techniques, resulting in improved performance. + +# 5.2. Results + +Performance on Fine-Grained metric. Table 1 shows the fine-grained performance of various methods on the FGVEdit benchmark. For the specialized metric Specificity, IKE and SERAC show notably limited performance, as they rely exclusively on textual data. Our MSCKE, which incorporates a multimodal classifier leveraging both visual and textual information, significantly outperforms SERAC. Specifically, for BLIP-2 OPT, Specificity increases from 31.92 to 61.60, and for MiniGPT-4, from 37.85 to 57.20. While MEND, which updates both visual and language + +![](images/96c13a782e33dd63651975df869b1f3f9735892cff113f686fbaf91b8a4d3526.jpg) + +![](images/b832e951d229c6a22a8e6655699786c0af868f02e62c41b559703b3249c2a07a.jpg) +Figure 4. Comparison of classification performance between MSCKE's multimodal scope classifier and SERAC's classifier on text and image samples. + +modules, performs better than other baselines, MSCKEMEND further enhances performance, improving Specificity from 65.85 to 68.38 for BLIP-2 OPT and from 67.39 to 71.98 for MiniGPT-4. These results underline the effectiveness of our approach in tackling complex multimodal editing tasks. + +Performance on common metrics. Table 2 shows the results on three common metrics. Most editing methods perform excellently in terms of Reliability. However, our methods, MSCKE and MSCKE-MEND, show consistent improvements over their respective baselines, SERAC and MEND, demonstrating superior reliability. For Locality, both FT-LLM and IKE exhibit subpar performance due to the absence of relevant constraints during the editing process. In contrast, only FT-Visual and our MSCKE achieve a perfect score of 100 across both models. This indicates that our method excels in precisely targeting edits without inadvertently altering unrelated information in the text. For Generality, although our method is optimized for multimodal tasks, it outperforms SERAC and MEND even in purely text-based tasks, showcasing its versatility. + +# 5.3. Effect of Multimodal Scope Classifier + +Next, we analyze the performance of our multimodal scope classifier compared to the SERAC scope classifier. Our classifier can process both text-only and text-image inputs, whereas the SERAC classifier is limited to text-only input. We evaluate their editing scope classification performance on two tasks: a text editing scope classification and a text-image editing scope classification task, both using our + +
sourcetargetReliabilityLocalityGeneralitySpecificity
methodmodelmethodmodel
MSCKEBLIP-2 OPT//99.1310098.5661.60
MSCKEBLIP-2 OPTMSCKE-MENDBLIP-2 OPT97.40/97.40100/10096.52/96.5068.35/68.38
MSCKEBLIP-2 OPTMSCKEMiniGPT-499.50/99.50100/10093.11/93.0057.09/57.20
MSCKEBLIP-2 OPTMSCKE-MENDMiniGPT-497.10/97.05100/10096.82/96.7072.16/71.98
+ +Table 3. Transferability of the trained multimodal scope classifier to different editing methods and base models. Results are presented as transferred/retrained, where applicable. + +
ReliabilityT-GeneralityT-LocalityM-Locality
SERAC99.9099.90100.02.91
MSCKE99.8095.60100.098.08
MEND98.5197.5199.9496.65
MSCKE-MEND98.2092.10100.099.23
+ +Table 4. Experimental results on MMEdit. Results are presented as percentages. + +FGVEdit dataset. The results are shown in Fig. 4. + +For text editing scope classification, both classifiers receive text-only inputs and achieve high accuracy, demonstrating the robustness of our multimodal scope classifier in maintaining strong classification performance for text-based tasks. However, in the text-image editing scope classification, the SERAC scope classifier shows limited effectiveness, largely classifying most samples as in-scope. This limitation arises from the fine-grained nature of our FGVEdit dataset, where both in-scope and out-of-scope samples are closely related to the text in the editing sample. In contrast, our multimodal scope classifier excels by leveraging both textual and visual information, significantly improving classification accuracy. + +# 5.4. Transferability of Multimodal Scope Classifier + +To further highlight the advantages of our multimodal scope classifier, we investigate its transferability across different editing methods and base models. Specifically, we transfer the multimodal scope classifier, initially trained with FGVEdit dataset on BLIP-2 OPT, to various settings, such as using MEND as the counterfactual model and MiniGPT4 as the base model. The results are presented in Table 3. Our multimodal scope classifier demonstrates exceptional transferability. In all experiments, the transferred classifier performs nearly identically to a classifier retrained for each specific setting. This decoupled design and robust transferability enable our method to be quickly adapted for deployment across new MLLMs. + +# 5.5. Experimental Results on Other Benchmark + +To verify the adaptability of our method, we also test MSCKE on another benchmark, i.e., MMEdit [4]. Table 4 shows the results of editing BLIP-2 OPT on the VQA task. Notably, our MSCKE also achieves outstanding performance across MMEdit. Additionally, some baselines (e.g., MEND) demonstrate strong performance across all + +![](images/8428529f23eb772bfbfb16be5603aa7565720653ab45a81526a615e86c751de9.jpg) +Figure 5. An editing case of MSCKE on FGVEdit. The outputs before and after knowledge editing are shown. + +metrics in MMEdit, highlighting both the challenges and the importance of the task we propose. The complete results are provided in Section D of the supplementary material. + +# 5.6. Case Study + +Fig. 5 shows an editing case obtained by MSCKE on FGVEdit. Before editing, the MLLM incorrectly identified the "horse" as "zebra". After updating this knowledge by MSCKE, MLLM revises its answers to the in-scope question (as shown in the image, people are looking at the target animal). For the question involving unrelated content (i.e., the color of the stripes on the women's shirt), the response remains unchanged. + +# 6. Conclusion + +This paper identifies the unique challenges of knowledge editing in MLLMs by introducing a visual-oriented, fine-grained multimodal knowledge editing task. To drive progress in this area, we present the FGVEdit benchmark and propose the MSCKE framework by specifically designing a multimodal scope classifier. Extensive experiments reveal that existing methods perform poorly on the FGVEdit benchmark, indicating their inability to effectively address the task's challenges. In contrast, MSCKE achieves strong performance, demonstrating its effectiveness and potential as a robust solution for advancing multimodal knowledge editing. We hope this work inspires further research in this direction, fostering advancements in fine-grained multimodal knowledge editing. + +# Acknowledgements + +Funded by National Natural Science Foundation of China (No.62472138, No.72188101, No.62020106007) and Anhui Postdoctoral Scientific Research Program Foundation (No.2022B657). + +# References + +[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 2 +[2] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. NeurIPS, 2022. 1, 2 +[3] Tom B Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. In NeurIPS, 2020. 2 +[4] Siyuan Cheng, Bozhong Tian, Qingbin Liu, Xi Chen, Yongheng Wang, Huajun Chen, and Ningyu Zhang. Can we edit multimodal large language models? In EMNLP, 2023. 1, 2, 5, 6, 8 +[5] Damai Dai, Li Dong, Yaru Hao, Zhifang Sui, Baobao Chang, and Furu Wei. Knowledge neurons in pretrained transformers. In ACL, 2022. 2 +[6] N De Cao, W Aziz, and I Titov. Editing factual knowledge in language models. In EMNLP, 2021. 1, 2 +[7] Qingxiu Dong, Damai Dai, Yifan Song, Jingjing Xu, Zhifang Sui, and Lei Li. Calibrating factual knowledge in pretrained language models. In EMNLP, 2022. 2 +[8] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In CVPR, 2017. 6 +[9] Han Huang, Haitian Zhong, Qiang Liu, Shu Wu, Liang Wang, and Tieniu Tan. Kebench: A benchmark on knowledge editing for large vision-language models. arXiv preprint arXiv:2403.07350, 2024. 1, 2, 5, 6 +[10] Zeyu Huang, Yikang Shen, Xiaofeng Zhang, Jie Zhou, Wenge Rong, and Zhang Xiong. Transformer-patcher: One mistake worth one neuron. In ICLR, 2023. 2 +[11] Tom Kwiatkowski, Jennimaria Palomaki, Olivia Redfield, Michael Collins, Ankur Parikh, Chris Alberti, Danielle Epstein, Illia Polosukhin, Jacob Devlin, Kenton Lee, Kristina Toutanova, Llion Jones, Matthew Kelcey, Ming-Wei Chang, Andrew M. Dai, Jakob Uszkoreit, Quoc Le, and Slav Petrov. Natural questions: A benchmark for question answering research. Transactions of the Association for Computational Linguistics, 7:453–466, 2019. 6 +[12] Angeliki Lazaridou, Adhi Kuncoro, Elena Gribovskaya, Devang Agrawal, Adam Liska, Tayfun Terzi, Mai Gimenez, Cyprien de Masson d'Autume, Tomas Kocisky, Sebastian + +Ruder, et al. Mind the gap: Assessing temporal generalization in neural language models. NeurIPS, 34:29348-29363, 2021. 2 +[13] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730–19742. PMLR, 2023. 1, 6 +[14] Jiaqi Li, Miaozeng Du, Chuanyi Zhang, Yongrui Chen, Nan Hu, Guilin Qi, Haiyun Jiang, Siyuan Cheng, and Bozhong Tian. Mike: A new benchmark for fine-grained multimodal entity knowledge editing. arXiv preprint arXiv:2402.14835, 2024.1,2 +[15] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. NeurIPS, 2024. 1 +[16] Aman Madaan, Niket Tandon, Peter Clark, and Yiming Yang. Memory-assisted prompt editing to improve gpt-3 after deployment. In EMNLP, 2022. 1, 2 +[17] Kevin Meng, David Bau, Alex Andonian, and Yonatan Belinkov. Locating and editing factual associations in gpt. NeurIPS, 2022. 2 +[18] Kevin Meng, Arnab Sen Sharma, Alex J Andonian, Yonatan Belinkov, and David Bau. Mass-editing memory in a transformer. In ICLR, 2023. 2 +[19] Eric Mitchell, Charles Lin, Antoine Bosselut, Chelsea Finn, and Christopher D Manning. Fast model editing at scale. In ICLR, 2022. 1, 2, 6, 7 +[20] Eric Mitchell, Charles Lin, Antoine Bosselut, Christopher D Manning, and Chelsea Finn. Memory-based model editing at scale. In ICML, 2022. 1, 2, 6 +[21] Haowen Pan, Yixin Cao, Xiaozhi Wang, Xun Yang, and Meng Wang. Finding and editing multi-modal neurons in pre-trained transformers, 2024. 2 +[22] Haowen Pan, Xiaozhi Wang, Yixin Cao, Zenglin Shi, Xun Yang, Juanzi Li, and Meng Wang. Precise localization of memories: A fine-grained neuron-level knowledge editing technique for llms, 2025. 1 +[23] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 2, 4 +[24] Anton Sinitsin, Vsevolod Plokhotnyuk, Dmitry Pyrkin, Sergei Popov, and Artem Babenko. Editable neural networks. In ICLR, 2019. 1 +[25] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 2 +[26] Maria Tsimpoukelli, Jacob L Menick, Serkan Cabi, SM Eslami, Oriol Vinyals, and Felix Hill. Multimodal few-shot learning with frozen language models. NeurIPS, 2021. 1 +[27] Peng Wang, Ningyu Zhang, Xin Xie, Yunzhi Yao, Bozhong Tian, Mengru Wang, Zekun Xi, Siyuan Cheng, Kangwei Liu, Guozhou Zheng, et al. EASYedit: An easy-to-use knowledge editing framework for large language models. arXiv preprint arXiv:2308.07269, 2023. 6 + +[28] Zhengyuan Yang, Linjie Li, Jianfeng Wang, Kevin Lin, Ehsan Azarnasab, Faisal Ahmed, Zicheng Liu, Ce Liu, Michael Zeng, and Lijuan Wang. Mm-react: Prompting chatgpt for multimodal reasoning and action. arXiv preprint arXiv:2303.11381, 2023. 1 +[29] Yunzhi Yao, Peng Wang, Bozhong Tian, Siyuan Cheng, Zhoubo Li, Shumin Deng, Huajun Chen, and Ningyu Zhang. Editing large language models: Problems, methods, and opportunities. In EMNLP, 2023. 1 +[30] Shukang Yin, Chaoyou Fu, Sirui Zhao, Ke Li, Xing Sun, Tong Xu, and Enhong Chen. A survey on multimodal large language models. arXiv preprint arXiv:2306.13549, 2023. 1, 2 +[31] Junzhe Zhang, Huixuan Zhang, Xunjian Yin, Baizhou Huang, Xu Zhang, Xinyu Hu, and Xiaojun Wan. Mcmke: A fine-grained multimodal knowledge editing benchmark emphasizing modality consistency. arXiv preprint arXiv:2406.13219, 2024. 1, 2 +[32] Ningyu Zhang, Yunzhi Yao, Bozhong Tian, Peng Wang, Shumin Deng, Mengru Wang, Zekun Xi, Shengyu Mao, Jintian Zhang, Yuansheng Ni, et al. A comprehensive study of knowledge editing for large language models. arXiv preprint arXiv:2401.01286, 2024. 1 +[33] Ce Zheng, Lei Li, Qingxiu Dong, Yuxuan Fan, Zhiyong Wu, Jingjing Xu, and Baobao Chang. Can we edit factual knowledge by in-context learning? In EMNLP, 2023. 1, 2, 6 +[34] Chen Zhu, Daliang Li, Felix Yu, Manzil Zaheer, Sanjiv Kumar, Srinadh Bhojanapalli, and Ankit Singh Rawat. Modifying memories in transformer models. In ICML, 2021. 1 +[35] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. MiniGPT-4: Enhancing vision-language understanding with advanced large language models. In ICLR, 2024. 1, 4, 6 \ No newline at end of file diff --git a/ICCV/2025/Visual-Oriented Fine-Grained Knowledge Editing for MultiModal Large Language Models/images.zip b/ICCV/2025/Visual-Oriented Fine-Grained Knowledge Editing for MultiModal Large Language Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..40f0e5be2f27cded69ed07248724e2a3f298a5fd --- /dev/null +++ b/ICCV/2025/Visual-Oriented Fine-Grained Knowledge Editing for MultiModal Large Language Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f6f930db5ff27b5369ab52e29697b0ce813637b5b2475946e7a61108ffa3935a +size 619454 diff --git a/ICCV/2025/Visual-Oriented Fine-Grained Knowledge Editing for MultiModal Large Language Models/layout.json b/ICCV/2025/Visual-Oriented Fine-Grained Knowledge Editing for MultiModal Large Language Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..9f41f292d64b7904eb76f4c2d86be5c730535f67 --- /dev/null +++ b/ICCV/2025/Visual-Oriented Fine-Grained Knowledge Editing for MultiModal Large Language Models/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ab293bf5da8399e081e74ee415a48928c949ae56098515f4236ed22c65606b9d +size 353431 diff --git a/ICCV/2025/Visual-RFT_ Visual Reinforcement Fine-Tuning/b1758928-1f95-4aa1-882d-f17b8d4c58c5_content_list.json b/ICCV/2025/Visual-RFT_ Visual Reinforcement Fine-Tuning/b1758928-1f95-4aa1-882d-f17b8d4c58c5_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..b789d90013a7460f82b2943b9fa4f06207f3fafb --- /dev/null +++ b/ICCV/2025/Visual-RFT_ Visual Reinforcement Fine-Tuning/b1758928-1f95-4aa1-882d-f17b8d4c58c5_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f588520dba4526f21ffe6bc838eb1d3c6c95d2ac6e1adcd7a9302bb2fad37e37 +size 91220 diff --git a/ICCV/2025/Visual-RFT_ Visual Reinforcement Fine-Tuning/b1758928-1f95-4aa1-882d-f17b8d4c58c5_model.json b/ICCV/2025/Visual-RFT_ Visual Reinforcement Fine-Tuning/b1758928-1f95-4aa1-882d-f17b8d4c58c5_model.json new file mode 100644 index 0000000000000000000000000000000000000000..e69ef7a4df9b9e564e515d205b449d3a0fe349f6 --- /dev/null +++ b/ICCV/2025/Visual-RFT_ Visual Reinforcement Fine-Tuning/b1758928-1f95-4aa1-882d-f17b8d4c58c5_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:91990280b975e89923643aa3ca6e34d86e958949e107f8c56fecbc2d8327e95e +size 114610 diff --git a/ICCV/2025/Visual-RFT_ Visual Reinforcement Fine-Tuning/b1758928-1f95-4aa1-882d-f17b8d4c58c5_origin.pdf b/ICCV/2025/Visual-RFT_ Visual Reinforcement Fine-Tuning/b1758928-1f95-4aa1-882d-f17b8d4c58c5_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..246283d29d148a10763ae051178dd4bfbcc67d94 --- /dev/null +++ b/ICCV/2025/Visual-RFT_ Visual Reinforcement Fine-Tuning/b1758928-1f95-4aa1-882d-f17b8d4c58c5_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:127677128d2e8cf1be7000ba7c82a34db8eca0081edfa81825124ac47bcb5933 +size 3515005 diff --git a/ICCV/2025/Visual-RFT_ Visual Reinforcement Fine-Tuning/full.md b/ICCV/2025/Visual-RFT_ Visual Reinforcement Fine-Tuning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..1043ee9c46c01729fec7b8fcd308a6ec3c527bc3 --- /dev/null +++ b/ICCV/2025/Visual-RFT_ Visual Reinforcement Fine-Tuning/full.md @@ -0,0 +1,414 @@ +# Visual-RFT: Visual Reinforcement Fine-Tuning + +Ziyu Liu $^{1,2*}$ Zeyi Sun $^{1,2*}$ Yuhang Zang $^{2\boxtimes}$ Xiaoyi Dong $^{2,3}$ + +Yuhang Cao² Haodong Duan² Dahua Lin²,³,⁴ Jiaqi Wang² + +$^{1}$ Shanghai Jiao Tong University $^{2}$ Shanghai Artificial Intelligence Laboratory + +3The Chinese University of Hong Kong 4CPII under InnoHK + +{liuziyu77, szy2023}@sjtu.edu.cn, {zangyuhang, wangjiaqi}@pjlab.org.cn + +https://github.com/Liuziyu77/Visual-RFT + +# Abstract + +Reinforcement Fine-Tuning (RFT) in Large Reasoning Models like OpenAI o1 learns from feedback on its answers, which is especially useful in applications when fine-tuning data is scarce. Recent open-source work like DeepSeek-R1 demonstrates that reinforcement learning with verifiable reward is possibly one key direction in reproducing o1. While the R1-style model has demonstrated success in language models, its application in multi-modal domains remains under-explored. This work introduces Visual Reinforcement Fine-Tuning (Visual-RFT), which further extends the application areas of RFT on visual tasks. Specifically, Visual-RFT first uses Large Vision-Language Models (LVLMs) to generate multiple responses containing reasoning tokens and final answers for each input, and then uses our proposed visual perception verifiable reward functions to update the model via the policy optimization algorithm such as Group Relative Policy Optimization (GRPO). We design different verifiable reward functions for different perception tasks, such as the Intersection over Union (IoU) reward for object detection. Experimental results on fine-grained image classification, few-shot object detection, reasoning grounding, as well as open-vocabulary object detection benchmarks show the competitive performance and advanced generalization ability of Visual-RFT compared with Supervised Fine-tuning (SFT). For example, Visual-RFT improves accuracy by $24.3\%$ over the baseline in one-shot fine-grained image classification with around 100 samples. In few-shot object detection, Visual-RFT also exceeds the baseline by 21.0 on COCO's 4-shot setting and 15.4 on LVIS. Our Visual-RFT represents a paradigm shift in fine-tuning LVLMs, offering a data-efficient, reward-driven approach that enhances reasoning and adaptability for domain-specific tasks. + +![](images/4df96540ec403dd1f9fb84726aed36a0168987f12afd8cb46cefe936499ff748.jpg) +(a) OV Results on COCO + +![](images/992a2d60bc4f30f01df86107821bed794ad7ef2a97562a7100d0aa188a67aad6.jpg) +(b) OV Results on LVIS +(c) Few-shot Results on COCO + +![](images/b5abf5bce379cd9bd33e66a1abd8b20e22fc38fcfd8245b08d01c9923a7ab4c7.jpg) + +![](images/c10c16329891bc684ec2fb377c599822455a7f923ee2f0b3264ddf32f01eafd1.jpg) +(d) Reasoning Grounding on LISA + +![](images/72e21d7f799efcfbc83d8f454a2748e9d78777930be9cad73664c77ecfde2410.jpg) +(e) CLS on Fine-Grained Datasets +Figure 1. Our Visual Reinforcement Fine-Tuning (Visual-RFT) performs better than previous Supervised Fine-Tuning (SFT) on a variety of tasks, such as Open Vocabulary(OV)/Few-shot Detection, Reasoning Grounding, and Fine-grained Classification. + +# 1. Introduction + +Large Reasoning Models (LRMs) such as OpenAI o1 [7] represent frontier AI models designed to spend more time "thinking" before answering, and achieving excellent reasoning abilities. One impressive capability of OpenAI o1 is Reinforcement Fine-Tuning (RFT) $^{1}$ , which efficiently fine-tune the model with merely dozens to thousands of samples to excel at domain-specific tasks. Although the implementation details of o1 are not publicly available, recent open-source studies like DeepSeek R1 [4] reveal one promising direction in reproducing o1 is Verifiable Rewards [4, 14, 39]: the reward score in reinforcement learning is directly determined by pre-defined rules, rather than predicted by the separate reward model [19, 28, 49] trained on preference data. + +A primary distinction between the RFT and Previous Supervised Fine-Tuning (SFT) lies in data efficiency. Previous + +![](images/a24ddca00e136a98dede582cb00053cb916a7c27b9f17d846e91eaaccbafb865.jpg) +Figure 2. Overview of Visual-RFT. Compared to the (a) Visual Instruction Tuning that is data-hungry, (b) our Visual Reinforcement Fine-Tuning (Visual-RFT) is more data efficient with limited data. (c) We successfully empower Large Vision-Language Models (LVLMs) with RFT on a series of multi-modal tasks, and present examples of the model's reasoning process at the bottom. + +SFT paradigm (see Fig. 2 (a)) directly imitates the "ground-truth" answers provided in the high-quality, curated data, thus relying on a large amount of training data. By contrast, RFT evaluates the model's responses and adjusts based on whether they're correct, helping it learn through trial and error. Thus, RFT is particularly useful in domains where data is scarce [7, 26]. However, a previous common sense is that RFT is applied merely in tasks like scientific (e.g., mathematics) and code generation. That's because math and coding exhibit clear and objective final answers or test cases, making their rewards relatively straightforward to verify. In this paper, we demonstrate that RFT can be applied beyond math and code domains to visual perception tasks. Specifically, we introduce Visual Reinforcement Fine-Tuning (Visual-RFT), which successfully extends RFT to empower Large Vision-Language Models (LVLMs) in various multi-modal tasks (see Fig. 1), such as few-shot classification and open-vocabulary object detection. + +To extend RFT on visual tasks, we present the implementation details of Visual-RFT in Fig. 2 (b). For each input, Visual-RFT uses Large Vision-Language Models (LVLMs) to generate multiple responses (trajectories) that contain the reasoning tokens and final answers. Crucially, we define task-specific, rule-based verifiable reward functions to guide policy optimization, such as GRPO [33], in updata + +ing the model. For instance, we propose the Intersection over Union (IoU) reward for the object detection task. Our Visual-RFT contrasts with SFT, which relies on memorizing correct answers. Instead, our approach explores different possible solutions and learns to optimize for a desired outcome defined by our verified reward function. It's about discovering what works best, not just mimicking predefined answers. Our approach shifts the training paradigm from data scaling in SFT to the strategic design of variable reward functions tailored to specific multi-modal tasks. As shown in Fig. 2 (c), the synergistic combination of verifiable rewards and visual perception abilities (e.g., detection, grounding, classification) allows our model to achieve rapid and data-efficient mastery of new concepts, facilitated by a detailed reasoning process. + +We validate the effectiveness of Visual-RFT on the following tasks. In fine-grained image classification, the model utilizes its advanced reasoning capabilities to analyze fine-grained categories with high precision. In the one-shot setting with extremely limited data (e.g., around 100 samples), Visual-RFT boosts the accuracy with $24.3\%$ over the baseline, while SFT dropped by $4.3\%$ . In few-shot experiments, Visual-RFT also demonstrates exceptional performance with minimal training data, showcasing superior few-shot learning capabilities compared to SFT. In + +![](images/d09cc525464bf44cbce87c0af8893a725e940c61df0b3ba814fee4044b5fdadf.jpg) +Figure 3. Framework of Visual-RFT. Given the question and visual image inputs, the policy model generates multiple responses containing reasoning steps. Then the verifiable reward such as IoU reward and CLS reward is used with the policy gradient optimization algorithm to update the policy model. + +reasoning grounding, Visual-RFT excels in the LISA [13] dataset, which heavily relies on reasoning, outperforming specialized models like GroundedSAM [20]. Furthermore, in open vocabulary object detection, Visual-RFT quickly transfers recognition capabilities to new categories, including rare categories in LVIS [5], showing strong generalization. Specifically, the 2B model achieves nAP improvements from 9.6 to 27.9 on new classes of COCO [17] and from 2.7 to 20.7 on selected rare classes of LVIS [5]. These diverse visual perception tasks not only highlight Visual-RFT's robust generalization capabilities in visual recognition but also underscore the crucial role of reinforcement learning in enhancing visual perception and reasoning. + +In summary, our key contributions are as follows: + +(1) We introduce Visual Reinforcement Fine-tuning (Visual-RFT), which extends reinforcement learning with verifiable rewards on visual perception tasks that are effective with limited data for fine-tuning. +(2) We design different verifiable rewards for different visual tasks that enable efficient, high-quality reward computation at a negligible cost. This allows the seamless transfer of DeepSeek R1's style reinforcement learning to LVLMs. +(3) We conduct extensive experiments on various visual perception tasks, including fine-grained image classification, few-shot object detection, reasoning grounding, and open vocabulary object detection. On all the settings, VisualRFT achieves remarkable performance improvements, significantly surpassing the supervised fine-tuning baselines. + +# 2. Related Work + +Large Vision Language Models (LVLMs) like GPT-4o [25] achieves excellent visual understanding ability by integrating both visual and textual data. This integration enhances the models' ability to understand complex multi-modal inputs and enables more advanced AI systems [15, 18, 40, 52] capable of processing and responding to both images and text. Generally, the training of LVLMs involves two steps: (a) pre-training and (b) posttraining which contains supervised fine-tuning and reinforcement learning. Post-training is crucial in improving the model's response quality, instruction following, and reasoning abilities. While there has been significant research on using reinforcement learning to enhance LLMs during post-training [1, 3, 27, 30, 34, 35, 38, 44, 48, 62, 63], the progress for LVLMs has been slower. In this paper, we propose Visual-RFT, which used GRPO-based reinforcement algorithms and verifiable reward during the post-training phase to enhance the model's visual perception and reasoning capabilities. + +Reinforcement Learning Recently, with the emergence of reasoning models like OpenAI's o1 [7], the research focus in Large Language Models (LLMs) has increasingly shifted towards enhancing the models' reasoning capabilities through reinforcement learning (RL) techniques. Studies have explored improving LLMs' performance in reasoning tasks such as solving mathematical problems [2, 22, 33, 43, 45] and coding [6, 8, 51, 54]. A notable breakthrough in this area is Deepseek-R1-Zero [4], which introduced a + +![](images/57720b7418f0befadc4c81caefb5f4ead4d2344dd7154be9fd99bc9f1866ba29.jpg) + globe-flower + +![](images/51dbe52ee5faa3c63dc7cdcc7fef5e6da0b3de142834334fa1c5a1d0eed55044.jpg) + +![](images/d82cb71eecc72002074a29753439e8bee0895c3e8aa0c7a592d4e3edb7f30f4e.jpg) +Visual RFT + +![](images/0ceccc83fefb02e0ffc14502919095a2900a427698acf2c23be8ab85e6bacf01.jpg) +This is an image containing a flower. Please identify the species of the flower based on the image + +![](images/5d0a3058f97f574c7a4737608d81e141637341b50b1360a3eb047a5ae5514bbe.jpg) + +![](images/a2a7bf9fa92237876609d62a08c58c65c50979ef9b76ace2188cbfacc289598d.jpg) + + The image shows a yellow buttercup flower with a green stem and green leaves in the background. Buttercups are a type of flowering plant in the buttercup family. They are known for their bright yellow flowers and are commonly found in meadows and grasslands. + + buttercup + +![](images/4f3ecd6a9af7e2388b443735c60ab9f3f6e42873bd945685b0bedf5b12ebb8ac.jpg) +Figure 4. Qualitative results of Fine-Grained Image Classification. The thinking process significantly improves the reasoning ability of LVLMs, leading to higher image classification performance. + +![](images/cc8d67efb14346790c03444758af878e5abf1c9d510b56e68aacaf731ac69b71.jpg) + hedgehog + +![](images/8ed5c9feec2148fa3d7d95677496559de4e61b623a7652e7c194c20512e05f86.jpg) +SFT + +SFT + +![](images/9502eef6b4964bfeb9eb2fee433043af86fc0b190212da3fe3b626c898e3975b.jpg) +This is an image containing a pet. Please identify the species of the pet based on the image + +![](images/6626af9be26fe8e1f2537d250ce8e8137a25f2a83b54cf836d05839432044726.jpg) + +![](images/5db22543c15c404fa8005fd8ddcd7c22a169bd28ce0190e9461964c2f6dabba2.jpg) + + The dog in the image appears to be a Japanese Chin, also known as a Japanese Spaniel. This breed is known for its small size, fluffy coat, and distinctive black and white coloration. The Japanese Chin is a popular pet and is known for its friendly and affectionate nature. + + Japanese Chin (or Japanese Spaniel) + +![](images/e9918b99e5b0474fdde5d45b464fd216b92459afc582f1770a325c2f4e525613.jpg) + +Table 1. Prompts used to construct the dataset. We have listed the detection prompt and classification prompt separately. + +Detection Prompt: Detect all objects belonging to the category '\{category\}' in the image, and provide the bounding boxes (between 0 and 1000, integer) and confidence... The output answer format should be as follows: $< \text{think}>$ ... $$ [Position': [x1, y1, x2, y2], 'Confidence': number, ...] Please strictly follow the format. + +Classification Prompt: This is an image containing a plant. Please identify the species of the plant based on the image... The output answer format should be as follows: ... species name Please strictly follow the format. + +new approach to achieving robust reasoning capabilities using RL merely, eliminating the supervised fine-tuning (SFT) stage. However, current research on RL-based reasoning has largely been confined to the language domain, with limited exploration of its application in multi-modal settings. For LVLMs, RL has primarily been used for tasks like mitigating hallucinations and aligning models with human preference [21, 36, 37, 46, 47, 55, 60, 61], but there remains a significant gap in research focusing on enhancing reasoning and visual perception of Large Vision Language Models. To address this gap, our work introduces a novel reinforcement fine-tuning strategy Visual-RFT, applying verifiable rewards with GRPO-based [33] RL to a broad range of visual perception tasks. Our approach aims to improve the performance of LVLMs in processing various visual tasks, especially when the fine-tuning data is limited. + +# 3. Methodology + +# 3.1. Preliminary + +Reinforcement Learning with Verifiable Rewards. Reinforcement Learning with Verifiable Rewards (RLVR) [4, 14, 39] is a novel training approach designed to enhance language models in tasks with objectively verifiable outcomes, such as math and coding. Unlike previous + +Reinforcement Learning from Human Feedback (RLHF) [19, 28, 49], which relies on a trained reward model, RLVR instead uses a direct verification function to assess correctness. This approach simplifies the reward mechanism while maintaining strong alignment with the task's inherent correctness criteria. Given the input question $q$ , the policy model $\pi_{\theta}$ generates responses $o$ and receives the verifiable reward. More specifically, RLVR optimizes the following objective: + +$$ +\max _ {\pi_ {\theta}} \mathbb {E} _ {O \sim \pi_ {\theta} (q)} \left[ R _ {\mathrm {R L V R}} (q, o) \right] \tag {1} +$$ + +$$ += \left[ R (q, o) - \beta \mathrm {K L} \left[ \pi_ {\theta} (o | q) \| \pi_ {\text {r e f}} (o | q) \right] \right], \tag {2} +$$ + +where $\pi_{\mathrm{ref}}$ is the reference model before optimization, $R$ is the verifiable reward function, and $\beta$ is the hyperparameters to control the KL-divergence. The verifiable reward function $R$ takes the question and output pair $(q,o)$ as inputs, and checks if the ground-truth answer remains the same as the prediction $o$ : + +$$ +R (q, o) = \left\{ \begin{array}{l l} 1, & \text {i f o = g r o u n d t r u t h}, \\ 0, & \text {o t h e r w i s e}. \end{array} \right. \tag {3} +$$ + +DeepSeek R1-Zero and GRPO. The DeepSeek R1-Zero algorithm eliminates dependence on supervised fine-tuning (SFT) by employing reinforcement learning for training, specifically through its Group Relative Policy Optimization (GRPO) framework. Different from reinforcement learning algorithms such as PPO [32] that require a critic model to evaluate policy performance, GRPO compares groups of candidate responses directly, eliminating the need for an additional critic model. For a given question $q$ , GRPO first generates $G$ distinct responses $\{o_1, o_2, \dots, o_G\}$ from the current policy $\pi_{\theta_{\mathrm{old}}}$ . Then GRPO takes actions based on these responses and denotes the obtained rewards as $\{r_1, r_2, \dots, r_G\}$ . By computing their mean and standard deviation for normalization, GRPO determines the relative quality of these responses: + +![](images/11a3f9cd07b97fa213e5601c9c1631c9473d5e66cab212414729e17d0edc1a8d.jpg) + +![](images/ad2e25d4d2992770ae4f5e5d701d997224811830bce465fdf689652656003385.jpg) +The athlete in the picture seems to be very excited about the competition results. If she needs to maintain clear vision underwater, what item should she wear over her eyes? + +![](images/621ffc5f86cc8bebcbc02ad0a60221f0d8128a800e5d8f42b6851db5f2781c57.jpg) + +![](images/dc615c598963df98bcf446ee3008d623cfd53f808270fbcc101aeaee41377616.jpg) +Birds often need a stable surface to land on and rest. What in the picture could provide a suitable landing spot for a bird? + +![](images/00d846b76e7967f973a6c5fc19ca05f3fc17825b4b9b4ddb79e8884134c94d33.jpg) +SFI + +![](images/f4e9de322f33151cebb7c0fa61a983490bccac3679c8432dbe79a027d2e748eb.jpg) + + +![](images/876fa59c80c892a3b2cc81d1686f57ee329070b3ba259cb6946f2b04859ed95a.jpg) +SFT + +![](images/5637ec958166fd4d43e5c97dffacd9f3bc318eb49e4e7776bb40124b6b7070a9.jpg) + +![](images/9a7fcd1877e0d1fb2f7913386d28f86395515c63e3e06dedf327e16e605f1aed.jpg) + +![](images/7e547ffdabc5a56733c486da8838c6fd7d713a1736e51e697bd5c902b6e4aa44.jpg) +Visual RFT + +![](images/f37dded202d0ef18c112104e7b23e38ddf9187149b69f8b96e2f1364583a46c9.jpg) + The bird is perched on a concrete ledge, which provides a stable surface for the bird to land on and rest. The ledge is also visible in the image. + +![](images/639f9fae758f346500f0ad1c3234a6ae9a6c4d852fb0e7454a7141164cee992a.jpg) + +![](images/71e92566c34abf755b0c79a05239a24989ca2eac6e28f733af186765738aca19.jpg) +Output the bounding box in the image corresponding to the instruction: What object in the room is mainly used as a sitting tool for writing or using an electronic device? + +![](images/2e97a739bd42191a77739b347377d811289944aa8a9299ede0be67f240511d78.jpg) +Visual RFT + +![](images/093ff02d2d2d9cf62a695c43f64fc07c358c6c3305f75d8cfdb00ea87f55f6ad.jpg) +Figure 5. Qualitative results of reasoning grounding on LISA [13]. Thinking process significantly improves reasoning grounding ability with Visual-RFT. + +![](images/132ff6a17af2c785350e4935711dacccd65d4163bbbaa1d33ff4f4d4713c02e2.jpg) + +![](images/70a4dfecd923cd17ab54c3bdf64d3fb4b1dbc531ec483da71f2b92b1948f0a3e.jpg) + +![](images/b79e4d47f1967dbe39ad2219981188ec9541d5ab09173f486cec23cc66a87763.jpg) +The image shows a athlete in a pool expressing a strong emotion of surprise or excitement. The athlete is wearing goggles over her eyes. To maintain clear vision underwater, she should wear goggles over her eyes. (350,187),(574, 612)/ +SFT + +![](images/8bb42a6ba6c8bf78ae2de24b3eaa70416b62129ae2c572a49a66cb0bea8aa556.jpg) + +![](images/566ddbc92d629f5458495e67c81831e423439c083f57c7ad3ac0f49b33a945b6.jpg) + +![](images/56eacccb448c3508293e054b155c174af3fbabebc430904616ce93a4e4072f4d.jpg) +Visual RFT + The chair in the room is mainly used as a sitting tool for writing or using an electronic device. + +![](images/86f2ed44dad28415283a443007dd92251560c55fa7a6b6ba3fcead9618642b70.jpg) + +$$ +A _ {i} = \frac {r _ {i} - \operatorname {m e a n} \left(\left\{r _ {1} , \dots , r _ {G} \right\}\right)}{\operatorname {s t d} \left(\left\{r _ {1} , \dots , r _ {G} \right\}\right)}, \tag {4} +$$ + +where $A_{i}$ represents the relative quality of the $i$ -th answer. GRPO encourages the model to favor better answers with a high reward value within the group. + +# 3.2. Visual-RFT + +The framework of Visual-RFT is shown in Fig. 3. The multi-modal input data from the user consists of images and questions. The policy model $\pi_{\theta}$ outputs a reasoning process and generates a group of responses based on the input. Each response is passed through a verifiable reward function to compute the reward. After group computation of the rewards for each output, the quality of each response is evaluated and used to update the policy model. To ensure the stability of the policy model training, Visual-RFT uses KL divergence to limit the difference between the policy model and the reference model. We will further discuss how to design the verifiable reward for visual tasks in Sec. 3.2.1, and the data preparation steps in Sec. 3.2.2 + +# 3.2.1. Verifiable Reward in Visual Perception + +The reward model is a key step in reinforcement learning (RL) that aligns models with preference alignment algorithms, which can be as straightforward as a verification function that checks for exact matches between predictions and ground-truth answers. The RL training process in the recent DeepSeek-R1 [4] model achieves a significant improvement in the model's reasoning ability through the verifiable reward design. To transfer this strategy to the visual domain, we design different rule-based verifiable reward functions for various visual perception tasks. + +IoU Reward in Detection Tasks. For the detection task, the model's output consists of bounding boxes (bbox) and corresponding confidences. The reward function for the + +detection task should adequately consider the Intersection-over-Union (IoU) metric, which is used to compute the mean Average Precision (mAP) in evaluation. Therefore, we design an IoU and confidence-based reward function $R_{d}$ . First, for the model's output box and confidence, we sort these boxes based on their confidence, denoted as $\{b_1,b_2,\dots,b_n\}$ . We then match each $b_i$ with the ground truth bbox, $\{b_1^g,b_2^g,\dots,b_m^g\}$ , and calculate the IoU, while setting an IoU threshold $\tau$ . Bounding boxes with an IoU below this threshold $\tau$ are considered invalid, and unmatched bboxes have an IoU of 0. After matching, we obtain the IoU and confidence for each box from the initial set, denoted as $\{iou_1:c_1,iou_2:c_2,\dots,iou_n:c_n\}$ . + +We then use these IoU results and confidence to construct our reward $R_{d}$ . Our reward $R_{d}$ consists of three parts, including the IoU reward, Confidence reward, and Format reward: + +$$ +R _ {d} = R _ {\mathrm {I o U}} + R _ {\text {c o n f}} + R _ {\text {f o r m a t}}. \tag {5} +$$ + +The IoU reward $R_{\mathrm{IoU}}$ is the average IoU of all the bounding boxes in the model's output, + +$$ +R _ {\mathrm {I o U}} = \frac {i o u _ {1} + i o u _ {2} + \dots + i o u _ {n}}{n}. \tag {6} +$$ + +The confidence reward $R_{\mathrm{conf}}$ is related to IoU. For each bounding box, if the $iou_{i}$ is non-zero, indicating a successful match, the confidence reward for this single box $r_{c}$ as the predicted confidence in computed as: + +$$ +r _ {c i} = \left\{ \begin{array}{l l} c _ {i} & , \text {i f} \quad i o u _ {i} \neq 0, \\ 1 - c _ {i} & , \text {i f} \quad i o u _ {i} = 0. \end{array} \right. \tag {7} +$$ + +This means that for successfully matched boxes, the higher the confidence, the better. If the $iou_{i}$ is zero, indicating a failed match, the lower the confidence reward $r_{c}$ for this box, the better. The overall confidence reward $R_{\mathrm{conf}}$ for the model's output is the average of the confidence rewards of + +Table 2. Few-shot results on Fine-grained Classification dataset. We evaluated four fine-grained image classification datasets under 1-shot,4-shot and 8-shot. Baseline results are from InPK [58] under 4-shot setting. + +
ModelsAverageFlower102Pets37AircraftCars196
CoOp [56]62.770.789.424.965.7
MaPLe [9]67.780.892.129.068.7
CoCoOp [57]68.982.693.030.969.1
PromptSRC [10]72.391.393.232.871.8
Qwen2-VL-2B56.054.866.445.956.8
+SFT (1-shot)51.756.654.765.330.0
+Visual-RFT80.3 +24.370.8 +16.084.1 +17.772.5 +26.693.8 +37.0
+SFT (4-shot)55.658.555.567.940.5
+Visual-RFT81.9 +25.971.4 +16.686.1 +19.774.8 +28.995.3 +38.5
+SFT(8-shot)60.359.671.469.240.9
+Visual-RFT85.1 +29.177.7 +22.990.2 +23.875.9 +30.096.5 +39.7
+ +all the bounding boxes in that output, + +$$ +R _ {\text {c o n f}} = \frac {\sum_ {i = 1} ^ {n} r _ {c i}}{n}. \tag {8} +$$ + +The format reward $R_{\mathrm{format}}$ is used to force the model prediction to meet the required HTML tag format of and (will detailed in Sec. 3.2.2). + +CLS Reward in Classification Tasks. In classification tasks, the reward function we use consists of two parts: accuracy reward $R_{\mathrm{acc}}$ and format reward $R_{\mathrm{format}}$ . The accuracy reward is determined by comparing the model's output class with the ground truth class, yielding a value of 1 for correct classification and 0 for incorrect classification: + +$$ +R _ {\mathrm {c l s}} = R _ {\mathrm {a c c}} + R _ {\text {f o r m a t}}. \tag {9} +$$ + +# 3.2.2. Data Preparation + +To train the Visual-RFT on various visual perception tasks, we need to construct the multi-modal training dataset. Similar to DeepSeek-R1, to enhance the model's reasoning ability and apply this ability to improve visual perception, Visual-RFT designed a prompt format to guide the model to output its reasoning process before providing the final answer. The prompts used for detection and classification tasks are shown in Tab 1. + +During the training process, we use the format reward to guide the model to output its reasoning process and the final answer in a structured format. The reasoning process is key to the model's self-learning and improvement during reinforcement fine-tuning, while the reward determined by the answer directs the model's optimization. + +# 4. Experiments + +# 4.1. Experimental Setup + +Implementation Details Our method is adaptable to various visual perception tasks. We employ a few-shot learn + +Table 3. Few-Shot results on COCO dataset of 8 categories. We conducted 1-shot, 4-shot and 8-shot experiments on 8 categories from the COCO dataset. + +
ModelsmAPbustrainfire hydrantstop signcatdogbedtoilet
Qwen2-VL-2B19.619.015.825.818.429.923.214.69.8
+SFT (1-shot)19.518.317.423.118.228.023.417.310.4
+Visual-RFT33.623.435.739.123.854.342.519.530.8
Δ+14.0+4.4+19.9+13.3+5.4+24.4+19.3+4.9+21.0
+SFT(4-shot)25.225.423.626.621.535.630.618.419.9
+Visual-RFT40.630.040.645.735.060.944.924.643.1
Δ+21.0+11.0+24.8+19.9+16.6+31.0+21.7+10.0+33.3
+SFT (8-shot)30.225.835.129.421.944.539.022.623.5
+Visual-RFT47.436.247.950.447.765.257.030.444.0
Δ+27.8+17.2+32.1+24.6+29.3+35.3+33.8+15.8+34.2
+ +Table 4. Few-shot results on LVIS dataset of 6 rare categories. We conducted 10-shot experiments on 6 rare categories from the LVIS dataset. + +
ModelsmAPbuggydiekitchen-tableomeletpapayastepladder
Qwen2-VL-2B4.02.91.213.44.71.50.0
+SFT10.07.07.634.14.76.30.0
+Visual-RFT19.49.119.642.220.414.510.9
Δ+15.4+6.2+18.4+29.2+15.7+13.0+10.9
Qwen2-VL-7B15.419.721.914.518.118.50.0
+SFT27.626.921.949.729.225.212.7
+Visual-RFT33.826.227.870.623.521.229.3
Δ+18.4+6.5+5.9+56.1+5.4+2.7+29.3
+ +Table 5. Reasoning Grounding Results on LISA [13]. Visual-RFT surpasses SFT in reasoning grounding with 239 training images. + +
ModelmIoUtestmIoUvalgIoUtest
OV-Seg [16]28.430.526.1
X-Decoder [64]28.529.124.3
GroundedSAM [20]26.228.621.3
Qwen2-VL-2B26.930.125.3
+SFT28.329.725.3
+Visual-RFT37.6 +10.734.4 +4.334.4 +9.1
Qwen2-VL-7B40.445.238.0
+SFT39.143.937.2
+Visual-RFT43.9 +3.547.1 +1.943.7 +5.6
+ +ing approach, providing the model with a minimal number of samples for training. For the image classification and object detection task, we adopt a few-shot setting to evaluate the model's fine-grained discriminative and recognition capability, applying reinforcement learning on limited data. Then, for the LISA [13] dataset focusing on reasoning grounding, which demands strong reasoning abilities, we train the model using Visual-RFT and assess its + +Table 6. Open Vocabulary Object Detection Results on COCO dataset. We trained on 48 base categories for 10-shot and tested on 17 novel categories. + +
ModelsnAPbAPmAP
Qwen2-VL-2B9.65.66.7
+SFT16.712.413.5
+Visual-RFT27.9 +18.320.3 +14.722.2 +15.5
Qwen2-VL-7B25.017.619.5
+SFT26.918.420.6
+Visual-RFT32.3 +7.323.6 +6.025.9 +6.4
+ +![](images/25b6ca914ee65361f8c11f74f40958cc6587ac85ae532ba23a844917c58f38b1.jpg) +Figure 6. Open Vocabulary Results Compared with Specialized Classification Models. We display the performance of different models in the figure, using nAP50 as the evaluation metric. + +reasoning and perception performance. Lastly, for open-vocabulary object detection, we evaluate the model's generalization capability by training the Qwen2-VL-2/7B [40] using Visual-RFT on a subdivided COCO dataset containing 48 base classes. We then test it on 17 novel classes from COCO and 13 rare classes from LVIS [5]. The model's visual perception and reasoning abilities are assessed in an open-vocabulary detection setting. In our detection experiments, we first prompt the model to check whether the category is present in the image, then predict bound boxes for categories that exist in the images. + +# 4.2. Few-Shot Classification + +To demonstrate the extensive generalization ability of Visual-RFT in the visual domain, we conduct few-shot experiments on fine-grained image classification. We selected four datasets: Flower102 [24], Pets37 [29], FGVC-Aircraft [23], and Car196 [12], which contain dozens to hundreds of similar categories, adding significant difficulty to the classification task. + +As shown in Tab. 2, with just one-shot of data, Visual-RFT already delivers a significant performance boost $(+24.3\%)$ . In contrast, SFT shows a noticeable decline $(-4.3\%)$ with the same minimal amount of data. Under the 4-shot setting, the performance of SFT is still slightly lower than the baseline, while the reinforcement fine-tuned model with Visual-RFT achieves an average performance + +improvement of 25.9. Under the 8-shot settings, as the amount of data increases, SFT's performance slightly exceeds the baseline. However, SFT still lags significantly behind the performance of the Visual-RFT. In Fig.4, we present some inference cases of the model after reinforcement fine-tuning when handling fine-grained classification tasks. These results not only demonstrate the strong generalization ability of Visual-RFT and its capacity to learn from limited data but also confirm that reinforcement fine-tuning, compared to SFT, leads to a genuine understanding of the task and deeper learning from reasoning. We provide additional results under more shot settings, as well as the performance of the 7B model and results on out-of-domain datasets in Sec. D.2 and Sec. C. + +# 4.3. Few-Shot Object Detection + +Few-shot learning has always been one of the core challenges faced by traditional visual models and large-scale vision-language models (LVLMs). Reinforcement finetuning provides a new solution to this problem by enabling the model to quickly learn and understand with a small amount of data. We selected eight categories from the COCO dataset, with 1, 4 and 8 images per category, to construct training sets with limited data. For the LVIS dataset, we select 6 rare categories. Since the training images for these rare categories are very sparse, with each category having between 1 and 10 images, we approximated this as a 10-shot setting. We then train the Qwen2-VL-2/7B model for 200 steps using both reinforcement finetuning and SFT, to evaluate the model's learning ability with limited data. + +As shown in Tab. 3 and Tab. 4, although both SFT and reinforcement fine-tuning can improve the model's recognition accuracy under the few-shot setting, the model after reinforcement fine-tuning consistently outperforms the SFT model by a large margin, maintaining a significant lead. On the COCO [17] categories, as the training data increases, the SFT model reaches an average mAP of approximately 30, while the reinforcement fine-tuned model approaches 47. In the LVIS [5] few-shot experimental results shown in Tab. 4, for the six more challenging rare categories in LVIS, reinforcement fine-tuning still outperforms SFT. The results in Tab. 3 and Tab. 4 clearly demonstrate the exceptional performance of reinforcement fine-tuning in the few-shot setting, where the model achieves a significant improvement in visual perception capabilities through reinforcement learning with only a small amount of data. We present additional results under more shot settings and for the 7B model in Sec. D.3, along with results on domain-specific detection datasets in Sec. C. + +# 4.4. Reasoning Grounding + +Another crucial aspect of vision-language intelligence is grounding the exact object according to user needs. Pre + +vious specialized detection systems lack reasoning abilities and fail to fully understand the user's intentions. Pioneered by LISA [13], there have been works done to enable large language models (LLMs) to output control tokens for other models (such as SAM [11]) or directly predict bounding box coordinates [31, 40] through supervised fine-tuning. In our work, we explore the use of Visual-RFT in this task and find that reinforcement learning (RL) leads to significant improvements over supervised fine-tuning. + +We finetune Qwen2-VL 2B/7B model [40] using Visual-RFT and supervised fine-tuning (SFT) on the LISA training set, which consists of 239 images with reasoning grounding objects. We follow the same test setting with LISA and compare the results of SFT and our method, both with 500 fine-tuning steps. As shown in Tab. 5, Visual-RFT significantly improves the final results in terms of bounding box IoU compared to SFT. Additionally, we prompt SAM [11] with the Qwen2-VL predicted bounding box to generate the segmentation mask (evaluated using gIoU). Visual-RFT significantly enhances grounding ability and outperforms previous specialized detection systems. Qualitative results are visualized in Fig. 5, where the thinking process significantly improves the ability to reason and grounding accuracy. Through Visual-RFT, Qwen2-VL learns to think critically and carefully examine the image to produce accurate grounding results. + +# 4.5. Open Vocabulary Object Detection + +The advantage of Visual-RFT over SFT arises from the former's true deep understanding of the task, rather than merely memorizing the data. To further demonstrate the powerful generalization ability of reinforcement finetuning, we conduct open vocabulary object detection experiments. We constructed a 10-shot training set by selecting 10 images per category from the 48 base categories of the COCO dataset. We perform Visual-RFT and SFT on the Qwen2-VL-2/7B model [40] using this data, and test the model on 17 new categories it has never seen before. + +As shown in Tab. 6, after reinforcement fine-tuning, the Qwen2-VL-2/7B model achieves an average nAP increase of 18.3 and 7.3 on 17 new categories from the COCO dataset. The Visual-RFT transfers its detection capabilities from the COCO base categories to the new COCO categories. In Fig. 6, we compare Visual-RFT's performance with several other detection models [41, 42, 50, 53, 59] using nAP50 as the evaluation metric. Both Qwen2-VL-2B and Qwen2-VL-7B models show significant improvements, with the 7B model outperforming the specialized model DE-ViT [53] under 10-shot setting. This demonstrates that Visual-RFT has a significant impact on improving the performance and generalization ability in visual recognition for LVLMs. We conducted more extensive open vocabulary experiments in the supplementary material, including results + +![](images/c67beca65bde8a1914c4ef8dd7a98b93cc3b1269063d75f2f38ae58c1a2ed9e9.jpg) +Figure 7. Ablation Study. We evaluated Visual-RFT and SFT across different steps and classification datasets. Visual-RFT demonstrated superior performance and generalization capability compared to SFT. + +![](images/e9b7ce65fb4291b843db6aae56190545604afbb03d5befe43ac7217bf41cad0f.jpg) + +on the LVIS dataset, see Sec. D.1 for details. + +# 4.6. Ablation Study + +To further compare the performance and generalization capability of SFT and Visual-RFT, we conduct the following ablation experiment. We use a 2-shot setting on the Flower102 dataset, training models separately with Visual-RFT and SFT. The models are then tested on both the Flowers and Pets datasets. The results under different step settings are shown in Fig. 7. + +In the left chart, the accuracy of both Visual-RFT and SFT on the Flowers increases with the number of training steps, gradually reaching saturation. Notably, Visual-RFT demonstrates a significant advantage over SFT when working with a small amount of data. In contrast, in the right chart, the accuracy of SFT on the Pets dataset drops sharply, falling well below the baseline. This decline is due to the model being trained on the Flower dataset, highlighting SFT's limited generalization ability. Meanwhile, Visual-RFT maintains its ability to distinguish fine-grained categories on the Pets dataset, showcasing its strong generalization performance even when the training and testing data domains differ. + +# 5. Conclusion + +In this paper, we introduce Visual Reinforcement Finetuning (Visual-RFT), the first approach to adapt the GRPO-based reinforcement learning strategy for enhancing the visual perception and grounding ability of LVLMs. By using a rule-based verifiable reward system, Visual-RFT reduces the need for manual labeling and simplifies reward computation, achieving significant improvements across various visual perception tasks. Extensive experiments show that Visual-RFT excels in fine-grained classification, open vocabulary detection, reasoning grounding and few-shot learning tasks. It outperforms supervised fine-tuning (SFT) with minimal data and shows strong generalization. This work demonstrates the potential of reinforcement learning to enhance the capabilities of LVLMs, making them more efficient and effective in visual perception tasks. + +# Acknowledgments + +This work was supported by National Key R&D Program of China 2022ZD0161600, Shanghai Artificial Intelligence Laboratory, Hong Kong RGC TRS T41-603/20-R, the Centre for Perceptual and Interactive Intelligence (CPII) Ltd under the Innovation and Technology Commission (ITC)'s InnoHK. Dahua Lin is a PI of CPII under the InnoHK. + +# References + +[1] Marwa Abdulhai, Isadora White, Charlie Snell, Charles Sun, Joey Hong, Yuexiang Zhai, Kelvin Xu, and Sergey Levine. Lmrl gym: Benchmarks for multi-turn reinforcement learning with language models. arXiv preprint arXiv:2311.18232, 2023. 3 +[2] Zheng Cai, Maosong Cao, Haojiong Chen, Kai Chen, Keyu Chen, Xin Chen, Xun Chen, Zehui Chen, Zhi Chen, Pei Chu, et al. Internl m2 technical report. arXiv preprint arXiv:2403.17297, 2024. 3 +[3] Thomas Carta, Clément Romac, Thomas Wolf, Sylvain Lamprier, Olivier Sigaud, and Pierre-Yves Oudeyer. Grounding large language models in interactive environments with online reinforcement learning. In ICLR, 2023. 3 +[4] Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, et al. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948, 2025. 1, 3, 4, 5 +[5] Agrim Gupta, Piotr Dollar, and Ross Girshick. Lvis: A dataset for large vocabulary instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5356-5364, 2019. 3, 7, 12, 13 +[6] Binyuan Hui, Jian Yang, Zeyu Cui, Jiaxi Yang, Dayiheng Liu, Lei Zhang, Tianyu Liu, Jiajun Zhang, Bowen Yu, Keming Lu, et al. Qwen2. 5-coder technical report. arXiv preprint arXiv:2409.12186, 2024. 3 +[7] Aaron Jaech, Adam Kalai, Adam Lerer, Adam Richardson, Ahmed El-Kishky, Aiden Low, Alec Helyar, Aleksander Madry, Alex Beutel, Alex Carney, et al. Openai o1 system card. arXiv:2412.16720, 2024. 1, 2, 3 +[8] Fangkai Jiao, Geyang Guo, Xingxing Zhang, Nancy F Chen, Shafiq Joty, and Furu Wei. Preference optimization for reasoning with pseudo feedback. arXiv preprint arXiv:2411.16345, 2024. 3 +[9] Muhammad Uzair Khattak, Hanoona Abdul Rasheed, Muhammad Maaz, Salman H. Khan, and Fahad Shahbaz Khan. Maple: Multi-modal prompt learning. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19113-19122, 2022. 6, 17 +[10] Muhammad Uzair Khattak, Syed Talal Wasim, Muzammal Naseer, Salman Siddique Khan, Ming Yang, and Fahad Shahbaz Khan. Self-regulating prompts: Foundational model adaptation without forgetting. 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 15144-15154, 2023. 6, 17 + +[11] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4015-4026, 2023. 8 +[12] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In ICCV workshops, 2013. 7, 12, 13 +[13] Xin Lai, Zhuotao Tian, Yukang Chen, Yanwei Li, Yuhui Yuan, Shu Liu, and Jiaya Jia. Lisa: Reasoning segmentation via large language model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9579-9589, 2024. 3, 5, 6, 8, 13, 14, 19 +[14] Nathan Lambert, Jacob Morrison, Valentina Pyatkin, Shengyi Huang, Hamish Ivison, Faeze Brahman, Lester James V Miranda, Alisa Liu, Nouha Dziri, Shane Lyu, et al. T\”ulu 3: Pushing frontiers in open language model posttraining. arXiv preprint arXiv:2411.15124, 2024. 1, 4 +[15] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Peiyuan Zhang, Yanwei Li, Ziwei Liu, et al. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 3 +[16] Feng Liang, Bichen Wu, Xiaoliang Dai, Kunpeng Li, Yinan Zhao, Hang Zhang, Peizhao Zhang, Peter Vajda, and Diana Marculescu. Open-vocabulary semantic segmentation with mask-adapted clip. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7061-7070, 2023. 6 +[17] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 3, 7, 12, 13, 20 +[18] Aixin Liu, Bei Feng, Bing Xue, Bingxuan Wang, Bochao Wu, Chengda Lu, Chenggang Zhao, Chengqi Deng, Chenyu Zhang, Chong Ruan, et al. Deepseek-v3 technical report. arXiv preprint arXiv:2412.19437, 2024. 3 +[19] Chris Yuhao Liu, Liang Zeng, Jiacai Liu, Rui Yan, Jujie He, Chaojie Wang, Shuicheng Yan, Yang Liu, and Yahui Zhou. Skywork-Reward: Bag of tricks for reward modeling in llms. arXiv preprint arXiv:2410.18451, 2024. 1, 4 +[20] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Qing Jiang, Chunyuan Li, Jianwei Yang, Hang Su, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. In European Conference on Computer Vision, pages 38-55. Springer, 2024. 3, 6, 16 +[21] Ziyu Liu, Yuhang Zang, Xiaoyi Dong, Pan Zhang, Yuhang Cao, Haodong Duan, Conghui He, Yuanjun Xiong, Dahua Lin, and Jiaqi Wang. Mia-dpo: Multi-image augmented direct preference optimization for large vision-language models. arXiv preprint arXiv:2410.17637, 2024. 4 +[22] Trung Quoc Luong, Xinbo Zhang, Zhanming Jie, Peng Sun, Xiaoran Jin, and Hang Li. Reft: Reasoning with reinforced fine-tuning, 2024. 3 +[23] Subhransu Maji, Esa Rahtu, Juho Kannala, Matthew Blaschko, and Andrea Vedaldi. Fine-grained visual classi + +fication of aircraft. arXiv preprint arXiv:1306.5151, 2013. 7, 12, 13 +[24] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In ICVGIP, 2008. 7, 12, 13 +[25] OpenAI. Hello gpt-4o, 2024. 3 +[26] OpenAI. Openai o3-mini system card, 2025. 2 +[27] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. In NeurIPS, 2022. 3 +[28] Long Ouyang, Jeff Wu, Xu Jiang, Diogo Almeida, Carroll L. Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, John Schulman, Jacob Hilton, Fraser Kelton, Luke E. Miller, Maddie Simens, Amanda Askell, Peter Welinder, Paul Francis Christiano, Jan Leike, and Ryan J. Lowe. Training language models to follow instructions with human feedback. In NeurIPS, 2022. 1, 4 +[29] Omkar M Parkhi, Andrea Vedaldi, Andrew Zisserman, and CV Jawahar. Cats and dogs. In CVPR, 2012. 7, 13 +[30] Rajkumar Ramamurthy, Prithviraj Ammanabrolu, Kianté Brantley, Jack Hessel, Rafet Sifa, Christian Bauckhage, Hannaneh Hajishirzi, and Yejin Choi. Is reinforcement learning (not) for natural language processing: Benchmarks, baselines, and building blocks for natural language policy optimization. In ICLR, 2023. 3 +[31] Hanoona Rasheed, Muhammad Maaz, Sahal Shaji, Abdelrahman Shaker, Salman Khan, Hisham Cholakkal, Rao M Anwer, Eric Xing, Ming-Hsuan Yang, and Fahad S Khan. Glamm: Pixel grounding large multimodal model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13009-13018, 2024. 8 +[32] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. arXiv:1707.06347, 2017. 4 +[33] Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, YK Li, Y Wu, et al. Deepseekmath: Pushing the limits of mathematical reasoning in open language models. arXiv preprint arXiv:2402.03300, 2024. 2, 3, 4 +[34] Charlie Victor Snell, Ilya Kostrikov, Yi Su, Sherry Yang, and Sergey Levine. Offline RL for natural language generation with implicit language q learning. In *ICLR*, 2023. 3 +[35] Nisan Stiannon, Long Ouyang, Jeffrey Wu, Daniel Ziegler, Ryan Lowe, Chelsea Voss, Alec Radford, Dario Amodei, and Paul F Christiano. Learning to summarize with human feedback. In NeurIPS, 2022. 3 +[36] Zhiqing Sun, Sheng Shen, Shengcao Cao, Haotian Liu, Chunyuan Li, Yikang Shen, Chuang Gan, Liang-Yan Gui, Yu-Xiong Wang, Yiming Yang, et al. Aligning large multimodal models with factually augmented rlhf. arXiv preprint arXiv:2309.14525, 2023. 4 +[37] Zhiqing Sun, Sheng Shen, Shengcao Cao, Haotian Liu, Chunyuan Li, Yikang Shen, Chuang Gan, Liang-Yan Gui, + +Yu-Xiong Wang, Yiming Yang, et al. Aligning large multimodal models with factually augmented rlhf. arXiv preprint arXiv:2309.14525, 2023. 4 +[38] Zhiqing Sun, Sheng Shen, Shengcao Cao, Haotian Liu, Chunyuan Li, Yikang Shen, Chuang Gan, Liang-Yan Gui, Yu-Xiong Wang, Yiming Yang, et al. Aligning large multimodal models with factually augmented rlhf. In ACL, 2024. 3 +[39] Kimi Team, Angang Du, Bofei Gao, Bowei Xing, Changjiu Jiang, Cheng Chen, Cheng Li, Chenjun Xiao, Chenzhuang Du, Chonghua Liao, et al. Kimi k1.5: Scaling reinforcement learning with llms. arXiv preprint arXiv:2501.12599, 2025. 1, 4 +[40] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 3, 7, 8 +[41] Zishuo Wang, Wenhao Zhou, Jinglin Xu, and Yuxin Peng. Sia-ovid: Shape-invariant adapter for bridging the image-region gap in open-vocabulary detection. ArXiv, abs/2410.05650, 2024. 8 +[42] Size Wu, Wenwei Zhang, Lumin Xu, Sheng Jin, Xiangtai Li, Wentao Liu, and Chen Change Loy. Clipself: Vision transformer distills itself for open-vocabulary dense prediction. ArXiv, abs/2310.01403, 2023. 8 +[43] An Yang, Beichen Zhang, Binyuan Hui, Bofei Gao, Bowen Yu, Chengpeng Li, Dayiheng Liu, Jianhong Tu, Jingren Zhou, Junyang Lin, et al. Qwen2. 5-math technical report: Toward mathematical expert model via self-improvement. arXiv preprint arXiv:2409.12122, 2024. 3 +[44] Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik R Narasimhan, and Yuan Cao. React: Synergizing reasoning and acting in language models. In ICLR, 2023. 3 +[45] Huaiyuan Ying, Shuo Zhang, Linyang Li, Zhejian Zhou, Yunfan Shao, Zhaoye Fei, Yichuan Ma, Jiawei Hong, Kuikun Liu, Ziyi Wang, et al. Internlm-math: Open math large language models toward verifiable reasoning. arXiv preprint arXiv:2402.06332, 2024. 3 +[46] Tianyu Yu, Yuan Yao, Haoye Zhang, Taiwen He, Yifeng Han, Ganqu Cui, Jinyi Hu, Zhiyuan Liu, Hai-Tao Zheng, Maosong Sun, et al. RIHF-V: Towards trustworthy mllms via behavior alignment from fine-grained correctional human feedback. In CVPR, 2024. 4 +[47] Tianyu Yu, Haoye Zhang, Yuan Yao, Yunkai Dang, Da Chen, Xiaoman Lu, Ganqu Cui, Taiwen He, Zhiyuan Liu, Tat-Seng Chua, et al. RLAIF-V: Aligning mllms through open-source ai feedback for super gpt-4v trustworthiness. arXiv preprint arXiv:2405.17220, 2024. 4 +[48] Yuhang Zang, Wei Li, Jun Han, Kaiyang Zhou, and Chen Change Loy. Contextual object detection with multimodal large language models. IJCV, 2024. 3 +[49] Yuhang Zang, Xiaoyi Dong, Pan Zhang, Yuhang Cao, Ziyu Liu, Shengyuan Ding, Shenxi Wu, Yubo Ma, Haodong Duan, Wenwei Zhang, et al. A simple yet effective multi-modal reward model. arXiv preprint arXiv:2501.12368, 2025. 1, 4 +[50] Alireza Zareian, Kevin Dela Rosa, Derek Hao Hu, and Shih-Fu Chang. Open-vocabulary object detection using captions. + +2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14388-14397, 2020. 8 +[51] Kechi Zhang, Ge Li, Yihong Dong, Jingjing Xu, Jun Zhang, Jing Su, Yongfei Liu, and Zhi Jin. Codedpo: Aligning code models with self generated and verified source code. arXiv preprint arXiv:2410.05605, 2024. 3 +[52] Pan Zhang, Xiaoyi Dong, Yuhang Zang, Yuhang Cao, Rui Qian, Lin Chen, Qipeng Guo, Haodong Duan, Bin Wang, Linke Ouyang, et al. Internlm-xcomposer-2.5: A versatile large vision language model supporting long-contextual input and output. arXiv preprint arXiv:2407.03320, 2024. 3 +[53] Xinyu Zhang, Yuting Wang, and Abdeslam Boularias. Detect every thing with few examples. ArXiv, abs/2309.12969, 2023. 8 +[54] Yuxiang Zhang, Shangxi Wu, Yuqi Yang, Jiangming Shu, Jinlin Xiao, Chao Kong, and Jitao Sang. o1-coder: an o1 replication for coding. arXiv preprint arXiv:2412.00154, 2024.3 +[55] Zhiyuan Zhao, Bin Wang, Linke Ouyang, Xiaoyi Dong, Ji-aqi Wang, and Conghui He. Beyond hallucinations: Enhancing lvlms through hallucination-aware direct preference optimization. arXiv preprint arXiv:2311.16839, 2023. 4 +[56] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Zwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, 130:2337 - 2348, 2021. 6, 17 +[57] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Conditional prompt learning for vision-language models. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16795-16804, 2022. 6, 17 +[58] Shuchang Zhou. Inpk: Infusing prior knowledge into prompt for vision-language models. 2025. 6, 17 +[59] Xingyi Zhou, Rohit Girdhar, Armand Joulin, Phillip Krahenbuhl, and Ishan Misra. Detecting twenty-thousand classes using image-level supervision. *ArXiv*, abs/2201.02605, 2022. 8 +[60] Yiyang Zhou, Chenhang Cui, Rafael Rafailov, Chelsea Finn, and Huaxiu Yao. Aligning modalities in vision large language models via preference fine-tuning. arXiv preprint arXiv:2402.11411, 2024. 4 +[61] Yiyang Zhou, Chenhang Cui, Rafael Rafailov, Chelsea Finn, and Huaxiu Yao. Aligning modalities in vision large language models via preference fine-tuning. arXiv preprint arXiv:2402.11411, 2024. 4 +[62] Yifei Zhou, Andrea Zanette, Jiayi Pan, Sergey Levine, and Aviral Kumar. Archer: Training language model agents via hierarchical multi-turn rl. In ICML, 2024. 3 +[63] Daniel M Ziegler, Nisan Stiennon, Jeffrey Wu, Tom B Brown, Alec Radford, Dario Amodei, Paul Christiano, and Geoffrey Irving. Fine-tuning language models from human preferences. arXiv:1909.08593, 2019. 3 +[64] Xueyan Zou, Zi-Yi Dou, Jianwei Yang, Zhe Gan, Linjie Li, Chunyuan Li, Xiyang Dai, Harkirat Behl, Jianfeng Wang, Lu Yuan, et al. Generalized decoding for pixel, image, and language. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15116-15127, 2023. 6 \ No newline at end of file diff --git a/ICCV/2025/Visual-RFT_ Visual Reinforcement Fine-Tuning/images.zip b/ICCV/2025/Visual-RFT_ Visual Reinforcement Fine-Tuning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..7e516ea5f0d68f36967240db66953646e8f00c05 --- /dev/null +++ b/ICCV/2025/Visual-RFT_ Visual Reinforcement Fine-Tuning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:678a3099705940ea11650be4796706988f16a0c4de83191d4820955aee91951f +size 804794 diff --git a/ICCV/2025/Visual-RFT_ Visual Reinforcement Fine-Tuning/layout.json b/ICCV/2025/Visual-RFT_ Visual Reinforcement Fine-Tuning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..690dbdfc1a718dd28e34e6c099e2fde3c38aab60 --- /dev/null +++ b/ICCV/2025/Visual-RFT_ Visual Reinforcement Fine-Tuning/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c7c4235d6f8d52bf19ff64c6a0e47f6fe2f2670386ad87114d4ee3d0da70bf19 +size 458871 diff --git a/ICCV/2025/VisualCloze_ A Universal Image Generation Framework via Visual In-Context Learning/3aa630b3-d121-44bf-b130-2854f77fd1d9_content_list.json b/ICCV/2025/VisualCloze_ A Universal Image Generation Framework via Visual In-Context Learning/3aa630b3-d121-44bf-b130-2854f77fd1d9_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..7ed6a9006e509d83e97096fee64c724586c1fbce --- /dev/null +++ b/ICCV/2025/VisualCloze_ A Universal Image Generation Framework via Visual In-Context Learning/3aa630b3-d121-44bf-b130-2854f77fd1d9_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cba43f895947ddc8ff9f01bed034706120b396eb3658d963c6ca29e2adb5a5fc +size 79759 diff --git a/ICCV/2025/VisualCloze_ A Universal Image Generation Framework via Visual In-Context Learning/3aa630b3-d121-44bf-b130-2854f77fd1d9_model.json b/ICCV/2025/VisualCloze_ A Universal Image Generation Framework via Visual In-Context Learning/3aa630b3-d121-44bf-b130-2854f77fd1d9_model.json new file mode 100644 index 0000000000000000000000000000000000000000..5d09a6fb505eea91a694868fa4b42fb748beb857 --- /dev/null +++ b/ICCV/2025/VisualCloze_ A Universal Image Generation Framework via Visual In-Context Learning/3aa630b3-d121-44bf-b130-2854f77fd1d9_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1f0504a9e859e843a464f45e8017cbea8179abece1be4ad715404a740a0a3e25 +size 103208 diff --git a/ICCV/2025/VisualCloze_ A Universal Image Generation Framework via Visual In-Context Learning/3aa630b3-d121-44bf-b130-2854f77fd1d9_origin.pdf b/ICCV/2025/VisualCloze_ A Universal Image Generation Framework via Visual In-Context Learning/3aa630b3-d121-44bf-b130-2854f77fd1d9_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..6cdf8072112b075814bf65cc163edc781fb96667 --- /dev/null +++ b/ICCV/2025/VisualCloze_ A Universal Image Generation Framework via Visual In-Context Learning/3aa630b3-d121-44bf-b130-2854f77fd1d9_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:87f27dbb3aa2978188280d011cea2414988f9ec6dc0bbe32d9a9864a3457fdbe +size 6038808 diff --git a/ICCV/2025/VisualCloze_ A Universal Image Generation Framework via Visual In-Context Learning/full.md b/ICCV/2025/VisualCloze_ A Universal Image Generation Framework via Visual In-Context Learning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..e966d5b21f4fb7ae6fe6c0e03a72e6f94e6d5902 --- /dev/null +++ b/ICCV/2025/VisualCloze_ A Universal Image Generation Framework via Visual In-Context Learning/full.md @@ -0,0 +1,308 @@ +# VisualCloze: A Universal Image Generation Framework via Visual In-Context Learning + +Zhong-Yu Li $^{1,4,5*}$ Ruoyi Du $^{3,5*}$ Juncheng Yan $^{5,6}$ Le Zhuo $^{5}$ +Zhen Li $^{7\dagger}$ Peng Gao $^{2,5}$ Zhanyu Ma $^{3}$ Ming-Ming Cheng $^{1,4\dagger}$ $^{1}$ NKIARI, Shenzhen Futian + +$^{2}$ Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences $^{3}$ Beijing University of Posts and Telecommunications $^{4}$ VCIP, CS, Nankai University $^{5}$ Shanghai AI Laboratory $^{6}$ Tsinghua University $^{7}$ The Chinese University of Hong Kong + +# Abstract + +Recent advances in diffusion models have significantly advanced image generation; however, existing models remain task-specific, limiting their efficiency and generalizability. While universal models attempt to address these limitations, they face critical challenges, including generalizable instruction design, appropriate task distributions, and unified architectural design. In this work, we propose VisualCloze, a universal image generation framework, to tackle these challenges. Unlike existing methods that rely on language-based task descriptions, leading to task ambiguity and weak generalization, we integrate visual in-context learning, allowing models to identify tasks from demonstrations. Meanwhile, the inherent sparsity of visual task distributions hampers the learning of transferable knowledge across tasks. To this end, we introduce Graph200K, a graph-structured dataset that establishes various interrelated tasks, enhancing task density and knowledge transfer. Furthermore, we uncover an intrinsic alignment between image infilling and in-context learning, enabling us to leverage the strong generative priors of pre-trained infilling models without modifying their architectures. Experiments demonstrate that VisualCloze achieves strong performance across various indomain tasks while generalizing to unseen tasks in few-shot and zero-shot settings. Our codes and dataset are available at https://visualcloze.github.io/. + +# 1. Introduction + +Recent advancements in image generation, propelled by the progress of diffusion models [14, 30, 78], have led to a wide range of applications, including image editing [60], + +![](images/66bfdbb5092d9dd97cb5be73057844507ef3addd0d61043a1018f487e462485d.jpg) +Figure 1. Illustration of our universal image generation framework based on visual in-context learning. Given one query of a specific task, the generative model learns the task by observing a few in-context examples presented as demonstrations. + +style transfer [55, 71], virtual try-on [10, 11], and personalized generation [34, 48], among others. However, these tasks typically require task-specific models, which limit efficiency and scalability for real-world applications. In recent years, there has been growing interest in unified generative models [26, 35, 40], which aim to handle diverse image generation tasks within a single framework. However, key challenges remain despite significant progress, e.g., instruction design, task distribution, and model architecture. + +Regarding instructions that guide the model to differentiate distinct tasks and produce expected behavior, existing methods primarily rely on language instructions [26, 40] or task-specific tokens [35]. However, the task complexity and inherent gap between vision and language modalities make it hard for the model to understand the task intent based on language description alone. This issue leads to task confusion [35] and hinders generalization on unseen tasks [32, 61]. In contrast, large language models have + +![](images/0e946ce8b964ee08a09a1e9c7ea814aaf342c3b652142a6dc4afbffb8c8c278e.jpg) +Figure 2. Results of our universal image generation framework. The grid image follows the structure of Fig. 1, and the red box indicates the generated image. After training on a wide range of tasks, our framework exhibits a certain generalization ability on unseen tasks. + +successfully achieved unified multi-task modeling, partially due to the rise of in-context learning [5], which allows models to understand tasks using only a few demonstrations. In the vision modality, prior works [1, 3, 4, 39, 57, 61, 72] also show the potential of in-context learning, even though they are still limited to limited task domain (e.g., dense prediction, style transfer, etc.) [58, 77] or simple generation tasks that involve only one condition and one target image [39, 52]. Inspired by this, not only language instructions, we also adopt a few in-context task examples as visual instructions to guide the model toward desired behavior, as shown in Fig. 1. Then, the model supports various tasks with mitigated task ambiguity and supports few-shot generalization on unseen tasks, as shown in Fig. 2. The model also exhibits a certain degree of zero-shot generalization. + +From the perspective of task distribution, visual tasks are inherently sparse compared to those in natural language processing because task-specific datasets [61, 75] for different tasks have minimal overlap [18, 29, 69]. Such sparse task learning isolates the knowledge of each task and limits the model from learning shared features across tasks. Moreover, the weak correlations between tasks hinder knowledge transfer and adaptability of models. However, existing works in multi-task learning [9, 15, 28, 47] have verified the benefits of overlapping knowledge across related tasks. To alleviate the sparsity of visual tasks, we introduce a graph-structured dataset, Graph200K, where each image is associated with annotations spanning five meta-tasks, i.e., conditional generation [70], IP preservation [66], style transfer [71], image editing [60], and restoration [67]. By com + +bining different conditions, we train the model with a variety of tasks that overlap and interact with each other. Given this highly overlapping task space, our dataset significantly increases task density, allowing the model to learn shared and transferable representations more effectively. + +For the architecture design, to fully realize the potential of universal image generation, it is crucial to design structures that accommodate flexible task formats [26, 32, 61], especially when using in-context examples, and are also compatible with state-of-the-art models [30, 78] to leverage their strong generative priors. In this work, we find that the state-of-the-art image infilling model [30] has a consistent objective with our in-context learning based universal generative models. Specifically, we concatenate all input and output images as a grid layout image, where the objective of a task is to fill the unknown grid, as illustrated in Fig. 1. This alignment enables us to build our model upon advanced general-purpose infilling models without additional modifications, thereby achieving powerful universal generation capabilities with minimal data and training costs. + +In this work, we propose a universal image generation framework, VisualCloze, which fine-tunes FLUX.1-Filldev [30] with interrelated tasks sampled from Graph200K to learn transferable knowledge and support visual in-context learning. As the number of in-context examples increases, we observe enhanced performances and reduced task confusion, enabling the model to support a broad spectrum of in-domain tasks, including conditional generation, image restoration, editing, style transfer, IP-preservation, and their combinations. On unseen tasks, the model also shows a cer + +tain degree of generalization ability, as shown in Fig. 2. In summary, our main contributions are as follows: + +- We propose an in-context learning based universal image generation framework that supports a wide range of indomain tasks and exhibits generalization to unseen ones. +- We design a graph-structured dataset, Graph200K, which constructs a compact task space, enabling flexible online task sampling and promoting the models to learn shared and transferable representations across tasks. +- We find that the state-of-the-art infilling model shares a consistent objective with unified image generation, enabling us to achieve exceptional performance through minimal tuning without modifying the model structure. + +# 2. Related Work + +# 2.1. Image Generation + +Recent advances in text-to-image generation have achieved remarkable performance, largely driven by the development of autoregressive models [37, 50, 68] and diffusion models [2, 12, 14, 17, 23, 36, 38, 43, 46]. Among these, rectified flow transformers [14, 16, 30, 78] have shown great training efficiency and overall performance. Building on these foundational models, diverse applications have emerged, such as conditional generation [70], style transfer [55], and personalized generation [34]. More recently, universal models that address various tasks [32, 40, 73] have been explored. For example, unified models like OmniGen [61] leverage large vision language models to consolidate multiple tasks into a single framework. Similarly, UniReal [8] unifies image generation tasks as discontinuous video generation. However, they still face issues such as over-reliance on language instructions, isolation and sparsity of visual tasks, and architecture design that accommodates flexible task formats. To address these issues, we propose a universal image generation framework that unifies generation tasks as image infilling. Through visual in-context learning and our Graph200K dataset that constructs a denser task space to learn transferable representations, our method alleviates ambiguity to support a diverse set of in-domain tasks and generalizes to tasks unseen during training. + +# 2.2. Visual In-context Learning + +Along with the emergence of large language models, such as GPT-3 [5], in-context learning [13] has been an effective approach to allow the language model to understand and perform complex tasks given a few demonstrations. Early works [20, 21] in vision modality propose image analogies to automatically create an image filter from examples. In recent years, leveraging inpainting model [3, 4, 72], masked image modeling [39, 57, 58], or vision-language model [1, 76], visual in-context learning is proposed to handle more tasks. However, they mainly focus on dense pre + +diction [49, 51, 77] or visual understanding [54]. OmniGen [61] also leverages in-context learning to generalize to unseen domains, e.g., segmenting unseen concepts when the model has learned the segmentation task during training. However, it mainly focuses on simple tasks of dense prediction, and the gap between the unseen and training domains is still limited. Some recent works [31, 39, 52, 59] extend visual in-context learning to image generation, but they are still limited by simple tasks such as conditional generation and dense prediction. Moreover, the sparsity of visual tasks makes it difficult for models to learn transferable and overlapping knowledge across tasks, limiting the generation ability of in-context learning. In contrast, we introduce a graph-structured dataset that supports interrelated tasks and thus constructs a more dense task space, promoting the model to learn shared and transferable knowledge and enhance its adaptability. + +# 3. Dataset + +Recent works [25, 61] have made great progress in unified image generation. However, their ability to generalize to unseen tasks remains highly limited. We partially attribute this to the sparsity and isolation of visual tasks, making it difficult for the model to capture shared features across tasks and perform well on unseen ones. Moreover, weak correlations between tasks further hinder knowledge transfer, restricting the adaptability of models. Therefore, increasing task density or strengthening task inter-relations helps improve the generalization ability of models via a compact task distribution. In this paper, taking the Subject200K [53] dataset as a testbed, we extend each data point in this dataset into a data graph, constructing an expanded dataset dubbed Graph200K. As illustrated in Fig. 3 (a), we can flexibly construct various related tasks by sampling different paths within the data graph. + +# 3.1. Graph-Structured Multi-Task Dataset + +In natural language processing, tasks overlap significantly, facilitating strong cross-task learning ability. In contrast, visual tasks are inherently distinct, posing challenges for vision models to achieve similar generalization ability via instruction tuning. To ease this issue, we introduce a Graph-Structured Multi-Task Dataset. As illustrated in Fig. 3 (a), given a text-to-image dataset, each original image $O_{n}$ is treated as the central node of a graph $G_{n}$ around which diverse task annotations are constructed, including those for various spatial conditions $(C_n^i)$ , various degradations $(D_n^i)$ , various image editing results $(E_n^i)$ , reference image $(R_{n})$ for IP-preservation, and style transfer $(T_{n}^{i})$ with various reference styles $(S_{n}^{i})$ . The construction process for each task pair is detailed in the next section. + +As shown in Fig. 3 (a), each task annotation forms a bidirectional edge with the image. Thus, the graph is strongly + +![](images/6376e447297ef4a43b446b6b4383d28b4042efe28ca56ebdc55ce401dcc42e4d.jpg) +(a) + +![](images/a2c81a41b7822c1900892fd4aa528a4fe1b75542c062ca5cc83db4b7ed890d89.jpg) +(b) +Figure 3. (a) Graph200K consists of $N$ graph, defined as $\{G_1,\dots ,G_n,\dots ,G_N\}$ , where the $n$ -th graph contains an image $(O_{n})$ and its annotations of various tasks, such as conditional generation $(C_n^i)$ , restoration $(D_n^i)$ , editing $(E_n^i)$ , IP preservation $(R_n^i)$ , and style transfer $(T_n^i)$ . The superscript $i$ indicates the fine-grained classification of tasks. (b) Through sampling nodes along the graph, Graph200K supports diverse and flexible tasks. For a sample with $R + 1$ rows, the first $R$ rows mean in-context examples defined in Section 4.1. + +connected, which means that for any two nodes, bidirectional paths exist between them. In other words, a generation task can be formulated as a path within the graph. The nodes along these paths (except the end nodes) serve as condition images, which is analogous to the question in instruction fine-tuning, while the target image (the end nodes) plays the role of the answer. Specifically, there are 49 types of nodes in our Graph200K and we sample up to 134 highly overlapping tasks, making the model learn more compact and shared representations across tasks. Moreover, it enriches the diversity and flexibility of our instruction fine-tuning data. For example, the path $\{R_{n},E_{n},O_{n}\}$ from graph $G_{n}$ corresponds to the image editing with subject reference, as shown in Fig. 3 (b) bottom. + +# 3.2. Dataset Construction + +For convenience, we inherit subject-driven data from the Subjects200K [53]. Additionally, 32 different degradations are applied online to the images to acquire restoration data. We summarize the data construction methods in this section for the remaining three tasks. + +Conditional generation. Each image is paired with 12 distinct conditions generated by specialized models, including canny edges [6], HED edges [62], Hough lines [19], semantic segmentation maps [33], depth maps [64], shape normal maps [63], and human keypoints [7], following ControlNet [70]. This work extends the conditions by incorporating SAM2 [45] masks, foreground segmentation, and open-world boxes and masks. The foreground segmenta + +tion, derived from the RMBG [74], supports diverse tasks such as inpainting and foreground extraction. Open-world bounding boxes are generated through the grounding caption capability of Qwen2-VL [56], which are processed using SAM2 [45] to produce corresponding masks. + +Style transfer. We transfer the style of images according to reference in both semantic-variant and semantic-invariant settings. Specifically, the semantic-invariant transfer adopts InstantStyle [55] to preserve the semantic content, while the semantic-variant transfer relies on FLUX.1-Redux-dev [30], using the style embeddings and depth as conditions. For each image, we randomly generate five stylized versions. Mixing the two tasks pushes the model to better follow the in-context examples to avoid ambiguity. + +Image editing. We design two types of editing tasks, including background-variant and background-invariant editing. The background-invariant editing begins with localizing the subjects. Then, we leverage a large vision-language model, Qwen2-VL [56], to modify the image caption with a new object that replaces the original subject. The image, with the subject masked, is subsequently processed by the FLUX.1-Fill-dev [30] inpainting model to integrate the alternative object into the masked region. The above operation is repeated five times to enrich the dataset. For background-variant editing, the difference lies on the last step, which utilizes FLUX.1-Redux-dev [30] with depth as the condition and the modified caption as the text prompt. + +# 4. Method + +This paper identifies the core challenges in building a universal image generation model, including the need for a clearly defined and generalizable task formulation, the sparsity of visual tasks, and the lack of a unified framework for multi-task learning. In the previous section, we addressed the issue of task sparsity by constructing the compact Graph200K dataset. Section 4.1 introduces visual in-context learning as the ideal paradigm for universal task formulation. Afterward, Section 4.2 considers the image infilling model as a unified multi-task framework, achieving strong generalization capabilities with minimal cost. + +# 4.1. Visual In-context Learning + +Language instructions are usually used to specify the generation definition to handle multiple visual generation tasks with a single generative model. However, due to the gap between vision and language, the text comprehension ability of image generation models remains limited. This issue leads to task confusion [35] in existing universal generative models and weak generalization to unseen tasks. Inspired by the success of few-shot learning on large language models [5], we recognize that visual context may serve as a more friendly task instruction for visual generative models, given their superior visual understanding capabilities. + +Therefore, in this paper, we re-propose visual in-context learning to build a universal and generalizable image generation system. Formally, we define the image input-output of arbitrary conditional generation task as a query consisting of $L - 1$ condition images and a blank target $\varnothing$ to be completed by the model, i.e., $X = \mathrm{concat}(\{x_1,\dots,x_{L - 1},\varnothing \})$ . During training, we randomly provide 0 to $C$ demonstration contexts $(X = \mathrm{concat}(\{x_1,\dots,x_{L - 1},x_L\}))$ for each query. This strategy ensures the generalization ability of models across different numbers of contexts while also preserving its capability to infer solely based on text instructions. As a result, the model can handle various applications, i.e., users can efficiently process in-domain simple tasks using only text instructions or prompt the model with multiple visual contexts to execute complex and even unseen tasks. In our experiments, we show that providing task demonstrations not only helps alleviate task confusion and boost model performance across in-domain tasks [35], but also enhances the generalization ability on unseen tasks. + +# 4.2. Unified Multi-task Framework + +Unlike previous visual in-context learning models that primarily focus on scenarios with a single image condition and a single context [39, 52], in this work, we aim to construct a unified framework capable of handling varying numbers of conditions and contexts, allowing for flexible adaptation to diverse tasks. Assuming all images processed by the model share the same size $W \times H$ , the input-output of our model + +naturally forms an interface of size $(L\times W)\times ((C + 1)\times H)$ under our visual in-context learning formulation. Specifically, suppose we concatenate all inputs and outputs from a single interaction into a complete grid image (as shown in Fig. 1). In that case, the model we seek infills the last grid based on the surrounding context, akin to solving visual cloze puzzles. Therefore, we build our unified framework, VisualCloze, based on the general image infilling architecture capable of handling multiple resolutions. + +Consistent with common diffusion-based infilling model designs, our model can be formulated as: + +$$ +\hat {X} = f (X \mid T, M), \tag {1} +$$ + +where $X$ is the concatenated image, with the last grid left blank, $T$ is the language instruction, $M$ is the mask condition, and $\hat{X}$ represents the infilled result. The target image can be simply cropped from $\hat{X}$ . The mask $M$ is a binary matrix with the size of $(H\times (C + 1),W\times L)$ : + +$$ +M (i, j) = \left\{ \begin{array}{l l} 1 & \text {i f} i \in [ H \times (C - 1), H \times C) \\ & \text {a n d} j \in [ W \times (L - 1), W \times L), \\ 0 & \text {o t h e r w i s e}, \end{array} \right. \tag {2} +$$ + +where $M(i,j) = 1$ indicates that the pixel will be masked and generated by the infilling model. Specifically, we mask the region in the last row and column. + +One key benefit of this design is that it shares a highly consistent optimization objective with general image infilling models without architectural modifications or explicit input conditions. This allows us to directly fine-tune advanced image infilling models using the newly constructed dataset while maximizing the utilization of the prior knowledge of foundation models. In contrast, existing task-specific models often require additional learnable modules [34, 60] or adapting to extra condition inputs [53], which may compromise the native capabilities of the model. + +Note that the design of language instruction is also necessary for VisualClozebecause it is responsible for defining the grid image layout and specifying the task content when example contexts are unavailable. In our unified framework, the instruction consists of three parts: (1) layout instruction, which describes the $(C + 1)\times W$ layout of the grid image, (2) task instruction, which specifies the task type, and (3) content instruction, which describes the content of the target image. The details about the instructions are available in the supplementary material. By restructuring the three components $X,T$ , and $M$ in Eqn. (1), we achieve a unified multi-task framework for image generation with the general image infilling paradigm and support in-context learning. + +# 4.3. Implementation Details + +We use FLUX.1-Fill-dev [30] as our foundation model, considering its outstanding performance among open-source + +
ConditionMethodContextControllabilityQualityText Consistency
F1 ↑RMSE ↓FID [22] ↓SSIM ↑MAN-IQA [65] ↑MUSIQ [27] ↑CLIP-Score [44] ↑
CannyControlNet [70]0.13-46.060.340.3145.4534.10
OminiControl [53]0.47-29.580.610.4461.4034.40
OneDiffusion [32]0.39-32.760.550.4659.9934.99
OmniGen [61]0.43-51.580.470.4762.6633.66
Oursdev00.39-30.360.610.4861.1335.03
Oursfill00.35-30.600.550.4964.3934.98
Oursfill10.36-31.340.550.4964.1234.96
Oursfill20.36-31.150.560.4964.0834.85
DepthControlNet [70]-23.7036.830.410.4460.1734.49
OminiControl [53]-21.4436.230.520.4460.1834.08
OneDiffusion [32]-10.3539.030.490.4960.4934.71
OmniGen [61]-15.0786.080.260.4964.9029.72
Oursdev0-25.0642.140.530.4658.9534.80
Oursfill0-10.3133.880.540.4864.8535.10
Oursfill1-9.9134.440.540.4964.3234.95
Oursfill2-9.6834.880.540.4864.2934.89
DeblurControlNet [70]-37.8253.280.490.4561.9233.80
OminiControl [53]-19.7026.170.850.4560.7034.53
OneDiffusion [32]-------
OmniGen [61]-------
Oursdev0-25.0356.760.740.3846.6833.52
Oursfill0-26.5340.590.740.4659.6234.56
Oursfill1-25.8736.930.760.4861.5834.82
Oursfill2-25.5736.280.760.4861.7734.82
+ +Table 1. Quantitative comparison on conditioning generation and image restoration. The methods that train a specialist for each task are marked as gray color. Except for these methods, the best method is bolded, and the second-best method is underlined. + +
MethodContextDINOv2CLIP-ICLIP-T
OminiControl [53]73.1787.7033.53
OneDiffusion [32]73.8886.9134.85
OmniGen [61]67.7383.4334.53
Oursdev078.0587.6835.06
Oursfill080.4189.6335.16
Oursfill179.3389.2235.02
Oursfill280.3289.3635.01
+ +Table 2. Quantitative comparison for subject-driven image generation. We report clip scores on text alignment and style consistency. Specialist models for each task are shaded in gray. Among the remaining methods, the best one is emphasized in bold, while the second-best is underlined. + +models. We fine-tune the model using LoRA [24] with rank of 256. The model is tuned for 20,000 iterations with an accumulated batch size of 64 on $8 \times \mathrm{A}100$ GPUs. We employ the AdamW optimizer with a learning rate of $1e^{-4}$ . Following [30], we incorporate the lognorm noise strategy with dynamic time shifting. The number of contexts is set between 0 and 2, while the data stream length $L$ varies between 2 and 4 in the Graph200K dataset. To balance computational efficiency, each grid size is fixed at $384 \times 384$ . Although the generated resolution is relatively low, it is sufficient to + +
text↑image↑
InstantStyle [55]0.270.60
OmniGen [61]0.270.52
Oursdev0.300.53
Oursfill0.290.55
+ +Table 3. Quantitative comparison for style transfer. We report CLIP scores on text alignment and style consistency. The specialist are indicated in gray. Among the others, the top-performing one is highlighted in bold, and the second-best is underlined. + +validate the effectiveness of our proposed method. Moreover, in practical applications, high-resolution outputs can be obtained through simple post-upscaling techniques [41]. + +# 5. Experiments + +# 5.1. Main Results + +We compare our method with universal generative models, including OmniGen [61] and OneDiffusion [32], as well as specialized models, such as ControlNet [70] and OmniControl [53]. Additionally, we fine-tune FLUX.1-dev [30] using the same settings as FLUX.1-Fill-dev for comparison, and refer to the tuned models as Oursdev and Oursfill. + +For conditional generation and image restoration, we evaluate the models based on three criteria, i.e., controllabil + +![](images/08da35d6dc0f1dff56391202274dadafb9c3211bbf2108e72cb90194d47c8dfa.jpg) + +![](images/b7baaaa839ef138bdba2bc145d35c3843602f2a0a1320875f222c7363a0b211c.jpg) + +![](images/a2ec14ef14ae6b695150c6a996481f9fa243e5212ab81ee42bad6a2d940edfb8.jpg) + +![](images/18dd8bd30aa1323e6447a5448eedd59dc688a121bd46a6bb94358adafea5470b.jpg) + +![](images/0ba0b220ab9a0947b6d2dab789f23a1ff2b422fb588b5de326ba397bb18ef7f1.jpg) + +![](images/e64150fec44f7870a76db1ef22f26a78abefed1c6e566bc5b0a719d84c35ca84.jpg) + +![](images/70390cbd29ac8a1ce42a03ad45c4f2b1fc3c258e705594f7c0ea9c37657c4e12.jpg) + +![](images/b3a3c9616c6d3d1b3dfa49b9c825765b382053be2f56bfd75c1be56a0c1d6e53.jpg) +Figure 4. Illustration of conditional image generation. + +![](images/e2e39772fdc134efa22fc2b47da0478f8922539660478970ebf846ff3645084d.jpg) + +![](images/cf405feb21b71ac9431b8b7d305a12dbb5bc09247139092e982c92dec2e55097.jpg) + +![](images/03c022371581f6f8d360a722758e854be03440e8cd7734265438cb23544e3ab4.jpg) + +![](images/f402c3cc3c5547365671e41ee7b2a8633fda7fad2aa3b9cb3cbd02429857723a.jpg) + +![](images/ae01e7f62b7cf57e7990ba24047b27e1aaac52a1a22812373bab02fed0220509.jpg) +Figure 5. Subject-driven image generation. Based on a reference, we generate multiple images under different environments. + +ity, visual quality, and text consistency, following the evaluation approach of OminiControl [53]. As shown in Tab. 1, our framework demonstrates comparable controllability to existing universal methods while achieving superior visual quality and text consistency. When compared to specialized methods, our model also performs on par with the best results and even outperforms them on the depth-to-image. + +In the style transfer task, we measure text consistency and style alignment using the CLIP [44] model. As reported in Tab. 3, our method outperforms OmniGen [61] by $2\%$ and $3\%$ in text alignment and style consistency, respectively. Even when compared with InstantStyle-Plus [71], a specialized model, we achieve a $2\%$ improvement in text consistency, with only a slight decrease in style alignment. + +Furthermore, we evaluate the models on subject-driven image generation and report semantic alignment using the DINOv2 [42], CLIP-I [44], and CLIP-T [44] scores. Across all these metrics, our method consistently delivers improvements, as shown in Tab. 2. For example, compared to the specialized model, OmniControl [53], we achieve improvements of $7.15\%$ , $1.66\%$ , and $1.48\%$ in these three scores. + +Advantages of the infilling model. Our method (Oursfill) is built on FLUX.1-Fill-dev [30], which shares the same objective as our unified image generation framework. To verify its effectiveness, we also fine-tune Fill.1-dev [30] (Oursdev) using identical settings. Unlike Oursfill, which requires no modifications, Oursdev necessitates model adaptations for universal image generation, as shown in the supplementary material. Despite its simplicity, Oursfill achieves superior performance across multiple tasks. + +As shown in Tab. 1, Oursdev achieves a higher F1 score + +![](images/1a06bc88b6bca58066f738aaf695f5f7d8cbc38dc5e9ca6fefbbb9b5aa069f4a.jpg) +Figure 6. Comparison between Flux.1-dev (Oursdev) and Flux.1-Fill-dev (Oursfill). + +than Oursfill in canny-to-image generation. However, in other tasks, Oursfill demonstrates a significant advantage. For instance, in depth-to-image generation, Oursfill reduces RMSE from 25.06 to 10.31, indicating improved structural alignment. In the deblurring task, Oursfill achieves superior quality by lowering RMSE while maintaining a higher visual quality. In subject-driven image generation, Tab. 2 shows that Oursfill consistently outperforms Oursdev. Additionally, in semantic-invariant style transfer, Oursfill delivers comparable performance to Oursdev, as shown in Tab. 3. + +Furthermore, Fig. 6 presents a visual comparison, where Oursfill demonstrates clear advantages over Oursdev. Notably, in the depth-to-image generation, images produced by Oursdev frequently exhibit diagonal streak artifacts, which significantly degrade visual fidelity. Considering the advantages in performance, visual quality, and architectural efficiency, Oursfill stands out as the superior model. + +# 5.2. Effectiveness of In-Context Learning + +In this section, we analyze the impact of in-context learning on both seen and unseen tasks. + +Quantitative comparison. Tab. 1 demonstrates the impact of in-context learning on different image generation tasks. Under the canny condition, our method without in-context examples achieves an FID of 30.60, which improves to 31.15 with two in-context examples. When conditioned on depth, the RMSE decreases from 10.31 to 9.68 as the number of in-context examples increases, indicating enhanced structural consistency. Similarly, in the deblurring task, + +![](images/fdc6f40526c0828bd03a490243e160f00d4078c842701d795d553f0adc7b0242.jpg) +Figure 7. Effect of in-context learning on style transfer, best to zoom in to view. The images in the red box mean style references. Without in-context learning, semantic-variant style transfer (bottom) does not change the semantics as expected. + +RMSE decreases from 26.53 to 25.57, reflecting improved fidelity to the original content. These results highlight in-context learning as an effective guidance mechanism, enabling the model to better align with the task intent. + +In-domain task ambiguity. We observe that the model occasionally experiences task confusion, failing to accurately interpret the intended objective. In-context learning effectively alleviates this issue by providing task-specific demonstrates. As shown in Fig. 7, without in-context examples, the model struggles to distinguish between semantic-invariant and semantic-variant style transfers, leading to incorrect results where the background and semantics remain unchanged in semantic-variant cases. However, incorporating just a single in-context example enables the model to correctly perform semantic-variant style transfer. Similarly, as shown in Fig. 8, increasing the number of in-context examples enhances the performance. Additional visualizations are provided in the supplementary materials. + +Generalization to unseen tasks. Beyond mitigating task confusion, in-context learning also improves the ability of models to generalize to tasks unseen during training. As shown in Fig. 9, we evaluate the performance in generating frontal faces from side-view images, a task not encountered during training. Without in-context examples, the generated faces fail to align with the expected frontal view. However, as the number of in-context examples increases, the model progressively refines its outputs, producing more accurate frontal faces. These results demonstrate that in-context learning serves as an effective guidance mechanism, enabling adaptation to novel tasks without retraining. + +# 6. Conclusion + +In this work, we propose VisualCloze, a universal image generation framework that addresses key challenges in ex + +![](images/706b758e6cafec5cb4f98b4b30973d9f9c8c23cb1a56839fc4c27629e842847e.jpg) +Figure 8. Effect of in-context learning on dense prediction. + +![](images/aca6e91bb2e9a4a88261fdff67073fc5cdb9a83f7d5bc89f343a07f55313cddf.jpg) +Figure 9. Effect of in-context learning on the unseen task: generating frontal face conditioning on side faces. More in-context examples produce a more accurate frontal face. + +isting methods, including generalizable instruction design, appropriate task distributions, and unified architectural design. Rather than relying solely on language-based instructions to convey task intent, we re-propose visual in-context learning, enabling the model to learn tasks from demonstrations. This approach improves generalization to unseen tasks and reduce task ambiguity. To overcome the sparsity of visual task distributions, which limits the learning of transferable knowledge, we construct Graph200K, a graph-structured dataset that establishes interrelated tasks. In this compact task space, the model is promoted to learn transferable representations and improve adaptability. Meanwhile, we identify the consistent optimization objective between image infilling and universal generation, allowing us to seamlessly adapt general-purpose infilling models for universal generation without architectural modifications. Experimental results show that our approach supports a diverse set of in-domain tasks using in-context learning while demonstrating strong generalization to unseen tasks. + +Acknowledgments. This work is supported by the Shenzhen Science and Technology Program (JCYJ20240813114237048), and the National Natural Science Foundation of China (Grant No. 62206272). + +# References + +[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, Roman Ring, Eliza Rutherford, Serkan Cabi, Tengda Han, Zhitao Gong, Sina Samangooei, Marianne Monteiro, Jacob Menick, Sebastian Borgeaud, Andrew Brock, Aida Nematzadeh, Sahand Sharifzadeh, Mikolaj Binkowski, Ricardo Barreira, Oriol Vinyals, Andrew Zisserman, and Karen Simonyan. Flamingo: a visual language model for few-shot learning. In NeurIPS, 2022. 2, 3 +[2] Michael Samuel Albergo and Eric Vanden-Eijnden. Building normalizing flows with stochastic interpolants. In ICLR, 2023. 3 +[3] Ivana Balazevic, David Steiner, Nikhil Parthasarathy, Relja Arandjelovic, and Olivier J Henaff. Towards in-context scene understanding. In NeurIPS, 2023. 2, 3 +[4] Amir Bar, Yossi Gandelsman, Trevor Darrell, Amir Globerson, and Alexei A Efros. Visual prompting via image inpainting. In NeurIPS, 2022. 2, 3 +[5] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. *NeurIPS*, 2020. 2, 3, 5 +[6] John Canny. A computational approach to edge detection. IEEE TPAMI, 1986. 4 +[7] Z. Cao, G. Hidalgo Martinez, T. Simon, S. Wei, and Y. A. Sheikh. Openpose: Realtime multi-person 2d pose estimation using part affinity fields. IEEE TPAMI, 2019. 4 +[8] Xi Chen, Zhifei Zhang, He Zhang, Yuqian Zhou, Soo Ye Kim, Qing Liu, Yijun Li, Jianming Zhang, Nanxuan Zhao, Yilin Wang, Hui Ding, Zhe Lin, and Hengshuang. Unireal: Universal image generation and editing via learning real-world dynamics. arXiv preprint arXiv:2412.07774, 2024. 3 +[9] Zhao Chen, Vijay Badrinarayanan, Chen-Yu Lee, and Andrew Rabinovich. Gradnorm: Gradient normalization for adaptive loss balancing in deep multitask networks. In ICML, 2018. 2 +[10] Seunghwan Choi, Sunghyun Park, Minsoo Lee, and Jaegul Choo. Viton-hd: High-resolution virtual try-on via misalignment-aware normalization. In CVPR, 2021. 1 +[11] Zheng Chong, Xiao Dong, Haoxiang Li, shiyue Zhang, Wenqing Zhang, Hanqing Zhao, xujie zhang, Dongmei Jiang, and Xiaodan Liang. CatVTON: Concatenation is all you need for virtual try-on with diffusion models. In ICLR, 2025. 1 +[12] Prafulla Dhariwal and Alexander Quinn Nichol. Diffusion models beat GANs on image synthesis. In NeurIPS, 2021. 3 +[13] Qingxiu Dong, Lei Li, Damai Dai, Ce Zheng, Jingyuan Ma, Rui Li, Heming Xia, Jingjing Xu, Zhiyong Wu, Tianyu Liu, Baobao Chang, Xu Sun, Lei Li, and Zhifang Sui. A survey on in-context learning. arXiv preprint arXiv:2301.00234, 2024. 3 +[14] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, Dustin Podell, Tim + +Dockhorn, Zion English, Kyle Lacey, Alex Goodwin, Yannik Marek, and Robin Rombach. Scaling rectified flow transformers for high-resolution image synthesis. arXiv preprint arXiv:2403.03206, 2024. 1, 3 +[15] Christopher Fifty, Ehsan Amid, Zhe Zhao, Tianhe Yu, Rohan Anil, and Chelsea Finn. Efficiently identifying task groupings for multi-task learning. In NeurIPS, 2021. 2 +[16] Peng Gao, Le Zhuo, Dongyang Liu, Ruoyi Du, Xu Luo, Longtian Qiu, Yuhang Zhang, Chen Lin, Rongjie Huang, Shijie Geng, et al. Lumina-t2x: Transforming text into any modality, resolution, and duration via flow-based large diffusion transformers. arXiv preprint arXiv:2405.05945, 2024. 3 +[17] Shanghua Gao, Pan Zhou, Ming-Ming Cheng, and Shuicheng Yan. Mdtv2: Masked diffusion transformer is a strong image synthesizer. arXiv preprint arXiv:2303.14389, 2024. 3 +[18] Golnaz Ghiasi, Barret Zoph, Ekin D. Cubuk, Quoc V. Le, and Tsung-Yi Lin. Multi-task self-training for learning general representations. In ICCV, 2021. 2 +[19] Geonmo Gu, Byungsoo Ko, SeoungHyun Go, Sung-Hyun Lee, Jingeun Lee, and Minchul Shin. Towards light-weight and real-time line segment detection. In AAAI, 2022. 4 +[20] Aaron Hertzmann. Algorithms for rendering in artistic styles. PhD thesis, New York University, Graduate School of Arts and Science, 2001. 3 +[21] Aaron Hertzmann, Charles E. Jacobs, Nuria Oliver, Brian Curless, and David H. Salesin. Image analogies. In Proceedings of the 28th Annual Conference on Computer Graphics and Interactive Techniques, 2001. 3 +[22] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6 +[23] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 3 +[24] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In ICLR, 2022. 6 +[25] Lianghua Huang, Wei Wang, Zhi-Fan Wu, Huanzhang Dou, Yupeng Shi, Yutong Feng, Chen Liang, Yu Liu, and Jingren Zhou. Group diffusion transformers are unsupervised multitask learners. arXiv preprint arxiv:2410.15027, 2024.3 +[26] Lianghua Huang, Wei Wang, Zhi-Fan Wu, Yupeng Shi, Huanzhang Dou, Chen Liang, Yutong Feng, Yu Liu, and Jingren Zhou. In-context lora for diffusion transformers. arXiv preprint arxiv:2410.23775, 2024. 1, 2 +[27] Junjie Ke, Qifei Wang, Yilin Wang, Peyman Milanfar, and Feng Yang. Musiq: Multi-scale image quality transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5148-5157, 2021. 6 +[28] Alex Kendall, Yarin Gal, and Roberto Cipolla. Multi-task learning using uncertainty to weigh losses for scene geometry and semantics. In CVPR, 2018. 2 + +[29] Iasonas Kokkinos. Ethernet: Training a universal convolutional neural network for low-, mid-, and high-level vision using diverse datasets and limited memory. In CVPR, 2017. 2 +[30] Black Forest Labs. Flux. https://github.com/black-forest-labs/flux, 2024.1, 2, 3, 4, 5, 6, 7 +[31] Bolin Lai, Felix Juefei-Xu, Miao Liu, Xiaoliang Dai, Nikhil Mehta, Chenguang Zhu, Zeyi Huang, James M Rehg, Sang-min Lee, Ning Zhang, et al. Unleashing in-context learning of autoregressive models for few-shot image manipulation. arXiv preprint arXiv:2412.01027, 2024. 3 +[32] Duong H. Le, Tuan Pham, Sangho Lee, Christopher Clark, Aniruddha Kembhavi, Stephan Mandt, Ranjay Krishna, and Jiasen Lu. One diffusion to generate them all, 2024. 1, 2, 3, 6 +[33] Kunchang Li, Yali Wang, Junhao Zhang, Peng Gao, Guanglu Song, Yu Liu, Hongsheng Li, and Yu Qiao. Uniformer: Unifying convolution and self-attention for visual recognition. IEEE TPAMI, 2023. 4 +[34] Zhen Li, Mingdeng Cao, Xintao Wang, Zhongang Qi, MingMing Cheng, and Ying Shan. Photomaker: Customizing realistic human photos via stacked id embedding. In CVPR, 2024. 1, 3, 5 +[35] Weifeng Lin, Xinyu Wei, Renrui Zhang, Le Zhuo, Shitian Zhao, Siyuan Huang, Junlin Xie, Yu Qiao, Peng Gao, and Hongsheng Li. Pixwizard: Versatile image-to-image visual assistant with open-language instructions. arXiv preprint arXiv:2409.15278, 2024. 1, 5 +[36] Yaron Lipman, Ricky T. Q. Chen, Heli Ben-Hamu, Maximilian Nickel, and Matthew Le. Flow matching for generative modeling. In ICLR, 2023. 3 +[37] Dongyang Liu, Shitian Zhao, Le Zhuo, Weifeng Lin, Yu Qiao, Hongsheng Li, and Peng Gao. Lumina-mgpt: Illuminate flexible photorealistic text-to-image generation with multimodal generative pretraining. arXiv preprint arXiv:2408.02657, 2024. 3 +[38] Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow. arXiv preprint arXiv:2209.03003, 2022. 3 +[39] Yihao Liu, Xiangyu Chen, Xianzheng Ma, Xintao Wang, Jiantao Zhou, Yu Qiao, and Chao Dong. Unifying image processing as visual prompting question answering. arXiv preprint arXiv:2310.10513, 2023. 2, 3, 5 +[40] Chaojie Mao, Jingfeng Zhang, Yulin Pan, Zeyinzi Jiang, Zhen Han, Yu Liu, and Jingren Zhou. Ace++: Instruction-based image creation and editing via context-aware content filling. arXiv preprint arXiv:2501.02487, 2025. 1, 3 +[41] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equations. arXiv preprint arXiv:2108.01073, 2021. 6 +[42] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 7 +[43] William Peebles and Saining Xie. Scalable diffusion models with transformers. In ICCV, 2023. 3 + +[44] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PmLR, 2021. 6, 7 +[45] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, Eric Mintun, Junting Pan, Kalyan Vasudev Alwala, Nicolas Carion, Chao-Yuan Wu, Ross Girshick, Piotr Dólár, and Christoph Feichtenhofer. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024. 4 +[46] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, 2022. 3 +[47] Sebastian Ruder. An overview of multi-task learning in deep neural networks. arXiv preprint arXiv:1706.05098, 2017. 2 +[48] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In CVPR, 2023. 1 +[49] Dianmo Sheng, Dongdong Chen, Zhentao Tan, Qiankun Liu, Qi Chu, Jianmin Bao, Tao Gong, Bin Liu, Shengwei Xu, and Nenghai Yu. Towards more unified in-context visual understanding. In CVPR, 2024. 3 +[50] Peize Sun, Yi Jiang, Shoufa Chen, Shilong Zhang, Bingyue Peng, Ping Luo, and Zehuan Yuan. Autoregressive model beats diffusion: Llama for scalable image generation. arXiv preprint arXiv:2406.06525, 2024. 3 +[51] Yanpeng Sun, Qiang Chen, Jian Wang, Jingdong Wang, and Zechao Li. Exploring effective factors for improving visual in-context learning. arXiv preprint arXiv:2304.04748, 2023. 3 +[52] Yasheng SUN, Yifan Yang, Houwen Peng, Yifei Shen, Yuqing Yang, Han Hu, Lili Qiu, and Hideki Koike. Imagebrush: Learning visual in-context instructions for exemplar-based image manipulation. In NeurIPS, 2023. 2, 3, 5 +[53] Zhenxiong Tan, Songhua Liu, Xingyi Yang, Qiaochu Xue, and Xinchao Wang. *Omnicontrol: Minimal and universal control for diffusion transformer.* arXiv preprint arXiv:2411.15098, 3, 2024. 3, 4, 5, 6, 7 +[54] Alex Jinpeng Wang, Linjie Li, Yiqi Lin, Min Li, Lijuan Wang, and Mike Zheng Shou. Leveraging visual tokens for extended text contexts in multi-modal learning. NeurIPS, 2024. 3 +[55] Haofan Wang, Peng Xing, Renyuan Huang, Hao Ai, Qixun Wang, and Xu Bai. Instantstyle-plus: Style transfer with content-preserving in text-to-image generation. arXiv preprint arXiv:2407.00788, 2024. 1, 3, 4, 6 +[56] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Yang Fan, Kai Dang, Mengfei Du, Xuancheng Ren, Rui Men, Dayiheng Liu, Chang Zhou, Jingren Zhou, and Junyang Lin. Qwen2-vl: Enhancing vision-language model's + +perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 4 +[57] Xinlong Wang, Wen Wang, Yue Cao, Chunhua Shen, and Tiejun Huang. Images speak in images: A generalist painter for in-context visual learning. In CVPR, 2023. 2, 3 +[58] Xinlong Wang, Xiaosong Zhang, Yue Cao, Wen Wang, Chunhua Shen, and Tiejun Huang. Seggpt: Towards segmenting everything in context. In ICCV, 2023. 2, 3 +[59] Zhendong Wang, Yifan Jiang, Yadong Lu, yelong shen, Pengcheng He, Weizhu Chen, Zhangyang Wang, and Mingyuan Zhou. In-context learning unlocked for diffusion models. In NeurIPS, 2023. 3 +[60] Cong Wei, Zheyang Xiong, Weiming Ren, Xinrun Du, Ge Zhang, and Wenhu Chen. Omniedit: Building image editing generalist models through specialist supervision. arXiv preprint arXiv:2411.07199, 2024. 1, 2, 5 +[61] Shitao Xiao, Yueze Wang, Junjie Zhou, Huaying Yuan, Xingrun Xing, Ruiran Yan, Shuting Wang, Tiejun Huang, and Zheng Liu. Omnigen: Unified image generation. arXiv preprint arXiv:2409.11340, 2024. 1, 2, 3, 6, 7 +[62] Saining Xie and Zhuowen Tu. Holistically-nested edge detection. In CVPR, 2015. 4 +[63] Haofei Xu, Jing Zhang, Jianfei Cai, Hamid Rezatofighi, Fisher Yu, Dacheng Tao, and Andreas Geiger. Unifying flow, stereo and depth estimation. IEEE TPAMI, 2023. 4 +[64] Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything v2. arXiv:2406.09414, 2024. 4 +[65] Sidi Yang, Tianhe Wu, Shuwei Shi, Shanshan Lao, Yuan Gong, Mingdeng Cao, Jiahao Wang, and Yujiu Yang. Maniaq: Multi-dimension attention network for no-reference image quality assessment. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1191-1200, 2022. 6 +[66] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023. 2 +[67] Fanghua Yu, Jinjin Gu, Zheyuan Li, Jinfan Hu, Xiangtao Kong, Xintao Wang, Jingwen He, Yu Qiao, and Chao Dong. Scaling up to excellence: Practicing model scaling for photo-realistic image restoration in the wild. arXiv preprint arXiv:2401.13627, 2024. 2 +[68] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, Ben Hutchinson, Wei Han, Zarana Parekh, Xin Li, Han Zhang, Jason Baldridge, and Yonghui Wu. Scaling autoregressive models for content-rich text-to-image generation. Transactions on Machine Learning Research, 2022. 3 +[69] Hayoung Yun and Hanjoo Cho. Achievement-based training progress balancing for multi-task learning. In ICCV, 2023. 2 +[70] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, 2023. 2, 3, 4, 6 +[71] Yuxin Zhang, Nisha Huang, Fan Tang, Haibin Huang, Chongyang Ma, Weiming Dong, and Changsheng Xu. + +Inversion-based style transfer with diffusion models. In CVPR, 2023. 1, 2, 7 +[72] Yuanhan Zhang, Kaiyang Zhou, and Ziwei Liu. What makes good examples for visual in-context learning? In NeurIPS, 2023. 2, 3 +[73] Canyu Zhao, Mingyu Liu, Huanyi Zheng, Muzhi Zhu, Zhiyue Zhao, Hao Chen, Tong He, and Chunhua Shen. Disception: A generalist diffusion model for visual perceptual tasks. arXiv preprint arXiv:2502.17157, 2025. 3 +[74] Peng Zheng, Dehong Gao, Deng-Ping Fan, Li Liu, Jorma Laaksonen, Wanli Ouyang, and Nicu Sebe. Bilateral reference for high-resolution dichotomous image segmentation. CAAI Artificial Intelligence Research, 2024. 4 +[75] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In CVPR, 2017. 2 +[76] Yucheng Zhou, Xiang Li, Qianning Wang, and Jianbing Shen. Visual in-context learning for large vision-language models. arXiv preprint arXiv:2402.11574, 2024. 3 +[77] Muzhi Zhu, Yang Liu, Zekai Luo, Chenchen Jing, Hao Chen, Guangkai Xu, Xinlong Wang, and Chunhua Shen. Unleashing the potential of the diffusion model in few-shot semantic segmentation. In NeurIPS, 2024. 2, 3 +[78] Le Zhuo, Ruoyi Du, Han Xiao, Yangguang Li, Dongyang Liu, Rongjie Huang, Wenze Liu, Xiangyang Zhu, Fu-Yun Wang, Zhanyu Ma, Xu Luo, Zehan Wang, Kaipeng Zhang, Lirui Zhao, Si Liu, Xiangyu Yue, Wanli Ouyang, Yu Qiao, Hongsheng Li, and Peng Gao. Lumina next: Making lumina-t2x stronger and faster with next-dit. In NeurIPS, 2024. 1, 2, 3 \ No newline at end of file diff --git a/ICCV/2025/VisualCloze_ A Universal Image Generation Framework via Visual In-Context Learning/images.zip b/ICCV/2025/VisualCloze_ A Universal Image Generation Framework via Visual In-Context Learning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..0f9054106252f7e244f89491bf144fb62ff32233 --- /dev/null +++ b/ICCV/2025/VisualCloze_ A Universal Image Generation Framework via Visual In-Context Learning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c8316fd6841849bb5613d8b360bafcc3a1116b5c8aa6f08fe46de0225de97c3d +size 832478 diff --git a/ICCV/2025/VisualCloze_ A Universal Image Generation Framework via Visual In-Context Learning/layout.json b/ICCV/2025/VisualCloze_ A Universal Image Generation Framework via Visual In-Context Learning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..4929db5a9811e5fc7ff6c0f3d5c366e135fcb978 --- /dev/null +++ b/ICCV/2025/VisualCloze_ A Universal Image Generation Framework via Visual In-Context Learning/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:33a125207291ba8f6efd95db4b962b5c490fa9348248bc6de399a32eeb0b7c4a +size 396506 diff --git a/ICCV/2025/Vivid4D_ Improving 4D Reconstruction from Monocular Video by Video Inpainting/334ccb81-eef2-4518-a459-c4c4b31dd8c9_content_list.json b/ICCV/2025/Vivid4D_ Improving 4D Reconstruction from Monocular Video by Video Inpainting/334ccb81-eef2-4518-a459-c4c4b31dd8c9_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..5b3d1ac94d73ea87e80b51d237767c9b430b83e5 --- /dev/null +++ b/ICCV/2025/Vivid4D_ Improving 4D Reconstruction from Monocular Video by Video Inpainting/334ccb81-eef2-4518-a459-c4c4b31dd8c9_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1c234dcf62f5372b5d44dc57d599133807f1aa558e7d5012a3a305641a612daa +size 98520 diff --git a/ICCV/2025/Vivid4D_ Improving 4D Reconstruction from Monocular Video by Video Inpainting/334ccb81-eef2-4518-a459-c4c4b31dd8c9_model.json b/ICCV/2025/Vivid4D_ Improving 4D Reconstruction from Monocular Video by Video Inpainting/334ccb81-eef2-4518-a459-c4c4b31dd8c9_model.json new file mode 100644 index 0000000000000000000000000000000000000000..0f708c3c87ca3be70181249b9e49dee70a3f87a8 --- /dev/null +++ b/ICCV/2025/Vivid4D_ Improving 4D Reconstruction from Monocular Video by Video Inpainting/334ccb81-eef2-4518-a459-c4c4b31dd8c9_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:542a5324bc48225e66c747caf939494ff8dbf35782bc44d6f69e113470af5b7f +size 131054 diff --git a/ICCV/2025/Vivid4D_ Improving 4D Reconstruction from Monocular Video by Video Inpainting/334ccb81-eef2-4518-a459-c4c4b31dd8c9_origin.pdf b/ICCV/2025/Vivid4D_ Improving 4D Reconstruction from Monocular Video by Video Inpainting/334ccb81-eef2-4518-a459-c4c4b31dd8c9_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..39c953ddbf1aec6603837a1e3c2eeec888b62ac4 --- /dev/null +++ b/ICCV/2025/Vivid4D_ Improving 4D Reconstruction from Monocular Video by Video Inpainting/334ccb81-eef2-4518-a459-c4c4b31dd8c9_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6ebe67a7369f4d4de79c6708f7cf8c66f41535bcdc7ce9bcfcaa41e40dff92ff +size 19641114 diff --git a/ICCV/2025/Vivid4D_ Improving 4D Reconstruction from Monocular Video by Video Inpainting/full.md b/ICCV/2025/Vivid4D_ Improving 4D Reconstruction from Monocular Video by Video Inpainting/full.md new file mode 100644 index 0000000000000000000000000000000000000000..cf00cf84f7175a7c1fe93bd95eb0173e7cb4d192 --- /dev/null +++ b/ICCV/2025/Vivid4D_ Improving 4D Reconstruction from Monocular Video by Video Inpainting/full.md @@ -0,0 +1,400 @@ +# Vivid4D: Improving 4D Reconstruction from Monocular Video by Video Inpainting + +Jiaxin Huang $^{1}$ , Sheng Miao $^{1}$ , Bangbang Yang $^{2}$ , Yuewen Ma $^{2}$ , Yiyi Liao $^{1\boxtimes}$ + +1 Zhejiang University 2 ByteDance + +Project Page: https://xdimlab.github.io/Vivid4D/ + +![](images/87ba887f58ebdd3582b4b970b04d29d25dbe4fec3dd56763962c7b7f0e1106b5.jpg) +Figure 1. Vivid4D. We improve dynamic scene reconstruction from casually captured monocular videos by synthesizing augmented views. Our approach integrates both geometric and generative priors to reformulate the video augmentation as a video inpainting task. This enables our method to effectively complete invisible regions in the scene and enhance reconstruction quality. + +# Abstract + +Reconstructing 4D dynamic scenes from casually captured monocular videos is valuable but highly challenging, as each timestamp is observed from a single viewpoint. We introduce Vivid4D, a novel approach that enhances 4D monocular video synthesis by augmenting observation views — synthesizing multi-view videos from a monocular input. Unlike existing methods that either solely leverage geometric priors for supervision or use generative priors while overlooking geometry, we integrate both. This reformulates view augmentation as a video inpainting task, where observed views are warped into new viewpoints based on monocular depth priors. To achieve this, we train a video inpainting model on unposed web videos with synthetically generated masks that mimic warping occlusions, ensuring spatially and temporally consistent completion of missing regions. To further mitigate inaccuracies in monocular depth priors, we introduce an iterative view augmentation strategy and a robust reconstruction loss. Experiments demonstrate that our method effectively improves monocular 4D scene reconstruction and completion. + +# 1. Introduction + +Reconstructing dynamic 3D scenes from convenient monocular video captures is a highly valuable task in computer vision and graphics, yet remains a profound challenge. Recent advances in 4D reconstruction [9, 17, 21, 23, 44, 50, 61, 83, 96] have made significant progress in modeling dynamic scenes, but they typically rely on synchronized multiview cameras. + +Given the sparse observations of monocular videos where each timestamp is observed from a single viewpoint, existing methods attempt to address this challenging task by incorporating additional priors. One line of studies has explored various geometric priors for monocular 4D reconstruction, including geometric constraints [57, 58], optical flow [41, 42, 81], depth estimation [24, 47], and tracking [38, 69, 78]. However, existing methods yield suboptimal results for two main reasons. First, these auxiliary supervisions are not consistently reliable and may not correlate linearly with rendering performance; for example, a small depth error may lead to significant color shifts in regions with texture changes. Second, these priors, which are derived solely from input views, provide limited guidance for regions occluded or unobserved in the input views. More recently, video diffusion models have emerged as an + +![](images/e4c0ce98f481e7dd6fac93e27e2104c360da413a673bd66062391f359b60f96f.jpg) +Figure 2. Video Inpainting for 4D Reconstruction. To train the video inpainting model, we use 2D tracking to generate masked training pairs from unposed web videos. During 4D reconstruction, we warp the monocular video to novel viewpoints, creating masked videos that our inpainting diffusion model then completes. + +other type of prior for such tasks, owing to their impressive ability to generate realistic videos [7, 36, 48, 97]. These models can generate plausible RGB images for unknown viewpoints, overcoming the disadvantages of using geometry priors alone. However, existing methods are either limited to static scenes [46, 52, 85, 101, 103], or require a large amount of posed images for training [84, 109], while collecting such training data is highly challenging as dynamic scenes' camera poses are difficult to obtain. + +In this work, we propose Vivid4D, a novel approach for 4D monocular reconstruction that integrates geometric priors from input videos with the generative capabilities of video diffusion models. Instead of directly generating novel views, we reformulate view augmentation as a video inpainting task: first, observed views are warped to unseen viewpoints based on geometric priors; then, a video diffusion model inpaints missing regions to ensure spatial-temporal consistency. This allows us to train the video inpainting model using unposed web videos by mimicking warping occlusions based on 2D trackings. Specifically, our method includes two key steps as shown in Fig. 2. First, we train a video inpainting model tailored for our view augmentation task that learns to in-paint occluded regions in a spatial-temporally coherent manner (Fig. 2 top). This model is trained on unposed web videos, using 2D trackings to synthesize occlusion masks. Next, we use the trained video inpainting model for view augmentation in monocular-video-based 4D reconstruction, leveraging monocular depth priors to warp the video to unseen viewpoints, which our inpainting model then completes (Fig. 2 bottom). To maximize scene coverage while compensating for inaccurate monocular depth priors, we propose a progressive view expansion strategy that iteratively generates augmented views, effectively accumulating comprehensive multi-view observations. When further combined with a robust reconstruction loss, the augmented views enables Vivid4D to improve 4D reconstruction while maintaining geometric consistency and temporal coherence. + +The main contributions of our work can be summarized as follows: 1) We propose Vivid4D, a novel method for monocular 4D scene reconstruction that leverages multiview supervision from both video diffusion models and depth-guided warping. 2) We reformulate view augmentation as a video inpainting task, enabling efficient training on unposed web videos through synthesized occlusion masks based on 2D tracking. 3) We design an iterative view augmentation strategy to progressively expand the observation viewpoint of the scene, enabling robust reconstruction and novel view synthesis. 4) We conduct extensive experiments across various dynamic scenes and show that our method improves the quality of novel view synthesis compared to existing approaches on different types of motion and scene complexity. + +# 2. Related Work + +3D Novel View Synthesis. Recent advances in neural rendering, from NeRF [3-5, 53, 54, 77] to 3D GS [10, 28, 32, 34, 49, 89, 102, 110], have revolutionized static scene reconstruction and novel view synthesis. To improve reconstruction quality, especially in challenging scenarios such as sparse inputs, researchers have explored two main directions: integrating traditional geometric reconstruction techniques [14, 20, 105] and applying geometric regularizations [35, 39, 56, 61, 75, 92]. Recently, generative models have been leveraged for additional visual supervision. For instance, several methods [12, 88, 112] employ score distillation sampling (SDS) [60] to distill knowledge from pre-trained diffusion models, optimizing the 3D representation to align with learned priors. Another line of work [46, 52, 59, 67, 85, 99, 101, 103] projects dense scene representations [34, 54, 67, 80] and projects them into novel viewpoints, guiding reconstruction with geometrically consistent views generated by diffusion models. Meanwhile, [45] incorporates 3D structure conditioning into video diffusion for 3D-consistent video generation, and [26] generates multi-view images conditioned on input views and target camera poses to provide additional supervision. While these methods achieve impressive results, they are typically limited to static scenes and struggle with dynamic objects and temporal consistency. + +4D Monocular Novel View Synthesis. Many methods have extended neural rendering to dynamic scenes, typically based on canonical space modeling with time-dependent deformation fields [21, 44, 47, 50, 61, 83] or 4D time-varying representations [9, 17, 23, 62, 86, 96]. However, these methods typically require multi-view synchronized and calibrated videos, making them ill-posed when multiview data is limited. + +For monocular video reconstruction, some methods leverage data-driven priors like depth [68, 90, 93], + +tracks [91], and geometric constraints [38, 42, 47, 58, 69, 78, 81] for initialization or supervision, but these priors may not be fully accurate and do not always correlate linearly with RGB performance. More recently, diffusion models have been introduced for 4D reconstruction. For example, [15] utilizes pre-trained diffusion models [64] for scene completion, though the visual quality still requires improvement. Concurrent work [113] uses LGM [71] to generate pseudo novel views as additional supervision, but it is primarily applied at object level. Other methods [70, 84, 109] generate spatial-temporally consistent frames to optimize 4D scenes, but they require temporally and spatially decoupled data to learn disentangled control over camera motion and time. Additionally, [84, 109] rely on camera motion as input for training, which can be challenging to obtain for dynamic video datasets. In contrast, our approach only requires unposed web videos with simulated warping occlusion masks, enabling training on a wider range of casual video datasets without camera pose information. Moreover, [70] lacks precise camera control during inference, while our method implicitly incorporates precise pose information through warping. + +Video Diffusion Models. Video diffusion models have shown remarkable capabilities in video generation tasks. Recent works focused on video inpainting [27, 37, 40, 95, 107, 111, 114] and out-painting [11, 19, 74] have demonstrated the effectiveness of diffusion models in generating temporally consistent content for incomplete videos. However, these methods mainly deal with a single video without perspective changes. Other generative methods [1, 73] take camera poses as input to generate synchronized videos from novel viewpoints, but their quality needs improvement, and they are constrained by static input-output camera viewpoints. + +Some works [6, 108] attempt to overcome viewpoint constraints by inpainting videos after perspective changes. These methods utilize geometric constraints [63, 100, 108] or pseudo 4D Gaussian fields [6] to obtain masked videos after warping, and fine-tune video diffusion models [7, 30, 97] for multi-view synthesis. While these approaches show the potential of diffusion models in generating consistent content for partially observed videos, they rely on expensive training data from video warping or 4D Gaussian field creation. Moreover, they focus primarily on camera-controlled video generation [2, 29] rather than 4D reconstruction, which does not guarantee consistency or efficiency for novel view playback like Gaussian Splatting. In contrast, our method simply uses a pre-trained tracking model [33] to generate masked training data and finetunes the model to provide additional supervision for 4D monocular reconstruction through iterative view expansion. Note that [52, 101] also employ diffusion models to inpaint + +warped videos, but they are limited to static scenes. + +# 3. Method + +We present a novel approach for improving monocular 4D scene reconstruction through view augmentation using monocular depth priors and video diffusion models. The key to enhancing the 4D monocular video synthesis is to warp the monocular video to augmented views based on the depth prior, followed by a video inpainting model to fill occluded regions. Our key components include a video inpainting model tailored for the task of view augmentation from a single monocular video (Sec. 3.1), and an iterative augmented view generation strategy (Sec. 3.2). The original and the augmented videos are then used for 4D reconstruction (Sec. 3.3). Fig. 3 illustrates our 4D reconstruction based on view augmentation. + +# 3.1. Video Inpainting Diffusion Model + +We aim to learn a video inpainting diffusion model to inpaint the occluded or invisible areas in a masked video while ensuring spatial-temporal consistency. To align with our core objective of 4D reconstruction, it is essential to provide training data in which the regions requiring inpainting primarily result from camera motion. Unlike previous methods, which require known camera poses [84, 101], or apply random mask to input videos [52], our method leverages a pre-trained 2D tracking model [33] for generating the training data for the video inpainting model, allowing us to leverage casual, unposed videos for training. + +Training Data Preparation: Given an unposed input video sequence $\mathcal{V} = \{I_t\}_{t=1}^T$ with $T$ frames, we first employ a pre-trained 2D tracking model [33] to establish point correspondences across frames. Let $\mathcal{P}_1 = \{p_1^i\}_{i=1}^N$ denote the set of $N$ pixel points sampled from the first frame $I_1$ . The tracking model generates trajectories $\mathcal{T} = \{(p_t^i)_{t=1}^T\}_{i=1}^N$ that record the positions of these initial points throughout the video sequence. For each subsequent frame $I_t$ ( $t > 1$ ), we identify pixels that are not associated with any tracked point from $\mathcal{P}_1$ . These untracked pixels form our mask region $\mathcal{M}_t$ , as they likely correspond to areas that were previously occluded or newly revealed in the scene. Formally, we define the mask for frame $t$ as: + +$$ +\mathcal {M} _ {t} = \{q \in I _ {t} \mid q \notin \left\{p _ {t} ^ {i} \right\} _ {i = 1} ^ {N} \} +$$ + +where $\{p_t^i\}_{i = 1}^N$ represents the set of tracked points in frame $t$ . The masked frame can be obtained as: + +$$ +\mathcal {V} ^ {m} = \left\{I _ {t} ^ {m} \right\} _ {t = 1} ^ {T}, \quad I _ {t} ^ {m} = I _ {t} \odot \mathcal {M} _ {t} +$$ + +where $\odot$ indicates element-wise multiplication. This approach naturally captures regions that become visible due to object motion or camera movement, providing ideal training data for our video inpainting task. + +Anchor-Conditioned Video Inpainting Diffusion Model: + +![](images/1499863170fc464b37265d2ed7ed7a57224752369ecf909b1cdc16d56c7d5ba6.jpg) +Figure 3. 4D reconstruction based on view augmentation. Given an input monocular video, we first perform sparse reconstruction to obtain camera poses and align monocular depth to metric scale, forming an initial data buffer $\mathcal{D}_0$ . In each iterative view augmentation step, we select frames at each timestamp from the previous buffer $\mathcal{D}_{j-1}$ and warp them to novel viewpoints using pre-defined camera poses $\mathbf{T}$ , creating new perspective videos with continuous invisible region masks. These masked videos, along with binary masks and an anchor video, are fed into our pre-trained anchor-conditioned video inpainting diffusion model to produce completed novel-view videos. We update the buffer $\mathcal{D}_j$ with these enhanced videos, their metric depths and poses. Finally, both the original monocular video and all synthesized multi-view videos are used to supervise 4D scene reconstruction. + +Given the masked video $\mathcal{V}^m$ and corresponding binary mask $\mathcal{M}$ , we introduce a video diffusion model that inpaints occluded or invisible regions with spatial-temporal consistency while incorporating priors from unwarped videos. To leverage pre-trained video diffusion models effectively, we fine-tune an existing model [51, 76] and adapt its architecture for video inpainting in view augmentation. Our modifications include processing the masked input and conditioning the model on unwarped videos to utilize full observed frames for context. + +Specifically, we modify the standard video diffusion model in two ways. First, we extend the input processing to handle masked videos by encoding the incomplete video sequence through a VAE encoder, producing a 4-channel latent representation $\mathbf{z}_m\in \mathbb{R}^{T\times H\times W\times 4}$ that is concatenated with the corresponding resized binary mask $\mathcal{M}'\in \mathbb{R}^{T\times H\times W\times 1}$ [64] and the noisy latent $\mathbf{z}_t\in \mathbb{R}^{T\times H\times W\times 4}$ . Second, we introduce an anchor video mechanism, where frames sampled from the same video clip but at different timestamps serve as additional spatial-temporal context. Note that in our later 4D reconstruction stage, we use the original, unwarped monocular video as the anchor video. The anchor video is similarly encoded through the VAE into latent space $\mathbf{z}_a\in \mathbb{R}^{T\times H\times W\times 4}$ . During training, our modified U-Net receives a latent input $\mathbf{z} = [\mathbf{z}_t;\mathbf{z}_m;\mathcal{M}';\mathbf{z}_a]$ , where these maps are concatenated along the channel dimension. While $\mathbf{z}_a$ are not spatially aligned with other maps using this simple concatenation, our 3D U-Net can effectively learn spatio-temporal contexts by fusing local struc + +tures via spatio-temporal convolutions and identifying related regions through attention operations. The model is trained with the following objective: + +$$ +\mathcal {L} = \mathbb {E} _ {\mathbf {x}, t, \epsilon \sim \mathcal {N} (0, 1)} \| \epsilon - \epsilon_ {\theta} (\mathbf {z} _ {t}, t, \mathbf {z} _ {m}, \mathcal {M} ^ {\prime}, \mathbf {z} _ {a}) \| _ {2} ^ {2} +$$ + +where $\mathbf{z}_t = \alpha_t\mathbf{x} + \sigma_t\epsilon$ is the noised latent of the ground truth video $\mathbf{x}$ , and $t$ is the diffusion timestep. + +# 3.2. Diffusion-based View Augmentation + +Given the trained video inpainting model, we augment the monocular input used for 4D reconstruction to generate multi-view supervision, as illustrated in Fig. 3. + +Depth Estimation and Scale Alignment: Given a monocular video $\mathcal{V} = \{I_t\}_{t=1}^T$ with $T$ frames, we first utilize COLMAP [65, 66] to obtain its camera intrinsics $\mathbf{K}$ , camera poses $\{\mathbf{T}_t\}_{t=1}^T$ , and a sparse point cloud $\{X_i\}_{i=1}^M$ , where each point $X_i \in \mathbb{R}^3$ denotes the 3D coordinates reconstructed in the world coordinate system. + +For depth estimation, we adopt a two-step approach that combines learning-based prediction with geometric refinement, as presented in Fig. 3 (left). We first employ an off-the-shelf monocular depth estimation model (e.g., [94]) to obtain initial depth maps $\{\hat{D}_t\}_{t=1}^T$ . Next, we align them with the metric scale using the sparse point cloud $\{X_i\}_{i=1}^M$ from COLMAP via Least Square and RANSAC [22], obtaining the metric depth maps $\{D_t\}_{t=1}^T$ . Please refer to supplementary (Sec. 6.3) for details. + +Iterative View Augmentation: Using the scale-aligned metric depth, we can warp the observed frames $\mathcal{V} =$ + +![](images/b75344985bc0af8f13a40815d54228bd41367b9af4ca4acfec44d01117c2d780.jpg) +Figure 4. Difference between Direct Warping and Iterative Warping. $S$ and $\mathcal{T}$ denote spatial and temporal dimensions, with $\mathcal{V}$ being the input monocular video (yellow dots indicating camera poses in 4D space). Left: With $N = 1$ , we directly warp the input video to new perspectives (green dots) in a single iteration. Warped frames are ranked by warping distance per timestamp and organized into a video (purple underline). Right: With $N = 2$ , the video is first warped the closest pre-defined poses (green dots, $j = 1$ ), then additional perspectives (red dots, $j = 2$ ) are generated by selecting frames with minimal warping angles from existing frames. This iterative approach minimizes distortion and floaters caused by depth inaccuracies in large-angle warping, enhancing reconstruction quality. + +$\{I_t\}_{t=1}^T$ to unseen viewpoints for synthesizing multi-view observations, where each warped frame $I_t' \in \mathbb{R}^{H \times W \times 3}$ is accompanied by a binary mask $\mathcal{M}_t' \in \{0,1\}^{H \times W}$ . However, due to inaccuracies in the depth priors (e.g., bleeding edges), directly warping the videos by a large angle introduces noticeable artifacts, which degrade the performance of 4D reconstruction. To address this, we propose an iterative view augmentation strategy, as illustrated in Fig. 3 (middle), leveraging the fact that artifacts like bleeding edges are less severe when the warping angle is small. Fig. 4 shows the difference between direct warping and iterative warping. + +Specifically, we first pre-compute a set of target unseen camera viewpoints and perform iterative view augmentation over $N$ iterations. As presented in Fig. 4, when $N = 1$ (left), we directly warp the original video to all target frames in a single step. In contrast, when $N > 1$ (right), we progressively expand the warped angles over multiple iterations. We start with a data buffer $\mathcal{D}_0$ containing the input monocular video $\mathcal{V}^0$ , its aligned depths $D^0$ and camera poses $\mathbf{T}^0$ , formally expressed as $\mathcal{D}_0 = (\mathcal{V}^0, D^0, \mathbf{T}^0)$ . At each iteration $j$ ( $1 \leq j \leq N$ ), we select frames with the smallest warping angles from the data buffer $\mathcal{D}_{j-1}$ at each timestamp $t$ and warp them to target poses $\mathbf{T}^j$ , generating multiple new videos of $T$ frames. The warped videos and corresponding masks, together with the anchor video, are then fed into our diffusion model to synthesize completed videos $\hat{\mathcal{V}}^j$ . Next, we estimate the metric depth $D^j$ using our scale alignment strategy and update the data buffer as follows: $\mathcal{D}_j = \mathcal{D}_{j-1} \cup (\hat{\mathcal{V}}^j, D^j, \mathbf{T}^j)$ . This process is repeated until all pre-defined target poses have been covered by each frame. Furthermore, to avoid the inconsistency in 4D reconstruction supervisions caused by repeated inpainting across different iterations, we introduce a supervision + +mask $S_{t}^{j} \in \{0,1\}^{H \times W}$ for each augmented frame, in which pixel value 1 indicates valid for supervision. Please refer to supplementary (Sec. 6.5) for more details. + +# 3.3. 4D Reconstruction + +4D Scene Representation: Our method can effectively produce accurate, consistent, and high-fidelity novel view videos from monocular input, offering rich multi-view supervision for 4D scene reconstruction. It is compatible with various 4D reconstruction approaches, including those extending static representations with temporal dimensions [96] and those modeling dynamic scenes via deformation or motion fields [78, 83]. In our experiments, we adopt a motion field-based 3DGS representation [78] to capture both geometric structure and temporal dynamics effectively. + +Loss Function: We rely solely on the multi-view constraints from our synthesized novel views to guide reconstruction. To mitigate distortions from depth estimation errors and video diffusion artifacts, we propose to use IV RGB loss, a robust pixel-wise loss inspired by [18], as the supervisory signal for our augmented views. Specifically, we compute the L1 loss between each pixel in the rendered image and the closest $3 \times 3$ pixels in the corresponding augmented supervision image, backpropagating only through the pixel with the minimum error. This loss, denoted as $\mathcal{L}_{\mathrm{IV}}$ , improves reconstruction robustness by reducing sensitivity to slight misalignments. For the original frames, we use standard L1, SSIM and LPIPS loss to supervise the entire image. For our valid augmented frames, we use the loss function as follows: + +$$ +\mathcal {L} = \sum_ {j = 1} ^ {N} \sum_ {t = 1} ^ {T} \left(\lambda_ {r} \mathcal {L} _ {\mathrm {I V}} ^ {t, j} + \lambda_ {s} \mathcal {L} _ {\mathrm {s s i m}} ^ {t, j} + \lambda_ {l} \mathcal {L} _ {\mathrm {l p i p s}} ^ {t, j}\right) +$$ + +where each loss term is computed only in regions where the supervision mask $S_{t}^{j}$ equals 1. Specifically, for time step $t$ in iteration $j$ : + +$$ +\mathcal {L} _ {\mathrm {I V}} ^ {t, j} = \sum_ {p} \mathcal {S} _ {t} ^ {j} (p) \cdot \mathrm {I V} (\overline {{I}} _ {t} ^ {j} (p), \hat {I} _ {t} ^ {j} (p)) +$$ + +$$ +\mathcal {L} _ {\mathrm {s s i m}} ^ {t, j} = \operatorname {S S I M} \left(\bar {I} _ {t} ^ {j} \odot \mathcal {S} _ {t} ^ {j}, \hat {I} _ {t} ^ {j} \odot \mathcal {S} _ {t} ^ {j}\right) +$$ + +$$ +\mathrm {L} _ {\text {l p i p s}} ^ {t, j} = \operatorname {L P I P S} \left(\bar {I} _ {t} ^ {j} \odot S _ {t} ^ {j}, \hat {I} _ {t} ^ {j} \odot S _ {t} ^ {j}\right) +$$ + +Here, $p$ denotes pixel locations, $\overline{I}_t^j$ and $\hat{I}_t^j$ are rendered and our augmented supervisions respectively, and $\odot$ indicates element-wise multiplication. The hyperparameters $\lambda_r$ , $\lambda_s$ and $\lambda_l$ balance the contribution of each loss term. + +# 4. Experiments + +In this section, we begin by comparing our method with SOTA 4D scene reconstruction methods in Sec. 4.1. After that, we conduct ablation studies to evaluate our design choices in Sec. 4.2. We highlight the best, second-best, + +![](images/9e4294bba231010873aaaea2b0818c8c77ca3e50ae87f3a337596caa0db84047.jpg) + +![](images/a2953d334f1db939a710f5f43d87e0f30f37cba32eb9363dcd68346fcd92c643.jpg) + +![](images/63ca1592353325af5c6a813e70b783878923f5faa717b1fef256e8bcde91f05f.jpg) + +![](images/5bf6e384295b255f2d43b00a62265ab400159e5f5d7fc7d00f3a06ab91d2e87c.jpg) + +![](images/ad4973261e726fe6f89aef1cc21a2aa49e93a6bcf55bda62b1e7614672eee498.jpg) + +![](images/32c5882823c4b81246ac9327dad46fce9ae674034c840685469ebfce2b38e8c2.jpg) +4D GS + +![](images/91805859ce2dbe66084c7caecc3638d5fdddc80d97380d539b5bb7d1af1552bf.jpg) +Shape of Motion + +![](images/8d388fcf714c907829a83db96bdb67675effd0446aafd619c074a1aae72dbc1d.jpg) +CoCoCo + +![](images/5082387bc230746e81328ee354198759821ea1b3e4ca065381edc82583b015f8.jpg) +Ours +Figure 5. Qualitative comparison of dynamic scene reconstruction on iPhone dataset and HyperNeRF dataset. The black holes in 4D GS and the white areas in Shape of Motion indicate regions where the input video lacks visibility. In contrast, our method effectively fills these invisible areas within the scenes, leveraging multi-view constraints and spatiotemporal priors to enhance reconstruction quality. + +![](images/287d50a9c65100b88a92e4f14112e8eb4c6a4fb6f0c885881f3f6cc57037778b.jpg) +Ground Truth + +and third-best scores achieved on any metrics. Please refer to supplementary (Sec. 6.2) for more training details. + +# 4.1. Monocular 4D Reconstruction + +Dataset and Metrics: We evaluate dynamic scene reconstruction from monocular video using the iPhone [25] and HyperNeRF [58] datasets. Our test sets comprise 5 iPhone and 3 HyperNeRF scenes, each with synchronized static cameras for novel-view synthesis evaluation. From the original video sequences, we select continuous clips with a few dozen frames, chosen for their large perspective differences from the test view to emphasize our method's efficacy. We assess reconstruction quality using mPSNR, mSSIM, and mLPIPS [25] for dynamic foreground, static background, and the overall scene, employing a co-visible mask to specify visible areas for reconstruction. + +Comparison Baselines: We compare our method against 4D GS [96], Shape of Motion [78], CoCoCo [114], StereoCrafter [108], and ViewCrafter [101]. CoCoCo and StereoCrafter are designed for video inpainting, while ViewCrafter focuses on 3D-aware inpainting. We integrate their diffusion models into our pipeline for warped video inpainting to validate the necessity of our anchor-conditioned model. For fairness, we provide CoCoCo with the same text prompt for each scene and use the first frame of our anchor video as the image conditioning input for ViewCrafter. Additionally, we compare against the state-of-the-art monocular reconstruction method Shape of Motion [78] and the + +
ConditionPSNR ↑SSIM ↑LPIPS ↓FVD ↓
w/o anchor25.340.80530.105618.99
w anchor (ours)27.220.82230.080114.30
+ +Table 1. Quantitative ablation study on anchor video condition on our processed ${5K}$ videos from OpenVid-1M. + +general 4D reconstruction approach 4D GS [96]. + +Results: The results are shown in Fig. 5 and Tab. 2. It can be observed that our pipeline significantly improves reconstruction quality over 4D GS and Shape of Motion, which exhibit outliers or over-sharpening artifacts. In contrast, our method not only produces a smoother result but also better preserves dynamic foreground details. Additionally, it fills missing regions invisible in the input video, addressing the inner holes seen in 4D GS and Shape of Motion. Compared to CoCoCo, which also achieves high-quality reconstruction, our method offers greater overall clarity, finer details—especially in dynamic areas—and more contextually consistent inpainting, thanks to our anchor-based video inpainting model. Quantitatively, our approach clearly outperforms all baselines across all metrics, further demonstrating its effectiveness in 4D reconstruction. + +# 4.2. Ablation Study + +We ablate various components of our method on the iPhone dataset and HyperNeRF dataset for 4D reconstruction to verify design choices of our approach. + +
DatasetDynamic PartStatic PartAll
MethodmPSNR ↑mSSIM ↑mLPIPS ↓mPSNR ↑mSSIM ↑mLPIPS ↓mPSNR ↑mSSIM ↑mLPIPS ↓
iPhone Dataset
4D GS12.280.87570.524914.540.48700.559414.010.38770.5939
Shape of Motion11.870.88940.477815.420.57440.508214.560.45700.5292
CoCoCo12.200.89640.465015.850.58090.513014.990.47010.5280
StereoCrafter12.090.89890.481115.600.59890.559314.850.49450.5676
ViewCrafter12.130.89850.505315.820.59460.531514.940.48880.5772
Ours12.340.89900.451616.310.60880.472615.200.50040.4930
HyperNeRF Dataset
4D GS17.930.95560.381918.320.47370.440918.240.42000.4450
Shape of Motion17.700.96090.302719.120.56590.452018.820.51760.4589
CoCoCo18.470.96480.304819.110.55990.462719.000.51470.4692
StereoCrafter18.510.96600.334318.980.56770.517118.860.52310.5181
ViewCrafter18.530.96580.322818.980.56910.486018.910.52590.4888
Ours18.850.96660.302019.570.59100.440719.450.54460.4449
+ +Table 2. Quantitative comparison of dynamic scene reconstruction on iPhone dataset and HyperNeRF dataset. + +![](images/263a15c3e066fe6f044889a34d34b8d9a417f2fcf1a6877780fff5e6c556f316.jpg) + +![](images/be2bc50d097638687e64294a2751f0f01854683c97d5aa8a48c9d30643f6b87c.jpg) +Masked Sequence + +![](images/b815af2e789542673288cc347d91b1767f51d7f096c1cd0a64ed989710f7fc47.jpg) + +![](images/16a78a9902e623d9aac8ba3e7edc096929923db8ecf2dd8e17efab90f4e4620a.jpg) +w/o anchor + +![](images/542d27b19567d47ac08c266287180403e9895b0ea271c4da31eb26f2f101c72c.jpg) + +![](images/de9b6d0ad621e74bc8360f1b94cdfa4eb2c2aee1e87137d5935ae8e92bf1b4a6.jpg) +w anchor (ours) +Figure 6. Quantitative ablation study on anchor video condition on our processed $5K$ videos from OpenVid-1M. + +Effect of Anchor condition: We explored the impact of anchor video condition on the video inpaint results of our diffusion model, as presented in Tab. 1 and Fig. 6. Without the anchor condition, the model's filling capability weakens, resulting in noticeable artifacts. In contrast, after incorporating the anchor video, the filled areas become more consistent with the surrounding content, significantly reducing artifacts. This demonstrates that our model effectively leverages spatiotemporal priors from the anchor video. + +Effect of View Warping and Video inpainting: To assess the effectiveness of our view augmentation, which includes view warping and video inpainting, and to compare it with direct depth supervision loss $\mathcal{L}_D$ , we compare our full method against: (a) no view augmentation or depth supervision, (b) direct depth supervision without view augmentation, and (c) iterative view warping without video inpainting. For fairness, all experiments use the same Gaussian initialization, and our view augmentation is set to 6 itera + +![](images/8add57bac569c8d15592b926de417615c493c62a360ef9b14a21368e00d1cc15.jpg) + +![](images/588bd47be74572c9314d2133ede3c1919bbf9cf77c3645f62fbf2b8021d9396a.jpg) +(a) +Figure 7. Qualitative ablation study on view warping and video inpainting on HyperNeRF dataset. (a) w/o warping, w/o inpainting, w/o depth; (b) w/o warping, w/o inpainting, w/ depth; (c) w/ warping, w/o inpainting; (d) w/ warping, w/ inpainting (ours). + +![](images/942855a598bff0fb70e482696b1a13a05ef44834e5dc0f0b5ee30b19eddf03f0.jpg) + +![](images/8d87baed96d1b6cec9003b573228ae93e075812f521af199d2870ca8681f4637.jpg) +(b) + +![](images/4b73953d511671368773e97bc2641e15d35d6ceb9aad47806f2b52a75070b0dd.jpg) + +![](images/1e549a9b3d79e7d89f7096d2b2133ca5246e8e910735b907e1f2dd6f299d147b.jpg) +(c) + +![](images/dc831f8cc6e191b70b38d9009b676bd0e271755fdf76ac3fc004170df320085d.jpg) + +![](images/ffab011f82806bc740816c4bbf60e554bc1c1952438e362d9ce00652ced69304.jpg) +(d) + +tions. The results are presented in Tab. 3 and Fig. 7. + +It can be observed that directly leveraging the depth for supervision (b) only leads to minor improvement compared with (a). In contrast, using the same depth to synthesize augmented training views leads to larger improvement even without inpainting (c). This verifies that directly adopting geometric priors for supervision may not linearly correlate with the reconstruction performance in the RGB space. Moreover, with our inpainting model to complete the warped views (d), reconstruction further improves, correctly recovering previously invisible areas and reducing flying artifacts. This demonstrates that leveraging depth priors via warping benefits reconstruction, while integrating diffusion priors further enhances completion and quality in 4D reconstruction. + +Effect of Iterations: We investigated the impact of iter + +
MethodmPSNR ↑mSSIM ↑mLPIPS ↓
(a) w/o warp, w/o inpaint, w/o LD16.040.46170.5249
(b) w/o warp, w/o inpaint, w/ LD16.120.46990.5084
(c) w/ warp, w/o inpaint, w/o LD16.440.49130.4876
(d) w/ warp, w/ inpaint, w/o LD (ours)16.800.51700.4750
+ +Table 3. Quantitative ablation study on view warping and video inpainting on iPhone dataset and HyperNeRF dataset. + +
IterationsmPSNR ↑mSSIM ↑mLPIPS ↓
w/o view aug., w/o depth16.040.46170.5249
N=1, j=115.800.47110.4959
N=6, j=216.410.50540.4964
N=6, j=416.560.51040.4868
N=6, j=6 (ours)16.800.51700.4750
+ +![](images/e85601febb725b0b5702ebb03574797ef1e14ede8fae3bd828fa1b57a52dbdad.jpg) +Figure 8. Qualitative ablation study on iterations on HyperN-erRF dataset. + +ation count on reconstruction results, as shown in Tab. 4 and Fig. 8. When no iterations are performed (w/o view aug), meaning only a single monocular video is used for supervision, the rendered image exhibits holes and noticeable artifacts at object edges. After applying view augmentation, we obtain more comprehensive observations of the scene. However, directly warping the initial video into multiple views at once $(N = 1, j = 1)$ results in poor reconstruction due to large-angle warping, which introduces severe artifacts when transforming to distant viewpoints (e.g., bleeding artifacts are more noticeable at larger viewing angles). In contrast, our iterative strategy gradually enhances viewpoints, reducing warping angles and minimizing the impact of depth inaccuracies. Even with a small number of iterations (e.g., 2 or 4), the scene is better filled, and multiview supervision helps mitigate floaters. As iterations increase, more comprehensive views of the scene are obtained for each timestamp, leading to more robust reconstruction. Although the overall scene coverage remains the same for $N = 1$ , $j = 1$ and $N = 6$ , $j = 6$ , additional iterations reduce warping distortion and improve reconstruction fidelity. Quantitative metrics further confirm that increasing the number of iterations leads to progressively better reconstruction results. + +Effect of Depth Estimation Methods: Next, we evaluate various depth estimation methods [8, 13, 31, 68, 94] to assess their impact, as shown in Tab. 5. We observe that dif + +Table 4. Quantitative ablation study on iterations on iPhone dataset and HyperNeRF dataset. + +
ModelmPSNR ↑mSSIM ↑mLPIPS ↓
DepthCrafter16.760.51350.4835
ChronoDepth16.780.51330.4874
Depth Pro16.790.51210.4807
Video Depth Anything16.790.51590.4875
Depth Anything V2 (ours)16.800.51700.4750
+ +![](images/eedf994d79412f4c28c5d63bebccb071984aa8f7d85ad08145e33e1f5523e1ff.jpg) +Figure 9. Qualitative ablation on loss function on HyperNeRF dataset. + +Table 5. Quantitative ablation study on various depth estimation models on iPhone dataset. + +
LossmPSNR ↑mSSIM ↑mLPIPS ↓
L116.500.51180.5097
IV (ours)16.800.51700.4750
+ +Table 6. Quantitative ablation study on loss function on iPhone dataset and HyperNeRF dataset. + +ferent methods yield similar results, indicating our method's robustness against monocular depth estimation variations. This may be attributed to aligning monocular depth predictions with COLMAP's scale, along with our interactive view warping and robust RGB loss. By default, we use Depth Anything V2 [94], as it achieves slightly better performance than the others. + +Effect of Loss Function: Lastly, we also evaluated the effectiveness of IV RGB loss on the robustness of the results. Instead of using $\mathcal{L}_{\mathrm{IV}}$ for RGB supervision, we directly applied the L1 loss. The results, shown in Tab. 6 and Fig. 9 (best viewed zoomed in), reveal that L1 loss leads to blurrier renderings due to misaligned supervision. In contrast, our method backpropagates errors only to the pixel with the minimum difference within adjacent patches, effectively reducing blurring caused by distortion. + +# 5. Conclusion + +We introduced Vivid4D, a novel approach for 4D monocular reconstruction that combines geometric priors with video diffusion models. By reformulating view augmentation as a video inpainting task, our method leverages unposed web videos for training and employs an iterative view expansion strategy with a robust loss to minimize artifacts from depth errors. Experiments demonstrate that Vivid4D consistently outperforms existing methods across dynamic scenes, producing more coherent and detailed novel view synthesis results while effectively filling unobserved regions. Our approach bridges geometric-based methods and generative models, enabling high-quality 4D reconstruction from casual monocular videos, with applications in virtual reality and content creation. Future work will focus on enabling plausible scene extrapolation beyond observed regions, and improving robustness to complex camera motions and scene dynamics. + +# Acknowledgements + +This work is supported by NSFC under grant 62441223, 62202418 and U21B2004. + +# References + +[1] Jianhong Bai, Menghan Xia, Xintao Wang, Ziyang Yuan, Xiao Fu, Zuozhu Liu, Haoji Hu, Pengfei Wan, and Di Zhang. Syncammaster: Synchronizing multi-camera video generation from diverse viewpoints. arXiv preprint arXiv:2412.07760, 2024. 3 +[2] Jianhong Bai, Menghan Xia, Xiao Fu, Xintao Wang, Lianrui Mu, Jinwen Cao, Zuozhu Liu, Haoji Hu, Xiang Bai, Pengfei Wan, et al. Recammaster: Camera-controlled generative rendering from a single video. arXiv preprint arXiv:2503.11647, 2025. 3 +[3] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5855–5864, 2021. 2 +[4] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5470–5479, 2022. +[5] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Zip-nerf: Anti-aliased grid-based neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19697-19705, 2023. 2 +[6] Weikang Bian, Zhaoyang Huang, Xiaoyu Shi, Yijin Li, FuYun Wang, and Hongsheng Li. Gs-dit: Advancing video generation with pseudo 4d gaussian fields through efficient dense 3d point tracking. arXiv preprint arXiv:2501.02690, 2025.3 +[7] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 2, 3, 1 +[8] Aleksei Bochkovskii, Amaël Delaunoy, Hugo Germain, Marcel Santos, Yichao Zhou, Stephan R Richter, and Vladlen Koltun. Depth pro: Sharp monocular metric depth in less than a second. arXiv preprint arXiv:2410.02073, 2024. 8 +[9] Ang Cao and Justin Johnson. Hexplane: A fast representation for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 130-141, 2023. 1, 2 +[10] David Charatan, Sizhe Lester Li, Andrea Tagliasacchi, and Vincent Sitzmann. pixelsplat: 3d gaussian splats from image pairs for scalable generalizable 3d reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vi + +sion and Pattern Recognition, pages 19457-19467, 2024. 2 +[11] Qihua Chen, Yue Ma, Hongfa Wang, Junkun Yuan, Wenzhe Zhao, Qi Tian, Hongmei Wang, Shaobo Min, Qifeng Chen, and Wei Liu. Follow-your-canvas: Higher-resolution video outpainting with extensive content generation. arXiv preprint arXiv:2409.01055, 2024.3 +[12] Shen Chen, Jiale Zhou, and Lei Li. Optimizing 3d gaussian splatting for sparse viewpoint scene reconstruction. arXiv preprint arXiv:2409.03213, 2024. 2 +[13] Sili Chen, Hengkai Guo, Shengnan Zhu, Feihu Zhang, Zilong Huang, Jiashi Feng, and Bingyi Kang. Video depth anything: Consistent depth estimation for super-long videos. arXiv:2501.12375, 2025. 8 +[14] Yuedong Chen, Haofei Xu, Chuanxia Zheng, Bohan Zhuang, Marc Pollefeys, Andreas Geiger, Tat-Jen Cham, and Jianfei Cai. Mvsplat: Efficient 3d gaussian splatt-ting from sparse multi-view images. arXiv preprint arXiv:2403.14627, 2024. 2 +[15] Wen-Hsuan Chu, Lei Ke, and Katerina Fragkiadaki. Dreamscene4d: Dynamic multi-object scene generation from monocular videos, 2024. 3 +[16] Jaeyoung Chung, Suyoung Lee, Hyeongjin Nam, Jaerin Lee, and Kyoung Mu Lee. Luciddreamer: Domain-free generation of 3d gaussian splatting scenes. arXiv preprint arXiv:2311.13384, 2023. 2 +[17] Yuanxing Duan, Fangyin Wei, Qiyu Dai, Yuhang He, Wenzheng Chen, and Baoquan Chen. 4d gaussian splatting: Towards efficient novel view synthesis for dynamic scenes. arXiv preprint arXiv:2402.03307, 2024. 1, 2 +[18] Adrian Dziembowski, Dawid Mieloch, Jakub Stankowski, and Adam Grzelka. Iv-psnr—the objective quality metric for immersive video applications. IEEE Transactions on Circuits and Systems for Video Technology, 32(11):7575-7591, 2022. 5 +[19] Fanda Fan, Chaoxu Guo, Litong Gong, Biao Wang, Tiezheng Ge, Yuning Jiang, Chunjie Luo, and Jianfeng Zhan. Hierarchical masked 3d diffusion model for video outpainting. In Proceedings of the 31st ACM International Conference on Multimedia, pages 7890-7900, 2023. 3 +[20] Zhiwen Fan, Wenyan Cong, Kairun Wen, Kevin Wang, Jian Zhang, Xinghao Ding, Danfei Xu, Boris Ivanovic, Marco Pavone, Georgios Pavlakos, et al. Instantsplat: Unbounded sparse-view pose-free gaussian splatting in 40 seconds. arXiv preprint arXiv:2403.20309, 2024. 2 +[21] Jiemin Fang, Taoran Yi, Xinggang Wang, Lingxi Xie, Xiaopeng Zhang, Wenyu Liu, Matthias Nießner, and Qi Tian. Fast dynamic radiance fields with time-aware neural voxels. In SIGGRAPH Asia 2022 Conference Papers, pages 1-9, 2022. 1, 2 +[22] Martin A Fischler and Robert C Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM, 24(6):381-395, 1981. 4, 1 +[23] Sara Fridovich-Keil, Giacomo Meanti, Frederik Rahbæk Warburg, Benjamin Recht, and Angjoo Kanazawa. K-planes: Explicit radiance fields in space, time, and appear + +ance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12479-12488, 2023. 1, 2 +[24] Chen Gao, Ayush Saraf, Johannes Kopf, and Jia-Bin Huang. Dynamic view synthesis from dynamic monocular video. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5712-5721, 2021. 1 +[25] Hang Gao, Ruilong Li, Shubham Tulsiani, Bryan Russell, and Angjoo Kanazawa. Monocular dynamic view synthesis: A reality check, 2022. 6, 3 +[26] Ruiqi Gao, Aleksander Holynski, Philipp Henzler, Arthur Brussee, Ricardo Martin-Brualla, Pratul Srinivasan, Jonathan T Barron, and Ben Poole. Cat3d: Create anything in 3d with multi-view diffusion models. arXiv preprint arXiv:2405.10314, 2024. 2 +[27] Bohai Gu, Hao Luo, Song Guo, and Peiran Dong. Advanced video inpainting using optical flow-guided efficient diffusion. arXiv preprint arXiv:2412.00857, 2024. 3 +[28] Antoine Guédon and Vincent Lepetit. Sugar: Surface-aligned gaussian splattering for efficient 3d mesh reconstruction and high-quality mesh rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5354-5363, 2024. 2 +[29] Hao He, Ceyuan Yang, Shanchuan Lin, Yinghao Xu, Meng Wei, Liangke Gui, Qi Zhao, Gordon Wetzstein, Lu Jiang, and Hongsheng Li. Cameractr ii: Dynamic scene exploration via camera-controlled video diffusion models. arXiv preprint arXiv:2503.10592, 2025. 3 +[30] Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. Cogvideo: Large-scale pretraining for text-to-video generation via transformers. arXiv preprint arXiv:2205.15868, 2022. 3 +[31] Wenbo Hu, Xiangjun Gao, Xiaoyu Li, Sijie Zhao, Xiaodong Cun, Yong Zhang, Long Quan, and Ying Shan. Depthcrafter: Generating consistent long depth sequences for open-world videos. arXiv preprint arXiv:2409.02095, 2024.8 +[32] Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2d gaussian splatting for geometrically accurate radiance fields. In ACM SIGGRAPH 2024 conference papers, pages 1-11, 2024. 2 +[33] Nikita Karaev, Ignacio Rocco, Benjamin Graham, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. Co-tracker: It is better to track together. In European Conference on Computer Vision, pages 18-35. Springer, 2024. 3 +[34] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139-1, 2023. 2 +[35] Hamed Razavi Khosroshahi, Jaime Sancho, Gun Bang, Gauthier Lafruit, Eduardo Juarez, and Mehrdad Teratani. Da4nerf: Depth-aware augmentation technique for neural radiance fields. Journal of Visual Communication and Image Representation, 107:104365, 2025. 2 +[36] Weijie Kong, Qi Tian, Zijian Zhang, Rox Min, Zuozhuo Dai, Jin Zhou, Jiangfeng Xiong, Xin Li, Bo Wu, Jian- + +wei Zhang, et al. Hunyuanvideo: A systematic framework for large video generative models. arXiv preprint arXiv:2412.03603, 2024. 2 +[37] Minhyeok Lee, Suhwan Cho, Chajin Shin, Jungho Lee, Sunghun Yang, and Sangyoun Lee. Video diffusion models are strong video inpainter. arXiv preprint arXiv:2408.11402, 2024.3 +[38] Jiahui Lei, Yijia Weng, Adam Harley, Leonidas Guibas, and Kostas Daniilidis. Mosca: Dynamic gaussian fusion from casual videos via 4d motion scaffolds. arXiv preprint arXiv:2405.17421, 2024. 1, 3 +[39] Jiahe Li, Jiawei Zhang, Xiao Bai, Jin Zheng, Xin Ning, Jun Zhou, and Lin Gu. Dngaussian: Optimizing sparse-view 3d gaussian radiance fields with global-local depth normalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20775-20785, 2024. 2 +[40] Xiaowen Li, Haolan Xue, Peiran Ren, and Liefeng Bo. Diffueraser: A diffusion model for video inpainting. arXiv preprint arXiv:2501.10018, 2025.3 +[41] Zhengqi Li, Simon Niklaus, Noah Snavely, and Oliver Wang. Neural scene flow fields for space-time view synthesis of dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6498-6508, 2021. 1 +[42] Zhengqi Li, Qianqian Wang, Forrester Cole, Richard Tucker, and Noah Snavely. Dynibar: Neural dynamic image-based rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4273-4284, 2023. 1, 3 +[43] Zhengqi Li, Richard Tucker, Forrester Cole, Qianqian Wang, Linyi Jin, Vickie Ye, Angjoo Kanazawa, Aleksander Holynski, and Noah Snavelly. Megasam: Accurate, fast, and robust structure and motion from casual dynamic videos. arXiv preprint arXiv:2412.04463, 2024. 6 +[44] Youtian Lin, Zuozhuo Dai, Siyu Zhu, and Yao Yao. Gaussian-flow: 4d reconstruction with dynamic 3d gaussian particle. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21136-21145, 2024. 1, 2 +[45] Fangfu Liu, Wenqiang Sun, Hanyang Wang, Yikai Wang, Haowen Sun, Junliang Ye, Jun Zhang, and Yueqi Duan. Reconx: Reconstruct any scene from sparse views with video diffusion model. arXiv preprint arXiv:2408.16767, 2024. 2 +[46] Kunhao Liu, Ling Shao, and Shijian Lu. Novel view extrapolation with video diffusion priors. arXiv preprint arXiv:2411.14208, 2024. 2 +[47] Qingming Liu, Yuan Liu, Jiepeng Wang, Xianqiang Lv, Peng Wang, Wenping Wang, and Junhui Hou. Modgs: Dynamic gaussian splatting from causally-captured monocular videos. arXiv preprint arXiv:2406.00434, 2024. 1, 2, 3 +[48] Yixin Liu, Kai Zhang, Yuan Li, Zhiling Yan, Chujie Gao, Ruoxi Chen, Zhengqing Yuan, Yue Huang, Hanchi Sun, Jianfeng Gao, et al. Sora: A review on background, technology, limitations, and opportunities of large vision models. arXiv preprint arXiv:2402.17177, 2024. 2 + +[49] Tao Lu, Mulin Yu, Linning Xu, Yuanbo Xiangli, Limin Wang, Dahua Lin, and Bo Dai. Scaffold-gs: Structured 3d gaussians for view-adaptive rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20654-20664, 2024. 2 +[50] Jonathon Luiten, Georgios Kopanas, Bastian Leibe, and Deva Ramanan. Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. arXiv preprint arXiv:2308.09713, 2023. 1, 2 +[51] Zhengxiong Luo, Dayou Chen, Yingya Zhang, Yan Huang, Liang Wang, Yujun Shen, Deli Zhao, Jingren Zhou, and Tieniu Tan. Videofusion: Decomposed diffusion models for high-quality video generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 4, 1 +[52] Baorui Ma, Huachen Gao, Haoge Deng, Zhengxiong Luo, Tiejun Huang, Lulu Tang, and Xinlong Wang. You see it, you got it: Learning 3d creation on pose-free videos at scale. arXiv preprint arXiv:2412.06699, 2024. 2, 3 +[53] Sheng Miao, Jiaxin Huang, Dongfeng Bai, Weichao Qiu, Bingbing Liu, Andreas Geiger, and Yiyi Liao. Efficient depth-guided urban view synthesis. arXiv preprint arXiv:2407.12395, 2024. 2 +[54] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2 +[55] Kepan Nan, Rui Xie, Penghao Zhou, Tiehan Fan, Zhenheng Yang, Zhijie Chen, Xiang Li, Jian Yang, and Ying Tai. Openvid-1m: A large-scale high-quality dataset for text-to-video generation. arXiv preprint arXiv:2407.02371, 2024. 1, 3 +[56] Michael Niemeyer, Jonathan T Barron, Ben Mildenhall, Mehdi SM Sajjadi, Andreas Geiger, and Noha Radwan. Regnerf: Regularizing neural radiance fields for view synthesis from sparse inputs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5480-5490, 2022. 2 +[57] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5865-5874, 2021. 1 +[58] Keunhong Park, Utkarsh Sinha, Peter Hedman, Jonathan T. Barron, Sofien Bouaziz, Dan B Goldman, Ricardo MartinBrualla, and Steven M. Seitz. Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields, 2021. 1, 3, 6 +[59] Soumava Paul, Christopher Wewer, Bernt Schiele, and Jan Eric Lenssen. Sp2360: Sparse-view 360 scene reconstruction using cascaded 2d diffusion priors. In ECCV 2024 Workshop on Wild 3D: 3D Modeling, Reconstruction, and Generation in the Wild, 2024. 2 +[60] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 2 + +[61] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: Neural radiance fields for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10318-10327, 2021. 1, 2 +[62] Jiawei Ren, Kevin Xie, Ashkan Mirzaei, Hanxue Liang, Xiaohui Zeng, Karsten Kreis, Ziwei Liu, Antonio Torralba, Sanja Fidler, Seung Wook Kim, et al. L4gm: Large 4d gaussian reconstruction model. arXiv preprint arXiv:2406.10324, 2024. 2 +[63] Xuanchi Ren, Tianchang Shen, Jiahui Huang, Huan Ling, Yifan Lu, Merlin Nimier-David, Thomas Müller, Alexander Keller, Sanja Fidler, and Jun Gao. Gen3c: 3d-informed world-consistent video generation with precise camera control. arXiv preprint arXiv:2503.03751, 2025. 3 +[64] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 3, 4 +[65] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 4, 1 +[66] Johannes Lutz Schonberger, Enliang Zheng, Marc Pollefeys, and Jan-Michael Frahm. Pixelwise view selection for unstructured multi-view stereo. In European Conference on Computer Vision (ECCV), 2016. 4, 1 +[67] Junyoung Seo, Kazumi Fukuda, Takashi Shibuya, Takuya Narihira, Naoki Murata, Shoukang Hu, Chieh-Hsin Lai, Seungryong Kim, and Yuki Mitsufuji. Genwarp: Single image to novel views with semantic-preserving generative warping. Advances in Neural Information Processing Systems, 37:80220-80243, 2025. 2 +[68] Jiahao Shao, Yuanbo Yang, Hongyu Zhou, Youmin Zhang, Yujun Shen, Vitor Guizilini, Yue Wang, Matteo Poggi, and Yiyi Liao. Learning temporally consistent video depth from video diffusion priors, 2024. 2, 8 +[69] Colton Stearns, Adam Harley, Mikaela Uy, Florian Dubost, Federico Tombari, Gordon Wetzstein, and Leonidas Guibas. Dynamic gaussian marbles for novel view synthesis of casual monocular videos. arXiv preprint arXiv:2406.18717, 2024. 1, 3 +[70] Wenqiang Sun, Shuo Chen, Fangfu Liu, Zilong Chen, Yueqi Duan, Jun Zhang, and Yikai Wang. Dimensionx: Create any 3d and 4d scenes from a single image with controllable video diffusion. arXiv preprint arXiv:2411.04928, 2024. 3 +[71] Jiaxiang Tang, Zhaoxi Chen, Xiaokang Chen, Tengfei Wang, Gang Zeng, and Ziwei Liu. Lgm: Large multiview gaussian model for high-resolution 3d content creation. In European Conference on Computer Vision, pages 1-18. Springer, 2024. 3 +[72] Thomas Unterthiner, Sjoerd van Steenkiste, Karol Kurach, Raphael Marinier, Marcin Michalski, and Sylvain Gelly. Towards accurate generative models of video: A new metric and challenges, 2019. 3 +[73] Basile Van Hoorick, Rundi Wu, Ege Ozguroglu, Kyle Sargent, Ruoshi Liu, Pavel Tokmakov, Achal Dave, Changxi + +Zheng, and Carl Vondrick. Generative camera dolly: Extreme monocular dynamic novel view synthesis. In European Conference on Computer Vision, pages 313-331. Springer, 2024. 3 +[74] Fu-Yun Wang, Xiaoshi Wu, Zhaoyang Huang, Xiaoyu Shi, Dazhong Shen, Guanglu Song, Yu Liu, and Hongsheng Li. Be-your-outpainter: Mastering video outpainting through input-specific adaptation. In European Conference on Computer Vision, pages 153-168. Springer, 2024. 3 +[75] Guangcong Wang, Zhaoxi Chen, Chen Change Loy, and Ziwei Liu. Sparsenerf: Distilling depth ranking for few-shot novel view synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9065-9076, 2023. 2 +[76] Jiuniu Wang, Hangjie Yuan, Dayou Chen, Yingya Zhang, Xiang Wang, and Shiwei Zhang. Modelscope text-to-video technical report. arXiv preprint arXiv:2308.06571, 2023. 4, 1 +[77] Peng Wang, Yuan Liu, Zhaoxi Chen, Lingjie Liu, Ziwei Liu, Taku Komura, Christian Theobalt, and Wenping Wang. F2-nerf: Fast neural radiance field training with free camera trajectories. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4150–4159, 2023. 2 +[78] Qianqian Wang, Vickie Ye, Hang Gao, Jake Austin, Zhengqi Li, and Angjoo Kanazawa. Shape of motion: 4d reconstruction from a single video. arXiv preprint arXiv:2407.13764, 2024. 1, 3, 5, 6 +[79] Qianqian Wang, Yifei Zhang, Aleksander Holynski, Alexei A. Efros, and Angjoo Kanazawa. Continuous 3d perception model with persistent state, 2025. 6 +[80] Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. Dust3r: Geometric 3d vision made easy. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20697-20709, 2024. 2 +[81] Shizun Wang, Xingyi Yang, Qiuhong Shen, Zhenxiang Jiang, and Xinchao Wang. Gflow: Recovering 4d world from monocular video. arXiv preprint arXiv:2405.18426, 2024. 1, 3 +[82] Zhou Wang, A.C. Bovik, H.R. Sheikh, and E.P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4): 600-612, 2004. 3 +[83] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 4d gaussian splatting for real-time dynamic scene rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20310-20320, 2024. 1, 2, 5 +[84] Rundi Wu, Ruiqi Gao, Ben Poole, Alex Trevithick, Changxi Zheng, Jonathan T Barron, and Aleksander Holynski. Cat4d: Create anything in 4d with multi-view video diffusion models. arXiv preprint arXiv:2411.18613, 2024. 2, 3 +[85] Rundi Wu, Ben Mildenhall, Philipp Henzler, Keunhong Park, Ruiqi Gao, Daniel Watson, Pratul P Srinivasan, Dor + +Verbin, Jonathan T Barron, Ben Poole, et al. Reconfusion: 3d reconstruction with diffusion priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21551-21561, 2024. 2 +[86] Wenqi Xian, Jia-Bin Huang, Johannes Kopf, and Changil Kim. Space-time neural irradiance fields for free-viewpoint video. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9421-9431, 2021. 2 +[87] Zeqi Xiao, Wenqi Ouyang, Yifan Zhou, Shuai Yang, Lei Yang, Jianlou Si, and Xingang Pan. Trajectory attention for fine-grained video motion control. arXiv preprint arXiv:2411.19324, 2024. 3 +[88] Haolin Xiong. Sparsegs: Real-time $360^{\circ}$ sparse view synthesis using gaussian splatting. Master's thesis, University of California, Los Angeles, 2024. 2 +[89] Yunzhi Yan, Haotong Lin, Chenxu Zhou, Weijie Wang, Haiyang Sun, Kun Zhan, Xianpeng Lang, Xiaowei Zhou, and Sida Peng. Street gaussians for modeling dynamic urban scenes. arXiv preprint arXiv:2401.01339, 2024. 2 +[90] Zhiqiang Yan, Yupeng Zheng, Chongyi Li, Jun Li, and Jian Yang. Learnable differencing center for nighttime depth perception, 2023. 2 +[91] Jinyu Yang, Mingqi Gao, Zhe Li, Shang Gao, Fangjing Wang, and Feng Zheng. Track anything: Segment anything meets videos. arXiv preprint arXiv:2304.11968, 2023. 3 +[92] Jiawei Yang, Marco Pavone, and Yue Wang. Freenerf: Improving few-shot neural rendering with free frequency regularization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8254-8263, 2023. 2 +[93] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Ji-ashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10371–10381, 2024. 2 +[94] Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything v2. Advances in Neural Information Processing Systems, 37:21875-21911, 2025. 4, 8 +[95] Shiyuan Yang, Zheng Gu, Liang Hou, Xin Tao, Pengfei Wan, Xiaodong Chen, and Jing Liao. Mtv-inpaint: Multi-task long video inpainting. arXiv preprint arXiv:2503.11412, 2025. 3 +[96] Zeyu Yang, Hongye Yang, Zijie Pan, Xiatian Zhu, and Li Zhang. Real-time photorealistic dynamic scene representation and rendering with 4d gaussian splatting. arXiv preprint arXiv:2310.10642, 2023. 1, 2, 5, 6 +[97] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al. Cogvideox: Text-to-video diffusion models with an expert transformer. arXiv preprint arXiv:2408.06072, 2024. 2, 3 +[98] Meng You, Zhiyu Zhu, Hui Liu, and Junhui Hou. Nvssolver: Video diffusion model as zero-shot novel view synthesizer. arXiv preprint arXiv:2405.15364, 2024. 3 + +[99] Hanyang Yu, Xiaoxiao Long, and Ping Tan. Lm-gaussian: Boost sparse-view 3d gaussian splatting with large model priors. arXiv preprint arXiv:2409.03456, 2024. 2 +[100] Mark YU, Wenbo Hu, Jinbo Xing, and Ying Shan. Trajectorycrafter: Redirecting camera trajectory for monocular videos via diffusion models. arXiv preprint arXiv:2503.05638, 2025. 3 +[101] Wangbo Yu, Jinbo Xing, Li Yuan, Wenbo Hu, Xiaoyu Li, Zhipeng Huang, Xiangjun Gao, Tien-Tsin Wong, Ying Shan, and Yonghong Tian. Viewcrafter: Taming video diffusion models for high-fidelity novel view synthesis. arXiv preprint arXiv:2409.02048, 2024. 2, 3, 6, 4 +[102] Zehao Yu, Anpei Chen, Binbin Huang, Torsten Sattler, and Andreas Geiger. Mip-splatting: Alias-free 3d gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19447-19456, 2024. 2 +[103] Zhongrui Yu, Haoran Wang, Jinze Yang, Hanzhang Wang, Zeke Xie, Yunfeng Cai, Jiale Cao, Zhong Ji, and Mingming Sun. Sgd: Street view synthesis with gaussian splatting and diffusion prior. arXiv preprint arXiv:2403.20079, 2024. 2 +[104] Junyi Zhang, Charles Herrmann, Junhwa Hur, Varun Jampani, Trevor Darrell, Forrester Cole, Deqing Sun, and Ming-Hsuan Yang. Monst3r: A simple approach for estimating geometry in the presence of motion. arXiv preprint arXiv:2410.03825, 2024. 6 +[105] Jiawei Zhang, Jiahe Li, Xiaohan Yu, Lei Huang, Lin Gu, Jin Zheng, and Xiao Bai. Cor-gs: Sparse-view 3d gaussian splatting via co-regularization. arXiv preprint arXiv:2405.12110, 2024. 2 +[106] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric, 2018. 3 +[107] Zhixing Zhang, Bichen Wu, Xiaoyan Wang, Yaqiao Luo, Luxin Zhang, Yinan Zhao, Peter Vajda, Dimitris Metaxas, and Licheng Yu. Avid: Any-length video inpainting with diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7162-7172, 2024. 3 +[108] Sijie Zhao, Wenbo Hu, Xiaodong Cun, Yong Zhang, Xiaoyu Li, Zhe Kong, Xiangjun Gao, Muyao Niu, and Ying Shan. Stereocrafter: Diffusion-based generation of long and high-fidelity stereoscopic 3d from monocular videos. arXiv preprint arXiv:2409.07447, 2024. 3, 6, 4 +[109] Yuyang Zhao, Chung-Ching Lin, Kevin Lin, Zhiwen Yan, Linjie Li, Zhengyuan Yang, Jianfeng Wang, Gim Hee Lee, and Lijuan Wang. Genxd: Generating any 3d and 4d scenes. arXiv preprint arXiv:2411.02319, 2024. 2, 3 +[110] Hongyu Zhou, Jiahao Shao, Lu Xu, Dongfeng Bai, Weichao Qiu, Bingbing Liu, Yue Wang, Andreas Geiger, and Yiyi Liao. Hugs: Holistic urban 3d scene understanding via gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21336-21345, 2024. 2 +[111] Shangchen Zhou, Chongyi Li, Kelvin CK Chan, and Chen Change Loy. Propainter: Improving propagation and transformer for video inpainting. In Proceedings of the + +IEEE/CVF international conference on computer vision, pages 10477-10486, 2023. 3 +[112] Zhizhuo Zhou and Shubham Tulsiani. Sparsefusion: Distilling view-conditioned diffusion for 3d reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12588-12597, 2023. 2 +[113] Hanxin Zhu, Tianyu He, Xiqian Yu, Junliang Guo, Zhibo Chen, and Jiang Bian. Ar4d: Autoregressive 4d generation from monocular videos. arXiv preprint arXiv:2501.01722, 2025.3 +[114] Bojia Zi, Shihao Zhao, Xianbiao Qi, Jianan Wang, Yukai Shi, Qianyu Chen, Bin Liang, Kam-Fai Wong, and Lei Zhang. Cococo: Improving text-guided video inpainting for better consistency, controllability and compatibility. arXiv preprint arXiv:2403.12035, 2024. 3, 6 \ No newline at end of file diff --git a/ICCV/2025/Vivid4D_ Improving 4D Reconstruction from Monocular Video by Video Inpainting/images.zip b/ICCV/2025/Vivid4D_ Improving 4D Reconstruction from Monocular Video by Video Inpainting/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..96387af0b4246e10849adcbb3c62467473f29a38 --- /dev/null +++ b/ICCV/2025/Vivid4D_ Improving 4D Reconstruction from Monocular Video by Video Inpainting/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3463e45b43d513c086fd436dab46069b908e13f50ea1157284ad4f9412e57f99 +size 820101 diff --git a/ICCV/2025/Vivid4D_ Improving 4D Reconstruction from Monocular Video by Video Inpainting/layout.json b/ICCV/2025/Vivid4D_ Improving 4D Reconstruction from Monocular Video by Video Inpainting/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..32141745efd5879615b35c48540dc53f08cda557 --- /dev/null +++ b/ICCV/2025/Vivid4D_ Improving 4D Reconstruction from Monocular Video by Video Inpainting/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f42baef9a4dc4dbf722dfe10e0a819c3e1bbf12648141aba911fa4d1bbf9ae36 +size 529578 diff --git a/ICCV/2025/VoiceCraft-Dub_ Automated Video Dubbing with Neural Codec Language Models/68799307-ec96-4f0a-9ca1-088fb28e13e1_content_list.json b/ICCV/2025/VoiceCraft-Dub_ Automated Video Dubbing with Neural Codec Language Models/68799307-ec96-4f0a-9ca1-088fb28e13e1_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..6c3e3ea254332857ac363a857975318b68813518 --- /dev/null +++ b/ICCV/2025/VoiceCraft-Dub_ Automated Video Dubbing with Neural Codec Language Models/68799307-ec96-4f0a-9ca1-088fb28e13e1_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:63d3f122fa3a78618c85edc44b78781182c27d09ba40f503feb03ecd5a3265c4 +size 82309 diff --git a/ICCV/2025/VoiceCraft-Dub_ Automated Video Dubbing with Neural Codec Language Models/68799307-ec96-4f0a-9ca1-088fb28e13e1_model.json b/ICCV/2025/VoiceCraft-Dub_ Automated Video Dubbing with Neural Codec Language Models/68799307-ec96-4f0a-9ca1-088fb28e13e1_model.json new file mode 100644 index 0000000000000000000000000000000000000000..d69500d0a6369e5962464227b068399d7cca0bd7 --- /dev/null +++ b/ICCV/2025/VoiceCraft-Dub_ Automated Video Dubbing with Neural Codec Language Models/68799307-ec96-4f0a-9ca1-088fb28e13e1_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7095089a10bc115dd4237b1ea8ef9b9fc75fa7149b3e6cb3d6da6fada19dc55d +size 98975 diff --git a/ICCV/2025/VoiceCraft-Dub_ Automated Video Dubbing with Neural Codec Language Models/68799307-ec96-4f0a-9ca1-088fb28e13e1_origin.pdf b/ICCV/2025/VoiceCraft-Dub_ Automated Video Dubbing with Neural Codec Language Models/68799307-ec96-4f0a-9ca1-088fb28e13e1_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..3117fde41741c795293f81f2a124b5ee34382a1e --- /dev/null +++ b/ICCV/2025/VoiceCraft-Dub_ Automated Video Dubbing with Neural Codec Language Models/68799307-ec96-4f0a-9ca1-088fb28e13e1_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cb28e7085ef18fc60931e15f97e3390c9b73a5163f4d0efc07a01fab76864f38 +size 2442339 diff --git a/ICCV/2025/VoiceCraft-Dub_ Automated Video Dubbing with Neural Codec Language Models/full.md b/ICCV/2025/VoiceCraft-Dub_ Automated Video Dubbing with Neural Codec Language Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..72f4d85289455d94d12945955319f44de56b30aa --- /dev/null +++ b/ICCV/2025/VoiceCraft-Dub_ Automated Video Dubbing with Neural Codec Language Models/full.md @@ -0,0 +1,282 @@ +# VoiceCraft-Dub: Automated Video Dubbing with Neural Codec Language Models + +Kim Sung-Bin $^{1}$ Jeongsoo Choi $^{2}$ Puyuan Peng $^{3}$ Joon Son Chung $^{2}$ Tae-Hyun Oh $^{4}$ David Harwath $^{3}$ \ + $^{1}$ Department of Electrical Engineering, POSTECH + +$^{2}$ School of Electrical Engineering and $^{4}$ School of Computing, KAIST + +$^{3}$ Department of Computer Science, The University of Texas at Austin + +![](images/fe30fdeb063e5c064e456ed969d3c8aa2b1cac69375b3204cb3e0a0e76f4b4c2.jpg) +(a) Text-to-Speech vs. Automated video dubbing +Figure 1. Automated video dubbing. (a) Unlike text-to-speech, which generates diverse speech based on target text, automated video dubbing requires synthesized speech to be temporally and expressively aligned with the video while maintaining naturalness and intelligibility. (b) Examples of synthesized speech from VoiceCraft-Dub show that each speech is aligned with the lip movements of the input video. We strongly encourage listening to each of the synthesized samples in https://voicecraft-dub.github.io/. + +![](images/bd2b006cab90464c1d111f55e2f4c574b70203eda6501fb7bb11865e497d621b.jpg) +Synthesized speech + +![](images/5b28093645479c9d7564aaa3c9ccd939018cf937b7e5d27f28683ad6300da2e0.jpg) +(b) Examples of the synthesized speech with temporal alignment to the video + +# Abstract + +We present VoiceCraft-Dub, a novel approach for automated video dubbing that synthesizes high-quality speech from text and facial cues. This task has broad applications in filmmaking, multimedia creation, and assisting voice-impaired individuals. Building on the success of Neural Codec Language Models (NCLMs) for speech synthesis, our method extends their capabilities by incorporating video features, ensuring that synthesized speech is time-synchronized and expressively aligned with facial movements while preserving natural prosody. To inject visual cues, we design adapters to align facial features with the NCLM token space and introduce audio-visual fusion layers to merge audio-visual information within the NCLM framework. Additionally, we curate CelebV-Dub, a new dataset of expressive, real-world videos specifically designed for automated video dubbing. Extensive experiments show that our model achieves high-quality, intelligible, and natural speech synthesis with accurate lip synchronization, outperforming existing methods in human perception and performing favorably in objective evaluations. We also adapt VoiceCraft-Dub for the video-to-speech task, demonstrating its versatility for various applications. + +# 1. Introduction + +Recent advancements in Text-to-Speech (TTS) have significantly improved the ability to synthesize human-like speech with remarkable naturalness, content accuracy, and voice similarity. Specifically, Neural Codec Language Models (NCLMs) represent a new approach for speech generation that leverages language modeling with discrete codes derived from neural speech CODECs [14, 43]. This method has demonstrated superior performance across diverse speech generation tasks, such as TTS [7, 21, 38], speech editing [30], and voice conversion [3], due to the strong in-context learning ability of language models. Here, context refers to the speaker characteristics in the reference speech, enabling the generated speech to exhibit higher speaker similarity and naturalness compared to conventional models. + +While TTS traditionally synthesizes speech from input text, combining text with video input gives rise to a different application: automated video dubbing [19]. Automated video dubbing involves generating target speech given the source speech, target text, and target video, as shown in Fig. 1 (a). The source speech provides reference voice characteristics, while the target text and video determine the content to be synthesized, its temporal alignment to the speaker's + +facial and lip movements, and even provide prosodic and emotional cues for how the speech should be delivered. This task has applications in filmmaking, multimedia content creation, dubbing translations of films into different languages, redubbing silent movies, and voice generation for voice-impaired individuals [20, 29]. Automated video dubbing is challenging as it must meet criteria inherited from TTS: (1) Intelligibility: the speech should convey accurate content, (2) Naturalness: the speech should exhibit natural prosody and intonation, and (3) Speaker similarity: the voice should match the source speaker. Additionally, video-specific criteria must also be considered: (4) Lip synchronization: the synthesized speech should be precisely time-aligned with lip movements, and (5) Expressiveness: the speech should reflect natural expressions corresponding to facial cues. + +In this work, we introduce VoiceCraft-Dub, a novel approach for automated video dubbing using NCLMs. Building on the strengths of NCLMs in TTS, we extend their capabilities to incorporate video features, enabling the synthesis of natural and expressive speech that is synchronized with the target video. Specifically, we frame the task as an autoregressive speech token prediction problem, conditioned on the source speech, target text, and target video. Since NCLMs typically accept only text and speech tokens as input, we introduce an audio-visual fusion approach to inform the model of the desired alignment between output speech and facial features—such as lip movements and expressions—by merging speech and visual tokens at each timestep. These merged audio-visual tokens are autoregressively fed into a Transformer decoder to generate the next speech token, allowing the model to synthesize time-aligned and expressive speech by referencing the target video, as shown in Fig. 1 (b). + +We evaluate our proposed method using the existing audio-visual dataset LRS3 [1], which is effective for assessing model performance in real-world scenarios. To further test our model's ability to generate expressive speech in diverse, in-the-wild settings, we curate a video dataset, CelebV-Dub, built upon existing video sources [42, 45] that feature emotionally rich scenarios from vlogs, dramas, and movies. To ensure data quality, we design a curation pipeline that filters for content suitable for automated video dubbing. Compared to LRS3, CelebV-Dub includes a broader range of in-the-wild videos with expressive speech. + +Given the generative nature of this task, subjective evaluation is essential for assessing the quality of synthesized speech in terms of naturalness and lip synchronization. We conduct extensive human evaluations in which participants assess synthesized outputs from different approaches across various aspects, including naturalness, intelligibility, lip-sync accuracy, speaker similarity, and expressiveness. Additionally, we employ several objective metrics for quantitative evaluation. Experimental results show that our method achieves highly accurate lip synchronization while signifi + +icantly outperforming state-of-the-art approaches [10, 11] in naturalness and content accuracy across all metrics. Furthermore, we demonstrate the versatility of our approach by adapting it for the video-to-speech generation task using an off-the-shelf visual speech recognition model [26]. Although this is not our primary focus, our model performs favorably, emphasizing its potential for a wide range of applications. Our main contributions are summarized as follows: + +- We introduce VoiceCraft-Dub, the first work to extend Neural Codec Language Models (NCLMs) for automated video dubbing by integrating visual facial cues. +- We propose a novel audio-visual fusion approach in NCLMs to achieve precise lip synchronization while preserving speech quality and intelligibility. +- We curate CelebV-Dub, a dataset containing expressive real-world videos designed for automated video dubbing. +- We demonstrate superior performance in synthesizing human-like speech with high content accuracy, naturalness, and precise synchronization with video. + +# 2. Related work + +Neural codec language models (NCLMs). NCLMs are inspired by advancements in both neural audio CODEs and language modeling techniques. Neural audio CODEs [14, 43] have enabled high-fidelity audio compression by efficiently encoding raw waveforms into discrete tokens while preserving quality through residual vector quantization (RVQ) method. By applying language modeling techniques to these discrete audio tokens, NCLMs have been shown to generate high-quality speech. Moreover, NCLMs exhibit strong in-context learning capabilities, allowing them to replicate the vocal characteristics, emotion, and prosody of a prompt utterance, resulting in highly natural and expressive speech. In this regard, NCLMs significantly outperform conventional TTS methods. NCLMs have demonstrated strong performance in diverse applications, such as speech continuation [5], zero-shot TTS [7, 21, 38], speech editing [30], and voice conversion [3]. Several methods have also explored style-controlled synthesis [23, 41]. Furthermore, NCLMs have been successfully applied to other audio domains, such as music [2, 13, 15] and sound effects [22]. We further extend the capabilities of NCLMs by integrating them with visual information, specifically targeting the automated video dubbing task. By bridging the audio-visual domain, we significantly expand the potential and versatility of NCLMs. + +Automated video dubbing. Automated video dubbing is a task that synthesizes speech from both text and video inputs. A practical application lies in re-recording actors' lines in film and TV post-production to enhance audio quality and ensure synchronization when the original recordings are compromised. Since this task extends text-to-speech (TTS) by incorporating video, research in this area has largely evolved + +![](images/c9353c421717f852f8ab7c6a80dbed78aa564e024ef603b5c3cbf5a3603ac27d.jpg) +(a) Overall architecture + +![](images/a2f8ecd046c78d738f3d28d7331a330da7670986987650a3a48148bdca52e07b.jpg) +(b) Audio-visual (AV) fusion layers +(c) Speech synthesis +Figure 2. Our proposed approach. (a) The Transformer decoder autoregressively generates speech tokens from phonemized text tokens, source speech tokens (extracted via the Encodec encoder), and audio-visual fused tokens, which combine the target speech tokens with lip and facial tokens from the target video. The numbers in each token denote the timestep. (b) The audio-visual fusion layers align the generated target speech tokens with the preceding lip and face tokens, effectively merging their information for better synchronization. (c) The generated speech tokens are decoded by the Encodec decoder to synthesize high-quality speech that is time-aligned with the video. + +from advances in TTS. Early work [19, 24] primarily focuses on modifying TTS models to generate lip-synced speech using lip-cropped videos. Subsequent work [16] proposes reflecting both text and the speaker's emotional state by analyzing facial expressions for speech synthesis. Building on these foundations, HPMDubbing [10] introduces a multilevel approach that aligns visual cues with speech prosody across lip movements, facial expressions, and scene context, modifying FastSpeech [33] to incorporate visual cues. StyleDubber [11] employs a multi-scale style learning framework to improve pronunciation accuracy while maintaining consistent speech style. Zhang et al. [44] propose a two-stage method: first, learning pronunciation from a large-scale text-to-speech corpus, then synchronizing prosody with speech and visual emotions, while ensuring duration consistency between speech timing and lip movements. + +Despite these efforts, existing methods still struggle to synthesize natural and expressive human-like speech. To overcome these limitations, we bring NCLMs to bear for automated video dubbing, leveraging the strong in-context learning ability of these models for natural and high-fidelity speech synthesis. However, directly applying NCLMs off-the-shelf to this task is not possible, as existing NCLMs do not take video inputs as features, nor do they have mechanisms to precisely align the generated speech waveform with a target video of a talker's face. To address this, we propose a novel method for effectively injecting video features into NCLMs, enabling the synthesis of high-fidelity speech that aligns accurately with the video's timing and expressions. + +# 3. Modeling approach + +# 3.1. Overview + +Our primary goal is to generate high-quality, human-like speech while ensuring precise temporal alignment with the target video. Given the source speech $\mathbf{Y}_{\mathrm{src}}$ , text input $\mathbf{Z}$ , and + +target video $\mathbf{V}$ , we design a model that synthesizes the target speech $\mathbf{Y}_{\mathrm{tgt}}$ . To achieve this, we formulate the task as a next-token prediction problem using a Neural Codec Language Model (NCLM), which has been shown to produce natural and expressive speech [3, 5, 7, 21, 30, 38]. + +In conventional NCLM-based speech synthesis, the Transformer decoder processes either text alone or a combination of text and speech tokens to autoregressively generate speech. A straightforward extension for video dubbing is to preprocess video tokens alongside text and speech tokens in the Transformer decoder. However, we empirically found that this increases the sequence length to the point that it becomes difficult for the model to effectively attend to each modality. This leads to performance degradation in content accuracy, intelligibility, and lip synchronization (refer to Sec. 4.4). + +To address this issue, we propose VoiceCraft-Dub, a novel approach that extends the NCLM framework to incorporate visual cues for automated video dubbing. As shown in Fig. 2 (a), VoiceCraft-Dub takes text and discrete source speech tokens as inputs and autoregressively generates target speech tokens, with each token merged with visual features through audio-visual (AV) fusion layers. Specifically, as described in Fig. 2 (b), visual features, including lip movements and facial expressions, are mapped into the NCLM token space using separate adapters. These mapped tokens are fused with target speech tokens via AV fusion layers. Unlike simply prepending video tokens, this fusion mechanism directly aligns visual features with speech generation, enabling the Transformer decoder to produce speech that is temporally aligned with the video while maintaining intelligibility. Finally, the generated target speech tokens are fed into the neural codec decoder to synthesize high-quality speech, as in Fig. 2 (c). In the following subsections, we detail the input features, architecture, and techniques used to achieve high-quality automated video dubbing. + +# 3.2. Input feature extraction + +We describe the input representations for each modality that are provided to the Transformer decoder as inputs. + +Speech input. The source speech $\mathbf{Y}_{\mathrm{src}}$ provides the speaker's voice and prosodic characteristics. We use discrete neural speech tokens extracted from the pre-trained Encoder [14] as the speech input for the Transformer decoder. Encodec is a high-fidelity neural audio codec designed for efficient speech and audio compression. It encodes raw waveforms into a compact discrete representation using a residual vector quantization (RVQ) mechanism while preserving audio quality. The model consists of an encoder that transforms audio into quantized tokens, a decoder that reconstructs the waveform from these tokens, and codebooks that store learned representations for efficient compression. + +Given the source speech $\mathbf{Y}_{\mathrm{src}}$ , we apply the Encoder encoder $E$ to extract a sequence of discrete speech tokens: $\tilde{\mathbf{h}}_{\mathrm{src}} = E(\mathbf{Y}_{\mathrm{src}})$ , where $\tilde{\mathbf{h}}_{\mathrm{src}} \in \mathbb{R}^{T_{\mathrm{src}} \times D \times K}$ , $T_{\mathrm{src}}$ is the number of timesteps, $D$ is the dimension of each token, and $K$ is the number of RVQ codebooks. The Encoder model we use operates at a codec frame rate of $50\mathrm{Hz}$ on $16\mathrm{kHz}$ recordings. + +Text input. The text input $\mathbf{Z}$ specifies the content to be synthesized. To process text effectively, we convert it into phonemes based on the International Phonetic Alphabet (IPA) using the Phonemizer toolkit [4]. Each phoneme is mapped to a learnable vector representation, $\mathbf{h}_{\mathrm{text}} \in \mathbb{R}^{T_{\mathrm{text}} \times D}$ , where $T_{\mathrm{text}}$ is the length of the phoneme sequence and $D$ matches the dimension of the discrete speech tokens. These learnable representations are updated during model training. + +Video input. The target video $\mathbf{V}$ provides the necessary facial cues for speech synthesis. Specifically, the video input serves two primary roles: the full-face video $(\mathbf{V})$ captures facial expression context to enhance expressiveness, while the lip video $(\mathbf{V}_{\mathrm{lip}})$ , cropped from $\mathbf{V}$ , indicates when to speak or pause, ensuring temporal synchronization. + +We leverage off-the-shelf models to extract facial features: AV-HuBERT [35], a powerful audio-visual representation model, serves as a lip encoder to extract lip features from $\mathbf{V}_{\mathrm{lip}}$ and EmoFAN [36], a facial expression analysis model, serves as a face encoder to extract overall facial features from $\mathbf{V}$ . Since these extracted features exist in different representational spaces compared to speech inputs, we introduce adapters composed of MLP layers to align them with NCLM's token space, as shown in Fig. 2 (b): + +$$ +\mathbf {h} _ {\mathrm {l i p}} ^ {\prime} = \operatorname {A d a p t e r} _ {\mathrm {l i p}} (\operatorname {A V - H u B E R T} (\mathbf {V} _ {\mathrm {l i p}})) \tag {1} +$$ + +$$ +\mathbf {h} _ {\text {f a c e}} ^ {\prime} = \operatorname {A d a p t e r} _ {\text {f a c e}} (\operatorname {E m o F A N} (\mathbf {V})) \tag {2} +$$ + +where $\mathbf{h}_{\mathrm{lip}}^{\prime}, \mathbf{h}_{\mathrm{face}}^{\prime} \in \mathbb{R}^{M \times D}$ , $M$ is the number of video frames and $D$ matches the token dimension. We further duplicate the video features to match the temporal resolution of speech tokens: $\mathbf{h}_{\mathrm{lip}}, \mathbf{h}_{\mathrm{face}} \in \mathbb{R}^{T \times D}$ where $T = 2M$ , since video frames are at 25fps while speech tokens are at $50\mathrm{Hz}$ . + +# 3.3. Audio-visual (AV) fusion layers + +Since extending NCLMs to maintain temporal synchronization between synthesized speech and video is non-trivial, our AV fusion layers play a crucial role by effectively integrating visual and speech information, enabling accurate speech generation synchronized with video. As shown in Fig. 2 (a), the AV fusion layers take speech, lip, and facial tokens at each timestep and merge them into AV-fused tokens, which are then provided to the Transformer decoder. We design two AV fusion layers, each combining speech tokens with lip and facial tokens, as described in Fig. 2 (b). Specifically, given the currently generated target speech token $\mathbf{h}_{\mathrm{tgt}}^t$ , we fuse it with the lip and facial tokens as follows: + +$$ +\mathbf {r} _ {\mathrm {t g t}, \mathrm {l i p}} ^ {t} = \operatorname {A V F u s e} _ {\mathrm {l i p}} \left(\left[ \mathbf {h} _ {\mathrm {t g t}} ^ {t}; \mathbf {h} _ {\mathrm {l i p}} ^ {t + 1} \right]\right) \tag {3} +$$ + +$$ +\mathbf {r} _ {\mathrm {t g t}, \text {f a c e}} ^ {t} = \operatorname {A V F u s e} _ {\text {f a c e}} \left(\left[ \mathbf {h} _ {\mathrm {t g t}} ^ {t}; \mathbf {h} _ {\text {f a c e}} ^ {t + 1} \right]\right), \tag {4} +$$ + +where $\mathbf{h}_{\mathrm{tgt}}^t,\mathbf{r}_{\mathrm{tgt,lip}}^t,\mathbf{r}_{\mathrm{tgt,face}}^t\in \mathbb{R}^D$ , AVFuselip/face are the AV fusion layers, $t$ is the current timestep, and $[;]$ denotes channel-wise concatenation. Importantly, we fuse the currently generated speech token $\mathbf{h}_{\mathrm{tgt}}^{t}$ with the preceding lip and facial tokens at timestep $t + 1$ . This allows the model to preview the upcoming visual changes to help the model synthesize temporally and semantically aligned next speech token with next frame video. Finally, $\mathbf{r}_{\mathrm{tgt,lip}}^{t}$ and $\mathbf{r}_{\mathrm{tgt,face}}^{t}$ act as residual values, which are added to the current target token as $\mathbf{h}_{\mathrm{fuse}}^{t} = \mathbf{h}_{\mathrm{tgt}}^{t} + \mathbf{r}_{\mathrm{tgt,lip}}^{t} + \mathbf{r}_{\mathrm{tgt,face}}^{t}$ and then fed into the Transformer decoder for autoregressive speech generation. + +# 3.4. Transformer decoder + +The Transformer decoder follows a GPT-style architecture, with the goal of generating target speech tokens $\tilde{\mathbf{h}}_{\mathrm{tgt}}$ up to $T$ , aligned with the fixed-length video. Specifically, the decoder synthesizes the current target speech token as $\mathbf{h}_{\mathrm{tgt}}^t = G([ \mathbf{h}_{\mathrm{text}}; \mathbf{h}_{\mathrm{src}}; \mathbf{h}_{\mathrm{fuse}}^{< t}])$ , where $G$ is the Transformer decoder. As mentioned in Sec. 3.2, the source speech tokens $\tilde{\mathbf{h}}_{\mathrm{src}} \in \mathbb{R}^{T_{\mathrm{src}} \times D \times K}$ consist of $K$ residual codebooks. These tokens are summed along the codebook dimension, resulting in $\mathbf{h}_{\mathrm{src}} \in \mathbb{R}^{T_{\mathrm{src}} \times D}$ to serve as inputs for the decoder. The generated target speech token $\mathbf{h}_{\mathrm{tgt}}^t$ is passed through $K$ MLP heads to predict logits for each residual codebook, yielding $\tilde{\mathbf{h}}_{\mathrm{tgt}}^t \in \mathbb{R}^{D \times K}$ . This is summed along the codebook dimension before being passed into the audio-visual fusion layer in the next timestep. The text and source speech tokens are combined with their corresponding sinusoidal positional encoding [37] before being fed into the Transformer decoder, while the AV-fused tokens share the positional encoding of the source speech. Following existing approaches [12, 30], we adopt a delayed prediction strategy, effective for autoregressive generation over stacked RVQ tokens. At each timestep $t$ , each residual level is delayed, so the prediction of codebook $k - 1$ at timestep $t$ is conditioned by the prediction of codebook $k$ at the same timestep. + +Once generation is complete, the discrete speech tokens are passed to the Encodec decoder $D$ : $\mathbf{Y}_{\mathrm{tgt}} = D(\tilde{\mathbf{h}}_{\mathrm{tgt}})$ to reconstruct the waveform, as shown in Fig. 2 (c). + +Learning objective. The model is trained to minimize the negative log-likelihood of predicting the correct speech tokens, given the text, source speech, and fused audio-visual tokens. Specifically, following the insights in [30] that weighting the first residual codebook more for updates is effective, we assign higher weights to the first residual codebook compared to the later ones, resulting in the final loss: + +$$ +L (\theta) = - \sum_ {k = 1} ^ {K} \alpha_ {k} \log P _ {\theta} \left(\tilde {\mathbf {h}} _ {\mathrm {t g t}, k} \mid \mathbf {h} _ {\mathrm {s r c}}, \mathbf {h} _ {\mathrm {t e x t}}, \mathbf {h} _ {\mathrm {f u s e}}\right), \tag {5} +$$ + +where $\alpha_{k}$ denotes the weighting hyperparameters. The trainable parameters include the adapters, audio-visual fusion layers, and the Transformer decoder, while AV-HuBERT, EmoFAN, and Encodec remain frozen. + +# 3.5. Implementation details + +The Transformer decoder is initialized with pretrained weights from VoiceCraft [30] and fine-tuned, while the audio-visual fusion layers and adapters are trained from scratch. Encodec has $K = 4$ RVQ codebooks, each with a vocabulary size of $D = 2048$ . The Transformer decoder consists of 16 layers, with hidden and FFN dimensions of 2048 and 8192, and 12 attention heads. The weight hyperparameters for updating the codebook are set to $\alpha = [3,1,1,1]$ . The audio-visual fusion layers each contain a linear layer, and the adapters are designed as two-layer MLPs with GELU activation [17]. For training, we use the AdamW optimizer with a base learning rate of 1e-5 for the Transformer decoder and 1e-3 for the other modules. We train using four RTX 8000 GPUs, with a total batch size of 80K frames, and conduct training for 100K steps with early stopping. + +# 4. Experiments + +We validate the effectiveness of our proposed model through a series of subjective and objective assessments. In addition, we conduct ablation studies to analyze key design choices and demonstrate the extended applicability of our approach to video-to-speech generation. + +# 4.1. Experimental setup + +Dataset. We train and test our model on the LRS3 dataset [1] and our curated CelebV-Dub dataset. LRS3 is an in-the-wild video dataset in English, sourced from TED and TEDx talks. It comprises approximately 439 hours of video with unconstrained, long utterances from thousands of speakers. The pretrain split is utilized for training, while 1,174 utterances from the test split are reserved for testing. + +Since LRS3 videos predominantly feature neutral expressions, we introduce CelebV-Dub to enhance our evaluation. + +This dataset is built upon CelebV-HQ [45] and CelebV-Text [42], collected from diverse in-the-wild sources, such as vlogs, dramas, and movies. CelebV-Dub serves as a challenging benchmark by featuring unconstrained real-world settings and expressive variations. The video sources in CelebV-HQ and CelebV-Text initially contain considerable noise. To address this, we develop a dataset curation algorithm to extract expressive and high-quality in-the-wild videos. This curation results in videos featuring active speakers, segmented into utterances, with suppressed background noise and music, and labeled with pseudo-transcriptions. More details about the dataset construction are provided in the Sec. C of the supplementary material. + +Subjective metrics. Given the generative nature of the task, subjective evaluation is essential to assess the quality of synthesized speech. We conduct two types of evaluations: Mean Opinion Score (MOS) and A/B testing. In the MOS evaluation, participants assess each synthesized speech for naturalness, lip-sync accuracy, intelligibility, expressiveness, and speaker similarity. Each sample is presented alongside the target video, and participants rate it on a 5-point Likert scale, where 1 indicates poor quality and 5 indicates excellent quality. In the A/B testing, we compare the output of VoiceCraft-Dub with existing methods, asking participants to choose which utterance sounds preferable in terms of naturalness, expressiveness, and lip-sync. These evaluations are conducted on 150 utterance samples (100 from LRS3 and 50 from CelebV-Dub) using Amazon Mechanical Turk. + +Objective metrics. We employ several metrics to evaluate various aspects of synthesized speech. Following prior work [30, 38], Word Error Rate (WER) assesses content accuracy, while speaker similarity (spkSIM) measures voice consistency, computed using the Whisper [31] medium.en model and WavLM-TDNN [6], respectively. Lip-sync accuracy (LSE-D and LSE-C) is evaluated using SyncNet [9], which takes both speech and video as inputs. Emotion similarity (emoSIM) is assessed by measuring cosine similarity between synthesized and ground truth speech embeddings using Emotion2Vec [27]. We utilize automatic Mean Opinion Scores (MOS), DNSMOS [32] and UTMOS [34], to evaluate overall speech quality. Low-level metrics, including Mel-Cepstral Distortion (MCD), fundamental frequency distance (F0), and energy distance (Energy), are also employed. + +Competing methods. We compare our method with two open-source approaches: HPMDubbing [10] and StyleDubber [11], which are non-NCLM-based models. Both models use a FastSpeech-like decoder [33] that relies on melspectrogram representations for speech synthesis and requires explicit speech-text alignment through a forced aligner during training. To ensure a fair comparison with our model and improve their generalization, we train HPMDubbing and StyleDubber using both LRS3 and our proposed CelebV-Dub, as the original models were trained on relatively small + +
ModelNaturalnessIntelligibilityLip synchronizationExpressivenessSpeaker similarity
LRS3CelebV-DubLRS3CelebV-DubLRS3CelebV-DubLRS3CelebV-DubLRS3CelebV-Dub
Ground-Truth4.42±0.074.42±0.114.65±0.054.58±0.084.45±0.064.50±0.114.38±0.074.52±0.093.56±0.063.58±0.09
Ours4.30±0.074.18±0.114.52±0.064.42±0.104.37±0.064.42±0.114.33±0.074.41±0.093.36±0.073.44±0.09
HPMDubbing [10]3.12±0.111.65±0.093.33±0.091.74±0.103.96±0.073.34±0.152.93±0.102.34±0.141.66±0.090.72±0.10
StyleDubber [11]2.85±0.111.68±0.093.40±0.092.24±0.133.83±0.083.14±0.152.78±0.102.26±0.131.50±0.100.92±0.11
+ +Table 1. Comparison of Mean Opinion Score (MOS). Our proposed VoiceCraft-Dub significantly outperforms existing methods and performs on par with the ground truth in diverse aspects of human perception metrics. + +
A vs. BNaturalnessExpressivenessLip synchronization
A wins (%)NeutralB wins (%)A wins (%)NeutralB wins (%)A wins (%)NeutralB wins (%)
Ours vs. HPMDubbing [10]87.42.210.488.62.49.075.44.819.8
Ours vs. StyleDubber [11]96.60.62.897.40.42.285.36.68.1
Ground-Truth vs. HPMDubbing [10]95.01.63.495.81.82.483.67.19.3
Ground-Truth vs. Ours55.830.214.056.030.413.661.014.824.2
+ +Table 2. A/B testing results on LRS3. We report the preferences (%) between A and B across various aspects of synthesized speech. In rows 1 and 2, our model is significantly preferred by humans over existing methods. Comparing rows 3 and 4, HPMDubbing is significantly less preferred compared to the ground truth, while our model, highlighted with a gray background, is preferred more. + +datasets. Both models are trained with a batch size of 16 for over 500k steps until convergence, with the rest of the settings following the original configurations [10, 11]. + +Inference. For each test sample, we form {source speech, target text, target video} triplets as input for model inference, where the source speech is from a different utterance of the target speaker. For inference, we employ Nucleus sampling [18] with $p = 0.8$ and a temperature of 1 for all experiments. Although the model produces natural and accurately lip-synced speech, the stochastic nature of autoregressive generation can sometimes result in inaccurate sounds. Thus, similar to the sorting method described in [7], we generate 10 samples, sort them by WER and LSE-D, and select the optimal speech from the sorted examples. We apply this same sorting method to other methods during testing. + +# 4.2. Human evaluation + +As mentioned in Sec. 4.1, validating our model through human perception is the most crucial metric for genuinely evaluating performance. Table 1 shows the Mean Opinion Score (MOS) of our model, along with existing work [10, 11] and the ground truth, on the LRS3 and our collected CelebV-Dub dataset. As shown, our model outperforms existing methods across a wide range of evaluated criteria. Surprisingly, our model performs very closely to the ground truth in many cases. While the existing methods demonstrate favorable results in lip synchronization accuracy, our model significantly excels in naturalness, expressiveness, and speaker similarity. + +Furthermore, Table 2 reports the A/B testing results on LRS3, summarizing preferences between samples from two different methods. In the top two rows, our model is significantly preferred over existing methods, achieving over $88\%$ preference in naturalness and expressiveness, and $75.4\%$ in lip synchronization. Comparing rows 3 and 4, we observe + +that HPMDubbing is significantly less preferred than our model when compared to the ground truth. Interestingly, in row 4, over $40\%$ of the time, humans perceive our output as comparable to or better than the ground truth. After the human evaluation, several comments noted that the synthesized speech from VoiceCraft-Dub is natural and time-synced with the video, making it hard to distinguish from real recordings. These results highlight that the synthesized speech from our model is human-like, natural, and expressive enough to outperform existing methods while competing in quality with the ground truth, emphasizing the effectiveness of our approach. The A/B testing results on CelebV-Dub are available in the Sec. E.1 of the supplementary materials. + +# 4.3.Quantitative results + +We provide quantitative comparisons on each dataset to assess various aspects of the synthesized speech. Specifically, we present results from two variants of our model: one that uses both lip and face input and another that uses only lip video input. In addition to the results from our models and existing methods, we also include the performance of the ground truth and an NCLM-based zero-shot TTS model [30] for reference. Note that the following results are intended to supplement the human evaluation; therefore, we recommend watching the demos for a more accurate assessment. + +Table 3 summarizes the comparison on LRS3. In terms of content accuracy (WER), speaker similarity (spkSIM), and overall speech quality (UTMOS, DNSMOS), our models outperform existing methods by a significant margin. Although HPMDubbing achieves the best lip-sync accuracy (LSE-D/C), our models perform on par with the ground truth. The zero-shot TTS model shows the best WER; however, this approach alone is unsuitable for video dubbing, as it lacks visual cues, leading to poor lip-sync accuracy and + +
MethodWER (↓)LSE-D (↓)LSE-C (↑)spkSIM (↑)UTMOS (↑)DNSMOS (↑)MCD (↓)F0 (↓)Energy (↓)emoSIM (↑)
Ground-Truth1.796.887.63-3.593.18----
Zero-shot TTS [30]0.6810.414.040.3333.483.30---0.682
HPMDubbing [10]7.196.587.990.2193.082.988.7061.543.010.743
StyleDubber [11]3.259.335.210.2952.712.958.29112.802.310.779
Ours (lip-only)1.386.597.970.3613.663.207.8457.871.960.789
Ours (lip & face)1.686.877.740.3733.863.217.5854.011.920.789
+ +Table 3. Comparison on the LRS3 dataset. We highlight the best results in bold and underline the second best. + +
MethodWER (↓)LSE-D (↓)LSE-C (↑)spkSIM (↑)UTMOS (↑)DNSMOS (↑)MCD (↓)F0 (↓)Energy (↓)emoSIM (↑)
Ground-Truth4.157.446.73-2.903.38----
Zero-shot TTS [30]3.8311.682.780.3162.923.40---0.704
HPMDubbing [10]24.067.806.360.1462.102.879.80107.044.380.721
StyleDubber [11]9.4810.403.780.2641.862.948.18180.683.110.784
Ours (lip-only)7.268.275.930.3403.393.407.9775.883.210.778
Ours (lip & face)7.018.136.050.3333.373.407.9172.713.030.782
+ +Table 4. Comparison on our curated CelebV-Dub dataset. We highlight the best results in bold and underline the second best. + +
P.E.AV fuseFaceWER (↓)LSE-D (↓)LSE-C (↑)SIM (↑)UTMOS (↑)DNSMOS (↑)MCD (↓)F0 (↓)Energy (↓)Emotion (↑)
(a)2.857.616.950.3613.683.187.9853.022.110.770
(b)✓2.817.586.960.3603.663.157.9156.822.080.768
(c)✓✓1.386.597.970.3613.663.207.8457.871.960.789
(d)✓✓✓1.686.877.740.3733.863.217.5854.011.920.789
+ +Table 5. Ablation results on the LRS3 dataset [1]. We evaluate the effectiveness of various design choices by comparing different configurations of our method. "P.E." denotes shared positional encoding between speech and video tokens, "AV fuse" indicates using audio-visual fusion layers for merging speech and video tokens, and "Face" refers to using both facial and lip features as inputs. (c) and (d) are the final models used in all other experiments. + +emotion similarity. Our results show that our approach effectively extends NCLM-based models to synthesize accurate lip-synced speech while maintaining intelligibility and naturalness. When comparing our two models, we observe a slight degradation in WER and LSE-D/C with both lip and face input. Nevertheless, performance remains comparable to the ground truth and generally improves across other metrics. Since the LRS3 dataset mostly contains neutral expressions, no significant difference in emotion similarity (emoSIM) is observed between the two variants. + +Similar trends are observed in comparisons conducted on our curated CelebV-Dub dataset, as summarized in Table 4. Again, the zero-shot TTS model achieves the best WER but exhibits poor lip-sync accuracy, highlighting the need of properly integrating visual cues. Both of our models successfully incorporate these cues and generally perform favorably against existing methods. One observation is that our models show lower LSE-D/C compared to HPMDubbing, which we attribute to the instability of SyncNet, a limitation noted in related works [25, 39, 40]. We further provide an analysis in Sec. D of the supplementary materials showing that LSE-D has a low correlation with human evaluation, suggesting it should be used as a reference rather than a definitive metric. Consequently, we argue that the human evaluation described in Sec. 4.2 should be given greater weight as a more accurate measure. Since CelebV-Dub is a more expressive dataset compared to LRS3, we observe an improvement in emoSIM when using both lip and face input over the lip-only variant. + +# 4.4. Ablation study + +We conduct a series of experiments to verify our design choices, as detailed in Table 5. (a) extends the NCLM-based method by prepending video inputs, with video and speech inputs having separate positional encodings, while (b) is the same as (a) but with video tokens sharing the positional encoding of the speech. Comparing (a) and (b), we observe that using shared positional encoding does not significantly affect performance. Therefore, we introduce audio-visual fusion layers as in (c), which significantly improves WER and LSE-D/C compared to both (a) and (b). Using face input alongside lip input (d) shows comparable performance in WER and LSE-D/C to (c) and generally yields better performance. The impact of adding face input is further demonstrated in Table 4, and the generalization results in Sec. E.2 of the supplementary materials, particularly enhancing the emotion similarity (emoSIM). These insights led us to select (c) and (d) as the final model configurations. + +# 4.5. Qualitative results + +In Fig. 3, we visually compare the mel-spectrogram samples converted from the synthesized speech of existing work [10, 11] and our VoiceCraft-Dub, along with those from ground truth recordings. Focusing on the red boxes in columns 1 and 2, we observe that HPMDubbing produces incorrect speech. Although StyleDubber synthesizes accurate content, its mel-spectrograms lack detail and appear blurry. + +![](images/8baff3b35d149a389c5942cd0b4138fa981a0286765c1eb4e40c56e9c18febb8.jpg) +Figure 3. Qualitative results. We compare mel-spectrogram visualizations from ground truth recordings, our model, and prior methods [10, 11] on LRS3 (columns 1-2) and CelebV-Dub (columns 3-4). The texts below each mel-spectrogram represent time-aligned speech extracted using Whisper [31], with red text indicating incorrectly synthesized speech. + +![](images/23c5d83cc52a94cb39533a290d4a82c8cc17cb4bb707d191b7588669a9b55770.jpg) + +![](images/d5a0dbf16ab5065c56ff9f6c16c07d2c3f953a725094609c6d312c9f0540e612.jpg) + +![](images/b9090f343212200de79135d45f112b8214bf3564e9c8960898560d9366cc55a3.jpg) + +
MethodWER (↓)LSE-C (↑)spkSIM (↑)UTMOS (↑)DNSMOS (↑)
VSR [26]26.75----
VSR [26] & TTS [30]29.723.660.3213.483.30
SVTS [28]82.386.020.0771.282.38
Intelligible [8]30.008.020.3102.702.86
Ours (lip-only)28.836.310.3343.523.16
+ +Table 6. Results of video-to-speech generation on LRS3 [1]. Although not specifically designed for this task, our model performs favorably, particularly in content accuracy (WER) and overall quality (UTMOS and DNSMOS) compared to existing methods. + +These observations align with the results in Sec. 4.2 and Sec. 4.3, where StyleDubber received the lowest human and automatic MOS scores. In column 3, all models synthesize correct speech, but HPMDubbing and StyleDubber generate noticeable noise during brief pauses, as indicated by the red boxes. In column 4, existing methods produce incorrect content and blurry speech with considerable noise. In contrast, samples from our model across all experiments exhibit fine details in mel frequency and closely match the ground truth. These results demonstrate that our model synthesizes speech with accurate content and high quality. + +# 4.6. Application: Video-to-speech generation + +We demonstrate the extensibility of our model by adapting it to the video-to-speech synthesis task. Unlike automated video dubbing, video-to-speech generation requires only source speech and video inputs (without text) to synthesize speech. We use an expert Visual Speech Recognition (VSR) model [26] to extract text from the silent video. The extracted text, along with the provided video and source speech, is then fed into our model for video-to-speech generation. + +Table 6 compares the performance of our approach with several other methods, including standalone VSR, a combination of VSR and zero-shot TTS [30], and the existing methods [26, 28]. Although video-to-speech synthesis is not our primary focus, the results show that our model per + +forms favorably against the existing methods across various metrics. While the combination of VSR and TTS models achieves similar performance, as expected, it fails to synchronize with the video. This highlights that our approach effectively extends the capabilities of current TTS models to handle video inputs robustly. These results confirm that extending the NCLM-based model to accommodate visual inputs is both versatile and effective for speech synthesis. + +# 5. Discussion and conclusion + +Discussion. While our method shows high-quality speech synthesis, it also has limitations. Despite the effectiveness of full-face input, the model sometimes follows the characteristics of the source speech over the face input. Additionally, the autoregressive approach can produce incorrect samples if the next token prediction fails. Nonetheless, the results show that our approach is effective in synthesizing expressive and accurately time-aligned speech. Explicitly infusing emotional state or specifying speech length could address these limitations, which we plan to explore in future work. + +Conclusion. In this work, we propose VoiceCraft-Dub, a novel NCLM-based approach for automatic video dubbing. We extend the high-quality, human-like speech synthesis capabilities of NCLMs to incorporate visual cues as inputs. Specifically, we perform audio-visual fusion to provide direct alignment signals to NCLM for visually aligned speech synthesis. Additionally, we curate the expressive CelebV-Dub dataset, specifically designed for dubbing tasks. Our extensive human evaluations and quantitative results show that VoiceCraft-Dub outperforms existing methods and performs on par with the original recordings in terms of naturalness, intelligibility, and lip synchronization. We also demonstrate the versatility of our approach by adapting it to the video-to-speech generation task. We believe our approach will lead to a new paradigm in human-like video dubbing synthesis, enabling more immersive content creation. + +# Acknowledgments + +K. Sung-Bin and T.-H. Oh were partially supported by IITP grant funded by the Korea government (MSIT) (No. RS-2024-00457882, National AI Research Lab Project; No. RS-2021-II212068, Artificial Intelligence Innovation Hub) and the National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT) (No. NRF-2021R1C1C1006799). T.-H. Oh was partially supported by the KAIST Cross-Generation Collaborative Lab Project. J. Choi and J. S. Chung were partially supported by the IITP-ITRC grant funded by the Korean government (MSIT, IITP-2025-RS-2023-00259991). + +# References + +[1] Triantafyllos Afouras, Joon Son Chung, and Andrew Zisserman. Lrs3-ted: a large-scale dataset for visual speech recognition. arXiv preprint arXiv:1809.00496, 2018. 2, 5, 7, 8 +[2] Andrea Agostinelli, Timo I Denk, Zalán Borsos, Jesse Engel, Mauro Verzetti, Antoine Caillon, Qingqing Huang, Aren Jansen, Adam Roberts, Marco Tagliasacchi, et al. Musicl: Generating music from text. arXiv preprint arXiv:2301.11325, 2023. 2 +[3] Alan Baade, Puyuan Peng, and David Harwath. Neural codec language models for disentangled and textless voice conversion. In Conference of the International Speech Communication Association (INTERNSPSPEECH), 2024. 1, 2, 3 +[4] Mathieu Bernard and Hadrien Titeux. Phonemizer: Text to phones transcription for multiple languages in python. Journal of Open Source Software, 2021. 4 +[5] Zalán Borsos, Raphaël Marinier, Damien Vincent, Eugene Kharitonov, Olivier Pietquin, Matt Sharifi, Dominik Roblek, Olivier Teboul, David Grangier, Marco Tagliasacchi, et al. Audiolm: a language modeling approach to audio generation. IEEE/ACM transactions on audio, speech, and language processing, 2023. 2, 3 +[6] Sanyuan Chen, Chengyi Wang, Zhengyang Chen, Yu Wu, Shujie Liu, Zhuo Chen, Jinyu Li, Naoyuki Kanda, Takuya Yoshioka, Xiong Xiao, et al. Wavlm: Large-scale self-supervised pre-training for full stack speech processing. IEEE Journal of Selected Topics in Signal Processing, 2022. 5 +[7] Sanyuan Chen, Shujie Liu, Long Zhou, Yanqing Liu, Xu Tan, Jinyu Li, Sheng Zhao, Yao Qian, and Furu Wei. Vall-e 2: Neural codec language models are human parity zero-shot text to speech synthesizers. arXiv preprint arXiv:2406.05370, 2024. 1, 2, 3, 6 +[8] Jeongsoo Choi, Minsu Kim, and Yong Man Ro. Intelligible lip-to-speech synthesis with speech units. In Conference of the International Speech Communication Association (INTERSPEECH), 2023. 8 +[9] Joon Son Chung and Andrew Zisserman. Out of time: automated lip sync in the wild. In Computer Vision-ACCV 2016 Workshops: ACCV 2016 International Workshops, Taipei, Taiwan, November 20-24, 2016, Revised Selected Papers, Part II 13, 2017. 5 + +[10] Gaoxiang Cong, Liang Li, Yuankai Qi, Zheng-Jun Zha, Qi Wu, Wenyu Wang, Bin Jiang, Ming-Hsuan Yang, and Qingming Huang. Learning to dub movies via hierarchical prosody models. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 3, 5, 6, 7, 8 +[11] Gaoxiang Cong, Yuankai Qi, Liang Li, Amin Beheshti, Zhedong Zhang, Anton Hengel, Ming-Hsuan Yang, Chenggang Yan, and Qingming Huang. Stylehubber: Towards multi-scale style learning for movie dubbing. In Findings of the Association for Computational Linguistics: ACL 2024, 2024. 2, 3, 5, 6, 7, 8 +[12] Jade Copet, Felix Kreuk, Itai Gat, Tal Remez, David Kant, Gabriel Synnaeve, Yossi Adi, and Alexandre Défossez. Simple and controllable music generation. In Advances in Neural Information Processing Systems (NeurIPS), 2023. 4 +[13] Jade Copet, Felix Kreuk, Itai Gat, Tal Remez, David Kant, Gabriel Synnaeve, Yossi Adi, and Alexandre Défossez. Simple and controllable music generation. In Advances in Neural Information Processing Systems (NeurIPS), 2023. 2 +[14] Alexandre Défossez, Jade Copet, Gabriel Synnaeve, and Yossi Adi. High fidelity neural audio compression. arXiv preprint arXiv:2210.13438, 2022. 1, 2, 4 +[15] Hugo Flores Garcia, Prem Seetharaman, Rithesh Kumar, and Bryan Pardo. Vampnet: Music generation via masked acoustic token modeling. arXiv preprint arXiv:2307.04686, 2023. 2 +[16] Michael Hassid, Michelle Tadmor Ramanovich, Brendan Shillingford, Miaosen Wang, Ye Jia, and Tal Remez. More than words: In-the-wild visually-driven prosody for text-to-speech. 2022 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 3 +[17] Dan Hendrycks and Kevin Gimpel. Gaussian error linear units (gelus). arXiv preprint arXiv:1606.08415, 2016. 5 +[18] Ari Holtzman, Jan Buys, Li Du, Maxwell Forbes, and Yejin Choi. The curious case of neural text degeneration. arXiv preprint arXiv:1904.09751, 2019. 6 +[19] Chenxu Hu, Qiao Tian, Tingle Li, Wang Yuping, Yuxuan Wang, and Hang Zhao. Neural dubber: Dubbing for videos according to scripts. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 1, 3 +[20] Alexander B Kain, John-Paul Hosom, Xiaochuan Niu, Jan PH Van Santen, Melanie Fried-Oken, and Janice Staehely. Improving the intelligibility of dysarthric speech. Speech communication, 2007. 2 +[21] Eugene Kharitonov, Damien Vincent, Zalán Borsos, Raphaël Marinier, Sertan Girgin, Olivier Pietquin, Matt Sharifi, Marco Tagliasacchi, and Neil Zeghidour. Speak, read and prompt: High-fidelity text-to-speech with minimal supervision. Transactions of the Association for Computational Linguistics, 2023. 1, 2, 3 +[22] Felix Kreuk, Gabriel Synnaeve, Adam Polyak, Uriel Singer, Alexandre Défossez, Jade Copet, Devi Parikh, Yaniv Taigman, and Yossi Adi. Audiogen: Textually guided audio generation. arXiv preprint arXiv:2209.15352, 2022. 2 +[23] Guanghou Liu, Yongmao Zhang, Yi Lei, Yunlin Chen, Rui Wang, Zhifei Li, and Lei Xie. Promptstyle: Controllable style transfer for text-to-speech with natural language descriptions. arXiv preprint arXiv:2305.19522, 2023. 2 + +[24] Junchen Lu, Terrak Sisman, Rui Liu, Mingyang Zhang, and Haizhou Li. Visualtts: Tts with accurate lip-speech synchronization for automatic voice over. In IEEE International Conference on Acoustics, Speech, and Signal Processing (ICASSP), 2022. 3 +[25] Junxian Ma, Shiwen Wang, Jian Yang, Junyi Hu, Jian Liang, Guosheng Lin, Kai Li, Yu Meng, et al. Sayanything: Audiodriven lip synchronization with conditional video diffusion. arXiv preprint arXiv:2502.11515, 2025. 7 +[26] Pingchuan Ma, Alexandros Haliassos, Adriana Fernandez-Lopez, Honglie Chen, Stavros Petridis, and Maja Pantic. Auto-avsr: Audio-visual speech recognition with automatic labels. In IEEE International Conference on Acoustics, Speech, and Signal Processing (ICASSP), 2023. 2, 8 +[27] Ziyang Ma, Zhisheng Zheng, Jiaxin Ye, Jinchao Li, Zhifu Gao, Shiliang Zhang, and Xie Chen. emotion2vec: Self-supervised pre-training for speech emotion representation. arXiv preprint arXiv:2312.15185, 2023. 5 +[28] Rodrigo Mira, Alexandros Haliassos, Stavros Petridis, Björn W Schuller, and Maja Pantic. Svts: scalable video-to-speech synthesis. In Conference of the International Speech Communication Association (INTERSPEECH), 2022. 8 +[29] Keigo Nakamura, Tomoki Toda, Hiroshi Saruwatari, and Kiyohiro Shikano. Speaking-aid systems using gmm-based voice conversion for electrolaryngeal speech. Speech communication, 2012. 2 +[30] Puyuan Peng, Po-Yao Huang, Shang-Wen Li, Abdelrahman Mohamed, and David Harwath. Voicecraft: Zero-shot speech editing and text-to-speech in the wild. arXiv preprint arXiv:2403.16973, 2024. 1, 2, 3, 4, 5, 6, 7, 8 +[31] Alec Radford, Jong Wook Kim, Tao Xu, Greg Brockman, Christine McLeavey, and Ilya Sutskever. Robust speech recognition via large-scale weak supervision. In International Conference on Machine Learning (ICML), 2023. 5, 8 +[32] Chandan KA Reddy, Vishak Gopal, and Ross Cutler. Dnmos p. 835: A non-intrusive perceptual objective speech quality metric to evaluate noise suppressors. In IEEE International Conference on Acoustics, Speech, and Signal Processing (ICASSP), 2022. 5 +[33] Yi Ren, Yangjun Ruan, Xu Tan, Tao Qin, Sheng Zhao, Zhou Zhao, and Tie-Yan Liu. Fastspeech: Fast, robust and controllable text to speech. In Advances in Neural Information Processing Systems (NeurIPS), 2019. 3, 5 +[34] Takaaki Saeki, Detai Xin, Wataru Nakata, Tomoki Koriyama, Shinnosuke Takamichi, and Hiroshi Saruwatari. Utmos: Utokyo-sarulab system for voicemos challenge 2022. arXiv preprint arXiv:2204.02152, 2022. 5 +[35] Bowen Shi, Wei-Ning Hsu, Kushal Lakhotia, and Abdelrahman Mohamed. Learning audio-visual speech representation by masked multimodal cluster prediction. In International Conference on Learning Representations (ICLR), 2022. 4 +[36] Antoine Toisoul, Jean Kossaifi, Adrian Bulat, Georgios Tzimiropoulos, and Maja Pantic. Estimation of continuous valence and arousal levels from faces in naturalistic conditions. Nature Machine Intelligence, 2021. 4 +[37] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia + +Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems (NeurIPS), 2017. 4 +[38] Chengyi Wang, Sanyuan Chen, Yu Wu, Ziqiang Zhang, Long Zhou, Shujie Liu, Zhuo Chen, Yanqing Liu, Huaming Wang, Jinyu Li, et al. Neural codec language models are zero-shot text to speech synthesizers. arXiv preprint arXiv:2301.02111, 2023. 1, 2, 3, 5 +[39] Dogucan Yaman, Fevziye Irem Eyiokur, Leonard Bärmann, Seymanur Akti, Hazim Kemal Ekenel, and Alexander Waibel. Audio-visual speech representation expert for enhanced talking face video generation and evaluation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 7 +[40] Dogucan Yaman, Fevziye Irem Eyiokur, Leonard Bärmann, Hazim Kemal Ekenel, and Alexander Waibel. Audio-driven talking face generation with stabilized synchronization loss. In European Conference on Computer Vision (ECCV), 2024. 7 +[41] Dongchao Yang, Songxiang Liu, Rongjie Huang, Chao Weng, and Helen Meng. Instructtts: Modelling expressive tts in discrete latent space with natural language style prompt. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 2024. 2 +[42] Jianhui Yu, Hao Zhu, Liming Jiang, Chen Change Loy, Weidong Cai, and Wayne Wu. Celebv-text: A large-scale facial text-video dataset. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 5 +[43] Neil Zeghidour, Alejandro Luebs, Ahmed Omran, Jan Skoglund, and Marco Tagliasacchi. Soundstream: An end-to-end neural audio codec. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 2021. 1, 2 +[44] Zhedong Zhang, Liang Li, Gaoxiang Cong, Haibing Yin, Yuhan Gao, Chenggang Yan, Anton van den Hengel, and Yuankai Qi. From speaker to dubber: movie dubbing with prosody and duration consistency learning. In ACM International Conference on Multimedia (MM), 2024. 3 +[45] Hao Zhu, Wayne Wu, Wentao Zhu, Liming Jiang, Siwei Tang, Li Zhang, Ziwei Liu, and Chen Change Loy. Celebv-hq: A large-scale video facial attributes dataset. In European Conference on Computer Vision (ECCV), 2022. 2, 5 \ No newline at end of file diff --git a/ICCV/2025/VoiceCraft-Dub_ Automated Video Dubbing with Neural Codec Language Models/images.zip b/ICCV/2025/VoiceCraft-Dub_ Automated Video Dubbing with Neural Codec Language Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..eea5783298bba630d438456d311d1161cfbccabb --- /dev/null +++ b/ICCV/2025/VoiceCraft-Dub_ Automated Video Dubbing with Neural Codec Language Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ac3bc507537d716a9d1f389f751cae145d2de797bce2153d9d85be4ad7d3737c +size 532245 diff --git a/ICCV/2025/VoiceCraft-Dub_ Automated Video Dubbing with Neural Codec Language Models/layout.json b/ICCV/2025/VoiceCraft-Dub_ Automated Video Dubbing with Neural Codec Language Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..b0da66da5d3025360a9fec5a3e6864ff25862d8d --- /dev/null +++ b/ICCV/2025/VoiceCraft-Dub_ Automated Video Dubbing with Neural Codec Language Models/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f69375d1338ed13b436643db9acda20211a6c62cc5df8808caaad1a48bb082c8 +size 349893 diff --git a/ICCV/2025/VoluMe - Authentic 3D Video Calls from Live Gaussian Splat Prediction/aca36b55-c429-4d0d-a0b8-cd49d838955e_content_list.json b/ICCV/2025/VoluMe - Authentic 3D Video Calls from Live Gaussian Splat Prediction/aca36b55-c429-4d0d-a0b8-cd49d838955e_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..c0faeb1858112eaf9870ed4b6dd2e96bcef3910b --- /dev/null +++ b/ICCV/2025/VoluMe - Authentic 3D Video Calls from Live Gaussian Splat Prediction/aca36b55-c429-4d0d-a0b8-cd49d838955e_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ff141094bfc2a908d47b2c77345850d003995d9430c66048d50323491cc71910 +size 77465 diff --git a/ICCV/2025/VoluMe - Authentic 3D Video Calls from Live Gaussian Splat Prediction/aca36b55-c429-4d0d-a0b8-cd49d838955e_model.json b/ICCV/2025/VoluMe - Authentic 3D Video Calls from Live Gaussian Splat Prediction/aca36b55-c429-4d0d-a0b8-cd49d838955e_model.json new file mode 100644 index 0000000000000000000000000000000000000000..c2b771e2488058b8f70e792c497af733aeca1c36 --- /dev/null +++ b/ICCV/2025/VoluMe - Authentic 3D Video Calls from Live Gaussian Splat Prediction/aca36b55-c429-4d0d-a0b8-cd49d838955e_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a007a6f73846bea9729bce410f645296a0977001ff3a70f7f75512eb20514e28 +size 97937 diff --git a/ICCV/2025/VoluMe - Authentic 3D Video Calls from Live Gaussian Splat Prediction/aca36b55-c429-4d0d-a0b8-cd49d838955e_origin.pdf b/ICCV/2025/VoluMe - Authentic 3D Video Calls from Live Gaussian Splat Prediction/aca36b55-c429-4d0d-a0b8-cd49d838955e_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..b9f65ea3e72be112a3d402006be2274939d07cdb --- /dev/null +++ b/ICCV/2025/VoluMe - Authentic 3D Video Calls from Live Gaussian Splat Prediction/aca36b55-c429-4d0d-a0b8-cd49d838955e_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:982a3da5a4248b574ee0a37b75cc223b4fc913687c185c464afb420293908db2 +size 20606143 diff --git a/ICCV/2025/VoluMe - Authentic 3D Video Calls from Live Gaussian Splat Prediction/full.md b/ICCV/2025/VoluMe - Authentic 3D Video Calls from Live Gaussian Splat Prediction/full.md new file mode 100644 index 0000000000000000000000000000000000000000..8802ad6e282f96e8dff8fd1ce57d939f2ef3da80 --- /dev/null +++ b/ICCV/2025/VoluMe - Authentic 3D Video Calls from Live Gaussian Splat Prediction/full.md @@ -0,0 +1,313 @@ +# VoluMe – Authentic 3D Video Calls from Live Gaussian Splat Prediction + +Martin de La Gorce +Givi Meishvili + +Charlie Hewitt Marek Kowalski + +Tibor Takacs Thomas J. + +Robert Gerdisch +ashman An + +Zafiirah Hosenie +Onio Criminisi + +Microsoft, Cambridge, UK + +Input frame 3D reconstruction rendered from -40, -20, 0, +20, and +40 degrees + +Input frame 3D reconstruction rendered from -40, -20, 0, +20, and +40 degrees + +![](images/8b84510f20c06df910debd12979c7c07f8e7688e0bd78814c889a5080fb83a58.jpg) +Figure 1. Given a streaming monocular video of a person, our method reconstructs each frame as a realistic 3D Gaussian splat of the moment captured by that frame. Avatar systems build a model of the person's appearance in the past, and may not fully adapt to how a person looks in the present. Our method builds a realistic representation in real time and without enrolment. We use the term authenticity for the property that the input frame is faithfully represented, shown in this figure by the match between the input frame and the reconstruction when rendered to the original camera's viewpoint. Our method faithfully represents eye glasses, headwear and diverse head and facial hair. + +# Abstract + +Virtual 3D meetings offer the potential to enhance copresence, increase engagement and thus improve effectiveness of remote meetings compared to standard 2D video calls. However, representing people in 3D meetings remains a challenge; existing solutions achieve high quality by using complex hardware, making use of fixed appearance via enrolment, or by inverting a pre-trained generative model. These approaches lead to constraints that are unwelcome and ill-fitting for videoconferencing applications. + +We present the first method to predict 3D Gaussian reconstructions in real time from a single 2D webcam feed, where the 3D representation is not only live and realistic, but also authentic to the input video. By conditioning the 3D representation on each video frame independently, our reconstruction faithfully recreates the input video from the captured viewpoint (a property we call authenticity), while generalizing realistically to novel viewpoints. Additionally, we introduce a stability loss to obtain reconstructions that are temporally stable on video sequences. + +We show that our method delivers state-of-the-art accuracy in visual quality and stability metrics compared to existing methods, and demonstrate our approach in live one- + +to-one 3D meetings using only a standard 2D camera and display. This demonstrates that our approach can allow anyone to communicate volumetrically, via a method for 3D videoconferencing that is not only highly accessible, but also realistic and authentic. + +# 1. Introduction + +Online video meetings have transformed the ability for people to communicate and collaborate while physically remote, but standard 2D video meetings create challenges with fatigue [24], make it difficult to infer where attention is directed [28], and have limited effectiveness [22]. Recent evidence suggests that virtual 3D meetings offer the potential to enhance copresence, increase engagement, and thus improve communication effectiveness [13, 28, 34]. However, representing people in 3D meetings, with accuracy and realism in both appearance and movement, and on consumer devices, is an unsolved challenge. + +Existing solutions achieve high quality either with complicated hardware, such as multiple input cameras, dedicated sensors and powerful processors, e.g., Project Starline [13]. Alternatively, photoreal avatars [26] use a latent represent + +tation to animate a pre-enrolled model of the user, limiting the appearance to what was seen at enrolment time, and thus intentionally failing to adapt to the current moment. Finally, for approaches such as Live 3D Portrait [35] and Tri-PlaneNet [1], for which the 3D reconstruction is learned or predicted from the EG3D generative model [4], the 3D reconstructions are limited by the textures and geometry learned by that model (see Fig. 9). In each of these approaches, the constraints on the 3D representations are unwelcome for videoconferencing, where users need the rich, authentic and unconstrained representation that is familiar from 2D video, without the need for specialized hardware. In order to meet this challenge, we need a 3D representation that is + +1. Authentic: generates images that match the input video foreground (with every detail of, for example, clothes, hair, and glasses from the present moment), when rendered from the original camera's viewpoint; +2. Realistic: generates plausible and realistic reconstructions when rendered in new views, while supporting the full diversity of human communication (e.g. including diverse hair, headwear and accessories); +3. Live: runs in real time on consumer devices; +4. Stable: generates predictions that are stable with respect to time and as the viewpoint changes, to avoid flickering. + +. In this paper, we demonstrate the first method to satisfy all four properties. Examples of our results are shown in Fig. 1. Note that we can do so without the need for a full $360^{\circ}$ reconstruction, while addressing an important class of 3D video meetings (the capture volume for Project Starline [13], for example, necessitates novel views only up to $55^{\circ}$ from frontal). In this paper, our contributions are: + +- We demonstrate 3D Gaussian reconstructions from 2D video that are live, authentic and realistic for a range of novel views that are up to $40^{\circ}$ offset from the input. +- We show that these reconstructions are competitive with the state of the art for realism and quality, and excel at reconstructing the input view, even for a wide range of unconstrained inputs typical of videoconferencing scenarios (such as diverse hair, lighting, eyewear, clothing and accessories). We demonstrate state-of-the-art results on PSNR and LPIPS metrics on the Ava256 and Cafca datasets, and better results on jitter metrics on the Ava256 dataset. +- We demonstrate a real-time system that uses these 3D reconstructions to enable 3D videoconferencing at 30FPS with simulated motion parallax, using only a standard RGB camera and screen. + +# 1.1. Related work + +Photorealistic avatars. Avatar-based techniques rely on an enrolment process, where an appearance and geometry model is built for a given user. Enrolment may involve capturing multiple views of the user from a phone [3, 12] or multi-camera rig [26]; this process may be lengthy, as + +the user needs to show different head positions, facial expressions etc., and may be inaccessible or expensive in the case of multi-camera setups. Recent work on photorealistic avatars relies on radiance field representations, first by adapting NeRF [21] to represent a head model using deformation fields [43] or deformable volumes [2, 6, 18]. With Gaussian splatting, Kerbl et al. [10] demonstrate that radiance fields can be rendered more efficiently via rasterization, and inspire a large range of work that adapts Gaussian splatting to digital humans [5, 7, 23, 27, 37, 39, 40]. + +In this work, we draw inspiration from the realism demonstrated by Gaussian avatars, but apply this representation to a setting where authenticity is important; we cannot rely on a fixed enrolled appearance from the past. Instead we want to ensure that our 3D reconstruction adapts instantly to any appearance changes, as is true for all 2D video calls today. + +Single view reconstruction. A second class of related work predicts a 3D reconstruction from a single 2D image or video frame, without prior enrolment. These methods can offer high authenticity and realism, but the challenge of reconstructing 3D from a single 2D input makes it harder for these methods to be live and stable. + +Live 3D Portrait [35] and TriPlaneNet [1] both make use of the EG3D generative model [4] and predict an implicit 'triplane' radiance field representation. Live 3D Portrait runs in real time and achieves impressive results, satisfying many of our requirements, but by distilling knowledge from a generative 3D model, the resulting reconstructions are constrained by said model (see Fig. 9), and TriPlaneNet predicts a latent code for EG3D, making this a hard constraint. + +The most promising explicit representation methods are those that predict Gaussian splat [10] outputs, as they can capture thin structures, specularities and transparencies. Xu et al. [38] and Zou et al. [44] combine the benefits of (implicit) triplane and (explicit) Gaussian splat representations, but both approaches are based on transformer architectures that do not scale well to live high-resolution Gaussian splat prediction. Lyu et al. [19] train a diffusion model to predict multiple views from a single image, which are then used to predict 3D Gaussians using GS-LRM [41]; this grants high realism, but the multi-view diffusion prediction makes it harder to produce stable outputs for video sequences. + +Szymanowicz et al. [32] circumvent these problems using a 2D convolutional U-Net architecture [30] that directly predicts a Gaussian splat representation, or 'splatter image', in a method that is capable of live reconstruction but is not applied to human reconstruction, nor to video sequences. + +In this paper, we build upon the Splatter Image [32] method and demonstrate that direct Gaussian splat prediction can satisfy the requirements for realistic, authentic, live and stable 3D video calls. In contrast to Szymanowicz et al. [32], we use a planar homography to reduce perspective distortion when a person appears near the border of the input image, + +we incorporate a scaling step during model training to deal with varying distances between the camera and the subject, we output two Gaussians per pixel to increase the quality of the reconstruction, and we add a stability loss to reduce temporal flicker in reconstructions of video sequences. A concurrent work [31] also uses two Gaussians per pixel, but for scene reconstruction rather than human reconstruction. + +3D reconstruction with synthetic data. One interesting aspect of our work is the choice to train on purely synthetic data. This enables multi-view data with perfect camera parameters that emulates in-the-wild scenarios; all important factors for our method. However, recent work reconstructing 3D humans using methods trained on synthetic data have encountered issues generalizing to real images at test time. For example, Rodin [36] trains a generative model to create 3D avatar representations. Due to the synthetic training data their 3D results look noticeably synthetic, matching the domain of the training data. Similar issues are visible for Live 3D Portrait [35] and TriPlaneNet [1] (e.g., eyes following the camera), although the EG3D-based synthetic data has a smaller domain gap to real images. Other recent methods aim to explicitly handle the domain gap by training a generative prior model on synthetic data and fine-tuning on real enrolment data [2, 27]. This does not suit our application, as the results are neither live nor authentic. Instead we choose to follow Splatter Image [32], in which the output is much more directly tied to the input with a feed-forward network. This allows our network to generalize well to real images, despite being trained only on synthetic data. The skip connections and direct-sampling elements of the architecture are critical in allowing information to be forwarded directly from the input image to the output 3D representation, while the bottleneck can learn a strong prior over 3D geometry based on the synthetic training data. + +# 2. Training the 3D Reconstruction Model + +# 2.1. System Design + +Our method takes a single image as input, performs face detection on that image, extracts a region around the face and runs our model on this region. The output of the model is a set of 3D Gaussians that represent the 3D shape and appearance of the face in the image. The model architecture is based on Splatter Image [32]. We also use a U-Net architecture [25] to predict the Gaussians, though we choose a smaller backbone than SongUNet [30] to reduce execution time. We further reduce the number of convolution layers to only two per resolution and increase the number of resolutions to five to maintain a large receptive field, also omitting the self-attention layers. To retain high-quality results we add four new ingredients in our method, described below. + +Optimizable layers. The U-Net architecture uses only 2D convolutions and non-linear activation, which is limited in + +![](images/12b049f2f0c8613b65fed0907ce45496ba2aef912ed6be4f07acf86611e3403f.jpg) +(a) + +![](images/3fd16b3d81e177aae11d58cc96f67b1e5f196cfcc332c0ee30b8cab6fd5ad202.jpg) +(b) + +![](images/30391e586777ecb163c1b35dc54574ed8948cebf536959286f270584ae7808ac.jpg) +(c) +Figure 2. Splatter image components: a) input ROI, b) colour, c) opacity, d) log scale, e) z-depth, f) xyz, g) rotation (Euler angles) + +![](images/1eea09d6e1dbb7e055d16d1ad97a08b1764a725319794ecfeb753988479cc7ec.jpg) +(d) + +![](images/7c6c9ff3f12e0ba1b96c3c00bff21f1980f41d011d21c26cbcc1e98a2a030021.jpg) +(e) + +![](images/77a17cd3c08fea83524fa8107e414a20808df9723a89f0d2ff5c872164036e4a.jpg) +(f) + +![](images/6020509880996e940c20a7a5c2892bd4cc3f13dd8f65bac2eff641d88d921024.jpg) +(g) + +its ability to create output with high spatial frequencies. This limitation becomes stronger as we decrease the size of the network when aiming for a network that runs in real time. To mitigate this problem, we add explicit channels to the network that can take a different value for each pixel and whose values are optimized alongside the other parameters of the network. The estimizable channels enable a lightweight shallow network to generate high frequency patterns with a minor increase in compute, and so improve the quality of the final reconstruction. We add four channels at each stage of the decoding part of the U-Net architecture. + +Direct colour sampling. Following Szymanowicz et al. [32], we use offsets that allow 3D Gaussians to move away from the rays associated to their pixel in the splatter image. For a Gaussian that is visible with large offset, the information provided by the skip connection in the U-Net is not well collocated with the location that the Gaussian would reproject in the input image. As a result, the information most useful for predicting the Gaussian colour is not provided by the skip connection and needs to be propagated through the deeper layers, which limits the ability to maintain texture details. In order to mitigate this, we compute the 2D location of the Gaussian after applying offset and projection into the input image and then use direct bilinear sampling of the input image where the Gaussian reprojects. We provide this extra information as additional channels to a shallow convolutional residual block that predicts the Gaussian colour, scales, and rotation. + +Multiple Gaussians per pixel. In order to provide more flexibility to the representation, we generate multiple Gaussians per pixel [31]. This allows the network to predict multiple layers of the surface while only increasing the computational cost of the network by a small amount. We determined experimentally that using two Gaussians per pixel provided a good quality improvement, while three did not yield noticeable improvements and significantly increased rendering time. As visible in Fig. 2 the Gaussians in the first layer have larger scales and seem to contribute more to occluded regions, while the ones in the second layer have much smaller scales and tend to represent high-frequency details in the visible regions. + +ROI homography mapping. Our method runs on a region of interest (ROI) around the face, rather than the whole + +image. The simplest approach to extract an ROI is a simple image crop. However, this has important limitations when it comes to predicting 3D geometry; faces away from the image centre appear distorted in the cropped images due to the perspective projection. CLIFF [16] addresses this by providing the focal length and central point of a new virtual camera with oblique frustum that corresponds to the cropped region as input to the DNN. We instead create a virtual camera with a symmetric frustum, that is rotated to point to the centre of the face in the original image, with the FOV computed such that the face covers a third of the FOV angle. We then warp the input image to match this new camera using a homography. Because this camera has a symmetric frustum, we only need to provide the normalized focal length $f / w$ as input to the neural network. We concatenate the input image with the ray directions and two channels that contain the repeated normalized focal length and its inverse. + +# 2.2. Data generation + +Our method takes one image as input and generates a set of Gaussians that are rendered in new views using a differentiable renderer. These renders are compared to ground truth views to obtain a training loss. For training we therefore need a multi-view dataset, but capturing a diverse, multi-view, in-the-wild dataset at a scale suitable to train a Gaussian splat prediction model is a formidable challenge. Instead, we use exclusively synthetic data to train our method, employing the data generation pipeline described by Hewitt et al. [8]. This enables us to create multi-view data with perfect camera parameters, including diverse subjects, and emulating in-the-wild scenarios. Our synthetic dataset includes $N_{s} = 20,000$ subjects with varied face shape, skin texture, hair, clothing, glasses, body pose and facial expression. + +In order for the prediction to align well with the ground truth in the novel views we need the model to predict accurate absolute positions of the Gaussians. Due to the depth ambiguity that is inherent to single-view reconstruction, this is a very challenging task. To resolve this, previous Gaussian splat prediction methods train models using synthetic data with 3D objects positioned at fixed distance to the camera [32]. This approach has critical limitations: the same face looks very different when captured from cameras at different distances; by training the model with a fixed distance, it cannot learn strong priors about the 3D proportions of faces. Further, absolute depth prediction is not required for our video call application, where small offsets are acceptable as long as the face shape is predicted correctly. Requiring the trained model to predict very accurate distances during training is therefore wasteful in terms of model capacity. + +We instead resolve the depth ambiguity by rescaling the prediction during training (see Sec. 2.3). This is enabled by high-quality ground-truth depth generation from the synthetic pipeline. As such, we sample input cameras at various + +![](images/6e943afdab34a2e5b523ed7b019ad529b89a9bc504e77917266cb2ea55574b24.jpg) +(a) + +![](images/2c342e2bddb212ccfedf35419bc8dcc80ca404a859a16648b1e72f5400bcc349.jpg) +(b) + +![](images/97bb88df2bbcae86d9dbf5f84f1f19b0fce7a1c11f9f3bb967f09b9232a1726b.jpg) +(c) +Figure 3. Example images from our synthetic training data. a) Input image with background. b-d) Three renders (out of 10) from which the foreground image and alpha mask are used for supervision. Note the presence of glasses and different hair styles. + +![](images/41d2c40e434c4af9a1942439041db71016a58fccfdb7e80e416bc08fe670a330.jpg) +(d) + +distances to the subject (ranging from $40\mathrm{cm}$ to $1\mathrm{m}$ ). The input camera should emulate a webcam, so we position faces randomly within the image, ensuring that the angle of the face to the camera is within $30^{\circ}$ , and render at $1080 \times 720$ resolution. The images used for the supervision do not need to reflect the variability in the input data. We therefore render $512 \times 512$ pixel images from ten cameras closer to the face, pointing directly at the centre of the face, and whose positions are distributed randomly on a $45^{\circ}$ spherical cap centred around the input camera direction. Samples of our synthetic training data are shown in Fig. 3. + +# 2.3. Training loop + +The set of operations done on a single multi-view image sample in a batch during the training can be summarized as follows: (1) Using an ROI around the face based on the ground-truth 2D landmarks, define the rotated camera and warp the image using the homography as described in Sec. 2.1. Provide the warped image and ROI camera parameters as input to the DNN. (2) Predict the splatter image in the ROI camera coordinate system. (3) Render the splatter image Gaussian depth map into the input view using alpha composition to get differentiable rendering and gradients. Compute a global scale correction transform centred on the camera optical centre applied to the Gaussians by comparing the depth in the overlapping region of the ground truth mask and the rendered depth mask. The scale factor is obtained through linear least squares. (4) Apply the scale and transform the Gaussians back to the original camera coordinate system, render the corrected Gaussians in the virtual views and compare to ground truth images. The process is visualized in the top of Fig. 4. + +During training we minimize the following loss function + +$$ +L = L _ {d} + \lambda_ {\sigma} L _ {\sigma} + \lambda_ {m} L _ {m} + \lambda_ {c} L _ {c} + \lambda_ {j} L _ {j} \tag {1} +$$ + +with data term $L_{d}$ , regularizers $L_{\sigma}$ , $L_{m}$ and $L_{c}$ and stability term $L_{j}$ and corresponding weights defined below. + +Following Szymanowicz et al. [32] we define $L_{d}$ as: + +$$ +L _ {d} = \frac {1}{N _ {v} N _ {s}} \sum_ {v = 1} ^ {N _ {v}} \sum_ {s = 1} ^ {N _ {s}} L _ {i} \left(R \left(U \left(I _ {1} ^ {s}\right), C _ {v}\right), I _ {v} ^ {s}\right) \tag {2} +$$ + +![](images/6081e3b0e47bb6c8d1a38d8d528f054f49f41a8749de30a569a81bee8497186a.jpg) +Figure 4. Overview of our the training (top) and inference (bottom) pipelines of our method. + +where $L_{i}$ is an image pair loss, $U$ is the U-Net model and $R$ is the differentiable Gaussian splat rendering function [10]. $N_{v} = 11$ is the number of views used in each sample for the supervision and $N_{s}$ is the number of multi-view samples in the dataset. $C_{v}$ are the camera parameters for view $v$ and $I_{v}^{s}$ is the ROI image for view $v$ in sample $s$ . $v = 1$ is taken as the input view. The image comparison loss, $L_{i}(X,Y) = L_{e}(X,Y) + \lambda_{p}L_{p}(X,Y)$ , is composed of two terms: a Euclidean RGB distance term $L_{e}$ and a perceptual loss term $L_{p}$ with $\lambda_{p} = 0.5$ . We composite the rendered image and the ground truth RGBA image over a random background colour for each sample. + +This encourages the network to perform alpha matting by estimating the Gaussian transparency and subtracting the background colour from the predicted Gaussian colours. + +Euclidean RGB distance. We compute $L_{e}$ as the mean Euclidean distance in the RGB space, i.e., $L_{e}(X,Y) = \frac{1}{HW}\sum_{i=1}^{H}\sum_{j=1}^{W}\left\| X_{i,j} - Y_{i,j}\right\|_{2}$ . This loss differs from the classical mean of absolute differences by being invariant to rotations in the 3D RGB colour space which helps reduce chromaticity aberration in the reconstruction, and is more tolerant to outliers than $L_{2}$ loss which helps to not overpenalize small silhouette misalignments. + +Perceptual loss. We compute the perceptual loss $L_{p}$ using the VGG network [29] at two resolutions: full $(512 \times 512)$ and half $(256 \times 256)$ . We empirically found that the perceptual loss at half resolution helps to obtain better results qualitatively and quantitatively. + +Regularization Losses. We extend the original Splatter Image [32] method by predicting multiple Gaussians per pixel. With this modification we often encounter local minima during training where only one 'layer' of Gaussians is used, even though this does not provide the best results. We there + +fore introduce losses on the splatter image opacities to avoid falling into poor local minima by penalizing cases where the average opacity is near 0 in any of the layers. Denoting $\sigma_{i,j,k}$ the opacity of the $k^{\mathrm{th}}$ Gaussian represented at coordinate $(i,j)$ , we compute mean opacities for Gaussians layer $k$ as $\bar{\sigma}_k = \frac{1}{HW}\sum_{i,j}\sigma_{i,j,k}$ . We penalize the means $\bar{\sigma}_k$ near zero using a loss $L_{m} = \frac{1}{K}\sum_{k}e^{-\tau \bar{\sigma}_{k}}$ with weight $\lambda_{m} = 5$ and $\tau = 50$ , where $K$ is the number of Gaussians per pixel. We also encourage the prediction of opaque Gaussians by using an opacity loss $L_{\sigma} = \frac{1}{K}\sum_{k}(1 - \bar{\sigma}_{k})$ with weight $\lambda_{\sigma} = 10^{-4}$ . This term acts as a small bias and helps in making the best use of the allowed number of Gaussians. Finally, we add a regularization term $L_{c} = \log (s)^{2}$ on the scale correction, $s$ , obtained in the third step of the training loop to encourage the network to predict valid metric depths, with weight $\lambda_{c} = 1$ . This helps significantly with convergence during training. + +Jitter Loss. Temporal stability is important in the context of videoconferencing; flicker can be very distracting and degrade the experience. In order to improve the temporal stability of our prediction we use an approach similar to Zheng et al. [42]. Each sample, $s$ , in a training batch is duplicated and random perturbations added to the face box, independently for each sample, to form sample $s'$ . We then encourage stability by penalizing the difference in the renders between each pair of duplicated samples, $L_{j} = \frac{1}{N_{v}N_{s}}\sum_{v=1}^{N_{v}}\sum_{s=1}^{N_{s}}\|R(U(I_{1}^{s}),C_{v}) - R(U(I_{1}^{s'},C_{v})\|^{2}$ . This encourages the DNN to favour consistency in the prediction over accuracy, and we can control the trade-off with weight $\lambda_{j} = 1.0$ . + +![](images/46d8936bf25e37c2f6369865a87e14a4395d5ab42e01d15d177a8f67d328ec37.jpg) +Figure 5. Snapshot from our prototype 3D video conferencing application. See supplementary video for details. + +# 3. 3D Video Calls on Commodity Hardware + +This section describes the live 3D video conferencing prototype application we use to demonstrate our method. Our application exploits the 'magic window' effect [14] to create the illusion of 3D on a 2D display. We have two subjects, A and B, both equipped with a webcam, a 2D computer screen and a GPU. Given the symmetry of the setup, we describe only what happens on subject A's machine, the bottom of Fig. 4 provides a graphical representation. The camera-screen relative pose is assumed known through calibration. + +We render the appropriate new view of B's face on A's screen through the steps: (1) Detect A's face using an SSD-based detector [17] to get a bounding box, smooth it temporally to reduce jitter, and estimate the position of A's head assuming a fixed distance to the camera. (2) Detect B's face in the video feed received from B, and extract the ROI around B's face using a homography, as described in Sec. 2.1. (3) Predict a Gaussian splatter image in the ROI camera using the model described in Sec. 2.3. (4) Transform the Gaussians back to the original input camera coordinate system for B's video and render the 3D reconstruction of B in the view of subject A based on their estimated head position. + +We run the prototype on a laptop with a NVIDIA 4090 Mobile GPU and video sent over a standard network connection. The application runs at 28FPS, a snapshot of the prototype in action is shown in Fig. 5, further demonstration is provided in the supplementary video. + +# 4. Experiments and Results + +To validate our approach, we compare our method to recent works across a number of benchmarks, for ablation experiments please consult the supplementary material. Note that we limit comparison to reconstruction methods which satisfy our requirement of authenticity, i.e., those which do not rely on an enrolled representation. + +![](images/dcf6b06c263c884c90e2b494da68453208bc817786ea8c578c2988e9ad272250.jpg) +Figure 6. Qualitative comparison for images from the Cafca dataset [2]. Splatter Image (SI) has been retrained on our data. + +![](images/851557c67218d6eae361b456e0cee87c1927f9dbfe0de56dc4cdfb1755e477f2.jpg) +Figure 7. Qualitative comparison for images from the Ava-256 dataset [20]. Splatter Image (SI) has been retrained on our data. + +![](images/fcebb04609380f1215d38519fa01c3cfaca52ba4cca75b2e875aa36da6f74038.jpg) +Figure 8. Qualitative comparison for images from the FFHQ dataset [9] showing reconstructed input images, and visualized geometry. + +# 4.1. Experimental Setup + +We use three datasets for our evaluation. Firstly, the Cafca dataset [2] which contains multi-view renders of synthetic humans with variation in identity, expression, clothing and environment. This dataset provides a diverse range of settings which emulate an in-the-wild context, but with limitations on visual quality and realism. Secondly, the Ava-256 dataset [20] contains multi-camera captures of people in a studio with controlled clothing and lighting. This dataset + +![](images/ee3a8d9b9fc61cf2a88a6e836f2e9420e4a2d87630ac6f5e10a26fe57f195a3f.jpg) +Figure 9. Qualitative comparison for failure case of Live 3D Portrait (L3DP) [35]; our method is not constrained by EG3D [4], so can authentically represent all details from the input image, and can also create disconnected 3D structures like glasses frames. + +provides less diversity than Cafca, but is not synthetic and so provides far higher realism. Finally, the FFHQ dataset [9] includes frontal, in-the-wild photos of a variety of people. + +For Ava-256 and Cafca we follow the protocol of Lyu et al. [19] to evaluate novel view rendering (see supplementary material for details). We use PSNR, SSIM, LPIPS and DreamSim metrics to evaluate accuracy of image reconstruction, and ArcFace to assess identity preservation in novel views. For FFHQ we follow the protocol of Trevithick et al. [35] to evaluate input view reconstruction for a random subset of 500 images. We use LPIPS, DISTS and SSIM, but omit pose and identity metrics due to insufficient details to enable re-implementation. + +# 4.2. Results + +Results for novel view rendering on the Cafca dataset [2] are provided in Tab. 1. Our method outperforms others both quantitatively and qualitatively, and the results are of high quality. Compared to FaceLift [19] our results more accurately preserve the identity of the individual, although the visual quality of our results can degrade more significantly for extreme side views, as shown in Fig. 6. Overall, this produces numerically superior results, and is acceptable for our application where the viewing angle is limited. Results for novel view rendering on the Ava-256 dataset [20] are provided in Tab. 2. Results show a similar trend to those for Cafca, with superior identity preservation resulting in better overall metrics. See Fig. 7 for qualitative results. Qualitative results for in-the-wild images are shown in Figs. 1 and 10 and the supplementary material. Our method performs well across a diverse range of individuals and environments. + +Results for input view reconstruction on the FFHQ dataset [9] are provided in Tab. 3. Our method does not perform as well as Live 3D Portrait [35] for DISTS, but is state-of-the-art for LPIPS and SSIM. Qualitative results shown in Fig. 8 demonstrate that the perceived difference in quality is minimal, and that our geometry is smoother and more accurate. Fig. 9 also shows some failure cases of Live 3D Portrait [35] which our method is able to address, such as fine details which are not in the EG3D [4] latent space and disconnected structures such as glasses frames which their approach is not able to represent. Note that our method + +Table 1. Numerical results on Cafca [2] subset, following FaceLift protocol [19]. Splatter Image [32] is retrained on our data. + +
MethodPSNR ↑SSIM ↑LPIPS ↓DreamSim ↓ArcFace ↓
Era3D [15]13.690.72300.36620.28920.2978
TriPlaneNet [1]15.820.77050.53030.13390.2826
LGM [33]16.520.79330.30600.15520.2557
FaceLift [19]16.610.79680.26940.10960.1573
Splatter Image [32]20.040.83900.22440.02760.1202
Ours20.390.84620.18680.02860.1129
+ +Table 2. Numerical results on Ava-256 [20] subset, following FaceLift protocol [19]. Splatter Image [32] is retrained on our data. + +
MethodPSNR ↑SSIM ↑LPIPS ↓DreamSim ↓ArcFace ↓
Era3D [15]14.770.79630.25380.25150.3721
TriPlaneNet [1]14.320.68460.36590.16920.2795
LGM [33]14.050.81360.24760.14960.3142
FaceLift [19]16.520.82710.22770.10650.1871
Splatter Image [32]18.490.82580.25930.07250.1458
Ours19.110.82980.19080.07150.1216
+ +Table 3. Quantitative evaluation using LPIPS, DISTS, SSIM on 500 FFHQ images [9]. Evaluated only using the foreground on $256 \times 256$ images. * Use a different subset of 500 images; we find the deviation between different subsets to be negligible. + +
LPIPS↓DISTS↓SSIM↑
ROME* [11].1158.1058.8257
Live 3D Portrait* [35].0468.0407.8981
TriPlaneNet [1].2109.1580.7075
Splatter Image [32].0876.0916.9165
Ours.0413.0649.9456
+ +Table 4. Models trained for 100 epochs with and without stability loss $L_{j}$ with metrics computed with 250 subjects in the Ava256 dataset (more detail in the supplementary material), and our method compared with the Splatter Image method trained for 100 epochs. + +
MethodPSNR ↑SSIM ↑LPIPS ↓Jitter ↓
Splatter Image18.2240.86510.23000.006353
Ours no Lj18.949086580.15810.005020
Ours19.0690.86820.14970.004967
+ +is also able to seamlessly fill the 'holes' behind occluding objects such as glasses frames. + +Stability is an important attribute for our method. We devise a jitter metric to compare our method to Splatter Image [32]. We use images from 250 subjects in the Ava256 dataset with 10 views and 2 frames per subject. We align the reconstruction and ground truth in the first frame as described by Lyu et al. [19] and keep the same 2D alignment transform for the second frame to better measure jitter. The jitter metric for view $v$ at frame $t$ is defined as $||(I_t^v - I_{t-1}^v) - (R_t^v - R_{t-1}^v)||$ . We average this over the views and subjects. The quantitative results in Tab. 4 demonstrate that the jitter loss, $L_j$ , contributes to improved stability. + +![](images/ff8e36537e5e93095b39c153e9025651429b58a8fc4974158bd7dbd9915b8828.jpg) +Input frame +3D reconstruction rendered from -40, -20, 0, +20, and +40 degrees +Figure 10. Qualitative results of our method. Our approach is able to accurately reconstruct headwear, glasses and other accessories, as well as capturing fine details of hair and skin. + +![](images/b2a82d5620bdade993965b776064b9e879bb753f000cc742439d8d3758da1072.jpg) +Input frame +3D reconstruction rendered from -40, -20, 0, +20, and +40 degrees +Figure 11. Limitations of our method include: unusual texture on the face, extreme input angles, hands/props, segmentation failures. We suspect these are due to lack of representation in the training data; future work will aim to improve the training data to address these issues. + +# 4.3. Limitations + +Our method has a number of limitations which largely derive from the training data. Hands, held objects and face paint are uncommon in the training data and the range of angles of the face and novel view cameras, relative to the input camera, are limited. As shown in Fig. 11, face-paint can result in hallucination of glasses frame structures. Extreme input angles can cause blurry/visually degraded predictions, although this can be desirable in our scenario to make clear that the prediction in this area is not based on directly observed data. Reconstruction quality of hands and held objects can be poor when viewed in novel views, and segmentation failures can also occur. Future work will aim to address these failures through improvements to the training data. + +# 5. Potential societal impact + +This work aims to bring the experience of remote communication closer to in-person communication, reducing geographic inequality. By relying only on commodity hardware, we avoid exacerbating inequality in technology access. Our method cannot be used to impersonate someone else in an online meeting, due to our requirement for authenticity, supporting trust in online communication and media. This work also supports authentic representation regardless of factors such as hair colour and type, skin tone, and choice of headwear, aiding diverse self-expression (see Fig. 10). + +Like all human-centric computer vision, our 3D recon + +struction may have lower accuracy for some demographic groups. We find that our use of synthetic data is helpful in addressing any lack of fairness we discover, given our precise control over the training data distribution. Nevertheless, there are aspects of human diversity that are not yet represented by our synthetic datasets (see Fig. 11), and there may also be lack of fairness that we have not yet discovered. + +# 6. Conclusion + +Online communication is growing ever more important, and yet 2D video meetings have challenges that are not solvable within the 2D paradigm. Recently there has been considerable interest in developing photorealistic avatars that can be controlled and animated from a broad range of signals. However, we argue that 3D video, with an emphasis on authentic, realistic, live, and stable reconstructions, is needed to support online meetings that bring us closer to in-person interactions, while preserving the properties that make video-conferencing so widely adopted today. + +In this paper, we presented a method that fulfils the core properties required for authentic 3D video. There are many opportunities for improvement, and we hope to inspire future work that builds on these results. With this paper, we hope to highlight an open challenge for the community to demonstrate live 3D reconstruction from monocular inputs that can serve as a true analogue to 2D video, supporting the same diversity of expression that 2D video does today. + +# References + +[1] Ananta R. Bhattacharai, Matthias Nießner, and Artem Sevastopolsky. TriPlaneNet: An encoder for EG3D inversion. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3055-3065, 2024. 2, 3, 6, 7 +[2] Marcel C. Buehler, Gengyan Li, Erroll Wood, Leonhard Helminger, Xu Chen, Tanmay Shah, Daoye Wang, Stephan Garbin, Sergio Orts-Escalano, Otmar Hilliges, Dmitry Lagun, Jérémy Riviere, Paulo Gotardo, Thabo Beeler, Abhimitra Meka, and Kripasindhu Sarkar. Cafca: High-quality novel view synthesis of expressive faces from casual few-shot captures. In SIGGRAPH Asia 2024 Conference Papers, New York, NY, USA, 2024. Association for Computing Machinery. 2, 3, 6, 7 +[3] Chen Cao, Tomas Simon, Jin Kyu Kim, Gabe Schwartz, Michael Zollhoefer, Shun-Suke Saito, Stephen Lombardi, Shih-En Wei, Danielle Belko, Shouu-I Yu, Yaser Sheikh, and Jason Saragih. Authentic volumetric avatars from a phone scan. ACM Trans. Graph., 41(4), 2022. 2 +[4] Eric R. Chan, Connor Z. Lin, Matthew A. Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas Guibas, Jonathan Tremblay, Sameh Khamis, Tero Karras, and Gordon Wetzstein. Efficient geometry-aware 3D generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16123-16133, 2022. 2, 7 +[5] Yufan Chen, Lizhen Wang, Qijing Li, Hongjiang Xiao, Shengping Zhang, Hongxun Yao, and Yebin Liu. MonoGaussian-Avatar: Monocular Gaussian point-based head avatar. In ACM SIGGRAPH 2024 Conference Papers, New York, NY, USA, 2024. Association for Computing Machinery. 2 +[6] Stephan J. Garbin, Marek Kowalski, Virginia Estellers, Stanislaw Szymanowicz, Shideh RezaEIFar, Jingjing Shen, Matthew A. Johnson, and Julien Valentin. VolTeMorph: Realtime, controllable and generalizable animation of volumetric representations. Computer Graphics Forum, 43(6):e15117, 2024. 2 +[7] Simon Giebenhain, Tobias Kirschstein, Markos Georgopoulos, Martin Rünz, Lourdes Agapito, and Matthias Nießner. Learning neural parametric head models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2 +[8] Charlie Hewitt, Fatemeh Saleh, Sadegh Aliakbarian, Lohit Petikam, Shideh RezaEIFar, Louis Florentin, Zafiirah Hosenie, Thomas J Cashman, Julien Valentin, Darren Cosker, and Tadas Baltrusaitis. Look Ma, no markers: holistic performance capture without the hassle. ACM Transactions on Graphics (TOG), 43(6):#235:1-12, 2024. 4 +[9] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 6, 7 +[10] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3D Gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42(4), 2023. 2, 5 + +[11] Taras Khakhulin, Vanessa Sklyarova, Victor Lempitsky, and Egor Zakharov. Realistic one-shot mesh-based head avatars. In European Conference of Computer vision (ECCV), 2022. 7 +[12] Tobias Kirschstein, Javier Romero, Artem Sevastopolsky, Matthias Nießner, and Shunsuke Saito. Avat3r: Large animatable Gaussian reconstruction model for high-fidelity 3D head avatars, 2025. 2 +[13] Jason Lawrence, Dan B Goldman, Supreeth Achar, Gregory Major Blascovich, Joseph G. Desloge, Tommy Fortes, Eric M. Gomez, Sascha Haberling, Hugues Hoppe, Andy Huibers, Claude Knaus, Brian Kuschak, Ricardo MartinBrualla, Harris Nover, Andrew Ian Russell, Steven M. Seitz, and Kevin Tong. Project Starline: A high-fidelity telepresence system. ACM Transactions on Graphics (Proc. SIGGRAPH Asia), 40(6), 2021. 1, 2 +[14] Johnny Chung Lee. Head tracking for desktop virtual reality displays using the Wii remote, 2007. http://youtu.be/Jd3-eiid-Uw.6 +[15] Peng Li, Yuan Liu, Xiaoxiao Long, Feihu Zhang, Cheng Lin, Mengfei Li, Xingqun Qi, Shanghang Zhang, Wei Xue, Wenhan Luo, Ping Tan, Wenping Wang, Qifeng Liu, and Yike Guo. Era3D: High-resolution multiview diffusion using efficient row-wise attention. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 7 +[16] Zhihao Li, Jianzhuang Liu, Zhensong Zhang, Songcen Xu, and Youliang Yan. CLIFF: Carrying location information in full frames into human pose and shape estimation. In Computer Vision - ECCV 2022, pages 590-606, Cham, 2022. Springer Nature Switzerland. 4 +[17] Wei Liu, Dragomir Anguelov, Dumitru Erhan, Christian Szegedy, Scott Reed, Cheng-Yang Fu, and Alexander C Berg. SSD: Single shot multibox detector. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part I 14, pages 21-37. Springer, 2016. 6 +[18] Stephen Lombardi, Tomas Simon, Gabriel Schwartz, Michael Zollhoefer, Yaser Sheikh, and Jason Saragih. Mixture of volumetric primitives for efficient neural rendering. ACM Trans. Graph., 40(4), 2021. 2 +[19] Weijie Lyu, Yi Zhou, Ming-Hsuan Yang, and Zhixin Shu. FaceLift: Single image to 3D head with view generation and gs-lrm. arXiv preprint arXiv:2412.17812, 2024. 2, 6, 7 +[20] Julieta Martinez, Emily Kim, Javier Romero, Timur Bagautdinov, Shunsuke Saito, Shoou-I Yu, Stuart Anderson, Michael Zollhöfer, Te-Li Wang, Shaojie Bai, Chenghui Li, Shih-En Wei, Rohan Joshi, Wyatt Borsos, Tomas Simon, Jason Saragih, Paul Theodosis, Alexander Greene, Anjani Josyula, Silvio Mano Maeta, Andrew I. Jewett, Simon Venshtain, Christopher Heilman, Yueh-Tung Chen, Sidi Fu, Mohamed Ezzeldin A. Elshaer, Tingfang Du, Longhua Wu, Shen-Chi Chen, Kai Kang, Michael Wu, Youssef Emad, Steven Longay, Ashley Brewer, Hitesh Shah, James Booth, Taylor Koska, Kayla Haidle, Matt Andromalos, Joanna Hsu, Thomas Dauer, Peter Selednik, Tim Godisart, Scott Ardisson, Matthew Cipperly, Ben Humberston, Lon Farr, Bob Hansen, Peihong Guo, Dave Braun, Steven Krenn, He Wen, Lucas Evans, Natalia + +Fadeeva, Matthew Stewart, Gabriel Schwartz, Divam Gupta, Gyeongsik Moon, Kaiwen Guo, Yuan Dong, Yichen Xu, Takaaki Shiratori, Fabian Prada, Bernardo R. Pires, Bo Peng, Julia Buffalini, Autumn Trimble, Kevyn McPhail, Melissa Schoeller, and Yaser Sheikh. Codec Avatar Studio: Paired Human Captures for Complete, Driveable, and Generalizable Avatars. NeurIPS Track on Datasets and Benchmarks, 2024. 6, 7 +[21] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing scenes as neural radiance fields for view synthesis. In The European Conference on Computer Vision (ECCV), 2020. 2 +[22] Gary M. Olson and Judith S. Olson. Distance matters. Hum.-Comput. Interact., 15(2):139-178, 2000. 1 +[23] Shenhan Qian, Tobias Kirschstein, Liam Schoneveld, Davide Davoli, Simon Giebenhain, and Matthias Nießner. Gaussian-Avatars: Photorealistic head avatars with rigged 3D Gaussians. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20299-20309, 2024. 2 +[24] Alexander Raake, Markus Fiedler, Katrin Schoenenberg, Katrien De Moor, and Nicola Döring. Technological factors influencing videoconferencing and Zoom fatigue. arXiv:2202.01740 [cs], 2022. 1 +[25] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-Net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234-241. Springer, 2015. 3 +[26] Shunsuke Saito, Gabriel Schwartz, Tomas Simon, Junxuan Li, and Giljoo Nam. Relightable Gaussian codec avatars. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 130-141, 2024. 1, 2 +[27] Jack Saunders, Charlie Hewitt, Yanan Jian, Marek Kowalski, Tadas Baltrusaitis, Yiye Chen, Darren Cosker, Virginia Estellers, Nicholas Gydé, Vinay P Namboodiri, et al. Gasp: Gaussian avatars with synthetic priors. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 271-280, 2025. 2, 3 +[28] Martin Schuessler, Luca Hormann, Raimund Dachselt, Andrew Blake, and Carsten Rother. Gazing heads: Investigating gaze perception in video-mediated communication. ACM Trans. Comput.-Hum. Interact., 31(3), 2024. 1 +[29] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. CoRR, abs/1409.1556, 2014. 5 +[30] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations, 2021. 2, 3 +[31] Stanislaw Szymanowicz, Eldar Insafutdinov, Chuanxia Zheng, Dylan Campbell, Joao F Henriques, Christian Rupprecht, and Andrea Vedaldi. Flash3D: Feed-forward generalisable 3D scene reconstruction from a single image. arXiv preprint arXiv:2406.04343, 2024. 3 +[32] Stanislaw Szymanowicz, Christian Rupprecht, and Andrea Vedaldi. Splatter image: Ultra-fast single-view 3D reconstruc + +tion. In The IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2, 3, 4, 5, 6, 7 +[33] Jiaxiang Tang, Zhaoxi Chen, Xiaokang Chen, Tengfei Wang, Gang Zeng, and Ziwei Liu. LGM: Large multi-view gaussian model for high-resolution 3d content creation. In European Conference on Computer Vision, pages 1-18. Springer, 2024. 7 +[34] John C. Tang, Kori Inkpen, Sasa Junuzovic, Keri Mallari, Andrew D. Wilson, Sean Rintel, Shiraz Cupala, Tony Carbary, Abigail Sellen, and William A.S. Buxton. Perspectives: Creating inclusive and equitable hybrid meeting experiences. Proc. ACM Hum.-Comput. Interact., 7(CSCW2), 2023. 1 +[35] Alex Trevithick, Matthew Chan, Michael Stengel, Eric R. Chan, Chao Liu, Zhiding Yu, Sameh Khamis, Manmohan Chandraker, Ravi Ramamoorthi, and Koki Nagano. Real-time radiance fields for single-image portrait view synthesis. In ACM Transactions on Graphics (SIGGRAPH), 2023. 2, 3, 6, 7 +[36] Tengfei Wang, Bo Zhang, Ting Zhang, Shuyang Gu, Jianmin Bao, Tadas Baltrusaitis, Jingjing Shen, Dong Chen, Fang Wen, Qifeng Chen, and Baining Guo. Rodin: A generative model for sculpting 3D digital avatars using diffusion. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4563-4573, 2023. 3 +[37] Jun Xiang, Xuan Gao, Yudong Guo, and Juyong Zhang. FlashAvatar: High-fidelity head avatar with efficient Gaussian embedding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 2 +[38] Dejia Xu, Ye Yuan, Morteza Mardani, Sifei Liu, Jiaming Song, Zhangyang Wang, and Arash Vahdat. AGG: Amortized generative 3D Gaussians for single image to 3D. Transactions on Machine Learning Research, 2024. 2 +[39] Yuelang Xu, Benwang Chen, Zhe Li, Hongwen Zhang, Lizhen Wang, Zerong Zheng, and Yebin Liu. Gaussian head avatar: Ultra high-fidelity head avatar via dynamic Gaussians. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2 +[40] Yuelang Xu, Lizhen Wang, Zerong Zheng, Zhaoqi Su, and Yebin Liu. 3D Gaussian parametric head model. In Proceedings of the European Conference on Computer Vision (ECCV), 2024. 2 +[41] Kai Zhang, Sai Bi, Hao Tan, Yuanbo Xiangli, Nanxuan Zhao, Kalyan Sunkavalli, and Zexiang Xu. GS-LRM: Large reconstruction model for 3D Gaussian splatting. European Conference on Computer Vision, 2024. 2 +[42] Stephan Zheng, Yang Song, Thomas Leung, and Ian J. Goodfellow. Improving the robustness of deep neural networks via stability training. CoRR, abs/1604.04326, 2016. 5 +[43] Wojciech Zielonka, Timo Bolkart, and Justus Thies. Instant volumetric head avatars. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4574-4584, 2022. 2 +[44] Zi-Xin Zou, Zhipeng Yu, Yuan-Chen Guo, Yangguang Li, Ding Liang, Yan-Pei Cao, and Song-Hai Zhang. Triplane meets gaussian splatting: Fast and generalizable single-view 3d reconstruction with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10324–10335, 2024. 2 \ No newline at end of file diff --git a/ICCV/2025/VoluMe - Authentic 3D Video Calls from Live Gaussian Splat Prediction/images.zip b/ICCV/2025/VoluMe - Authentic 3D Video Calls from Live Gaussian Splat Prediction/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..0252bd418a1aea9108579b4f85d751f49c8ae093 --- /dev/null +++ b/ICCV/2025/VoluMe - Authentic 3D Video Calls from Live Gaussian Splat Prediction/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7b4c67e5793de30ad6d6cd9404d28b1a36cfdf94077946dc3478570319d2e897 +size 563743 diff --git a/ICCV/2025/VoluMe - Authentic 3D Video Calls from Live Gaussian Splat Prediction/layout.json b/ICCV/2025/VoluMe - Authentic 3D Video Calls from Live Gaussian Splat Prediction/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..819bf10fb7802b41f64837b5c30d801dc225e413 --- /dev/null +++ b/ICCV/2025/VoluMe - Authentic 3D Video Calls from Live Gaussian Splat Prediction/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fe6c3e180b5b90a8363ccd30b98563151cc6f1f25d45e85627ce37e0c78ddfdb +size 377036 diff --git a/ICCV/2025/VolumetricSMPL_ A Neural Volumetric Body Model for Efficient Interactions, Contacts, and Collisions/bb7a2147-d88a-4bc3-b1d6-b526e6f5b798_content_list.json b/ICCV/2025/VolumetricSMPL_ A Neural Volumetric Body Model for Efficient Interactions, Contacts, and Collisions/bb7a2147-d88a-4bc3-b1d6-b526e6f5b798_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..bf6c54367ebc29f4bfde21c4740a203983179b71 --- /dev/null +++ b/ICCV/2025/VolumetricSMPL_ A Neural Volumetric Body Model for Efficient Interactions, Contacts, and Collisions/bb7a2147-d88a-4bc3-b1d6-b526e6f5b798_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:74c92319449edad27e91c3abeaf96b8f81e273724b0d0d942b3f637e5a0d2ab8 +size 86216 diff --git a/ICCV/2025/VolumetricSMPL_ A Neural Volumetric Body Model for Efficient Interactions, Contacts, and Collisions/bb7a2147-d88a-4bc3-b1d6-b526e6f5b798_model.json b/ICCV/2025/VolumetricSMPL_ A Neural Volumetric Body Model for Efficient Interactions, Contacts, and Collisions/bb7a2147-d88a-4bc3-b1d6-b526e6f5b798_model.json new file mode 100644 index 0000000000000000000000000000000000000000..fc8b20a4198de0134e8d0a98d2c47ecdf69e03f2 --- /dev/null +++ b/ICCV/2025/VolumetricSMPL_ A Neural Volumetric Body Model for Efficient Interactions, Contacts, and Collisions/bb7a2147-d88a-4bc3-b1d6-b526e6f5b798_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:85a87008269fda56c329f3a4b9be85e94fc274b693d8cda7a80e7f64a57f1b84 +size 111373 diff --git a/ICCV/2025/VolumetricSMPL_ A Neural Volumetric Body Model for Efficient Interactions, Contacts, and Collisions/bb7a2147-d88a-4bc3-b1d6-b526e6f5b798_origin.pdf b/ICCV/2025/VolumetricSMPL_ A Neural Volumetric Body Model for Efficient Interactions, Contacts, and Collisions/bb7a2147-d88a-4bc3-b1d6-b526e6f5b798_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..f012ea87d5f912a063c7b17aac46acc91c22234f --- /dev/null +++ b/ICCV/2025/VolumetricSMPL_ A Neural Volumetric Body Model for Efficient Interactions, Contacts, and Collisions/bb7a2147-d88a-4bc3-b1d6-b526e6f5b798_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9a59bfbc5227664b14611ff17d02f60f36d7cce51f0122704b8aeeeb08005a36 +size 9764444 diff --git a/ICCV/2025/VolumetricSMPL_ A Neural Volumetric Body Model for Efficient Interactions, Contacts, and Collisions/full.md b/ICCV/2025/VolumetricSMPL_ A Neural Volumetric Body Model for Efficient Interactions, Contacts, and Collisions/full.md new file mode 100644 index 0000000000000000000000000000000000000000..f18843318ba9dba2118c13ec6e4894e1d6912559 --- /dev/null +++ b/ICCV/2025/VolumetricSMPL_ A Neural Volumetric Body Model for Efficient Interactions, Contacts, and Collisions/full.md @@ -0,0 +1,352 @@ +# VolumetricSMPL: A Neural Volumetric Body Model for Efficient Interactions, Contacts, and Collisions + +Marko Mihajlovic1, Siwei Zhang1, Gen Li1, Kaifeng Zhao1, Lea Müller2, Siyu Tang1 +1ETH Zürich 2UC Berkeley + +markomih.github.io/VolumetricSMPL + +![](images/9225b9f9278be2f6506320bb9790ab0324ba84a4f7754101459abd7f2ddc2d91.jpg) +Figure 1. We introduce VolumetricSMPL, a volumetric body model that represents human shape as a Signed Distance Field (SDF), enabling efficient, differentiable, and more accurate interactions across multiple downstream tasks: (1) [Sec. 4.2.1]: Human-object interaction reconstruction from in-the-wild images, achieving $500 \times$ speedup over prior work [59, 68]. (2) [Sec. 4.2.2]: $3 \times$ faster and more accurate human mesh recovery in 3D scenes from egocentric images compared to [69]. (3) [Sec. 4.2.3]: $7 \times$ faster scene-constrained human motion synthesis [78] with $20 \times$ lower GPU memory, leading to more realistic and collision-free motion compared to imposing scene-constraints via prior work [42]. (4) [Sec. 4.2.4]: Efficient volumetric constraints for resolving mesh self-intersections, achieving $3 \times$ speedup, $3 \times$ lower GPU memory, and improved intersection handling over [42]. VolumetricSMPL provides a powerful solution for learning-based human interaction modeling with improved efficiency, scalability, and accuracy. + +# Abstract + +Parametric human body models play a crucial role in computer graphics and vision, enabling applications ranging from human motion analysis to understanding human-environment interactions. Traditionally, these models use surface meshes, which pose challenges in efficiently handling interactions with other geometric entities, such as objects and scenes, typically represented as meshes or point clouds. To address this limitation, recent research has explored volumetric neural implicit body models. However, existing works are either insufficiently robust for complex human articulations or impose high computational and memory costs, limiting their widespread use. To this end, we introduce VolumetricSMPL, a neural volumetric body model that leverages Neural Blend Weights (NBW) to generate compact, yet efficient MLP decoders. Unlike + +prior approaches that rely on large MLPs, NBW dynamically blends a small set of learned weight matrices using predicted shape- and pose-dependent coefficients, significantly improving computational efficiency while preserving expressiveness. VolumetricSMPL outperforms prior volumetric occupancy model COAP with $10 \times$ faster inference, $6 \times$ lower GPU memory usage, enhanced accuracy, and a Signed Distance Function (SDF) for efficient and differentiable contact modeling. We demonstrate VolumetricSMPL's strengths across four challenging tasks: (1) reconstructing human-object interactions from in-the-wild images, (2) recovering human meshes in 3D scenes from egocentric views, (3) scene-constrained motion synthesis, and (4) resolving self-intersections. Our results highlight its broad applicability and significant performance and efficiency gains. + +# 1. Introduction + +Learned parametric body models [38, 46, 49, 63] are essential tools in computer graphics and vision, supporting diverse applications such as avatar creation [32, 58, 62, 64], motion analysis [52, 55, 65, 72, 75], human pose and shape reconstruction [56], and even medical treatments [15, 20]. These models are typically constructed by fitting parametric meshes to large-scale body scan datasets, providing a compact yet expressive representation of body shape and pose-dependent deformations. + +Despite their efficiency in rendering and memory usage, mesh-based deformable body models face challenges in efficient intersection testing and differentiable contact modeling, particularly when interacting with objects, environments, and other humans. Existing methods address this limitation with application-specific strategies. Some preprocess 3D scenes into volumetric representations [18, 72, 73, 76], which is computationally expensive and error-prone. Others constrain interactions to synthetic environments where analytic distance fields can be computed [19, 78], limiting real-world applicability. Techniques modeling human-human interactions [12, 45] often substantially downsample meshes to enable traditional graphics-based methods, such as winding numbers [22], but these are non-differentiable and difficult to generalize to other applications that require differentiable collision loss terms [35, 59, 68, 69]. Overall, these solutions tend to be computationally costly, application-specific, or nondifferentiable, hindering their broader adoption. + +To overcome these challenges, volumetric human body models [2, 42, 43] have emerged as a unified solution that facilitates more natural and efficient interactions between the human body and point clouds or meshes. LEAP [43] represents the body as a volumetric function in canonical space, using a deformation network to map between observed and canonical coordinates. imGHUM [2] models the body directly in observation space using sub-part learnable neural implicit fields, though such a flexible model requires large collections of human scans for training. COAP [42] introduces compositional neural fields in the canonical space of individual body parts and explicitly models deformations via skeletal transformations, improving pose generalization. Due to its strong performance, COAP has been widely adopted in applications such as human-scene interaction, human-object interaction, and digital human synthesis [3, 10, 27, 33-35, 60, 69]. + +However, COAP's large MLP decoder creates a computational bottleneck, restricting its widespread usability in learning- and optimization-based applications. Querying COAP with a batch size of 5 and $15\mathrm{k}$ query points fully saturates a 24 GB GPU, making it impractical for integration with additional neural modules. Moreover, COAP represents the body as a step occupancy function, which leads to + +ill-defined gradients for points far from the iso-surface, restricting its effectiveness in tasks requiring smooth distance approximation such as scene-constrained motion synthesis demonstrated in Sec. 4.2.3. + +To overcome these limitations, we introduce VolumetricSMPL, a novel volumetric body model that not only offers improved accuracy over COAP but also significantly enhances computational efficiency. VolumetricSMPL achieves around $10 \times$ faster inference, requires $6 \times$ less GPU memory, and models the human body as a Signed Distance Field (SDF) instead of an occupancy function, ensuring smooth gradients for downstream tasks. + +At the core of VolumetricSMPL is the Neural Blend Weights (NBW) Generator, a neural network that dynamically predicts compact MLP weights for an SDF decoder. Unlike prior volumetric body models [2, 42, 43], which rely on large MLP decoders with 256-512 neurons, NBW generates lightweight 64-neuron MLPs by blending compact learned weight matrices with predicted shape- and pose-dependent coefficients. This design preserves expressiveness while drastically reducing computational cost, leading to faster inference, improved accuracy, and lower memory usage. + +VolumetricSMPL builds upon COAP's compositional encoder, partitioning the body into kinematic chain-based parts [42], each represented in its own local canonical space and encoded using a lightweight PointNet MLP [51]. The NBW module then generates specialized network weights for each local SDF decoder, effectively creating part-specific networks, akin to the mixture of experts paradigm [67]. At inference time, each local decoder predicts a signed distance for a given query point, and the results are aggregated using a minimum operation, ensuring a smooth, globally consistent SDF representation. To further boost efficiency, we introduce an analytic SDF for regions far from the body, significantly accelerating queries in non-critical areas while maintaining accuracy near the surface. + +We extensively evaluate VolumetricSMPL and demonstrate its superior performance across multiple metrics compared to existing models. Furthermore, we showcase its effectiveness in four key downstream applications: 1) reconstructing human-object interactions from in-the-wild images [59], 2) recovering human meshes in 3D scenes from egocentric views [69], 3) scene-constrained human motion synthesis, and 4) resolving self-intersections in human body models [42]. These applications highlight significant improvements in speed, accuracy, and convergence, enabled by our efficient SDF representation (see Fig. 1). + +In summary, our contributions are: + +- A novel parametric Signed Distance Field body model, achieving around $10 \times$ faster inference, $6 \times$ lower GPU memory usage, and higher accuracy compared to COAP [42]. + +- The introduction of Neural Blend Weights Generator that enables efficient SDF learning with tiny MLPs. +- The deployment of VolumetricSMPL, leading to substantial computational efficiency and accuracy gains across downstream applications, including human-object interaction reconstruction, human mesh recovery, scene-constrained human motion synthesis, and self-intersection resolution, demonstrating its practicality and versatility. See Fig. 1 as an overview. + +We provide a lightweight and easy-to-use add-on module for the widely used SMPL-based body models [38, 49], enabling seamless integration of VolumetricSMPL with a single line of code, without compromising the generalization of our method. The code and models will be released under the MIT license. + +# 2. Related Work + +Data-Driven Body Models are important components in computer vision and graphics [8, 38, 49, 50, 63]. Trained on large-scale human scan datasets, these models generate human shapes using low-dimensional representations, such as shape coefficients and pose parameters. Many parametric models employ linear shape spaces [36, 38, 49, 53], ensuring compatibility with graphics pipelines while maintaining high-quality shape representation. We develop our volumetric representation based on the SMPL body model [38, 49], inheriting its generative capabilities. This choice ensures seamless integration with prior methods operating within the SMPL parameter space [56], facilitating compatibility with existing shape estimation and reconstruction pipelines. + +Modeling Human Interactions faces challenges for mesh body models [4, 18, 70, 74, 76, 79, 80]. Many approaches rely on application-specific solutions. To model human-scene interactions, several methods [18, 23, 26, 33, 76] employ additional preprocessing steps to convert 3D scans into volumetric representations, enabling efficient collision detection and resolution. However, these techniques are often computationally expensive, memory-intensive, and prone to inaccuracies. To mitigate these limitations, some works [24, 39, 65, 66, 77] constrain interactions to synthetic environments or simplified geometric primitives, though this often comes at the cost of reduced realism. Similarly, techniques for modeling human-human interactions [12, 45] often aggressively downsample meshes to make traditional graphics-based methods, such as winding numbers [22], computationally feasible. However, these methods remain non-differentiable and challenging to generalize to other applications that require differentiable collision loss terms [35, 59, 68, 69]. + +Volumetric Body Models address the limitations of mesh models in human interaction tasks. While classical tetrahedral meshes [17] are used for collision detection, they lack direct support for animatable body models, as skinning + +is typically defined only on surface points and are rather used to model rigid body parts [28, 37]. A more flexible alternative is neural volumetric body models [2, 11, 42, 43, 47, 48, 57]. BodyNet [57] regresses a volumetric body directly from images, while other methods [11, 47, 48] require subject-specific optimization, making them unsuitable for feed-forward inference modeling of human bodies akin to SMPL. More closely related to our work, LEAP [43], COAP [42], and imGHUM [2] enable feed-forward volumetric body modeling from low-dimensional shape and pose parameters. These models integrate seamlessly into perception applications [3, 10, 27, 34, 60], facilitating direct interactions with scene and object representations. + +Feed-Forward Volumetric Body Models [2, 42, 43] enable efficient human body modeling using low-dimensional shape and pose parameters while facilitating differentiable interactions with point clouds and meshes. Among them, imGHUM [2] learns expressive implicit neural fields directly in observation space, providing flexibility but requiring large-scale human scan datasets for training, which limits its adaptability to new shape spaces. In contrast, LEAP [43] and its more robust successor, COAP [42], restrict the learning space to the shape space of parametric mesh models such as SMPL [38] and SMPL-X [49]. This design choice allows for seamless integration into downstream tasks that rely on these parametric models and require efficient body-environment interactions. Notably, COAP has already been widely adopted in applications such as human-scene and human-object interaction modeling [3, 10, 27, 33–35, 60, 69]. + +However, learning- and optimization-based applications require frequent queries to the volumetric body representation to model collisions, making computational efficiency a critical concern. We analyze existing volumetric body models and identify the primary bottleneck as the computational cost of large MLP decoders. For instance, both LEAP and COAP employ 256-neuron MLP decoders, while imGHUM relies on even larger 512-neuron architectures. + +Generating Network Weights from a conditioning vector has been explored via direct weight regression using HyperNetworks [16] and weight modulation approaches such as FiLM [7]. HyperNetworks have also been applied for subject-specific avatar modeling [58]. However, we find these approaches unsuitable for feed-forward modeling of compositional volumetric bodies, as each body part is only weakly supervised, leading to unstable training. + +Instead, we build on insights from ResFields [41], which compose learned network weights blended through estimable coefficients for fitting temporal signals. Here, we repurpose this design and regress pose- and shape-dependent blending coefficients, introducing our Neural Blend Weights (NBW) Generator to predict compact MLPs for efficient feed-forward volumetric body modeling. + +![](images/20cfe3b6ed164875f147e84a19998345f388d5ad6024a99ddb92dbba3f00c5c6.jpg) +Figure 2. Overview. VolumetricSMPL models the human body volume as a Signed Distance Field (SDF) based on a parametric mesh body model [38, 49], controlled via shape and pose parameters $(\beta, \theta)$ . First, the human mesh is sampled into a point cloud, where each point is assigned to one of $K$ body parts (colored regions). Each part is then canonicalized based on its respective bone transformation $G_{k}$ and over-approximated by a bounding box $B_{k}$ following [42]. Next, each body part is encoded into a compact feature vector $\mathbf{z}_{k}$ via a PointNet MLP [51], which is further processed by the Neural Blend Weights (NBW) generator (Sec. 3.3) that predicts weights $\mathcal{W}_k$ for the corresponding local SDF decoder. For a given query point $\mathbf{x} \in \mathbb{R}^3$ , we first transform it into its canonicalized form $\mathbf{x}_k$ (Eq. 2) and optionally apply positional encoding $\gamma(\cdot)$ [44]. The SDF prediction $\tilde{d}_k$ is then obtained by querying each local decoder, which takes as input both $\mathbf{x}_k$ and the corresponding shape feature $\mathbf{z}_k$ . The local predictions are then aggregated using a min operation, yielding $d_{\mathrm{implicit}}$ . If the query point is far from the body, outside of all boxes $\mathcal{B} = [B_k]_{k=1}^K$ placed via $\mathcal{G} = [G_k]_{k=1}^K$ , an analytic SDF estimate $d_{\mathrm{analytic}}$ is used for efficiency. The final signed distance prediction $\tilde{d}(\mathbf{x})$ is computed by fusing the implicit and analytic estimates. Gray blocks indicate trainable neural modules. + +# 3. VolumetricSMPL + +VolumetricSMPL is a volumetric body model $f_{\Theta}$ , parameterized by learned parameters $\Theta$ and conditioned on human shape $\beta \in \mathbb{R}^{|\beta|}$ and pose $\theta \in \mathbb{R}^{|\theta|}$ , following the structure of widely used parametric human models [38, 49, 63]. We denote $K$ body parts using rigid bone transformation matrices $\mathcal{G} = [G_k \in \mathbb{R}^{4 \times 4}]_{k=1}^K$ , which define the location and orientation of each body part. Formally, the volumetric body model is defined as a level-set function that maps a query point $\mathbf{x} \in \mathbb{R}^3$ to a signed distance value: + +$$ +\tilde {d} (\mathbf {x}) = f _ {\Theta} (\mathbf {x} \mid \beta , \theta) \in \mathbb {R}, \tag {1} +$$ + +where $\tilde{d} (\mathbf{x})$ represents the distance to the closest surface, with its sign indicating whether the point lies inside or outside the human body. An overview is provided in Fig. 2. + +# 3.1. Model Representation and Encoding + +VolumetricSMPL extends a parametric mesh body model (e.g., SMPL [38, 49]) by representing human shape as a Signed Distance Field (SDF), controlled by its shape and pose parameters $(\beta, \theta)$ . + +First, the human mesh is converted into a point cloud by uniformly sampling points proportional to mesh face areas. Each point is then assigned to one of $K$ body parts using a kinematic chain partitioning scheme [42]. Each body part is then transformed into a local canonical space using its respective bone transformation $G_{k}$ and over-approximated by a bounding box $B_{k}$ , constructed by extending the extremities of the part with a $12.5\%$ padding. To efficiently encode shape information, the point cloud of each body part is processed using a compact shared PointNet encoder [51]. This + +produces a local feature vector $\mathbf{z}_k$ , which captures shape and deformation information. This vector then conditions the Neural Blend Weights (NBW) module (Sec. 3.3), which generates adaptive network weights $\mathcal{W}_k$ for the $k$ th SDF decoder, enabling an efficient and flexible volumetric representation. + +# 3.2. Efficient SDF Querying + +To evaluate the SDF efficiently, we transform the query point $\mathbf{x}$ into the canonical space of each body part using a rigid transformation matrix that preserves signed distances: + +$$ +\mathbf {x} _ {k} = \left(G _ {k} ^ {- 1} \mathbf {x} ^ {h}\right) _ {1: 3}, \quad \text {w h e r e} \quad \mathbf {x} ^ {h} = [ \mathbf {x}, 1 ] \in \mathbb {R} ^ {4}. \tag {2} +$$ + +Each local SDF decoder, using the generated weights $\mathcal{W}_k$ from the NBW module and conditioned on $\mathbf{z}_k$ , predicts a signed distance $\tilde{d}_k$ for the transformed query point. The final neural implicit SDF estimate $d_{\mathrm{implicit}}$ is obtained by taking the minimum over all local predictions. + +To improve computational efficiency, we introduce an analytic SDF approximation $d_{\mathrm{analytic}}$ for query points far from the human body. Specifically, if $\mathbf{x}$ lies outside all bounding boxes $\mathcal{B} = [B_k]_{k=1}^K$ , we approximate its signed distance using a bounding box-based SDF: + +$$ +\tilde {d} (\mathbf {x}) = \left\{ \begin{array}{l l} d _ {\text {a n a l y t i c}} (\mathbf {x} \mid \mathcal {B}, \mathcal {G}), & \text {i f} \mathbf {x} \notin B _ {k} \text {f o r a l l} k, \\ d _ {\text {i m p l i c i t}} (\mathbf {x} \mid \beta , \theta), & \text {o t h e r w i s e .} \end{array} \right. \tag {3} +$$ + +The analytic SDF computes the Euclidean distance to the closest bounding box surface, enabling a fast, memory-efficient approximation for distant points. This hybrid approach offers two key advantages: (i) Improved scalability + +![](images/3c9ce29a9f4ab84c51e5e9b2bd43c15466e56e5ac15b49857d4188522d7c71fb.jpg) +Figure 3. Neural Blend Weights (NBW) Generator takes a local latent code $\mathbf{z}_k$ as input and outputs layer-specific weights $\mathcal{W}_k$ for the $k$ th SDF decoder. First, $\mathbf{z}_k$ is transformed into a set of blending coefficients $\{\mathbf{v}_k^l\}_{l=1}^L$ using $L$ independent learnable linear layers. These coefficients modulate the shape weights $\mathbf{W}_k^l$ , which are then added to the base network weights $\mathbf{W}^l$ . The resulting neural blend weights $\mathcal{W}_k = \{\mathcal{W}_k^l\}_{l=1}^L$ are then used in the MLP-based SDF decoder. For brevity, bias terms are omitted. Gray blocks indicate trainable modules. + +and accuracy, as relying solely on neural models is computationally expensive and prone to inaccuracies in distant regions; and (ii) Faster querying, as the analytic SDF efficiently approximates distances, it reduces neural computation in non-critical areas. + +# 3.3. Neural Blend Weights Generator + +Following the success of neural fields [61] in representing continuous SDF functions via MLP architectures, we implement an MLP-based SDF decoder. This decoder maps a query point $\mathbf{x}_k$ (optionally positionally encoded as $\gamma (\mathbf{x}_k)$ [44]) and a conditioning latent code $\mathbf{z}_k$ to an SDF value. Formally, our $k$ -th $L$ -layer SDF decoder is defined as: + +$$ +\tilde {d} _ {k} = \sigma_ {L} \left(\mathcal {W} _ {k} ^ {L} \left(\phi_ {L - 1} \circ \dots \circ \phi_ {1}\right) \left(\gamma \left(\mathbf {x} _ {k}\right), \mathbf {z} _ {k}\right) + \mathbf {b} _ {L}\right), \tag {4} +$$ + +where each layer $\phi_l: \mathbb{R}^{M_l} \mapsto \mathbb{R}^{N_l}$ applies affine transformation over the layer input $\mathbf{x}_l' \in \mathbb{R}^{M_l}$ followed by a non-linear activation function $\sigma_l$ : + +$$ +\phi_ {l} \left(\mathbf {x} _ {l} ^ {\prime}\right) = \sigma_ {l} \left(\mathcal {W} _ {k} ^ {l} \mathbf {x} _ {l} ^ {\prime} + \mathbf {b} _ {l}\right). \tag {5} +$$ + +Here, $\mathcal{W}_k^l$ and $\mathbf{b}_l\in \mathbb{R}^{N_l}$ are the weights and biases specific to the $l$ th layer. + +The Challenge: Balancing Efficiency and Expressiveness. A naive approach would treat $\mathcal{W}_k^l\in \mathbb{R}^{N_l\times M_l}$ as fully trainable, as done in prior volumetric models [2, 42, 43]. However, to capture complex human shapes and poses, these models require wide MLPs (256-512 neurons per layer), leading to: 1) High computational cost - large networks slow down inference. 2) Memory inefficiency - wide weight matrices consume excessive GPU memory. + +To address this, we propose the Neural Blend Weights generator (NBW) that predicts efficient neural network + +weights for narrow MLPs that retain expressiveness while being compact. An overview of NBW is illustrated in Fig. 3, showing how $\mathbf{z}_k$ is transformed into MLP weights. + +How NBW Works. For each layer $l$ of the MLP, we define the weight matrix $\mathcal{W}_k^l$ as a linear combination of $R$ learnable base weight matrices $\mathbf{W}_k^l [r] \in \mathbb{R}^{N_l \times M_l}$ (where $r = 1, \ldots, R$ ): + +$$ +\mathcal {W} _ {k} ^ {l} = \mathbf {W} ^ {l} + \sum_ {r = 1} ^ {R} \mathbf {v} _ {k} ^ {l} [ r ] \mathbf {W} _ {k} ^ {l} [ r ]. \tag {6} +$$ + +Here, $\mathbf{W}^l\in \mathbb{R}^{N_l\times M_l}$ is the base MLP weight matrix. $\mathbf{W}_k^l [r]$ are $R$ learned shape weight matrices specific to the $k$ th body part and layer $l$ . $\mathbf{v}_k^l\in \mathbb{R}^R$ are blending coefficients that determine how much each base matrix contributes to the final weights. + +These blending coefficients $\mathbf{v}_k^l$ are not fixed; instead, they are computed dynamically by passing the latent code $\mathbf{z}_k$ - which encodes the shape and pose of the $k$ th body part - through an independent linear layer. This ensures that the MLP weights adapt to the specific body configuration. + +This approach draws inspiration from the ResField architecture [41], originally designed for temporal signal fitting, but here repurposed for a feed-forward inference of network weights for volumetric body modeling. + +Advantages of NBW: 1) Compact architecture - NBW allows small MLPs (64 × 64 weight matrices), drastically reducing memory usage and speeding up inference. 2) High expressiveness - Dynamic blending of weights enables accurate modeling of complex SDFs, rivaling larger MLPs. 3) Scalability - Lower GPU memory consumption enables practical use in real-world applications. + +# 3.4. Training + +We follow the training protocol established in [42, 43], using SMPL [38] body model meshes from the AMASS dataset [40]. For training, we sample every 5th body model from the MoVi dataset [13] and every 10th model from DFaust [6]. For validation, we use PosePrior [1], a dataset featuring challenging body poses, where every 300th model is selected, resulting in a total of 316 bodies. + +At each training step, we sample query points $\mathcal{D}$ following the approach in [42]. Half of the samples (256 per body part) are drawn uniformly inside the bounding boxes $\mathcal{B}$ , while the other half is sampled around the mesh surface using Gaussian noise $\sim \mathcal{N}(0, 0.1)$ . The ground truth SDF value for each query point $d(\mathbf{x})$ is estimated by computing its distance to the surface mesh, with the sign robustly determined via inside-outside checks as described in [11, 42, 43]. + +The model is trained by supervising both the sign and absolute value of the SDF. The loss function is defined as: + +$$ +\mathcal {L} = \sum_ {\mathbf {x} \in \mathcal {D}} l _ {2} \left(s g n (\tilde {d} (\mathbf {x})), s g n (d (\mathbf {x}))\right) + l _ {2} \left(| \tilde {d} (\mathbf {x}) |, | d (\mathbf {x}) |\right). \tag {7} +$$ + +Table 1. Comparisons with Volumetric Bodies. Compared to COAP [42] on unseen subjects from the challenging PosePrior [1] dataset. Our body model is more accurate while providing SDF representation, being around $10 \times$ times faster, and using $6 \times$ times less GPU memory. The resources are measured on a total of $60 \mathrm{k}$ query points uniformly sampled around the body and sampled around the surface; surf. and unif. IoU refers to points sampled around the surface and uniformly at random respectively. + +
ResourcesIoU [%]MSE ↓
t [ms] ↓GPU ↓meansurf.unif.SDF|SDF|
LEAP [43]797.7 G75.9869.9881.97--
COAP [42]14018.7 G94.3193.9894.63--
VolumetricSMPL153.1 G94.6794.2595.103.7 × 10-53.5 × 10-5
+ +We train the network using a batch size of 8, optimizing the weights by minimizing $\mathcal{L}$ (Eq. 7) with the Adam optimizer [29]. The learning rate is initialized at $10^{-4}$ and gradually annealed to $10^{-5}$ . Training is performed for 15 epochs, totaling 450k iterations, and takes approximately 20 hours on a single 24GB RTX 3090 GPU. + +# 4. Experiments + +We evaluate VolumetricSMPL by benchmarking it against recent volumetric body models (Sec. 4.1) and demonstrating its effectiveness across multiple downstream applications (Sec. 4.2) visualized in Fig. 1. + +# 4.1. Volumetric Body Models + +We train all models on mesh sequences from MoVi [13] and DFaust [6] and evaluate them on the challenging PosePrior dataset [1], which features extreme body poses, following the setup in Sec. 3.4. For evaluation, we report Mean Intersection over Union (IoU), Mean Squared Error (MSE) for SDF prediction, GPU memory usage, and inference speed. + +Tab. 1 compares VolumetricSMPL with recent volumetric body models, LEAP [43] and COAP [42], both of which learn the SMPL shape space. Our model achieves higher accuracy while being $6 \times$ more memory efficient and around $10 \times$ faster in inference compared to COAP. This substantial reduction in computational cost makes VolumetricSMPL particularly advantageous for optimization and learning-based applications, where frequent volumetric body queries are required to impose penalty terms, as demonstrated in the next section. + +A direct comparison with imGHUM [2] is infeasible under our evaluation setup, as their released model is trained on private data and does not operate within the SMPL shape space. However, running their model shows that its inference is $86\%$ slower than VolumetricSMPL. For a fair comparison, in Appendix B.3, we attempt to re-implement their architecture and train it under our setup. Notably, like LEAP, imGHUM lacks direct compatibility for resolving self-intersections (Sec. 4.2.4). + +Table 2. Ablation Study of design choices in our body model. We compare the use of regular MLP instead of the proposed neural blend weights (NBW), the positional encoding $(\gamma(\cdot))$ [44] (w. two levels) for the local query point, and the varying numbers of NBW components, from rank $R = 1$ to $R = 80$ . + +
t [ms]ResourcesIoU [%] ↑MSE [×10-5] ↓
GPU ↓#parammeansurf.unif.SDF|SDF|
Base MLP152.9G0.4M92.7592.2393.275.26.9
+ γ(·)153.1G0.4M93.0092.4293.598.311
+ NBW (R=1)153.1G0.8M94.0693.5594.584.85.2
+ NBW (R=5)153.1G1.0M94.1293.8294.424.44.6
+ NBW (R=10)153.1G1.2M94.0793.6794.464.13.9
+ NBW (R=20)153.1G1.6M94.6094.1095.103.73.5
+ NBW (R=40)153.1G2.4M94.7994.4395.143.83.7
+ NBW (R=80)153.1G4.0M94.6794.2595.103.73.5
+ +We further analyze the impact of key design choices in Tab. 2, evaluating: 1) Our base method without Neural Blend Weights (NBW) (Sec. 3.3), instead using a standard MLP decoder. 2) The effect of positional encoding $(\gamma(\cdot))$ [44], specifically retaining only low-frequency components (two levels). 3) The number of NBW components $R$ , ranging from 1 to 80. + +Following the evaluation setup in Tab. 1, we observe that low-frequency $\gamma (\cdot)$ slightly improves IoU but reduces SDF accuracy, leading us to retain only two levels of $\gamma (\cdot)$ . The most significant improvement comes from NBW, with performance saturating at $R = 80$ , which we adopt as the final model configuration as it strikes a good balance between IoU and SDF metrics. Notably, increasing the parameter count by increasing $R$ has a negligible impact on computational resources, as it does not expand the MLP decoder size while still improving learning capacity. Additional insights and ablation of model parameters are provided in Appendix C. + +# 4.2. Downstream Tasks + +# 4.2.1. In-the-Wild Human-Object Reconstruction + +Recovering 3D human-object layouts from a single image is crucial for action understanding. PHOSA [68] introduced an optimization framework that refines human and object poses by enforcing contact rules. Wang et al. [59] improved scalability using large language models, but optimization remains costly, requiring $\sim 40$ minutes (1k steps) due to repeated SDF computations for object meshes. + +Instead of recomputing object SDFs at each optimization step, we represent the human as an SDF with VolumetricSMPL. This eliminates the bottleneck, accelerating optimization by $\sim 500\times$ while maintaining or improving reconstruction accuracy. + +Following [59], we evaluate collision (object penetration ratio) and contact (Chamfer distance) scores, along with per-iteration and total optimization time. As shown in Tab. 3, VolumetricSMPL drastically reduces runtime $(\sim 500\times)$ while preserving reconstruction quality. The test + +Table 3. In-the-Wild Human-Object Reconstruction. VolumetricSMPL achieves a $500\times$ speedup in optimization time over [59], reducing per-iteration cost from $2128.2\mathrm{ms}$ to $4.28~\mathrm{ms}$ and total optimization time from $35.9\mathrm{min}$ to $0.57\mathrm{min}$ . Despite this substantial acceleration, VolumetricSMPL maintains physical plausibility, achieving lower object penetration (collision) while preserving contact accuracy. We report optimization efficiency and reconstruction quality under identical settings, using the same optimization iterations and hardware (NVIDIA RTX 3090) on the "interaction database" introduced by [59]. + +
Optimization Time ↓Reconstruction Quality ↓
Per IterationTotalCollisionContact
w. SDF [59]2128.2 ms35.9 min6.33%3.20 cm
w. Ours4.28 ms0.57 min5.73%3.96 cm
+ +data is sourced from the "interaction database" introduced by [59], with metrics averaged over 100 images. + +By improving efficiency and scalability, Volumetric-SMPL paves the way for large-scale 3D human-object interaction dataset construction from web-scraped images. + +# 4.2.2. Human Mesh Reconstruction in 3D Scenes + +Reconstructing 3D human meshes from egocentric images is crucial for AR/VR applications. Given a 3D scene and an egocentric image, the goal is to recover a physically plausible 3D human mesh. + +Egocentric images often exhibit severe truncations due to close interaction distances, making pose inference challenging. To address this, we adopt the recent scene-conditioned diffusion model EgoHMR [69], which integrates COAP to guide sampling toward collision-free poses. However, COAP is computationally expensive, requiring substantial GPU memory and processing time, making inference costly, with over 2 seconds per frame. + +By replacing COAP with VolumetricSMPL, we significantly improve efficiency while slightly enhancing accuracy. At each diffusion step, given a denoised SMPL body and a 3D scene point cloud, we compute signed distances via VolumetricSMPL to guide sampling toward collision-free configurations. + +Following [69], we evaluate our method on the EgoBody dataset [71], measuring reconstruction quality for visible and occluded joints using MPJPE (Mean Per-Joint Position Error). Additionally, we assess physical plausibility with collision (coll.) and contact (cont.) scores, where collision measures scene penetration and contact evaluates proximity-based scene interaction. + +As shown in Tab. 4, replacing COAP with Volumetric-SMPL enables EgoHMR to run with a $10 \times$ larger batch size on the same GPU (NVIDIA TITAN RTX, 24GB) while achieving over $3 \times$ faster inference per frame. Additionally, VolumetricSMPL reduces interpenetrations and provides a slight improvement in joint accuracy. These gains stem + +Table 4. Human Mesh Reconstruction in 3D Scenes using EgoHMR with VolumetricSMPL vs. COAP. VolumetricSMPL enables $10 \times$ larger batch sizes (30 vs. 3) and $3.4 \times$ faster inference (0.61s vs. 2.08s per frame) while reducing human-scene collision (coll.) (18.4% vs. 19.1%) and maintaining optimal contact (cont.). MPJPE remains comparable for both visible (vis.) and invisible (invis.) joints, demonstrating improved efficiency without sacrificing reconstruction quality. Experiments conducted on an NVIDIA TITAN RTX (24GB) and the EgoBody dataset. + +
ResourcesInteraction [%]MPJPE [mm] ↓
Batch ↑t ↓coll. ↓cont. ↑meanvis.invis.
w. COAP32.08 s0.1919986.7465.7107.8
w. Ours300.61 s0.1849986.6365.8107.5
+ +from VolumetricSMPL's more efficient SDF-based gradients, which enable smoother and more precise collision handling compared to COAP's occupancy step function. + +# 4.2.3. Scene-Constrained Human Motion Synthesis + +Synthesizing realistic human motion in 3D scenes is crucial for applications in gaming, AR/VR, and robotics. A key challenge is ensuring the generated motion respects scene constraints, such as avoiding collisions, while preserving natural behavior. The recent motion generation work DartControl [78] introduces an optimization-based framework for synthesizing scene-constrained motions. It relies on evaluating human-scene collisions using precomputed scene SDFs to enforce collision constraints, which requires complete and high-quality scene scans. This limits its use to synthetic environments, as real-world reconstructions are often noisy or incomplete. Instead of modeling the scene as an SDF, we propose modeling the human as an SDF using VolumetricSMPL. This allows direct interaction with raw scans and enables optimization-based motion synthesis without requiring precomputed scene SDFs. + +We evaluate this approach on an indoor navigation task, where the goal is to generate human motion from a start position to a goal while avoiding obstacles. We integrate VolumetricSMPL's collision loss into DartControl's optimization framework and compare it against a COAP-based baseline. Results, averaged over 15 reconstructed scenes (8 sequences each, 80-120 frames), are presented in Tab. 5. We evaluate: 1) Per-frame computation time and memory usage for collision evaluation. 2) Total resources required for optimizing an 80-frame sequence. 3) Motion quality, measured via scene collisions and final body-to-goal distance. + +VolumetricSMPL achieves a $7\times$ speedup and $20\times$ lower GPU memory usage per frame compared to COAP. Due to COAP's high computational cost, scene collisions can only be evaluated for 16 frames per sequence on an 80GB GPU, requiring 71GB of memory. In contrast, VolumetricSMPL processes full 80-frame sequences on a 24GB GPU using only 15.9GB. + +Table 5. Scene-Constrained Human Motion Synthesis. Comparison of indoor navigation motion synthesis using [78] with different collision evaluation methods. VolumetricSMPL significantly enhances motion quality and efficiency, reducing per-frame memory usage by $20 \times$ and computation time by $7 \times$ compared to COAP. For an 80-frame sequence, VolumetricSMPL enables full-sequence optimization on a single 24GB GPU, requiring only 15.9GB of memory. In contrast, COAP exceeds the capacity of an 80GB GPU, restricting collision checks to 16 frames (denoted with *) and still consuming 71GB. Motion quality is also improved, reducing scene collisions over $90\%$ and achieving more precise goal-reaching behavior. Experiments were conducted on an 80GB NVIDIA A100. + +
Per-Frame ↓Total ↓Motion Quality ↓
MemoryTimeMemoryCollisionGoal Dist.
w. COAP4.44 GB26.53 ms71.0 GB*2.78 cm0.16 m
w. Ours0.19 GB3.78 ms15.9 GB0.24 cm0.01 m
+ +Notably, motion synthesized with VolumetricSMPL exhibits fewer scene penetrations and improved goal-reaching accuracy. This improvement is attributed to Volumetric-SMPL's SDF-based collision score, which provides smooth, informative gradients for optimization. In contrast, COAP's occupancy-based collision score offers weaker guidance, often disrupting optimization and leading to suboptimal motion generation. + +# 4.2.4. Volumetric Constraints for Self-Intersections + +Self-intersections are a common issue in human mesh reconstruction. Like COAP, VolumetricSMPL leverages its compositional structure to impose volumetric constraints that resolve mesh intersections. Given a self-intersected body configuration $(\theta, \beta)$ , we optimize the pose $\theta^{*}$ to eliminate these intersections. + +Following [42], we approximate body parts using 3D bounding boxes $\mathcal{G}$ to efficiently detect potential collisions. Intersecting box pairs define overlapping volumes, from which we uniformly sample an initial set of points. This set is refined by selecting only points inside at least two body parts, determined via part-wise SDF predictions. Let this final set be $S$ ; the self-intersection loss is then: + +$$ +\arg \min _ {\theta^ {*}} \sum_ {x \in \mathcal {S}} \sigma (f _ {\Theta} (x | \theta , \beta)) + \mathcal {L} _ {\text {P o s e P r i o r}}, \tag {8} +$$ + +where $\mathcal{L}_{\mathrm{PosePrior}}$ regularizes the optimization, ensuring the refined pose remains close to the original. Collisions between kinematically connected parts are ignored, as they are naturally expected to intersect (e.g., armpits). + +We follow COAP's experimental setup and evaluate our method on self-intersected human meshes sampled from PROX [18], using a subset of released 16 cases. The pose prior loss $\mathcal{L}_{\mathrm{PosePrior}}$ is implemented as an $\ell_2$ loss between the initial and optimized pose. We retain COAP's hyperparameters, setting the learning rate to $10^{-5}$ with a maximum + +Table 6. Volumetric Constraints for Self-Intersections. We compare VolumetricSMPL and COAP in terms of per-step computational cost (runtime and memory), total optimization iterations and time, and final mesh self-intersections. VolumetricSMPL achieves $2 \times$ faster per-step optimization (14ms vs. 30ms), $3 \times$ lower GPU memory usage (1.8GB vs. 6.2GB), and faster convergence (7.37 vs. 9.75 iterations, 0.10s vs. 0.29s total time). Additionally, it reduces self-intersections, producing fewer triangle collisions (68 vs. 74). These improvements demonstrate VolumetricSMPL's superior efficiency and accuracy over COAP. + +
Per-Step ↓Total ↓Mesh Triangle +Collisions ↓
TimeMemoryIterationsTime
w. COAP30 ms6.2 GB9.750.29 s74
w. Ours14 ms1.8 GB7.370.10 s68
+ +of 200 optimization iterations, a pose prior weight of $10^{3}$ , and a self-penetration loss weight of 0.1. The only change is replacing COAP's volumetric self-penetration term with our implementation using VolumetricSMPL. + +As shown in Tab. 6, VolumetricSMPL achieves per-step $2 \times$ faster optimization (14 ms vs. 30 ms), $3 \times$ lower GPU memory usage (1.8 GB vs. 6.2 GB), and overall fewer self-intersections (68 vs. 74 triangle collisions [49]). By significantly reducing computational overhead, VolumetricSMPL enables the efficient use of self-penetration constraints in downstream tasks that require larger batch sizes, making it more scalable for real-world applications. + +# 5. Conclusion + +We introduced VolumetricSMPL, a novel volumetric body model that overcomes the computational limitations of prior neural implicit models while preserving accuracy and enabling efficient interaction modeling. Our approach leverages Neural Blend Weights Generator to significantly reduce the memory and compute requirements of SDF-based body representations, achieving $10 \times$ faster inference and $6 \times$ lower GPU memory usage compared to COAP. + +Through extensive evaluations, we demonstrated that VolumetricSMPL improves both body model performance and downstream applications, including human mesh recovery, human-object interaction, scene-constrained motion synthesis, and self-intersection handling. By providing a scalable, differentiable, and computationally efficient framework for human interactions, VolumetricSMPL enables more realistic and robust modeling, paving the way for broader adoption in AR/VR, robotics, and autonomous systems. We hope this work serves as a foundation for future research in learning-based interaction modeling and advances in differentiable human-environment reasoning. + +Acknowledgments. Marko Mihajlovic is supported by Hasler Stiftung via the ETH Zurich Foundation (2025-FS-334). + +# References + +[1] Ijaz Akhter and Michael J. Black. Pose-conditioned joint angle limits for 3D human pose reconstruction. In CVPR, 2015. 5, 6 +[2] Thiemo Alldieck, Hongyi Xu, and Cristian Sminchisescu. imGHUM: Implicit generative models of 3d human shape and articulated pose. In ICCV, 2021. 2, 3, 5, 6, 1 +[3] Rayan Armani, Changlin Qian, Jiaxi Jiang, and Christian Holz. Ultra inertial poser: Scalable motion capture and tracking from sparse inertial sensors and ultra-wideband ranging. In ACM SIGGRAPH, 2024. 2, 3 +[4] Bharat Lal Bhatnagar, Xianghui Xie, Ilya Petrov, Cristian Sminchisescu, Christian Theobalt, and Gerard Pons-Moll. Behave: Dataset and method for tracking human object interactions. In CVPR, 2022. 3 +[5] Federica Bogo, Angjoo Kanazawa, Christoph Lassner, Peter Gehler, Javier Romero, and Michael J. Black. Keep it SMPL: Automatic estimation of 3D human pose and shape from a single image. In ECCV, 2016. 3 +[6] Federica Bogo, Javier Romero, Gerard Pons-Moll, and Michael J. Black. Dynamic FAUST: Registering human bodies in motion. In CVPR, 2017. 5, 6 +[7] Eric R Chan, Marco Monteiro, Petr Kellnhofer, Jiajun Wu, and Gordon Wetzstein. pi-gan: Periodic implicit generative adversarial networks for 3d-aware image synthesis. In CVPR, 2021. 3 +[8] Zhi-Quan Cheng, Yin Chen, Ralph R Martin, Tong Wu, and Zhan Song. Parametric modeling of 3d human body shape—a survey. Computers & Graphics, 2018. 3 +[9] MMPose Contributors. Openmmlab pose estimation toolbox and benchmark. https://github.com/open-mmlab/mmpose, 2020.3 +[10] Sisi Dai, Wenhao Li, Haowen Sun, Haibin Huang, Chongyang Ma, Hui Huang, Kai Xu, and Ruizhen Hu. Interfusion: Text-driven generation of 3d human-object interaction. In ECCV, 2025. 2, 3 +[11] Boyang Deng, JP Lewis, Timothy Jeruzalski, Gerard Pons-Moll, Geoffrey Hinton, Mohammad Norouzi, and Andrea Tagliasacchi. NASA: Neural Articulated Shape Approximation. In ECCV, 2020. 3, 5, 1 +[12] Mihai Fieraru, Mihai Zanfir, Teodor Szente, Eduard Bazavan, Vlad Olaru, and Cristian Sminchisescu. Remips: Physically consistent 3d reconstruction of multiple interacting people under weak supervision. NeurIPS, 2021. 2, 3 +[13] Saeed Ghorbani, Kimia Mahdaviani, Anne Thaler, Konrad Kording, Douglas James Cook, Gunnar Blohm, and Nikolaus F Troje. MoVi: A large multipurpose motion and video dataset. arXiv preprint arXiv:2003.01888, 2020. 5, 6 +[14] Ross Girshick. Fast R-CNN. In ICCV, 2015. 6 +[15] Hengtao Guo, Benjamin Planche, Meng Zheng, Srikrishna Karanam, Terrence Chen, and Ziyan Wu. Smpl-a: Modeling person-specific deformable anatomy. In CVPR, 2022. 2 +[16] David Ha, Andrew M. Dai, and Quoc V. Le. Hypernetworks. In ICLR, 2017. 3 +[17] Si Hang. Tetgen, a delaunay-based quality tetrahedral mesh generator. TOMS, 2015. 3 +[18] Mohamed Hassan, Vasileios Choutas, Dimitrios Tzionas, and Michael J. Black. Resolving 3D human pose ambiguities with 3D scene constraints. In ICCV, 2019. 2, 3, 8 + +[19] Mohamed Hassan, Partha Ghosh, Joachim Tesch, Dimitrios Tzionas, and Michael J Black. Populating 3d scenes by learning human-scene interaction. In CVPR, 2021. 2 +[20] Nikolas Hesse, Sergi Pujades, Michael J Black, Michael Arens, Ulrich G Hofmann, and A Sebastian Schroeder. Learning and tracking the 3d body shape of freely moving infants from rgb-d sequences. PAMI, 2019. 2 +[21] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. NeurIPS, 2020. 5 +[22] Alec Jacobson, Ladislav Kavan, and Olga Sorkine-Hornung. Robust inside-outside segmentation using generalized winding numbers. TOG, 2013. 2, 3 +[23] Wen Jiang, Nikos Kolotouros, Georgios Pavlakos, Xiaowei Zhou, and Kostas Daniilidis. Coherent reconstruction of multiple humans from a single image. In CVPR, 2020. 3 +[24] Korrawe Karunratanakul, Konpat Preechakul, Emre Aksan, Thabo Beeler, Supasorn Suwajanakorn, and Siyu Tang. Optimizing diffusion noise can serve as universal motion priors. In CVPR, 2024. 3 +[25] Hiroharu Kato, Yoshitaka Ushiku, and Tatsuya Harada. Neural 3d mesh renderer. In CVPR, 2018. 5 +[26] Rawal Khirodkar, Shashank Tripathi, and Kris Kitani. Occluded human mesh recovery. In CVPR, 2022. 3 +[27] Hyeonwoo Kim, Sookwan Han, Patrick Kwon, and Hanbyul Joo. Beyond the contact: Discovering comprehensive affordance for 3d objects from pre-trained 2d diffusion models. In ECCV, 2024. 2, 3 +[28] Meekyoung Kim, Gerard Pons-Moll, Sergi Pujades, Seungbae Bang, Jinwook Kim, Michael J Black, and Sung-Hee Lee. Data-driven physics for human soft tissue animation. TOG, 2017. 3 +[29] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2015. 6 +[30] Alexander Kirillov, Yuxin Wu, Kaiming He, and Ross Girshick. Pointrend: Image segmentation as rendering. In CVPR, 2020. 5 +[31] Muhammed Kocabas, Chun-Hao P. Huang, Otmar Hilliges, and Michael J. Black. PARE: Part attention regressor for 3D human body estimation. In ICCV, 2021. 3 +[32] Nikos Kolotouros, Thiemo Alldieck, Andrei Zanfir, Eduard Gabriel Bazavan, Mihai Fieraru, and Cristian Sminchisescu. Dreamhuman: Animatable 3d avatars from text. In NeurIPS, 2023. 2 +[33] Gen Li, Kaifeng Zhao, Siwei Zhang, Xiaozhong Lyu, Mihai Dusmanu, Yan Zhang, Marc Pollefeys, and Siyu Tang. Egogen: An egocentric synthetic data generator. In CVPR, pages 14497-14509, June 2024. 2, 3 +[34] Ronghui Li, Hongwen Zhang, Yachao Zhang, Yuxiang Zhang, Youliang Zhang, Jie Guo, Yan Zhang, Xiu Li, and Yebin Liu. Lodge++: High-quality and long dance generation with vivid choreography patterns. arXiv preprint arXiv:2410.20389, 2024. 3 +[35] Ronghui Li, Youliang Zhang, Yachao Zhang, Yuxiang Zhang, Mingyang Su, Jie Guo, Ziwei Liu, Yebin Liu, and Xiu Li. Interdance: Reactive 3d dance generation with realistic duet interactions. arXiv preprint arXiv:2412.16982, 2024.2.3 +[36] Tianye Li, Timo Bolkart, Michael. J. Black, Hao Li, and Javier Romero. Learning a model of facial shape and ex + +pression from 4D scans. SIGGRAPH Asia, 2017. 3 +[37] Yuwei Li, Longwen Zhang, Zesong Qiu, Yingwenqi Jiang, Nianyi Li, Yuexin Ma, Yuyao Zhang, Lan Xu, and Jingyi Yu. Nimble: a non-rigid hand model with bones and muscles. TOG, 2022. 3 +[38] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. SMPL: A skinned multiperson linear model. TOG, 2015. 2, 3, 4, 5, 1 +[39] Zhengyi Luo, Ryo Hachiuma, Ye Yuan, and Kris Kitani. Dynamics-regulated kinematic policy for egocentric pose estimation. NeurIPS, 2021. 3 +[40] Naureen Mahmood, Nima Ghorbani, Nikolaus F. Troje, Gerard Pons-Moll, and Michael J. Black. AMASS: Archive of motion capture as surface shapes. In ICCV, 2019. 5 +[41] Marko Mihajlovic, Sergey Prokudin, Marc Pollefeys, and Siyu Tang. ResFields: Residual neural fields for spatiotemporal signals. In ICLR, 2024. 3, 5 +[42] Marko Mihajlovic, Shunsuke Saito, Aayush Bansal, Michael Zollhoefer, and Siyu Tang. Coap: Compositional articulated occupancy of people. In CVPR, 2022. 1, 2, 3, 4, 5, 6, 8 +[43] Marko Mihajlovic, Yan Zhang, Michael J Black, and Siyu Tang. LEAP: Learning articulated occupancy of people. In CVPR, 2021. 2, 3, 5, 6, 1 +[44] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 4, 5, 6, 1 +[45] Lea Müller, Vickie Ye, Georgios Pavlakos, Michael Black, and Angjoo Kanazawa. Generative proxemics: A prior for 3d social interaction from images. In CVPR, 2024. 2, 3 +[46] Ahmed AA Osman, Timo Bolkart, and Michael J Black. Star: Sparse trained articulated human body regressor. In ECCV, 2020. 2 +[47] Pablo Palafox, Aljaz Bozic, Justus Thies, Matthias Nießner, and Angela Dai. Neural parametric models for 3d deformable shapes. In ICCV, 2021. 3 +[48] Pablo Palafox, Nikolaos Sarafianos, Tony Tung, and Angela Dai. Spams: Structured implicit parametric models. In CVPR, 2022. 3 +[49] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed AA Osman, Dimitrios Tzionas, and Michael J Black. Expressive body capture: 3d hands, face, and body from a single image. In CVPR, 2019. 2, 3, 4, 8, 1 +[50] Leonid Pishchulin, Stefanie Wuhrer, Thomas Helten, Christian Theobalt, and Bernt Schiele. Building statistical shape spaces for 3d human modeling. Pattern Recognition, 2017. 3 +[51] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In CVPR, 2017. 2, 4, 1 +[52] Davis Rempe, Tolga Birdal, Aaron Hertzmann, Jimei Yang, Srinath Sridhar, and Leonidas J. Guibas. Humor: 3d human motion model for robust pose estimation. In ICCV, 2021. 2 +[53] Javier Romero, Dimitrios Tzionas, and Michael J. Black. Embodied hands: Modeling and capturing hands and bodies together. SIGGRAPH Asia, 2017. 3 +[54] Vincent Sitzmann, Julien Martel, Alexander Bergman, David Lindell, and Gordon Wetzstein. Implicit neural representations with periodic activation functions. NeurIPS, 2020. 3 +[55] Cristian Sminchisescu and Alexandru C Telea. Human pose + +estimation from silhouettes. a consistent approach using distance level sets. In WSCG, 2002. 2 +[56] Yating Tian, Hongwen Zhang, Yebin Liu, and Limin Wang. Recovering 3d human mesh from monocular images: A survey. PAMI, 2023. 2, 3, 1 +[57] Gul Varol, Duygu Ceylan, Bryan Russell, Jimei Yang, Ersin Yumer, Ivan Laptev, and Cordelia Schmid. Bodynet: Volumetric inference of 3d human body shapes. In ECCV, 2018. 3 +[58] Shaofei Wang, Marko Mihajlovic, Qianli Ma, Andreas Geiger, and Siyu Tang. Meta­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­­. +clothed human models from few depth images. In NeurIPS, 2021. 2, 3 +[59] Xi Wang, Gen Li, Yen-Ling Kuo, Muhammed Kocabas, Emre Aksan, and Otmar Hilliges. Reconstructing action-conditioned human-object interactions using commonsense knowledge priors. In 3DV, 2022. 1, 2, 3, 6, 7, 4, 5 +[60] Songpengcheng Xia, Yu Zhang, Zhuo Su, Xiaozheng Zheng, Zheng Lv, Guidong Wang, Yongjie Zhang, Qi Wu, Lei Chu, and Ling Pei. EnvPoser: Environment-aware realistic human motion estimation from sparse observations with uncertainty modeling. arXiv preprint arXiv:2412.10235, 2024. 2, 3 +[61] Yiheng Xie, Towaki Takikawa, Shunsuke Saito, Or Litany, Shiqin Yan, Numair Khan, Federico Tombari, James Tompkin, Vincent Sitzmann, and Srinath Sridhar. Neural fields in visual computing and beyond. In Computer Graphics Forum, 2022. 5 +[62] Hongyi Xu, Thiemo Alldieck, and Cristian Sminchisescu. H-nerf: Neural radiance fields for rendering and temporal reconstruction of humans in motion. NeurIPS, 2021. 2 +[63] Hongyi Xu, Eduard Gabriel Bazavan, Andrei Zanfir, William T Freeman, Rahul Sukthankar, and Cristian Sminchisescu. Ghum & ghuml: Generative 3d human shape and articulated pose models. In CVPR, 2020. 2, 3, 4 +[64] Ye Yuan, Xueting Li, Yangyi Huang, Shalini De Mello, Koki Nagano, Jan Kautz, and Umar Iqbal. Gavatar: Animatable 3d gaussian avatars with implicit mesh learning. CVPR, 2024. 2 +[65] Ye Yuan, Jiaming Song, Umar Iqbal, Arash Vahdat, and Jan Kautz. Physdiff: Physics-guided human motion diffusion model. In CVPR, 2023. 2, 3 +[66] Ye Yuan, Shih-En Wei, Tomas Simon, Kris Kitani, and Jason Saragih. Simpoe: Simulated character control for 3d human pose estimation. In CVPR, 2021. 3 +[67] Seniha Esen Yuksel, Joseph N Wilson, and Paul D Gader. Twenty years of mixture of experts. IEEE TNNLS, 2012. 2 +[68] Jason Y. Zhang, Sam Pepose, Hanbyul Joo, Deva Ramanan, Jitendra Malik, and Angjoo Kanazawa. Perceiving 3d human-object spatial arrangements from a single image in the wild. In ECCV, 2020. 1, 2, 3, 6, 4, 5 +[69] Siwei Zhang, Qianli Ma, Yan Zhang, Sadegh Aliakbarian, Darren Cosker, and Siyu Tang. Probabilistic human mesh recovery in 3d scenes from egocentric views. In ICCV, 2023. 1, 2, 3, 7, 5, 6 +[70] Siwei Zhang, Qianli Ma, Yan Zhang, Zhiyin Qian, Taein Kwon, Marc Pollefeys, Federica Bogo, and Siyu Tang. Ego-body: Human body shape and motion of interacting people from head-mounted devices. In ECCV, 2022. 3 +[71] Siwei Zhang, Qianli Ma, Yan Zhang, Zhiyin Qian, Taein + +Kwon, Marc Pollefeys, Federica Bogo, and Siyu Tang. Ego-body: Human body shape and motion of interacting people from head-mounted devices. In ECCV, 2022. 7, 6 +[72] Siwei Zhang, Yan Zhang, Federica Bogo, Pollefeys Marc, and Siyu Tang. Learning motion priors for 4d human body capture in 3d scenes. In ICCV, 2021. 2 +[73] Siwei Zhang, Yan Zhang, Qianli Ma, Michael J. Black, and Siyu Tang. PLACE: Proximity learning of articulation and contact in 3D environments. In 3DV, 2020. 2 +[74] Xiaohan Zhang, Bharat Lal Bhatnagar, Sebastian Starke, Vladimir Guzov, and Gerard Pons-Moll. Couch: Towards controllable human-chair interactions. In ECCV, 2022. 3 +[75] Yan Zhang, Michael J. Black, and Siyu Tang. We are more than our joints: Predicting how 3D bodies move. In CVPR, 2021. 2 +[76] Yan Zhang, Mohamed Hassan, Heiko Neumann, Michael J Black, and Siyu Tang. Generating 3d people in scenes without people. In CVPR, 2020. 2, 3 +[77] Yan Zhang and Siyu Tang. The wanderings of odysseus in 3d scenes. In CVPR, 2022. 3 +[78] Kaifeng Zhao, Gen Li, and Siyu Tang. A diffusion-based autoregressive motion model for real-time text-driven motion control. In ICLR, 2025. 1, 2, 7, 8, 6 +[79] Kaifeng Zhao, Shaofei Wang, Yan Zhang, Thabo Beeler, and Siyu Tang. Compositional human-scene interaction synthesis with semantic control. In ECCV, 2022. 3 +[80] Kaifeng Zhao, Yan Zhang, Shaofei Wang, Thabo Beeler, and Siyu Tang. Synthesizing diverse human motions in 3d indoor scenes. In ICCV, 2023. 3 \ No newline at end of file diff --git a/ICCV/2025/VolumetricSMPL_ A Neural Volumetric Body Model for Efficient Interactions, Contacts, and Collisions/images.zip b/ICCV/2025/VolumetricSMPL_ A Neural Volumetric Body Model for Efficient Interactions, Contacts, and Collisions/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..d14141b74fa236f99754d9553ce395385d92c097 --- /dev/null +++ b/ICCV/2025/VolumetricSMPL_ A Neural Volumetric Body Model for Efficient Interactions, Contacts, and Collisions/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3c6d571e1b352f32a3a74e70ec6fd55bed512b137e011ee7eef7781a227d29c1 +size 368170 diff --git a/ICCV/2025/VolumetricSMPL_ A Neural Volumetric Body Model for Efficient Interactions, Contacts, and Collisions/layout.json b/ICCV/2025/VolumetricSMPL_ A Neural Volumetric Body Model for Efficient Interactions, Contacts, and Collisions/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..eb9d0e1055c53dbae7b572a58020a4e84b1ba146 --- /dev/null +++ b/ICCV/2025/VolumetricSMPL_ A Neural Volumetric Body Model for Efficient Interactions, Contacts, and Collisions/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3008e6c7fd29ebe769ebda552458a5a326804c5f03460a42470887ae508324e0 +size 479895 diff --git a/ICCV/2025/VoteSplat_ Hough Voting Gaussian Splatting for 3D Scene Understanding/5802dd71-01ac-4867-bdce-70fd12fae3bc_content_list.json b/ICCV/2025/VoteSplat_ Hough Voting Gaussian Splatting for 3D Scene Understanding/5802dd71-01ac-4867-bdce-70fd12fae3bc_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..b16393ad93e82b500147930fc6bfedba5b49ff4e --- /dev/null +++ b/ICCV/2025/VoteSplat_ Hough Voting Gaussian Splatting for 3D Scene Understanding/5802dd71-01ac-4867-bdce-70fd12fae3bc_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5d2ab9a2000718bb15bce43055e255be2021daaae5a2d543110fdc81e4c80121 +size 72297 diff --git a/ICCV/2025/VoteSplat_ Hough Voting Gaussian Splatting for 3D Scene Understanding/5802dd71-01ac-4867-bdce-70fd12fae3bc_model.json b/ICCV/2025/VoteSplat_ Hough Voting Gaussian Splatting for 3D Scene Understanding/5802dd71-01ac-4867-bdce-70fd12fae3bc_model.json new file mode 100644 index 0000000000000000000000000000000000000000..ee9624ec3ad494dd5cad2d91938bf5900b040c47 --- /dev/null +++ b/ICCV/2025/VoteSplat_ Hough Voting Gaussian Splatting for 3D Scene Understanding/5802dd71-01ac-4867-bdce-70fd12fae3bc_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:af029b53a44a9fc288a41f838db9af0a00520f4950824522917daebf06ab224d +size 91287 diff --git a/ICCV/2025/VoteSplat_ Hough Voting Gaussian Splatting for 3D Scene Understanding/5802dd71-01ac-4867-bdce-70fd12fae3bc_origin.pdf b/ICCV/2025/VoteSplat_ Hough Voting Gaussian Splatting for 3D Scene Understanding/5802dd71-01ac-4867-bdce-70fd12fae3bc_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..71f146571abb6b895e84f4427fb3c7a28f0606d9 --- /dev/null +++ b/ICCV/2025/VoteSplat_ Hough Voting Gaussian Splatting for 3D Scene Understanding/5802dd71-01ac-4867-bdce-70fd12fae3bc_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7b02fe75830cc0244e941f3b74cbe8fce8f53cd9651b23825929efe3efd127f5 +size 4159299 diff --git a/ICCV/2025/VoteSplat_ Hough Voting Gaussian Splatting for 3D Scene Understanding/full.md b/ICCV/2025/VoteSplat_ Hough Voting Gaussian Splatting for 3D Scene Understanding/full.md new file mode 100644 index 0000000000000000000000000000000000000000..e31b247753fa56edbd3eb0d506342b9ea00c2c21 --- /dev/null +++ b/ICCV/2025/VoteSplat_ Hough Voting Gaussian Splatting for 3D Scene Understanding/full.md @@ -0,0 +1,322 @@ +# VoteSplat: Hough Voting Gaussian Splatting for 3D Scene Understanding + +Minchao Jiang $^{1*}$ , Shunyu Jia $^{1*}$ , Jiaming Gu $^{1,2}$ , Xiaoyuan Lu $^{3}$ , Guangming Zhu $^{1}$ , Anqi Dong $^{4}$ , Liang Zhang $^{1\dagger}$ + +1 School of Computer Science and Technology, Xidian University + +$^{2}$ Algorithm R&D Center, Qing Yi (Shanghai) + +$^{3}$ Shanghai Pudong Cryptography Research Institute + +4 Division of Decision and Control Systems and Department of Mathematics, KTH Royal Institute of Technology + +{jamchaos, syjia_2001}@stu.xidian.edu.cn, jiaming-gu_xidian@outlook.com, {gmzhu, liangzhang}@xidian.edu.cn, xylu@bnc.org.cn, anqid@kth.se + +![](images/f3f863b972e0cc6be3dbc322440327041c5045c0b4dc08ba62a6e366144fa55f.jpg) +(a)Rendered View + +![](images/8f4ce0c27b885bc19480dae8c7c0194e52c7053d11a9c45fb8ac8465e5b1b862.jpg) +(b)Voting +Figure 1. VoteSplat integrates 3DGS and Hough Voting for 3D scene understanding: (a) 3DGS retains its original novel view synthesis capability, (b) each Gaussian primitive encodes an offset vector which votes the point cloud to the instance center, (c) 3D vote clustering enables instance segmentation, and (d) open-vocabulary 3D instance localization and click-based 3D object localization are demonstrated. + +![](images/a69aef31d237a2db51682d1a2585d8521fad300d61ac4f46f01edb6eef0261d8.jpg) +(c)Clustering +(d)3D scene understanding + +![](images/f3b300cad41ff53b19a5bb89d05e440757d40fdc5c27ab8f7d1efa1d9475c614.jpg) +Prompt:"Green chair" +Hierarchical Segmentation + +![](images/3fb6e8f85972f14ac1b2c0af197480e83700489e419a5ec879aa2470ded3b649.jpg) + +# Abstract + +3D Gaussian Splatting (3DGS) has become horsepower in high-quality, real-time rendering for novel view synthesis of 3D scenes. However, existing methods focus primarily on geometric and appearance modeling, lacking deeper scene understanding while also incurring high training costs that complicate the originally streamlined differentiable rendering pipeline. To this end, we propose VoteSplat, a novel 3D scene understanding framework that integrates Hough voting with 3DGS. Specifically, Segment Anything Model (SAM) is utilized for instance segmentation, extracting objects, and generating 2D vote maps. We then embed spatial offset vectors into Gaussian primitives. These offsets construct 3D spatial votes by associating them with 2D image votes, while depth distortion constraints refine localization along the depth axis. For open-vocabulary object localiza + +tion, VoteSplat maps 2D image semantics to 3D point clouds via voting points, reducing training costs associated with high-dimensional CLIP features while preserving semantic unambiguity. Extensive experiments demonstrate VoteSplat's effectiveness in open-vocabulary 3D instance localization, 3D point cloud understanding, click-based 3D object localization, hierarchical segmentation, and ablation studies. Our code is available at VoteSplat. + +# 1. Introduction + +Localization and instance-level semantic understanding of 3D scenes are critical objectives in the computer vision community, especially with the rise of embodied intelligence. Traditional point cloud-based methods for 3D scene understanding have been widely explored, including classification [21], segmentation [22], and detection [23]. Recently, 3D Gaussian Splatting (3DGS) [10] has gained popularity for its ability to achieve real-time, photorealistic + +novel view synthesis (NVS) at high resolutions. Unlike implicit Neural Radiance Fields (NeRF) [20], which represent scenes as continuous volumetric functions, 3DGS reconstructs scene appearance and geometry using pointclouds-alike Gaussian primitives. Building on its success, 3DGS has been extended to various domains, including rendering [6, 18, 39], surface reconstruction [4, 9, 40], generation [37], and scene understanding [24, 41]. Moreover, scene reconstruction and semantic understanding can be seamlessly integrated with 3DGS rendering pipeline, facilitating more advanced intelligent agent interaction and decision-making. + +Efforts have aimed to integrate learnable semantic information into 3DGS, enhancing the language-grounded capabilities of 3DGS. Current approaches can be broadly categorized into the following two types: + +(i) The first approach [24, 25, 29, 36] directly embeds semantic vectors into Gaussian primitives, project semantic information onto images in the same way as rendering colors. This enables cross-frame association through semantics but has two key limitations [35]: (1) The high dimensionality of CLIP-extracted features leads to excessive training overhead, while dimensionality reduction introduces semantic ambiguity; (2) The object occlusion restricts them to pixel-level segmentation, where embedded semantic vectors are insufficient for point-level scene understanding. + +(ii) The second approach [5, 35] adopts a point cloud clustering strategy to improve instance differentiation and reduce ambiguity, following a multi-stage pipeline. 3DGS independently reconstructs the scene and embeds feature vectors into the trained Gaussian primitives, using contrastive learning for instance differentiation. The goal is to address the spatial adjacency of primitives belonging to different instances, with additional features for better separation. However, contrastive learning adds significant computational complexity to the rendering pipeline. + +To this end, we propose VoteSplat, that integrates Hough voting with Gaussian splatting for 3D scene understanding. VoteSplat defines a set of 3D Gaussians embedded with additional three-dimensional spatial offset vectors to compute spatial 3D votes. We expect 3D votes to be located at the centroid of the instances so that clustering methods can be directly applied to distinguish each instance without the need for additional learned features for differentiation. Given the Gaussian primitives forming a specific instance and its centroid are unknown, we propose 3D-2D Votes Association Learning as the centroid may exist in multi-view two-dimensional images. + +Specifically, for each image, we first apply the Segment Anything Model (SAM) [12] to obtain well-segmented masks and compute their centroids with precise object boundaries. The pixel coordinates of these centroids serve as 2D ground-truth votes to supervise projection of 3D + +votes, encouraging convergence toward the instance center while ensuring multiview consistency. Since projection transformations cause depth information loss, relying solely on 2D vote supervision can introduce noise in spatial voting points. To overcome this, we introduce a depth distortion regularization term to improve spatial vote aggregation along the depth dimension. + +In a trained VoteSplat scene, each Gaussian primitive surrounding an object has an offset vector pointing to a 3D vote near the instance centroid. Clustering these votes allows us to effectively determine instance IDs to the Gaussian primitives, which in turn establishes correspondences between point clouds and image semantics, enabling robust 3D scene understanding. The contributions can be thus summarized as + +1. Hough voting is first considered into 3D Gaussian Splating (3DGS) to achieve spatial clustering of point clouds belonging to the same instance. This enables point-level segmentation without requiring additional high-dimensional feature vectors, improving training efficiency and scene understanding accuracy. +2. 3D-2D Votes Association Learning is proposed, incorporating a custom depth distortion loss to enhance spatial aggregation of voting points along the depth dimension for denoising. +3. An instance ID-based approach is introduced to associate 2D image semantics with 3D point clouds, enabling the linking of CLIP features to individual 3D instances, and facilitating open-vocabulary scene understanding. + +Therefore, VoteSplat enables efficient point-level segmentation without requiring high-dimensional feature embeddings. Additionally, we introduce 3D-2D Votes Association Learning and depth distortion regularization to refine spatial clustering and improve localization accuracy. The rest of the paper is organized as follows: Section 2 reviews related works in 3D Gaussian Splatting and Hough Voting. Section 3 details our methodology, including 3D vote construction and semantic association. Section 4 presents the experimental setup and results across various tasks. Finally, Section 5 concludes the paper. + +# 2. Related Works + +3D Gaussian Splatting for Scene Understanding. 3D Gaussian Splatting has emerged as a promising method for real-time scene rendering, offering superior visual quality. Utilize 3DGS to jointly reconstruct the appearance and geometric information of a scene with instance and semantic information, to better support downstream tasks. + +NeRF, as an innovative 3D reconstruction method, has inspired numerous works [31, 42] to develop 3D language fields upon it. LERF[11] first integrated CLIP[3] features into NeRF, constructing language embedded radi + +ance fields to enable open-vocabulary 3D querying. Additionally, DINO features were used to enhance boundary accuracy. However, due to high computational costs, NeRF-based methods face rendering performance bottlenecks, leading to the adoption of 3DGS [10] with semantic information. Building on 3DGS, LangSplat[24] employs an autoencoder to reduce CLIP feature dimensionality, embedding the compressed representations into Gaussian primitives, while incorporating a semantic hierarchy. Shi et al. [29] use VQ-VAE to quantize high-dimensional CLIP features into discrete categories, converting semantic supervision into category-level supervision, thereby reducing computational overhead. Similarly, Shorinwa et al. [30] constructs a 3D language field by mapping semantic features to discrete categories. + +These methods achieve multi-view training through cross-frame semantic similarity. To reduce computational complexity, CLIP feature needs to be either dimensionally reduced or classified, which inevitably introduces semantic ambiguity. Other methods [5, 16, 35, 36] rely on learnable features to distinguish instances and ultimately assign complete semantic information. These methods ensure semantic accuracy; however, intra-class and inter-class contrastive learning can be highly time-consuming. + +Hough Voting for 3D Point Clouds. The Hough transform (also, Hough voting), originates in late 1950s [8], convert the detection of simple patterns in point samples as peak detection in a parametric space. The Generalized Hough Transform [1] extends this concept to image patches, enabling the identification of complex objects. Hough voting has been widely applied in various tasks, including the implicit shape model[15], plane extraction from 3D point clouds[2], 6D pose estimation [32], and so forth. + +Hough voting has been successfully integrated with advanced learning techniques. In 3D object detection, a common approach [13, 14, 33, 34] is to adapt mature 2D detection algorithms, such as Faster R-CNN[27] or YOLO[26], to 3D point clouds by generating proposals at each input point. However, a fundamental challenge arises: 3D sensors capture only surface data, meaning the object center often lies in empty space, far from the available points in the input point cloud. As a result, there are typically no input points near the object center, making it difficult for surface-based networks to extract meaningful contextual information, leading to inaccurate proposals. To overcome this, Qi et al. [23] introduced VoteNet, a Hough voting-based method. VoteNet first samples seed points from the input point cloud and votes for the target's center, generating voting points near the object center. These voting points are then used to generate bounding box proposals, effectively addressing the issue of inaccurate proposals when the object center is distant from the surface points. + +# 3. Method + +We now formally introduce VoteSplat, a 3D scene understanding framework for 3DGS representations, incorporating effective and distinctive voting mechanisms. We first employ SAM's automatic mask generation module to produce masks for training views of the scene. The resulting multi-level mask information is then used to compute 2D ground-truth votes (detailed in Section. 3.2). Next, we embed offset vectors into 3DGS and train them to generate 3D votes (Section. 3.3). To improve depth consistency across different views, we introduce a depth regularization term, ensuring vote point aggregation during training. Finally, we establish a mapping between Gaussians and semantic information (Section. 3.4). The complete VoteSplat pipeline is illustrated in Figure 2. + +# 3.1. Recap: 3D Gaussian Splitting + +Compared to implicit Neural Radiance Fields (NeRF) [20], 3D Gaussian Splatting (3DGS) [39] constructs 3D scenes using explicit 3D Gaussian primitives. These primitives are represented as point clouds with associated attributes and are rendered through a tile-based differentiable rasterizer. + +Given a training set of $K$ images $I = \{I_k\}_{k=1}^K$ with associated camera poses, and an image resolution of $H \times W$ , the goal is to learn a set of $N$ three-dimensional Gaussian, denoted as $G = \{g_i\}_{i=1}^N$ . Each Gaussian $g_i$ is characterized by quintet $g_i := \{\mathbf{p_i}, \mathbf{s_i}, \mathbf{q_i}, o_i, \mathbf{c_i}\}$ of trainable parameters: the center position $\mathbf{p_i} \in \mathbb{R}^3$ , the scaling factor $\mathbf{s_i} \in \mathbb{R}^3$ , quaternion representing Gaussian's 3D covariance $\mathbf{q_i} \in \mathbb{R}^4$ , opacity value $o_i \in \mathbb{R}$ , and $\mathbf{c_i} \in [0,1]^3$ encodes RGB color using spherical harmonics coefficients. + +After projecting 3D Gaussians onto the 2D image space under a given camera pose, 3DGS computes the pixel color $\mathbf{C}$ using its differentiable rasterizer. The color is determined via $\alpha$ -blending across $\mathcal{N}$ depth-ordered points overlapping the pixel, given by + +$$ +\mathbf {C} = \sum_ {i \in \mathcal {N}} \mathbf {c} _ {\mathrm {i}} \alpha_ {i} T _ {i}, \tag {1} +$$ + +where $\alpha_{i}$ is determined by evaluating the influence of each projected Gaussian based on its splatted 2D covariance $\Sigma$ [38], opacity $o_i$ , and distance $\mathbf{d}$ to the pixel that reads + +$$ +\alpha_ {i} = o _ {i} \exp \left(- \frac {1}{2} \mathbf {d} ^ {\mathbf {T}} \Sigma^ {- 1} \mathbf {d}\right). \tag {2} +$$ + +The transmittance $T_{i} \coloneqq \prod_{j=1}^{i-1}(1 - \alpha_{j})$ , represents the accumulated visibility of $i$ -th Gaussian, accounting for occlusion from previously processed Gaussians in-depth order. + +# 3.2. 2D Vote Construction + +3D point clouds, such as those obtained from radar or Structure-from-Motion (SFM) [28], typically exhibit a concentration of points on the surface of spatial instances, + +![](images/f3be40a136ac3cb74a478490187dfa2734d15470ebafe6ffd318a58bba481b69.jpg) +Figure 2. Three main steps in VoteSplat pipeline: (a) We first deploy SAM to automatically generate segmentation masks for all instances independently across different views.(b) For each segmented mask, we compute the instance center to construct the 2D Vote Map. (c) By projecting 3D votes into pixel space through splatting and computing the voting loss together with the previous 2D voting map, each Gaussian primitive forming an instance is ensured to learn an offset vector pointing toward the instance center. For simplicity, we omit the rendering process and density control of other Gaussian parameters, as these are inherited from [10]. + +with sparser distributions toward the instance centers. This property persists during the densification process of 3DGS, wherein the generated point clouds remain concentrated on the instance surfaces. Consequently, directly clustering 3DGS point clouds can introduce boundary ambiguity between instances, leading to clustering errors when instance boundaries are in close proximity. Choi et al. [5], Shi et al. [29] embed additional feature vectors into the Gaussian Splatting framework to distinguish instances to avoid this. While effective, these methods often rely on computationally expensive contrastive learning. In contrast, 2D images inherently preserve structural information about instances and consistently capture object centers (provided the instance lies within the field of view). Exploiting this advantage, we propose a method to infer 3D instance centers from their corresponding 2D instance centers. We then detail the computation of the 2D centers (2D votes). + +SAM effectively groups pixels belonging to the same instance and segment images into multiple object masks with well-defined boundaries. Following Lang-Splat, we utilize SAM to obtain precise hierarchical object masks. To improve the accuracy of 2D votes, we further filter out masks whose boundaries extend beyond the field of view (FoV). + +For a given instance mask at level $l$ , denoted as $M_{l}$ , we compute the $x$ -axis instance centroid $c_{x}^{l}$ by + +$$ +c _ {x} ^ {l} = \text {r o u n d} \left(\frac {\sum_ {y = 0} ^ {H - 1} \sum_ {x = 0} ^ {W - 1} x \cdot M _ {l} (x , y)}{\sum_ {y = 0} ^ {H - 1} \sum_ {x = 0} ^ {W - 1} M _ {l} (x , y)}\right), \tag {3} +$$ + +recall $H$ and $W$ are the resolution of the image and round $(\cdot)$ is the rounding function and $y$ -axis centroid $c_y^l$ is computed similarly. The 2D vote $\mathbf{V}_{\mathrm{i}}^{2\mathbf{d}}$ is thus defined as the doublet + +$$ +\mathbf {V} _ {\mathbf {i}} ^ {\mathbf {2 d}} (x, y) := \left\{c _ {x} ^ {l}, c _ {y} ^ {l} \right\}. \tag {4} +$$ + +Each pixel within the mask is then assigned to its corresponding 2D vote $\mathbf{V}_{\mathrm{i}}^{2\mathrm{d}}$ , forming the ground-truth vote map, that serves as supervision for subsequent learning stages. + +# 3.3. 3D Vote Construction + +3D Vote Splatting. In 3DGS, both the initial and densified point clouds predominantly lie on instance surfaces, making direct clustering-based instance separation challenging. To overcome this, we introduce an offset vector $\Delta \mathbf{p_i} \in \mathbb{R}^3$ for each Gaussian primitive, allowing the point cloud to vote toward instance centers. Consequently, the 3D vote reads + +$$ +\mathbf {V} _ {\mathrm {i}} ^ {3 \mathrm {d}} = \Delta \mathbf {p} _ {\mathrm {i}} + \mathbf {p} _ {\mathrm {i}}. \tag {5} +$$ + +The additional vector attributes $\mathbf{V}_{\mathrm{i}}^{3\mathrm{d}}$ are typically optimized using the same $\alpha$ -blending approach as color rendering, i.e., + +$$ +\mathbf {V} ^ {\mathbf {3 d}} = \sum_ {i \in \mathcal {N}} \mathbf {V} _ {\mathbf {i}} ^ {\mathbf {3 d}} \alpha_ {i} T _ {i}. \tag {6} +$$ + +While generally effective, this approach faces two key challenges when applied to spatial constraints: + +(i) Unequal depth-weighting contribution: In 3D space, all points forming an instance should contribute equally to its center vote. However, the traditional $\alpha$ -weighting mechanism disproportionately reduces the influence of farther points, leading to biased voting. +(ii) Incorrect occlusion Handling: In 3DGS, points behind an instance still affect the final color rendering but should not be involved in the voting process, as they do not belong to the instance. + +We introduce a distinct voting transmittance model to overcome this, applying uniform averaging in $\alpha$ -blending as + +$$ +\mathbf {V} ^ {\mathbf {3 d}} := \frac {1}{| \mathcal {M} |} \sum_ {i \in \mathcal {M}} \mathbf {V} _ {i} ^ {\mathbf {3 d}}, \tag {7} +$$ + +where $\mathcal{M}$ represents set of the depth-ordered points under the voting transmittance $\hat{T}_i$ and $|\mathcal{M}|$ as its cardinality. Next, we project the blended 3D votes $\mathbf{V}^{3\mathrm{d}}$ into screen space: + +$$ +\tilde {\mathbf {V}} ^ {2 \mathrm {d}} = \mathbf {H} \mathbf {V} ^ {3 \mathrm {d}}, \tag {8} +$$ + +![](images/1cde177cfe019cdd348f118dc5588bc66fc2cec8862b87de173ba14cff6269bd.jpg) +Figure 3. Since projection results in-depth information loss, projected 3D votes may align correctly on the pixel plane but exhibit significant depth deviations, as shown in the left subfigure. The proposed depth regularization enforces spatial proximity along the depth dimension, keeping voting points concentrated in 3D space and enhancing overall voting consistency. + +where $\mathbf{H}$ is 4-by-4 transformation matrix from world space to screen space. Notably, blending is performed in 3D space before projection, ensuring stable voting. Otherwise, distant votes would experience significant fluctuations, making convergence toward the instance center less reliable. + +The vote loss with respect to the precomputed 2D votes $\mathbf{V}^{2\mathrm{d}}$ is defined as + +$$ +\mathcal {L} _ {\text {v o t e}} := \frac {1}{| \mathcal {P} |} \sum_ {i \in \mathcal {P}} \left| \tilde {\mathbf {V}} _ {\mathrm {i}} ^ {2 \mathrm {d}} - \mathbf {V} _ {\mathrm {i}} ^ {2 \mathrm {d}} \right|, \tag {9} +$$ + +where $\mathcal{P}$ denotes the set of pixels within the mask that contain 2D votes. Enforcing this loss ensures that offset vectors effectively guide Gaussians toward instance centers, while preserving the efficiency of the rasterization pipeline. Depth Regularization. Projecting 3D votes to 2D votes inherently results in a loss of depth information. While training with multi-view images helps reduce depth uncertainty and encourages the point cloud to collectively vote toward the instance center, disturbances in the voting points may still persist, as shown in Figure 3. + +Inspired by depth distortion in 2DGS [9], which promotes the concentration of Gaussian primitives, a parallel strategy is applied to 3D votes to preserve their depth alignment around the instance center. The initial depth distortion formulation in VoteSplat reads + +$$ +\mathcal {L} _ {d} = \sum_ {i, j} \omega_ {i} \omega_ {j} | z _ {i} - z _ {j} |. \tag {10} +$$ + +where $z$ is the $\mathbf{V}^{3\mathrm{d}}$ depth in the camera coordinate system with weight $\omega_{i}\coloneqq \alpha_{i}T_{i}$ + +Recall that the votes of Gaussian primitives need to be equally concentrated in-depth (av), the weight should not be reduced, and the formula is modified as + +$$ +\mathcal {L} _ {d} ^ {*} = \sum_ {i, j} | z _ {i} - z _ {j} |. \tag {11} +$$ + +by eliminating the weights from (10). + +Reconstruction-Voting-Depth (RVD) Loss: Lastly, the model optimizes the sparse point cloud, shifting it from its initial positions in the posed images toward a more concentrated target with respect to the RVD loss, i.e., + +$$ +\mathcal {L} = \mathcal {L} _ {c} + \lambda_ {\text {v o t e}} \mathcal {L} _ {\text {v o t e}} + \lambda_ {L _ {d}} \mathcal {L} _ {d} ^ {*}, \tag {12} +$$ + +where $\mathcal{L}_c(\mathcal{L}_1,\mathrm{D - SSIM})$ is the combination of RGB reconstruction loss $\mathcal{L}_1$ [10, Eq. 7] and D-SSIM term from [10, Eq. 7]. The vote loss (9) and depth regularization loss (11) penalized by their corresponding weights $\lambda_{vote}$ and $\lambda_{\mathcal{L}_d}$ , respectively, both depends on instances' size in the scene. + +# 3.4. Semantic Construction + +By constructing 3D votes, spatially separated voting points are obtained. To facilitate natural, open-vocabulary interactions, effectively associating 3D Gaussians with language features is essential. Thus, we propose an instance-level 3D-2D semantic association method based on voting points and instance IDs. The approach can be detailed as follows. + +- Background Filtering: Point clouds with an offset vector $\Delta \mathbf{p} = \mathbf{0}$ are removed, as they are considered background and do not contribute to instance construction, receiving no gradient updates. +- Clustering and Instance Dictionary Construction: The remaining point clouds are clustered using HDBSCAN [19] based on 3D votes, which performs density-based clustering while filtering out outliers. A dictionary is then built, where instance IDs serve as keys and the corresponding point cloud IDs as values. +- Rendering Instance ID Maps: The 3DGS rasterization pipeline is used to render the instance ID map. Combining this with the original RGB images, the pixel regions corresponding to each instance ID are identified. +- Semantic Association with CLIP Features: CLIP image features are extracted from the associated pixels, establishing a mapping between instance IDs and CLIP features, with multi-view feature integration incorporated for improved consistency. + +# 4. Experiments + +# 4.1. Open-Vocabulary Object Selection in 3D Space + +Experimental Setup. (i) Objective: Given an open-vocabulary text query, CLIP extracts textual features, and cosine similarity is computed with each instance ID's language features. The most relevant instances are selected, and their Gaussian primitives are rendered into multi-view images via the 3DGS rasterization pipeline; (ii) Baseline: Our method is compared against LangSplat, Gaussian Grouping, and OpenGaussian. VoteSplat follows the method described in Section 3.4 to associate each instance + +
MethodsmIoU ↑mAcc. ↑
figurinesteatimeramenwaldo_kitchenMeanfigurinesteatimeramenwaldo_kitchenMean
LangSplat10.1611.387.929.189.668.9320.3411.279.0912.41
OpenGaussian60.1165.8031.0122.7044.9082.1479.6642.2531.9258.99
VoteSplat68.6266.7139.2425.8450.1085.7188.1461.9733.6867.38
+ +![](images/38be2eb1956b23c7981954a859852645867c69bf67fd4c8c6ce49c965d74a715.jpg) +Figure 4. Click-based 3D object selection and scene editing results. VoteSplat enables complete 3D object selection without issues of incompleteness or redundancy. Moreover, after removing the selected Gaussian primitives, the scene can be effectively edited. + +Table 1. Performance of semantic segmentation on the LeRF dataset compared to LangSplat and OpenGaussian based on text query. Accuracy is measured by mAcc@0.25. + +
Methodssnacksfigurinesteatimeramen
LS(Level1)~ 67~ 116~ 87~ 56
OG~ 114~ 117~ 104~55
GG-~ 150~122~ 90
VS~ 57~ 54~ 43~ 53
+ +Table 2. The training time (in minutes) of LangSplat (LS), Open-Gaussian (OG), GaussianGrouping (GG), and VoteSplat (VS).2 + +with a 512-dimensional CLIP feature and select the corresponding Gaussian primitives for rendering. For LangSplat and OpenGaussian, we adhere to their respective procedures. In LangSplat, the 512-dimensional CLIP feature is reconstructed from the low-dimensional language feature of each Gaussian. In OpenGaussian, cosine similarity is computed to select the corresponding Gaussian primitives for rendering. Since Gaussian Grouping does not inherently support semantic queries on Gaussian primitives and is limited to instance segmentation, it is excluded from semantic query comparisons; (iii) Dataset and Metrics: Experiments are conducted on 3D-OVS[17] and Lerf-OVS[11], with all datasets annotated by LangSplat. Performance is evaluated using average IoU and segmentation accuracy, measuring the alignment between rendered images (from selected 3D Gaussian points) and ground-truth object masks. Additionally, we report training time across different methods and provide feature visualizations of the point clouds. + +Result. Table 1 shows VoteSplat outperforms other methods in both mIoU and mAcc. LangSplat, with weaker 3D understanding, struggles to accurately associate 3D Gaussian points with query text. This suggests that 2D image semantics derived from $\alpha$ -blending fail to effectively + +capture 3D semantics encoded by Gaussian primitives. Table 2 reports the training time for each model, evaluated over 60,000 iterations on NVIDIA RTX 3090 GPU. VoteS-plat achieves the shortest training time among all methods, benefiting from its efficient voting mechanism. In contrast, OpenGaussian's intra-/inter-class contrastive learning significantly increases training time. Gaussian Grouping [36] relies on KNN for feature consistency, leading to high computational and memory complexity during training. + +Qualitative results are presented in Figure 5. Given a text query, VoteSplat selects relevant Gaussian points and renders them into multi-view images. Due to the ambiguity of 3D point features, LangSplat struggles to accurately recognize target objects, while OpenGaussian fails to capture fine-grained details as effectively as VoteSplat. For example, with the prompt "old camera," VoteSplat successfully clusters finer details, such as the rope. Additionally, VoteSplat outperforms other methods in rendering occluded objects. In the teatime scene, it reconstructs the bear's lower body, despite being partially obscured by the table. Moreover, VoteSplat generates images with fewer noise artifacts, enhancing overall rendering quality. + +Figure 6 visualizes the point cloud features, where VoteSplat assigns distinct colors to instance categories for clarity. In OpenGaussian, colors are derived by applying PCA to reduce feature dimensions to three, and then mapping them to RGB. In contrast, LangSplat directly uses its three-dimensional features as point cloud colors. The feature visualization for Gaussian Grouping is provided in the supplementary material. The well-segmented instances in VoteSplat demonstrate its superior performance. + +![](images/b1031371191bc647847783bf2837385985280e98220dd2dd347a1fd3fe23d0a4.jpg) +Figure 5. Open-vocabulary 3D object selection on the LERF dataset. VoteSplat outperforms LangSplat and OpenGaussian in accurately identifying 3D objects corresponding to text queries. + +![](images/e3690f6182a81f1a69266c4e92a0ad8f49e290ff05137c8d9026913c729aa1dd.jpg) +Figure 6. Comparison of point cloud feature visualizations. From left to right, the scenes correspond to figurines, teatime, ramen, and snacks. The first three scenes are from LeRF, and the last scene is from 3D-OVS. Our proposed method, VoteSplat, demonstrates superior performance in terms of feature granularity and accuracy. + +# 4.2. Click-based 3D Object Selection and Editing + +Given an image from any viewpoint, clicking on a 2D pixel selects the corresponding 3D Gaussian points. The instance ID associated with the selected Gaussian primitive is then retrieved, enabling click-based object selection. Additionally, removing the entire instance allows for scene editing effects. Figure 4 demonstrates click-based object selection + +and scene editing on the LERF dataset. The left image highlights instance segmentation under occlusion, while the right image focuses on small object selection. + +# 4.3. Hierarchical Segmentation + +In some cases, instance segmentation is sufficient, but certain applications require finer segmentation, such as part + +![](images/1fcaea9773daa657f840f3da62346d38a0208b79af174df37b0f3657fa4822f8.jpg) +Figure 7. Using SAM, objects are divided into multiple parts, each assigned a 2D vote. After training, VoteSplat generates a corresponding 3D vote for each component. The rendering results of these components are then visualized. + +![](images/13ddb6c1894dc1201a2212f50a72ed574827717e323910c945154fef162f1edc.jpg) +Figure 8. Instance segmentation in complex scenes. Compared with Open-Gaussian, VoteSplat can handle more complex scenarios, such as instance overlap and contain each other. + +![](images/4f54e81dd7b567287eb3e875f3f5dfeb731f8029efdfa75d67e1c60feed2193e.jpg) +Figure 9. Comparison of ablation experiments. (a) shows 3D votes of VoteSplat are precisely located at the instance center. (b) reflects the effects of $w / o \mathcal{L}_d$ and the consequence of projecting first and then accumulating $\mathbf{V}_{\mathrm{i}}^{3\mathrm{d}}$ . (c) illustrates the impact of using transmittance $T$ . + +segmentation for more precise analysis. VoteSplat supports this functionality by utilizing multi-level masks from SAM to compute hierarchical 2D votes. These votes generate layered 2D voting maps, which then supervise 3D votes, enabling finer-grained segmentation. As shown in Figure 7, hierarchical 3D votes and rendering results are presented on the LLFF [20] and LeRF datasets. + +# 4.4. Instance Segmentation in Complex Scenes + +To evaluate VoteSplat in complex scenarios, experiments is conducted on GraspNet dataset [7], where instances are overlapping, adjacent, and contained. Despite instances' proximity, VoteSplat successfully segments them. As shown in Figure 8(b), 3D votes remain well-separated, demonstrating the method's effectiveness. + +# 4.5. Ablation Study + +As discussed in Section 3.3, projection inherently leads to depth information loss. While the projected points may align with 2D votes on the pixel plane, they can exhibit sig- + +nificant depth discrepancies. This issue is more pronounced in forward-facing data compared to 360-degree captures, as the absence of side-view images prevents voting points from consistently converging toward the instance center. Figure 9 illustrates the impact of $\mathcal{L}_d$ on the 3D-OVS dataset. Without $\mathcal{L}_d$ , the point cloud appears dispersed along the depth dimension, whereas with $\mathcal{L}_d$ , the points are more tightly clustered. Similarly, applying projection before accumulation results in 3D vote dispersion. Additionally, when using transmittance $T$ , background points participate in voting, introducing clustering disturbances, as in Figure 9(c). + +# 5. Concluding Remarks + +We introduce VoteSplat, a novel 3D scene understanding method that integrates Hough voting with 3D Gaussian Splitting (3DGS). Utilizing SAM for image instance segmentation, we generate 2D vote maps to supervise 3D votes, which are computed through embedded spatial offset vectors. To further refine clustering, we introduce depth distortion, constraining spatial offsets along the depth dimension, ensuring Gaussian primitives are well-clustered in 3D space. Additionally, by projecting instance IDs, VoteSplat establishes a precise correspondence between Gaussian primitives and 2D image semantics, effectively resolving semantic ambiguities. Experimental results confirm its effectiveness across various tasks. The voting mechanism encounters challenges when instances have projected sizes that significantly exceed the field of view (FoV), leading to inaccuracies in 2D votes. Additionally, it may also struggle with instances enclosed in concave containers, where spatial proximity makes precise separation difficult. + +# Acknowledgement + +# References + +[1] Dana H. Ballard. Generalizing the Hough transform to detect arbitrary shapes. Pattern recognition, 13(2):111-122, 1981. 3 +[2] Dorit Borrmann, Jan Elseberg, Kai Lingemann, and Andreas Nüchter. The 3D Hough transform for plane detection in point clouds: A review and a new accumulator design. 3D Research, 2(2):1-13, 2011. 3 +[3] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 2 +[4] Danpeng Chen, Hai Li, Weicai Ye, Yifan Wang, Weijian Xie, Shangjin Zhai, Nan Wang, Haomin Liu, Hujun Bao, and Guofeng Zhang. PGSR: Planar-based Gaussian Splitting for Efficient and High-Fidelity Surface Reconstruction. arXiv: 2406.06521, 2024. 2 +[5] Seokhun Choi, Hyeonseop Song, Jaechul Kim, Taehyeong Kim, and Hoseok Do. Click-gaussian: Interactive segmentation to any 3D gaussians. In European Conference on Computer Vision, pages 289-305. Springer, 2024. 2, 3, 4 +[6] Yitong Dong, Yijin Li, Zhaoyang Huang, Weikang Bian, Jingbo Liu, Hujun Bao, Zhaopeng Cui, Hongsheng Li, and Guofeng Zhang. A Global Depth-Range-Free Multi-View Stereo Transformer Network with Pose Embedding. arXiv: 2411.01893, 2024. 2 +[7] Hao-Shu Fang, Chenxi Wang, Minghao Gou, and Cewu Lu. Grasnet-1billion:A large-scale benchmark for general object grasping. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11444–11453, 2020. 8 +[8] Paul V.C. Hough. Machine analysis of bubble chamber pictures. In International Conference on High Energy Accelerators and Instrumentation, CERN, 1959, pages 554-556, 1959. 3 +[9] Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2D gaussian splatting for geometrically accurate radiance fields. In ACM SIGGRAPH 2024 conference papers, pages 1-11, 2024. 2, 5 +[10] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3D gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139-1, 2023. 1, 3, 4, 5 + +The authors sincerely thank the anonymous reviewers for their valuable comments and suggestions. This work was supported by grants from the Natural Science Foundation of Shanxi Province (2024JCJQN-66), Science and Technology Commission of Shanghai Municipality (NO.24511106900), Key R&D Program of Zhejiang (2024SSYS0091) and is partially supported by the National Natural Science Foundation of China under grant Nos. 62072358 and 62072352. +[11] Justin Kerr, Chung Min Kim, Ken Goldberg, Angjoo Kanazawa, and Matthew Tancik. Leref: Language embedded radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19729-19739, 2023. 2, 6 +[12] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dollár, and Ross Girshick. Segment anything. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4015-4026, 2023. 2 +[13] Jan Knopp, Mukta Prasad, and Luc Van Gool. Orientation invariant 3D object classification using Hough transform based methods. In Proceedings of the ACM workshop on 3D object retrieval, pages 15-20, 2010. 3 +[14] Jan Knopp, Mukta Prasad, and Luc Van Gool. Scene cut: Class-specific object detection and segmentation in 3D scenes. In 2011 International Conference on 3D Imaging, Modeling, Processing, Visualization and Transmission, pages 180-187. IEEE, 2011. 3 +[15] Bastian Leibe, Ales Leonardis, and Bernt Schiele. Robust object detection with interleaved categorization and segmentation. International journal of computer vision, 77:259-289, 2008. 3 +[16] Hajjie Li, Yanmin Wu, Jiarui Meng, Qiankun Gao, Zhiyao Zhang, Ronggang Wang, and Jian Zhang. Instancegaussian: Appearance-Semantic Joint Gaussian Representation for 3D Instance-Level Perception. arXiv: 2411.19235, 2024. 3 +[17] Kunhao Liu, Fangneng Zhan, Jiahui Zhang, Muyu Xu, Yingchen Yu, Abdulmotaleb El Saddik, Christian Theobalt, Eric Xing, and Shijian Lu. Weakly supervised 3D open-vocabulary segmentation. Advances in Neural Information Processing Systems, 36:53433-53456, 2023. 6 +[18] Tao Lu, Mulin Yu, Linning Xu, Yuanbo Xiangli, Limin Wang, Dahua Lin, and Bo Dai. Scaffold-GS: Structured 3D gaussians for view-adaptive rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20654-20664, 2024. 2 +[19] Leland McInnes, John Healy, and Steve Astels. HDBSCAN: Hierarchical density based clustering. J. Open Source Softw., 2(11):205, 2017. 5 +[20] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2, 3, 8 +[21] Charles R. Qi, Hao Su, Kaichun Mo, and Leonidas J. Guibas. Pointnet: Deep learning on point sets for 3D classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017. 1 +[22] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J. Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 30, 2017. 1 +[23] Charles R. Qi, Or Litany, Kaiming He, and Leonidas J. Guibas. Deep Hough voting for 3D object detection in point + +clouds. In proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9277-9286, 2019. 1, 3 +[24] Minghan Qin, Wanhua Li, Jiawei Zhou, Haoqian Wang, and Hanspeter Pfister. Langsplat: 3D language gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20051-20060, 2024. 2, 3 +[25] Jiaxiong Qiu, Liu Liu, Zhizhong Su, and Tianwei Lin. GLS: Geometry-aware 3D language gaussian splatting. arXiv:2411.18066, 2024. 2 +[26] Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. You only look once: Unified, real-time object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 779-788, 2016. 3 +[27] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster R-CNN: Towards real-time object detection with region proposal networks. IEEE transactions on pattern analysis and machine intelligence, 39(6):1137-1149, 2016. 3 +[28] Johannes L. Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4104-4113, 2016. 3 +[29] Jin-Chuan Shi, Miao Wang, Hao-Bin Duan, and Shao-Hua Guan. Language embedded 3D gaussians for open-vocabulary scene understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5333-5343, 2024. 2, 3, 4 +[30] Ola Shorinwa, Jiankai Sun, and Mac Schwager. Fast-Splat: Fast, Ambiguity-Free Semantics Transfer in Gaussian Splatting. arXiv:2411.13753, 2024. 3 +[31] Yawar Siddiqui, Lorenzo Porzi, Samuel Rota Buló, Norman Müller, Matthias Nießner, Angela Dai, and Peter Kontschieder. Panoptic lifting for 3D scene understanding with neural fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9043-9052, 2023. 2 +[32] Min Sun, Gary Bradski, Bing-Xin Xu, and Silvio Savarese. Depth-encoded Hough voting for joint object detection and shape recovery. In European Conference on Computer Vision, pages 658-671. Springer, 2010. 3 +[33] Alexander Velizhev, Roman Shapovalov, and Konrad Schindler. Implicit shape models for object detection in 3D point clouds. ISPRS Annals of the Photogrammetry, Remote Sensing and Spatial Information Sciences, 1:179-184, 2012. 3 +[34] Oliver J. Woodford, Minh-Tri Pham, Atsuto Maki, Frank Perbet, and Björn Stenger. Demisting the Hough transform for 3D shape recognition and registration. International Journal of Computer Vision, 106:332-341, 2014. 3 +[35] Yanmin Wu, Jiarui Meng, Hajjie Li, Chenming Wu, Yahao Shi, Xinhua Cheng, Chen Zhao, Haocheng Feng, Errui Ding, Jingdong Wang, and Jian Zhang. OpenGaussian: Towards Point-level 3D Gaussian-based Open Vocabulary Understanding. arXiv: 2406.02058, 2024. 2, 3 +[36] Mingqiao Ye, Martin Danelljan, Fisher Yu, and Lei Ke. Gaussian grouping: Segment and edit anything in 3D scenes. In European Conference on Computer Vision, pages 162-179. Springer, 2024. 2, 3, 6 + +[37] Taoran Yi, Jiemin Fang, Guanjun Wu, Lingxi Xie, Xiaopeng Zhang, Wenyu Liu, Qi Tian, and Xinggang Wang. Gaussiandreamer: Fast generation from text to 3D gaussian splattering with point cloud priors. arXiv: 2310.08529, 2023. 2 +[38] Wang Yifan, Felice Serena, Shihao Wu, Cengiz Öztireli, and Olga Sorkine-Hornung. Differentiable surface splatting for point-based geometry processing. ACM Transactions On Graphics (TOG), 38(6):1-14, 2019. 3 +[39] Zehao Yu, Anpei Chen, Binbin Huang, Torsten Sattler, and Andreas Geiger. Mip-splatting: Alias-free 3D gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19447-19456, 2024. 2, 3 +[40] Zehao Yu, Torsten Sattler, and Andreas Geiger. Gaussian opacity fields: Efficient and compact surface reconstruction in unbounded scenes. arXiv: 2404.10772, 2024. 2 +[41] Hongjia Zhai, Hai Li, Zhenzhe Li, Xiaokun Pan, Yijia He, and Guofeng Zhang. PanoGS: Gaussian-based Panoptic Segmentation for 3D Open Vocabulary Scene Understanding. In Proceedings of the Computer Vision and Pattern Recognition Conference (CVPR), pages 14114-14124, 2025. 2 +[42] Shuaifeng Zhi, Tristan Laidlow, Stefan Leutenegger, and Andrew J. Davison. In-place scene labelling and understanding with implicit scene representation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15838-15847, 2021. 2 \ No newline at end of file diff --git a/ICCV/2025/VoteSplat_ Hough Voting Gaussian Splatting for 3D Scene Understanding/images.zip b/ICCV/2025/VoteSplat_ Hough Voting Gaussian Splatting for 3D Scene Understanding/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..b8dae9cb6f714819b66607651b38bfacb87c6c0d --- /dev/null +++ b/ICCV/2025/VoteSplat_ Hough Voting Gaussian Splatting for 3D Scene Understanding/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:372c3f8cf1afe4cbfc2dff8cbb455c700ff4d49c6fa990ac4e86e42e4090d27d +size 667288 diff --git a/ICCV/2025/VoteSplat_ Hough Voting Gaussian Splatting for 3D Scene Understanding/layout.json b/ICCV/2025/VoteSplat_ Hough Voting Gaussian Splatting for 3D Scene Understanding/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..754bd68582a0d5fc01c8a2abfb9a650a92567df2 --- /dev/null +++ b/ICCV/2025/VoteSplat_ Hough Voting Gaussian Splatting for 3D Scene Understanding/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7e5657e0daae82f0da52e2c6edb9ddc08047957609ae4f2212ec313813ee7008 +size 362578 diff --git a/ICCV/2025/VoxelKP_ A Voxel-based Network Architecture for Human Keypoint Estimation in LiDAR Data/f5de6d36-8311-4c83-8578-66ba5526c326_content_list.json b/ICCV/2025/VoxelKP_ A Voxel-based Network Architecture for Human Keypoint Estimation in LiDAR Data/f5de6d36-8311-4c83-8578-66ba5526c326_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..c6c51cdbbdb77831ceff01e9356efe6a32e4d3c7 --- /dev/null +++ b/ICCV/2025/VoxelKP_ A Voxel-based Network Architecture for Human Keypoint Estimation in LiDAR Data/f5de6d36-8311-4c83-8578-66ba5526c326_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:083b8c97a5de63f0873eae3faa0224fd74c870963634b2c115cedb1f96cc82af +size 75251 diff --git a/ICCV/2025/VoxelKP_ A Voxel-based Network Architecture for Human Keypoint Estimation in LiDAR Data/f5de6d36-8311-4c83-8578-66ba5526c326_model.json b/ICCV/2025/VoxelKP_ A Voxel-based Network Architecture for Human Keypoint Estimation in LiDAR Data/f5de6d36-8311-4c83-8578-66ba5526c326_model.json new file mode 100644 index 0000000000000000000000000000000000000000..7390eca410f347ac60f15e6a9065cc81e158b536 --- /dev/null +++ b/ICCV/2025/VoxelKP_ A Voxel-based Network Architecture for Human Keypoint Estimation in LiDAR Data/f5de6d36-8311-4c83-8578-66ba5526c326_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:15f33e22a92973793b44be358699624625f4b785b0a5d8bc38b69b3c2bbdf2f9 +size 90260 diff --git a/ICCV/2025/VoxelKP_ A Voxel-based Network Architecture for Human Keypoint Estimation in LiDAR Data/f5de6d36-8311-4c83-8578-66ba5526c326_origin.pdf b/ICCV/2025/VoxelKP_ A Voxel-based Network Architecture for Human Keypoint Estimation in LiDAR Data/f5de6d36-8311-4c83-8578-66ba5526c326_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..7596340edf4a6b997d0f287b34aeb592ec6f522e --- /dev/null +++ b/ICCV/2025/VoxelKP_ A Voxel-based Network Architecture for Human Keypoint Estimation in LiDAR Data/f5de6d36-8311-4c83-8578-66ba5526c326_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3dfd17b962c0dbc0212c605b1975ec5fbd907b0e858b4b7ca7fcf3e613894a4b +size 2985896 diff --git a/ICCV/2025/VoxelKP_ A Voxel-based Network Architecture for Human Keypoint Estimation in LiDAR Data/full.md b/ICCV/2025/VoxelKP_ A Voxel-based Network Architecture for Human Keypoint Estimation in LiDAR Data/full.md new file mode 100644 index 0000000000000000000000000000000000000000..2304f4a82c5cff54e4a3fb429656a4476f997d3f --- /dev/null +++ b/ICCV/2025/VoxelKP_ A Voxel-based Network Architecture for Human Keypoint Estimation in LiDAR Data/full.md @@ -0,0 +1,267 @@ +# VoxelKP: A Voxel-based Network Architecture for Human Keypoint Estimation in LiDAR Data + +Jian Shi KAUST + +jian.shi@kaust.edu.sa + +Peter Wonka KAUST + +peter.wonka@kaust.edu.sa + +# Abstract + +We present VoxelKP, a novel fully sparse network architecture tailored for human keypoint estimation in LiDAR data. The key challenge is that objects are distributed sparsely in 3D space, while human keypoint detection requires detailed local information wherever humans are present. First, we introduce a dual-branch fully sparse spatial-context block where the spatial branch focuses on learning the local spatial correlations between keypoints within each human instance, while the context branch aims to retain the global spatial information. Second, we use a spatially aware multi-scale BEV fusion technique to leverage absolute 3D coordinates when projecting 3D voxels to a 2D grid encoding a bird's eye view for better preservation of the global context of each human instance. We evaluate our method on the Waymo dataset and achieve an improvement of $27\%$ on the MPJPE metric compared to the state-of-the-art, HUM3DIL, trained on the same data, and $12\%$ against the state-of-the-art, GC-KPL, pretrained on a $25\times$ larger dataset. To the best of our knowledge, VoxelKP is the first single-staged, fully sparse network that is specifically designed for addressing the challenging task of 3D keypoint estimation from LiDAR data, achieving state-of-the-art performance. Our code is available at https://github.com/shijianjian/VoxelKP. + +# 1. Introduction + +Human pose estimation (HPE) is a critical area of research with applications spanning computer vision, robotics, human-computer interaction, and augmented/virtual reality. Previous works [23, 32, 36] are mostly based on 2D images and videos. Compared to regular RGB input, LiDAR sensors provide detailed 3D structural information by measuring the distance to objects using laser light. Apart from its robustness under occlusion and illumination changes, LiDAR also offers privacy protection as it can not retain facial details. Existing 3D HPE approaches often rely on multi-stage pipelines [40], or require additional supervision, either from synthetic LiDAR data [37] or RGB im + +ages [42, 44]. Despite significant progress in 3D object detection from LiDAR point clouds [26-28], our experiments (in supplementary material) show that simply introducing a keypoint estimation head to these models often leads to suboptimal performance in 3D HPE tasks. Based on this observation, we suspect the object detection methods focus on capturing objects scattered sparsely across the 3D space, while the keypoints tend to be distributed densely within localized regions around the human body. This fundamental discrepancy in the context captured by existing detectors limits their suitability for precise 3D keypoint prediction due to the lack of fine-grained spatial information. In addition, most 3D object detectors use projected bird's-eye view (BEV) features to reduce computational costs. These height-compressed features complicate the precise keypoint localization. To address these gaps, we aim to extend the success of 3D object detection to 3D HPE for LiDAR point cloud data by introducing novel components to preserve fine-grained spatial information while retaining height information. As shown in Figure 1, our method significantly improves the precision of the estimated keypoints. + +We introduce VoxelKP, a novel fully sparse neural network tailored specifically for human pose estimation within LiDAR point clouds. We introduce a dual-branch fully sparse spatial-context block that integrates local dense features with the global spatial context from the sparse representations of LiDAR scans. More precisely, the spatial branch is used to extract local spatial correlation between keypoints, whilst the context branch is used to preserve the global spatial details. In this work, we refer to local features as fine-grained keypoint details within a detected human, and global features as larger-scale scene context. Second, we propose a spatially aware multi-scale BEV fusion module that aims to effectively encode absolute 3D coordinates to light-weight BEV representations. To the best of our knowledge, VoxelKP is the first single-staged, fully sparse network that is specifically designed for addressing the challenging task of 3D HPE from LiDAR data, achieving $27\%$ on the MPJPE metric compared to the current state-of-the-art trained on the same data. + +![](images/540043707a72a54859fd173d15bb51043b66a094e66ae8d0966cc5ada37d6471.jpg) +Figure 1. A visual demonstration of our baseline model (top) and the proposed VoxelKP (bottom). Our VoxelKP offers improved estimated human poses with precise locations and fewer false positives. The insets are color-coded according to the legend in the figure. In the green-colored insets, a comparison with the ground truth is shown, with ground truth in red and predictions in blue. Our baseline model is VoxelNeXt with additional human pose outputs. + +# 2. Related Work + +# 2.1. Deep Learning on Point Clouds + +Many neural network architectures have been proposed for processing smaller-scale point clouds [16, 22, 24, 25, 35]. However, typical LiDAR-generated point clouds contain more than 100,000 points, making point-by-point computations overwhelming due to the massive data scale. VoxelNet [45] proposed a voxel feature encoding (VFE) layer as a workaround for the high computational and memory issues brought by point-by-point computations. Meanwhile, sparse and submanifold sparse convolution operations [6] exploit sparsity in the voxel grid to reduce computations. SECOND [38] introduced an efficient sparse convolutional approach that benefits from the sparse operations. Following SECOND, subsequent works like PointPillars [13], 3DSSD [39], PV-RCNN [28], CenterPoint [41] further advanced sparse convolutional detection on point clouds. VoxelNeXt [2] further demonstrates a fully sparse voxel-based method without sparse-to-dense conversion or NMS postprocessing. However, these approaches are targeted at improving bounding box localization accuracy, which does not require fine-grained spatial features for precise pose estimation tasks. Instead, we propose VoxelKP, a novel sparse convolutional architecture tailored for learning discriminative local features from sparse LiDAR data for accurate human pose estimation. + +# 2.2. Human Pose Estimation on Point Clouds + +Human pose estimation has been extensively studied in images, with methods like DeepPose [36], Stacked Hourglass [23], and HRNet [32] achieving high accuracy on benchmarks like COCO-wholebody [10]. However, compared to RGB images, point clouds provide explicit 3D structural information about the shape and depth of objects. [29] pioneered point cloud human pose estimation from a single depth image. Recent works such as [21, 46] proposed a deep learning-based 3D human pose estimation from depth images. Waymo [33] has released keypoint annotations for LiDAR-collected point cloud scenes, while only $3\%$ of the frames are annotated with keypoint human poses. Due to the scarcity of the keypoint annotations within LiDAR point cloud data, many works have taken semi-supervised or weak-supervised approaches to compensate for the limited availability of labeled 3D pose data. For example, [42, 44] took a multi-modal approach to utilize the enriched image annotations to assist the recognition from point clouds. [37] proposed an unsupervised approach that generates pseudo ground truth without using annotated keypoint data, along with a fine-tuning approach that pretrains the model with synthetic data and then finetunes on the training set. A concurrent work [40] adopted a fine-tuning strategy that used a frozen backbone pretrained on a large-scale dataset as a feature extractor, achieving + +![](images/1d8445d1d47f529eb3e8620f738f3fc986101723d7e8f1892a1075208a146a1c.jpg) +Figure 2. The overall architecture of VoxelKP. The model begins with voxelizing a point cloud scene, followed by feature extraction using a stem module (Appendix A.1). Subsequently, the extracted sparse voxel features are processed through four fully sparse spatial-context blocks (Sec. 3.2) for capturing local and global spatial information. Lastly, we utilize a spatially-aware multi-scale BEV representation (Sec. 3.3) for accurate human pose estimation. + +plausible performance. In general, multi-person pose estimation from sorely point clouds remains relatively unexplored due to the lack of ground-truth 3D human pose annotations. This work proposes a single-staged 3D HPE method with only LiDAR point clouds, achieving comparable performance without extra training data. + +# 3. Method + +We propose a single-stage, fully sparse neural network, designed for 3D HPE within LiDAR point clouds. Unlike object detection, 3D HPE not only requires awareness of sparsely distributed human locations within the point cloud but also demands the ability to capture intricate spatial relationships between human keypoints. As illustrated in Figure 2, our VoxelKP framework contains two key components: 1) fully sparse spatial-context blocks to enhance the capturing of local sparse voxel features, and 2) spatially aware multi-scale BEV fusion to encode height information from the 3D voxel grid into 2D BEV maps. Our components are based on sparse operations that improve computational efficiency by skipping the computation for empty voxels. In this section, we first present the formulation of the task, then introduce the key components proposed in our network, and finally elaborate on the details of the network architecture. + +# 3.1. Problem Formulation + +Given a 3D point cloud scanned by LiDAR sensors, our goal is to estimate the 3D locations of $K$ keypoints that represent the human pose. Let the input point cloud $P$ be + +$\mathbb{R}^{N\times C}$ where $N$ is the number of points and $C$ is the number of features (e.g. x, y, z, intensity, elongation). We use a sparse voxel representation to represent point clouds, which consists of two separate tensors: one feature tensor $\mathbb{R}^{V\times C}$ and one index tensor $\mathbb{R}^{V\times 4}$ where $V$ is the number of nonempty voxels and 4 dimensions are used for batch sample index and the three coordinates of each voxel. We define the ground truth pose for the $i^{th}$ human as a set of 3D keypoint locations $G_{i} = \{g_{i}^{1}, g_{i}^{2}, \dots, g_{i}^{K}\}$ where $g_{i}^{k} \in \mathbb{R}^{3}$ is the location of the $k^{th}$ keypoint in the global coordinate frame. The set of $K$ keypoints corresponds to anatomical joints of interest such as shoulders, elbows, wrists, hips, knees, and ankles. Our objective is to predict the 3D keypoint locations from the input point cloud, i.e. to learn a function F such that $\hat{G} = F(P)$ , where $\hat{G} \in \mathbb{R}^{M\times K\times 3}$ is the tensor of predicted 3D keypoint locations of $M$ humans. + +# 3.2. Fully Sparse Spatial-Context Block + +We design specialized building blocks to process spatial information of varying densities effectively. Each building block starts with a basic sparse 3D block, subsequently branching into two distinct pathways for local and global spatial feature learning. Spatial branches are used for learning local spatial correlations between keypoints, incorporating sparse selective kernel modules (Sec. 3.2.1) and sparse box-attention modules (Sec. 3.2.2). Context branches use a straightforward MLP to maintain per-voxel information, ensuring the retention of global spatial details. The hybrid feature learning (Sec. 3.2.3) strategy captures the nuanced + +![](images/9c96ecc549eec1d3053d8ac07e7be4d8e409ecff12108fb1f79bb19786c3b506.jpg) + +![](images/72ec85573edd9c266728bc3dcdd08f62357195cdfe74091c0939d5c89a34b6ec.jpg) +Figure 3. Sparse selective kernel module with one sample input. The SSK module selects the best kernels from different receptive fields with a softmax-based channel-wise attention mechanism. + +local details with an understanding of the global context. Unlike $PVCNN$ [19] and $PVT$ [43], our approach does not require voxelization and devoxelization operations between modules, where the input is voxelized once, and all operations remain sparse throughout. + +# 3.2.1. Sparse Selective Kernel Module + +The sparse nature of voxel data means that only a limited set of non-empty voxels contribute to feature extraction, making it challenging to capture all relevant spatial details. Inspired by [15], we propose the sparse selective kernel (SSK) module that selectively aggregates multi-scale sparse features to improve spatial context. As demonstrated in Fig. 3, we first use two sparse 3D submanifold convolution branches with varied receptive field sizes of $3 \times 3 \times 3$ and $5 \times 5 \times 5$ . A submanifold convolution computes output values only if the convolution kernel is centered on a non-empty voxel, i.e., the number of non-empty voxels remains the same. Next, the sparse features from each branch are summed up and then fed into a selection module that compresses the spatial dimension by a sparse global average pooling (sGAP) to compute the average feature of all non-empty voxels, then a feature squeeze and expansion are applied [7]. The squeeze and expansion process compresses a feature map from $C$ channels to $Z$ channels, then expands it back to $C$ channels, where $Z$ is $25\%$ of $C$ in our implementation. We denote a voxel position as $p$ , the set of voxel positions within a voxel grid as $P_{s}$ , and the feature corresponding to voxel position $p$ as $f_{p}$ . The sGAP is achieved by: $\bar{F}_{s} = \{\frac{1}{|S_{\bar{P}}|}\sum_{p \in S_{\bar{P}}} f_{p}|p \in P_{s}\}$ , where $S_{\bar{P}}$ and $|S_{\bar{P}}|$ are the set of valid voxels and the number of valid voxels, respectively. This produces channel-wise attention weights after a softmax activation, allowing the network to emphasize or suppress the features from each branch selectively. + +# 3.2.2. Sparse Box-Attention Module + +Intuitively, the keypoint location is only determined by its local surrounding region, and the global context can barely + +help. In fact, as demonstrated in Sec. 5.1, the integration of global features can even harm the estimation accuracy. Thus, unlike the previous works that tried to capture a wider range of global features with self-attention methods for segmentation tasks [11, 12], we focus on localized feature attention to resolve the densely distributed keypoints in local regions. The key idea is to partition the sparse 3D voxel space into non-overlapping boxes by aggregating voxels by their spatial coordinates. Within each local box, we apply self-attention to capture dependencies between the voxels inside the box. The sparse voxel features in each box go through a linear layer for the queries $Q$ , keys $K$ , and values $V$ , where $Q, K, V \in \mathbb{R}^{n_b \times h \times d}$ and $n_b, h, d$ are the number of valid voxels in the $b$ -th box, attention heads, and feature dimensions. Since we are using sparse tensor representations, each box partition may contain a varying number of voxels. Referring to [11, 43], we then compute the attention map by the following equation: + +$$ +\operatorname {a t t n} _ {i, j, h} = Q _ {i, h} \cdot K _ {j, h}, \quad \operatorname {a t t n} _ {i.., h} = \operatorname {s o f t m a x} (\operatorname {a t t n} _ {i.., h}) +$$ + +$$ +y _ {i, h} = \sum_ {j = 1} ^ {n _ {b}} \operatorname {a t t n} _ {i, j, h} \times V _ {j, h}. \tag {1} +$$ + +The output is then obtained by applying a projection layer and a residual connection, as shown in Fig. 4. + +![](images/088a28489004523e4b6125bbae2912c3619308b3763c54c13c5b9c71e4088eca.jpg) +Figure 4. Sparse box-attention. This attention mechanism selects the voxel features that correspond to one box partition referring to the index tensor and then performs self-attention on the selected voxels. The functions $q, k, v$ , and $j$ are linear layers. + +# 3.2.3. Hybrid Feature Learning + +The sparse convolutional operations aggregate neighborhood information via spatial coordinates, which can overlook the fine-grained per-voxel details essential for keypoint localization. Concurrently, inspired by the previous point-voxel networks [19, 28, 43], we include an MLP branch for each stage. Differently, as a benefit of the fully sparse design, our approach avoids repeated voxelization and devoxelization between modules. The integration of an MLP branch alongside a convolutional branch is a strategic approach to capture both fine-grained per-voxel details and relatively coarse-grained local neighborhood information. Each MLP branch is composed of three sequential blocks, each consisting of a linear layer, batch normalization, and + +a ReLU activation function. The number of channels in each linear layer is set to match the channels of the incoming tensor. We then merge the output sparse voxel features from the MLP and sparse convolutional branches through element-wise summation to create hybrid features of the per-voxel and per-neighborhood information. + +# 3.3. Spatially Aware Multi-Scale BEV Fusion + +Compressing features into bird's eye view (BEV) maps is a common practice for object detection [1, 38] to collapse the point cloud to 2D for efficiency. However, different from object detection tasks, height information is essential for 3D HPE tasks to precisely locate each keypoint. For a sparse 3D voxel grid of size $C \times X \times Y \times Z$ , we use $C$ to denote the number of features per voxel, $X$ and $Y$ as the spatial extent in the ground plane, and $Z$ as the up axis. Ignoring the height information such as [2] is clearly unreasonable for 3D HPE. Another intuitive approach is to directly deploy 3D feature maps. Unfortunately, this direct 3D approach does not lead to a decent performance as training does not converge well due to the large amount of irrelevant features, as shown in Tab. 5. We, therefore, propose a spatially aware multi-scale BEV fusion approach for fusing sparse voxel features from multiple encoder layers in a way that retains spatial information, as illustrated in Fig. 5. Specifically, we use height encoding and scale-wise feature alignment to compensate for the loss of spatial information during the 2D projection. + +![](images/613bd4ed34adfb2d25cda366df76e6872197dcdd83e673122387f283b80b597b.jpg) +Figure 5. Spatially aware multi-scale BEV Fusion. We use a dense representation for a better visual illustration of the method. + +Height Encoding Transforming 3D data into BEV is often used in 3D object detection and segmentation tasks, for reducing the dimensionality of point clouds and making them more manageable for processing. An object detection method may project the 3D voxel grid to a 2D BEV repre + +sensation by adding features from voxels that share the same x and y position, losing the information about which height a feature was taken from. Instead, we use a height encoding method that compresses the height dimension to 1 using convolution kernels of size $(1,1,h)$ where $h$ is the height of each 3D voxel grid. Meanwhile, we increase the number of resulting channels to retain more spatial details and features from the 3D representation. This provides a richer representation for the 2D regression heads to work with. + +Multi-scale Feature Alignment After obtaining $z$ multiscale height-encoded BEV maps from the last few stages of the network, we then fuse those feature maps to create a feature map that contains multi-scale sparse voxel features. Unlike working with dense tensors, the direct interpolation of the feature maps in the sparse case is computationally complex, as it requires specialized algorithms to efficiently navigate through the predominantly empty voxels to find and interpolate the adjacent non-empty voxels. Instead, we directly modify the feature position of the sparse tensor by multiplying the voxel position by its scale $r$ . To avoid overlapping feature positions of $(x_{p} * 2^{r}, y_{p} * 2^{r})$ $r \in \{0,1,2\dots\}$ during the scale multiplication, we align the xy-plane positions $(x_{p}, y_{p})$ using scale offsets $(x_{p} * 2^{r} + r, y_{p} * 2^{r} + r)$ , where $p$ is the position of a voxel in a voxel grid. + +By stacking the $r$ -scaled feature maps together, we obtain a multi-scale 3D feature map with a height of $r$ . To obtain a BEV feature map, instead of collapsing with $1 \times 1 \times r$ convolutions, we simply apply an intuitive scaling for each scale of the feature map. The scaling factor $\hat{r}_p$ is proportional to the height (scale) of the 3D feature map for each feature $f_p$ at position $p$ , then we obtain scaled feature $\bar{f}_p = f_p \cdot \hat{r}_p$ . Given $\bar{P}$ as the set of the 2D xy-plane positions in the voxel grid, the compressed sparse features $\bar{F}$ and their positions $\bar{P}$ are obtained as: + +$$ +\bar {P} = \{(x _ {p}, y _ {p}) | p \in P \}, \quad \bar {F} = \{\sum_ {p \in S _ {\bar {p}}} \bar {f} _ {p} | \bar {p} \in \bar {P} \}, \quad (2) +$$ + +where $S_{\bar{p}} = \{p|x_p = x_{\bar{p}},y_p = y_{\bar{p}},p\in P\}$ contains voxels that are put onto the same 2D xy-plane position $\bar{p}$ . + +# 3.4. Network Architecture + +The input is a point cloud $\mathbb{R}^{N\times C}$ where $N$ is the number of points and $C$ is the number of features (e.g. x, y, z, intensity). We voxelize the point cloud into a sparse voxel representation. We use an input stem network and four stages with gradually decreased feature map size, where each stage reduces the spatial shape of the sparse voxel space by a factor of two. The input stem network is a simple stack of convolution layers, as shown in Appendix A.1, to extract low-level features from the voxelized point cloud. Next, four fully sparse spatial-context blocks are used for each subsequent stage for capturing features for accurate keypoint + +
MethodDatasetDescriptionMPJPE cm.
With Extra Training Data
Zheng et al. [44](CVPR 22)Internal dataset + Waymo v.?Trained on 155, 182 objects from internal data. Generated pseudo labels from 2D image labels.10.80 (-18%)
GC-KPL [37](CVPR 23)Waymo v.?Pre-trained on synthetic data. Fine-tuned on ground truth11.27 (-21%)
Waymo v.?Pre-trained on 200, 000 Waymo objects. Fine-tuned on ground truth10.10 (-12%)
Without Extra Training Data
HUM3DIL [42](CoRL 22)Waymo v.1.3.2Randomly initialized12.21 (-27%)
VoxelKPWaymo v.1.4.2Randomly initialized8.87
+ +Table 1. Benchmark results. The numbers in the table are taken from their corresponding papers aside from HUM3DIL, which is taken from GC-KPL paper. It is unclear about the exact training dataset used for Zheng et al. and GC-KPL. Waymo v1.3.2 and Waymo v1.4.2 share the same data for the human pose estimation task. + +localization. The sparse box-attention modules are only applied for our last two blocks to emphasize local-region features. Note that we do not increase the number of channels for the last three stages. We then convert the resulting 3D feature maps from the last three blocks to 2D spatial-encoded BEV representations. Note that we increase the number of channels for the BEV representation to compensate for the information loss of the BEV conversion. These 2D features are further refined with 2D convolutions to aggregate spatial context. In the end, we obtain the estimated keypoints $Y_{kp} \in \mathbb{R}^{K \times 3}$ and the corresponding predicted visibilities $Y_{kp} \in \mathbb{R}^{K}$ , where $K$ is the number of keypoints. The details of the used heads are provided in Appendix A.1. + +# 4. Experiments + +# 4.1. Implementation Details + +Dataset We use the Waymo v1.4.2 dataset [33]. During the training, we merged "Pedestrian" and "Cyclist" classes together as a "Human" class. Note that there are only 8, 125 human examples with keypoint annotations whilst over 1 million bounding box annotations. We, therefore, removed the points inside those bounding boxes without keypoint annotations. Each human object is labeled with 14 3D keypoints (nose, left/right shoulders, left/right elbows, left/right wrists, left/right hips, left/right knees, and left/right ankles, head). Additional evaluations on SLOPER4D [5] are provided in the supplementary. + +Network The architecture of the network is composed of a stem module followed by four stages, with output channels set to 64, 128, 256, 256, and 256, respectively. Given the high resolution (e.g. $1504 \times 1504 \times 61$ ) of the voxelized point cloud input, we employ larger sparse convolution kernels (kernel size $k = 5$ ) for the downsampling block in both the stem module and the initial stage. For the subsequent three stages, we revert to a smaller kernel size ( $k = 3$ ). To compensate for the information loss in the BEV projection, we increased the channels from 256 to 384. + +Training We use the point cloud range of $(150.4m, 150.4m, 6m)$ for the Waymo dataset and we transform them into voxel representations by a voxel size of + +$(0.1m, 0.1m, 0.1m)$ . We directly use the global keypoint locations without any encoding. Due to the limited number of training samples, we first apply a ground truth sampling technique [3, 38] to concatenate target objects from other frames into the sampled frames. Next, we apply global augmentations on the whole point cloud, including random flips on the $x$ and $y$ axes, random scale of the range of $[0.95, 1.05]$ , and random rotation ranged from $[- \pi / 4, \pi / 4]$ . Additionally, we apply local augmentations on each annotated object, including the random scale of the range of $[0.95, 1.05]$ , random rotation ranged from $[- \pi / 20, \pi / 20]$ , random frustum dropout [8] with an intensity range from $[0., 0.2]$ , and random noise around the object. Our model is trained using AdamW [20] optimizer plus OneCycle [31] learning rate scheduler to mitigate overfitting [30]. Specifically, we use a learning rate of 0.003, weight decay of 0.01, and 0.9 momentum. The details of the used loss functions can be found in Appendix A.3. + +# 4.2. Evaluation + +Benchmark Methods There is a limited amount of relevant research for this task. Most prior works utilize additional training data beyond the 3D keypoint data within the Waymo dataset. To provide a fair comparison, we consider approaches that use extra data and those that rely solely on Waymo separately. Zheng et al. [44] adopted a pseudolabel generation approach to provide stronger supervision. It utilizes an internal dataset for training and uses Waymo for evaluation. GC-KPL [37] pre-trains its backbone model with extra synthetic or real-world data, then fine-tunes the model with the full Waymo training set. Given the reliance on extra data in these methods, we consider the LiDAR-only version of HUM3DIL [42] as our primary competitor. HUM3DIL shares the exact same training data as our approach, allowing a direct comparison of techniques. + +Results Previous methods like GC-KPL use a subset of the validation data for evaluation, while we evaluate our method with the full validation set for better reproducibility. We report MPJPE on matched keypoints for our benchmark, following prior works. As shown in Tab. 1, we outperform the baseline HUM3DIL by approximately $27\%$ in MPJPE. + +
ComponentsMPJPEPEM
Spatial BEVSSKAttentionHybrid Feat.headshoulderselbowswristhipskneesanklesallall
✓0.06590.11270.16930.20200.09610.13430.19820.1394(0%)0.1973(0%)
0.07370.10260.14570.20130.08780.12850.19540.1332(+4%)0.1953(+1%)
✓✓0.06030.08480.12320.17150.07590.10840.16080.1118(+20%)0.1889(+4%)
✓✓✓0.05580.06040.09030.16790.06200.10910.18340.1039(+25%)0.1791(+9%)
✓✓✓✓0.05700.06690.09480.14670.06700.08200.10840.0887(+36%)0.1695(+14%)
+ +Our approach achieves state-of-the-art results among methods trained solely on Waymo ground truth. We also surpass the approaches leveraging extra synthetic data, beating Zheng et al. with synthetic pseudo labels by around $18\%$ and GC-KPL with synthetic point clouds by about $21\%$ . We achieve better performances as the SOTA GC-KPL approach which is pre-trained on 200,000 real-world samples by about $12\%$ . Overall, we demonstrate significant improvements over both the baseline solely using Waymo 3D keypoint data, as well as other techniques relying on extra data. Notably, our method can even achieve better results than previous multi-modal methods, e.g. [44]'s multimodal approach obtained 10.32 MPJPE with both LiDAR and RGB data while we achieved 8.87 with LiDAR only. A visual demonstration is presented in Fig. 1. Please find the accompanying video in the supplementary for a visualization of the results. In addition, we report the full spectrum of the evaluation in Appendix B, including keypointwise MPJPE, OKS@AP, and PEM. Our qualitative results in Fig. 6 clearly show that our method offers more reasonable estimations on the human objects that are not annotated by Waymo without pose distortions. + +# 5. Ablations + +We demonstrate the effectiveness of each proposed component in Tab. 2. We use the architecture of VoxelNext [2] as the baseline model, then gradually update the baseline model with the proposed components. We start with VoxelNeXt for two reasons: 1) it is one of the state-of-the-art point cloud object detection models with a fully sparse architecture design, and 2) it provides a good balance between computational costs and performance. We report the MPJPE for our ablations. The results indicate that all the individual components can contribute to improving 3D HPE. Compared to the baseline architecture, our proposed VoxelKP framework improves MPJPE by $36\%$ and PEM by $14\%$ . Next, we further present the ablation studies to show the alternative design choices of the individual component. + +# 5.1. Different Attention Mechanism + +Recent advancements, such as the stratified self-attention [11], focus on aggregating long-range contextual information, particularly beneficial for segmentation tasks. However, for human pose estimation tasks, capturing global + +![](images/56b4b57447aaeb57c1bfb7b6b38f4e118c67daef34b7887177ffd995c3288c86.jpg) + +![](images/a6909bc620a8624165797ef1e6277b7e48eaee945650646db35212efd2d6f93d.jpg) + +![](images/6e846c3d114941465229aad2e68954ed916c40405d980f6ef90b19cc29277dff.jpg) + +![](images/db87b87f980b41ef56acc99a8f987f1ce1b24b09bc5c8e364641d741ab18af0c.jpg) + +![](images/bd1fb375375ce827cc6d2092fd288ceae96e058069765f211f4c9eb9f1b81871.jpg) +Figure 6. A visual demonstration of the baseline model (top row) and VoxelKP (bottom row) on human objects that are not annotated by Waymo. The baseline model tends to produce keypoints that are squeezed together, whereas our approach yields a better result. + +![](images/bd1943f2af70bc3e367e38a90c677318afdb9db42bf90693163359f471542faa.jpg) + +![](images/5be302d54937aa2b9379b7143917d28d276e2f7c4fff4321bdf4cbfc96884b49.jpg) + +![](images/9c4204c9001527da2ed096bad0b3949859fbb89f2d079015b1729d1bdd39c6c0.jpg) + +dependencies is less crucial. Instead, our approach utilizes local box-attention, which concentrates on adjacent local regions. The results, as presented in Tab. 3, demonstrate that local box-attention outperforms other methods. Interestingly, we found that the stratified attention mechanism could slightly impair performance. We suspect that the box-based approach concentrates on areas most relevant to each keypoint location, whereas long-range attention may cause the network to overlook local, dense details. As a result, the box-based attention mechanism allows efficient modeling of local keypoint distributions, without excessive computation or over-smoothing from global aggregation. + +Table 2. Overall ablation for the effectiveness of each component. The first and last row represent the baseline method and our proposed VoxelKP architecture, respectively. + +
headshoulderselbowswristhipskneesanklesall
w/o0.06590.09560.14050.18550.08310.10770.15150.1181
stratified0.06500.09110.13470.19950.08190.12450.19190.1266
box0.05700.06690.09480.14670.06700.08200.10840.0887
+ +Table 3. Different attention methods. $w / o$ denotes no attention. + +# 5.2. The MLP Branch For Hybrid Feature Learning + +As shown in Fig. 2, the MLP is used for all the blocks, whilst the SSK and attention modules are only used for the + +last two blocks. Compared expensive 3D operations SSK and sparse attention, the MLP branch is a lightweight linear module that can be used to ensure global awareness for our model. Intuitively, the sparse box-attention aims to handle high-level feature patterns while the early-stage blocks are low-level feature extractors. If we deploy the sparse attention in the first stage, it will immediately encounter OOM on A100 80G GPUs, even though a CUDA optimized version is already used. Meanwhile, deploying it in the second stage significantly slows down the speed (over $5 \times$ slower) without performance boosts. In addition, from Tab. 4, we can see that the SSK and sparse attention create significant computational overhead with very high MACs, especially for the first two stages. + +
Speed (ms)MACs(G)
Stage12341234
MLP1.5941.5911.5841.6248.18.12.20.5
SSK42.73135.8813.8856.3885633.42816.7352.144.0
Attention22.98010.7594.9654.570136638.026671.42103.9468.6
+ +Table 4. MACs for each sparse module are estimated by assuming a dense voxel grid input. + +# 5.3. Spatially Aware BEV + +This ablation evaluates the effectiveness of the proposed spatially aware BEV module. We first evaluate the direct use of a naive 2D and 3D representations, followed by experiments with the spatially aware BEV. The findings, as shown in Tab. 5, indicate that our spatially aware BEV yields superior performance. The direct deployment of the 3D representation results in severe overfitting and, therefore, low performance. In addition, we also show that increasing the number of channels from 256 to 384 during the BEV projection can effectively improve the model performance, by compensating for information loss during projection. Overall, our spatially aware BEV strikes a balance that retains spatial acuity beyond basic BEV for resolving keypoint relationships while avoiding the complexity of full 3D convolutions. + +
cp.headshoulderselbowswristhipskneesanklesall
3D-2.46202.45592.44922.44492.43942.42642.4192.4422
V. BEV-0.06630.28100.45290.38020.19390.21480.25070.2614
OursX0.06880.07140.09820.16570.07230.10290.15950.1053
Ours✓0.05700.06690.09480.14670.06700.08200.10840.0887
+ +Table 5. Ablation study for the spatially aware BEV module. $Cp$ denotes if to expand the number of channels to compensate for the information loss during the 2D projection. $V$ . $BEV$ denotes the vanilla BEV method. + +# 5.4. Pose Estimation & Detection Trade-Off + +Although the PEM metric accounts for penalties in both box and keypoint mismatches, to provide a clearer understanding of the performance, we report both detection metrics and pose estimation metrics in Table 6 under different NMS + +thresholds. We reported the best MPJPE score in Tab. 1 at an NMS threshold of 0.3, while using a threshold of 0.1 improves detection performance with a slight sacrifice in MPJPE. Compared to the VoxelNeXt architecture, our method achieves similar detection performance (less than $1\%$ decrease) at an NMS threshold of 0.1, while significantly improving MPJPE performance (approximately $35\%$ increase). At an NMS threshold of 0.3, the MPJPE performance of VoxelKP can be further enhanced, although it results in a more substantial loss in detection performance. + +
ModelNMS ThresholdMPJPEPEMAP/L1AP/L2Recall@0.3Recall@0.5
VoxelNeXt0.10.14100.21200.71470.70960.97220.9375
VoxelKP0.10.09080.19000.70830.70490.96580.9354
VoxelNeXt0.30.13940.19730.66650.65860.86710.8404
VoxelKP0.30.08870.16940.60600.59980.78160.7565
+ +Table 6. Pose estimation and object detection performances under different NMS thresholds. + +# 6. Conclusion + +In this work, we identify the challenge of learning locally dense features within a sparse environment for human pose estimation. We proposed a new 3D fully sparse neural network for estimating dense human poses from point clouds. We present a comprehensive solution to the intricate challenges posed by spatial information of varying densities in the context of human pose estimation. Our method combines several novel components based on sparse operations to accurately and efficiently predict human body keypoints. Experiments on the Waymo dataset demonstrate the advantages of our approach compared to prior art and we demonstrate improved performance compared to other approaches trained on the same data as well as other approaches trained with additional data. While Waymo remains the sole publicly available dataset for these tasks, we eagerly anticipate the development of more datasets in the future. Through the identified challenge and the innovative framework, we pave the way for more nuanced and adaptable systems in this area. + +Despite these advancements, we further identify certain areas for future exploration and improvement. As mentioned above, this work used a small volume of training data, but it could benefit from a larger-scale dataset. While we focus on single-frame point clouds, future work could leverage temporal information across sequences of LiDAR point clouds. Additionally, instead of the straightforward estimation of keypoints, future work may adopt inverse kinematics to include physical constraints on human body movement. Aside from refining estimated keypoint locations, this may especially be useful to handle real-world challenges such as occlusion within motion. + +# Acknowledgments + +This work was supported by funding from King Abdullah University of Science and Technology (KAUST) - Center of Excellence for Generative AI, under award number 5940. + +# References + +[1] Xiaozhi Chen, Huimin Ma, Ji Wan, Bo Li, and Tian Xia. Multi-view 3d object detection network for autonomous driving. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 1907-1915, 2017. 5 +[2] Yukang Chen, Jianhui Liu, Xiangyu Zhang, Xiaojuan Qi, and Jiaya Jia. Voxelnext: Fully sparse voxelnet for 3d object detection and tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 2, 5, 7, 1 +[3] Zehui Chen, Zhenyu Li, Shiquan Zhang, Liangji Fang, Qinhong Jiang, and Feng Zhao. Autoalignv2: Deformable feature aggregation for dynamic multi-modal 3d object detection. ECCV, 2022.6 +[4] Spconv Contributors. Spconv: Spatially sparse convolution library. https://github.com/traveller59/spconv, 2022.1 +[5] Yudi Dai, Yitai Lin, Xiping Lin, Chenglu Wen, Lan Xu, Hongwei Yi, Siqi Shen, Yuexin Ma, and Cheng Wang. Sloper4d: A scene-aware dataset for global 4d human pose estimation in urban environments. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 682-692, 2023. 6 +[6] Benjamin Graham, Martin Engelcke, and Laurens Van Der Maaten. 3d semantic segmentation with submanifold sparse convolutional networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 9224-9232, 2018. 2 +[7] Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7132-7141, 2018. 4 +[8] Jordan SK Hu and Steven L Waslander. Pattern-aware data augmentation for lidar 3d object detection. In 2021 IEEE International Intelligent Transportation Systems Conference (ITSC), pages 2703-2710. IEEE, 2021. 6 +[9] Yihan Hu, Zhuangzhuang Ding, Runzhou Ge, Wenxin Shao, Li Huang, Kun Li, and Qiang Liu. Afdetv2: Rethinking the necessity of the second stage for object detection from point clouds. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 969-979, 2022. 1 +[10] Sheng Jin, Lumin Xu, Jin Xu, Can Wang, Wentao Liu, Chen Qian, Wanli Ouyang, and Ping Luo. Whole-body human pose estimation in the wild. In Proceedings of the European Conference on Computer Vision (ECCV), 2020. 2 +[11] Xin Lai, Jianhui Liu, Li Jiang, Liwei Wang, Hengshuang Zhao, Shu Liu, Xiaojuan Qi, and Jiaya Jia. Stratified transformer for 3d point cloud segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8500-8509, 2022. 4, 7 +[12] Xin Lai, Yukang Chen, Fanbin Lu, Jianhui Liu, and Jiaya + +Jia. Spherical transformer for lidar-based 3d recognition. In CVPR, 2023. 4 +[13] Alex H Lang, Sourabh Vora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom. Pointpillars: Fast encoders for object detection from point clouds. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12697-12705, 2019. 2 +[14] Hei Law and Jia Deng. Cornernet: Detecting objects as paired keypoints. In Proceedings of the European conference on computer vision (ECCV), pages 734-750, 2018. 1, 2 +[15] Xiang Li, Wenhai Wang, Xiaolin Hu, and Jian Yang. Selective kernel networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 510-519, 2019. 4 +[16] Yangyan Li, Rui Bu, Mingchao Sun, Wei Wu, Xinhan Di, and Baoquan Chen. Pointcnn: Convolution on x-transformed points. Advances in neural information processing systems, 31, 2018. 2 +[17] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 2 +[18] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. In Proceedings of the IEEE international conference on computer vision, pages 2980-2988, 2017. 1 +[19] Zhijian Liu, Haotian Tang, Yujun Lin, and Song Han. Pointvoxel cnn for efficient 3d deep learning. In Advances in Neural Information Processing Systems, 2019. 4, 1 +[20] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6 +[21] Xiaoxuan Ma, Jiajun Su, Chunyu Wang, Hai Ci, and Yizhou Wang. Context modeling in 3d human pose estimation: A unified perspective. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6238-6247, 2021. 2 +[22] Daniel Maturana and Sebastian Scherer. Voxnet: A 3d convolutional neural network for real-time object recognition. In 2015 IEEE/RSJ international conference on intelligent robots and systems (IROS), pages 922-928. IEEE, 2015. 2 +[23] Alejandro Newell, Kaiyu Yang, and Jia Deng. Stacked hourglass networks for human pose estimation. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part VIII 14, pages 483-499. Springer, 2016. 1, 2 +[24] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017. 2 +[25] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 30, 2017. 2 + +[26] Shaoshuai Shi, Xiaogang Wang, and Hongsheng Li. Poincarc: 3d object proposal generation and detection from point cloud. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 770-779, 2019. 1 +[27] Shaoshuai Shi, Zhe Wang, Xiaogang Wang, and Hongsheng Li. Part-a $^2$ net: 3d part-aware and aggregation neural network for object detection from point cloud. arXiv preprint arXiv:1907.03670, 2(3), 2019. +[28] Shaoshuai Shi, Chaoxu Guo, Li Jiang, Zhe Wang, Jianping Shi, Xiaogang Wang, and Hongsheng Li. Pv-rcnn: Pointvoxel feature set abstraction for 3d object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10529–10538, 2020. 1, 2, 4 +[29] Jamie Shotton, Andrew Fitzgibbon, Mat Cook, Toby Sharp, Mark Finocchio, Richard Moore, Alex Kipman, and Andrew Blake. Real-time human pose recognition in parts from single depth images. In CVPR 2011, pages 1297-1304. IEEE, 2011. 2 +[30] Leslie N Smith. A disciplined approach to neural network hyper-parameters: Part 1-learning rate, batch size, momentum, and weight decay. arXiv preprint arXiv:1803.09820, 2018.6 +[31] Leslie N Smith and Nicholay Topin. Super-convergence: Very fast training of neural networks using large learning rates. In Artificial intelligence and machine learning for multi-domain operations applications, pages 369-386. SPIE, 2019. 6 +[32] Ke Sun, Bin Xiao, Dong Liu, and Jingdong Wang. Deep high-resolution representation learning for human pose estimation. In CVPR, 2019. 1, 2 +[33] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, Vijay Vasudevan, Wei Han, Jiquan Ngiam, Hang Zhao, Aleksei Timofeev, Scott Ettinger, Maxim Krivokon, Amy Gao, Aditya Joshi, Yu Zhang, Jonathon Shlens, Zhifeng Chen, and Dragomir Anguelov. Scalability in perception for autonomous driving: Waymo open dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2, 6 +[34] OpenPCDet Development Team. Openpcdet: An opensource toolbox for 3d object detection from point clouds. https://github.com/open-mmlab/OpenPCDet, 2020.1 +[35] Hugues Thomas, Charles R Qi, Jean-Emmanuel Deschaud, Beatrix Marcotegui, François Goulette, and Leonidas J Guibas. Kpconv: Flexible and deformable convolution for point clouds. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6411–6420, 2019. 2 +[36] Alexander Toshev and Christian Szegedy. Deeppose: Human pose estimation via deep neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1653-1660, 2014. 1, 2 +[37] Zhenzhen Weng, Alexander S Gorban, Jingwei Ji, Mahyar Najibi, Yin Zhou, and Dragomir Anguelov. 3d human keypoints estimation from point clouds in the wild without human labels. In Proceedings of the IEEE/CVF Conference + +on Computer Vision and Pattern Recognition, pages 1158-1167, 2023. 1, 2, 6 +[38] Yan Yan, Yuxing Mao, and Bo Li. Second: Sparsely embedded convolutional detection. Sensors, 18(10):3337, 2018. 2, 5, 6 +[39] Zetong Yang, Yanan Sun, Shu Liu, and Jiaya Jia. 3dssd: Point-based 3d single stage object detector. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11040-11048, 2020. 2 +[40] Dongqiangzi Ye, Yufei Xie, Weijia Chen, Zixiang Zhou, and Hassan Foroosh. Lpformer: Lidar pose estimation transformer with multi-task network. arXiv preprint arXiv:2306.12525, 2023. 1, 2 +[41] Tianwei Yin, Xingyi Zhou, and Philipp Krahenbuhl. Center-based 3d object detection and tracking. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11784-11793, 2021. 2 +[42] Andrei Zanfir, Mihai Zanfir, Alex Gorban, Jingwei Ji, Yin Zhou, Dragomir Anguelov, and Cristian Sminchisescu. Hum3dil: Semi-supervised multi-modal 3d humanpose estimation for autonomous driving. In Conference on Robot Learning, pages 1114-1124. PMLR, 2023. 1, 2, 6 +[43] Cheng Zhang, Haocheng Wan, Xinyi Shen, and Zizhao Wu. Pvt: Point-voxel transformer for point cloud learning. International Journal of Intelligent Systems, 37(12):11985-12008, 2022. 4, 1 +[44] Jingxiao Zheng, Xinwei Shi, Alexander Gorban, Junhua Mao, Yang Song, Charles R Qi, Ting Liu, Visesh Chari, Andre Cornman, Yin Zhou, et al. Multi-modal 3d human pose estimation with 2d weak supervision in autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4478-4487, 2022. 1, 2, 6, 7 +[45] Yin Zhou and Oncel Tuzel. Voxelnet: End-to-end learning for point cloud based 3d object detection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 2 +[46] Yufan Zhou, Haiwei Dong, and Abdulmotaleb El Saddik. Learning to estimate 3d human pose from point cloud. IEEE Sensors Journal, 20(20):12334-12342, 2020. 2 \ No newline at end of file diff --git a/ICCV/2025/VoxelKP_ A Voxel-based Network Architecture for Human Keypoint Estimation in LiDAR Data/images.zip b/ICCV/2025/VoxelKP_ A Voxel-based Network Architecture for Human Keypoint Estimation in LiDAR Data/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..5f0f4240d769b5e5a25e9bd2a562914f9592a0b0 --- /dev/null +++ b/ICCV/2025/VoxelKP_ A Voxel-based Network Architecture for Human Keypoint Estimation in LiDAR Data/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fa7977a66aa3db935d18aadbc00c98be551dee9133d80f6f38500bc68a382cce +size 640813 diff --git a/ICCV/2025/VoxelKP_ A Voxel-based Network Architecture for Human Keypoint Estimation in LiDAR Data/layout.json b/ICCV/2025/VoxelKP_ A Voxel-based Network Architecture for Human Keypoint Estimation in LiDAR Data/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..74bd8ebd45f2d30bbad9ad4c843a038b90fc0937 --- /dev/null +++ b/ICCV/2025/VoxelKP_ A Voxel-based Network Architecture for Human Keypoint Estimation in LiDAR Data/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e0897e0fd1adfbe3ea721d2f9032c74628193ef2f6c5b1df4eba7be1b3f145bd +size 364320 diff --git a/ICCV/2025/Voyaging into Perpetual Dynamic Scenes from a Single View/c95a7c1e-c392-41cc-85d4-4436d0739bfa_content_list.json b/ICCV/2025/Voyaging into Perpetual Dynamic Scenes from a Single View/c95a7c1e-c392-41cc-85d4-4436d0739bfa_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..5aa66bfb49d49bb8b358a3a85072ccd92ac3d4a9 --- /dev/null +++ b/ICCV/2025/Voyaging into Perpetual Dynamic Scenes from a Single View/c95a7c1e-c392-41cc-85d4-4436d0739bfa_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:38f2e781338c0f53328aef64807a08058d9785caef41b5f6e8fd9649c2ca8214 +size 93163 diff --git a/ICCV/2025/Voyaging into Perpetual Dynamic Scenes from a Single View/c95a7c1e-c392-41cc-85d4-4436d0739bfa_model.json b/ICCV/2025/Voyaging into Perpetual Dynamic Scenes from a Single View/c95a7c1e-c392-41cc-85d4-4436d0739bfa_model.json new file mode 100644 index 0000000000000000000000000000000000000000..a0230ccb423576341d706da560b36cda4bec657a --- /dev/null +++ b/ICCV/2025/Voyaging into Perpetual Dynamic Scenes from a Single View/c95a7c1e-c392-41cc-85d4-4436d0739bfa_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4773b0db034476a16315a56a8ff2340a5ea13e8c93a73c401ada6257fb436f08 +size 118326 diff --git a/ICCV/2025/Voyaging into Perpetual Dynamic Scenes from a Single View/c95a7c1e-c392-41cc-85d4-4436d0739bfa_origin.pdf b/ICCV/2025/Voyaging into Perpetual Dynamic Scenes from a Single View/c95a7c1e-c392-41cc-85d4-4436d0739bfa_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..e744324fe8822f91d1a65ad458c4f0163bfd41fa --- /dev/null +++ b/ICCV/2025/Voyaging into Perpetual Dynamic Scenes from a Single View/c95a7c1e-c392-41cc-85d4-4436d0739bfa_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:184f624c96dafd29b7740e0aceb159e77422c8d94e4203bd916421edd625547f +size 16182963 diff --git a/ICCV/2025/Voyaging into Perpetual Dynamic Scenes from a Single View/full.md b/ICCV/2025/Voyaging into Perpetual Dynamic Scenes from a Single View/full.md new file mode 100644 index 0000000000000000000000000000000000000000..e6790b34fde28a6f931152348e81d486a08f2798 --- /dev/null +++ b/ICCV/2025/Voyaging into Perpetual Dynamic Scenes from a Single View/full.md @@ -0,0 +1,461 @@ +# Voyaging into Perpetual Dynamic Scenes from a Single View + +Fengrui Tian Tianjiao Ding Jinqi Luo Hancheng Min René Vidal University of Pennsylvania + +{tianfr,tjding,jinqiluo,hanchmin,vidalr}@upenn.edu + +![](images/d3cbe6660df44f38814f046cf189368763c7a07b700660fb5cb7cb74fdc67048.jpg) +Voyaging into Dynamic Scenes with Fly-through Camera Trajectories + +![](images/daddf7ae1b6d5f2e281585f189fbb818aae240cef3d8d9d1c8ea154e3caae9b3.jpg) +Controllable Dynamic Scene Generation with Scene Prompts +Figure 1. Dynamic Voyager generates 4D point clouds of perpetual dynamic scenes by our dynamic scene outpainting process. Given a fixed viewpoint video (or an image with a motion prompt) with dynamic scene prompts and fly-through camera trajectories, Dynamic Voyager can generate dynamic scenes along the trajectories (Top) and control scene generation contents (Bottom). + +# Abstract + +The problem of generating a perpetual dynamic scene from a single view is an important problem with widespread applications in augmented and virtual reality, and robotics. However, since dynamic scenes regularly change over time, a key challenge is to ensure that different generated views be consistent with the underlying 3D motions. Prior work learns such consistency by training on multiple views, but the generated scene regions often interpolate between training views and fail to generate perpetual views. To address this issue, we propose DynamicVoyager, which reformulates dynamic scene generation as a scene outpainting problem with new dynamic content. As 2D outpainting models struggle at generating 3D consistent motions from a single 2D view, we enrich 2D pixels with information from their 3D rays that facilitates learning of 3D motion consistency. + +More specifically, we first map the single-view video input to a dynamic point cloud using the estimated video depths. We then render a partial video of the point cloud from a novel view and outpaint the missing regions using ray information (e.g., the distance from a ray to the point cloud) to generate 3D consistent motions. Next, we use the outpainted video to update the point cloud, which is used for outpainting the scene from future novel views. Moreover, we can control the generated content with the input text prompt. Experiments show that our model can generate perpetual scenes with consistent motions along fly-through cameras. Project page: https://tianfr.github.io/DynamicVoyager. + +# 1. Introduction + +Perpetual scene generation [27, 30, 58] aims to create a virtual 3D scene as the camera moves along arbitrary tra + +jectories, typically starting from a single view observation. While recent studies [27, 57, 58] have achieved significant progress in perpetual generation of static scenes by leveraging image outpainting models [23, 24, 41], these approaches cannot generate scenes with dynamic content (e.g., waving hands, flowing rivers). Such dynamic content has important applications in augmented reality (AR), virtual reality (VR) [3, 4] and robotics [7]. For example, AR/VR game designers need to build a perpetual dynamic scene for players to explore and interact, while roboticists use virtual scenes containing commonly seen dynamic objects in natural environments for training embodied agents [9, 52] by self-exploration. + +A key challenge in generating perpetual dynamic scenes is to ensure that the generated dynamic content has 3D consistent motions: i.e., the motions observed in any two views must correspond to the same underlying 3D dynamics. Generating consistent motions from a single view is inherently ambiguous due to the limited 3D motion information contained in a 2D image. Previous dynamic scene generation methods [26, 28, 42, 50, 56] address this challenge by learning from multiple views surrounding the scenes so that 3D motion information can be implicitly learned from the 2D motions observed in these views. However, such a learning strategy severely restrains models from creating perpetual dynamic content, because the generated dynamic regions often interpolate between training views and fail to generate perpetual views, as exemplified in Figure 2. + +In this paper, we propose DynamicVoyager, a novel approach to perpetual dynamic scene generation that reformulates the task as a scene outpainting problem, enabling the synthesis of new dynamic content in previously unseen regions of the scene. Rather than relying on multiple 2D views as input, DynamicVoyager learns 3D-consistent motion dynamics from single-view observations by treating pixels as rays, thereby enriching each pixel with the contextual information of its corresponding camera ray. + +More specifically, given an input image, we exploit image-to-video diffusion models [14, 54] to synthesize a fixed pose video from that image. To initialize a dynamic scene from the video input, we estimate the video depth maps and backproject the fixed viewpoint video frames into 3D space as dynamic point clouds with the estimated depths. After that, we move the camera to a partially unseen area of the scene and rasterize the reconstructed point clouds into an incomplete video from that view. To outpaint the video with 3D consistent motions, we consider pixels as rays to complement the video input with 3D scene information. For pixels in the visible area, we obtain 3D motion information by sampling ray depth maps from the point clouds. For pixels in the unseen area, we backproject rays from pixels and use the distance between these rays and the point cloud to infer the 3D spatial relationship between + +![](images/599647b0f70f378030d4078e1f657d50063c321c8d3f04baf7b1abbc167ae9ee.jpg) +Input + +![](images/096562335e9b312c1e5ad830d703014252238a4e3f714e6bc130e4413b914ebe.jpg) +Move Right + +![](images/b9e578f83f7bd6dcec314f2e717c12cc7e6998c9362af924e488403f21ec0775.jpg) +Figure 2. Failure examples of previous dynamic scene generation methods. While the generated dynamic scenes from previous works [42, 49, 50] are strictly bounded by the input views, DynamicVoyager successfully generates dynamic scenes with large camera motions by the proposed scene outpainting process. + +![](images/d3d1fae165b743c2c303921ac947a05cb43fdb38d2193470f8ca6265e7692263.jpg) +Move Right + +the unseen area and the visible area. We then outpaint the incomplete video with the sampled ray depth and ray distance maps for the visible and unseen parts, respectively. Finally, we estimate the depth maps of the outpainted video and update the unseen area of the dynamic point clouds. Figure 1 visualizes examples of dynamic scenes generated by our method using input from a video with a fixed camera pose or a single image. It also shows that we can control the perpetual generation with input scene prompts. + +To summarize, our contributions are as follows: + +1. We reformulate the dynamic scene generation problem as a scene outpainting problem, so that our generated scene can be explored from a single view to any place through fly-through camera trajectories. +2. We generate view-consistent motions in 3D space by treating pixels as rays, which allows us to use 3D information to enrich the outpainting model. +3. We present experiments showing that, unlike other bounded scene generation methods, our model can generate perpetual dynamic scenes and control the generation with scene prompts. + +# 2. Related Work + +Static scene generation. With the development of vision foundation models [18, 20, 37, 40, 47], many works have started to generate static scenes from input images [2, 6, 8] or text prompts [13, 31]. Early efforts focused on indoor scene generation [2, 6, 13, 25] or object-centric scene generation [8]. These methods firstly generate multiview images of the target scene and fuse the multiview images by using 3DGS [19] or NeRF [32, 43, 44] representations. However, they can only generate static scenes with limited ranges of camera motions. When the camera moves drastically, these methods fail to generate unseen regions with consistent and clear scene structures. To overcome this issue, later studies have started to explore perpetual scene generation [30] that allows larger camera movements [17, 27, 30, 35, 57, 58]. Notably, WonderJourney [58] and WonderWorld [57] propose to generate an infinite world with diverse contents by using a combination of state-of-the-art depth estimation + +![](images/eb6c60da57f2c36bb6f663bcd33d84d83a7dee5d87f049dd0fd25e3918a9c6f4.jpg) +Figure 3. The overview of Dynamic Voyager. First, for an initial camera pose, we build the dynamic point clouds from the input image by employing the image-to-video diffusion model [54], depth model [37] and foreground segmentation model [18] (§3.1). Then given a new camera pose, we render the partial video and the corresponding ray depth and ray distance maps. We consider pixels as rays to enrich the pixel information of the partial video with the corresponding ray depth and ray distance information for outpainting the video with consistent motions (§3.2). Finally, we update the dynamic point cloud with the outpainted video and the corresponding video depth (§3.3). + +![](images/b7fbd2c213b99a2c26c2d46e22b2c2f1a87a9dfc1a5c25c6928300ac3fcb9314.jpg) + +[37], image segmentation [20] and object detection models [18]. However, these methods can only generate diverse scenes with static objects. In this paper, we target the problem of generating dynamic scenes with new dynamic contents with fly-through camera trajectories. + +Dynamic scene generation. Following the great success of static scene generation methods, recent studies have started to address the dynamic scene generation problem [26, 28, 34, 38, 42, 49, 50, 55, 56]. Due to the limited resources of real-world dynamic scene data, many researchers try to leverage the pretrained video diffusion models [14, 39, 51, 54] to generate multiple views of dynamic scenes. A key challenge in this setting is maintaining 3D motion consistency in multiple generated views. CAT4D [49] and DimensionX [42] learn to generate 3D consistent motion from multiview datasets [5, 29]. However, as the multiview images in these datasets are taken from cameras that are very close to each other, the generated dynamic scenes in these methods are strictly limited by the input images or videos. 4K4DGen [26] animates 3D consistent motion by building a 3D consistent noise space in the diffusion models. GEN3C [39] achieves camera control of dynamic scenes with 4D point cloud representations. However, since these methods only interpolate novel views surrounding the input image or videos, they still cannot generate new dynamic contents in the unseen regions of the input images. + +In this paper, we consider dynamic scene generation as a scene outpainting problem so that our method can generate new dynamic content with fly-through camera motions in the unseen regions of the input video. + +Video diffusion models. Our method relies on the power of the recent video diffusion models [51, 54] to outpaint dynamic scenes with diverse contents. While some researchers propose video diffusion models with controllable camera trajectories [1, 10, 11, 15, 22, 45, 53, 60], these trajectories are bounded by the regions of the input views (a failure case is shown in Figure 8). Besides, since the outputs of these methods are generated videos without any 4D scene structures, they always encounter view inconsistency issues when fusing 4D representation with the output videos. Instead, our method adopts 4D point clouds as the dynamic scene representation and outpaints the dynamic point clouds with video outpainting models, where the unseen regions in the input video can be generated by the outpainting procedure, and the motion consistency in different views can be achieved by introducing ray information in 3D space. + +# 3. Approach + +In this section, we present the proposed DynamicVoyager approach for generating perpetual dynamic scenes, which is summarized in Figure 3. DynamicVoyager consists of + +three steps, described in §3.1, §3.2 and §3.3, respectively: + +- Dynamic Scene Initialization: Dynamic Voyager accepts either a video with a fixed camera pose or an image with a prompt describing object motions as input—in the latter case, it first generates a video with a fixed pose. A depth map is then estimated from each frame of the video. We segment the video into a dynamic foreground and a static background, reconstruct the point clouds separately, and aggregate these into one dynamic point cloud. +- Ray Outpainting: At a new given pose, the goal is to generate a video consistent with the existing dynamic point cloud and a given scene prompt. We first rasterize the point cloud at the new pose, which is consistent by construction. That said, the rasterization only fills in part of the image, so the question is how to outpaint the rest consistently. To do so, we propose to guide the outpainting process with 3D information by providing the model with ray depth in the observed region and the distance of rays to the point cloud in the unseen region. +- Dynamic Scene Update: We update the dynamic point cloud using the depth maps of the video frames from the new pose. We repeat the steps in §3.2 and §3.3 to iteratively update the dynamic point cloud from novel poses. + +# 3.1. Dynamic Scene Initialization + +We begin by describing how to generate an initial dynamic point cloud given a video at a fixed camera pose $\boldsymbol{\Pi}^{(0)}$ or an image with a motion prompt. Notation-wise, a superscript $(i)$ denotes objects related to the $i^{\mathrm{th}}$ camera pose $\boldsymbol{\Pi}^{(i)}$ . + +For illustration, suppose we are given an image $I_0^{(0)} \in \mathbb{R}^{h \times w \times 3}$ , where $h, w, 3$ are respectively height, width and number of color channels, as well as a prompt describing the desired motion for the video. To generate a video at the fixed pose $\Pi^{(0)}$ , we employ a pretrained image-to-video diffusion model [54] by providing both $I_0^{(0)}$ and the motion prompt, where the motion prompt is further prepended with Camera is strictly fixed. We let $I^{(0)} = \{I_t^{(0)}\}_{t=0}^{N-1}$ denote the generated video, where $I_t^{(0)}$ denotes its $t^{\text{th}}$ frame. + +Given the generated or user-input fixed viewpoint video $\pmb{I}^{(0)}$ , we reconstruct the underlying dynamic scene, represented as a 4D point cloud $\mathcal{P}^{(0)} = \{\pmb{p} = (\pmb{x}, t, \pmb{c})\}$ , where $\pmb{x} \in \mathbb{R}^3$ , $t \in \mathbb{R}$ and $\pmb{c} \in \mathbb{R}^3$ denote the 3D position, times-amp and color respectively. Since the objects in the video foreground and the scene in the background have different dynamic natures, our framework reconstructs point cloud $\mathcal{P}^{(0,\mathrm{f})}$ for the foreground scene and $\mathcal{P}^{(0,\mathrm{b})}$ for background separately. To this end, we extract the binary video foreground masks $M^{(0)} = \{M_t^{(0)} \in \{0,1\}^{h \times w}\}_{t=0}^{N-1}$ with the foreground segmentation method [18] to segment the video $I^{(0)}$ into foreground video $I^{(0,\mathrm{f})} = \{I_t^{(0,\mathrm{f})}\}_{t=0}^{N-1}$ and background video $I^{(0,\mathrm{b})} = \{I_t^{(0,\mathrm{b})}\}_{t=0}^{N-1}$ . We employ the depth model [37] on video $I^{(0)}$ to obtain the video depth maps + +![](images/04dae07582f9d829c7fb681cb0f7725da599518128e34412a179881e0196e107.jpg) +Figure 4. Qualitative ablation studies of the proposed foreground layer (§3.1) and background completion (§3.3) strategy. + +$\pmb{D}^{(0)} = \{\pmb{D}_t^{(0)} \in \mathbb{R}^{h \times w}\}_{t=0}^{N-1}$ . Next, we describe the foreground and background scene initialization steps. + +Scene foreground initialization. We exploit the depth maps $D^{(0)}$ and foreground masks $M^{(0)}$ to obtain the foreground depth maps $D^{(0,\mathrm{f})} = \{D_t^{(0,\mathrm{f})}\}_{t=0}^{N-1}$ . Since the video is captured with a fixed camera, we initialize the video camera pose $\Pi^{(0)}$ as the center of the dynamic scene. We reconstruct the initial foreground point cloud by employing the foreground video frames $I^{(0,\mathrm{f})}$ , foreground depth maps $D^{(0,\mathrm{f})}$ and the initial camera pose $\Pi^{(0)}$ , + +$$ +\mathcal {P} _ {t} ^ {(0, \mathrm {f})} = \phi ([ \boldsymbol {I} _ {t} ^ {(0, \mathrm {f})}, \boldsymbol {D} _ {t} ^ {(0, \mathrm {f})} ], \boldsymbol {\Pi} ^ {(0)}, t), t \in \{0, \dots , N - 1 \}, (1) +$$ + +where $\mathcal{P}_t^{(0,f)}$ denotes the foreground point cloud at timestamps $t$ . $\phi$ denotes the mapping from an image with its depth map to the point cloud. We obtain the foreground point cloud at all timestamps $\mathcal{P}^{(0,f)} = \bigcup_{t=0}^{N-1} \mathcal{P}_t^{(0,f)}$ . + +Scene background initialization. Although the background geometry of a dynamic scene remains static over time, its appearance can vary across frames due to dynamic textures such as flowing water, moving clouds, or changing lighting. Hence, we model the background as having constant depth but varying color across time. Also, notice that foreground objects may occlude parts of the background at certain timestamps, making some background pixels unobservable at those frames. To address this, we leverage foreground masks to identify non-occluded regions and compute the background depth at each pixel location. Specifically, let $D_{t}^{(0)}(x,y)$ and $M_{t}^{(0)}(x,y)$ denote the depth and foreground mask values at pixel $(x,y)$ of the video frame $I_{t}^{(0)}$ . The background depth can be computed element-wise as + +$$ +\boldsymbol {D} ^ {(0, \mathrm {b})} (x, y) = \frac {\sum_ {t = 0} ^ {N - 1} \boldsymbol {D} _ {t} ^ {(0)} (x , y) \cdot \left(1 - \boldsymbol {M} _ {t} ^ {(0)} (x , y)\right)}{\sum_ {t = 0} ^ {N - 1} \left(1 - \boldsymbol {M} _ {t} ^ {(0)} (x , y)\right)}, \tag {2} +$$ + +where $D^{(0,\mathrm{b})}$ is the refined background depth map. We employ the same mapping function $\phi$ to map the background regions on the video frames to the 4D point clouds, + +$$ +\mathcal {P} _ {t} ^ {(0, \mathrm {b})} = \phi ([ \boldsymbol {I} _ {t} ^ {(0, \mathrm {b})}, \boldsymbol {D} ^ {(0, \mathrm {b})} ], \boldsymbol {\Pi} ^ {(0)}, t), t \in \{0, \dots , N - 1 \}. \tag {3} +$$ + +In this way, the obtained background point clouds $\mathcal{P}^{(0,b)} = \bigcup_{t=0}^{N-1} \mathcal{P}_t^{(0,b)}$ have time-invariant positions and time-varying colors. Finally, we obtain the scene point cloud by + +merging the foreground and background point clouds, + +$$ +\mathcal {P} ^ {(0)} = \mathcal {P} ^ {(0, f)} \cup \mathcal {P} ^ {(0, b)}, \tag {4} +$$ + +which is then introduced for outpainting the scene at a novel view with 3D consistent motions. + +# 3.2. Ray Outpainting at a Novel View + +To outpaint the scene at a given novel pose $\Pi^{(1)}$ , we follow the assumption of previous works [21, 57, 58] that the view of $\Pi^{(1)}$ overlaps moderately with that of $\Pi^{(0)}$ to allow for outpainting a substantial portion of the scene. The question is, how can we outpaint a dynamic scene containing consistent motions with the existing dynamic point cloud $\mathcal{P}^{(0)}$ ? + +To address this question, we first rasterize the point cloud at $\Pi^{(1)}$ . As shown in the top row of Figure 5, we obtain a partial video whose observed region is the projection of the portion of the point cloud visible from $\Pi^{(1)}$ . To handle the unseen region, a naive approach would be to directly outpaint the partial video. However, this typically leads to inconsistencies across the boundary of observed and unseen regions, as shown in Figure 5. Our intuition is that pixels in each video frame only provide 2D information, which is not enough to reconstruct the 3D motion. To address this issue, we consider the 3D ray from the camera origin through each pixel. If the pixel is in the observed region, we compute its depth. If the pixel is in the unseen region, we compute the distance from its ray to the point cloud. Both the depth and distance maps serve as guidance for how to outpaint a pixel while respecting information from the existing point cloud. + +Ray information sampling in observed region. Given a novel pose $\Pi^{(1)}$ , we first rasterize the dynamic point cloud $\mathcal{P}^{(0)}$ : at every time step $t \in \{0, \dots, N-1\}$ , we compute + +$$ +\left(\hat {\boldsymbol {I}} _ {t} ^ {(1)}, \hat {\boldsymbol {D}} _ {t} ^ {(1)}, \hat {\boldsymbol {M}} _ {t} ^ {(1)}\right) = \varphi \left(\mathcal {P} _ {t} ^ {(0)}, \boldsymbol {\Pi} ^ {(1)}\right). \tag {5} +$$ + +Here, $\varphi$ denotes the image rasterization function. $\hat{M}_t^{(1)}$ is a binary mask, where $\hat{M}_t^{(1)}(x,y) = 1$ if at least one 3D point from $\mathcal{P}_t^{(0)}$ hits the image at pixel $(x,y)$ during rasterization and 0 otherwise. Correspondingly, $\hat{I}_t^{(1)}$ is the rasterized partial video and $\hat{D}_t^{(1)}$ the rasterized ray depth map (to be distinguished from the depth map generated by depth models as in §3.1). A hat on the notation emphasizes that the signals are not observed for all locations, and we denote $\hat{D}^{(1)} = \{\hat{D}_t^{(1)}\}_{t=0}^{N-1}$ , $\hat{M}^{(1)} = \{\hat{M}_t^{(1)}\}_{t=0}^{N-1}$ and $\hat{I}^{(1)} = \{\hat{I}_t^{(1)}\}_{t=0}^{N-1}$ . A few remarks are in order. First, by doing the rasterization, the observed regions are obviously consistent with the point cloud. Second, the ray depth maps encode the 3D information of the observed scene, which as we shall see will be included as guidance for outpainting. Third, for the regions not observed at $\Pi^{(1)}$ , we introduce the distance between rays and the point cloud to complement the 3D information in these regions, as described next. + +![](images/e8cd9c01ce2c3e35166a910fbb6e3905034096fb9aa20203dc5578a870b49622.jpg) +Figure 5. Ablation studies of scene outpainting with ray information. We also compare with the 2D outpainting model MOTIA [46]. Detailed visualization demonstrates that with ray information, our model successfully outpaints the dynamic scenes with consistent motions. + +Ray-point-cloud distance computation in unseen region. Fix the reference coordinate system at the origin of a camera. Let $\pmb{r} \in \mathbb{R}^3$ be a unit vector denoting the direction of a ray from the origin pointing at a pixel location, and $\pmb{p} \in \mathbb{R}^3$ a point in the 3D space. We recall that the distance between the ray and the 3D point can be computed as + +$$ +\operatorname {d i s t} _ {\mathrm {r} 2 \mathrm {p}} (\boldsymbol {r}, \boldsymbol {p}) = \sqrt {\| \boldsymbol {p} \| _ {2} ^ {2} - \left(\boldsymbol {r} ^ {\top} \boldsymbol {p}\right) ^ {2}}. \tag {6} +$$ + +With the above said, we can compute the distance between camera rays and the 4D point cloud for the complementary 3D information of the unseen regions. Namely, we compute + +$$ +\bar {\boldsymbol {D}} _ {t} ^ {(1)} (x, y) = \min _ {\boldsymbol {p} \in \mathcal {P} _ {t} ^ {(0)}} \operatorname {d i s t} _ {\mathrm {r} 2 \mathrm {p}} \left(\boldsymbol {r} ^ {(1)} (x, y), \boldsymbol {p} - \boldsymbol {o} ^ {(1)}\right), \tag {7} +$$ + +where $\pmb{r}^{(1)}(x,y) \in \mathbb{R}^3$ is the unit-norm ray vector starting from the origin of $\pmb{\Pi}^{(1)}$ pointing at pixel $(x,y)$ , and $\pmb{o}^{(1)}$ is the camera center of $\pmb{\Pi}^{(1)}$ , both in the reference coordinate system of $\pmb{\Pi}^{(0)}$ . As we will see, such a distance will be used to guide the outpainting. + +Ray outpainting. To outpaint the dynamic scene at the novel pose $\Pi^{(1)}$ from the partial video, we desire a few properties for the outpainted video: 1) the outpainted video needs to have a static pose fixed at $\Pi^{(1)}$ and be consistent with the existing dynamic point cloud in the observed region, 2) the user should be able to control the generation + +![](images/44f7f3a9817971c86c0e5727d25a410aabee2ed25050720d98f4b1c57ada673e.jpg) +Input Image + +![](images/4a2d2a785b052940b4e0116b1ad7674d1adf59ad7f8ad190fa747319ac842741.jpg) +Rotate Left + +![](images/a866ee03201a1cd9aae926504528ff303679643730babb22ce8aee1f40d1a9fd.jpg) +Outpaint Scene + +![](images/927baace3dafc467a11860284e4650ff78ca06144c8e88aefabed3635b14b542.jpg) +Time Start + +![](images/43f5bcf3e7e3fc2ff620afd14244c9eec37b8715129a2dde2cb2ba7a5767435c.jpg) +$\frac{1 + u}{7} = \frac{u}{7} + 1 > \frac{\left( {1 + u}\right) u}{7} = u$ + +![](images/4d927ac3adbb593855a27d79e7b21b5486578e5e32b009e5c8a9353b62bcd5ac.jpg) + +![](images/9f0b2851f3233e5e47b3190f3e69adf0a6f6c9a1499327ca3b25a2e43a54c356.jpg) +Time End + +![](images/e6a1f247da9a31cc8e4f392f4bd7f7814ee49a30620c1f4afc61bd13cdf3a318.jpg) +Rotate Left + +![](images/99dceb9ef154b5e38bd674d3c5c71766a154a8a12d59d0d3759afd243bcbe479.jpg) +Initial Point Cloud + +![](images/19b1c77c4db4ed46c85a15f54cf30f0f8b76f6617dae934019e38ee2e3803143.jpg) +Outpaint Scene + +![](images/daa3e4401904b608e6116fc1377e3c741d5f0990ae8b86ae296d39bf98555dc8.jpg) +Move Back + +![](images/2251113b1e62d9b867c00db4b237a23a60da806243d9c9c9aab3d47956c37ccd.jpg) +Outpaint Scene + +![](images/5bcb998387083a16c496d121c64fb63d0bc95b049eac194dbf5aea49c586c0cf.jpg) +Move Back + +![](images/8590a809ccfa415cb9d975b3081ea7778295028cc4eb1a647804da29782c0960.jpg) +Outpaint Scene + +![](images/898210030d66f65955f5bbacfaa7128c611673b0730891de39d9d85b8f4e2d2e.jpg) +Rotate Left + +![](images/6d223c13ea1d15914875c40fc6f1e0510c4edb53313c4b84fc5c4686179572d0.jpg) +Outpaint Scene + +![](images/8ca600fd6fbf40ac8813a4f84793df7647481505d078a99f8f786d074ac0b36e.jpg) + +![](images/8f89d497b80ef71a3bf12b0ac9bb057952a52d9d3e58705dd8580b7ae4a07dad.jpg) +Rotate Left + +![](images/3b03072166f49cd2f49cd0e461ce24fbcf24b571e0d214ac40e1180874ed9c63.jpg) + +![](images/8e1d5d061ce16c2bd71c17e5994aeea9c128609b348ca747ad6e1f1295948b8a.jpg) +Time Start + +![](images/11080c2dd8feb84b89749292e890ed23b5802dddd9e74f05100005274d45e116.jpg) +${180}^{ \circ }$ +${180}^{ \circ }$ + +![](images/a0070c489c16c235f0a50970c0ad7efdc4f93398408cee38a842f91a4b24bdbc.jpg) + +![](images/de6cf8db4add38fac0cd85a4a473424964935728ad90cd8fd87ac9ed1dd91f8f.jpg) +# + +![](images/dcab1abf1cad0ed4feef4d5bc976a43ffd28f8ddceaec2c8bf9170080f4c9742.jpg) +Time End + +![](images/1b6d0284552f57f5309952792b852a529a0271bd6e07e0eeb884edcb0ee30e74.jpg) +Input Video +Initial Point Cloud +Figure 6. Voyaging into perpetual dynamic scenes with long camera trajectories from a single image or fixed viewpoint video. We iteratively show the unseen regions through large camera motions, the dynamic scene outpainting results and the consistent motions from novel views. By enriching the pixel information with ray contexts, our method successfully builds large dynamic scenes from input views. + +![](images/d0151d3ba92031448bd20df6e27b0df3c664eb0c59b387454f44d137b17b947b.jpg) +Move Back + +![](images/eb8373a0a710d4fc379486c6c8fcbb229cc4d63df62225bd629a6bb4e6fede46.jpg) +Outpaint Scene +Outpaint Scene + +![](images/e280bcc3229a00de9574f096f3426a751a121389c9813d953cc71ffaa7efd526.jpg) +Time Start + +![](images/179f47f5081961eadff632d69aaa71fd4c7374362d0a84abaa252f7dd4cf7e49.jpg) +$\frac{1 + u}{7} = \frac{u}{7} + 1 > \frac{\left( {1 + u}\right) u}{7} = \frac{u}{7} + 1 > \frac{u}{\left( {1 + u}\right) u} = \frac{u}{2}$ +$\left( {x - {2x}}\right) t - x{y}^{2} = \left( {x - {2x}}\right) {f}^{\prime }t$ + +![](images/54a0dd0b9f4d9c78dcf8f97233b90d3ede8dd5996fd6c4b9191325f279b95efc.jpg) + +![](images/f1f580d291db9d8729374637c2b86e8f92edc5e50bf0b9de735e752d629dd2b0.jpg) +# + +![](images/d2592184f47fe3e8c3ee513435916527e1b29166ce81ad51de1d10dfc1b737b5.jpg) +Time End + +in the outpainted region (i.e., the unseen region), and 3) the motions in the outpainted region need to be consistent with the motions in the observed region. To fulfill properties 1) and 2), we design our video outpainting model to take as input the extracted partial video $\hat{I}_t^{(1)}$ , as well as a prompt narrating the desired motions in the generated scene. For property 3), we propose to employ ray depth map $\hat{D}^{(1)}$ and ray distance map $\bar{D}^{(1)}$ as guidance on 3D information in the observed region and how much the outpainted ray in the unseen region should respect the existing point cloud. To sum up, our video outpainting model accepts the partial video $\hat{I}^{(1)}$ , scene prompts, ray depth map $\hat{D}^{(1)}$ , ray distance map $\bar{D}^{(1)}$ and generates an outpainted video $I^{(1)} = \{I_t^{(1)}\}_{t=0}^{N-1}$ . Next we describe how to train such a model. + +Training the outpainting model. We collect a small video dataset for training (details in supplementary material), and reconstruct point clouds from it to compute ray information for supervision. More specifically, given a training video $\pmb{I}$ we extract depth maps $\pmb{D}$ using the depth model [37] and backproject the video from a camera $\pmb{\Pi}$ at the world center to reconstruct a point cloud. We then move the camera closer to the scene and filter out the 3D points that are no longer visible from the current viewpoint. Then we move the camera back to the original camera pose $\pmb{\Pi}$ and com + +pute a partial video $\hat{I}$ , a partial depth map $\hat{D}$ , and a ray distance map $\bar{D}$ following §3.2. To train the model, we add noise to $\pmb{I}$ to obtain $z_{\tau}$ at step $\tau$ , and employ ControlNet [59] to inject ray information. Let $\theta_{\mathrm{v}}, \theta_{\mathrm{c}}$ denote the training parameters of the video diffusion model and ControlNet model. The training objective is + +$$ +\mathcal {L} = \mathbb {E} _ {z _ {0}, I, \hat {I}, \tau , \hat {D}, \bar {D}, \mathbf {s}, \epsilon} \left\| \epsilon - \epsilon_ {\theta_ {\mathrm {v}}, \theta_ {\mathrm {c}}} (I, \hat {I}, z _ {\tau}, \tau , \hat {D}, \bar {D}, \mathbf {s}) \right\| _ {2} ^ {2}, \quad (8) +$$ + +where $\epsilon \sim \mathcal{N}(0,1)$ is the diffusion noise. + +# 3.3. Dynamic Scene Update + +Now that we have an outpainted video $\mathbf{I}^{(1)}$ at the new pose $\Pi^{(1)}$ , the next step is to produce a new dynamic point cloud and merge it with the previous one $\mathcal{P}^{(0)}$ . + +Point cloud merging. We exploit the depth model [37] to obtain the video depth maps $D^{(1)}$ of the outpainted video $I^{(1)}$ . Noting that there is a depth inconsistency problem between the estimated depths and the rendered ray depth map $\hat{D}^{(1)}$ in the mask regions $\hat{M}^{(1)}$ , we follow common practices [57, 58] to align the video depth maps $D^{(1)}$ with the ray depth maps $\hat{D}^{(1)}$ by finetuning the depth estimation model. After obtaining the aligned depth maps from the finetuned depth model, we follow §3.1 to initialize foreground point cloud $\mathcal{P}^{(1,f)}$ and background point cloud + +![](images/a635fa1bc6146983a2f33417b0fc7cebce11601bd738b3d35d7471d77e6a9fe2.jpg) +Input View + +![](images/46f0435dc8cc050c4a43b11f2cf484f983af372050833ccd0446fbbc3722b7ef.jpg) + +![](images/e0ac863280b5cc2815278d2da6beeb9c5a1491a5906cea17e2b5aa262370be1f.jpg) + +![](images/fe3ed1d46afc9446434c8e53eba1328672926cfec94bf5c1e81e21007a23418c.jpg) +View extrapolation results from WonderJourney + +![](images/fe1728ca5a46297d77bfa371a3f9d91f968b9caf6c3f8027e43f7645542079fa.jpg) + +![](images/8d31968c606f9810b6c28d59ba075a5a4b19e877ac129ee310cebb541001f445.jpg) + +![](images/118cac2e8be7083e6f8d7e6dda1eee93ae69fe3c884e68ce12d133907f1367b7.jpg) + +![](images/2295bd888b846de5a0b42fe49be07f41b2a9fb4f30fe6510879c742f707129ba.jpg) + +![](images/144d4cc0dd4bce40b17a7bc748f24d3447f7597f8c6f627ca12293bf7fc175ce.jpg) + +![](images/f4fa17d1b73b0961e6a76013ad9545a12a268cc596dad9a573b3601a2040b293.jpg) +Target View + +![](images/7bcb765e9b0d3222b4e80dc4e22d94f0328fce2105c4c906efe278b8a4d365ca.jpg) + +![](images/6853a6c9d51f9787fe3aa10aa00fc515b2bfbe0137a2682a11400e358cab7c01.jpg) + +![](images/e06857b9aa9716d201fd3f6ef9691d054f3bf575dab95f271f7ca2aaad96ea58.jpg) +View extrapolation results from ours + +![](images/74401b8dcc21c124d98955704090d24564d6f9dd5cf1773346b83d65031720ef.jpg) + +![](images/08430a03e8c7f521afe192b30a9aecb48dda8ba9c35c161734ca6bf2848bf0d7.jpg) + +![](images/56f7aaa65b56a7a3db7fc9dbd58021c11620e80650e1958c638c3e7dd0ab5a16.jpg) + +![](images/e759cc285bc6a581dacfdc93011f89e89fd7a71d233b92d324f85a91db9274ce.jpg) + +![](images/5b13139c0a13175be579892329c351e2998bfaacc5d1bcc7d7a6235d25e48c8e.jpg) + +![](images/9cb368f25c8f11e466861b037b49805d09845b8df8ab8dee5d805e5df47e4a01.jpg) +Figure 7. View extrapolation results. While Wonderjourney [58] only extrapolates views of a static character in 3D space, DynamicVoyager renders extrapolated views with 3D consistent motions by leveraging the ray information for scene outpainting. + +![](images/810ba339a45c066ec24c540da800cbacc4ebf6e0aff1b04a183bf412c0012d56.jpg) +Table 1. Quantitative results of controllable dynamic scene generation on the dynamic degree (DD), factual consistency (FC) [12] and CLIP [36] scores. While WonderJourney [58] only controls static scene generation with scene prompts, our model achieves controllable dynamic scene generation with higher performance. + +![](images/8c72e2642b9e69d000a4c905f2e75dc10c3a0555269cec1664f3049281e9007b.jpg) +Figure 8. Visual comparisons with video diffusion model MotionCtrl [48]. Full details are in supplementary material. DynamicVoyager outpaints the scenes with large camera motions. + +![](images/2bd824a5f26a2567f673a0ba344f5fd3f2e0e9ac4f3327c1417db8c7d9709f6c.jpg) +MotionCtrl + +![](images/0907163f55102f6d1c50fccdae2216c821a1687be6c23605bd052e7e61d01534.jpg) + +![](images/546129943dd66590a800ad103c6405894c4c3b981930c9fa7be9e305c8ef1515.jpg) +Ours + +![](images/eb48d2383f59cc07788baf24b74b69e05d553068fa06cf8812b8e0a81f9b6419.jpg) + +$\mathcal{P}^{(1,\mathrm{b})}$ . Then we merge the new point clouds with the previous one to update the dynamic scenes, + +$$ +\mathcal {P} ^ {(1)} = \mathcal {P} ^ {(1, f)} \cup \mathcal {P} ^ {(1, b)} \cup \mathcal {P} ^ {(0)}, \tag {9} +$$ + +where $\mathcal{P}^{(1)}$ is the updated point cloud at the pose $\Pi^{(1)}$ . In this way, we could generate perpetual dynamic scenes along fly-through camera trajectories by looping §3.2 and §3.3 with new camera poses and scene prompts. + +Background completion. In practice, we find that the missing background point clouds occluded by the moving foreground in the scene cause inconsistency when rendering along fly-through camera trajectories (as shown in Figure 4). We hence propose to employ our video outpainting model to generate background videos of the occluded parts. After that, we extract depth maps of the background videos and reconstruct the background point clouds with the extracted depth maps. Details are in supplementary material. We also update the scene with these point clouds. + +# 4. Experiments + +In this section, we first describe the details of our experimental setup, we then discuss the results of the scene gener- + +
methodsWonderJourney [58]ours
CLIP-SIM ↑ / DD ↑ / FC ↑24.98 / 2.23 / 1.3025.23 / 3.13 / 1.32
+ +ation with fly-through cameras and controllable scene generation. Finally, we conduct ablation studies of our model. + +# 4.1. Experiment Setups + +Data. We trained our outpainting model with the OpenVid dataset [33], which contains high-resolution videos with detailed video prompts. In order to train with daily dynamic scene videos, we sampled 5,000 videos that contain outdoor dynamic scenes, including natural scenes (such as waterfalls, rivers and moving clouds) and urban scenes (such as walking people and moving cars). Additional data curation details can be found in the supplementary material. + +Training details. We employed the LoRA [16] to fine-tune the CogvideoX-5B-I2V model [54]. The LoRA rank was 256 and the learning rate was $1 \times 10^{-4}$ with a cosine schedule. The video resolution is $16 \times 512 \times 512$ , and hence $N = 16$ and $h = w = 512$ . For calculating the ray information, we used the depth model [37] to extract the depth map of each video frame. Then, we followed the state-of-the-art video outpainting model MOTIA [46] and exploited the same mask proportion for outpainting. + +Camera details. To set up dynamic scenes with fly-through camera trajectories, we followed the common practices in [21, 35, 58], where cameras move along a straight line or rotate. In the former case, we moved the camera backward by 0.0005 units between two adjacent camera poses, where one unit corresponds to the normalized 3D coordinate defined in PyTorch3D, and in the latter case, we rotated the + +![](images/d1e20ef236422b5b7bc85d004305d5c155ffb525db6bb59f27adaa2867270328.jpg) +Figure 9. Controllable scene generation from input images with scene prompts. Full details are shown in supplementary material. DynamicVoyager successfully controls the dynamic scene outpainting content with the corresponding scene prompt. + +camera by 0.45 radians. We generated camera paths by linearly interpolating translation and approximating the rotation using uniform angular steps. + +# 4.2. Scene Generation with Fly-through Cameras + +Figure 6 shows the dynamic scene generation results with long fly-through trajectories. We tested our model on both the background dynamics (water flowing in the waterfall) and the foreground dynamics (person walking on the path). Compared to the initial point cloud, our model generated perpetual dynamic scenes by exploiting the proposed scene outpainting process. As there is no direct baseline for perpetual dynamic scene generation, we compare our method with the perpetual static scene generation method WonderJourney [58] in Figure 7, state-of-the-art 4D scene generation model CAT4D [49] in Figure 2 and camera-controllable video diffusion model MotionCtrl [48] in Figure 8. These figures demonstrate that our model can better generate perpetual dynamic scenes with 3D consistent motions by exploiting our ray outpainting model. + +# 4.3. Scene Generation Controlled by Text + +To test the controllable dynamic scene generation ability of our model, we generated dynamic scenes from village and volcano images with scene prompts. As shown in Figure 9, for the village image, we control the outpainting content by introducing two different text prompts: waterfall and huge lake. For the volcano image, we prompted the model to generate glacier and magma. Our model successfully controls the perpetual dynamic scene generation with the given scene prompts. Following the common practices in [21, 56-58], we employed the CLIP [36] scores between the novel view images and the input scene prompts and the dynamic degree (DD) and factual consistency (FC) indices + +Table 2. Quantitative results of 3D consistent motions. VQ, TC and TA respectively denote the visual quality, temporal consistency and text-to-video alignment matrices [12]. Our model achieves better performance by achieving 3D consistent motions. + +
methodsCLIP-SIM ↑VQ ↑TC ↑TA ↑FC ↑
MOTIA [46]23.802.2811.8522.6561.859
w/o ray info23.122.5621.9842.6721.977
DynamicVoyager24.702.5782.2342.9382.188
+ +Table 3. Quantitative ablation studies of the proposed foreground layer (§3.1) and background completion (§3.3) strategy. + +
methodsCLIP-SIM ↑VQ ↑TC ↑TA ↑FC ↑
w/o fg layer23.881.8910.8712.2191.102
w/o bg completion24.621.9060.8712.5471.195
DynamicVoyager25.481.9771.0312.5621.320
+ +proposed by VideoScore [12] to thoroughly evaluate the generation quality of the video. Table 1 shows that, compared to WonderJourney, our model achieves better performance by controlling dynamic scene generation. + +# 4.4. Ablation Study + +Ray outpainting. To fairly evaluate the proposed ray outpainting model, we conducted experiments on five scenes containing common dynamic objects: cars, people, water and clouds. For each scene, we sampled 16 novel view images from the generated dynamic point cloud at 16 different timestamps along the fly-through camera trajectories. As there are no 3D video outpainting baselines, we chose the 2D outpainting model MOTIA [46] for comparisons. Figure 5 shows that by introducing the ray information for outpainting, our model successfully outpaints 3D consistent motions at the outpainting boundaries. In this way, our ray outpainting model achieves better performance in Table 2. + +4D representation. We tested the proposed foreground representation and background completion strategy in Figure 4 and Table 3. It can be seen that our model renders consistent foregrounds with foreground layers and generates a plausible background with the background completion strategy. + +# 5. Conclusion + +We proposed DynamicVoyager, a new model for perpetual dynamic scene generation, which reformulates this task as an outpainting problem. By considering pixels as rays and conditioning on ray depth and distance maps, DynamicVoyager can generate scenes with 3D consistent motions. Experimental results demonstrate the superiority of our model. + +Limitations. Challenges remain in handling reflections, shadows, fine structures, multi-view rendering, and depth discontinuities, which can be explored in future work. + +# Acknowledgments + +The authors sincerely thank Prof. Kostas Daniilidis, Dr. Jiahui Lei, Qiao Feng, Chen Wang, Ziqing Xu, and Uday Kiran for their generous and insightful feedback on this work. We also acknowledge the support of a Penn Engineering graduate fellowship, Penn startup funds, and the Research Collaboration on the Mathematical and Scientific Foundations of Deep Learning under grants NSF 2031985 and Simons 814201. + +# References + +[1] Sherwin Bahmani, Ivan Skorokhodov, Aliaksandr Siarohin, Willi Menapace, Guocheng Qian, Michael Vasilkovsky, Hsin-Ying Lee, Chaoyang Wang, Jiaxu Zou, Andrea Tagliasacchi, et al. VD3D: Taming large video diffusion transformers for 3D camera control. In ICLR, 2025. 3 +[2] Miguel Angel Bautista, Pengsheng Guo, Samira Abnar, Walter Talbott, Alexander Toshev, Zhuoyuan Chen, Laurent Dinh, Shuangfei Zhai, Hanlin Goh, Daniel Ulbricht, et al. Gaudi: A neural architect for immersive 3D scene generation. In NeurIPS, 2022. 2 +[3] Michael Broxton, John Flynn, Ryan Overbeck, Daniel Erickson, Peter Hedman, Matthew Duvall, Jason Dourgarian, Jay Busch, Matt Whalen, and Paul Debevec. Immersive light field video with a layered mesh representation. In ToG, 2020. 2 +[4] Alvaro Collet, Ming Chuang, Pat Sweeney, Don Gillett, Dennis Evseev, David Calabrese, Hugues Hoppe, Adam Kirk, and Steve Sullivan. High-quality streamable free-viewpoint video. In ToG, 2015. 2 +[5] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. In CVPR, 2023. 3 +[6] Terrance DeVries, Miguel Angel Bautista, Nitish Srivastava, Graham W Taylor, and Joshua M Susskind. Unconstrained scene generation with locally conditioned radiance fields. In ICCV, 2021. 2 +[7] C Gan, J Schwartz, S Alter, M Schrimpf, J Traer, J De Freitas, J Kubilius, A Bhandwaldar, N Haber, M Sano, et al. ThreeDWorld: A platform for interactive multi-modal physical simulation. In NeurIPS, 2021. 2 +[8] Ruiqi Gao*, Aleksander Holynski*, Philipp Henzler, Arthur Brussee, Ricardo Martin-Brualla, Pratul P. Srinivasan, Jonathan T. Barron, and Ben Poole*. CAT3D: Create anything in 3D with multi-view diffusion models. In NeurIPS, 2024. 2 +[9] Jiayuan Gu, Fanbo Xiang, Xuanlin Li, Zhan Ling, Xiqiang Liu, Tongzhou Mu, Yihe Tang, Stone Tao, Xinyue Wei, Yunchao Yao, Xiaodi Yuan, Pengwei Xie, Zhiao Huang, Rui Chen, and Hao Su. ManiSkill2: A unified benchmark for generalizable manipulation skills. In ICLR, 2023. 2 +[10] Yuwei Guo, Ceyuan Yang, Anyi Rao, Zhengyang Liang, Yaohui Wang, Yu Qiao, Maneesh Agrawala, Dahua Lin, and + +Bo Dai. *AnimateDiff: Animate your personalized text-to-image diffusion models without specific tuning.* In *ICLR*, 2024. 3 +[11] Hao He, Yinghao Xu, Yuwei Guo, Gordon Wetzstein, Bo Dai, Hongsheng Li, and Ceyuan Yang. CameraCtrl: Enabling camera control for text-to-video generation. In ICLR, 2025. 3 +[12] Xuan He, Dongfu Jiang, Ge Zhang, Max Ku, Achint Soni, Sherman Siu, Haonan Chen, Abhranil Chandra, Ziyan Jiang, Aaran Arulraj, Kai Wang, Quy Duc Do, Yuansheng Ni, Bohan Lyu, Yaswanth Narsupalli, Rongqi Fan, Zhiheng Lyu, Yuchen Lin, and Wenhu Chen. VideoScore: Building automatic metrics to simulate fine-grained human feedback for video generation. In EMNLP, 2024. 7, 8 +[13] Lukas Hollein, Ang Cao, Andrew Owens, Justin Johnson, and Matthias Nießner. Text2Room: Extracting textured 3D meshes from 2D text-to-image models. In ICCV, 2023. 2 +[14] Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. CogVideo: Large-scale pretraining for text-to-video generation via transformers. In ICLR, 2022. 2, 3 +[15] Chen Hou, Guoqiang Wei, Yan Zeng, and Zhibo Chen. Training-free camera control for video generation. In ICLR, 2025. 3 +[16] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. LoRA: Low-rank adaptation of large language models. In ICLR, 2022. 7 +[17] Ronghang Hu, Nikhila Ravi, Alexander C Berg, and Deepak Pathak. Worldsheet: Wrapping the world in a 3d sheet for view synthesis from a single image. In ICCV, 2021. 2 +[18] Jitesh Jain, Jiachen Li, Mang Tik Chiu, Ali Hassani, Nikita Orlov, and Humphrey Shi. OneFormer: One transformer to rule universal image segmentation. In CVPR, 2023. 2, 3, 4 +[19] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3D gaussian splatting for real-time radiance field rendering. In TOG, 2023. 2 +[20] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In ICCV, 2023. 2, 3 +[21] Hanyang Kong, Dongze Lian, Michael Bi Mi, and Xinchao Wang. DreamDrone: Text-to-image diffusion models are zero-shot perpetual view generators. In ECCV, 2024. 5, 7, 8 +[22] Zhengfei Kuang, Shengqu Cai, Hao He, Yinghao Xu, Hongsheng Li, Leonidas Guibas, and Gordon Wetzstein. Collaborative video diffusion: Consistent multi-video generation with camera control. In NeurIPS, 2024. 3 +[23] Black Forest Labs. Flux. https://github.com/black-forest-labs/flux, 2024. 2 +[24] Black Forest Labs, Stephen Batifol, Andreas Blattmann, Frederic Boesel, Saksham Consul, Cyril Diagne, Tim Dockhorn, Jack English, Zion English, Patrick Esser, Sumith Kulal, Kyle Lacey, Yam Levi, Cheng Li, Dominik Lorenz, Jonas Müller, Dustin Podell, Robin Rombach, Harry Saini, Axel Sauer, and Luke Smith. Flux.1 kontext: Flow matching for in-context image generation and editing in latent space, 2025. 2 + +[25] Jiabao Lei, Jiapeng Tang, and Kui Jia. RGBD2: Generative scene synthesis via incremental view inpainting using RGBD diffusion models. In CVPR, 2023. 2 +[26] Renjie Li, Panwang Pan, Bangbang Yang, Dejia Xu, Shijie Zhou, Xuanyang Zhang, Zeming Li, Achuta Kadambi, Zhangyang Wang, Zhengzhong Tu, et al. 4K4DGen: Panoramic 4D generation at 4K resolution. In ICLR, 2025. 2, 3 +[27] Zhengqi Li, Qianqian Wang, Noah Snavely, and Angjoo Kanazawa. Infinitenature-Zero: Learning perpetual view generation of natural scenes from single images. In ECCV, 2022. 1, 2 +[28] Huan Ling, Seung Wook Kim, Antonio Torralba, Sanja Fidler, and Karsten Kreis. Align your gaussians: Text-to-4D with dynamic 3D gaussians and composed diffusion models. In CVPR, 2024. 2, 3 +[29] Lu Ling, Yichen Sheng, Zhi Tu, Wentian Zhao, Cheng Xin, Kun Wan, Lantao Yu, Qianyu Guo, Zixun Yu, Yawen Lu, et al. DL3DV-10K: A large-scale scene dataset for deep learning-based 3D vision. In CVPR, 2024. 3 +[30] Andrew Liu, Richard Tucker, Varun Jampani, Ameesh Makadia, Noah Snavely, and Angjoo Kanazawa. Infinite Nature: Perpetual view generation of natural scenes from a single image. In ICCV, 2021. 1, 2 +[31] Yuan Liu, Cheng Lin, Zijiao Zeng, Xiaoxiao Long, Lingjie Liu, Taku Komura, and Wenping Wang. SyncDreamer: Generating multiview-consistent images from a single-view image. In ICLR, 2024. 2 +[32] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 2 +[33] Kepan Nan, Rui Xie, Penghao Zhou, Tiehan Fan, Zhenheng Yang, Zhijie Chen, Xiang Li, Jian Yang, and Ying Tai. Openvid-1M: A large-scale high-quality dataset for text-to-video generation. In ICLR, 2025. 7 +[34] Zijie Pan, Zeyu Yang, Xiatian Zhu, and Li Zhang. Efficient4D: Fast dynamic 3D object generation from a single-view video. arXiv preprint arXiv 2401.08742, 2024. 3 +[35] Stefan Popov, Amit Raj, Michael Krainin, Yuanzhen Li, William T. Freeman, and Michael Rubinstein. CamCtrl3D: Single-image scene exploration with precise 3D camera control. In 3DV, 2025. 2, 7 +[36] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 7, 8 +[37] René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. In IEEE TPAMI, 2020. 2, 3, 4, 6, 7 +[38] Jiawei Ren, Liang Pan, Jiaxiang Tang, Chi Zhang, Ang Cao, Gang Zeng, and Ziwei Liu. DreamGaussian4D: Generative 4D gaussian splatting. arXiv preprint arXiv:2312.17142, 2023. 3 +[39] Xuanchi Ren, Tianchang Shen, Jiahui Huang, Huan Ling, Yifan Lu, Merlin Nimier-David, Thomas Müller, Alexander + +Keller, Sanja Fidler, and Jun Gao. GEN3C: 3d-informed world-consistent video generation with precise camera control. In CVPR, 2025. 3 +[40] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2 +[41] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 2 +[42] Wenqiang Sun, Shuo Chen, Fangfu Liu, Zilong Chen, Yueqi Duan, Jun Zhang, and Yikai Wang. DimensionX: Create any 3D and 4D scenes from a single image with controllable video diffusion. arXiv preprint arXiv:2411.04928, 2024. 2, 3 +[43] Fengrui Tian, Shaoyi Du, and Yueqi Duan. MonoN-erF: Learning a generalizable dynamic radiance field from monocular videos. In ICCV, 2023. 2 +[44] Fengrui Tian, Yueqi Duan, Angtian Wang, Jianfei Guo, and Shaoyi Du. Semantic Flow: Learning semantic fields of dynamic scenes from monocular videos. In ICLR, 2024. 2 +[45] Basile Van Hoorick, Rundi Wu, Ege Ozguroglu, Kyle Sargent, Ruoshi Liu, Pavel Tokmakov, Achal Dave, Changxi Zheng, and Carl Vondrick. Generative camera dolly: Extreme monocular dynamic novel view synthesis. In ECCV, 2024. 3 +[46] Fu-Yun Wang, Xiaoshi Wu, Zhaoyang Huang, Xiaoyu Shi, Dazhong Shen, Guanglu Song, Yu Liu, and Hongsheng Li. Be-your-outpainter: Mastering video outpainting through input-specific adaptation. In ECCV, 2024. 5, 7, 8 +[47] Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. DUS3R: Geometric 3D vision made easy. In CVPR, 2024. 2 +[48] Zhouxia Wang, Ziyang Yuan, Xintao Wang, Yaowei Li, Tianshui Chen, Menghan Xia, Ping Luo, and Ying Shan. Motionctrl: A unified and flexible motion controller for video generation. In ACM SIGGRAPH, 2023. 7, 8 +[49] Rundi Wu, Ruiqi Gao, Ben Poole, Alex Trevithick, Changxi Zheng, Jonathan T Barron, and Aleksander Holynski. Cat4D: Create anything in 4D with multi-view video diffusion models. arXiv preprint arXiv:2411.18613, 2024. 2, 3, 8 +[50] Yiming Xie, Chun-Han Yao, Vikram Voleti, Huaizu Jiang, and Varun Jampani. SV4D: Dynamic 3D content generation with multi-frame and multi-view consistency. arXiv preprint arXiv:2407.17470, 2024. 2, 3 +[51] Jinbo Xing, Menghan Xia, Yong Zhang, Haoxin Chen, Wangbo Yu, Hanyuan Liu, Gongye Liu, Xintao Wang, Ying Shan, and Tien-Tsin Wong. Dynamiccafter: Animating open-domain images with video diffusion priors. In ICLR, 2024. 3 +[52] Jihan Yang, Shusheng Yang, Anjali W. Gupta, Rilyn Han, Li Fei-Fei, and Saining Xie. Thinking in Space: How multimodal large language models see, remember and recall spaces. arXiv preprint arXiv:2412.14171, 2024. 2 +[53] Shiyuan Yang, Liang Hou, Haibin Huang, Chongyang Ma, Pengfei Wan, Di Zhang, Xiaodong Chen, and Jing Liao. + +Direct-a-Video: Customized video generation with user-directed camera movement and object motion. In SIGGRAPH, 2024. 3 +[54] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al. CogVideoX: Text-to-video diffusion models with an expert transformer. In ICLR, 2024. 2, 3, 4, 7 +[55] Yuyang Yin, Dejia Xu, Zhangyang Wang, Yao Zhao, and Yunchao Wei. 4DGen: Grounded 4D content generation with spatial-temporal consistency. arXiv preprint arXiv:2312.17225, 2023. 3 +[56] Heng Yu, Chaoyang Wang, Peiye Zhuang, Willi Menapace, Aliaksandr Siarohin, Junli Cao, Laszlo A Jeni, Sergey Tulyakov, and Hsin-Ying Lee. 4Real: Towards photorealistic 4D scene generation via video diffusion models. In NeurIPS, 2024. 2, 3, 8 +[57] Hong-Xing Yu, Haoyi Duan, Charles Herrmann, William T Freeman, and Jiajun Wu. WonderWorld: Interactive 3D scene generation from a single image. arXiv preprint arXiv:2406.09394, 2024. 2, 5, 6 +[58] Hong-Xing Yu, Haoyi Duan, Junhwa Hur, Kyle Sargent, Michael Rubinstein, William T Freeman, Forrester Cole, Deqing Sun, Noah Snavely, Jiajun Wu, et al. WonderJourney: Going from anywhere to everywhere. In CVPR, 2024. 1, 2, 5, 6, 7, 8 +[59] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, 2023. 6 +[60] Yuyang Zhao, Chung-Ching Lin, Kevin Lin, Zhiwen Yan, Linjie Li, Zhengyuan Yang, Jianfeng Wang, Gim Hee Lee, and Lijuan Wang. GenXD: Generating any 3D and 4D scenes. In ICLR, 2025. 3 \ No newline at end of file diff --git a/ICCV/2025/Voyaging into Perpetual Dynamic Scenes from a Single View/images.zip b/ICCV/2025/Voyaging into Perpetual Dynamic Scenes from a Single View/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..5c9113f15340ece590c1ce7e8833859c42f2c321 --- /dev/null +++ b/ICCV/2025/Voyaging into Perpetual Dynamic Scenes from a Single View/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9453655bafac5d2c2d1a95ec4d66932eb5bed88e98e1e0b0edefe6057b035e41 +size 935452 diff --git a/ICCV/2025/Voyaging into Perpetual Dynamic Scenes from a Single View/layout.json b/ICCV/2025/Voyaging into Perpetual Dynamic Scenes from a Single View/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..c5d6d37402449c0ce3562d5070411d99f69c0f3b --- /dev/null +++ b/ICCV/2025/Voyaging into Perpetual Dynamic Scenes from a Single View/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:243393181aa95810c8542ec63648cf835f1796efbc3e5d148a73c5a427f8f13b +size 559765 diff --git a/ICCV/2025/Vulnerability-Aware Spatio-Temporal Learning for Generalizable Deepfake Video Detection/66ddbb2b-1283-449c-8670-af8534967206_content_list.json b/ICCV/2025/Vulnerability-Aware Spatio-Temporal Learning for Generalizable Deepfake Video Detection/66ddbb2b-1283-449c-8670-af8534967206_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..29abab4953d87a40410dfe3aee62ec1c2fc8a3d7 --- /dev/null +++ b/ICCV/2025/Vulnerability-Aware Spatio-Temporal Learning for Generalizable Deepfake Video Detection/66ddbb2b-1283-449c-8670-af8534967206_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7bf11e3febe746be7aef1b2520107bc4e17ce3fd941b52a170f7f645f719e5ac +size 94676 diff --git a/ICCV/2025/Vulnerability-Aware Spatio-Temporal Learning for Generalizable Deepfake Video Detection/66ddbb2b-1283-449c-8670-af8534967206_model.json b/ICCV/2025/Vulnerability-Aware Spatio-Temporal Learning for Generalizable Deepfake Video Detection/66ddbb2b-1283-449c-8670-af8534967206_model.json new file mode 100644 index 0000000000000000000000000000000000000000..b665de65a87195a15730578e1e6af9ee915689a2 --- /dev/null +++ b/ICCV/2025/Vulnerability-Aware Spatio-Temporal Learning for Generalizable Deepfake Video Detection/66ddbb2b-1283-449c-8670-af8534967206_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a6cd663f61b9b74e1bab1646ee4dee127610cbf3a9aaa645028ab945c1ecf3f3 +size 118479 diff --git a/ICCV/2025/Vulnerability-Aware Spatio-Temporal Learning for Generalizable Deepfake Video Detection/66ddbb2b-1283-449c-8670-af8534967206_origin.pdf b/ICCV/2025/Vulnerability-Aware Spatio-Temporal Learning for Generalizable Deepfake Video Detection/66ddbb2b-1283-449c-8670-af8534967206_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..bf55c98a2546074761e7f57f6cc7243877269baf --- /dev/null +++ b/ICCV/2025/Vulnerability-Aware Spatio-Temporal Learning for Generalizable Deepfake Video Detection/66ddbb2b-1283-449c-8670-af8534967206_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5d3221c853c8ea3feef0d2dcf7133b042aef66b18e020ba5b1796b2c3a0bdd4d +size 7709664 diff --git a/ICCV/2025/Vulnerability-Aware Spatio-Temporal Learning for Generalizable Deepfake Video Detection/full.md b/ICCV/2025/Vulnerability-Aware Spatio-Temporal Learning for Generalizable Deepfake Video Detection/full.md new file mode 100644 index 0000000000000000000000000000000000000000..7c914e00b412e8220d9a6e8975dbf24205a6627f --- /dev/null +++ b/ICCV/2025/Vulnerability-Aware Spatio-Temporal Learning for Generalizable Deepfake Video Detection/full.md @@ -0,0 +1,365 @@ +# Vulnerability-Aware Spatio-Temporal Learning for Generalizable Deepfake Video Detection + +Dat NGUYEN1, Marcella ASTRID1, Anis KACEM1, Enjie GHORBEL1, 2, Djamila AOUADA1, CVI2, SnT, University of Luxembourg1 + +Cristal Laboratory, National School of Computer Sciences, University of Manouba + +{dat.nguyen,marcella.astrid,anis.kacem,djamila.aouada}@uni.lu enjie.ghorbel@isamm.uma.tn + +# Abstract + +Detecting deepfake videos is highly challenging given the complexity of characterizing spatio-temporal artifacts. Most existing methods rely on binary classifiers trained using real and fake image sequences, therefore hindering their generalization capabilities to unseen generation methods. Moreover, with the constant progress in generative Artificial Intelligence (AI), deepfake artifacts are becoming imperceptible at both the spatial and the temporal levels, making them extremely difficult to capture. To address these issues, we propose a fine-grained deepfake video detection approach called FakeStormer that enforces the modeling of subtle spatio-temporal inconsistencies while avoiding overfitting. Specifically, we introduce a multi-task learning framework that incorporates two auxiliary branches for explicitly attending artifact-prone spatial and temporal regions. Additionally, we propose a video-level data synthesis strategy that generates pseudo-fake videos with subtle spatio-temporal artifacts, providing high-quality samples and hand-free annotations for our additional branches. Extensive experiments on several challenging benchmarks demonstrate the superiority of our approach compared to recent state-of-the-art methods. The code is available at https://github.com/10Ring/FakeStormer. + +# 1. Introduction + +With the advances in generative modeling [22, 43], deepfake videos have become alarmingly realistic. Despite their interest in several applications, such as entertainment and education, this type of technology also raises societal concerns [5, 45, 53]. There is therefore an urgency for developing effective deepfake detection methods. + +In the literature, several deepfake detection techniques aim to model spatial artifacts by treating each frame independently [3, 6, 7, 15, 33, 39, 41, 42, 46, 48, 69]. While this is reasonable when dealing with frame-level genera + +![](images/cca0d112633abdafc5f141aaaa00707244e1efbf6f42602e03d482666acfb4d0.jpg) + +![](images/23dd47657ef58849d6b93d338960e11d0c7dfdb280666aaef44ca409dbd421e6.jpg) + +![](images/8bd48a4fd90603e7649efb7afaf0f4f0dbb29fc1d6fa923d0bc55e170528c448.jpg) +Figure 1. a) Traditional video-based methods [20, 23, 25, 56, 57, 64, 67, 71] versus b) the proposed multi-task learning framework; c) Visualization of the temporal vulnerabilities. Note that only some temporal locations are shown. + +tion methods [32, 51, 73], it becomes less adequate in the presence of video-level manipulation techniques [1, 19, 58], where temporal and spatial artifacts are intertwined. + +For that reason, researchers have explored video-based deepfake detection methods capable of modeling spatiotemporal artifacts [20, 23, 25, 56, 64, 67, 71]. Those methods mainly rely on a deep neural network formed by a single binary classification branch that is trained using a fixed dataset with real and fake data (see Figure 1-a). As a result, they suffer from two main limitations, namely: (1) The lack of generalizability - As highlighted in [33, 39, 42, 48], models trained with a standalone binary classifier tend to overfit the type of deepfakes they are trained on, resulting in poor generalization to unseen manipulations; (2) The lack of robustness to high-quality (HQ) deepfake videos - The + +quality of deepfake videos is improving continuously, resulting in subtle spatio-temporal artifacts. As such, vanilla single-branch architectures trained solely with binary supervision fail to fully capture them, necessitating the design of appropriate attention mechanisms. + +To address the generalization issue, video-level data synthesis approaches [24, 34, 62] have been introduced to encourage models to learn more generic representations. However, these methods usually simulate exaggerated temporal variations that are inherently different from artifacts in hyper-realistic deepfake videos. On the other hand, to model localized spatio-temporal inconsistencies, some recent methods [63, 67] have used dedicated architectures integrating an implicit attention mechanism. Nevertheless, these models still rely solely on a binary classifier with no guarantee of extracting artifact-prone fine-grained traces. + +Interestingly, in image-based deepfake detection, it has been recently demonstrated that the use of a tailored multitask learning framework for explicitly attending artifact-prone small regions coupled with a subtle data synthesis strategy can be a way to enhance generalization and, at the same time, robustness to high-quality deepfakes [41, 42]. Nevertheless, such an approach has been disregarded in the field of video-level deepfake detection, as its extension to the video level is not straightforward. In particular, it would necessitate the characterization of subtle temporal artifacts that are inherently different from spatial ones, within both the multi-task learning framework and the data synthesis. + +In this paper, we redefine deepfake video detection as a fine-grained detection task by proposing a multi-branch network that leverages synthesized data and incorporates specialized learning objectives specifically targeting both subtle spatial and temporal artifacts. As shown in Figure 1-b, a novel multi-task learning framework, termed FakeSTormer is introduced. It is formed by two auxiliary parallel branches in addition to the standard classification head, namely: (1) a regression temporal branch incorporating an explicit attention that aims at locating the vulnerability-prone temporal locations. It has been shown that regressing spatial vulnerabilities in specific points [42] or patches [41] can help improve the generalizability of a deepfake detector model. We refer to the definitions given in [41, 42] which describe: "vulnerable patches/points as the patches/points that are the most likely to embed blending artifacts". To generalize this concept to the temporal domain, we propose locating temporal high changes in spatial vulnerable patches (see Figure 1-c). (2) a spatial branch to ensure a balance between the spatial and the temporal domains. In fact, detecting spatial artifacts in addition to temporal ones is crucial [56, 67]. For that purpose, we propose predicting frame-wise spatial vulnerabilities. + +To create hand-free ground truths for the proposed branches, we introduce a HQ video-level data synthesis al + +gorithm, called "Self-Blended Video (SBV)", inspired by "Self-Blended Image (SBI) [48]", enforcing temporal coherence using two proposed modules on top of SBI (detailed in Section. 3.1). Our experiments demonstrate that simply training a baseline classification model on SBV enables achieving on par performance w.r.t. state-of-the-art (SOTA), highlighting the effectiveness of SBV. Finally, for enhancing spatial and temporal modeling, we revisit the TimeSformer [2] architecture that we use as our backbone. In particular, we leverage TimeSformer's decomposed temporal and spatial attention on embedded patches, appending classification tokens for each frame and for each patch across frames, rather than a single token for the entire video. These classification tokens are then used within the spatial and classification heads, while the embedded patches are used within the temporal head. Extensive experiments on several well-known deepfake detection benchmarks show that our method outperforms the existing SOTA approaches. + +# Contributions. In summary, we propose in this paper: + +- A novel multi-task learning framework using only real data for fine-grained video-based deepfake detection. +- Two auxiliary branches that capture both temporal and spatial vulnerabilities, that are fined-grained by definition. +- A video-level data synthesis technique called SBV that generates high-quality pseudo-fakes and is supported by a vulnerability-driven cutout augmentation strategy to avoid overfitting specific artifact-prone regions. +- A revisited version of the TimeSformer [2], specifically tailored for the proposed video-based deepfake detector. +- Extensive experiments and analyses conducted on several challenging datasets. + +Paper organization. Section 2 reviews related work on video deepfake detection. Section 3 describes the proposed FakeSTormer method. Section 4 presents experiments and results. Finally, Section 5 concludes with future work. + +# 2. Related Work + +Video-based Deepfake Detection. As highlighted in [56, 71], using a naive spatio-temporal binary classification model for video-level deepfake detection can lead the model to overfit obvious artifacts, resulting in poor generalization to unseen manipulations. To address this, FTCN [71] proposes a fully temporal convolution network by reducing the spatial kernel size to one, hence decreasing the likelihood of focusing only on spatial artifacts. LipForensics [25] considers solely the mouth region, while spatiotemporal dropout [64] randomly removes parts of the input frames in both spatial and temporal domains. AltFreezing [56] separates convolution layers into spatial and temporal ones, failing to model long-term dependencies. Instead of using convolution layers, ISTVT [67] utilizes a video + +based Vision Transformer [16] with self-attention to extract longer-range correlations. Meanwhile, [9] decomposes features into spatial and temporal components. TALL [57] employs an image-level deepfake detector by converting video frames into a thumbnail layout. Despite being promising, most of the aforementioned methods solely rely on a single binary classifier that implicitly guides the feature extraction. As highlighted in the literature on image-based deepfake detection [6, 15, 33, 42], this approach might lead to overfitting specific artifacts present in training datasets. Moreover, the absence of an explicit attention mechanism to spatio-temporal artifact-prone regions can lead to poor robustness to high-quality artifacts. + +Data Synthesis. A highly effective approach for enhancing the generalizability of deepfake detectors is training models with synthesized data. While frame-level solutions have been extensively studied [6, 15, 33, 48, 68], video-level augmentations remains relatively underexplored. In recent works, STC [34] generates pseudo-fake samples via time-shuffling, VB [62] perturbs landmarks per frame without imposing temporal coherence, while ST-SBV [24] injects temporal artifacts through random face scaling and blurring over time. However, these methods often introduce exaggerated temporal distortions that differ from HQ deepfakes typically exhibiting finer temporal inconsistencies. + +# 3. Methodology + +Let $\mathcal{V} \triangleq \cup_{i=1}^{N} \left\{ (\mathbf{X}_i, y_i) \right\}$ be a training dataset formed by $N$ videos, where $\mathbf{X}_i$ denotes the $i^{\text{th}}$ video sample and $y_i$ its associated label indicating whether the clip is real ( $y_i = 0$ ) or fake ( $y_i = 1$ ). Traditional methods [20, 23, 25, 56, 57, 64, 67, 71] aim to learn jointly a feature extractor $\Phi: \mathcal{V} \mapsto \mathcal{F}$ and a binary classifier $f: \mathcal{F} \rightarrow \{0,1\}$ by minimizing the standard binary cross-entropy (BCE) loss $\mathcal{L}_{BCE}(f(\Phi(\mathbf{X}_i)), y_i)$ using the entire training set $\mathcal{V}$ , with $\mathcal{F}$ being the learned feature space. As previously discussed in [33, 39, 42, 48] and also highlighted in Section 1, such a strategy might lead to poor generalization capabilities to unseen generation methods while providing only binary outputs that are not interpretable. + +To tackle these issues, inspired by the literature on image-level deepfake detection [6, 15, 33, 42, 68], we introduce a novel multi-task learning framework called FakeSTormer that only relies on the real data subset denoted as $\mathcal{V}^r\subset \mathcal{V}$ . Specifically, in addition to the binary classifier $f$ , our framework includes two additional branches $h:\mathcal{F}\to \mathcal{H}$ and $g:\mathcal{F}\rightarrow \mathcal{G}$ that aim at triggering the learning of localized temporal and spatial artifact-prone features, respectively, through relevant auxiliary tasks. Note that $\mathcal{H}$ and $\mathcal{G}$ denote respectively the output spaces of $h$ and $g$ . The proposed branches are depicted further in Section 3.2. To provide ground truth to those branches and at the same time avoid overfitting to specific manipulations, we apply to each + +video belonging to $\mathcal{V}^r$ a data synthesis method described in Section 3.1, resulting in a pseudo-fake subset denoted as $\mathcal{V}^{\tilde{r}}$ . Hence, our framework is trained using $\tilde{\mathcal{V}} \triangleq \{\mathcal{V}^r \cup \mathcal{V}^{\tilde{r}}\}$ . + +# 3.1. Video-Level Data Synthesis and Augmentation + +Self-Blended Video. Blending-based data synthesis methods have demonstrated great performance in image-based deepfake detection [6, 7, 33, 41, 42, 48, 68]. In fact, as the blending step is common to different manipulation types, they contribute to the improvement of the generalization aspect in deepfake detection [33, 42]. Nevertheless, such an approach has been overlooked in the context of video-based deepfake detection. Hence, we propose to extend blending-based data synthesis to the video level. In particular, we revisit Self-Blended Image (SBI) [48], given its ability to produce high-quality pseudo-fake images. The proposed data synthesis approach, termed Self-Blended Video (SBV), is constituted of two main components building on top of SBI, i.e., a Consistent Synthesized Parameters (CSP) module followed by a Landmark Interpolation module (LI) for preserving the temporal coherence of synthesized videos, which is essential for producing high-quality synthesized videos. + +Specifically, given a real video $\mathbf{X}_i^r\in \mathcal{V}^r$ formed by $T$ consecutive frames, we start by extracting a set of 2D landmarks $\mathbf{L}_i(t) = \{\mathbf{l}_{ij}(t)\}_{1\leq j\leq n}$ at each instant $t$ from $\mathbf{X}_i^r (t)$ where $n$ refers to the number of landmarks and $\{\mathbf{l}_{ij}(t)\} \in \mathbb{R}^2$ . Then, we apply SBI to the $1^{\mathrm{st}}$ video frame denoted as $\mathbf{X}_i^r (t_0)$ to obtain a pseudo-fake image, i.e., $\mathbf{X}_i^{\tilde{r}}(t_0)$ and a blending mask $\mathbf{M}_i(t_0)$ . All related blending parameters $\theta^{(sbi)}$ (e.g., ConvexHull type, Mask deformation kernels, blending ratio, etc.) are then conserved for synthesizing the remaining video frames. However, using those parameters solely cannot guarantee the temporal consistency of pseudo-fake videos since the geometry of landmarks can significantly vary over time. To mitigate the issue, we propose to re interpolate each landmark $\mathbf{l}_i(t)$ based on $\mathbf{l}_i(t - 1)$ for $t > t_0$ as follows, + +$$ +\mathbf {l} _ {i} (t) = \left\{ \begin{array}{c} \mathbf {l} _ {i} (t - 1) + \frac {\mathbf {l} _ {i} (t) - \mathbf {l} _ {i} (t - 1)}{\operatorname {r o u n d} (d / \bar {d})}, \text {i f} d > \tau \\ \text {w h e r e} d = \| \mathbf {l} _ {i} (t) - \mathbf {l} _ {i} (t - 1) \| _ {2} / n \\ \mathbf {l} _ {i} (t), \text {o t h e r w i s e ,} \end{array} \right. \tag {1} +$$ + +where $d$ represents the normalized distance between the position of a given landmark at the instants $t$ and $t - 1$ , $\tau$ is a constant threshold for determining when to interpolate (interpolation intervenes only in the presence of drastic changes), and $\bar{d}$ is empirically chosen. Overall, a higher $d$ value will push the updated point to move closer to the previous landmark position $\mathbf{l}_i(t - 1)$ , hence contributing to smooth the landmark position over time. However, excessive smoothing can be disadvantageous, as it can discard temporal artifacts. To address this, we use a round operator + +![](images/07b62d8c66e0250b061a853b5449976d2044f117f6ee244ff66a15306f66fb57.jpg) +Figure 2. I) Overview of the proposed framework: Our multi-task learning framework, FakeSTormer, consists of three branches, i.e., the temporal branch $(h)$ , the spatial branch $(g)$ , and the standard classification branch $(f)$ . Those branches are specially designed to facilitate the disentanglement learning of spatial-temporal features. The hand-free ground-truth data to train the framework are generated based on our proposed video-level data synthesis algorithm coupled with a vulnerability-driven Cutout strategy. II) Overview of generating a self-blended video: It contains two main components, including a landmark interpolation module $(LI)$ and the consistent utilization of synthesized parameters $(CSP)$ . III) Examples of pseudo-fake videos: with $(w/)$ and without $(w/o)$ vulnerability-driven Cutout and their corresponding soft labels. We apply the Cutout data augmentation at the same spatial locations throughout video frames. IV) Extraction of temporal vulnerabilities: We compute derivatives of the spatial vulnerabilities over time. + +to incorporate slight errors. Hence, the proposed SBV data synthesis produces high-quality pseudo-fake videos incorporating subtle temporal artifacts. + +As a result, we obtain the pseudo-fake video $\mathbf{X}_i^{\tilde{r}}\in \mathcal{V}^{\tilde{r}}$ and its blending mask sequence $\mathbf{M}_i\in \mathbb{R}^{T\times H\times W}$ , with $H$ and $W$ being the image height and width, respectively. An illustration of SBV is given in Figure 2-II. Additional samples, as well as the detailed algorithm, are provided in supplementary materials. It is important to note that, despite being simple, SBV is generic and applicable to any existing video-level deepfake detection approach. + +Vulnerability-Driven Cutout Augmentation. Previous works [54, 72] have demonstrated that deep learning methods are often impacted by overfitting. Deepfake detectors might even be more sensitive to this phenomenon as deep-fakes are typically characterized by localized artifacts [42]. One solution to regularize training is data augmentation. As such, we propose, in addition to SBV, a novel Cutout data augmentation driven by vulnerable patches, i.e., image patches that are prone to blending artifacts [41]. We posit that by masking the most vulnerable regions, overfitting risks will be reduced, as the model will be pushed to learn from other areas. This masking strategy has already been explored in other computer vision fields such as Classification [8, 12], Object Detection [21] demonstrating great + +potential. + +Specifically, similar to [33], we create a set of blending boundaries $\mathbf{B}$ using a randomly generated blending mask $\mathbf{M}$ as follows, + +$$ +\mathbf {B} = (\mathbf {1} - \mathbf {M}) * \mathbf {M} * 4, \mathbf {B} \in \mathbb {R} ^ {T \times H \times W}, \tag {2} +$$ + +with $*$ being the element-wise multiplication and $\mathbf{1}$ an all-one matrix. Inspired by [41], vulnerability values are then quantified at the patch level in a non-overlapping manner by applying a MaxPooling function as follows, + +$$ +\bar {\mathbf {B}} = \operatorname {M a x P o o l i n g} (\mathbf {B}), \quad \bar {\mathbf {B}} \in \mathbb {R} ^ {T \times \sqrt {N} \times \sqrt {N}}, \tag {3} +$$ + +where $N$ indicates the number of patches. + +After that, we define a threshold $\tau_{cutout}$ that is randomly selected from the range (0.5, 1.0]. We use the latter to define the set of patches to be masked $\mathcal{P} = \{(l,m) \mid \tilde{\mathbf{B}}_{l,m}(t_0) > \tau_{cutout}\}$ within the first frame. The set $\mathcal{P}$ is then used to mask out patches at those locations not only in the first frame but also in the entire video to enforce temporal consistency that is crucial for generating high-quality pseudo-fakes. After masking those patches over time, we finally obtain the masked blending boundary denoted as $\tilde{\mathbf{B}}$ . This results in masking the most vulnerable regions, i.e., the regions that are the most likely to include blending artifacts. Figure 2-III shows some examples of the proposed cutout augmentation. + +# 3.2. FakeStormer + +Our multi-task framework, called FakeSTormer, is inspired by [41, 42], where auxiliary branches are designed to push the feature extractor to focus on vulnerabilities. As discussed earlier, the vulnerability is defined in [41, 42] as the pixels/patches that are the most likely to be impacted by blending artifacts. This strategy is therefore claimed to allow the detection of subtle artifacts that are generic across different types of manipulations. While such a vulnerability-driven approach has shown very promising results [41, 42], it does not take into account the temporal nature of videos. Therefore, in addition to spatial vulnerabilities, we argue that there is a need to model temporal vulnerabilities, which we define as significant temporal changes in the blending boundary. Specifically, we introduce two additional branches, namely a temporal head $h$ and a spatial one $g$ . The branch $h$ predicts the derivatives of the blending boundary over time which can reflect high changes, typically characterizing temporal artifacts. Moreover, we suggest the use of a spatial branch $g$ which enables predicting soft labels representing the forgery intensity encoded in each frame, computed from vulnerability information. The proposed framework relies on the TimeSformer backbone [2], which we revisit for better modeling spatial and temporal information. + +Herein, we first describe the proposed revisited TimeSformer-based feature extractor $\Phi$ in Section 3.2.1. We then detail the two additional temporal and spatial heads in Section 3.2.2 and Section 3.2.3, respectively. Finally, we give the overall training details in Section 3.2.4. + +# 3.2.1. Backbone: Revisited TimeSformer + +We choose TimeSformer [2] as our feature extractor given its ability to effectively capture separate long-range temporal information and spatial features. In TimeSformer, a video input $\mathbf{X} \in \mathbb{R}^{C \times T \times H \times W}$ results in an embedding matrix input $\mathbf{Z}^0 \in \mathbb{R}^{T \times N \times D}$ . A global class token $\mathbf{z}_{cls}$ attends all patches and is then used for classification. This mechanism implicitly captures mixed spatio-temporal features, which might lead to overfitting one type of artifact. We revisit it slightly in order to decouple the spatial and temporal information by considering two sorts of additional tokens (one spatial and one temporal). + +For that purpose, we attach in each dimension of $\mathbf{Z}^0$ , a spatial token $\mathbf{z}_s^0 \in \mathbb{R}^D$ and a temporal token $\mathbf{z}_t^0 \in \mathbb{R}^D$ , respectively. These tokens will independently interact only with patch embeddings belonging to their dimension axis by leveraging the decomposed SA [2]. This mechanism not only facilitates the disentanglement learning process of spatio-temporal features but is also beneficial to optimize the computational complexity of $\mathcal{O}(T^2 + N^2)$ as compared to $\mathcal{O}(T^2 \cdot N^2)$ in vanilla SA. Those tokens will be then fed into $L$ ( $L = 12$ as default) transformer encoder blocks, + +as described in Figure 2-I. Formally, the feature extraction process can be summarized as follows, + +$$ +\left[ \mathbf {Z} ^ {L}, \mathbf {z} _ {s} ^ {L}, \mathbf {z} _ {t} ^ {L} \right] = \Phi (\mathbf {X}), \tag {4} +$$ + +where $\mathbf{Z}^L$ is the final patch embedding matrix, $\mathbf{z}_s^L$ the resulting set of spatial tokens, and $\mathbf{z}_t^L$ the resulting set of temporal tokens that will be respectively sent to the temporal head $h$ , the spatial head $g$ , and the classification head $f$ . More details about the implementation of the proposed revisited TimeSformer are given in supplementary materials. + +# 3.2.2.Temporal Head $h$ + +Ground Truths. Our temporal head $h$ aims to model fine-grained temporal vulnerabilities in deepfake videos through a regression task. First, to generate ground truth data for the branch $h$ , we hypothesize that temporal high-changes in the blending boundary can reflect the presence of temporal artifacts (see Figure 2-IV). To achieve this, we compute $\mathbf{D}$ based on $\tilde{\mathbf{B}}$ such that: + +$$ +\mathbf {D} = \frac {\partial \tilde {\mathbf {B}}}{\partial t}, \quad \mathbf {D} \in \mathbb {R} ^ {T \times \sqrt {N} \times \sqrt {N}}. \tag {5} +$$ + +More details regarding the derivative calculation are provided in supplementary materials. To stabilize training, $\mathbf{D}$ is standardized resulting in $\hat{\mathbf{D}}\in \mathbb{R}^{T\times \sqrt{N}\times \sqrt{N}}$ . Experiments with different normalization strategies are reported in supplementary materials. + +Architecture Design. In order to construct the regression head for predicting $\hat{\mathbf{D}}$ , we take the patch embedding matrix $\mathbf{Z}^L$ as input and process them to produce 3D features as follows, + +$$ +\mathbf {F} = \operatorname {R e s h a p e} \left(\mathbf {Z} ^ {L}\right), \mathbf {F} \in \mathbb {R} ^ {D \times T \times \sqrt {N} \times \sqrt {N}}. \tag {6} +$$ + +To estimate temporal derivatives, we employ two 3D convolution blocks (3DCnvB) with 3-dimensional temporal kernels and 1-dimensional spatial kernels [71] as follows, + +$$ +\tilde {\mathbf {D}} = h (\mathbf {F}) = 3 \mathrm {D C n v B} _ {3 \times 1 \times 1} \left(3 \mathrm {D C n v B} _ {3 \times 1 \times 1} (\mathbf {F})\right), \tag {7} +$$ + +where $\tilde{\mathbf{D}}\in \mathbb{R}^{T\times \sqrt{N}\times \sqrt{N}}$ . Each convolution block comprises a 3D convolution layer, followed by a BatchNorm and a GELU layer. + +Objective Function. For training the temporal branch, we optimize the following Mean Squared Error (MSE) loss, + +$$ +\mathcal {L} _ {h} = \frac {1}{T \times N} \| \hat {\mathbf {D}} - \tilde {\mathbf {D}} \| _ {2} ^ {2}, \tag {8} +$$ + +with $\| .\| _2$ referring to the $L_{2}$ norm. + +# 3.2.3. Spatial Head $g$ + +Ground Truths. To avoid overfitting one type of artifact, we enforce the model to explicitly predict soft labels representing the intensity level of spatial artifacts for each video + +frame. Note that several works [4, 40, 49, 52, 66] have leveraged soft labels for training regularization. Given a pseudo-fake video $\mathbf{X} = (\mathbf{X}(t))_{t\in [[1,T]]}$ formed by $T$ frames and $\tilde{\mathbf{B}} = (\tilde{\mathbf{B}} (t))_{t\in [[1,T]]}$ its associated cutout blending boundary, the ground truth for these soft labels is generated for each frame $t$ as follows, + +$$ +p (t) = \max _ {l, m \in [ [ 1, \sqrt {N} ] ]} (\tilde {\mathbf {B}} (t)), \tag {9} +$$ + +resulting in the ground truth for training the spatial branch denoted as $\mathbf{p} = (p(t))_{t\in [[1,T]]}$ . We note that $\mathbf{p} = \mathbf{1}^T$ if cutout is not applied and $\mathbf{p} = \mathbf{0}^T$ for a real video. + +Architecture Design. To predict the proposed soft labels, a Multi-Layer Perceptron (MLP) is applied to the set of spatial tokens $\mathbf{z}_s^L$ , as follows, + +$$ +\tilde {\mathbf {p}} = g \left(\mathbf {z} _ {s} ^ {L}\right) = \operatorname {M L P} \left(\mathbf {z} _ {s} ^ {L}\right), \quad \tilde {\mathbf {p}} \in \mathbb {R} ^ {T}. \tag {10} +$$ + +Objective Function. To train the spatial branch, we optimize the following Binary Cross Entropy (BCE) loss similar to [52, 66], + +$$ +\mathcal {L} _ {g} = \operatorname {B C E} (\tilde {\mathbf {p}}, \mathbf {p}). \tag {11} +$$ + +# 3.2.4. Overall Training Objective + +Finally, for the standard classification head $f$ , we use the set of temporal tokens $\mathbf{z}_t^L$ such that the predicted label $\tilde{y}$ is given by, + +$$ +\tilde {y} = f \left(\mathbf {z} _ {t} ^ {L}\right) = \operatorname {M L P} \left(\mathbf {z} _ {t} ^ {L}\right). \tag {12} +$$ + +The classification loss $\mathcal{L}_c$ is then given by applying a BCE between the ground-truth label $y$ and the predicted label $\tilde{y}$ . + +Overall, the network is trained by optimizing the following loss: + +$$ +\mathcal {L} = \lambda_ {c} \mathcal {L} _ {c} + \lambda_ {h} \mathcal {L} _ {h} + \lambda_ {g} \mathcal {L} _ {g}, \tag {13} +$$ + +where $\lambda_c, \lambda_h, \lambda_g$ are hyper-parameters to balance the training of the three branches. + +# 4. Experiment + +# 4.1. Settings + +Datasets. We set up our datasets following several works [3, 9, 55-57, 59, 67, 71]. For both training and validation, we employ FaceForensics++ (FF++) [46], which consists of four manipulation methods for the fake data (Deepfakes (DF) [10], FaceSwap (FS) [31], Face2Face (F2F) [51], and NeuralTextures (NT) [50]). It can be noted that, for training, we use only the real videos and generate pseudo-fake data using our synthesized method, SBV. By default, the c23 version of $FF++$ is adopted, following the recent literature [9, 56, 57, 67, 71]. For further validation, we also evaluate on the following datasets: CelebDFv2 (CDF) [36], DeepfakeDetection (DFD) [17], Deepfake Detection Challenge Preview (DFDCP) [13], Deepfake Detection Challenge (DFDC) [14], WildDeepfake + +
MethodTrainingTest set AUC (%)
RealFakeCDFDFDDFDCPDFDCDFWDiffSwap
Xception [46]✓✓73.7--70.9--
MATT [69]✓✓68.392.963.0-65.7-
RECCE [3]✓✓70.998.2--68.2-
SBI [48]✓×90.6--72.4--
SFDG [55]✓✓75.888.073.6-69.3-
LSDA [60]✓✓91.1-77.0---
STIL [23]✓✓75.6-----
LipForensics [25]✓✓82.4--73.5--
RealForensics [26]✓✓86.982.275.9---
FTCN [71]✓✓86.994.474.071.0--
ISTVT [67]✓✓84.1-74.2---
AltFreezing [56]✓✓89.598.5----
Swin+TALL [57]✓✓90.8-76.8---
StyleLatentFlows [9]✓✓89.096.1----
LFGDIN [63]✓✓90.4-80.8--85.7
FakeStormer (T=4)✓×92.498.590.074.674.296.9
FakeStormer (T=8)✓×92.498.290.074.975.997.1
FakeStormer (T=16)✓×92.898.690.275.175.397.2
+ +Table 1. Generalization to unseen datasets. AUC (\%) comparisons at video-level on multiple unseen datasets [13, 14, 17, 36, 70, 74]. All detectors are trained on FF++(c23). Results are directly extracted from the original papers and from [25, 42]. Bold and Underlined text, respectively highlight the best and the second best performance, excluding the variants of our framework with $T = 8$ and $T = 16$ . + +
MethodTraining setCross-datasetDF40 subset
RealFakeCDFDFDCPDiffSwapBlendFaceFSGANMobileSwap
Face X-ray [33]✓✓79.5-----
PCL+1G [68]✓✓90.074.3----
SLADD [6]✓✓79.7-----
SBI [48]✓×93.286.290.686.585.486.6
LAA-Net [42]✓×95.486.992.191.294.293.9
STC-Scratch [34]✓×83.486.8----
STC-Pretrain [34]✓×95.889.4----
ST-SBV [24]✓×90.391.2----
StA+VB [62]✓×94.790.9-90.696.494.6
TimeSformer [2] + SBV✓×94.993.093.389.794.694.6
FakeStormer✓×96.594.197.791.196.495.0
+ +Table 2. AUC(%) comparison at video-level with other data synthesis methods. For fair comparison, we train our FakeSTormer on raw data of FF++(c0), and test cross-dataset on [13, 36, 70] and cross-manipulation on three subsets of [61]. + +
MethodTraining setFF++LQ (%)
RealNTDFFS
Xception [46]✓✓58.751.7
Face X-ray [33]✓✓57.151.0
F3Net [44]✓✓58.351.9
RFM [54]✓✓55.851.6
SRM [37]✓✓55.552.9
SLADD [6]✓✓62.856.8
TALL-Swin [57]✓✓63.251.4
ResNet3D* [27]✓✓66.860.6
TimeSformer* [2]✓✓73.354.4
Ours✓×85.362.1
+ +Table 3. Generalization on heavily compressed data (LQ). AUC (\%) comparisons on FF++ (LQ) [46] with a high compression level (c40). The results for comparison are directly extracted from [6, 38]. The symbol * denotes our implementation. + +(DFW) [74], DF40 [61], and DiffSwap [63, 70] generated using a recent diffusion-based approach [70]. Further details on these datasets are provided in supplementary materials. + +Data Pre-processing. Following the splitting convention [46], we extract 256, 32, and 32 consecutive frames for training, validation, and testing, respectively. Facial regions are cropped using Face-RetinaNet [11] and resized to a fixed resolution of $224 \times 224$ . Additionally, we store 81 facial landmarks for each frame, extracted using Dlib [30]. + +
MethodTraining setFF++ (%)Avg.
RealFakeDFFSF2FNT
Xception [46]✓✓93.951.286.879.777.9
Face X-ray [33]✓✓99.593.294.592.594.9
SBI [48]✓×98.695.492.682.392.2
LSDA [60]✓✓96.995.196.494.995.8
LipForensics [25]✓✓99.790.199.799.197.1
FTCN [71]✓✓99.899.698.295.698.3
RealForensics [26]✓✓10097.199.799.299.0
AltFreezing [56]✓✓99.899.798.696.298.6
StyleLatentFlows [9]✓✓99.798.898.696.498.4
NACO [65]✓✓99.999.799.899.499.7
LFGDIN [63]✓✓96.280.590.581.787.2
Ours (c23)✓×99.997.898.597.298.4
Ours (c0)✓×10099.899.999.799.9
+ +Further details are provided in the supplementary materials. Evaluation Metrics. For fair comparisons with SOTA methods, we use the widely adopted Area Under the Curve (AUC) metric at the video level [9, 25, 26, 56, 57, 67, 71]. Implementation Details. Our framework is initialized with pretrained MAE weights [28] and trained for 100 epochs using the SAM optimizer [18] with a weight decay of $10^{-4}$ and a batch size of 32. The learning rate starts at $5 \times 10^{-4}$ for the first quarter of training and decays to 0 thereafter. The backbone is frozen for the first 5 epochs for warm-up, then all layers are unfrozen. Data augmentation includes ColorJittering at the video level and our proposed Cutout. Experiments are conducted on four NVIDIA A100 GPUs, with $\tau = 0.35$ and $\bar{d} = 0.2$ (Eq. (1)), and $T = 4$ frames in most experiments. + +# 4.2. Comparison with State-of-the-art Methods + +Generalization to Unseen Datasets. To assess the generalization capabilities of our method, we conduct evaluations using the challenging cross-dataset setup [3, 42, 55, 56, 71], validating on unseen datasets (i.e., datasets other than $\mathrm{FF}++$ ). The results are detailed in Table 1 and Table 2. + +As shown, our method achieves comparable results on DFD while surpassing SOTA methods on other datasets. Specifically, it significantly outperforms prior video deepfake detection techniques, including spatiotemporal learning-based methods like AltFreezing [56] and ISTVT [67], as well as various data synthesis approaches. Moreover, our method exhibits superior performance on the large-scale DFDC dataset and the challenging in-the-wild DFW dataset. These results further confirm the enhanced generalization ability of FakeSTormer compared to recent methods. + +Generalization on Heavily Compressed Data. Following previous work [6, 38], we also evaluate FakeSTormer on heavily compressed FF++(c40) data. In addition to comparing with several SOTA methods, we train ResNet3D [27], commonly used in deepfake video detection [9, 56, 71], and TimeSformer [2] on NT, then test on DF and FS. The + +Table 4. Generalization to unseen manipulations. AUC (%) comparisons on FF++ [46], which consists of four manipulation methods (DF, FS, F2F, NT). + +
SBVV-CutOutghTest set AUC (%)Avg.
CDFDFDDFDCPDFDCDFWDiffSwap
××××61.562.859.458.565.271.663.2
✓×××90.795.787.972.270.992.985.1(↑21.9)
✓✓××91.196.087.672.671.093.185.2(↑22.0)
✓✓✓×92.295.488.572.871.393.885.7(↑22.5)
✓××✓93.498.588.572.869.697.386.7(↑23.5)
✓✓✓✓92.498.590.074.674.296.987.8(↑24.6)
+ +Table 5. Ablation study of framework's components. Gray indicates the use of original fake data for training. + +comparison results are presented in Table 3. Our method achieves notably higher AUC scores than other methods across both testing subsets, highlighting its robust generalization capability under various data compression conditions. + +Generalization to Unseen Manipulations. Table 4 compares our framework with SOTA methods on $\mathrm{FF}++$ . Other methods [9, 25, 56, 71] use a cross-manipulation setup, training on three forgery types and evaluating on the remaining one. In contrast, our approach trains only on real videos, treating all manipulations as unseen. Despite this, our method shows competitive performance with the others, even without being trained on specific forgery types. + +Robustness to Unseen Perturbations. Deepfakes are widely shared on social media, where various perturbations can affect their appearance. Following [29], we evaluate FakeSTormer's robustness across six unseen degradation types at five levels, comparing it with other augmented-based methods [33, 35, 42, 48]. Figure 3 shows AUC scores for each method on these perturbations, using models trained on $\mathrm{FF} + +$ . Our results demonstrate that FakeSTormer outperforms prior methods on most distortions, with a slight drop compared to LAA-Net [42] for Change Saturation. Nonetheless, FakeSTormer achieves higher performance on average, especially at higher severity levels, highlighting its superior generalization and robustness. Detailed scores are in supplementary materials. + +# 4.3. Additional Discussions + +Ablation Study of the FakeSTormer's Components. We conduct ablation studies to assess the impact of each component in our framework, as shown in Table 5. Using TimeSformer trained on $\mathrm{FF}++$ as the baseline, we experiment with different combinations of components: Self-Blended Video (SBV), Vulnerability-driven CutOut (V-CutOut), the spatial branch $(g)$ , and the temporal branch $(h)$ . Each component improves performance, with SBV providing the most significant boost by generating high-quality pseudo-fake data that aids generalization to unseen datasets. Ablation Study of the SBV's Components. SBV enhances SBI [48] with CSP and LI for robust pseudo-fake generation in video data. Table 6 shows that without these components, simply stacking frame-wise SBIs fails to produce consistent temporal features, leading to overfitting on more obvious artifacts and poor generalization [56]. A qualitative comparison is provided in supplementary materials. + +![](images/f94cfbe0177f0d60ce4d8060b96d0d25b1db9e079d120419bd42eb12b41b9cfa.jpg) +Figure 3. Robustness to unseen perturbations. AUC (\%) under five different degradation levels for various types of perturbations [29] on $\mathrm{FF}++$ [46]. "Average" denotes the mean across all corruptions at each level. Best viewed in color. + +
MethodComp.CSPLITest set AUC (%)
CDFDFDDFDCPDFDCDFWDiffSwap
Stacked SBIsc23××48.449.148.951.752.855.5
+ CSPc23✓×84.189.286.169.565.585.7
SBVc23✓✓90.795.787.972.270.992.9
SBVc0✓✓94.997.693.076.475.393.3
+ +Table 6. Ablation study of SBV's components. Performance analyses of different SBV's components using cross-evaluation on multiple datasets [13, 14, 17, 36, 70, 74]. + +
λcλhλgTest set AUC (%)Avg.
CDFDFDDFDCPDFDCDFWDiffSwap
0.910.189.591.093.271.574.790.685.1
0.9100.192.595.786.871.772.794.585.7
0.91000.191.698.087.373.670.696.286.2
0.81000.292.498.590.074.674.296.987.8
0.51000.592.498.088.174.572.197.187.0
+ +Table 7. Impact of loss balancing factors. AUC (\%) comparisons of FakeSTormer trained with different values of $\lambda_{c}$ , $\lambda_{h}$ , and $\lambda_{g}$ on cross-dataset setup, demonstrating robustness to varying hyperparameter settings. + +Influence of Number of Frames. Increasing the number of frames $T$ provides more fine-grained temporal information. In Table 1, we vary $T$ values by fixing it to 4, 8, and 16. Our results show a consistent performance improvement with more frames, confirming our hypothesis. However, increasing $T$ also incurs a higher computational cost. + +Impact of Loss Balancing Factors. We introduce three hyperparameters, $\lambda_c$ , $\lambda_g$ , and $\lambda_h$ in Eq (13) to balance the training among the three branches of our framework. In Table 7, we analyze the impact of these hyperparameters using various values. Our results show that the method is robust to a range of hyperparameter values, with the best performance achieved when $\lambda_c$ , $\lambda_g$ , and $\lambda_h$ are set to 0.8, 0.2, and 100, respectively. + +# 4.4. Visualization of Saliency Maps + +To analyze the contribution of the two proposed branches $h$ and $g$ in the detection performance of FakeSTormer, we visualize the input regions activated by those branches. For that purpose, we adopt Grad-CAM [47] for the temporal branch $h$ and utilize the final SA scores of spatial tokens for the spatial branch $g$ . The visualization results from various datasets are presented in Figure 4. It can be observed that FakeSTormer can discriminate between real and fake videos by focusing on very few, different local areas, even without + +![](images/6a86d5dfff4ea90b475f275e52bc95fb26c1de09319135e94d56d718cefe1fee.jpg) +Figure 4. Visualization of Saliency Maps. The second-fifth and sixth-ninth columns represent temporal heatmaps and spatial heatmaps on different frames in the video, respectively. All datasets are unseen during validation. + +having seen those types of forgeries during training. + +# 5. Conclusion + +This paper introduces a fine-grained approach for generalizable deepfake video detection with two main contributions. First, we propose a multi-task learning framework that targets both subtle spatial and fine-grained temporal vulnerabilities in high-fidelity deepfake videos, incorporating a standard classification branch along with two new auxiliary branches (temporal and spatial). These proposed branches help the model focus on vulnerable regions and provide more valuable insights into how the network sees the data while offering more robustness to high-quality deepfakes. This framework is further supported by the introduction of a high-quality pseudo-fake generation technique. Extensive experiments on several challenging benchmarks demonstrate that FakeSTormer achieves superior performance compared to SOTA methods. + +# Acknowledgment + +This work is supported by the Luxembourg National Research Fund, under the BRIDGES2021/IS/16353350/FaKeDeTeR, and by POST Luxembourg. Experiments were performed on the Luxembourg national supercomputer MeluXina. The authors gratefully acknowledge the LuxProvide teams for their expert support. + +# References + +[1] Yuval Alaluf, Or Patashnik, Zongze Wu, Asif Zamir, Eli Shechtman, Dani Lischinski, and Daniel Cohen-Or. Third time's the charm? image and video editing with stylegan3. CoRR, abs/2201.13433, 2022. 1 +[2] Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is space-time attention all you need for video understanding? In Proceedings of the International Conference on Machine Learning (ICML), 2021. 2, 5, 6, 7 +[3] Junyi Cao, Chao Ma, Taiping Yao, Shen Chen, Shouhong Ding, and Xiaokang Yang. End-to-end reconstruction-classification learning for face forgery detection. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4103-4112, 2022. 1, 6, 7 +[4] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. CoRR, abs/2104.14294, 2021. 6 +[5] Heather Chen and Kathleen Magramo. Finance worker pays out $25 million after video call with deepfake "chief financial officer". https://edition.cnn.com/2024/02/04/asia/deepfake-cfo-scam-hong-kong-int1-hnk/index.html, 2024. [Online; accessed 4-February-2024]. 1 +[6] Liang Chen, Yong Zhang, Yibing Song, Lingqiao Liu, and Jue Wang. Self-supervised learning of adversarial example: Towards good generalizations for deepfake detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18710-18719, 2022. 1, 3, 6, 7 +[7] Liang Chen, Yong Zhang, Yibing Song, Jue Wang, and Lingqiao Liu. Ost: Improving generalization of deepfake detection via one-shot test-time training. In Advances in Neural Information Processing Systems, pages 24597-24610. Curran Associates, Inc., 2022. 1, 3 +[8] Juhwan Choi and YoungBin Kim. Colorful cutout: Enhancing image data augmentation with curriculum learning, 2024. 4 +[9] Jongwook Choi, Taehoon Kim, Yonghyun Jeong, Seungryul Baek, and Jongwon Choi. Exploiting style latent flows for generalizing deepfake video detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1133-1143, 2024. 3, 6, 7 +[10] Deepfakes. Faceswapdevs. https://github.com/deepfakes/faceswap, 2019.6 +[11] Jiankang Deng, Jia Guo, Yuxiang Zhou, Jinke Yu, Irene Kotsia, and Stefanos Zafeiriou. Retinaface: Single-stage dense + +face localisation in the wild. CoRR, abs/1905.00641, 2019. 6 +[12] Terrance DeVries and Graham W. Taylor. Improved regularization of convolutional neural networks with cutout, 2017. 4 +[13] Brian Dolhansky, Russ Howes, Ben Pflaum, Nicole Baram, and Cristian Canton-Ferrer. The deepfake detection challenge (DFDC) preview dataset. CoRR, abs/1910.08854, 2019. 6, 8 +[14] Brian Dolhansky, Joanna Bitton, Ben Pflaum, Jikuo Lu, Russ Howes, Menglin Wang, and Cristian Canton-Ferrer. The deepfake detection challenge dataset. CoRR, abs/2006.07397, 2020. 6, 8 +[15] Shichao Dong, Jin Wang, Renhe Ji, Jiajun Liang, Haoqiang Fan, and Zheng Ge. Implicit identity leakage: The stumbling block to improving deepfake detection generalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3994-4004, 2023. 1, 3 +[16] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. CoRR, abs/2010.11929, 2020. 3 +[17] Nick Dufour and Andrew Gully. Contributing data to deepfake detection research. https://ai.googleblog.com/2019/09/contributing-data-to-deepfake-detection.html, 2019.6, 8 +[18] Pierre Foret, Ariel Kleiner, Hossein Mobahi, and Behnam Neyshabur. Sharpness-aware minimization for efficiently improving generalization. CoRR, abs/2010.01412, 2020. 7 +[19] Tsu-Jui Fu, Xin Eric Wang, Scott T. Grafton, Miguel P. Eckstein, and William Yang Wang. M3l: Language-based video editing via multi-modal multi-level transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10513-10522, 2022. 1 +[20] Ipek Ganiyusufoglu, L. Minh Ngô, Nedko Savov, Sezer Karaoglu, and Theo Gevers. Spatio-temporal features for generalized detection of deepfake videos. CoRR, abs/2010.11844, 2020. 1, 3 +[21] Chengyue Gong, Dilin Wang, Meng Li, Vikas Chandra, and Qiang Liu. Keepaugment: A simple information-preserving data augmentation approach. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1055-1064, 2021. 4 +[22] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2014. 1 +[23] Zhihao Gu, Yang Chen, Taiping Yao, Shouhong Ding, Jilin Li, Feiyue Huang, and Lizhuang Ma. Spatiotemporal inconsistency learning for deepfake video detection. Proceedings of the 29th ACM International Conference on Multimedia, 2021. 1, 3, 6 + +[24] Weinan Guan, Wei Wang, Bo Peng, Jing Dong, and Tieniu Tan. St-sbv: Spatial-temporal self-blended videos for deep-fake detection. In Chinese Conference on Pattern Recognition and Computer Vision (PRCV), pages 274-288. Springer, 2024. 2, 3, 6 +[25] Alexandros Haliassos, Konstantinos Vougioukas, Stavros Petridis, and Maja Pantic. Lips don't lie: A generalisable and robust approach to face forgery detection. CoRR, abs/2012.07657, 2020. 1, 2, 3, 6, 7 +[26] Alexandros Haliassos, Rodrigo Mira, Stavros Petridis, and Maja Pantic. Leveraging real talking faces via self-supervision for robust forgery detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14950-14962, 2022. 6, 7 +[27] Kensho Hara, Hirokatsu Kataoka, and Yutaka Satoh. Can spatiotemporal 3d cnns retrace the history of 2d cnns and imagenet? In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 6546-6555, 2018. 6, 7 +[28] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16000-16009, 2022. 7 +[29] Liming Jiang, Ren Li, Wayne Wu, Chen Qian, and Chen Change Loy. DeeperForensics-1.0: A large-scale dataset for real-world face forgery detection. In CVPR, 2020. 7, 8 +[30] Davis E. King. Dlib-ml: A machine learning toolkit. J. Mach. Learn. Res., 10:1755-1758, 2009. 6 +[31] Marek Kowalski. Faceswap. https://github.com/MarekKowalski/FaceSwap, 2018.6 +[32] Lingzhi Li, Jianmin Bao, Hao Yang, Dong Chen, and Fang Wen. Advancing high fidelity identity swapping for forgery detection. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5073-5082, 2020. 1 +[33] Lingzhi Li, Jianmin Bao, Ting Zhang, Hao Yang, Dong Chen, Fang Wen, and Baining Guo. Face x-ray for more general face forgery detection. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 1, 3, 4, 6, 7 +[34] Maosen Li, Xurong Li, Kun Yu, Cheng Deng, Heng Huang, Feng Mao, Hui Xue, and Minghao Li. Spatio-temporal catcher: A self-supervised transformer for deepfake video detection. In Proceedings of the 31st ACM International Conference on Multimedia, pages 8707-8718, 2023. 2, 3, 6 +[35] Yuezun Li and Siwei Lyu. Exposing deepfake videos by detecting face warping artifacts. In IEEE Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), 2019. 7 +[36] Yuezun Li, Xin Yang, Pu Sun, Honggang Qi, and Siwei Lyu. Celeb-df: A large-scale challenging dataset for deepfake forensics. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 6, 8 + +[37] Yuchen Luo, Yong Zhang, Junchi Yan, and Wei Liu. Generalizing face forgery detection with high-frequency features. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16317-16326, 2021. 6 +[38] Qingxuan Lv, Yuezun Li, Junyu Dong, Sheng Chen, Hui Yu, Huiyu Zhou, and Shu Zhang. Domainforensics: Exposing face forgery across domains via bi-directional adaptation. IEEE Transactions on Information Forensics and Security, 19:7275-7289, 2024. 6, 7 +[39] Nesryne Mejri, Enjie Ghorbel, and Djamila Aouada. Untag: Learning generic features for unsupervised type-agnostic deepfake detection. In ICASSP 2023 - 2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1-5, 2023. 1, 3 +[40] Rafael Müller, Simon Kornblith, and Geoffrey E. Hinton. When does label smoothing help? CoRR, abs/1906.02629, 2019. 6 +[41] Dat Nguyen, Marcella Astrid, Enjie Ghorbel, and Djamila Aouada. Fakeformer: Efficient vulnerability-driven transformers for generalisable deepfake detection. arXiv preprint arXiv:2410.21964, 2024. 1, 2, 3, 4, 5 +[42] Dat Nguyen, Nesryne Mejri, Inder Pal Singh, Polina Kuleshova, Marcella Astrid, Anis Kacem, Enjie Ghorbel, and Djamila Aouada. Laa-net: Localized artifact attention network for quality-agnostic and generalizable deepfake detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17395–17405, 2024. 1, 2, 3, 4, 5, 6, 7 +[43] Yuval Nirkin, Yosi Keller, and Tal Hassner. FSGANv2: Improved subject agnostic face swapping and reenactment. IEEE, 2022. 1 +[44] Yuyang Qian, Guojun Yin, Lu Sheng, Zixuan Chen, and Jing Shao. Thinking in frequency: Face forgery detection by mining frequency-aware clues. In European conference on computer vision, pages 86-103. Springer, 2020. 6 +[45] Reuters. South korea to criminalize watching or possessing sexually explicit deepfakes. https://edition.cnn.com/2024/09/26/asia/south-korea-deepfake-bill-passed-intl-hnk/index.html, 2024. [Online; accessed 26-September-2024]. 1 +[46] Andreas Rössler, Davide Cozzolino, Luisa Verdoliva, Christian Riess, Justus Thies, and Matthias Nießner. FaceForensics++: Learning to detect manipulated facial images. In International Conference on Computer Vision (ICCV), 2019. 1, 6, 7, 8 +[47] Ramprasaath R. Selvaraju, Abhishek Das, Ramakrishna Vedantam, Michael Cogswell, Devi Parikh, and Dhruv Batra. Grad-cam: Why did you say that? visual explanations from deep networks via gradient-based localization. CoRR, abs/1610.02391, 2016. 8 +[48] Kaede Shiohara and Toshihiko Yamasaki. Detecting deepfakes with self-blended images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18720-18729, 2022. 1, 2, 3, 6, 7 +[49] Guocong Song and Wei Chai. Collaborative learning for deep neural networks. In Neural Information Processing Systems, 2018. 6 + +[50] Justus Thies, Michael Zollhöfer, and Matthias Nießner. Deferred neural rendering: Image synthesis using neural textures. CoRR, abs/1904.12356, 2019. 6 +[51] Justus Thies, Michael Zollhöfer, Marc Stamminger, Christian Theobalt, and Matthias Nießner. Face2face: Real-time face capture and reenactment of RGB videos. CoRR, abs/2007.14808, 2020. 1, 6 +[52] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. CoRR, abs/2012.12877, 2020. 6 +[53] Jane Wakefield. Deepfake presidents used in Russia-Ukraine war. https://www.bbc.com/news/technology-60780142, 2022. [Online; accessed 7-March-2023]. 1 +[54] Chengrui Wang and Weihong Deng. Representative forgery mining for fake face detection. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 4, 6 +[55] Yuan Wang, Kun Yu, Chen Chen, Xiyuan Hu, and Silong Peng. Dynamic graph learning with content-guided spatial-frequency relation reasoning for deepfake detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7278-7287, 2023. 6, 7 +[56] Zhendong Wang, Jianmin Bao, Wengang Zhou, Weilun Wang, and Houqiang Li. Altfreezing for more general video face forgery detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4129-4138, 2023. 1, 2, 3, 6, 7 +[57] Yuting Xu, Jian Liang, Gengyun Jia, Ziming Yang, Yanhao Zhang, and Ran He. Tall: Thumbnail layout for deepfake video detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 22658-22668, 2023. 1, 3, 6, 7 +[58] Zhiliang Xu, Zhibin Hong, Changxing Ding, Zhen Zhu, Junyu Han, Jingtuo Liu, and Errui Ding. Mobilefaceswap: A lightweight framework for video face swapping. In Proceedings of the AAAI Conference on Artificial Intelligence, 2022. 1 +[59] Zhiyuan Yan, Yong Zhang, Yanbo Fan, and Baoyuan Wu. Ucf: Uncovering common features for generalizable deepfake detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 22412-22423, 2023. 6 +[60] Zhiyuan Yan, Yuhao Luo, Siwei Lyu, Qingshan Liu, and Baoyuan Wu. Transcending forgery specificity with latent space augmentation for generalizable deepfake detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8984-8994, 2024. 6, 7 +[61] Zhiyuan Yan, Taiping Yao, Shen Chen, Yandan Zhao, Xinghe Fu, Junwei Zhu, Donghao Luo, Chengjie Wang, Shouhong Ding, Yunsheng Wu, and Li Yuan. Df40: Toward next-generation deepfake detection. In Advances in Neural Information Processing Systems, pages 29387-29434. Curran Associates, Inc., 2024. 6 +[62] Zhiyuan Yan, Yandan Zhao, Shen Chen, Mingyi Guo, Xinghe Fu, Taiping Yao, Shouhong Ding, Yunsheng Wu, and + +Li Yuan. Generalizing deepfake video detection with plug-and-play: Video-level blending and spatiotemporal adapter tuning. In Proceedings of the Computer Vision and Pattern Recognition Conference (CVPR), pages 12615-12625, 2025, 2, 3, 6 +[63] Pengfei Yue, Beijing Chen, and Zhangjie Fu. Local region frequency guided dynamic inconsistency network for deepfake video detection. Big Data Mining and Analytics, 7(3): 889-904, 2024. 2, 6, 7 +[64] Daichi Zhang, Fanzhao Lin, Yingying Hua, Pengju Wang, Dan Zeng, and Shiming Ge. Deepfake video detection with spatiotemporal dropout transformer. In Proceedings of the 30th ACM international conference on multimedia, pages 5833-5841, 2022. 1, 2, 3 +[65] Daichi Zhang, Zihao Xiao, Shikun Li, Fanzhao Lin, Jianmin Li, and Shiming Ge. Learning natural consistency representation for face forgery video detection. In European Conference on Computer Vision, pages 407-424. Springer, 2024. 7 +[66] Hongyi Zhang, Moustapha Cisse, Yann N. Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. CoRR, abs/1710.09412, 2017. 6 +[67] Cairong Zhao, Chutian Wang, Guosheng Hu, Haonan Chen, Chun Liu, and Jinhui Tang. Istvt: Interpretable spatial-temporal video transformer for deepfake detection. IEEE Transactions on Information Forensics and Security, 18: 1335-1348, 2023. 1, 2, 3, 6, 7 +[68] Eric Zhao, Xiang Xu, Mingze Xu, Hui Ding, Yuanjun Xiong, and Wei Xia. Learning self-consistency for deepfake detection. In ICCV 2021, 2021. 3, 6 +[69] Hanqing Zhao, Wenbo Zhou, Dongdong Chen, Tianyi Wei, Weiming Zhang, and Nenghai Yu. Multi-attentional deepfake detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2185–2194, 2021. 1, 6 +[70] Wenliang Zhao, Yongming Rao, Weikang Shi, Zuyan Liu, Jie Zhou, and Jiwen Lu. Diffswap: High-fidelity and controllable face swapping via 3d-aware masked diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8568-8577, 2023. 6, 8 +[71] Yinglin Zheng, Jianmin Bao, Dong Chen, Ming Zeng, and Fang Wen. Exploring temporal coherence for more general video face forgery detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 15044-15054, 2021. 1, 2, 3, 5, 6, 7 +[72] Zhun Zhong, Liang Zheng, Guoliang Kang, Shaozi Li, and Yi Yang. Random erasing data augmentation. CoRR, abs/1708.04896, 2017. 4 +[73] Yuhao Zhu, Qi Li, Jian Wang, Chengzhong Xu, and Zhenan Sun. One shot face swapping on megapixels. In Proceedings of the IEEE conference on computer vision and pattern recognition (CVPR), pages 4834-4844, 2021. 1 +[74] Bojia Zi, Minghao Chang, Jingjing Chen, Xingjun Ma, and Yu-Gang Jiang. Wilddeepfake: A challenging real-world dataset for deepfake detection. Proceedings of the 28th ACM International Conference on Multimedia, 2020. 6, 8 \ No newline at end of file diff --git a/ICCV/2025/Vulnerability-Aware Spatio-Temporal Learning for Generalizable Deepfake Video Detection/images.zip b/ICCV/2025/Vulnerability-Aware Spatio-Temporal Learning for Generalizable Deepfake Video Detection/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..a00204c0a8fa9f24840631297e14c8834d5f271d --- /dev/null +++ b/ICCV/2025/Vulnerability-Aware Spatio-Temporal Learning for Generalizable Deepfake Video Detection/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:73193c52dbca0012daf3ef8592966461a438ad343a723e461743ebd13bcaedd9 +size 552424 diff --git a/ICCV/2025/Vulnerability-Aware Spatio-Temporal Learning for Generalizable Deepfake Video Detection/layout.json b/ICCV/2025/Vulnerability-Aware Spatio-Temporal Learning for Generalizable Deepfake Video Detection/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..b3ad8bab52392257c3d280f61ee910ea555408aa --- /dev/null +++ b/ICCV/2025/Vulnerability-Aware Spatio-Temporal Learning for Generalizable Deepfake Video Detection/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:67188dc08a5787151c0eae0bf164dd5a5fa2c3ad5ab8c7a2d2e6cc0efa58e3e1 +size 485425 diff --git a/ICCV/2025/WAVE_ Warp-Based View Guidance for Consistent Novel View Synthesis Using a Single Image/d2ddfab3-354c-4385-b1b5-8f76650ff96d_content_list.json b/ICCV/2025/WAVE_ Warp-Based View Guidance for Consistent Novel View Synthesis Using a Single Image/d2ddfab3-354c-4385-b1b5-8f76650ff96d_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..6b4c6375ad98d4137c419366af92341cb2a7746a --- /dev/null +++ b/ICCV/2025/WAVE_ Warp-Based View Guidance for Consistent Novel View Synthesis Using a Single Image/d2ddfab3-354c-4385-b1b5-8f76650ff96d_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:da2be98005691c4f1d3f244cdda40416adb307f0a0e33c136b6fff046cd132f3 +size 75968 diff --git a/ICCV/2025/WAVE_ Warp-Based View Guidance for Consistent Novel View Synthesis Using a Single Image/d2ddfab3-354c-4385-b1b5-8f76650ff96d_model.json b/ICCV/2025/WAVE_ Warp-Based View Guidance for Consistent Novel View Synthesis Using a Single Image/d2ddfab3-354c-4385-b1b5-8f76650ff96d_model.json new file mode 100644 index 0000000000000000000000000000000000000000..09920d220f98c66fa742b1d78342ac5ec50e285a --- /dev/null +++ b/ICCV/2025/WAVE_ Warp-Based View Guidance for Consistent Novel View Synthesis Using a Single Image/d2ddfab3-354c-4385-b1b5-8f76650ff96d_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a82be6a10ec1d05b7f455c0994daeb5152d135ffca586383cafa99a8b5fb3e84 +size 94373 diff --git a/ICCV/2025/WAVE_ Warp-Based View Guidance for Consistent Novel View Synthesis Using a Single Image/d2ddfab3-354c-4385-b1b5-8f76650ff96d_origin.pdf b/ICCV/2025/WAVE_ Warp-Based View Guidance for Consistent Novel View Synthesis Using a Single Image/d2ddfab3-354c-4385-b1b5-8f76650ff96d_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..fa8bf0c566af5c3cc39ac955232d25155c6b9e93 --- /dev/null +++ b/ICCV/2025/WAVE_ Warp-Based View Guidance for Consistent Novel View Synthesis Using a Single Image/d2ddfab3-354c-4385-b1b5-8f76650ff96d_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3f03971d18f3dbaad39b6c4bedb26f372541c57c846444f93a5fea32f83f1226 +size 4488077 diff --git a/ICCV/2025/WAVE_ Warp-Based View Guidance for Consistent Novel View Synthesis Using a Single Image/full.md b/ICCV/2025/WAVE_ Warp-Based View Guidance for Consistent Novel View Synthesis Using a Single Image/full.md new file mode 100644 index 0000000000000000000000000000000000000000..00b478650cfa78a0e569ff1a0b3cef8fec7e6a8a --- /dev/null +++ b/ICCV/2025/WAVE_ Warp-Based View Guidance for Consistent Novel View Synthesis Using a Single Image/full.md @@ -0,0 +1,323 @@ +# WAVE: Warp-Based View Guidance for Consistent Novel View Synthesis Using a Single Image + +Jiwoo Park Tae Eun Choi Youngjun Jun Seong Jae Hwang* + +Yonsei University + +{wldn1677, teunchoi, youngjun, seongjae}@yonsei.ac.kr + +![](images/5f1cca1f87bdc3b65429faf39114693ac2efe3b72939425e4241acb2576bc1ab.jpg) +LPIPS- $n e x t$ 1:0.221 + +![](images/86d62447120d87ccd164f622cf35de45157ef55e23a5976dfbcbf554915f2b4d.jpg) +Input View +WAVE + +![](images/e39285eaac325871f9cd591e225a3d9de2e419707319234559dcec585b7c2e29.jpg) + +![](images/c660cef8182bca9922c3a448cd5fd7128ddc3d5ba29881b76de4d45f375001c8.jpg) +Inference: 67.31s + +![](images/078cb67c9875885f6e182fd88a30cfb9e41535545065f416d58a7e18fa050f18.jpg) + +![](images/aac52c22b758256467ebfdb152328b060e46b9a99cb4e08ecdb951e042e5b793.jpg) +LPIPS- $n e x t$ 1 $(\downarrow)$ :0.359 + +![](images/a0583a9df0ef1203c28f0a10e09432715a6a112863ae855cf3c7f9a366a82578.jpg) + +![](images/3d7d500c681a0c9f88c2f41b9f8484955fe1a2d9a3da8480cc425e744514f68b.jpg) + +![](images/437589fe9bfee8583677205a50a7fbc8309a08a0639bae0a990ec616411e1534.jpg) +MegaScenes + +![](images/4c0314ef869e5387561b0e827b8adf46dc1ab71a4f2d0c89cce6cfa8f3d75de4.jpg) + +![](images/3a52844613ff3bf53da56c718aba8ae1cb896aae5e64f7f4c036fefc9f617556.jpg) + +![](images/16970f26f28ce1d88492821192f99507ffd32ecc3959831d824664171d6ec76d.jpg) + +![](images/7ab392d371a1e19cfdbc03574d29538346e26ba410f0be39efeb85492ee538ac.jpg) +LPIPS- $n e x t$ (↓): 0.234 + +![](images/d20f0bc32a9ba0430f07562eacb023f0b9e275a298290522816936cfd9abf1bc.jpg) +Figure 1. Teaser. We present WAVE, a training-free approach that enhances view consistency both among generated images and between the generated images and the input view. This approach is designed for novel view synthesis diffusion models using a input view. WAVE generates consistent images (lowest LPIPS- $^{next}$ ) across views without extra modules, enabling roughly 67s of inference time for 19 images. + +![](images/c6662edb30cceade2341ece14e39c6e585d8c6aaaeaf07510217c13090d75a98.jpg) + +![](images/dc08a986cba740ccfd6059d7e0b70b2203d6cfd4ffdc5e7c0f41b8230f909de3.jpg) +VistaDream + +![](images/04cad092b02f83673fb84941aaa17f57de1064c2f047053b449f521d758b7153.jpg) + +![](images/2f226fb00d24f5d6d747799799095e2f77aefa535d96cbeef2c1d4bebb3348df.jpg) +Inference: 851.13s + +![](images/b537b35d301d22d7031c6dfcc8a55642d43dfaf6cc14cfff92f70099cf241843.jpg) + +# Abstract + +Generating high-quality novel views of a scene from a single image requires maintaining structural coherence across different views, referred to as view consistency. While diffusion models have driven advancements in novel view synthesis, they still struggle to preserve spatial continuity across views. Diffusion models have been combined with 3D models to address the issue, but such approaches lack efficiency due to their complex multi-step pipelines. This paper proposes a novel view-consistent image generation method which utilizes diffusion models without additional modules. Our key idea is to enhance diffusion models with a training-free method that enables adaptive attention manipulation and noise reinitialization by leveraging view-guided warping to ensure view consistency. Through our comprehensive metric framework suitable for novel-view datasets, we show that our method improves view consistency across various diffusion models, demonstrating its broader applicability1 + +# 1. Introduction + +Scene-level novel view synthesis from a single image has received significant interest for its potential to generate unseen views from only one image [5, 13, 35, 38]. In single-image scenarios where multi-view information is unavailable, generative models have become a key component in novel view synthesis compared to rendering-based models [10, 19]. Specifically, diffusion models with strong generative capabilities have been adopted to generate novel views by utilizing view prior knowledge [6, 30]. Such approaches have led studies to explore diffusion models trained on large-scale datasets of images and camera parameters to generate novel views from a single image [25, 32]. Leveraging their generative power, diffusion models can synthesize unseen scenes with zero-shot capability. + +While diffusion-based models such as ZeroNVS [25] and MegaScenes [32] have demonstrated strong performance, a key limitation of these models is that they cannot maintain view consistency among generated images + +from different viewpoints, as shown in Fig. 2. For example, MegaScenes shows view inconsistencies due to variations in object appearance and color saturation, as illustrated in the second row of Fig. 1. To achieve view consistency, other studies [29, 35] have explored integrating diffusion models with additional models such as 3D models and large language models. In particular, VistaDream [35] leverages the 3D structures and contextual knowledge of vision language models to enhance consistency. Nevertheless, their applicability remains limited due to high computational costs and the complexity of multi-step pipelines. As shown in the third row of Fig. 1, despite the inference time being nearly ten times longer than that of MegaScenes, VistaDream still struggles to achieve satisfactory image quality. + +Even with various efforts to complement diffusion models, inherent factors within diffusion models that compromise view consistency remain unresolved [11, 20]. A major factor is that existing diffusion models generate each image separately [25, 32], disregarding the relationship between images. As a result, generated images lack interdependencies, making it challenging to maintain view consistency. Additionally, diffusion models exhibit high sensitivity to noise randomness during the generation process [8, 24, 37], leading to notable variations in the generated outputs. This variability impairs the model's ability to maintain overall visual coherence. Addressing these intrinsic factors directly could be a practical framework for achieving both view consistency and comparable inference time. Therefore, we introduce a streamlined method, WAVE, for view consistency that addresses the challenges inherent in diffusion models without the need for extra modules or training. + +To strengthen image interdependencies, we propose warp-guided adaptive attention, which integrates 3D warping [18] into batch self-attention to generate all viewpoints simultaneously. Batch self-attention computes attention by aggregating different keys and values across the batch, enabling information sharing between images [21, 31, 41]. However, batch self-attention fails to account for viewpoint changes, limiting its ability to capture view-dependent information in novel view synthesis. Thus, we utilize 3D warping which shifts pixels of an image from the input view to an arbitrary view. Integrated as a prior in batch self-attention, 3D warping provides cues for viewpoint transformations, improving view consistency. Moreover, noise randomness introduces unintended variations in generated images, diminishing view consistency. We propose pose-aware noise initialization which embeds the target image information into the initial diffusion noise to reduce noise randomness. Providing low-frequency information to the initial noise of diffusion models has proven effective in mitigating randomness [21, 37]. Building on this, our method adopts the low-frequency of warped images to inject pose-aware information into the initial noise. + +![](images/32e5e7f854b2b7228be77ae8cd128405068798d4040807600cdb6fbf7e91284e.jpg) +Figure 2. View inconsistency in diffusion models. Generated novel views using the single image on the left are presented. WAVE maintains view consistency, whereas other diffusion-based models like MegaScenes [32] and ZeroNVS [25] struggle to preserve consistency across viewpoints. + +As warped images inherently retain partial low-frequency details of the target view, they help alleviate noise randomness, ultimately enhancing view consistency. + +In a evaluation of view consistency improvement, we construct the comprehensive view consistency metric framework. This framework combines video metrics [7, 41] and camera pose accuracy metrics as reconstruction-based metrics are unreliable for consistency evaluation [12]. Additionally, we evaluate image quality on a sequential dataset and validate our method in a downstream task using 3D Gaussian splatting [10]. The contributions are as follows: + +- We propose WAVE, a training-free method for scene-level novel view synthesis, adaptable to various diffusion models without additional training. +- We present two approaches to enhance view consistency: (1) warp-guided adaptive attention to facilitate inter-view information flow, and (2) pose-aware noise initialization to mitigate the negative impact of noise randomness. +- Through our comprehensive metric framework along with extensive qualitative results, WAVE achieves improved performance over diffusion-based models, demonstrating comparable results to the rendering model. + +# 2. Related Work + +# 2.1. Novel View Synthesis from Single Scene Image + +Scene-level novel view synthesis from a single image has emerged as a crucial area of research, offering a feasible solution for generating realistic images on novel viewpoints while relying only on a single image. Unlike object-centric approaches [4, 15, 27, 34] that focus on generating isolated objects, scene-level synthesis [13, 14, 38] aims to generate complex and complete environments. + +Recent advancements in this domain have leveraged diffusion-based models, as they possess the generative capacity to generate novel views [11, 14, 20]. Consequently, efforts have been made to develop large-scale trained models [25, 32] on Stable Diffusion [23]. ZeroNVS [25] introduces scene-scale variables for fine-tuning at scene level. MegaScenes [32] has since addressed the viewpoint inaccuracies and limited generalization ability of ZeroNVS by employing a warping strategy for scene images. Various works [11, 20] have also focused on enforcing consistency between the input view and the generated novel view images. Despite these contributions, ensuring robust consistency across multiple views synthesized by diffusion models remains a fundamental challenge. + +To ensure view consistency, approaches such as VistaDream [35] and RealmDreamer [29] have integrated diffusion models with 3D models, reconstructing 3D from a single image or text inputs by leveraging 3D-aware representations. Although these approaches mitigate view inconsistency, they are made of multi-step pipelines that increases computational cost, making them less practical. While previous works have focused on auxiliary modules to assist diffusion models, the factors inherent to diffusion models which negatively impact view consistency remain unexplored. Accordingly, a practical approach that resolves these inherent factors needs to be explored. + +# 2.2. Image Consistency in Diffusion Models + +Diffusion models struggle to maintain consistency across image sequences in the fields of video generation [21, 41], text-to-image synthesis [8, 17, 31], and image synthesis [3, 36]. Various methods have been proposed for consistency by leveraging attention or cross-view conditioning without additional training [31, 39, 41]. ConsiStory [31] proposes a subject attention masking method to ensure consistent subject generation across different prompts, while StoryDiffusion [41] introduces an attention mechanism that enables the model to share contextual information. + +Other image-to-video generation methods [9, 21, 37] have emphasized the negative effect of noise randomness on consistency, attributing this issue to the discrepancy between noise in the training and inference stages. FreeInit [37] introduces iterative sampling to mix low-frequency information from latent variables into the noise, while ConsistI2V [21] incorporates low-frequency information from the first frame into the initial noise to mitigate randomness. + +These methods have primarily focused on preserving structural continuity. In novel view synthesis, however, attempts to consistently maintain view-dependent spatial information have yet to be made. Therefore, to ensure view consistency in novel view synthesis where view variation is a key factor, integrating geometric transformations for view changes into consistency methods is essential. + +![](images/9495cd902556fc9a68df76b7dee04afad9c6538df3b277e8f02e34a40bcbdd52.jpg) +Figure 3. Effectiveness of adaptive warp-range selection. We present examples of applying our method alongside the conventional batch self-attention method [31, 39] that references all viewpoints. The differences in the two camera pose accuracy metrics, Angular Consistency and Frobenius Norm, show that the previous method generates images that are less accurately aligned with the camera poses than ours. + +# 3. Methods + +We propose a training-free method that leverages warped images for consistent scene-level novel view synthesis using diffusion models pre-trained on large-scale datasets with camera poses and images [25, 32]. + +# 3.1. Preliminaries + +3D warping. 3D warping [18] is the process of transforming an input view image to match a desired camera pose using its depth. The transformation involves unprojecting image pixels into 3D space using depth values, followed by reprojecting them according to the specified camera poses. This approach enables the generation of warped images that accurately reflect the new viewpoints. + +Batch self-attention. Self-attention [33] extracts image features by applying linear projections with the attention matrices $W_{q}, W_{k}, W_{v}$ to produce $Q, K, V$ . The attention map $A$ is then computed as $\text{softmax}(QK^T / \sqrt{d_k})$ and multiplied by $V$ to generate hidden features $h$ . Otherwise, batch self-attention [31, 39] aggregates the keys and values from all $N$ generations in a batch. Batch self-attention is defined as: + +$$ +\begin{array}{l} Q _ {i} \in \mathbb {R} ^ {p \times d _ {q}}, \quad K _ {i} \in \mathbb {R} ^ {p \times d _ {k}}, \quad V _ {i} \in \mathbb {R} ^ {p \times d _ {v}}, \\ K ^ {*} = \left[ K _ {1}, K _ {2}, \dots , K _ {N} \right] \in \mathbb {R} ^ {N \cdot p \times d _ {k}}, \\ V ^ {*} = \left[ V _ {1}, V _ {2}, \dots , V _ {N} \right] \in \mathbb {R} ^ {N \cdot p \times d _ {v}}, \\ A _ {i} = \operatorname {s o f t m a x} \left(Q _ {i} K ^ {* ^ {\top}} / \sqrt {d _ {k}}\right) \in \mathbb {R} ^ {p \times N \cdot p}, \\ h _ {i} = A _ {i} \cdot V ^ {*} \in \mathbb {R} ^ {p \times d _ {v}}. \tag {1} \\ \end{array} +$$ + +Batch self-attention enables the model to generate images from multiple viewpoints in a single forward pass, preventing diffusion models from generating images independently. However, referencing all views with nonoverlapping and unnecessary information reduces the reliance on the desired camera pose view features. This makes + +![](images/ba806a5978225607f5cd2c4577f31270da22df12f625ddd48febe55bf5e90380.jpg) +Figure 4. Illustration of WAVE. Given an input view, depth map and continuous camera poses, our method generates scene images with smooth view transitions through three distinct processes: (a) adaptive warp-range selection utilizes warped region masks, from which the relevance between viewpoints is determined to compute the reference range for attention. (b) pose-aware noise initialization (PANI) re-initializes the diffusion initial noise by leveraging warped images and initial noise, incorporating frequency domain information. (c) warp-guided adaptive attention (WGAA) utilizes the warped region masks and reference range obtained from the adaptive warp-range selection, performing masked batch self-attention. + +it difficult to generate an image that matches the desired camera pose. To adequately reflect view variations, instead of referring to all views, we adaptively select the reference views (Section 3.2) and develop an attention mechanism that masks non-overlapping regions within these selected views (Section 3.3). + +# 3.2. Adaptive Warp-range Selection + +Referencing all other viewpoints in the batch as in batch self-attention [31, 39] reduces viewpoint accuracy, as shown in Fig. 3 and Supp. Table 3. Fig. 3 also shows examples of misalignment with the desired viewpoint. Aggregating information from non-overlapping views ultimately degrades viewpoint accuracy. Therefore, we set a reference range of viewpoints within the attention mechanism to utilize information from only a relevant subset of viewpoints. + +To determine the optimal reference range, we perform the adaptive warp-range selection shown in Fig. 4a. We utilize warped images from 3D warping to compute viewpoint relevance. As the viewpoint shifts, regions that are not visible from the input view become missing regions in the warped images, assigned a value of zero. In order to create warped region masks $M$ , we convert warped images into binary regions $W_{1:N}$ , where pixels with a value of zero remain 0, and non-zero pixels are set to 1. Then, to measure the relevance between viewpoints (e.g., $i$ and $j$ ), we compute the IoU $u_{ij}$ on warped region masks to construct the matrix $U = \{u_{ij}\} \in \mathbb{R}^{N \times N}$ , which provides a precise mea + +sure of their spatial relevance [22]. Based on this matrix, the adaptive range $\left[\mu_i - \sigma_i, \mu_i + \sigma_i\right]$ is selected for each viewpoint, for mean $\mu_i$ and standard deviation $\sigma_i$ . Applying the standard deviation of the IoU matrix enables the dynamic determination of reference points, making our approach applicable across various sets of viewpoints. + +# 3.3. Warp-Guided Adaptive Attention + +Moreover, in batch self-attention, as the number of referenced images increases, each image becomes overly dependent on information from other images rather than its own. To prevent excessive reliance on external information, previous work proposed a method of subject attention masking based on the subject mask to facilitate the selection of information [31]. Attention masking allows the model to selectively focus on the important information. Building on this insight, we design warp-guided attention masking, which employs warped region masks in the diffusion decoder as shown in Fig. 4c, as follows: + +$$ +A _ {i} = \operatorname {s o f t m a x} \left(Q _ {i} K ^ {* ^ {\top}} / \sqrt {d _ {k}}\right), +$$ + +$$ +W _ {1: N} = \mathcal {T} (\mathcal {I}, \text {p o s e} _ {1: N}) \odot \mathbf {1} _ {\mathcal {T} (\mathcal {I}, \text {p o s e} _ {1: N}) \neq 0}, +$$ + +$$ +M _ {i} = \left[ W _ {1}, W _ {2}, \dots , \mathbf {1}, \dots , W _ {N} \right], +$$ + +$$ +h _ {i} = \left(A _ {i} \odot M _ {i}\right) \cdot V ^ {*} \in \mathbb {R} ^ {p \times d _ {v}}, \tag {2} +$$ + +where $\mathcal{T}$ is a 3D warping operation. This approach leverages the assumption that the decoder's attention mechanism encodes spatial information relevant to different views. For details of the assumption, please refer to the supplementary + +![](images/1c641f9f7ed5a6d0de3d68578584aaf45b0f82280bcb8a2a438367d16d27c446.jpg) +(a) Without noise + +![](images/296c907bff17d127844fe8db55c941915deddfa86fcb00212099c79a02a29bd1.jpg) +(b) Fill noise +Figure 5. Analysis of warping methods. We present the generation results of the methods attempted to mitigate the negative impact of missing regions in warped images : (a) not filling missing regions with noise, (b) only filling missing regions with noise, (c) our method (low-frequency normalization and filling noise). + +![](images/629ac558636147e3b2995e804355443ea40d4a7d84655446bdf27a4cbd769795.jpg) +(c) WAVE + +material. These warped region masks encapsulate the viewpoint cues of the target view image to be generated. The warped region masks effectively serve as a masking technique by capturing the spatial differences between viewpoints, allowing for a focused integration of important information. Warp-guided adaptive attention performs warped region mask attention based on the adaptive range obtained through adaptive warp-range selection. This enables the attention mechanism to effectively reflect viewpoint changes. + +# 3.4. Pose-Aware Noise Initialization + +Beyond independent generation, noise randomness in diffusion models is another key issue. The stochastic nature of noise adversely impacts image consistency [8, 24, 37], which extends similarly to view consistency. + +View consistency, unlike image consistency, requires integrating information regarding view variations. Therefore, we propose pose-aware noise initialization which incorporates such information for noise reinitialization using warped images to enhance view consistency, as illustrated in Fig. 4b. Warped images capture the partial low-frequency of the target images, reflecting variations across different viewpoints. Leveraging them for noise reinitialization mitigates noise randomness by integrating geometric information from view changes into the noise. The process is summarized in Algorithm 1. Warped images $\mathcal{W}$ are generated, aligned with the given camera poses. These images are passed through the VAE encoder $\mathcal{E}$ , and DDPM noise is then added to make $z_{T}$ . Low-frequency information is extracted from $z_{T}$ and mixed with the high-frequency components of the noise to preserve essential information in warped images while introducing randomness in the high-frequency parts to enhance visual details [21, 37]. + +Warped images contain missing regions due to the absence of information from unseen viewpoints. Injecting low-frequency information directly from warped images results in undesirable black holes, as shown in Fig. 5a. This occurs because missing regions are reflected in the generated outputs. To address this, we apply random noise to fill the missing regions $\mathcal{W}'$ before injection, suppressing low-frequency associated with these regions. Otherwise, the low-frequency components of the random noise interfere + +Algorithm 1 Pose-Aware Noise Initialization + +Input: Reference image $\mathcal{I}$ , camera poses pose $_{1:N}$ + +Depth map $\mathcal{D}$ + +Output: re-initialized noise $\epsilon^{\prime}$ + +Note: Gaussian filter $\mathcal{G}(\mathcal{D}_0)$ , VAE encoder $\mathcal{E}$ + +$$ +\mathcal {W} \leftarrow \{\mathcal {W} _ {1}, \mathcal {W} _ {2}, \dots , \mathcal {W} _ {n} \} +$$ + +1: Apply 3D Warping Algorithm using $\mathcal{I},\mathcal{D}$ , and pose $1:N$ +2: $\mathcal{W}' \gets \mathcal{W} + \eta \cdot \mathbf{1}_{W=0} \quad \triangleright$ Add noise $\eta$ to zero regions of $W$ +3: $z_0\gets \mathcal{E}(W)$ , Sample $\epsilon \sim \mathcal{N}(0,I)$ +4: $z_{T} \gets \mathrm{DDPM~Forward}(z_{0}, T)$ ▷ noise scheduler +5: $F_T^{\mathrm{low}}\gets \mathrm{FFT}(z_T)\odot \mathcal{G}(\mathcal{D}_0)$ +6: $F_T^{\mathrm{low}} \gets F_T^{\mathrm{low}} / \mathrm{mean}(|F_T^{\mathrm{low}}|)$ ▷ mean of $F_T^{\mathrm{low}}$ 's magnitude +7: $F_{\epsilon}^{\mathrm{high}} \gets \mathrm{FFT}(\epsilon) \odot (1 - \mathcal{G}(\mathcal{D}_0))$ +8: $\epsilon^{\prime}\gets \mathrm{IFFT}(F_T^{\mathrm{low}} + F_\epsilon^{\mathrm{high}})$ +9: Return $\epsilon'$ + +with the original low-frequency of the warped images, leading to artifacts, as illustrated in Fig. 5b. To mitigate this, we normalize low-frequency components based on their magnitudes, controlling the impact of the random noise. This suppression of noise, as shown in Fig. 5c, ensures a more consistent generation process. + +# 4. Experiments + +Baselines & Datasets. Our method is compared with other diffusion-based approaches, including MegaScenes [32] and ZeroNVS [25], as well as a multi-step pipelines model VistaDream [35]. Experiments conducted on the datasets MegaScenes [32], RealEstate10K (RE10K) [40], DTU [1], and Mip-NeRF 360 [2] encompass synthetic, indoor, and outdoor scenes to ensure our method's generalizability. + +Metric framework. Several novel-view datasets [1, 32] lack sequential viewpoints and ground-truth images, complicating the use of reconstruction metrics for consistency evaluation. The reliability of reconstruction metrics in evaluating consistency has been questioned [12]. To address this, we design a view consistency metric framework that jointly evaluates image consistency and camera pose accuracy. This framework includes video metrics (LPIPS- $n e x t$ / first and CLIPSIM- $n e x t$ / first) alongside camera pose accuracy metrics (Frobenius Norm of rotation matrix, Rotation Angle Difference, and Angular Consistency). For further details, please refer to Supp. Section C. + +Experimental results based on our view consistency metric framework, along with results for reconstruction quality using the video dataset [40] are presented in Section 4.1. Additionally, Section 4.2 presents qualitative comparisons against existing methods, Section 4.3 covers the ablation study, and Section 4.4 evaluates our method on a down + +Table 1. Consistency and camera accuracy metrics. We conduct experiments on the three datasets MegaScenes, DTU, and RE10K and evaluate consistency using two metrics: Next metric, which assesses consistency with neighboring viewpoints, and First metric, which measures consistency with the input view. Additionally, to evaluate camera accuracy, a crucial aspect of novel view synthesis, we report the accuracy of camera extrinsic parameters using Frobenius Norm, Rotation Angle Difference, and Angular Consistency. + +
MethodNextFirstFrobenius Norm (Rotation) ↓Rotation Angle Difference ↓Angular Consistency ↓
LPIPS ↓CLIPSIM ↑LPIPS ↓CLIPSIM ↑
MegaScenes
ZeroNVS0.5610.8870.6020.8260.5630.41023.53
ZeroNVS + WAVE0.4040.9330.5450.8510.5570.40423.17
MegaScenes0.6070.8610.6090.8130.4810.35520.35
MegaScenes + WAVE0.3970.9200.5610.8260.3820.27715.91
DTU
ZeroNVS0.4890.8830.7260.7170.5680.41423.77
ZeroNVS + WAVE0.3440.9250.7040.7390.5500.40123.00
MegaScenes0.3790.8980.6920.6970.3080.22713.04
MegaScenes + WAVE0.1300.9560.6340.7160.1550.1106.32
RE10K
ZeroNVS0.4780.8970.7640.7450.6240.45926.30
ZeroNVS + WAVE0.3880.9210.7610.7430.6200.45526.10
MegaScenes0.3470.9120.7220.6760.2420.17910.25
MegaScenes + WAVE0.2620.9480.7280.6950.1490.1086.20
+ +![](images/adac0b37cd8680b27b8f28dce2f4a09b48349e6753f91a1593f512cf7aac3135.jpg) +Figure 6. Overall comparison of methods. The size of the circles represents inference time. Applying our method improves performance with comparable computation costs. + +stream task. Additional results and experimental settings can be found in the supplementary material. + +# 4.1. Comparison with Baselines + +Table 1 compares diffusion-based models with our method. Our method steadily outperforms diffusion-based models across diverse datasets, showing adaptability to diffusion models. Specifically, on the MegaScenes dataset, our method achieves 0.397 in LPIPS- ${nxt}$ and 0.920 in CLIPSIM- ${nxt}$ ,while generating images with higher accuracy in camera viewpoints. VistaDream [35] incorporates a rendering method giving it an advantage in consistency metrics over diffusion models. Thus, VistaDream is compared based on consistency metrics and inference time, unlike other models. Notably, as shown in Fig. 6, the result of MegaScenes+WAVE exhibits significantly lower inference time than VistaDream while achieving comparable performance. This shows that solving the inherent factors + +Table 2. RE10K sequence evaluation. We present the results of image reconstruction and generation quality on RE10K, a video sequence dataset with varying viewpoints, demonstrating that our method outperforms the baseline. + +
MethodPSNR ↑LPIPS ↓FID ↓KID ↓
VistaDream10.880.56729.480.011
ZeroNVS10.840.61224.130.010
ZeroNVS + WAVE11.380.60225.670.009
MegaScenes12.020.46918.200.006
MegaScenes + WAVE13.270.45015.780.004
+ +of diffusion models can achieve a performance similar to rendering-based methods without excessive computational cost, demonstrating the practicality of our method. + +Moreover, to evaluate image quality not captured by our metric framework, we employ the image sequence (i.e., video) dataset, RE10K, to assess the reconstruction and generation capabilities (Table 2). This dataset contains both ground-truth images and camera poses for varying views, allowing the evaluation of image quality. The results show that our method outperforms existing models in both image generation (FID, KID) and reconstruction (PSNR, LPIPS) metrics. The results demonstrate that our method produces consistent images while achieving higher image quality. + +# 4.2. Qualitative Comparison + +We compare our method against baselines with qualitative results. As shown in Fig. 7, MegaScenes struggles to maintain consistency across different viewpoints. In contrast, WAVE produces more coherent results, addressing the limitations of existing diffusion models. While VistaDream maintains view consistency, it exhibits prominent artifacts in generated images. WAVE achieves superior performance compared to VistaDream, demonstrating that better-quality images can be generated without relying on multiple mod + +![](images/29b3ca1ae96920b2bbd78e5f8ee224855d377d0f21b515652f07fc81070b31cc.jpg) +Figure 7. Qualitative results. We provide samples for our method and baseline models [25, 32]. The results are obtained by generating images from an input view and continuous camera poses, selecting representative samples from them. Our method exhibits consistent images among generated images. The middle column presents the input view, while the remaining columns display the generated results. + +els. More qualitative results and camera parameter visualization can be found in the supplementary materials. + +# 4.3. Ablation Study + +We conduct ablative experiments to study the effect of each component in our method. Ablation experiments are performed on the MegaScenes dataset and MegaScenes pretrained model [32]. The quantitative and qualitative results are shown in Table 3 and Fig. 8, respectively. + +Effect of warp-guided adaptive attention. Warp-guided adaptive attention improves performances across all metrics + +compared to the base model in Table 3. As shown in Fig. 8, it helps maintain the spatial structure of most scenes, although color variations persist. This indicates that our attention mechanism preserves structural features by sharing information from other generative processes. Additionally, a detailed comparison with conventional batch self-attention is provided in the supplementary material. + +Effect of pose-aware noise initialization. Applying pose-aware noise initialization also improves performance over the diffusion model in Table 3. This shows that pose-aware noise initialization effectively reflects view transformations, + +Table 3. Ablation results. We analyze different components of our model to assess their impact on the overall performance. PANI refers to the pose-aware noise initialization in Section 3.4, while WGAA denotes the warp-guided adaptive attention in Section 3.2. + +
Exp. #PANIWGAAFirstNextAngular Consistency ↓
LPIPS ↓CLIPSIM ↑LPIPS ↓CLIPSIM ↑
10.6090.8130.6070.86120.35
2✓0.5840.8190.5600.87319.76
3✓0.5810.8220.4160.91816.49
4✓✓0.5610.8260.3970.92015.91
+ +![](images/6af1d7159db4951097e8bf0143cda290213bd5ed4a871fd7a472ad362349c770.jpg) +Figure 8. Visual comparisons in ablation study. We compare of the generated results when PANI and WGAA are removed from our method, respectively. We generate images along an orbital trajectory, ranging from a specific leftward angle to a specific rightward angle relative to the input image. + +helping to reduce noise randomness. As shown in Fig. 8, while objects fail to appear across generated views, color variations do not occur. These results imply that the randomness in noise contributes to color inconsistencies. + +# 4.4. Downstream Task + +Additionally, we validate that the consistent images generated using our method effectively enhance the performance of 3D rendering tasks on various datasets [1, 2, 32, 40]. Since 3D rendering tasks rely on multiple images, ensuring view consistency is crucial for providing high-quality multi-view information to 3D models, ultimately improving 3D rendering performance [16, 28]. Thus, this task can be used to evaluate whether the models generate consistent images. Previous studies [25, 26, 30] on diffusion-based 3D rendering with a single image have explored training 3D models using novel view synthesis diffusion model's guidance. However, this approach makes it challenging to assess the diffusion model itself, as it jointly optimizes the 3D model's loss and diffusion guidance loss. + +We conduct an evaluation by directly using the generated outputs from the diffusion model as inputs to a 3D model to assess rendering performance. In this experiment, VistaDream is excluded since its pipeline already includes 3D rendering. 3D rendering experiments are conducted using + +Table 4. 3D rendering downstream task. We evaluate 3D rendering performance using 3D Gaussian Splatting [10] with images generated by our method and baselines. In Mip-NeRF 360, ZeroNVS [25] fails to generate reliable images without a rendering technique, making it unsuitable for 3D rendering. + +
MethodPSNR ↑SSIM ↑LPIPS ↓
MegaScenes
ZeroNVS23.380.8690.098
ZeroNVS + WAVE24.230.8520.094
MegaScenes22.110.7980.128
MegaScenes + WAVE24.690.8360.096
DTU
ZeroNVS23.530.8470.127
ZeroNVS + WAVE25.560.8710.101
MegaScenes20.430.8090.142
MegaScenes + WAVE26.950.8950.078
Mip-NeRF 360
ZeroNVS---
ZeroNVS + WAVE---
MegaScenes25.070.8290.074
MegaScenes + WAVE26.300.8390.082
+ +3D Gaussian Splatting [10] with 19 image sets generated by our method and other baseline models [25, 32]. As shown in Table 4, our approach outperforms the original methods. These results indicate that our generated consistent images enhance geometric accuracy and visual coherence in 3D reconstruction and rendering tasks. The potential for future application of our method in diffusion-based 3D rendering tasks is also demonstrated. Results for datasets not included in Table 4 are in the supplementary material. + +# 5. Conclusion + +We propose a training-free method to improve view consistency in diffusion models for scene-level novel view synthesis from a single image, offering a practical solution for this task. By leveraging inter-view information through attention and noise reinitialization with warped images, our approach overcomes key limitations of diffusion models, including independent generation and noise randomness. Lastly, we address the lack of reliable evaluation metrics by introducing a comprehensive metric process, providing a valuable benchmark for future research. + +Limitations and future work. The use of warped images makes the method sensitive to large viewpoint changes, leading to significant missing regions. An autoregressive approach, generating images within a specific view range and iteratively using them for subsequent generations, could address this issue through gradual warping and offer a promising direction for future work. + +# References + +[1] Henrik Aanaes, Rasmus Ramsbøl Jensen, George Vogiatzis, Engin Tola, and Anders Bjorholm Dahl. Large-scale data for multiple-view stereopsis. International Journal of Computer Vision, pages 1-16, 2016. 5, 8 +[2] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5470-5479, 2022. 5, 8 +[3] Giannis Daras, Yuval Dagan, Alex Dimakis, and Constantinos Daskalakis. Consistent diffusion models: Mitigating sampling drift by learning to be consistent. Advances in Neural Information Processing Systems, 36, 2024. 3 +[4] Matt Deitke, Ruoshi Liu, Matthew Wallingford, Huong Ngo, Oscar Michel, Aditya Kusupati, Alan Fan, Christian Laforte, Vikram Voleti, Samir Yitzhak Gadre, et al. Objverse-xl: A universe of $10\mathrm{m} + 3\mathrm{d}$ objects. Advances in Neural Information Processing Systems, 36, 2024. 2 +[5] Ruiqi Gao, Aleksander Holynski, Philipp Henzler, Arthur Brussee, Ricardo Martin-Brualla, Pratul Srinivasan, Jonathan T Barron, and Ben Poole. Cat3d: Create anything in 3d with multi-view diffusion models. arXiv preprint arXiv:2405.10314, 2024. 1 +[6] Jiatao Gu, Alex Trevithick, Kai-En Lin, Joshua M Susskind, Christian Theobalt, Lingjie Liu, and Ravi Ramamoorthi. Nerfdiff: Single-image view synthesis with nef-guided distillation from 3d-aware diffusion. In International Conference on Machine Learning, pages 11808-11826. PMLR, 2023. 1 +[7] Yuwei Guo, Ceyuan Yang, Anyi Rao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Sparsectrl: Adding sparse controls to text-to-video diffusion models. In European Conference on Computer Vision, pages 330-348. Springer, 2024. 2 +[8] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 2, 3, 5 +[9] Yuming Jiang, Tianxing Wu, Shuai Yang, Chenyang Si, Dahua Lin, Yu Qiao, Chen Change Loy, and Ziwei Liu. Videobooth: Diffusion-based video generation with image prompts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6689-6700, 2024. 3 +[10] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42 (4), 2023. 1, 2, 8 +[11] Seungryong Kim. Genwarp: Single image to novel views with semantic-preserving generative warping. In 38th Conference on Neural Information Processing Systems Workshop, NeurIPS 2024. Neural information processing systems foundation, 2024. 2, 3 +[12] Jeong-gi Kwak, Erqun Dong, Yuhe Jin, Hanseok Ko, Shweta Mahajan, and Kwang Moo Yi. Vivid-1-to-3: Novel view synthesis with video diffusion models. In Proceedings of + +the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6775-6785, 2024. 2, 5 +[13] Zhengqi Li, Qianqian Wang, Noah Snavely, and Angjoo Kanazawa. Infinitenature-zero: Learning perpetual view generation of natural scenes from single images. In European Conference on Computer Vision, pages 515-534. Springer, 2022. 1, 2 +[14] Andrew Liu, Richard Tucker, Varun Jampani, Ameesh Makadia, Noah Snavely, and Angjoo Kanazawa. Infinite nature: Perpetual view generation of natural scenes from a single image. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14458-14467, 2021. 2, 3 +[15] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9298–9309, 2023. 2 +[16] Xiaoxiao Long, Yuan-Chen Guo, Cheng Lin, Yuan Liu, Zhiyang Dou, Lingjie Liu, Yuexin Ma, Song-Hai Zhang, Marc Habermann, Christian Theobalt, et al. Wonder3d: Single image to 3d using cross-domain diffusion. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9970-9980, 2024. 8 +[17] Oscar Manas, Pietro Astolfi, Melissa Hall, Candace Ross, Jack Urbanek, Adina Williams, Aishwarya Agrawal, Adriana Romero-Soriano, and Michal Drozdal. Improving text-to-image consistency via automatic prompt optimization. arXiv preprint arXiv:2403.17804, 2024. 3 +[18] William R Mark, Leonard McMillan, and Gary Bishop. Postrendering 3d warping. In Proceedings of the 1997 symposium on Interactive 3D graphics, pages 7-ff, 1997. 2, 3 +[19] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 1 +[20] Norman Müller, Katja Schwarz, Barbara Rössle, Lorenzo Porzi, Samuel Rota Bulò, Matthias Nießner, and Peter Kontschieder. Multidiff: Consistent novel view synthesis from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10258-10268, 2024. 2, 3 +[21] Weiming Ren, Huan Yang, Ge Zhang, Cong Wei, Xinrun Du, Wenhao Huang, and Wenhu Chen. Consisti2v: Enhancing visual consistency for image-to-video generation. Transactions on Machine Learning Research. 2, 3, 5 +[22] Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian Reid, and Silvio Savarese. Generalized intersection over union: A metric and a loss for bounding box regression. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 658-666, 2019. 4 +[23] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 3 + +[24] Subham Sekhar Sahoo, Aaron Gokaslan, Chris De Sa, and Volodymyr Kuleshov. Diffusion models with learned adaptive noise. arXiv preprint arXiv:2312.13236, 2023. 2, 5 +[25] Kyle Sargent, Zizhang Li, Tanmay Shah, Charles Herrmann, Hong-Xing Yu, Yunzhi Zhang, Eric Ryan Chan, Dmitry Lagun, Li Fei-Fei, Deqing Sun, et al. Zeronvs: Zero-shot 360-degree view synthesis from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9420–9429, 2024. 1, 2, 3, 5, 7, 8 +[26] Junyoung Seo, Wooseok Jang, Min-Seop Kwak, Hyeonsu Kim, Jaehoon Ko, Junho Kim, Jin-Hwa Kim, Jiyoung Lee, and Seungryong Kim. Let 2d diffusion model know 3d-consistency for robust text-to-3d generation. arXiv preprint arXiv:2303.07937, 2023. 8 +[27] Ruoxi Shi, Hansheng Chen, Zhuoyang Zhang, Minghua Liu, Chao Xu, Xinyue Wei, Linghao Chen, Chong Zeng, and Hao Su. Zero123++: a single image to consistent multi-view diffusion base model, 2023. 2 +[28] Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d generation, 2024. 8 +[29] Jaidev Shriram, Alex Trevithick, Lingjie Liu, and Ravi Ramamoorthi. Realmdreamer: Text-driven 3d scene generation with inpainting and depth diffusion. 2025. 2, 3 +[30] Jiaxiang Tang, Jiawei Ren, Hang Zhou, Ziwei Liu, and Gang Zeng. Dreamgaussian: Generative gaussian splatting for efficient 3d content creation. arXiv preprint arXiv:2309.16653, 2023. 1, 8 +[31] Yoad Tewel, Omri Kaduri, Rinon Gal, Yoni Kasten, Lior Wolf, Gal Chechik, and Yuval Atzmon. Training-free consistent text-to-image generation. ACM Transactions on Graphics (TOG), 43(4):1-18, 2024. 2, 3, 4 +[32] Joseph Tung, Gene Chou, Ruojin Cai, Guandao Yang, Kai Zhang, Gordon Wetzstein, Bharath Hariharan, and Noah Snavely. Megascenes: Scene-level view synthesis at scale. In European Conference on Computer Vision, pages 197-214. Springer, 2025. 1, 2, 3, 5, 7, 8 +[33] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 3 +[34] Vikram Voleti, Chun-Han Yao, Mark Boss, Adam Letts, David Pankratz, Dmitry Tochilkin, Christian Laforte, Robin Rombach, and Varun Jampani. Sv3d: Novel multi-view synthesis and 3d generation from a single image using latent video diffusion. In European Conference on Computer Vision, pages 439-457. Springer, 2024. 2 +[35] Haiping Wang, Yuan Liu, Ziwei Liu, Wenping Wang, Zhen Dong, and Bisheng Yang. Vistadream: Sampling multiview consistent images for single-view scene reconstruction. arXiv preprint arXiv:2410.16892, 2024. 1, 2, 3, 5, 6 +[36] Chen Henry Wu and Fernando De la Torre. A latent space of stochastic diffusion models for zero-shot image editing and guidance. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7378-7387, 2023. 3 +[37] Tianxing Wu, Chenyang Si, Yuming Jiang, Ziqi Huang, and Ziwei Liu. Freeinit: Bridging initialization gap in video diffusion models. In European Conference on Computer Vision, pages 378-394. Springer, 2025. 2, 3, 5 + +[38] Hong-Xing Yu, Haoyi Duan, Junhwa Hur, Kyle Sargent, Michael Rubinstein, William T Freeman, Forrester Cole, Deqing Sun, Noah Snavely, Jiajun Wu, et al. Wonderjourney: Going from anywhere to everywhere. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6658-6667, 2024. 1, 2 +[39] Chuanxia Zheng and Andrea Vedaldi. Free3d: Consistent novel view synthesis without 3d representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9720-9731, 2024. 3, 4 +[40] Tinghui Zhou, Richard Tucker, John Flynn, Graham Fyffe, and Noah Snavely. Stereo magnification: learning view synthesis using multiplane images. ACM Transactions on Graphics (TOG), 37(4):1-12, 2018. 5, 8 +[41] Yupeng Zhou, Daquan Zhou, Ming-Ming Cheng, Jiashi Feng, and Qibin Hou. Storydiffusion: Consistent self-attention for long-range image and video generation. arXiv preprint arXiv:2405.01434, 2024. 2, 3 \ No newline at end of file diff --git a/ICCV/2025/WAVE_ Warp-Based View Guidance for Consistent Novel View Synthesis Using a Single Image/images.zip b/ICCV/2025/WAVE_ Warp-Based View Guidance for Consistent Novel View Synthesis Using a Single Image/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..e76a89e5b4a67c3cf2bf5abeb1986bde4038166e --- /dev/null +++ b/ICCV/2025/WAVE_ Warp-Based View Guidance for Consistent Novel View Synthesis Using a Single Image/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:198ff34c413291853d3bdf960b648643fefe9f9a74b327fbc0f4c8fb0857782f +size 882325 diff --git a/ICCV/2025/WAVE_ Warp-Based View Guidance for Consistent Novel View Synthesis Using a Single Image/layout.json b/ICCV/2025/WAVE_ Warp-Based View Guidance for Consistent Novel View Synthesis Using a Single Image/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..5c16c018af49b169e2b7c28eab6392ea91f101fd --- /dev/null +++ b/ICCV/2025/WAVE_ Warp-Based View Guidance for Consistent Novel View Synthesis Using a Single Image/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:58071f02d13ffdb80311508cabe9b3e30e0af387c53080474fe9a6df2b433f4c +size 367689 diff --git a/ICCV/2025/WINS_ Winograd Structured Pruning for Fast Winograd Convolution/c1353ff7-9892-4a0b-8c5f-7b946b4ec7c3_content_list.json b/ICCV/2025/WINS_ Winograd Structured Pruning for Fast Winograd Convolution/c1353ff7-9892-4a0b-8c5f-7b946b4ec7c3_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..bc813a217735dc68c6c04d1e8be088977871a1cd --- /dev/null +++ b/ICCV/2025/WINS_ Winograd Structured Pruning for Fast Winograd Convolution/c1353ff7-9892-4a0b-8c5f-7b946b4ec7c3_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:59424afc32910e6cc7006f9724d9eea3db7367f3a354897fd60e74cf95c82da2 +size 74733 diff --git a/ICCV/2025/WINS_ Winograd Structured Pruning for Fast Winograd Convolution/c1353ff7-9892-4a0b-8c5f-7b946b4ec7c3_model.json b/ICCV/2025/WINS_ Winograd Structured Pruning for Fast Winograd Convolution/c1353ff7-9892-4a0b-8c5f-7b946b4ec7c3_model.json new file mode 100644 index 0000000000000000000000000000000000000000..e1b40b46e6bfb6c5c88a39f29405639a22f9f90b --- /dev/null +++ b/ICCV/2025/WINS_ Winograd Structured Pruning for Fast Winograd Convolution/c1353ff7-9892-4a0b-8c5f-7b946b4ec7c3_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cbb6c4f7d63772f4544a67747a06292b3cc6452d9ba3aa6f7275ba2724749e91 +size 95993 diff --git a/ICCV/2025/WINS_ Winograd Structured Pruning for Fast Winograd Convolution/c1353ff7-9892-4a0b-8c5f-7b946b4ec7c3_origin.pdf b/ICCV/2025/WINS_ Winograd Structured Pruning for Fast Winograd Convolution/c1353ff7-9892-4a0b-8c5f-7b946b4ec7c3_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..f4aa879be7a9610bf4fcaf45381830692fac11b1 --- /dev/null +++ b/ICCV/2025/WINS_ Winograd Structured Pruning for Fast Winograd Convolution/c1353ff7-9892-4a0b-8c5f-7b946b4ec7c3_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3b8938789bb6c1921868c99b5ecf5da8b87c710ffb9a082ebca367ad2f850314 +size 2600535 diff --git a/ICCV/2025/WINS_ Winograd Structured Pruning for Fast Winograd Convolution/full.md b/ICCV/2025/WINS_ Winograd Structured Pruning for Fast Winograd Convolution/full.md new file mode 100644 index 0000000000000000000000000000000000000000..5664e44d1909e136ce78d89a49115bd76dad289d --- /dev/null +++ b/ICCV/2025/WINS_ Winograd Structured Pruning for Fast Winograd Convolution/full.md @@ -0,0 +1,283 @@ +# WINS: Winograd Structured Pruning for Fast Winograd Convolution + +Cheonjun Park1, Hyun Jae Oh3, Mincheol Park2, Hyunchan Moon4, Minsik Kim3, Suhyun Kim5, Myung Kuk Yoon6, and Won Woo Ro3* + +$^{1}$ Hankuk University of Foreign Studies, $^{2}$ Samsung Advanced Institute of Technology, $^{3}$ Yonsei University, $^{4}$ LG Electronics, $^{5}$ Kyung Hee University, $^{6}$ Ewha Womans University + +{cheonjun.park, hyunjae.oh, mincheol.park, minsik.kim, wro}@yonsei.ac.kr, {mhcqwe92, dr.suhyun.kim}@gmail.com, myungkuk.yoon@ewha.ac.kr + +# Abstract + +Recent GPUs leverage Winograd convolution and structured pruning to significantly accelerate inference. First, Winograd convolution is theoretically $2.25 \times$ faster than standard convolution. Second, structured pruning reduces inference time without additional overhead as the pruning ratio increases. However, applying conventional structured pruning alongside Winograd convolution is inefficient. Existing structured pruning methods, which do not account for how GPUs process Winograd convolution, require large pruning unit sizes, leading to significant information loss. In this paper, we propose Winograd Structured Pruning (WINS), the first approach to employ optimized structured pruning for Winograd convolution. WINS is designed based on an in-depth analysis of Winograd convolution's computational characteristics on GPUs. Additionally, we introduce two variants, WINS-B and WINS-AB, which further enhance performance. Experimental results show that WINS-AB achieves up to $2.8 \times$ practical speedup in baseline inference on GPUs while preserving the accuracy of ResNet-18 on ImageNet. + +# 1. Introduction + +CNNs are a powerful tool for computer vision tasks [29, 51]. However, their high computational demands have hindered their practical, on-device applications in mobile devices [45]. To overcome these challenges, industrial leaders, such as AMD and NVIDIA, have accepted network pruning [42, 50] and Winograd convolution [30, 57] to yield compact CNNs and process them rapidly on their GPGPU features. Both optimization methods could achieve a greater than a two-fold reduction in the processing time of CNNs [1, 60, 61, 68]. Model compression techniques such as pruning are essential for the advancement of on-device + +![](images/02e0adad973e594c75d2cc13f162c317c9017e10e29b6013aa1331f7c322fddc.jpg) +Figure 1. Comparison of speedup among (a) structured pruning, (b) Winograd weight pruning, and (c) WINS on ResNet-18. + +AI, and their importance is increasingly emphasized. In particular, CNNs still achieve the highest performance in accuracy-critical applications, and Winograd convolution is used in the latest models [5, 54]. + +Despite the success of network pruning and Winograd convolution in moving towards on-device applications, combining both techniques is still limited. Fine-grained pruning (e.g., weight pruning [15] and balanced pruning [41, 63]) shows strong sparsity by zeroing extremely large weights for the compression of CNNs. However, Winograd convolution [39] manipulates the sparse weights using linear transformations, which removes the sparsity and generates dense matrices. On the other hand, middle-grained pruning (e.g., tile pruning [36] and block sparse [43]) can avoid such a manipulation owing to its bigger pruning unit size than the transformation unit size. Despite its compatibility, the holistic use of middle-grained pruning and Winograd convolution poses a challenge in model deployment, as it makes irregular, sparse transformed matrices that need special-purpose hardware for efficient processing [26, 64]. Coarse-grained pruning (e.g., filter pruning [19, 32]) makes regularly coalesced data in the Winograd convolution. However, this technique is limited to relieving a significant information loss due to its large pruning unit size [49] (see Fig. 1-(a)). + +One prior work [39] proposes Winograd weight pruning to tackle the vanishing sparsity of fine-grained pruning in + +![](images/ac29bd32cf0bdf867f990069b962044ca18fb3f4834944c639be29c7b6d2d6cb.jpg) +Figure 2. Overview of Winograd convolution: Conversion from (a) theoretical computation to (b) GPU-accelerated execution. + +Winograd convolution. This technique applies the pruning process in the Winograd domain directly to the transformed data to preserve static sparsity. Hence, it can achieve high pruning ratios, but its element-wise computations lead to irregular memory access on GPUs. Therefore, its real-world speedups are restricted [64] (see Fig. 1-(b)). + +This paper proposes Winograd Structured Pruning (WINS), which proactively prunes the Winograd-transformed matrices to accelerate their processing in the GPU kernel (see Fig. 1-(c)). WINS considers three crucial factors to obtain GPU-aware sparsity in Winograd convolution: the optimal pruning unit size, workload balance between matrices, and layer importance. + +- The optimal pruning unit size: WINS starts with deciding the optimal pruning unit size. The pruning unit size is determined based on the basic unit of execution on GPUs (e.g., warp and wavefront) to utilize GPU parallelism while minimizing accuracy drops [48]. +- Workload balance between matrices: WINS then obtains a uniform pruning ratio (i.e., WINS-B) to make equal-sized segmented matrices. These matrices are executed simultaneously on multiple streaming multiprocessors (SMs) within the GPUs while satisfying the load-balanced executions over GEMMs. Balancing GEMMs is critical. If the execution times differ, SMs that terminate GEMMs earlier should be waiting for terminations of other SMs' GEMMs. This waiting time leads to delays in executing subsequent layers. +- Layer importance: When WINS-B is applied to all layers, the accuracy drops can outweigh the speedups, and these speedups may vary layer by layer. In some layers, replacing WINS-B with WINS can benefit from retaining accuracy while maintaining similar speedups. To address this, WINS-AB exploits our proposed selection algorithm that automatically chooses WINS or WINS-B layer-by-layer + +using the Adaptive Batch Normalization-based evaluation technique [31]. This paper demonstrates that this strategy effectively reduces accuracy drops while achieving a $2.8 \times$ speedup. + +# 2. Preliminary: Winograd Convolution + +Prerequisites The $n$ filters of $l^{th}$ convolution layer are as $\mathcal{F}^{(l)}\in \mathbb{R}^{n\times c\times k_h\times k_w}$ , where $n$ , $c$ , $k_{h}$ , and $k_{w}$ are the number of filters, the number of channels, height, and width of the filters, respectively. + +# 2.1. Process of Winograd Convolution + +As shown in Fig. 2-(a) and Equation 1, there are five steps in Winograd convolution [30]: ① input transformation (ITrans), ② filter transformation (FTrans), ③ Element-Wise Matrix Multiplication (EWMM), ④ channel-wise summation, and ⑤ output transformation. The $p \times p$ input patch $(d)$ , the basic input matrix of Winograd convolution, is extracted with a stride of $(p - 2)$ from the original input feature map. The input feature map has $ip$ number of input patches. The $p \times p$ input patch is convolved with a $3 \times 3$ filter $(g)$ to produce an $(p - 2) \times (p - 2)$ output patch $(S)$ [39]. + +$$ +S = A ^ {T} \left[ \left(G g G ^ {T}\right) \odot \left(B ^ {T} d B\right) \right] A \tag {1} +$$ + +Input and filter transformation transform spatial-domain input patch $(d)$ and filter $(g)$ to Winograd-domain input patch $(B^T dB)$ and filter $(GgG^T)$ using predefined matrices $B$ and $G$ , respectively [30]. After input and filter transformation, there is EWMM $(\odot)$ between $B^T dB$ and $GgG^T$ . Channelwise summation reduces the channel of EWMM output. Output transformation transforms Winograd-domain output patch to spatial-domain output patch $S$ using a predefined mapping matrix $A$ . The output patches are assembled into an output feature map. The matrices $B$ , $G$ , and $A$ used + +![](images/00b242f30633495e8570c51feff9c47fdddbc5587683052d4b7757de4e70535f.jpg) +Figure 3. Overview of previous pruning methods and the proposed WINS for Winograd convolution. + +in input, filter, and output transformations are made from Winograd's minimal algorithm [57]. + +# 2.2. Conversion for Winograd Convolution + +Inefficient process of Winograd convolution According to ARM [17, 23], EWMM accounts for most of the inference time of Winograd convolution, so EWMM is a target for acceleration. Moreover, EWMM even results in an inefficient operation on the GPU. GPU consists of dozens of independent functional modules called Streaming Multiprocessors (SMs) for parallel processing. During EWMM on GPUs, a 3D input patch $(B^T dB)$ and a 3D filter $(GgG^T)$ are loaded and processed in a single SM. As illustrated in Fig. 2-(a), to calculate filter $GgG^T$ #1 and input $B^T dB$ #1, element $\underline{\mathbf{i}}$ of filter $GgG^T$ #1 is calculated only once with element $\underline{\mathbf{l}}$ of input patch $B^T dB$ #1 and is not used again in the SM #1. Element $\underline{\mathbf{i}}$ of filter $GgG^T$ #1 is not reused within the SM #1, so the required data (element $\underline{\mathbf{i}}\mathbf{x}$ of input patch $B^T dB$ #2) is obtained from another SM #2. However, overhead occurs when the SMs exchange data because the memory of each SM is independent. + +Efficient Winograd convolution on GPU To solve the low computational intensity and frequent memory access overhead of EWMM, NVIDIA [24, 61] and ARM [23] convert and process Winograd convolution's EWMM and channel-wise summation to BGEMM (Batched General Matrix Multiplication) operations. BGEMM maintains a structured form to maximize data reuse on the GPU. + +This conversion process is called EC2B (conversion from EWMM and Channel-wise summation to BGEMM). As shown in Fig. 2-(b), BGEMM consists of a total of $p^2$ GEMM operations with the same shape. A GEMM of BGEMM is a matrix product operation between the weight matrix of $\mathbb{R}^{n \times c}$ shape and the input matrix of $\mathbb{R}^{c \times ip}$ shape. For parallelism, one GEMM is assigned to each SM. In GEMM1, the weight vector $(\underline{\mathbf{i}} \sim \mathbf{i}\mathbf{v})$ and the input vector $(\underline{\mathbf{1}} \sim \mathbf{4})$ generate output element $(\mathbf{I})$ by performing MAC operations at once. The BGEMM rearranges the input vectors $(\underline{\mathbf{9}} \sim \underline{\mathbf{12}})$ just next the weight vectors $(\underline{\mathbf{1}} \sim \underline{\mathbf{4}})$ in the same GEMM1 to increase the reuse of the weight vectors $(\underline{\mathbf{i}} \sim \underline{\mathbf{i}}\mathbf{v})$ . Thus, EC2B is a data rearrangement method that maximizes data reuse within a single SM. + +# 3. Related Work + +Compatibility between Winograd convolution and network pruning is challenging, but the following techniques can be compatible with Winograd convolution. + +Filter Pruning (FP) In standard convolution, filter pruning [18, 32, 53, 66] removes rows (filter pruning) in converted weight matrix using im2col [3] for multiple 3D filter tensors. According to the Spatial-Winograd pruning [65], which is the appliance of filter pruning on Winograd convolution, the pruning unit size is a 3D filter size. Therefore, as shown in Fig. 3-(a), the removal of filters is not influenced by filter transformations into 2D matrices within Winograd convolution. The FP pruned model shows + +![](images/9740faacf4871ab68e70be6083750d14ca805055a2665d615cfba8601f7554f9.jpg) +(a) + +![](images/e9472ab7b7caf355fc91b86a7c660346f43a2510441b6dd169a976d36940a06e.jpg) +(b) + +![](images/409144df71708e801b9987073b72ea6bc1051a6b2dc7164cc77cc353a2e30bd2.jpg) +(c) +Figure 4. (a) Comparison of pruning unit size on ResNet-18. (b) Comparison of accuracy with respect to pruning ratio for four pruning methods of ResNet-18 on ImageNet. The models were fine-tuned for one epoch. (c) Comparison of GEMM speedups by pruning between irregular data (cuSPARSE [44]) and regular data (cuBLAS [47]). The speedups are measured on the 6th convolution layer of VGG-16. + +dense weights in Winograd convolution, so FP has no index computation overhead in GPU kernel [56]. + +Winograd Weight Pruning (WWP) WWP removes redundant values of Winograd-transformed matrices in a unit of an element to reduce the matrix size of Winograd convolution. Sparse Winograd CNN [38] introduces the first attempt at pruning and re-training method in the Winograd domain. Winograd CNN Native Pruning [33] presents the $90.0\%$ sparsity of the matrices obtained within AlexNet's Winograd domain, demonstrating less than $0.1\%$ accuracy drops. Winograd-ReLU CNN Pruning [39] exploits dynamic sparsity on activated features and transformed matrices in the Winograd-domain. However, this work requires post-hoc conversion in the network architecture to process Winograd convolution rapidly. All the WWP methods are element-level pruning, which creates irregular, sparse data in the Winograd domain. The sparse data requires significant index computation with different latency of memory access. Hence, this memory divergence [27] indices hurt the massive throughput of the GPU, which requires a specialized hardware kernel or a dedicated accelerator for WWP [40, 55, 58, 62] for practical use. Therefore, we apply a GPU-friendly sparsity in Winograd convolution to reduce the overhead of index computation. + +# 4. Winograd Structured Pruning (WINS) + +(a) Motivation 1: Large pruning unit size problem of FP As shown in Fig. 3-(a) and Equation 2, in FP, the matrix multiplication has remained in the structured form, which minimizes the index computation overhead [56], even after the EC2B conversion is applied. + +$$ +S = A ^ {T} \left[ \left(G (\text {F i l t e r P r u n e} (g)) G ^ {T}\right) \odot \left(B ^ {T} d B\right) \right] A \tag {2} +$$ + +Such a regular structured form can be incredibly beneficial when general-purpose computing devices such as GPUs are used for inference. Another benefit of the structured form is that this can be easily calculated using the GEMM of existing GPU library [25, 47], well-optimized. However, + +the side-effect of filter pruning is representation loss when the models are aggressively pruned. As shown in Fig. 3-(a), the pruning unit size of filter pruning on the Winograd-domain is $p^2 \times \mathbb{R}^{1 \times c}$ , which tends to be too large, so it removes the important weights with non-important weights together. For example, on ResNet-18, the pruning unit sizes of filter pruning on the Winograd-domain are either 1024, 2048, 4096, and 8192 parameters depending on the layers as shown in Fig. 4-(a). In Fig. 4-(b), when evaluating filter pruning with pruning ratios of $20\%$ and $40\%$ in ResNet-18, the Top-1 accuracy is reduced by $5\%$ and $10\%$ , respectively. Although filter pruning keeps the matrices in a structured form to leverage the full computing capabilities of existing general-purpose computing devices, it shows significant accuracy drops in the high pruning ratio, even with retention. + +(b) Motivation 2: Irregular data problem of WWP As shown in Fig. 3-(b) and Equation 3, a pruned model from WWP shows irregular data access patterns before and after applying EC2B conversion because WWP is an elementwise pruning method $(\mathbb{R})$ + +$$ +S = A ^ {T} \left[ \text {W e i g h t P r u n e} \left(G g G ^ {T}\right) \odot \left(B ^ {T} d B\right) \right] A \tag {3} +$$ + +Irregular data pattern is inefficient for parallel processors, especially on GPUs [11]. This pattern is calculated by converting weights into a sparse data format (e.g., CSR) not to calculate the zero values in hardware. Therefore, as shown in Fig. 4-(c), the inference time of the pruned model (relying on the sparse GEMM) is not faster than the inference time of the unpruned model (relying on the dense GEMM) due to the considerable amount of index computation [21, 64]. To obtain the computational benefits from WWP pruned models, the support of special-purpose hardware, such as NPU [40], is necessary. + +(c) Method We propose WINS that satisfies both regularity and pruning unit size as small as possible for efficient processing on GPUs. We apply structured pruning in the Winograd domain, as shown in Fig. 3-(c) and Equation 4. + +![](images/cd4b967b44c91a37e528df998d5e145187a1111a4d80398727aa9ed9eff55695.jpg) +Figure 5. Overview of the three proposed methods: (a) WINS, (b) WINS-B, and (c) WINS-AB. ResNet-18 on ImageNet. (d) Comparison of performance improvements in layers obtained from applying balancing on WINS. (e) Comparison of Top-1 accuracy between WINS and WINS-B. The fine-tuning epoch is one. (f) Correlation of outlook accuracy (100 iterations) and fine-tuning accuracy (40 epochs). + +To preserve the sparsity by FTrans, the WINS applies pruning after FTrans. As described in Sec. 2.2, Winograd convolution is transformed into BGEMM for efficient processing in GPUs (consists of $16(p^{2})$ large-sized sub-GEMMs.). To achieve the regular weight structure, our proposed WINS removes weight parameters vector by vector from each sub-GEMM weight matrix (see Fig. 5-(a)). The pruning unit size $(\mathbb{R}^{1\times c})$ of the WINS is $16(p^{2})$ times smaller than that $(p^2\times \mathbb{R}^{1\times c})$ of the filter pruning, which provides delicate removal of weights for maintaining high model representation. The WINS calculates the threshold value using the target pruning ratio $(PR)$ . The WINS obtains the representative value of the vector using the $\ell_2$ -norm method [32] (see Sec. 6.1). If the absolute value of the representative value is less than a threshold value, it is set to 0 to remove the weight vector. WINS can be divided into row-wise WINS (WINS-R) and column-wise WINS (WINS-C), which remove the row and column vectors, respectively. These methods remove the row vector $(\mathbb{R}^{1\times c})$ and column vector $(\mathbb{R}^{n\times 1})$ of the sub-GEMM, respectively. In the Appendix and Fig. 4(b), we analyze two techniques (WINS-R and WINS-C) and found that row type is more effective. + +$$ +S = A ^ {T} \left[ S t r u c t u r e d P r u n e \left(G g G ^ {T}\right) \odot \left(B ^ {T} d B\right) \right] A \tag {4} +$$ + +
ResNet-18 on ImageNet
Pruning Ratio (PR)30%40%50%60%70%
Relative Speedup (WINS-B ÷ WINS)1.15x1.22x1.35x1.38x1.43x
+ +Table 1. Comparison between WINS and WINS-B with respect to relative speedup (inference time) over various pruning ratios. + +# 5. Enhanced Method + +# 5.1. Balanced WINS (WINS-B) + +(a) Motivation WINS exhibits an imbalanced characteristic attributed to the different $PR$ in each of the 16 ( $p^2$ ) sub-HEMAIs. This imbalance leads to an under-throughput on GPUs, as each sub-HEMA is assigned to an SM for independent execution [34]. In particular, SMs processing sub-HEMAIs with larger $PRs$ tend to complete their computations before SMs assigned to sub-HEMAIs with smaller $PRs$ , leading to idle periods (see Fig. 5-(a)). This idle period can induce a delay in the execution of the subsequent layers. + +(b) Method To address the imbalance, we improve WINS by introducing a new approach called Balanced WINS (i.e., WINS-B) (see Fig. 5-(b)). With WINS-B, we ensure that + +all sub-GEMMs have equal $PR$ . We employ a specific order for pruning to achieve this equal $PR$ across all sub-GEMMs. During WINS, the pruning threshold value is calculated using the $\ell_2$ -norm, taking into account all sub-GEMMs, resulting in different PR values for each sub-GEMM. However, in WINS-B, these representative values are considered within each sub-GEMM using the $\ell_2$ -norm rather than being assessed collectively across all GEMMs. In this context, Table. 1 shows that the WINS-B has a speed gain of averaged $30.6\%$ on ResNet-18 compared to WINS. + +# 5.2. Adaptive Balanced WINS (WINS-AB) + +(a) Motivation WINS-B enhances the load balancing in BGEMM operations by emphasizing regularity in the sparse weights of the WINS concept. Fig. 5-(d) illustrates the advantages of WINS-B, up to $25\%$ better. A distinctive acceleration is achieved across entire layers by ensuring an evenly distributed sparsity in the weight matrices involved in BGEMM. However, Fig. 5-(e) reveals a trade-off: The accuracy of WINS-B exhibits a disadvantage compared to relying on WINS alone. This observation is understandable since significant values need to be unintentionally excluded in network pruning. To maximize accelerating inference and preserving accuracy, we introduce an adaptive framework that applies WINS-B or WINS selectively, simultaneously considering evaluation accuracy and speedup due to workload balance. + +(b) Method We introduce the Adaptive Balanced WINS (i.e., WINS-AB) method (see Fig. 5-(c)). WINS-AB is an auto algorithm that automatically adjusts the number of layers applied to WINS-B or WINS when the accuracy loss of the WINS-B pruned model becomes significant. WINS-AB operates in two main ways: Pre-processing and Evaluation & Pruning. + +[1] Pre-processing In Fig. 5-(c), first, we measure the speedup of all the layers from WINS-B (called Balanced Speedup). Second, we search for pruning candidates that may result from applying WINS or WINS-B to each layer. If the number of pruning candidates is too large, we use Monte Carlo sampling to reduce the cases. Additionally, we use layer-wise outlook accuracy and balanced speedup data to pre-determine and fix certain layers to WINS or WINS-B, significantly reducing the number of cases. + +[2] Evaluation & Pruning Initially, the layer-by-layer balanced speedup obtained from pre-processing is used to calculate the total balanced speedup for each case. Next, an Adaptive Batch Normalization-based evaluation method [31] is applied to WINS-AB to measure the "outlook accuracy" of each candidate. This evaluation method uses minimal training iteration and validation data (e.g., 100-iteration). As seen in Fig. 5-(f), outlook accuracy and accuracy are highly correlated. Finally, the one with the + +highest balanced speedup and within the top $\alpha \%$ of outlook accuracy is chosen as the winner and undergoes fine-tuning. After fine-tuning, if the accuracy drop of the winner is negligible, the winner is added to the pruned model list. Since there might be a better case than the winner case, we halve the $\alpha$ value to search for a new winner case. Conversely, if the winner case shows a significant decrease in accuracy, we delete that winner case and double the $\alpha$ value to search for a new winner case. After about $\beta$ -iterations of the WINS-AB auto algorithm, the best model from the pruned model list is selected as the final pruned model. + +(c) Discussion about $\alpha$ and $\beta$ Through multiple heuristic experiments, we found that setting $\alpha$ and $\beta$ to 30 and 5, respectively, yields optimal results. The values of $\alpha$ and $\beta$ are interdependent; an excessively large $\alpha$ necessitates a sufficient number of iterations to identify a valid winner. Meanwhile, increasing $\beta$ improves the likelihood of finding a winner that meets the criteria. However, this comes at the cost of significantly increased training time, making it impractical to increase $\beta$ arbitrarily. + +# 6. Experiments + +We evaluate the performance of WINS-AB using CIFAR-10 [28], ImageNet [4], Pascal VOC [6], and COCO [37] (see Appendix). We exploit the GPU kernel based on the API [12] to accelerate Winograd convolution. We perform fine-tuning with only 65 epochs after conducting pruning methods. We set a batch size of 256. We use SGD optimizer with the weight decay, $1 \times 10^{-4}$ , and the momentum as 0.9 for fine-tuning. The initial learning rate is set to 0.001 and divided by 10 every 30 epochs. + +# 6.1. Comparison with Representative Value + +Structured pruning removes unnecessary vectors by leveraging a representative value of each vector. The representative value can be determined in various ways, such as $\ell_1$ -norm [15], $\ell_2$ -norm [15], and Geometric Median [20], with $\ell_2$ -norm being widely used in recent LLM model pruning [52]. In Table.2, We conducted experiments by applying these three methods to WINS pruning. The results showed that for ResNet-20 and ResNet-32, the WINS pruned model using $\ell_2$ -norm achieved the highest accuracy under the same conditions, while for ResNet-56, the WINS pruned model using the Geometric Median method demonstrated better accuracy. Based on these findings, we decided to use $\ell_2$ -norm-based WINS pruning in our experiments, as it consistently showed strong performance. + +# 6.2. Pruning Strategies for Point-wise Convolutions + +WINS is a structured pruning method specialized for Winograd convolution. Previous studies have indicated that + +
ModelMethodRepresentative ValuePruning RatioBaseline AccuracyPruned AccuracyGAP
ResNet-20WINSl1-norm [15]80%91.77%91.39%-0.38%
WINSl2-norm [32]80%91.77%91.51%-0.26%
WINSGeometric Median [20]80%91.77%91.42%-0.35%
ResNet-32WINSl1-norm [15]90%92.63%91.33%-1.30%
WINSl2-norm [32]90%92.63%91.84%-0.79%
WINSGeometric Median [20]90%92.63%91.82%-0.81%
ResNet-56WINSl1-norm [15]90%93.01%91.85%-1.16%
WINSl2-norm [32]90%93.01%92.08%-0.93%
WINSGeometric Median [20]90%93.01%92.56%-0.45%
+ +Table 2. Comparison of accuracy on CIFAR-10. GAP is the accuracy difference between baseline and pruned model. + +
Res2Net-50 [10] on ImageNet
MethodPruning Ratio (PR)Top-1 Accuracy
3×3 convolution1×1 convolution3×3 convolution1×1 convolutionBaselinePrunedGAP
WINS-ABFPGM [20]65%30%77.16%76.16%-1.00%
WINS-ABBCBP [48]65%30%77.16%76.32%-0.84%
WINS-ABFPGM [20]85%50%77.16%75.23%-1.93%
WINS-ABBCBP [48]85%50%77.16%75.87%-1.29%
+ +Table 3. Comparison of pruning strategies suitable for $3 \times 3$ and $1 \times 1$ convolutions. + +Winograd convolution is most effective for $3 \times 3$ convolution layers. However, recent research suggests that Winograd convolution can also be effective for depth-wise separable convolution [7] and large-kernel convolution [22]. Meanwhile, for point-wise convolution $(1 \times 1)$ , conventional structured pruning methods remain appropriate. Based on this, we applied WINS along with two structured pruning methods [20, 48] simultaneously. As shown in the Table. 3, applying WINS together with BCBP achieves higher accuracy than combining it with filter pruning. + +# 6.3. Comparison between WINS-B and WINS-AB + +As shown in Fig. 6-(a), we comprehensively compare the proposed models - WINS, WINS-B, and WINS-AB, considering multiple metrics such as $PR$ , accuracy, and speed. In Fig. 6-(a), at a $PR$ of $80\%$ , WINS and WINS-AB have an accuracy drop of less than $1\%$ , while WINS-B experiences the most significant accuracy drop at $1.27\%$ . Interestingly, at the same $80\%$ $PR$ , WINS-AB outperforms WINS in terms of speed, being approximately 1.2 times faster. Based on this comprehensive comparison, WINS-AB is the most effective approach for ResNet-18 compared with the other methods. + +# 6.4. Comparison of WINS-AB and Other Prunings + +We evaluate our WINS-AB method against previous Winograd-compatible pruning techniques (i.e., WWP and FP) based on accuracy, $PR$ , and speedup. + +Comparison with FP As shown in Fig. 6-(b) and (c), FP and WINS-AB are faster than a dense model because + +FP and WINS-AB keep a structured form after filter transformation. However, FP's pruning unit size is too large to avoid an accuracy drop because FP creates a structured form without considering the Winograd convolution. Since WINS-AB is a pruning method considering Winograd convolution, the pruning unit size of WINS-AB is about $p^2$ times less than that of FP. Fig. 6-(b) shows that FP exceeds $2.0\%$ accuracy drop from $30\%$ $PR$ on ResNet-18. On the other hand, even if WINS-AB prunes up to $80\%$ $PR$ , the accuracy drop of WINS-AB does not exceed $0.6\%$ in ResNet-18. In ResNet-18 and VGG-16, within $1.5\%$ accuracy drop, WINS-AB is $1.84\times$ and $1.27\times$ faster than FP, respectively (see Fig. 6-(b) and (c)). + +Comparison with WWP Both WWP and WINS-AB are pruning techniques that consider Winograd convolution. Since WINS-AB's pruning unit size is about $c$ times larger than that of WWP, WWP can prune more precisely. As shown in Fig. 6-(c), when the $PR$ is $70\%$ in VGG-16, WWP has $2.0\%$ higher accuracy than WINS-AB, respectively. However, despite pruning up to $90\% PR$ in VGG-16, WWP is slower than the unpruned model due to the significant index computation of WWP. WINS-AB maintains a structural form, so WINS-AB is $2.01\times$ faster than Winograd convolution when $PR$ is $84\%$ (see Fig. 6-(b)). + +# 6.5. Further Comparison with Various FP + +Table.4 illustrates the evaluation of our WINS-AB method on the ImageNet, utilizing various CNNs such as ResNet18, ResNet-50, SqueezeNet-1.0, and DenseNet-121. For ResNet-18, FPGM [20] and PFP [35] exhibit accuracy reductions exceeding $1.8\%$ at $PR$ of $41.8\%$ and $43.8\%$ , re + +![](images/6871934641523fa367d559371f830d32099734dca66562c1a0e3efcf86797e9c.jpg) + +![](images/3db78c39cc11b3574bb44138be413585d515b00f429e94e4d6ca14f602d08b86.jpg) + +![](images/368a8d68cfafaf967afb610c8f5596bdeb61e1c581362f26ab69dab738d99dc8.jpg) + +Figure 6. Comparison of speedup (inference time) and accuracy in various pruning ratios. We set $\alpha$ and $\beta$ of WINS-AB to 30 and 5. + +
MethodPruning Ratio (%)Pruned FLOPs (%)Top-1 Accuracy (%)Inference timeSpeed up
BaselinePrunedGAP
ResNet-18 (baseline)--69.70--3.16ms1.00×
UDSP [9]20.0-69.7669.48-0.282.51ms1.26×
SOSP [46]39.1-69.8069.60-0.201.88ms1.68×
WINS-AB [ours]40.1-69.7069.900.201.87ms1.69×
FPGM [20]41.8-70.2068.40-1.801.85ms1.71×
PFP [35]43.8-69.7067.30-2.401.83ms1.73×
WHC [2]-41.869.8068.50-1.30--
RLAL [8]-50.069.8069.00-0.80--
WINS-AB [ours]84.075.169.7069.00-0.701.12ms2.82×
ResNet-50 (baseline)--76.10--12.53ms1.00×
UDSP [9]25.0-76.1376.510.389.74ms1.29×
ATO [59]-61.776.1376.07-0.06--
APIB [13]58.062.076.1575.37-0.786.06ms2.07×
Torque [14]64.5-76.0774.58-1.495.34ms2.35×
WINS-AB* [ours]80.069.576.1076.430.333.61ms3.47×
SqueezeNet-1.0 (baseline)--58.20--1.81ms1.00×
\( \ell_2 \)-norm [32]9.7-58.2055.30-2.901.61ms1.12×
WINS-AB [ours]28.1-58.2057.00-1.201.48ms1.23×
DenseNet-121 (baseline)--74.60--9.21ms1.00×
C-SGD [16]29.9-74.5073.70-0.808.10ms1.14×
LWP [67]35.2-74.6072.80-1.806.92ms1.33×
WINS-AB‡ [ours]85.0-74.6074.20-0.405.55ms1.66×
+ +Table 4. Comparison with FP on ImageNet [4]. The GEMM operation with the unpruned weight parameters of WINS-AB and input values is accelerated using customized CUTLASS API [25]. This symbol $(\star)$ means 'prune using BCBP [48] at $1\times 1$ convolution layer'. This symbol $(\ddagger)$ means 'no prune at $1\times 1$ convolution layer'. GAP is the Top-1 accuracy difference between baseline and pruned model. + +spectively. In contrast, WINS-AB shows a minimal accuracy drop of $0.7\%$ at $84.0\%$ PR. While WINS-AB demonstrates a slight speed limitation against FPGM [20] at a comparable $40\%$ PR due to the time required for index processing of sparse matrices, it significantly excels in speed by 2.82 times at an $84\%$ PR with better Top-1 accuracy. For ResNet-50, WINS-AB achieves $1.82\%$ higher accuracy and has 1.47 times faster inference time, even with up to $15.5\%$ more PR than Torque [14] because it has a smaller pruning unit size compared to FP. + +# 7. Conclusion + +We propose WINS that accelerates Winograd convolution in consideration of Winograd convolution computational characteristics on a parallel processor. We perform vector-wise pruning in BGEMM operation, which brings delicate structured data pattern compared to WWP and FP, for accuracy maintenance and speedup. WINS-AB also makes balanced matrices for computing utilization, and has realistic CNN representation power at a high pruning ratio. + +# Acknowledgements + +This work was supported by Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government(MSIT) (RS-2021-II212068, Artificial Intelligence Innovation Hub). + +# References + +[1] Renzo Andri, Beatrice Bussolino, Antonio Cipolletta, Lukas Cavigelli, and Zhe Wang. Going further with winograd convolutions: Tap-wise quantization for efficient inference on 4x4 tile. arXiv preprint arXiv:2209.12982, 2022. 1 +[2] Shaowu Chen, Weize Sun, and Lei Huang. Whc: Weighted hybrid criterion for filter pruning on convolutional neural networks. In ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1-5. IEEE, 2023. 8 +[3] Sharan Chetlur, Cliff Woolley, Philippe Vandermersch, Jonathan Cohen, John Tran, Bryan Catanzaro, and Evan Shelhamer. codnn: Efficient primitives for deep learning. arXiv preprint arXiv:1410.0759, 2014. 3 +[4] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 6, 8 +[5] Kaiwen Duan, Song Bai, Lingxi Xie, Honggang Qi, Qingming Huang, and Qi Tian. Centernet++ for object detection. IEEE transactions on pattern analysis and machine intelligence, 46(5):3509-3521, 2023. 1 +[6] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International journal of computer vision, 88:303-338, 2010. 6 +[7] Yonggan Fu, Haichuan Yang, Jiayi Yuan, Meng Li, Cheng Wan, Raghuraman Krishnamoorthi, Vikas Chandra, and Yingyan Lin. Depthshrinker: a new compression paradigm towards boosting real-hardware efficiency of compact neural networks. In International Conference on Machine Learning, pages 6849-6862. PMLR, 2022. 7 +[8] Alireza Ganjdanesh, Shangqian Gao, and Heng Huang. Jointly training and pruning cnns via learnable agent guidance and alignment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16058-16069, 2024. 8 +[9] Shangqian Gao, Yanfu Zhang, Feihu Huang, and Heng Huang. Bilevelpruning: Unified dynamic and static channel pruning for convolutional neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16090-16100, 2024. 8 +[10] Shang-Hua Gao, Ming-Ming Cheng, Kai Zhao, Xin-Yu Zhang, Ming-Hsuan Yang, and Philip Torr. Res2net: A new multi-scale backbone architecture. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(2):652-662, 2021. 7 +[11] Joseph L Greathouse and Mayank Daga. Efficient sparse matrix-vector multiplication on gpus using thecsr storage + +format. In SC'14: Proceedings of the International Conference for High Performance Computing, Networking, Storage and Analysis, pages 769-780. IEEE, 2014. 4 +[12] Cong Guo, Bo Yang Hsueh, Jingwen Leng, Yuxian Qiu, Yue Guan, Zehuan Wang, Xiaoying Jia, Xipeng Li, Minyi Guo, and Yuhao Zhu. Accelerating sparse dnn models without hardware-support via tile-wise sparsity. In Proceedings of the International Conference for High Performance Computing, Networking, Storage and Analysis. IEEE Press, 2020. 6 +[13] Song Guo, Lei Zhang, Xiawu Zheng, Yan Wang, Yuchao Li, Fei Chao, Chenglin Wu, Shengchuan Zhang, and Rongrong Ji. Automatic network pruning via hilbert-schmidt independence criterion lasso under information bottleneck principle. In Proceedings of the IEEE/CVF international conference on computer vision, pages 17458-17469, 2023. 8 +[14] Arshita Gupta, Tien Bau, Joonsoo Kim, Zhe Zhu, Sumit Jha, and Hrishikesh Garud. Torque based structured pruning for deep neural network. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2711-2720, 2024. 8 +[15] Song Han, Jeff Pool, John Tran, and William Dally. Learning both weights and connections for efficient neural network. In Advances in neural information processing systems, pages 1135-1143, 2015. 1, 6, 7 +[16] Tianxiang Hao, Xiaohan Ding, Jungong Han, Yuchen Guo, and Guiguang Ding. Manipulating identical filter redundancy for efficient pruning on deep and complicated cnn. IEEE Transactions on Neural Networks and Learning Systems, 2023. 8 +[17] Peter Harris. The mali gpu: An abstract machine, part 2-tilebased rendering, 2014. 3 +[18] Yihui He, Xiangyu Zhang, and Jian Sun. Channel pruning for accelerating very deep neural networks. In Proceedings of the IEEE International Conference on Computer Vision, pages 1389-1397, 2017. 3 +[19] Yang He, Guoliang Kang, Xuanyi Dong, Yanwei Fu, and Yi Yang. Soft filter pruning for accelerating deep convolutional neural networks. arXiv preprint arXiv:1808.06866, 2018. 1 +[20] Yang He, Ping Liu, Ziwei Wang, Zhilan Hu, and Yi Yang. Filter pruning via geometric median for deep convolutional neural networks acceleration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4340-4349, 2019. 6, 7, 8 +[21] Parker Hill, Animesh Jain, Mason Hill, Babak Zamirai, Chang-Hong Hsu, Michael A Laurenzano, Scott Mahlke, Lingjia Tang, and Jason Mars. Deftnn: Addressing bottle-necks for dnn execution on gpus via synapse vector elimination and near-compute data fission. In Proceedings of the 50th Annual IEEE/ACM International Symposium on Microarchitecture, pages 786-799. ACM, 2017. 4 +[22] Di Huang, Xishan Zhang, Rui Zhang, Tian Zhi, Deyuan He, Jiaming Guo, Chang Liu, Qi Guo, Zidong Du, Shaoli Liu, et al. Dwm: A decomposable winograd method for convolution acceleration. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 4174-4181, 2020. 7 +[23] Gian Marco Iodice. Even Faster CNNs: Exploring the New Class of Winograd Algorithms, Embedded Vision Summit + +2018. https://www(edge-ai-vision.com/2018/06/even-faster\-\cnns-exploring-the-new-class-of\-\winograd-algorithms-a-pre-sentation-\from-arm/, 2018. Accessed: 2022-11-10.3 +[24] Liancheng Jia, Yun Liang, Xiuhong Li, Liqiang Lu, and Shengen Yan. Enabling efficient fast convolution algorithms on gpus via megakernels. IEEE Transactions on Computers, 69(7):986-997, 2020. 3 +[25] JD Andrew Kerr, Duane Merrill, and J Tran. Cutlass: Fast linear algebra in CUDA c++. NVIDIA Developer Blog, 2017. 4, 8 +[26] Heehoon Kim, Hyoungwook Nam, Wookeun Jung, and Jaejin Lee. Performance analysis of cnn frameworks for gpus. In 2017 IEEE International Symposium on Performance Analysis of Systems and Software (ISPASS), pages 55-64. IEEE, 2017. 1 +[27] John Kloosterman, Jonathan Beaumont, Mick Wollman, Ankit Sethia, Ron Dreslinski, Trevor Mudge, and Scott Mahlke. Warppool: Sharing requests with inter-warp coalescing for throughput processors. In 2015 48th Annual IEEE/ACM International Symposium on Microarchitecture (MICRO), pages 433-444. IEEE, 2015. 4 +[28] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 6 +[29] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. Advances in neural information processing systems, 25:1097-1105, 2012. 1 +[30] Andrew Lavin and Scott Gray. Fast algorithms for convolutional neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4013-4021, 2016. 1, 2 +[31] Bailin Li, Bowen Wu, Jiang Su, and Guangrun Wang. Eagleeye: Fast sub-net evaluation for efficient neural network pruning. In Computer Vision - ECCV 2020, pages 639-654, Cham, 2020. Springer International Publishing. 2, 6 +[32] Hao Li, Asim Kadav, Igor Durdanovic, Hanan Samet, and Hans Peter Graf. Pruning filters for efficient convnets. In International Conference on Learning Representations, 2017. 1, 3, 5, 7, 8 +[33] Sheng Li, Jongsoo Park, and Ping Tak Peter Tang. Enabling sparse winograd convolution by native pruning. arXiv preprint arXiv:1702.08597, 2017. 4 +[34] Xiuhong Li, Yun Liang, Shengen Yan, Liancheng Jia, and Yinghan Li. A coordinated tiling and batching framework for efficient gemm on gpus. In Proceedings of the 24th symposium on principles and practice of parallel programming, pages 229-241, 2019. 5 +[35] Lucas Liebenwein, Cenk Baykal, Harry Lang, Dan Feldman, and Daniela Rus. Provable filter pruning for efficient neural networks. In International Conference on Learning Representations, 2020. 7, 8 +[36] Mingbao Lin, Yuxin Zhang, Yuchao Li, Bohong Chen, Fei Chao, Mengdi Wang, Shen Li, Yonghong Tian, and Rongrong Ji. 1xn pattern for pruning convolutional neural networks. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 1 + +[37] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 6 +[38] Xingyu Liu, Song Han, Mao Huizi, and William J Dally. Efficient sparse-winograd convolutional neural networks. In International Conference on Learning Representations (ICLR) Workshop, 2017. 4 +[39] Xingyu Liu, Jeff Pool, Song Han, and William J Dally. Efficient sparse-winograd convolutional neural networks. In International Conference on Learning Representations, 2018, 1, 2, 4 +[40] Liqiang Lu and Yun Liang. Spwa: An efficient sparse winograd convolutional neural networks accelerator on fpgas. In Proceedings of the 55th Annual Design Automation Conference, pages 1-6, 2018. 4 +[41] Asit Mishra, Jorge Albericio Latorre, Jeff Pool, Darko Stosic, Dusan Stosic, Ganesh Venkatesh, Chong Yu, and Paulius Micikevicius. Accelerating sparse deep neural networks. arXiv preprint arXiv:2104.08378, 2021. 1 +[42] Pavlo Molchanov, Arun Mallya, Stephen Tyree, Iuri Frosio, and Jan Kautz. Importance estimation for neural network pruning. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11256-11264, 2019. 1 +[43] Sharan Narang, Eric Undersander, and Gregory Diamos. Block-sparse recurrent neural networks. arXiv preprint arXiv:1711.02782, 2017. 1 +[44] M Naumov, LS Chien, P Vandermersch, and U Kapasi. Cusparse library. In GPU Technology Conference, 2010. 4 +[45] Wei Niu, Xiaolong Ma, Sheng Lin, Shihao Wang, Xuehai Qian, Xue Lin, Yanzhi Wang, and Bin Ren. Patdnn: Achieving real-time dnn execution on mobile devices with pattern-based weight pruning. In Proceedings of the Twenty-Fifth International Conference on Architectural Support for Programming Languages and Operating Systems, pages 907-922, 2020. 1 +[46] Manuel Nonnenmacher, Thomas Pfeil, Ingo Steinwart, and David Reeb. SOSP: Efficiently capturing global correlations by second-order structured pruning. In International Conference on Learning Representations, 2022. 8 +[47] CUDA Nvidia. Cublas library. NVIDIA Corporation, Santa Clara, California, 15(27):31, 2008. 4 +[48] Cheonjun Park, Mincheol Park, Hyun Jae Oh, Minkyu Kim, Myung Kuk Yoon, Suhyun Kim, and Won Woo Ro. Balanced column-wise block pruning for maximizinggpu parallelism. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 9398-9407, 2023. 2, 7, 8 +[49] Mincheol Park, Dongjin Kim, Cheonjun Park, Yuna Park, Gyeong Eun Gong, Won Woo Ro, and Suhyun Kim. Reprune: Channel pruning via kernel representative selection. arXiv preprint arXiv:2402.17862, 2024. 1 +[50] Maying Shen, Hongxu Yin, Pavlo Molchanov, Lei Mao, Jianna Liu, and Jose M Alvarez. Structural pruning via latency-saliency knapsack. Advances in Neural Information Processing Systems, 35:12894-12908, 2022. 1 + +[51] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. In 3rd International Conference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Proceedings, 2015. 1 +[52] Mingjie Sun, Zhuang Liu, Anna Bair, and J Zico Kolter. A simple and effective pruning approach for large language models. In The Twelfth International Conference on Learning Representations, 2024. 6 +[53] Hidenori Tanaka, Daniel Kunin, Daniel L Yamins, and Surya Ganguli. Pruning neural networks without any data by iteratively conserving synaptic flow. Advances in Neural Information Processing Systems, 33:6377-6389, 2020. 3 +[54] Chien-Yao Wang, I-Hau Yeh, and Hong-Yuan Mark Liao. Yolov9: Learning what you want to learn using programmable gradient information. In European conference on computer vision, pages 1-21. Springer, 2024. 1 +[55] Haonan Wang, Wenjian Liu, Tianyi Xu, Jun Lin, and Zhongfeng Wang. A low-latency sparse-winograd accelerator for convolutional neural networks. In ICASSP 2019-2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1448-1452. IEEE, 2019. 4 +[56] Wei Wen, Chunpeng Wu, Yandan Wang, Yiran Chen, and Hai Li. Learning structured sparsity in deep neural networks. In Advances in neural information processing systems, pages 2074-2082, 2016. 4 +[57] Shmuel Winograd. Arithmetic complexity of computations. Siam, 1980. 1, 3 +[58] Di Wu, Xitian Fan, Wei Cao, and Lingli Wang. Svm: A high-performance sparse-winograd matrix multiplication cnn accelerator. IEEE Transactions on Very Large Scale Integration (VLSI) Systems, 29(5):936-949, 2021. 4 +[59] Xidong Wu, Shangqian Gao, Zeyu Zhang, Zhenzhen Li, Runxue Bao, Yanfu Zhang, Xiaoqian Wang, and Heng Huang. Auto-train-once: Controller network guided automatic network pruning from scratch. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16163-16173, 2024. 8 +[60] Xinghua Xue, Haitong Huang, Cheng Liu, Tao Luo, Lei Zhang, and Ying Wang. Winograd convolution: A perspective from fault tolerance. In Proceedings of the 59th ACM/IEEE Design Automation Conference, pages 853-858, 2022. 1 +[61] Da Yan, Wei Wang, and Xiaowen Chu. Optimizing batched winograd convolution on gpus. In Proceedings of the 25th ACM SIGPLAN symposium on principles and practice of parallel programming, pages 32-44, 2020. 1, 3 +[62] Tao Yang, Zhezhi He, Tengchuan Kou, Qingzheng Li, Qi Han, Haibao Yu, Fangxin Liu, Yun Liang, and Li Jiang. Biswsrbs: A winograd-based cnn accelerator with a fine-grained regular sparsity pattern and mixed precision quantization. ACM Transactions on Reconfigurable Technology and Systems (TRETS), 14(4):1-28, 2021. 4 +[63] Zhuliang Yao, Shijie Cao, Wencong Xiao, Chen Zhang, and Lanshun Nie. Balanced sparsity for efficient dnn inference ongpu. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 5676-5683, 2019. 1 + +[64] Jiecao Yu, Andrew Lukefahr, David Palframan, Ganesh Dasika, Reetuparna Das, and Scott Mahlke. Scalpel: Customizing dnn pruning to the underlying hardware parallelism. In 2017 ACM/IEEE 44th Annual International Symposium on Computer Architecture (ISCA), pages 548-560. IEEE, 2017. 1, 2, 4 +[65] Jiecao Yu, Jongsoo Park, and Maxim Naumov. Spatialwinograd pruning enabling sparse winograd convolution. arXiv preprint arXiv:1901.02132, 2019. 3 +[66] Ruichi Yu, Ang Li, Chun-Fu Chen, Jui-Hsin Lai, Vlad I Morariu, Xintong Han, Mingfei Gao, Ching-Yung Lin, and Larry S Davis. Nisp: Pruning networks using neuron importance score propagation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 9194-9203, 2018. 3 +[67] Xuanyang Zhang, Hao Liu, Zhanxing Zhu, and Zenglin Xu. Learning to search efficient densenet with layer-wise pruning. In 2020 International Joint Conference on Neural Networks (IJCNN), pages 1-8, 2020. 8 +[68] Maohua Zhu, Tao Zhang, Zhenyu Gu, and Yuan Xie. Sparse tensor core: Algorithm and hardware co-design for vectorwise sparse neural networks on modern gpus. In Proceedings of the 52nd Annual IEEE/ACM International Symposium on Microarchitecture, pages 359-371. ACM, 2019. 1 \ No newline at end of file diff --git a/ICCV/2025/WINS_ Winograd Structured Pruning for Fast Winograd Convolution/images.zip b/ICCV/2025/WINS_ Winograd Structured Pruning for Fast Winograd Convolution/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..1df2b6ff83f2be1f66cd3c4167b72fbc366e72fe --- /dev/null +++ b/ICCV/2025/WINS_ Winograd Structured Pruning for Fast Winograd Convolution/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f1f28520ac9bfd49b54e0686274baf2c1935f309b863a6e8109a6f7a798ff45a +size 863582 diff --git a/ICCV/2025/WINS_ Winograd Structured Pruning for Fast Winograd Convolution/layout.json b/ICCV/2025/WINS_ Winograd Structured Pruning for Fast Winograd Convolution/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..46758eccb85d36e515d291cda580238b65939741 --- /dev/null +++ b/ICCV/2025/WINS_ Winograd Structured Pruning for Fast Winograd Convolution/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:df1c46c1051d80b02cf0cbd3c7d260864e66b1a262f8b7031e413eb670922b3f +size 431908 diff --git a/ICCV/2025/WIPES_ Wavelet-based Visual Primitives/5c2a3453-b8e9-4c45-a9c7-d765229caa07_content_list.json b/ICCV/2025/WIPES_ Wavelet-based Visual Primitives/5c2a3453-b8e9-4c45-a9c7-d765229caa07_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..0ed5ff07a526dc7a8e1633a88df382f19d4178eb --- /dev/null +++ b/ICCV/2025/WIPES_ Wavelet-based Visual Primitives/5c2a3453-b8e9-4c45-a9c7-d765229caa07_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f6558e09e9ffb2a75ee7927403c22ab3ec71d102a89366d70c2e9685d18ac0c7 +size 93516 diff --git a/ICCV/2025/WIPES_ Wavelet-based Visual Primitives/5c2a3453-b8e9-4c45-a9c7-d765229caa07_model.json b/ICCV/2025/WIPES_ Wavelet-based Visual Primitives/5c2a3453-b8e9-4c45-a9c7-d765229caa07_model.json new file mode 100644 index 0000000000000000000000000000000000000000..cb7fbae4968357b8880d8ed1c1cbb7965a52491c --- /dev/null +++ b/ICCV/2025/WIPES_ Wavelet-based Visual Primitives/5c2a3453-b8e9-4c45-a9c7-d765229caa07_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:459c4f4f10a12a1149f86e38d031526f2e7d3ae87a4d5211d8d99a2b3e4eea8e +size 111044 diff --git a/ICCV/2025/WIPES_ Wavelet-based Visual Primitives/5c2a3453-b8e9-4c45-a9c7-d765229caa07_origin.pdf b/ICCV/2025/WIPES_ Wavelet-based Visual Primitives/5c2a3453-b8e9-4c45-a9c7-d765229caa07_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..c86ab19de574eb176c70b192e9738054ccba98ca --- /dev/null +++ b/ICCV/2025/WIPES_ Wavelet-based Visual Primitives/5c2a3453-b8e9-4c45-a9c7-d765229caa07_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a307e3eeedf71875b8d5522b6e5d07fc85b5639c0ce2cc03fa9f325f7ad5d45c +size 4761733 diff --git a/ICCV/2025/WIPES_ Wavelet-based Visual Primitives/full.md b/ICCV/2025/WIPES_ Wavelet-based Visual Primitives/full.md new file mode 100644 index 0000000000000000000000000000000000000000..2c17d5c4fdc6bffca13cc2e9bb081d1c481a60e0 --- /dev/null +++ b/ICCV/2025/WIPES_ Wavelet-based Visual Primitives/full.md @@ -0,0 +1,452 @@ +# WIPES: Wavelet-based Visual Primitives* + +Wenhao Zhang $^{1,\dagger}$ , Hao Zhu $^{1,\dagger}$ , Delong Wu $^{1,\dagger}$ , Di Kang $^{2}$ , Linchao Bao $^{2}$ , Xun Cao $^{1,*}$ , Zhan Ma $^{1}$ $^{1}$ Nanjing University, $^{2}$ Tencent + +$^{\dagger}$ Equal contribution. * Corresponding author: caoxun@nju.edu.cn + +# Abstract + +Pursuing a continuous visual representation that offers flexible frequency modulation and fast rendering speed has recently garnered increasing attention in the fields of 3D vision and graphics. However, existing representations often rely on frequency guidance or complex neural network decoding, leading to spectrum loss or slow rendering. To address these limitations, we propose WIPES, a universal Wavelet-based Visual PrimitivES for representing multi-dimensional visual signals. Building on the spatial-frequency localization advantages of wavelets, WIPES effectively captures both the low-frequency "forest" and the high-frequency "trees." Additionally, we develop a wavelet-based differentiable rasterizer to achieve fast visual rendering. Experimental results on various visual tasks, including 2D image representation, 5D static and 6D dynamic novel view synthesis, demonstrate that WIPES, as a visual primitive, offers higher rendering quality and faster inference than INR-based methods, and outperforms Gaussian-based representations in rendering quality. + +# 1. Introduction + +Representing visual signals as continuous functions has garnered increasing attention in recent years. Owing to the inherent advantages of end-to-end differentiability and scalable querying, these continuous representations can be seamlessly integrated with various imaging and rendering functions. This integration facilitates the optimization of inverse problems, whether the resolution is down-sampled or the dimensionality is reduced. Since these representations are optimized independently for different scenarios, pursuing a universal visual primitive, which offers complex frequency compatibility and fast rendering speed, is of paramount importance. + +However, existing representations suffer from spectral loss or slow rendering. For instance, the representational capacity of popular implicit neural representations (INRs) + +![](images/ead90e64c4f421b3af1e93ff1e7a80d3acd084d5d0077f4160b289825f8219f8.jpg) +(a) Comparison of Gaussians and WIPES + +![](images/f4beef75df43fc2a34e2eaae393cc593152e069a0a1e8f8a59f9641d9a738294.jpg) + +![](images/ae53cdc5112e2e51a441b0a22ab0a9a0d3c2e9f6c7e6e7b6a8c4bce524d164b0.jpg) +(b) PSNR improvement of WIPES over Gaussian primitives [44] +High-complexity Image +(c) Performance comparison of WIPES and Gaussian primitives [44] +Figure 1. Spatial-Frequency Localization Benefits of WIPES. (a) shows the spatial and frequency distributions of Gaussian and WIPES primitives with various parameters. Adjusting $\sigma$ only affects the Gaussian frequency bandwidth, unable to overcome its low-frequency constraint. In contrast, WIPES flexibly modulates frequency, making it advantageous for complex visual tasks. (b) displays image fitting on the Kodak dataset, where WIPES achieves higher PSNR with the same primitives count, and this advantage grows with more primitives. (c) illustrates that Gaussian primitives smooth complex textures due to low-frequency limitation, while WIPES accurately reconstructs rooftop patterns and railings through adaptable frequency properties. + +![](images/cfe9ae1fcc60c3994c05033e59341badd48c058b168ddb8f7974e33121618d80.jpg) + +![](images/f14782de79c2a86ddce6104216f2c1e58a68643ae4ddbbcfe2e78d7c2076af4d.jpg) +GT +WIPES +GSImage + +![](images/a3b6996e07f331d6f3a5f949e52145399b9d052784b7d5273003fdc67318b2bb.jpg) + +![](images/bb6fddccf1a00b822c68e311e5999537ea1e6157e64d0aca1807f1428f59e43a.jpg) + +[34] is often constrained by spectral bias [27], only the frequencies defined by hyperparameters are effectively learned [35, 41, 46], resulting in the loss of information in other frequency bands. Furthermore, the complex network architec + +tures of INRs require substantial computational resources [24] to decode signal attributes, thereby impeding processing speed across various visual tasks. Explicit representations, among which the 3D Gaussian Splatting (3DGS) [14] specifically models the visual world by employing multiple Gaussian functions. However, the inherent low-pass nature of Gaussian functions introduces inaccuracies in capturing high-frequency details. To mitigate these low-pass characteristics, frequency regularization [43] or modulation losses [11] are often employed, which necessitates extensive expertise to tune hyperparameters for different scenes. + +These problems arise from a mismatch between the characteristics of the visual primitives used and the visual signals themselves. According to Marr's vision theory, visual signals are perceived hierarchically, and intensity changes are detected at multiple scales. Wavelet, designed for multiscale signal analysis, aligns well with the human visual system and can capture both the "forest" and the "trees" in visual signals [10]. Consequently, we developed WIPES, a universal primitive for representing visual signals. + +WIPES leverages the Morlet wavelet—known for its continuous differentiability and robust spatial-frequency localization—to effectively capture signals exhibiting significant frequency variations. We extend the Morlet wavelet by developing a wavelet primitives adaptable to arbitrary dimensions, making it well-suited for various visual signals in real-world applications (e.g., 2D images, 5D static and 6D dynamic radiance fields). Additionally, we develop a fast, differentiable rasterizer for WIPES by utilizing an efficient GPU sorting algorithm [18]. To mitigate the substantial computational overhead associated with automatic differentiation during training, we derive explicit gradients for all parameters in both forward and backward propagations. Our rasterizer effectively replaces Gaussian-based rasterizers, enabling seamless integration with different loss functions [5, 43] and subsequent visual tasks [38, 42, 44] built upon Gaussian splatting techniques. + +In summary, we make the following contributions: + +- We design a universal differentiable wavelet primitive that adaptively models various frequency features in real-world scenes without relying on handcrafted frequency guidance. +- We develop a fast differentiable rasterizer for our WIPES, enabling anisotropic splatting and efficient backpropagation across different dimensions to achieve general and effective visual representation. +- The proposed wavelet-based approach benefits superior reconstruction performance across multiple task dimensions compared to other visual representations. + +# 2. Related Work + +Implicit Neural Representations (INRs) play a pivotal role in modeling continuous functions of signals found in + +images, scenes and videos [19]. They leverage MLPs to accurately map spatial coordinates to their corresponding values [24, 34]. This capability enables INRs to be applied in various domains, including novel view synthesis [24], freehand 3D/4D echocardiography [32, 33], microscopy imaging [45], and nano-optics design [4]. + +However, INRs are constrained by spectral bias, favoring low-frequency information over high-frequency details [27]. To mitigate this bias, strategies like positional encoding [35] and specifically activation functions [20, 28, 31, 34] have been developed. Despite these advancements, the spectral bias remains fundamentally unaddressed and is merely transformed into a "frequency"-specified spectral bias [46]. Additionally, INRs suffer from slow training and inference due to their reliance on computationally intensive MLPs. To enhance efficiency, recent approaches have integrated explicit storage with implicit decoding, including sparse voxel grids [9], hash tables [25, 47, 48], and multiscale/grid INRs [29, 30, 49]. While these approaches offer improvements, they introduce additional complexity, making it challenging to balance memory consumption with real-time performance. Furthermore, INRs are highly sensitive to parameter changes. Modifications to the input or internal parameters can lead to significant variations in the output, complicating the process of making local adjustments [23]. Although strategies such as explicit encoding and latent disentanglement [17, 36] have been proposed to address these issues, achieving precise control over local features in INR-based approaches remains challenging. + +Explicit Signal Representations with Gaussian Primitives, conversely, overcome several INR limitations and are growing in popularity in computer vision and graphics. They use multiple independent continuous Gaussian functions to render scenes, which offers marked benefits like enhanced expressiveness and editability, and superior fitting efficiency than conventional INRs. + +These Gaussian-based approaches have achieved top-tier performance in rendering quality and speed [14]. Early research on splatting introduced Gaussian primitives for scene representation [50, 51]. Recent advancements such as FreGS [43], mip-splatting [40], and scaffold-GS [21] enhance frequency management, mitigate aliasing, and incorporate voxel-based initialization, thereby demonstrating the robustness and efficiency of Gaussian-based approaches. + +However, Gaussian primitives are constrained by their inherent low-pass filtering characteristics, which are effective for smooth signals but hinder the capture of high-frequency details like sharp edges or intricate textures. Real-world scenes frequently contain abrupt changes, such as the complex textures of grass or nylon fabric. Accurately representing these details requires multiple small-scale Gaussians [11, 13], which significantly increase model complexity, thereby reducing scalability and efficiency. + +![](images/e04afdce3feaf9a25b4b075d9fdd5f95bfc23b7e88f45654d7e18734b29abc4e.jpg) +Figure 2. Pipeline of WIPEs. WIPEs can be seamlessly integrated into various visual reconstruction tasks, including 2D image representation, 5D static and 6D dynamic novel view synthesis (NVS). + +Wavelet-based Signal Representations have been proven to be highly effective in analyzing complex signals within vision tasks. Traditional Fourier approaches use sine functions for their smoothness and clear characteristics in the Fourier domain. However, they decompose signals into global sine bases, which lack localization in both space and frequency [22]. This limitation makes Fourier approaches less suitable for visual data with local features and abrupt changes. In contrast, wavelets provide optimal localization in both domains, effectively capturing local high-frequency variations and sudden signal changes while maintaining a compact representation. Recent studies have leveraged wavelets to enhance vision models. MFN [8] employs wavelets as nonlinear filters to approximate complex functions with fewer parameters, improving frequency information representation. TriNeRFLet [15] integrates wavelets into Triplane's multi-scale representations, enhancing cross-scale information sharing and high-frequency regularization for better 3D reconstructions. WavePlanes [2] uses wavelet compression on the spatiotemporal feature planes of dynamic NeRFs, reducing model size and enhancing high-frequency details. + +However, recent wavelet-inspired representations are implicitly integrated into neural networks, which significantly reduce the rendering speed. On the contrary, the proposed wavelet-based primitives operate on the visual signal explicitly and directly, thus improving the efficiency while maintaining the advantages of spatial-frequency location. + +# 3. Preliminaries + +# 3.1. 3D Gaussian Splitting + +3DGS [14] explicitly represents 3D scenes using Gaussian primitives, enabling efficient scene reconstruction. Specifically, each Gaussian primitive is parameterized by a spatial mean vector $\vec{\mu}$ , a covariance matrix $\boldsymbol{\Sigma}$ (expressed as $\boldsymbol{\Sigma} = \mathbf{RSS}^{\top}\mathbf{R}^{\top}$ , where $\mathbf{R}$ denotes rotation and $\mathbf{S}$ denotes scaling), an opacity value $\alpha$ , and color attributes encoded by spherical harmonic coefficients $c$ . In the world space, + +the Gaussian primitive centered at $\vec{\mu}$ with $\Sigma$ is defined as: + +$$ +\mathcal {G} (\vec {x}, \vec {\mu}, \boldsymbol {\Sigma}) = e ^ {- \frac {1}{2} (\vec {x} - \vec {\mu}) ^ {\top} \boldsymbol {\Sigma} ^ {- 1} (\vec {x} - \vec {\mu})}, \tag {1} +$$ + +where $\vec{x}$ and $\vec{\mu}$ are column vectors in the world space, denoted as $(x,y,z)^{\top}$ and $(\mu_x,\mu_y,\mu_z)^{\top}$ , respectively, the covariance matrix $\pmb{\Sigma}$ is a $3\times 3$ symmetric matrix. + +When employing the splatting method for rendering, the 3D reconstruction kernel must be integrated along the $z$ -axis to evaluate its contribution to each pixel on the image plane. Since integrating a 3D Gaussian function along a given axis still results in a Gaussian function, we can effectively transform a 3D Gaussian $\mathcal{G}(\vec{x})$ into a 2D Gaussian $\mathcal{G}'(\vec{x}')$ on the image plane using the ray-coordinate transformation described in [50]: + +$$ +\int_ {\mathbb {R}} \mathcal {G} (\vec {x}, \vec {\mu}, \boldsymbol {\Sigma}) d z = \mathcal {G} ^ {\prime} \left(\vec {x} ^ {\prime}, \vec {\mu} ^ {\prime}, \boldsymbol {\Sigma} ^ {\prime}\right), \tag {2} +$$ + +where $\vec{x}^{\prime} = (x,y)^{\top}$ and $\vec{\mu}^{\prime} = (\mu_{x},\mu_{y})^{\top}$ represent coordinates in the projected 2D image plane. To achieve accurate perspective projection, the original covariance matrix $\Sigma$ undergoes transformation through the world-to-camera transformation matrix $\mathbf{W}$ , followed by projection via the local affine Jacobian matrix $\mathbf{J}$ . This yields a ray-space covariance matrix $\hat{\Sigma} = \mathbf{J}\mathbf{W}\Sigma \mathbf{W}^{\top}\mathbf{J}^{\top}$ . Ultimately, the 2D covariance matrix $\Sigma^{\prime}$ in image space can be easily obtained by taking the upper left $2\times 2$ submatrix of $\hat{\Sigma}$ . + +Finally, the pixel values on the 2D image are obtained by $\alpha$ -blending: + +$$ +C \left(\vec {x} ^ {\prime}\right) = \sum_ {i \in N} c _ {i} \alpha_ {i} \mathcal {G} _ {i} ^ {\prime} \left(\vec {x} ^ {\prime}, \vec {\mu} ^ {\prime}, \boldsymbol {\Sigma} ^ {\prime}\right) \prod_ {j = 1} ^ {i - 1} \left(1 - \alpha_ {j} \mathcal {G} _ {j} ^ {\prime} \left(\vec {x} ^ {\prime}, \vec {\mu} ^ {\prime}, \boldsymbol {\Sigma} ^ {\prime}\right)\right), \tag {3} +$$ + +where $\vec{x}^{\prime}$ denotes the queried pixel position and $N$ represents the set of sorted 2D Gaussians associated with $\vec{x}^{\prime}$ . + +# 3.2. Expansion of 3D Gaussian Splatting + +Gaussian-based representations have demonstrated exceptional performance not only in 3D scene signal modeling + +Table 1. Quantitative comparison of image fitting on Kodak and DIV2K datasets. We color code each cell as best and second best . + +
Methods | MetricsKodakDIV2K
PSNR↑SSIM↑LPIPS↓FPS↑PSNR↑SSIM↑LPIPS↓FPS↑
WIRE [31]38.020.95410.047719.5533.200.95650.130311.63
GSImage-RS [44]43.090.99830.02162129.1839.850.99780.03712190.40
GSImage-Cholesky44.060.99850.01882061.9939.530.99750.04302248.74
WIPEs-RS45.760.99880.01321708.2840.340.99760.03601817.48
WIPEs-Cholesky45.870.99870.01201778.7540.320.99780.03611830.10
GTWIPEGSImageWIRE
+ +Figure 3. Qualitative comparison on Kodak and DIV2K datasets. Our approach effectively captures both low- and high-frequency details. + +but also in various other domains. For instance, Gaussian Image [44] explicitly parameterizes 2D image signals using 2D Gaussian primitives defined directly on the image plane, replacing traditional $\alpha$ -blending with a simplified weighted summation mechanism: + +$$ +C \left(\vec {x} ^ {\prime}\right) = \sum_ {i \in N} c _ {i} \alpha_ {i} \mathcal {G} _ {i} ^ {\prime} \left(\vec {x} ^ {\prime}, \vec {\mu} ^ {\prime}, \boldsymbol {\Sigma} ^ {\prime}\right). \tag {4} +$$ + +Moreover, to handle dynamic 3D content, D-3DGS [38] employs an MLP network $\mathcal{F}_{\theta}$ to model the deformation field of 3D Gaussian primitives. Thus, the primitives efficiently adapt to dynamic variations over time, yielding a high-fidelity representation of temporally-varying 3D scenes: + +$$ +\left(\delta \vec {x}, \delta \boldsymbol {\Sigma}\right) = \mathcal {F} _ {\theta} (\gamma (\vec {x}), \gamma (t)), \tag {5} +$$ + +where $\gamma$ denotes positional encoding [24]. + +# 4. Wavelet-based Visual Primitives + +# 4.1. 3D Splitting with Wavelet Primitives + +The wavelet transform decomposes a signal into a linear combination of translated and scaled short oscillatory pulses, effectively capturing localized variations in both spatial and frequency domains. Due to their compact support in these domains, wavelets achieve superior accuracy in signal approximation [6]. This property substantiates the feasibility of using differentiable wavelet basis functions in our methodology. Inspired by the Morlet wavelet, we obtain + +the 3D wavelet primitive by applying cosine modulation to the 3D Gaussian primitive. A 3D wavelet primitive is described by: + +$$ +\mathcal {W} (\vec {x}, \vec {\mu}, \vec {f}, \boldsymbol {\Sigma}) = \frac {1}{2} [ \cos (\vec {f} \cdot (\vec {x} - \vec {\mu})) + 1 ] \mathcal {G} (\vec {x}, \vec {\mu}, \boldsymbol {\Sigma}), (6) +$$ + +where $\vec{f} = (f_x, f_y, f_z)^\top$ is the frequency of cosine modulation. Notably, the Gaussian primitive can be considered a subset of the proposed 3D wavelet primitive by setting the parameter $\vec{f}$ to $\vec{0}$ . + +Integrating the 3D wavelet primitive along the $z$ -axis produces results similar to those described in Eq. (2): + +$$ +\int_ {\mathbb {R}} \mathcal {W} (\vec {x}, \vec {\mu}, \vec {f}, \boldsymbol {\Sigma}) d z = \mathcal {W} ^ {\prime} (\vec {x} ^ {\prime}, \vec {\mu} ^ {\prime}, \vec {f} ^ {\prime}, \boldsymbol {\Sigma} ^ {\prime}), \tag {7} +$$ + +where $\vec{f}^{\prime} = (f_{x},f_{y})^{\top}$ and $\vec{x}^{\prime} - \vec{\mu}^{\prime}$ is the same as the 3D Gaussian primitive in Eq. (2). + +The projected 2D covariance matrix $\Sigma^{\prime}$ for each wavelet primitive is computed via a similar spatial transformation used in projecting Gaussian primitives from 3D world space to the 2D image plane. Specifically, the covariance matrix $\Sigma$ and modulation frequency vector $\vec{f}$ of the wavelet primitive undergo the identical ray-space transformation defined by matrices $\mathbf{W}$ and $\mathbf{J}$ , resulting in the transformed frequency vector $\vec{f}^{\prime}$ and the projected covariance matrix $\Sigma^{\prime}$ . + +For rasterization, we follow a similar pipeline as in 3D GS while substituting the primitives with our proposed 3D wavelets. In the end, the volumetric $\alpha$ -blending can be + +![](images/8b89bc613dbd278d25008626168fcaa4dd9e64283362ace2c947fc7e2d541388.jpg) + +![](images/2fce84c29bddba9fb498bfde571df7cd225fd0fa01a6d8fc31bc6f04e043a586.jpg) + +![](images/4070408d69c155918707f6fc4531efbffee47c40876bcc38975b1a3cdbb32d1c.jpg) + +![](images/62f819fc1427e1df47a3d2057221d1ace78ffbaf9225acb316a49fc58ad09b6f.jpg) +GT + +![](images/faa06940224bc18ab45eafdbfaa3bb8379f1e59b5db594ffac10470ba189af24.jpg) +WIPES + +![](images/ed2391211dd28ba6e834408e70e8435b0c0756a56161c12b8e677ca249b5c563.jpg) +Gaussian +Figure 4. Comparisons of Gaussian and wavelet primitives for representing complex texture patterns. Results indicate that wavelet primitives effectively overcome the low-frequency bias of Gaussian primitives and demonstrate excellent performance in high-frequency texture regions. + +easily obtained by replacing the Gaussian primitive $\mathcal{G}'$ in Eq. (3) with the wavelet primitive $\mathcal{W}'$ . + +# 4.2. Wavelet-Based Frequency Flexibility + +Gaussian functions inherently act as low-pass filters, as illustrated in Fig. 1a. Specifically, in the frequency domain, Gaussians with large variance exhibit narrow bandwidths concentrated around zero frequency, while those with small variance produce wider bandwidths. This bell-shaped Gaussian spectrum rapidly attenuates high-frequency signals, thus limiting its capacity to represent signals rich in high-frequency content. Consequently, a large number of Gaussian basis functions with smaller variances are inevitably required to accurately capture intricate high-frequency details. + +In contrast, wavelets provide significantly enhanced frequency flexibility. By employing adaptive frequency modulation, wavelet primitives simultaneously encode both low- and high-frequency components within a single basis function. This modulation introduces a controllable shift in the wavelet spectrum, positioning the Gaussian-like envelope around a desired center frequency. The modulation frequency defines this spectral center, and the Gaussian envelope controls the bandwidth, ensuring optimal coverage of the spatial-frequency domain. + +We validate this advantage by comparing Gaussian and wavelet primitives on two images exhibiting complex moiré patterns (Fig. 4). Under identical conditions and equal numbers of primitives, Gaussian primitives consistently failed to represent detailed high-frequency structures, resulting in noticeable modeling voids. In contrast, the wavelet-based representation accurately captured intricate frequency transitions, effectively modeling sharp discontinuities and fine details. This result demonstrates that wavelets, through their inherent frequency adaptability, offer superior effi + +ciency and accuracy in representing signals characterized by rapid spatial variations. + +# 4.3. Expansion of 3D Wavelet Splatting + +We further leverage wavelet primitives to enhance representation quality, especially for signals exhibiting complex spatial-frequency content. Specifically, we validate the effectiveness of wavelet primitives in two key applications: image fitting and dynamic novel view synthesis. + +For the image fitting task, we replace the original 2D Gaussian primitives $\mathcal{G}'$ defined in Eq. (4) with frequency-adaptive wavelet primitives $\mathcal{W}'$ . This substitution introduces modulation frequencies $\vec{f}'$ , significantly expanding frequency coverage compared to Gaussian-based models. In dynamic novel view synthesis, we further generalize wavelet primitives by making them temporally adaptive. The modified formulation is given by: + +$$ +\left(\delta \vec {x}, \delta \vec {f}, \delta \boldsymbol {\Sigma}\right) = \mathcal {F} _ {\theta} \left(\gamma (\vec {x}), \gamma (t)\right). \tag {8} +$$ + +Comprehensive experimental validation demonstrating the superiority of our wavelet-based approach is provided in Sec. 5, confirming their ability to efficiently represent high-frequency details and complex temporal variations. + +# 5. Experiments + +Our experiments demonstrate the robustness and versatility of WIPES in representing real-world scenes, effectively capturing both high-frequency details (e.g., edge variations and intricate textures) and low-frequency contours. To validate WIPES as a generalized signal representation framework, we systematically evaluate its performance across three distinct tasks: 2D image fitting, 5D static novel view synthesis, and 6D dynamic novel view synthesis. + +All experiments are consistently executed on a single NVIDIA A6000 GPU. To ensure a fair comparison, we adopt the same initialization schemes, optimization strategies, and hyperparameter configurations as those of the baseline approaches. Specifically, frequency modulation coefficients are randomly initialized in alignment with competing methods to exclude initialization bias. Additionally, our implementation builds upon the recent gsplat library [39], augmented by custom CUDA kernels that enable wavelet-based differentiable rasterization blending. This adaptation is critical for efficiently leveraging wavelet primitives in rendering tasks. + +# 5.1.2D Image Fitting + +Configurations. In this section, we utilized the Kodak [7] and DIV2K [1] datasets. The DIV2K dataset was down-scaled by a factor of 2 using bicubic interpolation, resulting in images with resolutions ranging from $408 \times 1020$ to $1020 \times 1020$ . For our WIPES, we conducted experiments + +Table 2. Quantitative comparison of static novel view synthesis on Mip-NeRF360, Tanks & Temples and DeepBlending datasets. Results marked with † were adopted from the 3DGS paper [14], all others come from our experiments. + +
Methods | MetricsMip-NeRF360Tanks & TemplesDeepBlending
PSNR↑SSIM↑LPIPS↓FPS↑PSNR↑SSIM↑LPIPS↓FPS↑PSNR↑SSIM↑LPIPS↓FPS↑
Plenoxels [9]23.320.6460.4516.7921.080.7190.3791323.060.7950.51011.2
INGP [25]25.190.6670.44011.721.720.7230.33017.123.620.7970.4233.26
Mip-NeRF360†[3]27.690.7920.2370.0622.220.7590.2570.1429.400.9510.2450.09
3DGS [14]27.280.8110.220118.223.350.8350.182162.429.430.8980.246124.1
WIPES27.550.8160.21595.723.780.8520.171126.329.820.9070.23892.6
+ +![](images/34c33fc623b55d591fe508bfac0b277333b66d18be4282376fca60f504f78442.jpg) +GT + +![](images/c94b9c361088bcdc68db42279ba3df37565a521623d958644f147365b6fc3fed.jpg) +WIPES + +![](images/028c32d385b5ffba0f58e8569a8cdb7fa123ba5833c37faafcc33343bf5723bf.jpg) +3DGS + +![](images/bba8450140d9366250747d3109d6d8d823b8785b698340a2c0e79823e1c97590.jpg) +Mip-NeRF360 + +![](images/44b4894f6bf07a8f2c17df93b6642b6112d26db13a4ef0e3660e2a05e1bddac5.jpg) +Plenoxels + +![](images/e38e817bb094438c38374b19c27605516f73972f5c00e4292ad209a82f674dcd.jpg) +INGP + +![](images/9b1d43e7bd52f2850a6db9aa371b10f926affd34690d7d8b8bd3ab6b8a309ebf.jpg) + +![](images/771d417b542f5cf3c728ba76740c74f2d6400f7f797a85e1cab51b962458fdd7.jpg) + +![](images/a7566be6386dbe2f21cfa200d376958e29fc042617c6bb9b0d594241dceda44e.jpg) + +![](images/f8eb97f5bb856814176d2da7f5bcc5f6c0b3ec6b332a6b6fbb0143f8bca4f26e.jpg) + +![](images/8644a6393f7bfaa5aa85739c43b35d511f190f5f7bd5d55f796daa20910f9118.jpg) + +![](images/c963088948c35de1088e6f3b21045f32c6ffcfbec8c043fd4a93669c152fb2ef.jpg) + +![](images/5881f057bd03d17de48cfec2a6dca6cc5aad833e17da1e3e18036da4ecd9f3dc.jpg) + +![](images/cd866a5f417cc7c52d11ce51b1460c5ed615329dc694a699721da057e175982a.jpg) + +![](images/38c477091fa130e5be122bff8f92c9763a06700ec29abf1d0cf0366be92be1de.jpg) + +![](images/f186fae2140109ab943933366d511bd066de376da7c4eeb767e92cc12ec97d81.jpg) + +![](images/19ec45fcdb78b7a9d32e406b2aa65d440add78592beefc51251189059109cd76.jpg) + +![](images/7d396d0951471d9f17390623c3917ace964389dbd7fb3bd4cc64f05595a63b28.jpg) + +![](images/bfed0323455c049b1428ac92594236c13768fe129a1749126d0f518907464c6c.jpg) + +![](images/d0aa77473ee33a6b3ca0efcede3bb7df7e8603081b4e875bc4d76fffbfa318f0.jpg) + +![](images/fa128d699033e180c7ded3b2cd24c360926b2e2cda8d22e8d81e66594fe844a4.jpg) + +![](images/e49102b002df34af43a6d5fa1c817d7c18ea08a4b6c667d1857c83c4aebe00fd.jpg) + +![](images/40d44236eec531be0f1e25ca1246b285d1e13bd18b5b1634372926ccc3022f8d.jpg) + +![](images/fcda2d0f943ba031b409bdcb80cd850a9de0218f795fa90726fd36e481f61706.jpg) + +![](images/c55d1a0763280128f76adc623a3664cb8a3a947703d105a6c72ab6f7fea4d2df.jpg) + +![](images/18e09eca91b06302f4e122f70323e2e315619c24c5268187e8302a101ac0cc89.jpg) + +![](images/0dad8b9e590909ed72763c52831726b025884d1ea098a680c0c13a55cec6c428.jpg) + +![](images/cf2e6e044419f3b7b85238eaf3f2c8d7c3ae9006da29de71f2151fed7af7f04e.jpg) + +![](images/39c4993969b819beb2e93ccc3039197c8e8301641a0e7740a6118c06423eb652.jpg) + +![](images/20a3754e1b74e8d4c2d85813c892721b4d326bcee7e8c60e996b343f5508dbf1.jpg) + +![](images/c5a787fac0727fbc9e17a030dfba00262064991ce74c1908036d3134c1b71d26.jpg) + +![](images/28269968711bdf5b883eb6eb665da6e626ea67af4a43070c3a3d2542fc11a8f2.jpg) + +![](images/a2e8359b0ff631fe34f802791a74e8d0b420be291c54886245d99a69a143cba5.jpg) + +![](images/78cbf12bcd074606eed5adbc8c4d3137296f9c5af909d2e5d9d6a6f8b89cf2a3.jpg) + +![](images/8a1417b7428a798f8cbe67d4c94122c49808f48a272d310c24dc7f6dad3a0f10.jpg) + +![](images/9a855f27744065467606beb7683f639a5032ee48126e1cb45e514047bdd15f1c.jpg) + +![](images/0d4f278f69ba800d4fcb4647ae933df2b8fd3980bd97da63450fa0b8f867d6d7.jpg) + +![](images/00c432692f5bc8a70070fc9eee16cd5f5893dca72b58c4fe19bfc26632940c26.jpg) + +![](images/9ec7da8732878da78fcd4543a95cfe3e9cdc1b6a353225dbfce14e6798086983.jpg) + +![](images/7be284f37b6027cc55f119b42ca1bb85ea98abd907bccf723b1484674148987a.jpg) + +![](images/6f7fb65ec6c483a7a4970f4a9e597698d6ddec2137811d1bb4de0a9012858e5b.jpg) + +![](images/2ddc8cc61ccfd160c8ef0a0242ac6344e3d94cbb19db1c2dc5c8dfa686417876.jpg) + +![](images/18b468aba07c4967eeb19ad063372c778e888540d3c8e7c83583d9092e153506.jpg) +Detail Error map + +![](images/d3231e8ca9615e07e3c8c61e736a3c9425717ca3f7699d2dfe75c29cca587bba.jpg) + +![](images/a42d116726386c0b86880248bae54879916d75aaab420c298b8616db4491a7f3.jpg) + +![](images/24039e57bbd9eb0038c8087799337a61b8706592505bc001fe84b454bfb4c782.jpg) + +![](images/a6ca57667fc5bed746645cddf8db3263c4185b797d9bf98372d06b8c63a65bea.jpg) + +![](images/07f3e56cca11a1f9ee3285e888c40049786bae8d1c41afb22628018a2e46c37a.jpg) + +![](images/6c192bcc500c7406695e26d727697f4deb8ac1fbfac2e7c92bd016ea8666f6e3.jpg) +Figure 5. Qualitative comparison of static novel view synthesis on Mip-NeRF360, Tanks & Temples and DeepBlending datasets. + +using the two covariance construction methods proposed in GSImage, namely Cholesky and RS. The frequency coefficients were randomly initialized with a normal distribution, while all other learnable parameters followed the initialization strategy described in GSImage [44]. To ensure a fair comparison with WIRE [31] and GSImage [44], all hyperparameters were kept consistent with those used in GSImage. + +Results. Tab. 1 presents a quantitative comparison between + +our proposed WIPES and existing advanced INR-based and Gaussian-based approaches. Our approach achieves the best performance across all metrics, including PSNR, SSIM, and LPIPS. Additionally, as illustrated in Fig. 3, WIPES effectively captures both low-frequency features (e.g., walls and gates) and high-frequency details (e.g., tiles and streetlamps) within the scenes. Although GSImage demonstrates the highest rendering speed, our approach maintains a competitive frame rate, ensuring efficient per + +Table 3. Quantitative comparison on static novel view synthesis with GES [11], which incorporates frequency guidance, across the Mip-NeRF360, Tanks & Temples and DeepBlending datasets. GES(3e-4) uses the default parameters from the original GES paper, while GES(2e-4) adopts the same settings as 3DGS [14]. +(a) Mip-NeRF360 + +
Methods | MetricsPSNR↑SSIM↑LPIPS↓Points↓
GES(3e-4)26.980.72080.32571.521M
WIPES(3e-4)27.110.79580.25181.401M
GES(2e-4)27.360.81480.21723.055M
WIPES(2e-4)27.550.81640.21492.787M
+ +(b)Tanks & Temples + +
Methods | MetricsPSNR↑SSIM↑LPIPS↓Points↓
GES(3e-4)23.440.83940.19540.913M
WIPEs(3e-4)23.760.84330.19140.931M
GES(2e-4)23.590.84810.17721.601M
WIPEs(2e-4)23.780.85230.17091.591M
+ +(c) DeepBlending + +
Methods | MetricsPSNR↑SSIM↑LPIPS↓Points↓
GES(3e-4)29.590.90530.24921.243M
WIPEs(3e-4)29.720.90620.25011.137M
GES(2e-4)29.550.90410.24342.059M
WIPEs(2e-4)29.820.90690.23831.933M
+ +formance without sacrificing quality. This balance between high-quality image reconstruction and rendering efficiency underscores the practical effectiveness of our approach. + +# 5.2. 5D Static Novel View Synthesis + +Configurations. In this section, we evaluated WIPEs on 13 scenes sourced from the Mip-NeRF360 [3], Tanks & Temples [16], and DeepBlending [12] datasets. Specifically, we set the learning rate for the frequency coefficients to $2.5 \times 10^{-3}$ in WIPEs. Differentiable rasterization was implemented via custom CUDA kernels, while all other learnable parameters and experimental settings remained aligned with those of 3DGS [14]. This configuration facilitated a direct comparison with state-of-the-art approaches including Plenoxels [9], INGP [25], Mip-NeRF360, and 3DGS. + +Results. Tab. 2 presents a quantitative comparison between WIPES and leading 5D novel view synthesis approaches, demonstrating that our approach outperforms existing methods in most scenarios. Notably, WIPES exhibits significant advantages in environments with complex frequency components, such as large-scale outdoor scenes. Additionally, Fig. 5 provides qualitative comparisons for selected indoor and outdoor scenes from the dataset. The results indicate that WIPES excels at reconstructing reflective environments (e.g., mirror frames and truck windshields) and effectively captures high-frequency details (e.g., potted plants and distant transmission towers). + +Comparison with frequency-constrained approaches. We compare WIPES with Gaussian-based approaches that utilize frequency priors or guidance. Since FreGS [43] has not released its source code, we present quantitative results against GES [11]. GES employed a densify gradient threshold of $3e^{-4}$ , whereas the original 3DGS [14] used $2e^{-4}$ . To ensure fairness, we evaluated GES with both thresholds. Experiments on the Mip-NeRF360, Tanks & Temples, and DeepBlending datasets (Tab. 3) demonstrate that WIPES achieves superior reconstruction without requiring frequency guidance. Additionally, when using identical densify gradient thresholds, WIPES matches GES in parameter count, highlighting its ability to deliver enhanced results without relying on frequency priors or guidance while maintaining comparable parameter efficiency. + +# 5.3. 6D Dynamic Novel View Synthesis + +Configurations. In this section, we employed the D-NeRF [26] and NeRF-DS [37] datasets to assess performance under varying dynamic conditions. WIPES utilized a frequency coefficient learning rate of $2.5 \times 10^{-3}$ , consistent with the settings of D-3DGS [38]. All other learnable parameters were maintained in alignment with D-3DGS to ensure a fair comparison. Differentiable rasterization was achieved through the integration of custom CUDA kernels, mirroring the configurations used in the 6D tasks. + +Results. Tab. 4 presents the quantitative experimental results of WIPES in comparison with INR-based approaches and Gaussian-based approaches referenced in our work. WIPES exhibits enhanced expressive capabilities for various frequency components in dynamic scenes, enabling similar reconstruction performance to D-3DGS with fewer primitives. This indicates that the deformable INR in WIPES more effectively learns the dynamic features of primitives. Consequently, WIPES achieves superior performance in both reconstruction quality and rendering speed. Fig. 6 showcases qualitative results of WIPES across different dynamic datasets. It is evident that WIPES delivers more detailed and accurate results for both low-frequency elements, such as cloth, and high-frequency textures. + +# 6. Conclusion + +In this paper, we have proposed WIPES, a universal wavelet-based visual primitive for representing multidimensional visual signals. We have enhanced the Morlet function by designing a universally applicable and continuously differentiable wavelet primitive. Without relying on any frequency guidance or priors, WIPES has adaptively captured various real-world visual signals with rich frequency components, achieving impressive results in reconstruction tasks for 2D image representation, 5D static, and 6D dynamic novel view synthesis. To facilitate WIPES, we have developed a fast differentiable wavelet rasterizer that + +Table 4. Quantitative comparison of dynamic novel view synthesis on D-NeRF and NeRF-DS datasets. + +
Methods|MetricsD-NeRFNeRF-DS
PSNR↑SSIM↑LPIPS↓FPS↑PSNR↑SSIM↑LPIPS↓FPS↑
D-NeRF [26]30.430.95720.07240.117----
NeRF-DS [37]----23.580.85820.24960.008
D-3DGS [38]39.160.98950.013271.723.690.84230.197638.36
WIPES39.520.98990.012784.023.950.85270.176242.45
+ +![](images/0dd2feaee69e65a63d859c4c4fb2db550235e7d505697d44535921668afcccae.jpg) +GT +GT + +![](images/564687a84071f43926fa3c574e5ab2d250a99462489a8e139684e5f7b528c475.jpg) +WIPES +WIPES + +![](images/fd159988b28f11f2205edc03ae94b24c8a5d4c86a8c7d8acb1d1d0ca6ec1c770.jpg) +D-3DGS +D-3DGS + +![](images/a21d240ee4cc3a05be0ada33a2ee187ddf69d1d009edfede78d9df686f6118bc.jpg) +D-NeRF +NeRF-DS + +![](images/7aefab6bd447a4e2485efe604c356d031b0e29235c8f30367bcca4ff8837b93c.jpg) + +![](images/0f9a0e00c4be21d31ae43852e1020be8c2a76db680e9da35209ca13bccaaf0de.jpg) +Figure 6. Qualitative comparison on D-NeRF and NeRF-DS datasets. Our approach demonstrates more accurate registration, delivering outstanding performance in both static and dynamic regions. + +![](images/f3f46b7de9497d15afa808ca43704b7c363741eb555a324bb3bcbfcfb19a086b.jpg) + +![](images/2d9f1c5ccd781adc839470e40502ecf4c898a3a9390501c9c082134cd97247eb.jpg) + +![](images/7c0b75cd32ec0213d2d7a548a9539758b50904db3553e425303b23f486cc6913.jpg) + +![](images/69316c3fc6b8ce2d961db30f60247fc3bff728e322cb1182843759e9eca3b74a.jpg) + +![](images/e2f927a728d5e0fbe646bef95202fa9f646f2c78c73688d69261573dbcfb2ce0.jpg) + +![](images/e5a9ca079fe6cd7ba6f967890f7369617e6645b184b569f27e7f2dc2c6e84b48.jpg) + +seamlessly integrates into Gaussian-based pipelines, enhancing rendering speed and efficiency during reconstruction. We have conducted extensive experiments on several challenging datasets across different dimensions, demonstrating the generality and effectiveness of our approach through both qualitative and quantitative evaluations. + +However, our framework has limitations. The current Gaussian-based training pipeline has restricted scene + +representation due to fixed hyperparameters. Additionally, gradient-based densification in Gaussian approaches has caused instability and performance variability. Future work will aim to develop a more robust and efficient wavelet-based training and rendering framework to address these issues, fully leveraging WIPES's potential and advancing continuous signal representation in computer vision. + +# References + +[1] Eirikur Agustsson and Radu Timofte. Ntire 2017 challenge on single image super-resolution: Dataset and study. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 126-135, 2017. 5 +[2] Adrian Azzarelli, Nantheera Anantrasirichai, and David R. Bull. Waveplanes: A compact wavelet representation for dynamic neural radiance fields. arXiv preprint arXiv:2312.02218, 2023. 3 +[3] Jonathan T. Barron, Ben Mildenhall, Dor Verbin, Pratul P. Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5460-5469, 2022. 6, 7 +[4] Yuyao Chen, Lu Lu, George Em Karniadakis, and Luca Dal Negro. Physics-informed neural networks for inverse problems in nano-optics and metamaterials. Opt. Express, 28(8): 11618-11633, 2020. 2 +[5] Jaeyoung Chung, Jeongtaek Oh, and Kyoung Mu Lee. Depth-regularized optimization for 3d gaussian splatting in few-shot images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 811–820, 2024. 2 +[6] Ronald A. DeVore. Nonlinear approximation. Acta Numerica, 7:51-150, 1998. 4 +[7] Eastman Kodak Company. Kodak lossless true color image suite. https://r0k.us/graphics/kodak/, 1999.5 +[8] Rizal Fathony, Anit Kumar Sahu, Devin Willmott, and J. Zico Kolter. Multiplicative filter networks. In International Conference on Learning Representations (ICLR), 2021. 3 +[9] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In CVPR, 2022. 2, 6, 7 +[10] A. Graps. An introduction to wavelets. IEEE Computational Science and Engineering, 2(2):50-61, 1995. 2 +[11] Abdullah Hamdi, Luke Melas-Kyriazi, Jinjie Mai, Guocheng Qian, Ruoshi Liu, Carl Vondrick, Bernard Ghanem, and Andrea Vedaldi. Ges: Generalized exponential splatting for efficient radiance field rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19812-19822, 2024. 2, 7 +[12] Peter Hedman, Julien Philip, True Price, Jan-Michael Frahm, George Drettakis, and Gabriel Brostow. Deep blending for free-viewpoint image-based rendering. ACM Transactions on Graphics, 37(6):257, 2018. 7 +[13] Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2d gaussian splatting for geometrically accurate radiance fields. In *Special Interest Group on Computer Graphics and Interactive Techniques Conference Papers* '24. ACM, 2024. 2 +[14] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics (TOG), 42(4), 2023. 2, 3, 6, 7 + +[15] Rajaei Khatib. Trinerflet: A wavelet-based triplane nerf representation. In ECCV 2024 Workshop on Wild 3D: 3D Modeling, Reconstruction, and Generation in the Wild, 2024. 3 +[16] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Benchmarking large-scale scene reconstruction. Association for Computing Machinery, 36 (4):78, 2017. 7 +[17] Amit Pal Singh Kohli, Vincent Sitzmann, and Gordon Wetzstein. Semantic implicit neural scene representations with semi-supervised training. In 2020 International Conference on 3D Vision (3DV), pages 423-433, 2020. 2 +[18] Christoph Lassner and Michael Zollhöfer. Pulsar: Efficient sphere-based neural rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1440-1449, 2021. 2 +[19] David B. Lindell, Dave Van Veen, Jeong Joon Park, and Gordon Wetzstein. Bacon: Band-limited coordinate networks for multiscale scene representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2 +[20] Zhen Liu, Hao Zhu, Qi Zhang, Jingde Fu, Weibing Deng, Zhan Ma, Yanwen Guo, and Xun Cao. Finer: Flexible spectral-bias tuning in implicit neural representation by variable-periodic activation functions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2713-2722, 2024. 2 +[21] Tao Lu, Mulin Yu, Linning Xu, Yuanbo Xiangli, Limin Wang, Dahua Lin, and Bo Dai. Scaffold-gs: Structured 3d gaussians for view-adaptive rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20654-20664, 2024. 2 +[22] Stéphane Mallat. A Wavelet Tour of Signal Processing. Academic Press, 1999. 3 +[23] Julien N. P. Martel, David B. Lindell, Connor Z. Lin, Eric R. Chan, Marco Monteiro, and Gordon Wetzstein. Acorn: Adaptive coordinate networks for neural scene representation. ACM Transactions on Graphics (TOG), 40(4), 2021. 2 +[24] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In European Conference on Computer Vision, pages 405-421. Springer, 2020. 2, 4 +[25] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics, 41 (4):1-15, 2022. 2, 6, 7 +[26] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: Neural radiance fields for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 7, 8 +[27] Nasim Rahaman, Aristide Baratin, Devansh Arpit, Felix Draxler, Min Lin, Fred Hamprecht, Yoshua Bengio, and Aaron Courville. On the spectral bias of neural networks. In Proceedings of the International Conference on Machine Learning (ICML), pages 5301-5310. PMLR, 2019. 1, 2 + +[28] Sameera Ramasinghe and Simon Lucey. Beyond periodicity: Towards a unifying framework for activations in coordinate-mpls. In European Conference on Computer Vision, pages 142-158. Springer, 2022. 2 +[29] Christian Reiser, Songyou Peng, Yiyi Liao, and Andreas Geiger. Kilonerf: Speeding up neural radiance fields with thousands of tiny mlps. In Proceedings of the IEEE/CVF international conference on computer vision, pages 14335-14345, 2021. 2 +[30] Vishwanath Saragadam, Jasper Tan, Guha Balakrishnan, Richard G. Baraniuk, and Ashok Veeraraghavan. Miner: Multiscale implicit neural representation. In Proceedings of the European Conference on Computer Vision (ECCV), pages 318-333. Springer, 2022. 2 +[31] Vishwanath Saragadam, Daniel LeJeune, Jasper Tan, Guha Balakrishnan, Ashok Veeraraghavan, and Richard G. Baraniuk. Wire: Wavelet implicit neural representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18507-18516, 2023. 2, 4, 6 +[32] Chengkang Shen, Hao Zhu, You Zhou, Yu Liu, Si Yi, Lili Dong, Weipeng Zhao, David J Brady, Xun Cao, Zhan Ma, et al. Cardiac field: computational echocardiography for automated heart function estimation using two-dimensional echocardiography probes. European Heart Journal-Digital Health, page ztae072, 2024. 2 +[33] Chengkang Shen, Hao Zhu, You Zhou, Yu Liu, Si Yi, Lili Dong, Weipeng Zhao, David J Brady, Xun Cao, Zhan Ma, et al. Continuous 3d myocardial motion tracking via echocardiography. IEEE Transactions on Medical Imaging, 2024. 2 +[34] Vincent Sitzmann, Julien Martel, Alexander Bergman, David Lindell, and Gordon Wetzstein. Implicit neural representations with periodic activation functions. In Advances in Neural Information Processing Systems, pages 7462-7473, 2020. 1, 2 +[35] Matthew Tancik, Pratul Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan Barron, and Ren Ng. Fourier features let networks learn high frequency functions in low dimensional domains. Advances in Neural Information Processing Systems (NeurIPS), 33:7537-7547, 2020. 1, 2 +[36] Delong Wu, Hao Zhu, Qi Zhang, You Li, Zhan Ma, and Xun Cao. Neural poisson solver: A universal and continuous framework for natural signal blending. In Computer Vision - ECCV 2024, pages 259-275, Cham, 2025. Springer Nature Switzerland. 2 +[37] Zhiwen Yan, Chen Li, and Gim Hee Lee. Nerf-ds: Neural radiance fields for dynamic specular objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8285-8295, 2023. 7, 8 +[38] Ziyi Yang, Xinyu Gao, Wen Zhou, Shaohui Jiao, Yuqing Zhang, and Xiaogang Jin. Deformable 3d gaussians for high-fidelity monocular dynamic scene reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20331-20341, 2024. 2, 4, 7, 8 +[39] Vickie Ye, Ruilong Li, Justin Kerr, Matias Turkulainen, Brent Yi, Zhuoyang Pan, Otto Seiskari, Jianbo Ye, Jeffrey + +Hu, Matthew Tancik, and Angjoo Kanazawa. gsplat: An open-source library for gaussian splatting. arXiv preprint arXiv:2409.06765, 2024. 5 +[40] Zehao Yu, Anpei Chen, Binbin Huang, Torsten Sattler, and Andreas Geiger. Mip-splatting: Alias-free 3d gaussian splattering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19447-19456, 2024. 2 +[41] Gizem Yüce, Guillermo Ortiz-Jiménez, Beril Besbinar, and Pascal Frossard. A structured dictionary perspective on implicit neural representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19228-19238, 2022. 1 +[42] Ruyi Zha, Tao Jun Lin, Yuanhao Cai, Jiwen Cao, Yanhao Zhang, and Hongdong Li. $\mathbb{R}^2$ -gaussian: Rectifying radiative gaussian splatting for tomographic reconstruction. In Advances in Neural Information Processing Systems (NeurIPS), 2024. 2 +[43] Jiahui Zhang, Fangneng Zhan, Muyu Xu, Shijian Lu, and Eric Xing. Fregs: 3d gaussian splatting with progressive frequency regularization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 21424-21433, 2024. 2, 7 +[44] Xinjie Zhang, Xingtong Ge, Tongda Xu, Dailan He, Yan Wang, Hongwei Qin, Guo Lu, Jing Geng, and Jun Zhang. Gaussianimage: 1000 fps image representation and compression by 2d gaussian splatting. In Proceedings of the European Conference on Computer Vision (ECCV), 2024. 1, 2, 4, 6 +[45] Hao Zhu, Zhen Liu, You Zhou, Zhan Ma, and Xun Cao. Dnf: Diffractive neural field for lensless microscopic imaging. Optics Express, 30(11):18168-18178, 2022. 2 +[46] Hao Zhu, Zhen Liu, Qi Zhang, Jingde Fu, Weibing Deng, Zhan Ma, Yanwen Guo, and Xun Cao. Finer++: Building a family of variable-periodic functions for activating implicit neural representation. arXiv preprint arXiv:2407.19434, 2024. 1, 2 +[47] Hao Zhu, Shaowen Xie, Zhen Liu, Fengyi Liu, Qi Zhang, You Zhou, Yi Lin, Zhan Ma, and Xun Cao. Disorder-invariant implicit neural representation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 2 +[48] Hao Zhu, Fengyi Liu, Qi Zhang, Zhan Ma, and Xun Cao. Rhino: Regularizing the hash-based implicit neural representation. SCIENCE CHINA Information Sciences, pages -, 2025. 2 +[49] Junyu Zhu, Hao Zhu, Qi Zhang, Fang Zhu, Zhan Ma, and Xun Cao. Pyramid nerf: Frequency guided fast radiance field optimization. International Journal of Computer Vision, pages 1-16, 2023. 2 +[50] Matthias Zwicker, Hanspeter Pfister, Jeroen van Baar, and Markus Gross. Ewa volume splatting. In Proceedings Visualization, 2001. VIS '01., pages 29-538, 2001. 2, 3 +[51] Matthias Zwicker, Hanspeter Pfister, Jeroen van Baar, and Markus Gross. Surface splatting. In Proceedings of the 28th Annual Conference on Computer Graphics and Interactive Techniques, pages 371-378, New York, NY, USA, 2001. Association for Computing Machinery. 2 \ No newline at end of file diff --git a/ICCV/2025/WIPES_ Wavelet-based Visual Primitives/images.zip b/ICCV/2025/WIPES_ Wavelet-based Visual Primitives/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..7171623170bf428cb539977fc22073427caad440 --- /dev/null +++ b/ICCV/2025/WIPES_ Wavelet-based Visual Primitives/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e868652cf75737491938888a24b842e6a1e50e5aa5cf6a3f44bd2349b612b40f +size 1091975 diff --git a/ICCV/2025/WIPES_ Wavelet-based Visual Primitives/layout.json b/ICCV/2025/WIPES_ Wavelet-based Visual Primitives/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..375c918a599d5898e02af602fb104c47d08c29ed --- /dev/null +++ b/ICCV/2025/WIPES_ Wavelet-based Visual Primitives/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:36609edb71222240109a155728886f42fd29133e3513c4477be68359f372021f +size 500531 diff --git a/ICCV/2025/WIR3D_ Visually-Informed and Geometry-Aware 3D Shape Abstraction/49efbaa1-4dc1-4f95-8026-29529291070b_content_list.json b/ICCV/2025/WIR3D_ Visually-Informed and Geometry-Aware 3D Shape Abstraction/49efbaa1-4dc1-4f95-8026-29529291070b_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..e2c5af22a2a86e6e269e51d0baf365fa43f48939 --- /dev/null +++ b/ICCV/2025/WIR3D_ Visually-Informed and Geometry-Aware 3D Shape Abstraction/49efbaa1-4dc1-4f95-8026-29529291070b_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:598601636ac5b614d5ce25444b4344db87323eb85b09214f1801afb4fc16c0b0 +size 77768 diff --git a/ICCV/2025/WIR3D_ Visually-Informed and Geometry-Aware 3D Shape Abstraction/49efbaa1-4dc1-4f95-8026-29529291070b_model.json b/ICCV/2025/WIR3D_ Visually-Informed and Geometry-Aware 3D Shape Abstraction/49efbaa1-4dc1-4f95-8026-29529291070b_model.json new file mode 100644 index 0000000000000000000000000000000000000000..ef4bdde455db239599fca738ee16f1185e67af69 --- /dev/null +++ b/ICCV/2025/WIR3D_ Visually-Informed and Geometry-Aware 3D Shape Abstraction/49efbaa1-4dc1-4f95-8026-29529291070b_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:07343252398378ec955cc7a68265f19c6c268994461b1dd11e1354ee20bd7417 +size 100918 diff --git a/ICCV/2025/WIR3D_ Visually-Informed and Geometry-Aware 3D Shape Abstraction/49efbaa1-4dc1-4f95-8026-29529291070b_origin.pdf b/ICCV/2025/WIR3D_ Visually-Informed and Geometry-Aware 3D Shape Abstraction/49efbaa1-4dc1-4f95-8026-29529291070b_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..3c91e18c8b0cdba4adc4742d9ac2a60891d120c7 --- /dev/null +++ b/ICCV/2025/WIR3D_ Visually-Informed and Geometry-Aware 3D Shape Abstraction/49efbaa1-4dc1-4f95-8026-29529291070b_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3a23a160420a8a299476d536f2a811ab96d2aa0c5f84092ad1d923e97a6c6063 +size 33549903 diff --git a/ICCV/2025/WIR3D_ Visually-Informed and Geometry-Aware 3D Shape Abstraction/full.md b/ICCV/2025/WIR3D_ Visually-Informed and Geometry-Aware 3D Shape Abstraction/full.md new file mode 100644 index 0000000000000000000000000000000000000000..d40b4085ecbd9e4f82447b5dddbe56b4de3e9c52 --- /dev/null +++ b/ICCV/2025/WIR3D_ Visually-Informed and Geometry-Aware 3D Shape Abstraction/full.md @@ -0,0 +1,357 @@ +# WIR3D: Visually-Informed and Geometry-Aware 3D Shape Abstraction + +Richard Liu +University of Chicago +guanzhi@uchicago.edu + +Daniel Fu +University of Chicago +danielfu@uchicago.edu + +Noah Tan +University of Chicago +tntan@uchicago.edu + +Itai Lang +University of Chicago +itailang@uchicago.edu + +Rana Hanocka +University of Chicago +ranahanocka@uchicago.edu + +![](images/48131e9b2e8f6b59cd9bed11bb6e08299586180d161632ad6dda96f3ff76f6dc.jpg) +Figure 1. WIR3D produces 3D shape abstractions in the form of 3D strokes. The abstractions retain the overall shape structure and capture texture concepts (e.g. dragon scales and watermelon seeds) as well as key salient features (e.g. facial features). + +# Abstract + +In this work we present WIR3D, a technique for abstracting 3D shapes through a sparse set of visually meaningful curves in 3D. We optimize the parameters of Bézier curves such that they faithfully represent both the geometry and salient visual features (e.g. texture) of the shape from arbitrary viewpoints. We leverage the intermediate activations of a pre-trained foundation model (CLIP) to guide our optimization process. We divide our optimization into two phases: one for capturing the coarse geometry of the shape, and the other for representing fine-grained features. Our second phase supervision is spatially guided by a novel localized keypoint loss. This spatial guidance enables user control over abstracted features. We ensure fidelity to the original surface through a neural SDF loss, which allows the curves to be used as intuitive deformation handles. We successfully apply our method for shape abstraction over a broad dataset of shapes with varying complexity, geometric structure, and texture, and demonstrate downstream applications for feature control and shape deformation. + +# 1. Introduction + +In this work, we explore whether it is possible to abstract a 3D shape into a sparse set of semantically-informed curves. + +A key challenge lies in finding a sparse set of curves that best represents the shape's visual features. We use this term deliberately, to encompass both the geometry and texture features which are salient to humans. This problem cannot be solved with surface analysis alone, which focuses on low-level geometric contours. Our task is much broader, in that we wish to capture high-level concepts, visually salient geometry, and textures (e.g. Fig. 1). + +Existing works have dealt with the problem of computing occluding contours for non-photorealistic rendering [5, 29, 31]. Occluding contours relies entirely on surface analysis, which is subject to the aforementioned limitations. Furthermore, occluding contours is a 2D representation, while we specifically seek a view-consistent, 3D representation. Occluding contours is view-inconsistent by construction, which results in the commonly observed flickering artifact when rendering dense views of a 3D shape [21]. + +Our objective is to abstract visual shape features into a sparse set of 3D strokes. We optimize the parameters of a set of Bezier curves as our stroke representation, where the number of curves determines the level of abstraction. + +We show a motivating example in Fig. 2, where we abstract a cylinder with a partial texture, and compare against a naive solution of back-projecting 2D occluding contour curves. The naive solution calculates a geometric contour at each view (inset), backprojects the result to 3D curves, + +![](images/870e9b1e1153435dcce8d3392e06648b0eb86072fba1f91de5f6c79adbd1b24b.jpg) +Figure 2. Motivating example. Abstracting even a simple shape like a cylinder is nontrivial. Rendering the contours of the cylinder (inset) results in a static image from every view, removing any sense of 3D volume. Backprojecting these contours into 3D results in a dense cluster of lines spanning the body of the cylinder, which is both non-sparse and unsatisfying aesthetically. In contrast, our method effectively abstracts the 3D geometry of the cylinder with a few strokes along with the texture. + +and aggregates across all views. The naive solution only accounts for the cylinder geometry in a view-dependent manner, resulting in a dense body of strokes. Importantly, the naive solution cannot handle the texture at all. + +Our solution represents the cylinder volume with sparse vertical curves. Different views of the shape are visually distinguished, yet the cylinder's overall geometry is consistent and clear from any given view. Our method also preserves the high-level pattern of the texture. + +To achieve our abstraction, we leverage spatially-conditioned semantic supervision by utilizing the intermediate activations of a 2D pre-trained foundation model (CLIP [48]). Our choice of model is justified by previous work that demonstrates CLIP's strong performance in tasks requiring high-level abstract understanding across different modalities [16, 18, 62]. We observe that both different layers and architectures enable different levels of control over geometry and texture elements, so we split our optimization into separate geometry and texture abstraction stages. + +Our spatial conditioning comes from 3D keypoints, which determine the weight specific visual features are given in our abstraction. The keypoints being in 3D ensures their multi-view consistency and enables fine-grained user control. We show in Fig. 11 an application where users can iteratively add detail to the abstraction through selecting keypoints corresponding to salient features. + +We encourage curve adherence to the input geometry using a neural SDF loss. The surface compliance enables an application where the curves are used as intuitive deformation handles, shown in Fig. 10 and the supplemental. Our curves' effectiveness as control handles stems from their alignment to semantically meaningful regions of the surface and texture. + +In summary, we present a novel technique for abstracting 3D shapes through a sparse set of visually-informed 3D curves. WIR3D can abstract a myriad of shapes from different domains with various visual concepts, geometric structures, and textures. The abstractions are sparse yet effective and maintain high fidelity across arbitrary views. Our novel localized weighting framework enables interactive user control over the features represented in the abstraction, and our adherence to the input geometry allows for the + +curves' use as intuitive deformation handles. We encourage the reader to examine the supplemental material, which contains 360-degree videos of our results and an interactive demo of the deformation application using our optimized curves. We plan to release the code for the method and proposed applications in the near future. + +# 2. Related Work + +Shape decomposition. Shape decomposition is a longstanding problem in 3D geometry analysis, where the goal is to represent the shape by a set of elements, such as primitives [34, 55, 59, 67], convex parts [12, 24, 40, 45], or Gaussians [19, 20]. Tulsiani et al. [59] assemble 3D objects from cuboids and obtained simple shape abstractions with consistent structure. In Cvxnet [12], the authors reconstruct a 3D shape by a collection of convex polytopes. Paschalidou et al. [44] extend this approach by learning the decomposition and the parts for a set of domain-specific meshes. + +This line of work aims at reconstruction, whereas our goal is abstraction. Moreover, our method does not depend on a dataset and is not limited to a specific domain. It is robust to shapes of arbitrary quality and complexity. + +Non-photorealistic rendering. A classic problem in graphics is identifying visual contours from 3D geometry to create non-photorealistic renderings. A popular version of this problem is occluding contours [1, 2, 10, 25, 37, 49], in which contours representing visible shape regions are delineated from occluded ones. In the classic setting, visible contours are exactly defined to mean the visible surface points tangent to the view vector. Recent efforts have aimed at improving view-consistency and alignment with professional artists [5], and some apply neural techniques [30, 31]. While such works can successfully depict visual features based on the shape geometry, they lack 3D consistency, and frequently suffer from flickering artifacts [21]. + +A recent paper [69] and its follow-up [26] introduce implicit edge fields for 3D curve reconstruction from posed images. These works emphasize geometric boundaries, which is distinct from our focus on abstraction. A high-level abstraction may reduce a shape feature to a single stroke (e.g. Fig. 8 chair), which is not achievable by these + +![](images/7e2ab5b7ff89bf7d8e7acee1e2dcfde33c4065cc28d2630491ea46739538ca42.jpg) +Stage 1: Geometry Abstraction + +![](images/2c46b59f3dc3df0225760f7889eefa683a044070effed02e7c30f9c8e307d0ec.jpg) +Stage 2: Texture Abstraction +Figure 3. Method overview. In the first stage, our WIR3D learns to abstract the underlying geometry of the shape. In the second stage, we freeze the curves from the first stage and add new curves that are optimized to represent the shape's texture. + +methods. Furthermore, these methods cannot account for texture features which do not induce a prominent edge map. + +Curve-based abstraction. Our work is in the domain of sketch abstraction, which aims to depict a scene through 2D or 3D curve primitives. [7, 14, 22, 38, 61]. Most prior work is based on 2D sketch abstractions [6, 13, 15, 23, 27, 43, 46, 52, 56, 61, 62], though several works have aimed to represent 2D images through a 3D structures by projection along orthogonal axes [22, 39, 47]. Our task is different in that we aim to represent a single shape from all possible viewing directions, not simply three orthogonal views. + +Another line of work aims to reconstruct fabricable 3D wire structures from different modalities, including images [8, 32, 33], video [64], and 3D data [4, 57, 68]. These works are constrained by the fabrication objective, and thus do not abstract fine-grained visual details. + +A final strain of literature deals with analysis of 2D sketches of 3D structures to identify part information and aid in user sketch generation [17, 42, 50]. + +A recent work proposed optimizing strokes in 3D to match a text description or a guidance image [71]. This work is generative in nature (e.g. text-to-3D sketch) and does not involve spatial nor 3D inputs as in our work. 3Doodle [7] presents a technique for optimizing view-dependent and view-independent curves to represent a set of multiview images. Our work is orthogonal in spirit - we aim to use a set of view-independent curves for 3D abstraction of an input shape, and we build our framework leveraging the geometric and semantic shape information. Notably, our use of spatially-driven guidance and SDF loss enables our detail control and deformation applications. + +We use a variant of 3Doodle with only view-independent curves as our main baseline and show results in Fig. 8. + +# 3. Method + +WIR3D optimizes a set of 3D cubic Bezier curves to abstract a target (potentially textured) shape from all viewing angles. The method takes as input a 3D model and optional user-selected keypoints on the surface. When no keypoints are provided, we automatically detect keypoints using latent backprojection and clustering (Sec. 3.3). + +# 3.1. Curve Representation + +Our 3D strokes are modeled as a set of cubic Bézier curves $\{B_i\}_{i=1}^n$ , with control points $B_i = (p_i^0, p_i^1, p_i^2, p_i^3)$ , $p_i^j \in \mathbb{R}^3$ . Points on the curve are sampled through polynomial interpolation of the four control points + +$$ +\begin{array}{l} B (t) = (1 - t) ^ {3} p ^ {0} + (1 - t) ^ {2} t p ^ {1} \tag {1} \\ + (1 - t) t ^ {2} p ^ {2} + t ^ {3} p ^ {3} \\ \end{array} +$$ + +where $0\leq t\leq 1$ + +We make the same assumption as 3Doodle [7] that the camera is sufficiently far from the shape such that orthographic and perspective projection are nearly identical. Theorem 1 from 3Doodle thus applies to our method, which establishes equivalence between the normal 3D Bézier curves we optimize and the space of 2D rational Bézier curves, generated from projecting the 3D curve control points into 2D. Our pipeline renders the 3D Bézier curves by first perspective-projecting the 3D control points, then rasterizing the 2D cubic Bézier defined from the projected control points (interpolated as in Eq. (1)). This process is encapsulated by the "Differentiable Render" step in Fig. 3. The differentiable rasterizer we use is DiffVG [28]. + +The resulting image from this rasterization process is $I_{\mathrm{curve}} = \mathcal{R}(\{B_i\})$ . We refer to target shape renders as $I_{\mathrm{target}}$ . + +![](images/f9d3d00686199fb194244aea501590690b6f166e69e3fe9639cf72157dbb38cd.jpg) +Figure 4. Localized keypoint loss. Our localized keypoints weight the loss between the intermediate feature maps of the encoded curve render $I_{\mathrm{curve}}$ and the target shape render $I_{\mathrm{target}}$ . This weight is obtained through projecting 3D keypoints (red), followed by a Gaussian filter to obtain the weight map $I_{\mathrm{weight}}$ . This loss focuses the optimization on visual features local to the keypoint. + +# 3.2. Losses + +We leverage the priors of 2D pretrained image encoders to define our semantic losses. Specifically, we design specialized perceptual losses using CLIP [48] and LPIPS [70] to encourage our rendered 3D strokes to visually match the corresponding renders of the target shape. + +The basic structure of our semantic loss is adapted from CLIPasso [61], which compares both the intermediate spatial activations and final global activations between the rendered strokes and the target shape: + +$$ +\begin{array}{l} \mathcal {L} _ {\text {s e m a n t i c}} = \lambda_ {\mathrm {f c}} \operatorname {d i s t} \left(\operatorname {C L I P} \left(I _ {\text {c u r v e}}\right), \operatorname {C L I P} \left(I _ {\text {t a r g e t}}\right)\right) \tag {2} \\ + \sum_ {l = 3, 4} | | \mathrm {C L I P} _ {l} \left(I _ {\text {c u r v e}}\right) - \mathrm {C L I P} _ {l} \left(I _ {\text {t a r g e t}}\right) | | _ {2} ^ {2} \\ \end{array} +$$ + +where $\mathrm{dist}(\cdot, \cdot)$ measures cosine similarity, CLIP is the global CLIP encoding, and $\mathrm{CLIP}_l$ is the layer $l$ activation. + +We find this semantic loss, however, to be lacking when it comes to representing fine details, including textures. Thus, we introduce a spatial weighting framework, which directs the optimization towards specific visual features. + +Localized Keypoint Loss. Our spatial weighting is based on previous work that establishes the spatial correlation between CLIP's intermediate activations and the input image. [51]. Features of interest can be emphasized in these intermediate layers by identifying their location in the render and tracing the correspondence to the downsampled feature maps, hence localizing the features in the feature maps. + +We assume as optional input user-defined 3D keypoints. If no input keypoints are given then our method automatically detects keypoints, as described in Sec. 3.3 and the supplemental. These keypoints should indicate salient visual features on the input shape or texture, and we leverage that information to guide abstraction of the specific features. + +For every sampled view during optimization, we project the keypoints to the same views and identify their positions + +in the corresponding renders. Once we have the keypoint locations in the rendered image, we construct a weight map at the same resolution as the image based on a Gaussian dropoff from the keypoint center. Specifically, we construct the weight image $I_{\mathrm{weight}}$ such that + +$$ +I _ {\text {w e i g h t}} (x, y) = 1 + \sum_ {p} e ^ {\frac {- | | (x , y) - p | | ^ {2}}{2 \sigma^ {2}}} \tag {3} +$$ + +where $x, y \in [0,1]$ index the normalized image pixels, and $p \in [0,1]$ is the keypoint projected to normalized coordinates. We add 1 to the weights so that regions far from the keypoints still contribute to the loss. We use these constructed weight maps to weight our semantic loss Eq. (2) for each render, such that each L2 term in the intermediate layer losses $||\mathrm{CLIP}_l(I_{\mathrm{curve}}) - \mathrm{CLIP}_l(I_{\mathrm{target}})||_2^2$ becomes $||I_{\mathrm{weight}} \cdot (\mathrm{CLIP}_l(I_{\mathrm{curve}}) - \mathrm{CLIP}_l(I_{\mathrm{target}}))||_2^2$ . The weight map is downsampled to match the resolution of each intermediate feature map. We visualize this process in Fig. 4. + +We maintain a z-buffer for the shape renders, such that if a keypoint's projected depth is higher than the shape depth, then the keypoint is occluded by the surface from that view and does not contribute to the weight image. This prevents keypoints from being attributed to incorrect surface regions. + +To provide additional structural constraints, we include an LPIPS loss term [70]. LPIPS is a popular perceptual loss function which is known to be sensitive to geometric layouts [7]. The final localized keypoint loss becomes: + +$$ +\begin{array}{l} \mathcal {L} _ {\text {l o c a l}} = \lambda_ {\mathrm {f c}} \bar {I} _ {\text {w e i g h t}} \operatorname {d i s t} \left(\operatorname {C L I P} \left(I _ {\text {c u r v e}}\right), \operatorname {C L I P} \left(I _ {\text {t a r g e t}}\right)\right) \tag {4} \\ + \sum_ {l = 3, 4} | | I _ {\text {w e i g h t}} \cdot \left(\operatorname {C L I P} _ {l} \left(I _ {\text {c u r v e}}\right) - \operatorname {C L I P} _ {l} \left(I _ {\text {t a r g e t}}\right)\right) | | _ {2} ^ {2} \\ + \lambda_ {\mathrm {l p i p s}} \mathrm {L P I P S} \left(I _ {\text {c u r v e}}, I _ {\text {t a r g e t}}\right) \\ \end{array} +$$ + +where $\bar{I}_{\mathrm{weight}}$ indicates the mean-pooled weights. + +SDF Regularization. To encourage adherence of the curves to the target shape geometry, we use a loss based on the shape's Signed Distance Field (SDF). We fit an MLP $\Phi$ on the shape's SDF (Unsigned Distance Field in the case of shapes with boundaries) to obtain a neural SDF. During stroke optimization, we densely sample each 3D curve and query their SDF values using the neural SDF, penalizing values outside of the zero level set: + +$$ +\mathcal {L} _ {\mathrm {S D F}} = \frac {1}{n \cdot k} \sum_ {i = 1} ^ {n} \sum_ {k = 1} ^ {s} | \phi (B _ {i} (t _ {k}) | \tag {5} +$$ + +where $t_k \in [0,1]$ are random samples along the Bezier curve for $s$ total samples. This loss helps to anchor abstracted features to the surface implied by the curve set. + +View Regularization. We further regularize the abstraction by enforcing that all curves are visible from all sampled viewing angles, which prevents curves from hiding outside of the camera view. With $\mathcal{P}$ indicating the perspective projection of the Bezier curve control keypoints, we have: + +$$ +\mathcal {L} _ {\mathrm {n d c}} = \sum_ {i = 1} ^ {n} \sum_ {t} \operatorname {R e L U} \left(\mathcal {P} \left(B _ {i} (t)\right) - 1\right) + \operatorname {R e L U} \left(- \mathcal {P} \left(B _ {i} (t)\right)\right) \tag {6} +$$ + +where ReLU is a Rectified Linear Unit. This penalizes all curve points with image coordinates outside of the render frame (0-1). The first term has non-zero value when the coordinates are greater than 1, and the second term has nonzero value when the coordinates are less than 0. + +# 3.3. Two-Stage Optimization + +In our experiments we find that different CLIP architectures are sensitive to geometry or semantic shape features. To exploit this, we use a two-stage training pipeline, in which the first stage optimizes for the shape geometry and the second stage abstracts the shape texture, and each stage leverages different CLIP architectures and semantic losses. + +Keypoint Initialization. When keypoints are not included in the input, we automatically identify keypoints of interest on the shape's surface using the 2D to 3D feature back-projection method introduced in Backto3D [66] with KMeans clustering [35]. See supplemental for more details. + +Geometry Abstraction. During stage I optimization, the Bézier curves are initialized using furthest point sampling, with the control points drawn from small Gaussians around each sampled point. These curves are optimized towards the shape geometry with a combination of our original (non-localized) semantic (2), SDF (5), and NDC (6) losses, supervised with Freestyle renders (see supplemental) $I_{\mathrm{target}}^{\mathrm{free}}$ of the target shape. The stage I loss is: + +$$ +\mathcal {L} _ {\mathrm {I}} = \mathcal {L} _ {\text {s e m a n t i c}} \left(I _ {\text {c u r v e}}, I _ {\text {t a r g e t}} ^ {\text {f r e e}}\right) + 0. 1 \cdot \mathcal {L} _ {\text {S D F}} + \mathcal {L} _ {\text {n d c}} \left(\left\{B _ {i} \right\}\right). +$$ + +Texture Abstraction. In our second stage, we freeze the set of geometry curves optimized in the first stage and initialize a new set of curves in the same way as the first stage, using furthest distance sampling. These curves are then optimized to represent the semantic texture of the shape using our localized keypoint loss (4), SDF (5), and NDC (6) losses, supervised with surface (potentially textured) renders of the shape $I_{\mathrm{target}}^{\mathrm{surface}}$ . The stage II loss is: + +$$ +\mathcal {L} _ {\mathrm {I I}} = \mathcal {L} _ {\text {l o c a l}} \left(I _ {\text {c u r v e}}, I _ {\text {t a r g e t}} ^ {\text {s u r f a c e}}\right) + \mathcal {L} _ {\text {S D F}} + \mathcal {L} _ {\text {n d c}} \left(\left\{B _ {i} \right\}\right). +$$ + +We use CLIP architectures RN101 and RN50x64 for $\mathcal{L}_{\mathrm{semantic}}$ and $\mathcal{L}_{\mathrm{local}}$ , respectively. We've found empirically + +![](images/366edff74ee510e863065d51c3ea1d4cbc9abdf249b417ec8bf7391e29e37681.jpg) +Figure 5. Qualitative results for textured objects. We show WIR3D's the result on a collection of textured meshes. + +that RN101 tends to be more sensitive to geometric structures, whereas RN50x64 is more sensitive to higher-level visual concepts. + +# 4. Experiments + +We evaluate WIR3D across a wide variety of shapes and demonstrate multi-view fidelity and abstraction control in Sec. 4.1. We compare WIR3D quantitatively and qualitatively to relevant baselines in Sec. 4.2. Finally, Sec. 4.4 showcases the interactive feature control and shape deformation applications enabled by our method. + +The shapes in our experiments are aggregated from COSEG [60], the Meta Digital Twin Catalog [41], and Keenan Crane's 3D model repository [9]. WIR3D is robust to meshes with varying topology, complexity and quality. + +For all experiments, we randomly sample views spanning 0 to 30 degrees elevation and 0 to 360 degrees azimuth. We use 20 curves for stage 1 and initialize 20 additional curves for stage 2. + +# 4.1. Qualitative Results + +Shape Abstractions. Fig. 5 shows 3D stroke abstractions generated by WIR3D for textured shapes. WIR3D successfully abstracts textures and small salient structures, such as watermelon seeds (Row 1) and spots (Rows 5/6), as well as facial features (Rows 5/7). We show results on untextured + +![](images/37e70e1f40f9423be741becfcddac6dc12367b49779a26c85308960c7cf7db2d.jpg) +Figure 6. Abstraction control. The level of abstraction is implicitly controlled by the number of curves. As the curve count grows, the abstraction captures more fine details. + +shapes in Fig. 15 in the supplemental, which demonstrate that our method can also represent complex geometries. + +Abstraction Control. Our method automatically adapts the level of abstraction based on the number of strokes being optimized. Fig. 6 shows that adding more strokes adds progressively more detail to the abstraction, and our method produces high quality abstractions across all levels. + +In-The-Wild Abstraction. WIR3D has no requirements on the quality of the input geometry. Only the Freestyle rendering will be affected by poor geometry, but CLIP supervision is robust to noisy render artifacts due to its global semantic understanding. We show in Fig. 7 that WIR3D successfully abstracts shapes built from photo reconstruction software, which have many geometric defects. + +# 4.2. Baseline Comparisons + +The baselines we compare to are NEF [69] and 3Doodle [7]. For both methods, we use the publicly available repositories published by the respective authors. We standardize the sampled views and use the default hyperparameters otherwise. 3Doodle initializes its curves using SFM, but given the method's sensitivity to poor initialization, we standardize initialization to the same furthest point samples as our method, finding that they perform better than SFM. We use 40 curves for 3Doodle to match WIR3D, but NEF has no ability to determine the curve count. + +NEF often struggles to fit reasonable point clouds, which results in meaningless curves. We show qualitative comparison to NEF in the supplementary, where we train NEF on a set of viewpoints optimized for the method. + +Qualitative Comparison. Fig. 8 compares against WIR3D against 3Doodle for both textured and untextured + +![](images/511ad16e1a56668026d49f9bd0dbb17395b7726c49131a2d2862eed48a3a0ccd.jpg) +Figure 7. In-The-Wild Reconstructions. We abstract two shapes from the Real-World-Textured Things dataset [36], which provides textured models reconstructed using modern off-the-shelf photo-reconstruction tools on in-the-wild images. The shoe (top) contains 360 boundary edges, 31 non-manifold edges, and 9 non-manifold vertices. The sculpture (bottom) contains 113601 boundary edges and 1274 non-manifold edges. + +
MethodLPIPS (↓)CLIPimg (↑)User Rank (↑)Coverage (↓)
NEF0.3130.86-0.056
3Doodle0.2460.9000.120.020
WIR3D (Ours)0.2270.9090.880.008
+ +Table 1. We compare WIR3D, 3Doodle [7], and NEF [69] through 3 perceptual metrics computed over novel views of the curves and the target shape. "LPIPS" measures the average LPIPS similarity score using AlexNet. "CLIPimg" measures the average cosine similarity scaled between (0,1) using the ViT/B-32 model. "User Rank" measures the percentage of user responses that prefer one method over the other ( $N = 96$ ). We also compute a geometric metric "Coverage", which measures the 1-direction Chamfer distance between the curves and the target 3D surface. + +shapes. Our method consistently captures the global geometry, whereas 3Doodle struggles in a low signal context such as textured shapes (missing back left chair leg, flattened bird geometry). 3Doodle is consistently insensitive to facial structures, whereas our localized weighting ensures that we abstract them (rows 2, 6). Similarly, we are able to represent fine-grained texture structures as high-level visual motifs, such as the dragon scales (row 3). + +Quantitative Comparison. We use three perceptual metrics and one geometric metric, reported in Tab. 1. The first two, LPIPS [70] and $\mathrm{CLIP}^{\mathrm{img}}$ [63], are the same metrics reported by 3Doodle. Our method outperforms 3Doodle on both metrics, though numerically our $\mathrm{CLIP}^{\mathrm{img}}$ is not much higher. We note that CLIP has been shown to be insensitive to large visual differences, as systematized in [58, 65]. We further illustrate this point in Fig. 14 (supplemental), where we show that a randomly oriented cow outline obtains a similar $\mathrm{CLIP}^{\mathrm{img}}$ score as the edge map [3] of the ground + +![](images/e2c5d3a6d92165d4fb1fa3bc69cf967c2a729961ce61a997d4d7bb3378e11d1f.jpg) +Figure 8. Qualitative comparison. We compare a subset of our textured and untextured results against 3Doodle. 3Doodle produces reasonable abstractions that capture the overall structure but lack precision when it comes to the feature details. + +truth render. Thus, the numerical difference in $\mathrm{CLIP}^{\mathrm{img}}$ is an unreliable indicator of the difference in visual quality. + +This motivates our user perceptual study ( $N = 96$ ), which takes all the results from our dataset and compares Wir3D results against 3Doodle. Specifically, we display rotating gifs of both results side by side, along with a rotating gif of the target shape, and ask users to rank the 3D strokes based on how well they represent the target shape. We collected responses from 96 users and compute the frequency our method is ranked higher than 3Doodle. Our curves are ranked higher $88\%$ of the time. Screenshots from the study are shown in the supplemental. + +Our final metric, "Coverage", is a geometric metric that quantifies how well the stroke abstractions cover the original surface. We measure coverage by sampling 100k surface points and computing the 1-direction Chamfer distance from the points to the strokes. This evaluates whether each point on the surface has a curve reasonably close to it. Our method significantly outperforms the baselines in terms of coverage, with a $>2\mathrm{x}$ reduction in coverage distance relative to 3Doodle, and $>4\mathrm{x}$ reduction relative to NEF. + +
MethodLPIPS (↓)CLIPimg (↑)Coverage (↓)
WIR3D0.2270.9090.008
No SDF0.2290.9040.012
No Local0.2330.9050.009
w/o Stage 10.2480.9000.016
No CLIP Layers0.2940.8910.012
+ +Table 2. Ablation Quantitative Metrics. "WIR3D" is our full method. "w/o Stage 1" is the Stage 1 ablation. "No CLIP layers" is the ablation on CLIP intermediate layer activations. "No SDF" is the ablation on the SDF loss. "No Local" is the ablation on the localized keypoint loss. + +# 4.3. Ablations + +We perform a thorough ablation study demonstrating the importance of our design choices in WIR3D. Figures for each ablation are shown in the supplemental. We report the ablation metrics in Tab. 2. Removing the CLIP intermediate layers ("No CLIP Layers") from our supervision results in the highest reduction in quality (supplemental Fig. 18), followed by removing the stage 1 training ("w/o Stage 1"). Removing this stage results in the optimization biasing towards certain views and creating an overall "flattened" effect of the geometry (supplemental Fig. 17). Removing the localized keypoint loss ("No Local") also results in some reduction in the metrics, but as discussed in Sec. 4.2, perceptual metrics may not be sensitive to the fine visual detail this loss is designed to capture. Removing the SDF loss results in worse geometric coverage, as expected ("No SDF"). + +Keypoint Robustness. We test the robustness of our keypoint loss to noisy keypoints. In Fig. 9, we randomly sample keypoints and compare against the original results ("With Keypoints") and the results without using the localization weighting ("No Keypoints"). When noisy keypoints are given, the results are similar to the results without localization weighting (becomes Eq. (2)). With meaningful keypoints, our method produces strictly improved results. + +![](images/26a790faccd73720aefc75c99727326e575588eb5958e16e8876627028aa622d.jpg) +Figure 9. Keypoint Robustness. Our localized keypoint loss is robust. When keypoints are random ("Random Keypoints"; positions in insets shown), the result is similar to when the localized keypoint loss is not used at all ("No Keypoints"). + +![](images/b5171a6f05135c093c98f78aa235587801b257e701eef8cbba82ca6b08c496d8.jpg) +Figure 10. Deformation application. The curve abstractions make for intuitive control handles for shape deformation. The curves wrap salient features such that deformations are transferred smoothly through a simple L2 distance skinning procedure. + +![](images/4a026604b9794617942c133fa58667a5150432cd7e2bd198af95af4d2e3cc809.jpg) +Figure 11. Keypoint control. Our spatial weighting framework enables user-interactive detail refinement. + +# 4.4. Applications + +We demonstrate two applications enabled by our 3D curve representation. The first is user-interactive feature control. By selecting additional keypoints, new curves can be added and quickly optimized to add detail to any feature of interest. We show in Fig. 11 that keypoints can be used to make structures more explicit (e.g. wheels on the plane) or to add texture detail (e.g. nefertiti headband). The refinement is completed in a few hundred iterations (1 min). + +The curves' adherence to the input geometry makes them intuitive deformation handles. Fig. 10 shows examples of interactive deformations, where Euclidean distance-based skinning weights map deformations from the curves to the shape surface. The curves naturally wrap salient features so that the deformations are smoothly transferred, and relevant parts are easily manipulated. To verify the effectiveness of these curves as deformation handles, we conduct a perceptual study comparing deformations generated by + +WIR3D against ARAP [53], a classic deformation method. We drag a single control point on a WIR3D curve and compare the result to dragging the nearest vertex with ARAP. We ask respondents to "pretend to be 3D modelers in charge of sculpting various shapes", and ask them which deformation result is most desirable from that perspective. + +The respondents $(N = 42)$ found our WIR3D-based deformations to be more desirable $80\%$ of the time. We show one example in the inset, with remaining study details in the supple + +![](images/24a8fceb9f0fdb0cda38fd613c33169e9f26841df4f782bae4ecbc60dcfce593.jpg) + +mental. Our curve-based handles preserve the underlying shape semantics better, such as deforming only the cow's horn while leaving its head intact. + +# 5. Conclusion + +WIR3D is a method for 3D model abstraction into sparse strokes, parameterized as cubic Bézier curves in 3D. We introduce a novel localized keypoint loss, which allows the abstractions to represent fine-grained geometry and texture details. This localized weighting framework enables user control over the local abstraction detail through interactive keypoint selection and detail refinement. Our method is robust across models of arbitrary topology and quality. We encourage the 3D curves to maintain close correspondence to the original surface, which can enable intuitive shape editing applications such as curve-based deformation. + +Limitations. The quality of WIR3D is conditional on how good the keypoints are (Fig. 9). Another limitation is the preprocessing. Fitting a neural SDF, running our automatic keypoint algorithm, and generating the Freestyle renders may take a lot of time depending on the input (2 hours for Nefertiti at 100k faces). + +# References + +[1] Edmond Boyer and Marie-Odile Berger. 3D Surface Reconstruction Using Occluding Contours. International Journal of Computer Vision, 22(3):219-233, 1997. 2 +[2] Pierre Bénard and Aaron Hertzmann. Line Drawings from 3D Models: A Tutorial. Foundations and Trends® in Computer Graphics and Vision, 11(1-2): 1-159, 2019. 2 +[3] John Canny. A Computational Approach to Edge Detection. IEEE Transactions on Pattern Analysis and Machine Intelligence, PAMI-8(6):679-698, 1986. Conference Name: IEEE Transactions on Pattern Analysis and Machine Intelligence. 6, 2 +[4] Li Cao, Yike Xu, Jianwei Guo, and Xiaoping Liu. WireframeNet: A novel method for wireframe generation from point cloud. Computers & Graphics, 115: 226-235, 2023. 3 +[5] Ryan Capouellez, Jiacheng Dai, Aaron Hertzmann, and Denis Zorin. Algebraic Smooth Occluding Contours. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-10, 2023. 1, 2 +[6] Hong Chen, Ying-Qing Xu, Heung-Yeung Shum, Song-Chun Zhu, and Nan-Ning Zheng. Example-based facial sketch generation with non-parametric sampling. In Proceedings Eighth IEEE International Conference on Computer Vision. ICCV 2001, pages 433-438 vol.2, 2001. 3 +[7] Changwoon Choi, Jaeah Lee, Jaesik Park, and Young Min Kim. 3Doodle: Compact Abstraction of Objects with 3D Strokes. ACM Transactions on Graphics (TOG), 43(4):1-13, 2024. 3, 4, 6 +[8] Hyelim Choi, Minji Lee, Jiseock Kang, and Dongjun Lee. Online 3D Edge Reconstruction of Wiry Structures From Monocular Image Sequences. IEEE Robotics and Automation Letters, 8(11):7479-7486, 2023. Conference Name: IEEE Robotics and Automation Letters. 3 +[9] Keenan Crane, Ulrich Pinkall, and Peter Schröder. Robust fairing via conformal curvature flow. ACM Transactions on Graphics (TOG), 32(4):1-10, 2013. 5 +[10] Doug DeCarlo, Adam Finkelstein, Szymon Rusinkiewicz, and Anthony Santella. Suggestive contours for conveying shape. ACM Trans. Graph., 22(3):848-855, 2003. 2 +[11] Doug DeCarlo, Adam Finkelstein, Szymon Rusinkiewicz, and Anthony Santella. Suggestive Contours for Conveying Shape. In *Seminal Graphics Papers: Pushing the Boundaries*, Volume 2, pages 401-408. ACM New York, NY, USA, 2023. 1 +[12] Boyang Deng, Kyle Genova, Soroosh Yazdani, Sofien Bouaziz, Geoffrey Hinton, and Andrea Tagliasacchi. + +Cvxnet: Learnable Convex Decomposition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (CVPR), pages 31-44, 2020. 2 +[13] Kevin Frans, Lisa Soros, and Olaf Witkowski. CLIP-Draw: Exploring Text-to-Drawing Synthesis through Language-Image Encoders. Advances in Neural Information Processing Systems, 35:5207-5218, 2022. 3 +[14] Ran Gal, Olga Sorkine, Niloy Mitra, and Daniel Cohen-Or. iWiRES: An analyze-and-edit approach to shape manipulation. ACM Transactions on Graphics (proceedings of ACM SIGGRAPH), 28(3):33:1-33:10, 2009. 3 +[15] Rinon Gal, Yael Vinker, Yuval Alaluf, Amit Bermano, Daniel Cohen-Or, Ariel Shamir, and Gal Chechik. Breathing Life Into Sketches Using Text-to-Video Priors. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4325–4336, Seattle, WA, USA, 2024. IEEE. 3 +[16] Huy Ha and Shuran Song. Semantic Abstraction: Open-World 3D Scene Understanding from 2D Vision-Language Models, 2022. arXiv:2207.11514 [cs]. 2 +[17] James W. Hennessey, Han Liu, Holger Winnemöller, Mira Dontcheva, and Niloy J. Mitra. How2sketch: Generating easy-to-follow tutorials for sketching 3d objects. Symposium on Interactive 3D Graphics and Games, 2017. 3 +[18] Pablo Hernandez-Camara and Jorge Vila-Tomas. MEASURING HUMAN-CLIP ALIGNMENT AT DIFFERENT ABSTRACTION LEVELS. 2024. 2 +[19] Amir Hertz, Rana Hanocka, Raja Giryes, and Daniel Cohen-Or. PointGMM: a Neural GMM Network for Point Clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12054–12063, 2020. 2 +[20] Amir Hertz, Or Perel, Raja Giryes, Olga Sorkine-Hornung, and Daniel Cohen-Or. SPAGHETTI: Editing Implicit Shapes Through Part Aware Generation. ACM Transactions on Graphics (TOG), 41(4):1-20, 2022. 2 +[21] Aaron Hertzmann. New insights in smooth occluding contours for nonphotorealistic rendering. IEEE Computer Graphics and Applications, 44(1):76-85, 2024. 1, 2 +[22] Kai-Wen Hsiao, Jia-Bin Huang, and Hung-Kuo Chu. Multi-view Wire Art. ACM Transactions on Graphics (TOG), 37(6):1-11, 2018. 3 +[23] Felix Hahnlein, Changjian Li, Niloy J. Mitra, and Adrien Bousseau. CAD2Sketch: Generating Concept Sketches from CAD Sequences. ACM Transactions on Graphics, 41(6):1-18, 2022. 3 + +[24] R Kenny Jones, Aalia Habib, and Daniel Ritchie. SHRED: 3D Shape Region Decomposition with Learned Local Operations. ACM Transactions on Graphics (TOG), 41(6):1-11, 2022. 2 +[25] Jan J Koenderink. What does the occluding contour tell us about solid shape? Perception, 13(3):321-330, 1984. 2 +[26] Lei Li, Songyou Peng, Zehao Yu, Shaohui Liu, Remi Pautrat, Xiaochuan Yin, and Marc Pollefeys. 3D Neural Edge Reconstruction. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 21219-21229, Seattle, WA, USA, 2024. IEEE. 2 +[27] Mengtian Li, Zhe Lin, Radomir M'ech, , Ersin Yumer, and Deva Ramanan. Photo-sketching: Inferring contour drawings from images. WACV, 2019. 3 +[28] Tzu-Mao Li, Michal Lukáč, Gharbi Michael, and Jonathan Ragan-Kelley. Differentiable vector graphics rasterization for editing and learning. ACM Trans. Graph. (Proc. SIGGRAPH Asia), 39(6):193:1-193:15, 2020. 3 +[29] Chenxi Liu, Pierre Bénard, Aaron Hertzmann, and Shayan Hoshyari. ConTesse: Accurate Occluding Contours for Subdivision Surfaces. ACM Transactions on Graphics, 42(1):1-16, 2023. 1 +[30] Difan Liu, Mohamed Nabail, Aaron Hertzmann, and Evangelos Kalogerakis. Neural Contours: Learning to Draw Lines from 3D Shapes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5428-5436, 2020. 2 +[31] Difan Liu, Matthew Fisher, Aaron Hertzmann, and Evangelos Kalogerakis. Neural Strokes: Stylized Line Drawing of 3d Shapes. In Proceedings of the IEEE/CVF International Conference on Computer Vision (CVPR), pages 14204-14213, 2021. 1, 2 +[32] Lingjie Liu, Duygu Ceylan, Cheng Lin, Wenping Wang, and Niloy J. Mitra. Image-based reconstruction of wire art. ACM SIGGRAPH 2017, 2017. 3 +[33] Lingjie Liu, Nenglun Chen, Duygu Ceylan, Christian Theobalt, Wenping Wang, and Niloy J. Mitra. Curvefusion: Reconstructing thin structures from rgbd sequences. ACM Trans. Graph., 37(6):218:1-218:12, 2018. 3 +[34] Weixiao Liu, Yuwei Wu, Sipu Ruan, and Gregory S Chirikjian. Marching-Primitives: Shape Abstraction from Signed Distance Function. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8771–8780, 2023. 2 +[35] Stuart Lloyd. Least squares quantization in pmc. IEEE transactions on information theory, 28(2):129-137, 1982. 5, 1 +[36] Andrea Maggiordomo, Federico Ponchio, Paolo Cignoni, and Marco Tarini. Real-World Textured + +Things: A repository of textured models generated with modern photo-reconstruction tools. Computer Aided Geometric Design, 83:101943, 2020. 6 +[37] David Marr. Analysis of Occluding Contour. Proceedings of the Royal Society of London. Series B. Biological Sciences, 197(1129):441-475, 1977. 2 +[38] Ravish Mehra, Qingnan Zhou, Jeremy Long, Alla Sheffer, Amy Gooch, and Niloy J Mitra. Abstraction of man-made shapes. In ACM SIGGRAPH Asia 2009 papers, pages 1-10. ACM New York, NY, USA, 2009. 3 +[39] Niloy J Mitra and Mark Pauly. Shadow Art. ACM Transactions on Graphics, 28(5):156-1, 2009. 3 +[40] Alessandro Muntoni, Marco Livesu, Riccardo Scateni, Alla Sheffer, and Daniele Panozzo. Axis-Aligned Height-Field Block Decomposition of 3D Shapes. ACM Transactions on Graphics (TOG), 37(5):1-15, 2018. 2 +[41] Xiaqing Pan, Nicholas Charron, Yongqian Yang, Scott Peters, Thomas Whelan, Chen Kong, Omkar Parkhi, Richard Newcombe, and Carl Yuheng Ren. Aria digital twin: A new benchmark dataset for egocentric 3d machine perception, 2023. 5, 4 +[42] Karran Pandey, Fanny Chevalier, and Karan Singh. Juxtaposition: interactive visual summarization for exploratory shape design. ACM Trans. Graph., 42(4): 52:1-52:14, 2023. 3 +[43] Wamiq Para, Shariq Bhat, Paul Guerrero, Tom Kelly, Niloy Mitra, Leonidas J Guibas, and Peter Wonka. SketchGen: Generating Constrained CAD Sketches. In Advances in Neural Information Processing Systems, pages 5077-5088. Curran Associates, Inc., 2021. 3 +[44] Despoina Paschalidou, Angelos Katharopoulos, Andreas Geiger, and Sanja Fidler. Neural Parts: Learning Expressive 3D Shape Abstractions with Invertible Neural Networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3204-3215, 2021. 2 +[45] Ofek Pearl, Itai Lang, Yuhua Hu, Raymond A Yeh, and Rana Hanocka. GeoCode: Interpretable Shape Programs. arXiv preprint arXiv:2212.11715, 2022. 2 +[46] Yonggang Qi, Guoyao Su, Pinaki Nath Chowdhury, Mingkang Li, and Yi-Zhe Song. SketchLattice: Latticed Representation for Sketch Manipulation. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 933-941, Montreal, QC, Canada, 2021. IEEE. 3 +[47] Zhiyu Qu, Lan Yang, Honggang Zhang, Tao Xiang, Kaiyue Pang, and Yi-Zhe Song. Wired Perspectives: Multi-View Wire Art Embraces Generative AI. In Proceedings of the IEEE/CVF Conference on Com + +puter Vision and Pattern Recognition, pages 6149-6158, 2024.3 +[48] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning Transferable Visual Models from Natural Language supervision. In International Conference on Machine Learning (ICML), pages 8748-8763. PMLR, 2021. 2, 4 +[49] W Brent Seales and Charles R Dyer. Viewpoint from Occluding Contour. CVGIP: Image Understanding, 55(2):198-211, 1992. 2 +[50] Tianjia Shao, Wilmot Li, Kun Zhou, Weiwei Xu, Baining Guo, and Niloy J. Mitra. Interpreting concept sketches. ACM Transactions on Graphics, 32(4), 2013. 3 +[51] Gil Shomron and Uri Weiser. Spatial correlation and value prediction in convolutional neural networks. IEEE Comput. Archit. Lett., 18(1):10-13, 2019. 4 +[52] Jifei Song, Kaiyue Pang, Yi-Zhe Song, Tao Xiang, and Timothy M. Hospedales. Learning to Sketch with Shortcut Cycle Consistency. In 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 801-810, Salt Lake City, UT, 2018. IEEE. 3 +[53] Olga Sorkine and Marc Alexa. As-rigid-as-possible surface modeling. In Proceedings of the fifth Eurographics symposium on Geometry processing, pages 109–116, Goslar, DEU, 2007. Eurographics Association. 8, 6 +[54] Anjana Susarla, Ram Gopal, Jason Bennett Thatcher, and Suprateek Sarker. The Janus Effect of Generative AI: Charting the Path for Responsible Conduct of Scholarly Activities in Information Systems. Information Systems Research, 34(2):399-408, 2023. Publisher: INFORMS. 1 +[55] Jean-Marc Thiery, Émilie Guy, and Tamy Boubekeur. Sphere-Meshes: shape approximation using spherical quadric error metrics. ACM Transactions on Graphics, 32(6):1-12, 2013. 2 +[56] Yingtao Tian and David Ha. Modern Evolution Strategies for Creativity: Fitting Concrete Images and Abstract Concepts. In Artificial Intelligence in Music, Sound, Art and Design: 11th International Conference, EvoMUSART 2022, Held as Part of EvoStar 2022, Madrid, Spain, April 20-22, 2022, Proceedings, pages 275-291, Berlin, Heidelberg, 2022. Springer-Verlag. 3 +[57] Kenji Tojo, Ariel Shamir, Bernd Bickel, and Nobuyuki Umetani. Fabricable 3D Wire Art. In *Special Interest Group on Computer Graphics and Interactive Techniques Conference Papers* '24, pages 1-11, Denver CO USA, 2024. ACM. 3 + +[58] Shengbang Tong, Zhuang Liu, Yuexiang Zhai, Yi Ma, Yann LeCun, and Saining Xie. Eyes Wide Shut? Exploring the Visual Shortcomings of Multimodal LLMs, 2024. arXiv:2401.06209 [cs]. 6 +[59] Shubham Tulsiani, Hao Su, Leonidas J Guibas, Alexei A Efros, and Jitendra Malik. Learning Shape Abstractions by Assembling Volumetric Primitives. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2635-2643, 2017. 2 +[60] Oliver van Kaick, Andrea Tagliasacchi, Oana Sidi, Hao Zhang, Daniel Cohen-Or, Lior Wolf, and Ghassan Hamarneh. Prior knowledge for part correspondence. Computer Graphics Forum, 30(2):553-562, 2011. 5 +[61] Yael Vinker, Ehsan Pajouheshgar, Jessica Y Bo, Roman Christian Bachmann, Amit Haim Bermano, Daniel Cohen-Or, Amir Zamir, and Ariel Shamir. CLIPasso: Semantically-Aware Object Sketching. ACM Transactions on Graphics (TOG), 41(4):1-11, 2022. 3, 4, 2 +[62] Yael Vinker, Yuval Alaluf, Daniel Cohen-Or, and Ariel Shamir. CLIPascene: Scene Sketching with Different Types and Levels of Abstraction. In 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 4123–4133, Paris, France, 2023. IEEE. 2, 3 +[63] Jianyi Wang, Kelvin CK Chan, and Chen Change Loy. Exploring clip for assessing the look and feel of images. In AAAI, 2023. 6 +[64] Peng Wang, Lingjie Liu, Nenglun Chen, Hung-Kuo Chu, Christian Theobalt, and Wenping Wang. Vid2Curve: simultaneous camera motion estimation and thin structure reconstruction from an RGB video. ACM Trans. Graph., 39(4):132:132:1-132:132:12, 2020. 3 +[65] Wenxuan Wang, Quan Sun, Fan Zhang, Yepeng Tang, Jing Liu, and Xinlong Wang. Diffusion Feedback Helps CLIP See Better, 2024. arXiv:2407.20171 [cs]. 6 +[66] Thomas Wimmer, Peter Wonka, and Maks Ovsjanikov. Back to 3d: Few-shot 3d keypoint detection with back-projected 2d features. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 5, 1 +[67] Yuwei Wu, Weixiao Liu, Sipu Ruan, and Gregory S Chirikjian. Primitive-based Shape Abstraction via Nonparametric Bayesian Inference. In European Conference on Computer Vision (ECCV), pages 479-495. Springer, 2022. 2 +[68] Zhijin Yang, Pengfei Xu, Hongbo Fu, and Hui Huang. WireRoom: model-guided explorative design of abstract wire art. ACM Transactions on Graphics, 40(4): 1-13, 2021. 3 + +[69] Yunfan Ye, Renjiao Yi, Zhirui Gao, Chenyang Zhu, Zhiping Cai, and Kai Xu. NEF: Neural Edge Fields for 3D Parametric Curve Reconstruction from Multi-view Images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8486-8495, 2023. 2, 6 +[70] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 4, 6 +[71] Yibo Zhang, Lihong Wang, Changqing Zou, Tieru Wu, and Rui Ma. Diff3DS: Generating View-Consistent 3D Sketch via Differentiable Curve Rendering. arXiv preprint arXiv:2405.15305, 2024. 3 \ No newline at end of file diff --git a/ICCV/2025/WIR3D_ Visually-Informed and Geometry-Aware 3D Shape Abstraction/images.zip b/ICCV/2025/WIR3D_ Visually-Informed and Geometry-Aware 3D Shape Abstraction/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..a0fc8733b45005bb942fccf425acad3673f8a784 --- /dev/null +++ b/ICCV/2025/WIR3D_ Visually-Informed and Geometry-Aware 3D Shape Abstraction/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b189dc864167b504a9a319f22695bf049f9e24001093c8676d6199d915cf40fc +size 735618 diff --git a/ICCV/2025/WIR3D_ Visually-Informed and Geometry-Aware 3D Shape Abstraction/layout.json b/ICCV/2025/WIR3D_ Visually-Informed and Geometry-Aware 3D Shape Abstraction/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..288b2bb646a5d4c1f97c968de1799582f45ef902 --- /dev/null +++ b/ICCV/2025/WIR3D_ Visually-Informed and Geometry-Aware 3D Shape Abstraction/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:67fe6f5f9be97e29b1338e892353037ceb9d5cb614929f553b5ad104656e0604 +size 369430 diff --git a/ICCV/2025/WSI-LLaVA_ A Multimodal Large Language Model for Whole Slide Image/5c32ce28-a3b0-4e1d-be20-7568b907d88a_content_list.json b/ICCV/2025/WSI-LLaVA_ A Multimodal Large Language Model for Whole Slide Image/5c32ce28-a3b0-4e1d-be20-7568b907d88a_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..e1a201a3783e5394621c44c585941575643c4550 --- /dev/null +++ b/ICCV/2025/WSI-LLaVA_ A Multimodal Large Language Model for Whole Slide Image/5c32ce28-a3b0-4e1d-be20-7568b907d88a_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:84f39e2469897ceea016e00791c1fdf0e670ddb74baa872daf19fc790da0fe4f +size 69251 diff --git a/ICCV/2025/WSI-LLaVA_ A Multimodal Large Language Model for Whole Slide Image/5c32ce28-a3b0-4e1d-be20-7568b907d88a_model.json b/ICCV/2025/WSI-LLaVA_ A Multimodal Large Language Model for Whole Slide Image/5c32ce28-a3b0-4e1d-be20-7568b907d88a_model.json new file mode 100644 index 0000000000000000000000000000000000000000..c3fc6b6034afb6644b60967fa1e055d02a521458 --- /dev/null +++ b/ICCV/2025/WSI-LLaVA_ A Multimodal Large Language Model for Whole Slide Image/5c32ce28-a3b0-4e1d-be20-7568b907d88a_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ae070a2d490efb910cdb6d20b5569e0bfd73192cbe2c422bc968184ddfb7a29e +size 84498 diff --git a/ICCV/2025/WSI-LLaVA_ A Multimodal Large Language Model for Whole Slide Image/5c32ce28-a3b0-4e1d-be20-7568b907d88a_origin.pdf b/ICCV/2025/WSI-LLaVA_ A Multimodal Large Language Model for Whole Slide Image/5c32ce28-a3b0-4e1d-be20-7568b907d88a_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..70ec466237886212cdb247bde552b49848b937d0 --- /dev/null +++ b/ICCV/2025/WSI-LLaVA_ A Multimodal Large Language Model for Whole Slide Image/5c32ce28-a3b0-4e1d-be20-7568b907d88a_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4ded328b03c1d864223885cacaceaaeb4fe2afcc823c0198b780a2f1fbbb72f5 +size 9857427 diff --git a/ICCV/2025/WSI-LLaVA_ A Multimodal Large Language Model for Whole Slide Image/full.md b/ICCV/2025/WSI-LLaVA_ A Multimodal Large Language Model for Whole Slide Image/full.md new file mode 100644 index 0000000000000000000000000000000000000000..e1c499d66eed79e172d28fee2f64722e61a58841 --- /dev/null +++ b/ICCV/2025/WSI-LLaVA_ A Multimodal Large Language Model for Whole Slide Image/full.md @@ -0,0 +1,273 @@ +![](images/120232417015dbb3ee372a04c65fc6e7d6434923b0295db5a45e1fb321e884f2.jpg) + +# WSI-LLaVA: A Multimodal Large Language Model for Whole Slide Image + +Yuci Liang $^{1*}$ , Xinheng Lyu $^{1,2*}$ , Wenting Chen $^{3\dagger}$ , Meidan Ding $^{1}$ , Jipeng Zhang $^{5}$ , + +Xiangjian He $^{2}$ , Song Wu $^{1}$ , Xiaohan Xing $^{4}$ , Sen Yang $^{4}$ , Xiyue Wang $^{4}$ , Linlin Shen $^{1\dagger}$ + +1Shenzhen University, 2University of Nottingham Ningbo China, 3City University of Hong Kong, + +$^{4}$ Stanford University, $^{5}$ Hong Kong University of Science and Technology + +# Abstract + +Recent advances in computational pathology have introduced whole slide image (WSI)-level multimodal large language models (MLLMs) for automated pathological analysis. However, current WSI-level MLLMs face two critical challenges: limited explainability in their decision-making process and insufficient attention to morphological features crucial for accurate diagnosis. To address these challenges, we first introduce WSI-Bench, a large-scale morphology-aware benchmark containing 180k VQA pairs from 9,850 WSIs across 30 cancer types, specifically designed to evaluate MLLMs' understanding of morphological characteristics crucial for accurate diagnosis. To the best of our knowledge, WSI-Bench presents the first benchmarking systematically evaluate morphological understanding capabilities in WSI analysis. To enhance the model explainability, we present WSI-LLaVA, an MLLM framework for gigapixel WSI understanding with a three-stage training strategy, which can provide detailed morphological findings to explain its final answer. For more precise model assessment in pathological contexts, we develop two specialized WSI metrics: WSI-Precision and WSI-Relevance, focusing on clinical accuracy. Extensive evaluation on WSI-Bench reveals both the capabilities and limitations of current WSI MLLMs in morphological analysis and various pathology tasks, while demonstrating WSI-LLaVA's superior performance across all capabilities on both internal and external datasets. Source code and data are released. + +# 1. Introduction + +Recent computational pathology advances have introduced both patch-level [9, 18, 25, 30, 31] and whole slide image (WSI)-level [3-5, 12, 19, 28] multimodal large language models (MLLMs). While patch-level models analyze lo + +![](images/ccc7dd114f7bf0d1f187296ec79a1892f8e42f349978ec7a7805c1ab0087a508.jpg) + +what is the histologic grade of this tumor? + +SlideChat: A. Grade II + +WSI-VQA: Grade 2 + +![](images/e33ae5ace5736d3abd3fa866699fb4330f73f8a8f34ed7cccc63b7b051085255.jpg) +WSI + +WSI-LLaVA: The tumor is graded as SBR (Scarff-Bloom-Richardson) Grade 2, with an overall score of 6 out of 9. This grading considers the moderate differentiation of duct-like structures (differentiation score of 2), the pronounced nuclear pleomorphism (score of 3), and the low mitotic count (score of 1). The combination of these factors indicates a moderately differentiated tumor, aligning with SBR Grade 2, which suggests an intermediate prognosis in terms of aggressiveness and progression. + +![](images/1264aa65aa1da634901a53f738b83bdb03fa4c22163b15799981cbae661b8674.jpg) +Figure 1. Comparison of WSI MLLMs. Existing frameworks offer only brief answers without explanations, while WSI-LLaVA provides comprehensive grading with detailed morphological reasoning, showing the explainability gap in WSI analysis. + +calized tissue regions, WSI-level MLLMs leverage comprehensive contextual information from gigapixel images for diagnostics. WSI-level analysis is particularly valuable since histopathological diagnosis requires holistic analysis where pathologists integrate findings across extensive tissue areas [10, 27, 32]. The WSI-level MLLMs fall into two categories: transformer-based models [3, 4, 12] for pathology report generation and simple visual question answering (VQA), and LLM-based model [5] for complex VQA tasks. Despite their promising performance, these models still face significant challenges. + +The first challenge in current WSI-level MLLMs lies in their explainability: these models typically struggle to provide explanations for their final answers. Existing WSI-level models tend to offer short answers without supporting + +explanations. Some approaches [5] are trained exclusively on closed-ended visual question answering (VQA) pairs, conditioning models to produce only final results without justification. Others [4] are trained on open-ended VQA pairs that primarily elicit keyword responses, as illustrated in Fig. 1. Unlike human pathologists who articulate detailed observations and reasoning processes, these minimal responses prevent clinicians from confidently trusting or thoroughly validating model outputs [16, 18, 20]. Therefore, enhancing explainability in WSI-level MLLMs is crucial for improving their clinical utility and adoption. + +Another significant challenge with current pathology MLLMs is their tendency to map pathology images directly to diagnostic outcomes, overlooking crucial morphological features necessary for accurate diagnosis [18]. Pathologists depend on detailed observation of these morphological characteristics when interpreting pathology images. For example, distinguishing between invasive ductal carcinoma (IDC) and invasive lobular carcinoma (ILC) requires recognizing subtle differences in tumor cell arrangements [33]. While some pathology MLLMs incorporate questions about morphological features, many studies [13, 18, 25] are limited to patch-level analysis—preventing holistic assessment across the entire slide—or are trained on closed-ended VQA data that constrains their ability to produce rich, descriptive responses [4, 5]. As a result, they cannot analyze morphological features comprehensively and systematically with the detailed, nuanced descriptions characteristic of expert pathologists. Moreover, morphological observation capabilities remain unevaluated in current WSI multi-modal benchmarks [23], which primarily assess WSI MLLMs on common pathology tasks such as diagnosis, prognosis, and report generation. Therefore, integrating detailed morphological observations into WSI MLLMs is essential to better emulate human pathologists and properly evaluate their morphological understanding capabilities. + +WSI-Bench: To address these challenges, we introduce WSI-Bench, a morphology-aware gigapixel WSI benchmark that systematically evaluates WSI-level MLLMs on both morphological understanding and clinical pathology tasks. The benchmark assesses 4 core capabilities through 11 distinct tasks, comprising over 180k VQA pairs derived from 9,850 WSIs spanning 30 cancer types from 8,368 patients. As illustrated in Fig. 2, the core capabilities include morphological analysis, diagnosis, treatment planning, and report generation, comprehensively covering 11 specific clinical tasks in pathology. All open-ended tasks incorporate corresponding explanations, particularly for diagnosis tasks where models must provide both conclusions and supporting morphological descriptions. This design ensures interpretability by making the reasoning process transparent, enabling pathologists to understand and validate the model's decision-making pathway. For morpholog- + +ical analysis, WSI-Bench includes four hierarchical morphology tasks that progress from global to local perspectives, offering a comprehensive evaluation of morphological understanding capabilities. To the best of our knowledge, this represents the first and largest morphology-aware WSI benchmark with comprehensive clinical tasks specifically designed for WSI-level MLLM training and evaluation. + +WSI-LLaVA: To enhance explainability in pathology tasks, we introduce a novel WSI-level Large Language and Vision Assistant framework (WSI-LLaVA) that leverages detailed morphological observations to explain diagnostic reasoning. Trained on the WSI-Bench training set, WSI-LLaVA can first provide morphological findings related to the query before presenting the final answer, significantly enhancing model explainability. To mitigate the cross-modal gap between WSI and text, WSI-LLaVA employs a three-stage training approach: WSI-text alignment, feature space alignment, and task-specific instruction tuning. Our framework integrates both patch-level and slide-level encoders to capture local details and global context, providing a comprehensive representation of gigapixel WSIs for advanced pathological analysis and reasoning. + +WSI Metrics: Although natural language understanding (NLU) metrics are commonly used to evaluate medical language tasks, they do not accurately assess performance due to complex and often similar terminology in pathology.[11, 17]. To address these limitations, we introduce two specialized WSI metrics: WSI-Precision, which verifies the accuracy of each claim derived from the ground truth against the model's answers, and WSI-Relevance, which assesses the alignment of each claim in the model's responses with the ground truth to ensure their relevance. These metrics are specifically designed to evaluate the quality of morphological reasoning and explanatory content in model outputs, providing a more nuanced assessment and explanation than traditional metrics allow. + +Our contributions are summarized as follows: + +- We introduce WSI-Bench, the first large-scale morphology-aware benchmark for gigapixel WSI understanding and evaluation, encompassing 180k VQA pairs from 9,850 WSIs across 30 cancer types. This benchmark uniquely evaluates the morphological understanding capability and incorporates morphological descriptions as supporting evidence for each task. +- To enhance model explainability, we propose WSI-LLaVA, a novel framework for gigapixel WSI analysis and explanation that integrates patch-level and slide-level features through a three-stage training approach, which leverages detailed morphological observations to explain diagnostic reasoning. +- We develop WSI-specific evaluation metrics that provide a more accurate assessment and explanation of model performance in pathological contexts, which address the lim + +ities of traditional NLU metrics by verifying claim accuracy and response relevance. + +- We conduct a comprehensive evaluation of current WSI MLLMs and other MLLMs, shedding light on their strengths and weaknesses in WSI analysis, and prove the effectiveness of our WSI-LLaVA. + +# 2. Related Work + +Pathology VQA Benchmarks. Prior pathology VQA and report generation benchmarks have advanced the field but exhibit notable limitations. Patch-level benchmarks like PathVQA [13] and Quilt-Instruct [25] contribute valuable educational content but remain confined to patch analysis with multiple-choice formats, missing the holistic view required clinically. WSI-VQA [4] attempts slide-level analysis but uses simplistic formats with quality issues affecting clinical relevance. For report generation, WSICaption [3] pioneers WSI-text pair curation from TCGA data, which is later refined by HistGen[12]. While SlideBench-VQA [5] shows progress, it covers only 10 cancer types and lacks complex dialogues for thorough evaluation. Crucially, these benchmarks have not systematically evaluated morphological understanding capabilities essential for diagnosis. To address this issue, we introduce WSI-Bench, which curates WSI-level VQA pairs focused on morphological analysis, diagnostic reasoning, treatment planning, and report generation—enabling comprehensive assessment. + +WSI-Level MLLMs. Recent advances in WSI-level multi-modal models demonstrate efforts to tackle the gigapixel nature of whole slide images. Transformer-based architectures like WSI-Caption [3], HistGen [12], WSI-VQA [4], and PRISM [26] have made progress in report generation and basic slide understanding, but exhibit limited interaction capabilities beyond their designed tasks. Most existing models are limited to handling multiple-choice questions or producing extremely brief responses of only two or three words. SlideChat [5] represents the first attempt to integrate slide-level features with large language models, but lacks sufficient morphological reasoning capabilities and comprehensive output generation. Furthermore, existing models have been predominantly trained and evaluated in limited types of cancer, restricting their clinical applicability to various pathological conditions. WSI-LLaVA framework addresses these limitations through a systematic three-stage training approach that enables comprehensive gigapixel WSI analysis with enhanced explainability. + +# 3. WSI-Bench + +In clinical practice, pathologists rely on morphological features for diagnosis, but current WSI MLLMs overlook these crucial details, compromising precision [2]. Thus, we intro + +duce a morphology-aware WSI-Bench benchmark for gigapixel WSI evaluation to include morphological observations for a complete understanding of WSI. It can evaluate the WSI MLLMs across 4 pathological capabilities with 11 tasks for WSI analysis. As shown in Fig. 2, the pathological capabilities include morphological descriptions, diagnosis, and treatment planning, and report generation. + +# 3.1. Pathological Capabilities + +Morphological Analysis. To assess the morphological analysis capability of WSI MLLMs, the four hierarchical pathological tasks are introduced for a comprehensive evaluation, ranging from global to local perspectives. These tasks include global morphology description, key diagnostic description, regional structure description, and specific feature description. Each category is designed to capture distinct pathological aspects, enhancing the dataset's comprehensiveness while providing clinically relevant insights. + +1) Global Morphology Description evaluates the ability to provide a comprehensive overview of WSIs, establishing a context for global analysis. We construct the open-ended VQA pairs by a question with fixed content and a global morphological description. The question asks for the overall description of the WSI. The global morphological description is extracted from our WSI-report pairs, which should capture the general tissue architecture and overall tumor presentation, ensuring a broad understanding of the slide's features. + +2) Key Diagnostic Description focuses on the evaluation of crucial morphological features vital for accurate diagnosis, linking specific observations to their clinical implications. Each open-ended VQA pair includes a question about key morphological features for diagnosis and its answer. We construct the answer by extracting content related to the cancer type's diagnosis from the WSI report. + +3) Regional Structure Description aims to evaluate the ability of broad observations in specific regions, such as tumor infiltration,tissue architecture, and cellular morphology. We construct both open-ended and close-ended (multiple-choice questions, MCQs) VQA pairs, which include a question about a specific region and its answer. The answer is obtained by extracting relevant descriptions for the specific region from the WSI report. + +4) Specific Feature Description evaluates detailed morphological examinations of specific aspects (vascular invasion, gland formation patterns, cellular pleomorphism, etc.). We create both open-ended and close-ended (MCQs) VQA pairs about specific features, with answers indicating their presence in WSI reports. + +Diagnosis. To evaluate diagnostic capabilities, we examine four fundamental tasks: histological typing, grading, molecular subtyping, and staging. For interpretable and accurate diagnosis, each result includes both the answer and + +![](images/bc089ec5d2cb9f5a08ec9d155c5d5d0ce448efa8e81c9b59c1e976a492d9e267.jpg) +Pathological Tasks +Figure 2. Overview of WSI-Bench, a morphology-aware benchmark for gigapixel WSI evaluation across 4 pathological capabilities and 11 tasks, which encompasses about 180k VQA pairs from 9,850 WSIs across 30 cancer types, sourced from 8,368 patients. + +# WSI-Bench + +explanatory morphological analysis. + +1) Histological Typing evaluates the ability of histological type classification. It includes both open-ended and close-ended (MCQs) VQA pairs with a question and its answer, where the answer is obtained from the WSI report. +2) Grading assesses the ability to classify the grade with both open-ended and close-ended (MCQs) VQA pairs, where each VQA pair includes a question about grading and its answer obtained from the WSI report. +3) Molecular Subtyping evaluates identification of specific subtypes based on histological typing and immunohistochemical data. We create open-ended VQA pairs with questions containing histological typing and immunohistochemical data, with answers from original data labels. +4) Staging assesses prediction of Tumor-Node-Metastasis (TNM) [24] labels based on clinical scenarios. Both open-ended and close-ended (MCQs) VQA pairs present clinical information translated from TNM classifications, requiring inference of TNM classification from the provided context. + +Treatment Planing. In the context of treatment planning, WSI-Bench includes three pathological tasks, including prognosis, treatment recommendations, and report gen + +eration. To imitate real-world settings, we include both WSI and its comprehensive information like diagnostic and immunohistochemical results for treatment planning. + +1) Treatment recommendations evaluate treatment suggestion capabilities. We create open-ended VQA pairs with questions including diagnostic results (histological typing, grading, staging) and immunohistochemical data. GPT-4o generates initial recommendations based on WSI reports, refined by clinical experts. +2) Prognosis evaluates prediction of disease outcomes (good or bad). We create Yes/No VQA pairs with questions containing diagnostic and immunohistochemical data. GPT-4o generates initial prognosis results, refined by clinical experts. + +Report Generation. Report generation assesses the ability to create reports from input WSIs. Ground-truth reports are extracted from original pathology reports using GPT-4o. We exclude content not observable in WSIs, retaining only morphological descriptions and diagnostic results (histological typing and grading). + +![](images/ec0d9c06bb423018f27180229ad0c177c61db71b8d0fe0005b97fb9a61f7f7cc.jpg) +Figure 3. Two-step process for constructing WSI-Bench: Step 1 removes gross descriptions and immunohistochemical results from pathology reports, retaining morphological descriptions and diagnostic results, with enriched descriptions generated via reverse-engineering; Step 2 builds VQA pairs from the refined reports to support pathological tasks. + +# 3.2. Data Construction + +We curate WSI-Bench from The Cancer Genome Atlas (TCGA)[29], including 10,756 WSIs across 30 cancer types. After excluding cases with multiple reports per WSI, the final WSI-Bench contains 9,850 WSIs paired with their corresponding pathology reports. The specific distribution across different cancer types can be found in Appendix Section 3. In Fig. 3, our data preparation involved two main steps: (1) using GPT-4o to extract and clean text from original reports, removing gross descriptions and immunohistochemical results while retaining morphological descriptions and diagnostic results, and (2) implementing a reverse-engineer scheme to enrich morphological descriptions based on diagnostic labels and established standards like the Nottingham grading system [7]. All generated content was reviewed by clinical experts. Based on these finalized WSI reports, we construct the VQA pairs required for each pathological task, as detailed in Section 3.1. + +Data Statistics and Split. In Table 1, WSI-Bench comprises 179,569 VQA pairs. The dataset includes 175,450 VQA pairs (122,133 open-ended and 53,317 closed-ended) in the training set with 9,642 WSIs, plus 4,119 VQA pairs (2,838 open-ended and 1,281 closed-ended) in the testing set with 208 WSIs. Details on construction process, quality control measures, hallucination prevention methods, and data distribution are in Appendix Section 3. + +# 3.3. WSI Metrics + +Although NLU metrics are widely used to evaluate open-ended language tasks, they cannot accurately assess performance in pathology due to its complex and often similar terminology, and the scores themselves lack interpretability [14, 22]. Thus, we introduce two WSI metrics for open-ended questions: WSI-Precision and WSI-Relevance. The detailed evaluation workflow for these metrics can be found in Appendix Section Fig. 9 + +WSI-Precision measures the accuracy of model responses + +
CapabilitiesTasksVQA Pair Count
AnalysisGlobal Morphology Description9,850
Key Diagnostic Description9,850
Regional Structure Description55,181
Specific Feature Description48,170
DiagnosisHistological Typing19,697
Grading15,335
Molecular Subtyping977
Staging6,743
Treatment PlanningTreatment Recommendations1,958
Prognosis1,958
Report Generation9,850
Total179,569 (180k)
+ +Table 1. Statistics Analysis on the WSI-Bench. + +by comparing them directly to claims derived from the ground-truth (GT) response. It involves three main steps. 1) Claims Extraction: We extract distinct claims from the GT response using GPT-4o to break down complex medical information into assessable units. 2) Individual Score Calculation and Interpretation: We compare each extracted claim with the generated response using our scoring criterion. We assign Score 1 for complete accuracy with perfect alignment to established facts; Score 0.7 for mostly correct information with minor discrepancies that do not alter clinical significance; Score 0.3 for significant errors in critical content areas affecting the core message; and Score 0 for completely incorrect or unrelated information. Each score is accompanied by a corresponding explanation that justifies the reasoning behind the score. 3) Final Score Calculation: We calculate the overall WSI-Precision score by averaging all individual claim scores. + +WSI-Relevance measures the alignment of each claim in the model's responses with the ground truth to ensure their relevance. This metric uses a similar three-step process. The detailed step-by-step procedures, prompt designs, and + +![](images/fd0f1cb16ba9682425d10682a49aae87396f8c23b816d2914981037e638a4c79.jpg) +a +Figure 4. Overview of our gigapixel WSI-LLaVA framework with a three-stage training approach for cross-modal alignment and pathological tasks learning. + +implementation specifics for both WSI-Precision and WSI-Relevance metrics are provided in Appendix Section 4. + +Reliability Validation. To evaluate the reliability of our proposed WSI metrics, we randomly sample $2\%$ of open-ended responses for assessment by five medical experts. Results show strong alignment between automated metrics and human evaluation (Pearson correlation: 0.87 for WSI-Precision, 0.82 for WSI-Relevance). Further details are provided in Appendix Section 4.3. + +# 4. WSI-LLaVA + +To enhance the explainability, we introduce a WSI-LLaVA framework with a three-stage training approach for gigapixel WSI analysis (Fig. 4). Trained on the training set of WSI-Bench, it can first provide morphological findings related to the query before presenting the final answer, significantly enhancing model explainability. To mitigate the cross-modal gap, the three-stage training approach includes a WSI-text alignment, feature space alignment, and task-specific instruction tuning. In the first stage, we integrate a patch-level encoder using Prov-GigaPath[32]'s tile-level encoder (pre-trained with DINOv2[21]) to process $256 \times 256$ pixel patches, and a slide-level encoder based on LongNet[8] to generate global features. These are aligned with textual features encoded by Bio_ClinicalBERT[1] through contrastive learning on WSI-report pairs. In the second stage, we combine the pre-trained WSI encoder with a projection layer and LLM, freezing the pre-trained models + +while training only the projection layer on WSI-report pairs. In the last stage, we freeze WSI encoder while training the projection layer and LLM on all tasks in WSI-Bench. Please see Appendix Section 5 for more details. + +# 5. Experiment + +# 5.1.Quantitative Evaluation + +We evaluate various WSI MLLMs on WSI-Bench dataset, including specialized models for WSI report generation (MI-Gen [3] and Hist-Gen [12]) and pathological VQA tasks (Quilt-LLaVA [25] and WSI-VQA [4]), and the general-purpose MLLM like GPT-4o. For models with input size constraints, we resize the WSIs to $1024 \times 1024$ pixels. All WSI MLLMs are trained on WSI-Bench's training set and evaluated on its test set. For quantitative evaluation, Table 2 shows that our WSI-LLaVA achieves the highest average score of 0.754, indicating its superiority over both WSI-based models and general-purpose MLLMs. Moreover, the performance of closed-ended questions for different models is better than that of open-ended questions, suggesting that the open-ended questions are more challenging for current models. + +For the report generation task, Table 3 shows that WSI-LLaVA achieves superior performance across all metrics, leading in both NLU measures (BLEU-1: 0.480, ROUGE-L: 0.490) and our domain-specific metrics (WSI-Precision: 0.380, WSI-Relevance: 0.429), indicating its superiority over general-purpose and specialized models. + +
ModelInputMorphological AnalysisDiagnosisTreatment PlanningAverage
Open WSI-PWSI-RClose AccOpen WSI-PWSI-RClose AccOpen WSI-PWSI-RClose Acc
Quilt-LLaVA[25]Thumbnail0.4480.4470.9470.5860.6040.8490.7880.8161.0000.721
GPT-4o[15]Thumbnail0.2200.2040.4710.4720.4570.5300.4960.8410.8750.507
WSI-VQA[4]WSI0.3950.4620.7580.4360.5250.5770.7910.5950.7710.590
WSI-LLaVAWSI0.4880.6100.9510.6100.6120.8630.8100.8451.0000.754
+ +Table 2. Quantitative evaluation of WSI and general-purpose MLLMs on our WSI-Bench across three pathological capabilities. WSI-P: WSI-Precision, WSI-R: WSI-Relevance, Acc: accuracy, open: open-ended question, and close: close-ended question. + +
ModelsBLEU-1BLEU-2BLEU-3BLEU-4ROUGE-LMETEORWSI-PrecisionWSI-Relevance
GPT-4o[15]0.2020.0690.0300.0160.1320.1670.0670.138
Quilt-LLaVA[25]0.4740.3510.2820.2360.4750.4600.3240.333
MI-Gen[3]0.4030.3060.2480.2090.4460.4070.3100.377
Hist-Gen[12]0.4060.3070.2480.2080.4480.4160.3000.397
WSI-LLaVA0.4800.3580.2870.2400.4900.4650.3800.429
+ +Table 3. Quantitative evaluation of WSI and general-purpose MLLMs in report generation task on WSI-Bench. + +
Image
QuestionBased on the image, what is the histological classification?
Quilt-LLaVAThe histological classification is renal cell carcinoma, conventional (clear cell) type. This classification is supported by the presence of clear cells with distinct borders and abundant cytoplasm, arranged in nests and sheets, along with the noted nuclear features.
WSI-VQAThe histological classification based on the observed features is pheochromocytoma. This classification is supported by the presence of polygonal cells with clear cytoplasm and distinct cell borders along with the presence of sustentacular cells.
WSI-LLaVAThe histological classification is pheochromocytoma. This is supported by the presence of large polygonal cells with abundant granular cytoplasm and round to oval nuclei, along with the characteristic zellballen pattern and vascular stroma.
Ground TruthThe histological classification is pheochromocytoma. This is determined by the presence of the zellballen pattern and typical cytological features such as eosinophilic cytoplasm and finely stippled chromatin within the tumor cells.
+ +Table 4. Sample Comparison of VQA. Correct responses are highlighted in green, and incorrect responses are highlighted in red. + +# 5.2. Qualitative Evaluation + +In Table 4, all models trained with WSI-Bench gain diagnostic explanation capabilities. Quilt-LLaVA incorrectly diagnoses renal cell carcinoma due to misidentified morphology. WSI-VQA and WSI-LLaVA correctly identify pheochromocytoma, with WSI-LLaVA providing more accurate supporting features (e.g., Zellballen pattern) than WSI-VQA, demonstrating WSI-LLaVA's superior ability to identify key diagnostic features. + +# 5.3. Ablation Studies + +We conduct ablation studies to verify our WSI encoder and three-stage training scheme (Appendix Table 2). Without the WSI encoder, performance drops across all pathological capabilities, with overall average decreasing from $68.0\%$ to $63.2\%$ $(-5\%)$ , demonstrating its effectiveness for WSI feature extraction. For our training approach, removing the WSI-text alignment stage reduces overall performance by $2.2\%$ , while ablating both this and the feature space alignment stage decreases performance by $3.3\%$ , confirming each stage's contribution to cross-modal alignment. + +# 5.4. External Validation + +To assess WSI-LLaVA's generalization capability, we performed zero-shot external validation on three datasets: WSI-VQA dataset, SlideBench(BCNB), and CPTACNSCLC. In Table 6, WSI-LLaVA consistently outperforms existing methods across all datasets, showing robust generalization to non-TCGA sources. The performance gap widens significantly on tasks requiring detailed morphological interpretation, with WSI-LLaVA achieving $46.3\%$ vs. second-best $30.0\%$ for histological grading and $29.2\%$ vs. second-best $18.1\%$ for molecular subtyping. These results indicate WSI-LLaVA's strong morphological understanding enables more accurate zero-shot diagnostic assessments. + +# 5.5. Comparison of NLU Metrics and WSI Metrics + +Table 5 compares pathology report generation performance using traditional NLU and proposed WSI metrics. While Hist-Gen and WSI-LLaVA showed similar NLU scores (BLEU-1: 0.531 vs. 0.521, ROUGE-L: 0.536 vs. 0.534), their clinical accuracy differed significantly. Hist-Gen incorrectly reported "no evidence of vascular invasion or + +
ModelModel ResponsesWSI MetricsNLU Metrics
WSI-PWSI-R
Ground TruthMicroscopic observation of the pathology slide reveals poorly differentiated squamous cell carcinoma ex- +tensive invasion with minimal keratinization and significant cellular atypia. The tumor displays ex- +tensive necrosis alongside regions of calcification within the tissue. Additionally, there is evi- +dence of tumor infiltration into surrounding structures, with observable lymphovascular invasion. +Based on these observations, the diagnosis is squamous cell carcinoma, Grade 3 (of 3).N/AN/AN/A
WSI- +LLaVAMicroscopic observation of the pathology slide reveals poorly differentiated squamous cell carcinoma char- +acterized by ... with minimal keratinization. The tumor cells exhibit marked nuclear pleomorphism, +with enlarged, irregular nuclei and prominent nucleoli. ... Vascular invasion is evident, with tu- +mor cells infiltrating lymphovascular spaces. Areas of necrosis are present within the tumor mass. +Based on these observations, the diagnosis is poorly differentiated invasive squamous carcinoma.0.8000.830BLEU-1: 0.521 +BLEU-2: 0.389 +BLEU-3: 0.326 +BLEU-4: 0.283 +ROUGE-L: 0.534 +METEOR: 0.617
Hist-GenMicroscopic observation of the pathology slide reveals a poorly differentiated squamous cell carcinoma char- +acterized by ... atypical squamous cells. The tumor cells exhibit marked nuclear pleomorphism with promi- +nent nucleoli and prominent nucleoli. There is no evidence of vascular invasion or necrosis within the exam- +ined tissue. Based on these observations, the diagnosis is squamous cell carcinoma poorly differentiated.0.2330.520BLEU-1: 0.531 +BLEU-2: 0.402 +BLEU-3: 0.316 +BLEU-4: 0.259 +ROUGE-L: 0.536 +METEOR: 0.560
+ +Table 5. Comparison between WSI metrics and NLU metrics. Green: matched ground-truth content; Red: deviations; Orange: ground truth content missing in model response; Underlined: template language. WSI-P: Precision, WSI-R: Relevance. + +
ModelInputWSI-VQA DatasetSlideBench-VQA (BCNB)CPTAC-NSCLC
Tumor TypeHistological GradingMolecular SubtypeAverage
Quilt-LLaVA[25]Thumbnail0.1320.1130.1940.1810.1630.603
WSI-VQA[4]WSI0.4690.0390.3000.0000.1130.000
SlideChat[5]WSI-0.9020.2310.1750.436-
WSI-LLaVAWSI0.5460.9050.4630.2920.5530.721
+ +Table 6. Zero-shot external validation on three external datasets, i.e, WSI-VQA dataset, SlideBench-VQA (BCNB), and CPTAC-NSCLC. + +necrosis” despite their presence in ground truth. WSI metrics better reflected clinical accuracy, with WSI-LLaVA scoring much higher (WSI-Precision: 0.8) than Hist-Gen (0.233). This reveals NLU metrics' limitations in prioritizing surface similarity and n-gram matching over factual correctness, emphasizing the need for domain-specific metrics that evaluate clinical accuracy rather than linguistic similarity in computational pathology. + +# 5.6. Evaluation on Morphological Capability + +To further analyze WSI MLLMs' morphological capabilities, Fig. 5 compares performance across morphological tasks. Current WSI MLLMs and general MLLMs like GPT-40 show limited proficiency in both global and local morphological analysis. WSI-LLaVA consistently outperforms other WSI MLLMs across all six tasks, suggesting its robust morphological understanding enables superior performance at both analysis levels. + +# 6. Conclusion + +This work introduces WSI-Bench, the first morphology-aware benchmark containing 180k VQA pairs from 9,850 WSIs across 30 cancer types, alongside WSI-LLaVA, an MLLM for WSI analysis. Through three-stage training, WSI-LLaVA bridges the gap between WSIs and text, enabling detailed morphological reasoning and explainabil + +![](images/73e3c2739a24ff4301866ea1349baa02120c47f84ed45bcf31ee39089774f533.jpg) +Figure 5. Morphological capability evaluation. The regional structure description includes three description sub-tasks, i.e., infiltration, tissue, and cellular. + +ity. WSI-LLaVA outperforms existing methods across all pathological tasks on both internal and external datasets, while our WSI metrics provide more clinically relevant assessment than NLU metrics. By improving morphological understanding and explainability, WSI-LLaVA represents significant advancement toward trustworthy AI assistants for computational pathology. + +# Acknowledgments + +This study was supported by the National Natural Science Foundation of China (Grants 82261138629 and 12326610), the Guangdong Provincial Key Laboratory (Grant 2023B1212060076), the Yongjiang Technology Innovation Project (Grant 2022A-097-G), and the National Natural Science Foundation of China grant (UNNC: B0166). + +# References + +[1] Emily Alsentzer, John R Murphy, Willie Boag, Wei-Hung Weng, Di Jin, Tristan Naumann, and Matthew McDermott. Publicly available clinical bert embeddings. arXiv preprint arXiv:1904.03323, 2019. 6, 8 +[2] Richard S Bradbury, Sarah GH Sapp, Idzi Potters, Blaine A Mathison, John Frean, Abhishek Mewara, Harsha Sheorey, Francesca Tamarozzi, Marc Roger Couturier, Peter Chiodini, et al. Where have all the diagnostic morphological parasitologists gone? Journal of clinical microbiology, 60(11): e00986-22, 2022. 3 +[3] Pingyi Chen, Honglin Li, Chenglu Zhu, Sunyi Zheng, Zhongyi Shui, and Lin Yang. Wsicaption: Multiple instance generation of pathology reports for gigapixel whole-slide images. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 546-556. Springer, 2024. 1, 3, 6, 7 +[4] Pingyi Chen, Chenglu Zhu, Sunyi Zheng, Honglin Li, and Lin Yang. Wsi-vqa: Interpreting whole slide images by generative visual question answering. In European Conference on Computer Vision, pages 401-417. Springer, 2025. 1, 2, 3, 6, 7, 8 +[5] Ying Chen, Guoan Wang, Yuanfeng Ji, Yanjun Li, Jin Ye, Tianbin Li, Bin Zhang, Nana Pei, Rongshan Yu, Yu Qiao, et al. Slidechat: A large vision-language assistant for whole-slide pathology image understanding. arXiv preprint arXiv:2410.11761, 2024. 1, 2, 3, 8 +[6] Jacob Devlin. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 8 +[7] Lokuhetty Dilani, AW Valerie, W Beiko, et al. Who classification of tumours, breast tumours. Lyon: International Agency for Research on Cancer, 2019. 5, 4 +[8] Jiayu Ding, Shuming Ma, Li Dong, Xingxing Zhang, Shaohan Huang, Wenhui Wang, Nanning Zheng, and Furu Wei. Longnet: Scaling transformers to 1,000,000,000 tokens. arXiv preprint arXiv:2307.02486, 2023. 6, 8 +[9] Meidan Ding, Jipeng Zhang, Wenxuan Wang, Haiqin Zhong, Xiaoqin Wang, Xinheng Lyu, Wenting Chen, and Linlin Shen. Eagle: Expert-guided self-enhancement for preference alignment in pathology large vision-language model. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics. Association for Computational Linguistics, 2025. 1 +[10] Navid Farahani, Anil V Parwani, and Liron Pantanowitz. Whole slide imaging in pathology: advantages, limita + +tions, and emerging perspectives. Pathology and Laboratory Medicine International, pages 23-33, 2015. 1 +[11] Álvaro García-Barragán, Alberto González Calatayud, Oswaldo Solarte-Pabón, Mariano Provencio, Ernestina Menasalvas, and Víctor Robles. Gpt for medical entity recognition in spanish. Multimedia Tools and Applications, pages 1-20, 2024. 2 +[12] Zhengrui Guo, Jiabo Ma, Yingxue Xu, Yihui Wang, Liansheng Wang, and Hao Chen. Histgen: Histopathology report generation via local-global feature encoding and cross-modal context interaction. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 189-199. Springer, 2024. 1, 3, 6, 7 +[13] Xuehai He, Yichen Zhang, Luntian Mou, Eric Xing, and Pengtao Xie. Pathvqa: 30000+ questions for medical visual question answering. arXiv preprint arXiv:2003.10286, 2020. 2, 3 +[14] Zexue He, Yu Wang, An Yan, Yao Liu, Eric Y Chang, Amilcare Gentili, Julian McAuley, and Chun-Nan Hsu. Medeval: A multi-level, multi-task, and multi-domain medical benchmark for language model evaluation. arXiv preprint arXiv:2310.14088, 2023.5 +[15] Raisa Islam and Owana Marzia Moushi. Gpt-4o: The cutting-edge advancement in multimodal llm. Authorea Preprints, 2024. 7 +[16] Fenglin Liu, Hongjian Zhou, Boyang Gu, Xinyu Zou, Jina Huang, Jinge Wu, Yiru Li, Sam S Chen, Yining Hua, Peilin Zhou, et al. Application of large language models in medicine. Nature Reviews Bioengineering, pages 1-20, 2025. 2 +[17] Pilar López-Ábeda, Teodoro Martín-Noguerol, José Aneiros-Fernández, and Antonio Luna. Natural language processing in pathology: current trends and future insights. The American Journal of Pathology, 192(11):1486–1495, 2022. 2 +[18] Ming Y Lu, Bowen Chen, Drew FK Williamson, Richard J Chen, Kenji Ikamura, Georg Gerber, Ivy Liang, Long Phi Le, Tong Ding, Anil V Parwani, et al. A foundational multimodal vision language ai assistant for human pathology. arXiv preprint arXiv:2312.07814, 2023. 1, 2 +[19] Xinheng Lyu, Yuci Liang, Wenting Chen, Meidan Ding, Ji-aqi Yang, Guolin Huang, Daokun Zhang, Xiangjian He, and Linlin Shen. Wsi-agents: A collaborative multi-agent system for multi-modal whole slide image analysis. arXiv preprint arXiv:2507.14680, 2025. 1 +[20] Daniel McDuff, Mike Schaekermann, Tao Tu, Anil Palepu, Amy Wang, Jake Garrison, Karan Singhal, Yash Sharma, Shekoofeh Azizi, Kavita Kulkarni, et al. Towards accurate differential diagnosis with large language models. Nature, pages 1-7, 2025. 2 +[21] Maxime Oquab, Timothée Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 6, 8 +[22] Thiago Santos, Amara Tariq, Susmita Das, Kavyasree Vayalpati, Geoffrey H Smith, Hari Trivedi, and Imon Banerjee. + +Pathologybert-pre-trained vs. a new transformer language model for pathology domain. In AMIA annual symposium proceedings, page 962. American Medical Informatics Association, 2022. 5 +[23] Elisabetta Sauta, Flavio Sartori, Luca Lanino, Gianluca Asti, Saverio D'Amico, Mattia Delleani, Elena Riva, Matteo Zampini, Elena Zazzetti, Marilena Bicchieri, et al. Enhancing personalized prognostic assessment of myelodysplastic syndromes through a multimodal and explainable deep data fusion approach (magaera). Blood, 144:105, 2024. 2 +[24] Marta Scorsetti, Francesco Leo, Annalisa Trama, Rolando D'Angelillo, Danila Serpico, Marianna Macerelli, Paolo Zucali, Gemma Gatta, and Marina Chiara Garassino. Thymoma and thymic carcinomas. Critical reviews in oncology/hematology, 99:332-350, 2016. 4 +[25] Mehmet Saygin Seyfioglu, Wisdom O Ikezogwo, Fatemeh Ghezloo, Ranjay Krishna, and Linda Shapiro. Quilt-llava: Visual instruction tuning by extracting localized narratives from open-source histopathology videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13183–13192, 2024. 1, 2, 3, 6, 7, 8 +[26] George Shaikovski, Adam Casson, Kristen Severson, Eric Zimmermann, Yi Kan Wang, Jeremy D Kunz, Juan A Retamero, Gerard Oakley, David Klimstra, Christopher Kanan, et al. Prism: A multi-modal generative foundation model for slide-level histopathology. arXiv preprint arXiv:2405.10254, 2024.3 +[27] Xiyue Wang, Yuexi Du, Sen Yang, Jun Zhang, Minghui Wang, Jing Zhang, Wei Yang, Junzhou Huang, and Xiao Han. Retccl: Clustering-guided contrastive learning for whole-slide image retrieval. Medical image analysis, 83: 102645, 2023. 1 +[28] Xiyue Wang, Junhan Zhao, Eliana Marostica, Wei Yuan, Ji etian Jin, Jiayu Zhang, Ruijiang Li, Hongping Tang, Kanran Wang, Yu Li, Fang Wang, Yulong Peng, Junyou Zhu, Jing Zhang, Christopher R Jackson, Jun Zhang, Deborah Dillon, Nancy U Lin, Lynette M. Sholl, Thomas Denize, David Meredith, Keith L. Ligon, Sabina Signoretti, Shuju Ogino, Jeffrey A Golden, MacLean P Nasrallah, Xiao Han, Sen Yang, and Kun-Hsing Yu. A pathology foundation model for cancer diagnosis and prognosis prediction. Nature, 2024. 1 +[29] John N Weinstein, Eric A Collisson, Gordon B Mills, Kenna R Shaw, Brad A Ozenberger, Kyle Ellrott, Ilya Shmulevich, Chris Sander, and Joshua M Stuart. The cancer genome atlas pan-cancer analysis project. Nature genetics, 45(10):1113-1120, 2013. 5 +[30] Xiaomin Wu, Rui Xu, Pengchen Wei, Wenkang Qin, Peixiang Huang, Ziheng Li, and Lin Luo. Pathinsight: Instruction tuning of multimodal datasets and models for intelligence assisted diagnosis in histopathology. arXiv preprint arXiv:2408.07037, 2024. 1 +[31] Jinxi Xiang, Xiyue Wang, Xiaoming Zhang, Yinghua Xi, Feyisope Eweje, Yijiang Chen, Yuchen Li, Colin Bergstrom, Matthew Gopaulchan, Ted Kim, et al. A vision-language foundation model for precision oncology. Nature, 638 (8051):769-778, 2025. 1 + +[32] Hanwen Xu, Naoto Usuyama, Jaspreet Bagga, Sheng Zhang, Rajesh Rao, Tristan Naumann, Cliff Wong, Zelalem Gero, Javier González, Yu Gu, et al. A whole-slide foundation model for digital pathology from real-world data. Nature, pages 1-8, 2024. 1, 6, 8 +[33] Brian J Yoder, Edward J Wilkinson, and Nicole A Massoll. Molecular and morphologic distinctions between infiltrating ductal and lobular carcinoma of the breast. The breast journal, 13(2):172-179, 2007. 2 +[34] Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric Xing, et al. Judging llm-as-a-judge with mt-bench and chatbot arena. Advances in Neural Information Processing Systems, 36:46595-46623, 2023. 8 \ No newline at end of file diff --git a/ICCV/2025/WSI-LLaVA_ A Multimodal Large Language Model for Whole Slide Image/images.zip b/ICCV/2025/WSI-LLaVA_ A Multimodal Large Language Model for Whole Slide Image/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..e4db9a6e7f2bfed8ee7b6e983975c11c2b86617f --- /dev/null +++ b/ICCV/2025/WSI-LLaVA_ A Multimodal Large Language Model for Whole Slide Image/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f9950b1c0fb814456c4d6fc4c1c84473de5c7880f99720a76642db319acacc01 +size 713016 diff --git a/ICCV/2025/WSI-LLaVA_ A Multimodal Large Language Model for Whole Slide Image/layout.json b/ICCV/2025/WSI-LLaVA_ A Multimodal Large Language Model for Whole Slide Image/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..939916a1b104654dab485c08336a245faf207e63 --- /dev/null +++ b/ICCV/2025/WSI-LLaVA_ A Multimodal Large Language Model for Whole Slide Image/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5599c7a16fe8ad2af82e9838f7b10817e5c130ddff54f73c8f037916e6426f87 +size 279698 diff --git a/ICCV/2025/WalkVLM_ Aid Visually Impaired People Walking by Vision Language Model/f2f29bbd-c165-4e1f-b880-aba87bbd2c1f_content_list.json b/ICCV/2025/WalkVLM_ Aid Visually Impaired People Walking by Vision Language Model/f2f29bbd-c165-4e1f-b880-aba87bbd2c1f_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..d39413db563465f5f753d8b4f2dd9b5b6ecea672 --- /dev/null +++ b/ICCV/2025/WalkVLM_ Aid Visually Impaired People Walking by Vision Language Model/f2f29bbd-c165-4e1f-b880-aba87bbd2c1f_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4ce07b5f9ea2de4ef26e50403f05d859c2dbd642e824ca39d0e5df3ddabf4ec5 +size 90575 diff --git a/ICCV/2025/WalkVLM_ Aid Visually Impaired People Walking by Vision Language Model/f2f29bbd-c165-4e1f-b880-aba87bbd2c1f_model.json b/ICCV/2025/WalkVLM_ Aid Visually Impaired People Walking by Vision Language Model/f2f29bbd-c165-4e1f-b880-aba87bbd2c1f_model.json new file mode 100644 index 0000000000000000000000000000000000000000..ea7c4bbd6ce873afcb506dbadc47fbf58851b05a --- /dev/null +++ b/ICCV/2025/WalkVLM_ Aid Visually Impaired People Walking by Vision Language Model/f2f29bbd-c165-4e1f-b880-aba87bbd2c1f_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e8150c1ad2cfe9fb31d0b8b0864cd17608fce97828b6b7ca429183e42afeb1a8 +size 115109 diff --git a/ICCV/2025/WalkVLM_ Aid Visually Impaired People Walking by Vision Language Model/f2f29bbd-c165-4e1f-b880-aba87bbd2c1f_origin.pdf b/ICCV/2025/WalkVLM_ Aid Visually Impaired People Walking by Vision Language Model/f2f29bbd-c165-4e1f-b880-aba87bbd2c1f_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..f6993375c719f0db72b0c6d3f096ea9276a28f0c --- /dev/null +++ b/ICCV/2025/WalkVLM_ Aid Visually Impaired People Walking by Vision Language Model/f2f29bbd-c165-4e1f-b880-aba87bbd2c1f_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1c5206951c8679d5af9b8c83868f466c1936201df88e6e08790646e92f068b4d +size 15861733 diff --git a/ICCV/2025/WalkVLM_ Aid Visually Impaired People Walking by Vision Language Model/full.md b/ICCV/2025/WalkVLM_ Aid Visually Impaired People Walking by Vision Language Model/full.md new file mode 100644 index 0000000000000000000000000000000000000000..66ecde6167c48b00c63fb423184ecada1fbe2866 --- /dev/null +++ b/ICCV/2025/WalkVLM_ Aid Visually Impaired People Walking by Vision Language Model/full.md @@ -0,0 +1,366 @@ +# WalkyLM : Aid Visually Impaired People Walking by Vision Language Model + +Zhiqiang Yuan $^{\dagger}$ Ting Zhang $^{\dagger}$ Yeshuang Zhu Jiapei Zhang Ying Deng Zexi Jia Peixiang Luo Xiaoyue Duan Jie Zhou Jinchao Zhang\* Pattern Recognition Center, WeChat AI, Tencent + +# Abstract + +Approximately 200 million individuals around the world suffer from varying degrees of visual impairment, making it crucial to leverage AI technology to offer walking assistance for these people. With the recent progress of vision-language models (VLMs), applying VLMs to offer walking guidance has become popular. However, the existing methods of walking guidance are mainly based on self-curated question-answering datasets that are not publicly accessible, without a standardized benchmark for training or evaluation. Moreover, walking assistance often requires real-time streaming video analysis and the generation of concise yet informative reminders, making VLMs struggle due to excessive responses and low efficiency in inferences. In this paper, we introduce the first large-scale dataset dedicated to walking assistance, comprising 12,000 video-annotation pairs, to provide a unified benchmark for training and evaluating systems to help visually-impaired individuals walk. Furthermore, a WalkVLM model is proposed, which employs chain of thought for hierarchical planning to generate concise but informative reminders and utilizes temporal-aware adaptive prediction to reduce the temporal redundancy of reminders. Finally, we have established a solid benchmark for blind walking task and verified the advantages of WalkVLM in stream video processing for this task compared to other VLMs. Dataset and code are available at https://walkvlm2024.github.io. + +# 1. Introduction + +Approximately 200 million people worldwide suffer from varying degrees of visual impairment, with 36 million completely blind [5, 29]. These visually impaired people (VIPs) are facing severe challenges in daily activities such as walking, which may be alleviated by contemporary artificial intelligence technologies [19, 41]. + +The current walking assistance works are primarily based on electronic assistance devices, sensory substitution devices, and computer-vision-based assistance systems [10, 12, 16]. Furthermore, vision-based assistance systems can be roughly categorized into detection-based meth + +![](images/e65433659490f288c72ab466cfa1fbc5c69b0c8c54bace065e374ba183180f97.jpg) +Figure 1. WalkVLM delivers timely, concise, informative walking reminders for visually impaired individuals by utilizing hierarchical planning and temporal-aware adaptive prediction. + +ods and semantic-based methods [11, 38, 42]. Detection-based methods have been extensively studied, focusing on identifying potential obstacles within the field of view to help visually impaired persons (VIPs) walk safely [24, 35]. In contrast, semantic-based methods utilize vision-language models (VLMs) to analyze images and generate responses to the questions of VIPs [39, 48]. With the recent advancements in VLMs [4, 8], semantic-based methods have gained significant attention within the community. Some studies have applied VLMs in zero-shot settings to evaluate their effectiveness in blind walking assistance [35, 39], while another line of work fine-tuned VLMs using traditional visual question-answer (VQA) datasets or small-scale and self-constructed datasets to improve the quality of the responses [27, 41]. These efforts have demonstrated the potential of VLMs in walking guidance tasks, yielding promising results and paving the way for the application in this area. + +While VLM-based models for walking assistance have shown promise, several challenges remain as obstacles to effective application in real-world scenarios. First, most recent research [27, 45] relies on limited image-text pairs, and there lacks a comprehensive and standardized benchmark. Additionally, the existing datasets [7, 48] primarily adopt a question-and-answer paradigm, allowing VLMs to respond to explicit queries but making it challenging for VLMs to proactively generate context-aware guidance. Second, blind walking requires opportune streaming video analysis and the generation of concise yet informative reminders. However, existing approaches [2, 47] struggle with excessive re + +dundancy in the generated responses and suffer from low efficiency in inference, posing significant challenges for practical applications. + +To address these challenges, we propose WalkVLM, a vision-language model designed for blind walking assistance, and establish a new benchmark to facilitate the research in this field. Specifically, we introduce the Walking Awareness Dataset (WAD), a large-scale and diverse dataset comprising 12k annotated videos collected from Europe and Asia, providing a fair and standardized foundation for both training and evaluation of this task. As illustrated in Figure 1, WalkVLM processes streaming video inputs by incorporating a chain-of-thought reasoning framework to hierarchically guide the model in generating concise and informative reminders, and achieves opportune reminders by the proposed temporal-aware adaptive prediction. Comprehensive experiments demonstrate that WalkVLM generates more concise guidance and exhibits superior temporal adaptability when handling video streaming in blind walking task compared to existing VLM methods. + +Our main contributions can be summarized as follows: + +- We construct a diverse and extensive dataset named the Walking Awareness Dataset (WAD), providing extensive data support for the blind walking task. +- We propose WalkVLM, a vision-language model for streaming video analysis, which adaptively generates concise yet informative walking guidance for visually impaired people. +- To our knowledge, we are the first to utilize VLMs to provide opportune and reasoning walking guidance for visually impaired individuals, establishing a strong foundation for the practical application of VLMs in this field. + +# 2. Related Work + +Vision Datasets for Walking Guidance. Existing datasets for blind walking can be roughly divided into two types: detection-based [14, 32, 38, 46] and semantic-based [11, 48]. Detection-based datasets have been extensively studied in blind walking guidance, where researchers utilize these datasets to train models for obstacle detection, thereby reducing the accident rates of VIPs for this task. For example, Zhang et al. [46] recently developed a TP-Dataset for detecting visual tactile paving surfaces and offered walking route guidance for the visually impaired. Islam et al. [14] introduced a dataset containing 90 object annotations from 31 video clips to improve real-time object recognition systems to aid VIPs in navigation tasks. Compared with detection-based datasets, semantic-based datasets are relatively limited, primarily containing data for question answering and providing an enhanced human-computer interaction experience. Gurari et al. [11] constructed a visual question answering (VQA) dataset for VIPs containing 31k visual questions, each with 10 answers anno + +![](images/2c78a2cb059c53d1c3ff67a14f20f1cc749e7a1485723a27cefaa67f7250311c.jpg) +Figure 2. The data annotation pipeline for constructing the Walking Awareness Dataset (WAD). Appendix 1.5 provides more examples to illustrate the diversity of WAD dataset. + +tated by crowdsourcing. In addition, there have been several self-built question-answering datasets with specific attributes [41, 48], however, these datasets are not publicly available and are relatively small in scale, making them unsuitable for large-scale and unified benchmarking. + +Vision-based Methods for Walking Guidance. Similar to the categorization of datasets, the vision-based methods for walking guidance can also be divided into detection-based methods [24, 35] and semantic-based methods [48]. Detection-based methods typically use detectors to identify obstacles during walking and provide users with specific object locations. Liu et al. [24] proposed a system for understanding open scenes, which improves detection performance by using SAM [18] to generate pixel-level dense segmentation masks. Tian et al. [33] introduced a method to understand dynamic crosswalk scenes, including crosswalks, vehicles, and pedestrians, to provide VIPs with guidance of when and where to cross the road. In contrast, the semantic-based approach provides VIPs with the scene understanding in the form of question-answering. Merchant et al. [27] verified that vision-language models can generate correct and informative instructions for VIPs, and studied methods to provide users with context-related guidance. Yang et al. [41] explored how to utilize VLMs to provide reliable visual question answers for VIPs, and fine-tuned the VLMs with LoRA on a small amount of self-constructed dataset to generate detailed and practical suggestions. Moreover, a few applications such as Be My AI have adopted semantic-based methods to enable VIPs to take photos for visual question answering. However, these applications only support the question-and-answer paradigm and struggle to provide concise and timely guidance during walking. + +Vision-Language Models. With the advancements of large-language models (LMM), vision-language models have also gained significant attention [17, 45]. Liu et al. [22] introduced LLaVa, which integrates a ViT visual encoder to process images, followed by an MLP to map them to the LLM, achieving favorable performance in benchmark evaluations for vision-based question-answering tasks. Building on LLaVa, numerous studies have emerged, expanding its applications across diverse + +fields [6, 20, 31, 51]. Additionally, multimodal models such as Qwen, Gemini, and MiniCPM-V [3, 30, 43] have introduced support for multiframe image inputs and undergone optimizations for scenarios such as edge-device deployment, significantly improving the capacity of VLMs in a wide range of applications. Despite these advancements that demonstrated the potential of the large-scale multimodal models [34], the applications in domain-specific tasks remain limited. For example, only a limited number of studies [27, 41, 48] have explored the usage of VLMs for blind walking assistance, and a unified, systematic modeling approach for this task is still lacking. + +# 3. Walking Awareness Dataset + +# 3.1. Data Collection + +The dataset encompasses diverse geographical sources collected from 10 different locations across Europe and Asia. Approximately $20\%$ of the dataset is derived from the recordings made by annotators, while the remainder of the data is collected from YouTube2. Recorders positioned their cameras at chest level, using focal lengths of $13\mathrm{mm}$ , $20\mathrm{mm}$ , and $26\mathrm{mm}$ , with resolutions ranging from 1080p to 4K at 60fps. We collected approximately 13 hours of walking videos in total and see Appendix 1 for the details on the duration of data collected from various regions. + +# 3.2. Annotation Strategy + +Figure 2 shows the overall annotation pipeline of the walking awareness dataset. In the following sections, we detail the dataset construction process from two perspectives: scene annotation and response annotation. + +Scene annotation. Scene annotation aims to label the inherent attributes of the recorded scenes. We requested nine annotators to label each video scene in terms of weather conditions, location type, traffic flow rating, danger level, and scene description. Weather conditions are categorized into six types (e.g., sunny, rainy) for outdoor scenes, while indoor scenes are left unclassified. Location types are divided into eight categories, such as corridors and pedestrian walkways. The traffic flow rating is divided into three levels, determined based on the number of visible pedestrians in the video stream. The danger level reflects potential walking hazards in the scenes, which are qualitatively assessed by the traffic flow rating and road surface smoothness. Scene description provides a high-level summary of the environment, incorporating details such as pedestrian and vehicle movement, road conditions, and surrounding context. Subsequently, we employed the open-world detection model [50] for the preliminary object detection, followed by manual verification to ensure the data reliability. + +Response annotation. The response in our dataset refers + +![](images/3ece77a373712ea16a99e4c66ad7df2e18a82b203f0524d3c43e9b0eec35fa13.jpg) +Figure 3. Blind test experiment for analyzing the most critical information needed by users in blind walking. We required two individuals to collaborate as a team, where the participant at the rear provided directions to enable the individual at the front to arrive at a specific location safely in the absence of any visual information. + +to the concise reminders that the model aims to generate, as well as the answers that directly address the user's inquiries in the task of blind walking assistance. To identify the most critical information needed by users in this context, we conducted a blind test experiment, as illustrated in Figure 3. In this experiment, two participants collaborated in pairs: Person A, positioned behind, provided verbal directions, while Person B, wearing an eye mask, followed these instructions to reach a designated destination without collisions. This setup ensures that Person B relies solely on the verbal instructions from Person A, eliminating any pre-existing route knowledge that a VIP might possess and enabling us to analyze what information is necessary for this task. Through extensive trials, our findings confirm that the provided guidance enables visually impaired individuals to navigate safely, indicating that the information provided by Person A is sufficiently effective for Person B. To further refine our analysis, we recorded both video and audio of the experiments, examined the communication between participants, and subsequently identified key types of information essential for this task. These information serve as the basis for subsequent reminder and QA annotations. + +- Reminder type. Based on the blind test experiment, as shown in Figure 4, we categorize the reminders during walking into six types. (a) Obstacle reminder: Trigger a reminder when there is a non-moving obstacle on the walking route. (b) Intersection reminder: Trigger a re + +![](images/56dc7e558e2c33bca605134260a1a921c0d59b76406629168f0b04c098da10ea.jpg) +Figure 4. Visualization of six scenarios that require reminders, which were summarized through multiple blind experiments. + +
DatasetType#SampleModalityBounding BoxWeatherDanger levelScene SummaryQAReminderOpen
Obstacle Dataset (2023)[32]T8kImage✓XXXXX✓
WOTR (2023)[38]T13kImage✓XXXXX✓
ISLAM et al.(2024)[15]T31Image & Video✓XXXXX✓
Wang et al.(2024)[35]T50Video✓✓XXXXX
VizWiz (2018)[11]S31kImageXXXX✓X✓
Zain et al.(2024)[27]S48ImageXXXXX✓X
WAD (Ours)TS12k / 120kVideo / Image✓✓✓✓✓✓✓
+ +Table 1. Static information comparison of different datasets in blind walking. $\mathcal{T}$ and $S$ denote the dataset types,representing target-based and semantic-based datasets,respectively.WAD dataset holds a great advantage in terms of sample numbers, categories, and modalities. + +minder when the current road has intersections, turns, etc. (c) Road clear/narrow reminder: Provide reminders about the width and passability of the road. (d) Oncoming vehicle/person reminder: When there are moving obstacles on the walking route, trigger a reminder for potential dangers. (e) Road departure warning: Issue a warning when there is an angular offset between the walking route and the current road. (f) Identifier reminder: Provide reminders for prominent landmarks in the scene, such as road signs and traffic lights. + +- QA type. For QA type, we summarize based on three aspects: scene perception, road inquiry, and detailed consultation. (a) Scene perception: Macro-level insights such as the understanding of the scene. (b) Road inquiry: Route planning to reach a certain location within visible range. (c) Detailed consultation: Knowledge QA on local details, such as road sign content, shop names, etc. + +During annotating, we asked nine annotators to indicate the specific location of obstacles in the video. During annotating, the distances are represented by steps on a scale of 5, and the directions are indicated by clock positions to reduce the offset caused by the camera perspective. The annotation interface is in Appendix 1.3. After completing the annotations, we apply GPT [9] to rephrase the annotated content and conduct manual verification to further standardize the annotation to remove potential bias. + +# 3.3. Dataset Analysis + +Figure 5 shows an example of the WAD dataset, and we divide the annotations into three groups according to lower to higher semantic levels: perception, comprehension, and decision. The perception label reflects the basic attributes, such as obstacle location, weather conditions, etc., while the comprehension label indicates the model's understanding of the entire scene. The decision contains reminder and QA, reflecting the model's decision on the user's walking based on its understanding of the current scenario. + +Table 1 illustrates the comparison between the WAD dataset and other prevalent datasets utilized in blind walking tasks, with $\mathcal{T}$ representing the detection-based dataset and $S$ indicating the semantic-based dataset. Compared to other datasets, WAD includes more extensive data while containing more static attributes of the environment, scene summaries, QA, and reminders, thus providing more supervi + +![](images/ae6f23939c4bc469b61b6fd441498cdfb761f5b1abb0ce51e11211bd17ceceff.jpg) + +Objects: Frame $t_1 - [x_2, y_2, x_2, y_2, prob, cls], \ldots ]$ , Frame $t_2 - \ldots$ + +Weather: Sunny + +Location: Foot Path + +Danger Level: High + +Traffic Flow Rating: Low + +Scene Summary: The person is walking on the blind path. A street is on the left, and a pile of chairs at 11 o'clock in front of the right. There are two pedestrians passing by on the left. The pedestrian on the left is wearing a white shirt and gray shorts, and the pedestrian on the right is wearing a white short-sleeved shirt and black trousers. There is not much traffic on the road. + +Reminder (i): There are pedestrians about 5 steps ahead at 11 o'clock on the left, and a pile of chairs on the right. Be careful to avoid them. + +QA $(\mathbf{t}_4)$ + +Question-Is there a shop on my right? + +Answer: There are two shops on your right. One of them looks like it is under renovation and its door is open. The other is a gastrop pub, about 10 steps straight ahead. + +![](images/c25f1ae691385056f38c85df6998fd66e9de2c4ac1784473f6b662ccac05a236.jpg) +Perception +Figure 5. Visualization of the Walking Awareness Dataset. Each sample contains a video clip and multiple annotations, with the hierarchy divided into perception, comprehension, and decision. + +![](images/ff8de74ceb586ebe81c9580017c5a437023b2a227ea709a2c706a7d0fb08e8a7.jpg) +Brehension + +![](images/bb5d4944dd11c447e5b11d96903ce3081bf824a9550a376043d0025eb7a04490.jpg) +Decision + +![](images/c9e5029f36f7ae3f91841e6be238332699c9efb1f31b9ebbf58ba34bba6ee49e.jpg) +(a) Target category distribution +Figure 6. Visualization of the proportion of targets and categories in our Walking Awareness Dataset. + +![](images/0dc086cb644474c8b4517dc5c478e1ce8e7d412825363c48d852334375fcc68c.jpg) +(b) Static attribute categories and proportions + +sion to train the model. It is worth emphasizing that the samples we provide are exclusively video clips, which possess a greater volume of information in comparison to the images supplied by other datasets. Moreover, for each video clip, we extracted 10 keyframes to streamline the usage in broader research. The walking awareness dataset contains 3.47 million instances, and we show categories and the respective proportions in Figure 6(a). The distribution of categories in the WAD dataset is shown in Figure 6(b). We selected 1.5k samples as a test set based on different static tag types, reminder types, and QA types to ensure diversity and completeness in evaluation. + +# 3.4. Further Explanation of Data + +The WAD dataset is utilized for model training to generate auxiliary information that assists edge groups in making walking decisions. While guide dogs and canes are commonly used by VIPs in real-world scenarios, the WAD dataset primarily focuses on capturing the most critical vi + +![](images/3843a785beeafde1bcb929fcda10ac484fa659ac2894c937fb7002449acd3afc.jpg) +Figure 7. An overview of the proposed WalkVLM framework. WalkVLM employs CoT-based hierarchical planning to summarize the static attributes and understanding of scenes, thereby facilitating the subsequent reminder generation and QA tasks. Furthermore, temporal-aware adaptive prediction is proposed to calculate the trigger state of VLM, thereby reducing the temporal redundancy of outputs. + +sual information. Future iterations of the dataset will incorporate additional elements to further enhance the applicability. Given the limited availability of participants within the VIP community, we adopt a blind test experiment to determine appropriate annotation types. To assess the alignment of the obtained annotation types with the needs of VIPs, we conducted 34 validated surveys in which VIPs evaluated the applicability of the annotation types to walking. As + +
PositiveRelatively PositiveNeutralRelatively NegativeNegative
Ratio41.18%47.06%8.82%2.94%0.00%
+ +Table 2. The attitude proportion of the VIP group towards the annotation types of the WAD dataset. + +shown in Table 2, approximately $88.24\%$ of the participants expressed a positive attitude towards the WAD dataset, especially appreciating the reminders at traffic intersections. Respondents with a neutral or lower attitude indicated the potential redundancy in the annotations and expressed concerns about the risk of inaccurate prompts from the model. + +# 4. WalkVLM + +The overall architecture of WalkVLM is shown in Figure 7. We start with problem formulation and proceed with the hierarchical planning and temporal-aware adaptive prediction to generate concise and opportune walking reminders. + +# 4.1. Problem Formulation + +We aim to guide a VLM to process video streams, enabling it to provide walking reminders with temporal attributes and respond to specific queries in human-machine interactions. Specifically, at time $t_0$ , given a sequence of + +the newly appeared frames $[I_{t_{-N}},\dots,I_{t_{-1}},I_{t_0}]$ , and the corresponding category and obstacle positions in the image $[O_{t_{-N}},\dots,O_{t_{-1}},O_{t_0}]$ , VLM is expected to generate a concise and informative reminder $T_{t_0}^R$ based on visual context. During walking, VIPs can pose a question $Q_{t_0}$ to communicate with the VLM at any time to request information regarding the current scene and route. However, generating reminders at every frame may degrade the walking guidance experience and impose significant real-time processing pressure on the hardware. To address this, WalkVLM predicts the current VLM trigger state $s_{t_0}$ based on historical states $[s_{t_{-N}},\dots,s_{t_{-1}},s_{t_0}]$ and the previous $N$ frames, to determine optimal moments for issuing reminders. + +# 4.2. CoT-Based Hierarchical Planning + +To enhance the reasoning capabilities of the VLM, we use Chain of Thought (CoT) [37] to enable the model to derive step-by-step conclusions by summarizing comprehensive information, including static attributes and scene summary. This approach refines raw visual inputs into concise and contextually relevant reminders. The model architecture integrates a vision transformer encoder with a large language model (LLM). The vision encoder generates image tokens, while an attention-based extractor aligns these tokens with the LLM, facilitating comprehensive information processing and understanding. WalkVLM aggregates multi-frame information to ensure a holistic perception of the surroundings, thereby improving the accuracy and reliability of walking guidance. + +We structure the reminder generation process into three + +levels: perception, comprehension, and decision. At the perception level, the model extracts static visual attributes from the current frame, such as location type, weather conditions, and traffic flow rating. To enhance the VLM's ability to focus on critical elements and improve visual perception accuracy, we incorporate the Priori-Object Location Module (POLM). The POLM initially uses a generic object detector [50] to identify and locate objects in the scene. It then filters these objects based on size and confidence scores to prioritize key elements that reflect road conditions and potential hazards. The filtered information, along with fundamental environmental attributes, provides the model with sufficient representation of the external world. At the comprehension level, the model integrates all outputs from the perception stage, merging local detection results and fragmented scene information into a comprehensive global summary. Utilizing the capabilities of the VLM and the detailed attributes extracted in the perception stage, this phase ensures that the model attains a clear understanding of the current environment. At the decision level, we focus on training the WalkVLM model to perform visual QA and generate reminders. At this stage, the model possesses an understanding of the static attributes and overall context of the environment. With appropriate guidance, the model is expected to analyze potential hazards in the scene. + +During training, we adopt a CoT approach to progressively feed information from three levels into the VLM, and during testing, the model predicts the aforementioned attributes and generates the corresponding responses. + +# 4.3. Temporal-Aware Adaptive Prediction + +Despite VLMs are capable of scene parsing across multiple frames and generating the required output, directly applying them to video streaming introduces several challenges. First, generating walking reminders frame by frame or at fixed intervals can result in significant temporal redundancy, which can degrade the user experience. Additionally, continuous VLM inference imposes computational pressure on hardware devices, making real-time processing inefficient. Addressing these challenges is crucial for the effective deployment of VLMs in video streaming processing. + +To address the aforementioned issues, we introduce the Temporal-aware Adaptive Prediction (TAP) module that incorporates historical information to pre-determine whether the VLM should be triggered at a given moment, thereby alleviating the inference pressure on hardware. Specifically, as shown on the right of Figure 7, TAP utilizes a lightweight model to analyze historical $N$ frames and determine whether to trigger the VLM at the current moment based on the prior output states. Given a sequence of frames $[I_{t - N},\dots,I_{t - 1},I_{t_0}]$ , a 3D convolutional model extracts the features $f_{v}$ . Simultaneously, the predicted trigger states from the previous $N$ moments are independently embedded, concatenated, and then passed through multiple layers + +of perceptrons to generate the state feature $f_{s}$ . The final trigger probability $\mathcal{P}_t$ is obtained by integrating $f_{v}$ and $f_{s}$ through an additional MLP. To optimize reminder generation, TAP defines three trigger levels, corresponding to different degrees of environmental risk as categorized in the WAD dataset. Subsequent experiments demonstrate that TAP effectively reduces temporal redundancy by providing proper triggers for the VLM, ensuring that VLM-based walking guidance remains both efficient and effective. + +# 5. Experiments + +# 5.1. Settings + +Models & Details. WalkVLM is implemented with the MiniCPM-V2.6 model [43], which is an 8B multimodal model built upon Qwen2-7B [40]. We add LoRA to all the linear layers of MiniCPM-V2.6 with a rank of 64 while maintaining the video stream sampling rate of 2 FPS. The number of historical frames $N$ is set to 3, and the visual extraction backbone in the TAP module is ConvNext3D [25]. We compared WalkVLM with multiple popular multimodal models, including GPT-4o [13], Qwen2-VL (7B) [36], MiniCPM-V2.6 (8B) [43], DeepSeek-VL (1.3B&7B) [26], Yi-VL (6B) [44], LLaVa (7B) [23]. For open-source models with the same size, we use the WAD dataset to fine-tune them for comparison. All the prompts for the large model usage can be found in the Appendix 2. + +Metrics. We use the following metrics to evaluate the models: (a) ROUGE. This metric measures the similarity between the generated and the reference text by comparing overlapping words or phrases, including ROUGE-1, ROUGE-2, and ROUGE-L [21]. (b) TF-IDF Similarity. By combining term frequency and inverse document frequency to evaluate the weight of words, the text can be represented as a TF-IDF vector, and then the semantic similarity between texts is evaluated with TF-IDF similarity [28]. (c) GPT Score. GPT4 is used to evaluate the superiority ratio between the generation results of different multimodal models and the GPT-4o results [1, 49]. (d) Temporal Redundancy F1-Score (TRF). Given the historical model state and historical frames, the model predicts the danger level of the current moment, and the F1-Score is calculated between the prediction and the GT. The evaluation metrics capture different aspects of performance: TF-IDF and ROUGE evaluate semantic similarity and word-level granularity, respectively, while the GPT Score determines the optimal result by comparing model outputs with GT. Additionally, TRF quantifies the temporal redundancy of the model's output, where a higher TRF indicates lower redundancy in generated reminders. + +# 5.2. Quantitative Results + +Table 3 presents the quantitative metrics of different models on the reminder and QA task. On the ROUGE met + +
ModelReminder TaskQA Task
TF-IDFROUGE-1ROUGE-2ROUGE-LGPT ScoreTF-IDFROUGE-1ROUGE-2ROUGE-LGPT Score
LLaVa (7B) [23]0.0610.0620.0050.0700.0130.0740.0840.0110.0720.012
DeepSeek-VL (1.3B) [26]0.0730.0980.0150.0900.1480.1820.1030.0200.0950.336
DeepSeek-VL (7B) [26]0.1320.0730.0090.0680.0150.1890.0880.0210.0811.000
Yi-VL (6B) [44]0.1120.0930.0090.0850.1330.1130.0910.0120.0820.168
MiniCPM-V2.6 (8B) [43]0.1110.0710.0070.0640.0250.1920.1390.0250.1200.832
Qwen2-VL (7B) [36]0.1060.1070.0100.0970.0440.2320.1820.0370.1620.504
GPT-4o [13]0.1160.0780.0080.072-0.2420.1630.0340.145-
*DeepSeek-VL (7B) [26]0.1290.1520.0430.1410.9690.1930.1660.0370.1492.496
*MiniCPM-V2.6 (8B) [43]0.1520.1710.0560.1701.0240.2070.1760.0440.1603.172
*Qwen2-VL (7B) [36]0.1470.1630.0540.1651.0180.2460.1960.0470.1673.246
WalkVLM0.1660.1910.0620.1731.1030.1890.2020.0510.1744.168
+ +Table 3. Quantitative comparison of different methods on reminder generation and QA tasks. WalkVLM leads in almost all the TF-IDF, ROUGE, and GPT Score metrics. * indicates the fine-tuned model. Bold and underline indicate the best and the second-best, respectively. + +
ModelYi-VLMiniCPM-V2.6GPT-4oQwen2-VLWalkVLM
TRF0.3410.3960.4300.4490.505
+ +![](images/f2f85197fd5e6e1613797e383d4cf42f8de32f801d3ebabe5de002106c392542.jpg) +Figure 8. Visualization of triggering moments of GPT-4o and WalkVLM. WalkVLM triggers with less redundancy, providing information to users in a more timely manner. + +ric, WalkVLM achieved the highest scores in both tasks, confirming that the model's output is closest to the GT at the word granularity. For semantic similarity, measured by TF-IDF, WalkVLM performs best in reminder tasks, indicating that its outputs are more concise and accurate in alignment with GT. While in QA tasks, WalkVLM's performance on TF-IDF scores does not stand out significantly. This could be due to the training process, where the model is optimized to generate concise answers, potentially reducing its capacity to offer detailed explanations of the questions. The GPT score represents the overall evaluation of the LLM on the generated results and the GT. WalkVLM outperforms other models in GPT scores for both reminder and QA tasks, demonstrating that the model's outputs exhibit the most consistent distribution with the GT. + +We use TRF to evaluate the temporal redundancy of the output of various VLMs. Specifically, we provide the model with multiple frames of images along with historical dangerous states as inputs, allowing it to predict the current danger level and determine whether a reminder should be triggered. As shown in Table 4, WalkVLM achieved the highest TRF score among all models, indicating its superior ability to reduce reminder redundancy over time. + +# 5.3. Qualitative Results + +Figure 9 presents a visual comparison of the reminder task across different VLM models. Compared to other methods + +Table 4. Temporal redundancy assessment of the reminder task, our method achieved the highest TRF score. + +
ModelReminder TaskQA Task
Concise.Semantic.Concise.Semantic.
Yi-VL0.0850.0230.1210.022
GPT-4o0.0560.1950.0300.205
DeepSeek-VL(1.3B)0.0260.0800.0910.107
*DeepSeek-VL(7B)0.0020.1970.0440.114
*MiniCPM-V2.60.0260.1220.0610.190
*Qwen2-VL0.1210.1680.0770.192
WalkVLM0.6830.2160.5760.170
+ +Table 5. User study results on conciseness and semantic similarity across different tasks. Higher score indicates better performance. + +such as GPT-4o, WalkVLM generates more concise and informative responses, enhancing the user experience. In the left example, while other models provide highly detailed responses, WalkVLM delivers a concise prompt to the user, effectively emphasizing the most critical information. In the right example, WalkVLM successfully identifies a car approaching from the one o'clock direction and conveys this essential detail to the user, while other models fail to achieve. Figure 8 further illustrates a qualitative comparison of GPT-4o and WalkVLM in terms of temporal redundancy. Our model triggers the VLM less frequently, reducing temporal redundancy while ensuring the timely delivery of crucial information. Appendix 4 presents more qualitative results, including the comparison with other VLMs on real-world video streams. + +# 5.4. Subjective Results + +User Study. In Table 5, we recruited eight annotators to perform a subjective evaluation of various VLMs with respect to language conciseness and semantic similarity to the GT. Participants were required to rank the results independently, and we measured the model performance using the top-1 superiority ratio, where a higher score indicates better performance. Compared with other methods, WalkVLM demonstrates a substantial improvement in terms of conciseness, both in reminder and QA tasks. In terms of the semantic similarity to the GT, WalkVLM slightly outperforms GPT-4o in the reminder task but performs marginally worse in the QA task, which be attributed to its concise output style, which may occasionally omit details necessary to + +![](images/43947f873a83090da02ab671b3da4fbcefa51650019f7ba0a6ad6f2a12bda997.jpg) + +# Annotation + +Weather Condition: Sunny + +Area Type: Pedestrian Path + +Danger Level: High + +Summary: It is sunny. I am walking on the sidewalk. There is a pedestrian in a brown coat walking straight ahead. Five steps away at eleven o'clock, there is an adult in a red short-sleeved shirt and black shorts with a backpack leading a child in a white shirt and black pants walking in front. There are trees planted on both sides of the road. The road is narrow and there is a large flow of people. + +Alert: At three o'clock direction, there are pedestrians passing by, and in front there are pedestrians passing by. Be careful to avoid. + +# Qwen + +1 The video shows a group of people walking down a path. Based on the video, here are some pins for safe and comfortable walking: +1. Stay on the path: Make sure to stay on the designated path to avoid any potential hazards or obstacles. +2. Keep a safe distance: Maintain a safe distance from other people on the path to avoid collisions or accidents. +3. Watch your step: Be mindful of your footing and watch for any uneven surfaces or obstacles on the path. +4. Stay alert: Keep your eyes and ears open for any potential hazards or obstacles on the path. +5. Follow the rules: Follow any rules or guidelines. + +# MiniCPM + +To provide correct and unobstructed walking guidance based on the input video, follow these steps: +1. Observe the Pathway: The video shows a paved pathway surrounded by trees. Ensure you stay within this path to avoid any obstacles or uneven terrain. +2. Follow Traffic Flow: Walk in the same direction as others. In the video, people are walking around to maintain your position either behind or beside them, depending on available space. +3. Maintain Distance: Keep a safe distance from other individuals to ensure personal space is respected and to prevent potential collisions. +4. Stay Alert: Be aware of your surroundings. Look out for any changes in the path condition such as puddles, cracks, or debris that could cause tripping or falling. +5. Adapt to Environment: If the path becomes crowded, adjust your pace accordingly. In the video, there appears to be a moderate flow of people; however, it is not enough to show that you are in the center. +6. Use Handrails if Available: If there are hand rails along the path, use them for support, especially when navigating slopes or uneven sections. + +![](images/5abe976171002b2d738cad1250432a38cad4787068e5b040d2341607359e7645.jpg) + +# Annotation + +Weather Condition: Sunny Area Type: Road +Traffic Flow Rating: Low Danger Level: Low Summary: On a road, there are white and black cars passing at 1 o'clock direction. The main road is made of red and white. On the main road, there are white buildings ahead, a colorful car parked at 11 o'clock direction, a trash can ahead, a black car parked on the roadside at 10 o'clock direction, and pedestrians passing at 1 o'clock direction. Alert: At 1 o'clock direction, a car is passing by, be careful. + +# Ours + +There is a vehicle at 1 o'clock direction. Be careful to avoid. + +# GPT-40 + +To walk through this street safely and unobstructed, follow these steps: +1. Stay on the Left Side: The right side of the street is lined with parked vehicles, which could obstruct your path. Walking on the left side will provide a clearer path. +2. Watch for Obstacles: Keep an eye out for any obstacles like motorcycles, cars, or people that may be on the left side of the street. Navigate around them carefully. +3. Proceed Straight Ahead: Continue walking straight ahead of the main road, and head backgazl. This will keep you on a relatively clear path. +4. Be Aware of Moving Vehicles: Although the street is narrow, there may still be vehicles moving through Stay alert and be prepared to step aside if necessary. 5. Look for Pedestrian Walkways: If there are any designated pedestrian paths or sidewalks, use them for added safety. +By following these steps, you should be able to run through the street without any major obstructions. + +# Qwen + +The given images show a narrow street with parked cars on both sides. The street appears to be a residential area with buildings on either side. There are no visible pedestrians or other vehicles in the image. Based on this information, the correct and unobstructed walking guidance would be to walk down the street in the direction shown in the images. + +# MiniCPM + +To navigate through the narrow street in the video, follow these steps: +1. Begin Walking on the Sidewalk: Start by walking along the sidewalk to your left side, which is lined with a wall. +2. Proceed Down the Street: Continue walking straight down the center of the street as it stretches ahead. +3. Stay Aware of Surroundings: Be mindful of any vehicles that may be moving or parked closely to you. Pay attention to pedestrians who might also be using the sidewalk. +4. Turn Left If Necessary: If there are no obstructions or other signs the driver turning left at a junction where there is street forks. + +By following this guidance, you can safely walk through the main road and enter the main flow of traffic causing delays for other pedestrians. + +![](images/c175dcebef508ff41d734bc369c9fe39821f772d931456505d43cfad08683c30.jpg) +Figure 9. Visualization comparison of different VLM models. Compared to other models, WalkVLM is able to generate concise and informative answers, providing users with a good experience in blind walking. Refer to Appendix 3 for more visualization comparisons. + +Text in red is relevant objects, and Text with strikethrough is irrelevant or erroneous descriptions. + +fully capture the semantics of the GT in QA tasks. + +Closed-loop Analysis. We calculate the performance of WalkVLM when it is actually used for the VIPs. Evaluations were conducted on multiple roads unfamiliar to the BLVs, and then the "imminent crash rate" was compared before and after the deployment of the WalkVLM API. + +Table 6 presents the number of manual reminders required for three VIPs with varying degrees of visual impairment while traveling from point A to point B, and the comparative experiment evaluates the performance with and without WalkVLM, with experiments conducted first with WalkVLM to elim + +
p1p2p3
w/242228
w/o353337
+ +Table 6. Comparison of the number of manual reminders required by VIPs during walking when using and not using WalkVLM. + +inate potential route memory bias. In the experiment, the number of manual reminders required for the group with walkvlm was significantly less than that of the other experimental group, which fully demonstrates that our model can effectively support blind walking. + +# 5.5. Ablative Study + +Table 7 shows the ablation study of WalkVLM, to evaluate the effectiveness of CoT-based hierarchical planning (CHP) and POLM prior. We conducted three sets of ablation experiments: (a) w/o CHP. Remove the CHP mechanism and generate reminders directly from the input visual information. (b) w/o Pos Prior. Eliminates the approximate position of significant obstacles in POLM. (c) w/o POLM Prior. Excludes the filtered target's exact location and category as input. In these experiments, when the CHP mech- + +
ConfigurationTF-IDFROUGE-1ROUGE-2ROUGE-L
w/o CHP0.0940.0730.0070.066
w/o Pos Prior0.1510.1890.0620.171
w/o POLM Prior0.1520.1780.0560.164
Full0.1660.1910.0620.173
+ +Table 7. Ablation study on reminder task. CHP stands for CoT-based hierarchical planning, Pos Prior represents the general area where obstacles are located in POLM, and POLM Prior is the pixel point where the filtered target is exactly located. + +anism was removed, the performance degradation was significant, which may be due to the model's inability to fully perceive the scene, resulting in the discrepancies between the generated reminder distribution and the GT distribution. The inclusion of CHP allows the model to conduct more detailed analysis from static attributes and scene summaries, thereby obtaining more concise and accurate results. When removing POLM prior, the model's ROUGE performance is worse compared to lacking position prior, indicating that the model relies heavily on the visual details. + +# 6. Conclusion + +This paper introduces WalkVLM, a vision-language model to offer walking guidance for visual-impaired individuals, along with a diverse and extensive benchmark for the task. By utilizing the chain of thought for hierarchical planning, WalkVLM is able to generate concise yet effective guidance. In addition, the usage of temporal-aware adaptive prediction further improves the efficiency to facilitate real-time processing. Comprehensive experiments demonstrated the usage of the benchmark and the effectiveness of WalkVLM over other VLMs. + +# References + +[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023.6 +[2] Md Adnan Arefeen, Biplob Debnath, Md Yusuf Sarwar Uddin, and Srimat Chakradhar. Vita: An efficient video-to-text algorithm using vlm for rag-based video analysis system. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2266-2274, 2024. 1 +[3] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond. arXiv preprint arXiv:2308.12966, 2023. 3 +[4] Florian Bordes, Richard Yuanzhe Pang, Anurag Ajay, Alexander C Li, Adrien Bardes, Suzanne Petryk, Oscar Manas, Zhiqiu Lin, Anas Mahmoud, Bargav Jayaraman, et al. An introduction to vision-language modeling. arXiv preprint arXiv:2405.17247, 2024. 1 +[5] Erin Brady, Meredith Ringel Morris, Yu Zhong, Samuel White, and Jeffrey P Bigham. Visual challenges in the everyday lives of blind people. In Proceedings of the SIGCHI conference on human factors in computing systems, pages 2117-2126, 2013. 1 +[6] Shaoxiang Chen, Zequn Jie, and Lin Ma. Llava-mole: Sparse mixture of lora experts for mitigating data conflicts in instruction finetuning mllms. arXiv preprint arXiv:2401.16160, 2024.3 +[7] Maria De Marsico, Chiara Giacanelli, Clizia Giorgia Manganaro, Alessio Palma, and Davide Santoro. Vqask: a multimodal android gpt-based application to help blind users visualize pictures. In Proceedings of the 2024 International Conference on Advanced Visual Interfaces, pages 1-5, 2024. 1 +[8] Yifan Du, Zikang Liu, Junyi Li, and Wayne Xin Zhao. A survey of vision-language pre-trained models. arXiv preprint arXiv:2202.10936, 2022. 1 +[9] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 4 +[10] Nicholas A Giudice, Benjamin A Guenther, Toni M Kaplan, Shane M Anderson, Robert J Knuesel, and Joseph F Cioffi. Use of an indoor navigation system by sighted and blind travelers: Performance similarities across visual status and age. ACM Transactions on Accessible Computing (TACCESS), 13 (3):1-27, 2020. 1 +[11] Danna Gurari, Qing Li, Abigale J Stangl, Anhong Guo, Chi Lin, Kristen Grauman, Jiebo Luo, and Jeffrey P Bigham. Vizwiz grand challenge: Answering visual questions from blind people. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3608-3617, 2018. 1, 2, 4 + +[12] Marion A Hersh and Michael A Johnson. Assistive technology for visually impaired and blind people. Springer, 2008. 1 +[13] Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Welihinda, Alan Hayes, Alec Radford, et al. Gpt-4o system card. arXiv preprint arXiv:2410.21276, 2024. 6, 7 +[14] Md Touhidul Islam, Imran Kabir, Elena Ariel Pearce, Md Alimoor Reza, and Syed Masum Billah. A dataset for crucial object recognition in blind and low-vision individuals' navigation. arXiv preprint arXiv:2407.16777, 2024. 2 +[15] Md Touhidul Islam, Imran Kabir, Elena Ariel Pearce, Md Alimoor Reza, and Syed Masum Billah. Identifying crucial objects in blind and low-vision individuals' navigation. arXiv preprint arXiv:2408.13175, 2024. 4 +[16] Gaurav Jain, Yuanyang Teng, Dong Heon Cho, Yunhao Xing, Maryam Aziz, and Brian A Smith. "I want to figure things out": Supporting exploration in navigation for people with visual impairments. Proceedings of the ACM on Human-Computer Interaction, 7(CSCW1):1-28, 2023. 1 +[17] Yizhang Jin, Jian Li, Yexin Liu, Tianjun Gu, Kai Wu, Zhengkai Jiang, Muyang He, Bo Zhao, Xin Tan, Zhenye Gan, et al. Efficient multimodal large language models: A survey. arXiv preprint arXiv:2405.10739, 2024. 2 +[18] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 2 +[19] Askat Kuzdeuov, Olzhas Mukayev, Shakhizat Nurgaliyev, Alisher Kunbolsyn, and Huseyin Atakan Varol. Chatgpt for visually impaired and blind. In 2024 International Conference on Artificial Intelligence in Information and Communication (ICAIIC), pages 722-727. IEEE, 2024. 1 +[20] Chunyuan Li, Cliff Wong, Sheng Zhang, Naoto Usuyama, Haotian Liu, Jianwei Yang, Tristan Naumann, Hoifung Poon, and Jianfeng Gao. Llava-med: Training a large language-and-vision assistant for biomedicine in one day. Advances in Neural Information Processing Systems, 36, 2024. 3 +[21] Chin-Yew Lin. Rouge: A package for automatic evaluation of summaries. In Text summarization branches out, pages 74-81, 2004. 6 +[22] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning, 2023. 2 +[23] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 6, 7 +[24] Ruiping Liu, Jiaming Zhang, Kunyu Peng, Junwei Zheng, Ke Cao, Yufan Chen, Kailun Yang, and Rainer Stiefelhagen. Open scene understanding: Grounded situation recognition meets segment anything for helping people with visual impairments. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1857-1867, 2023. 1, 2 +[25] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In Proceedings of the IEEE/CVF conference on com + +puter vision and pattern recognition, pages 11976-11986, 2022. 6 +[26] Haoyu Lu, Wen Liu, Bo Zhang, Bingxuan Wang, Kai Dong, Bo Liu, Jingxiang Sun, Tongzheng Ren, Zhuoshu Li, Hao Yang, et al. Deepseek-vl: towards real-world vision-language understanding. arXiv preprint arXiv:2403.05525, 2024. 6, 7 +[27] Zain Merchant, Abrar Anwar, Emily Wang, Souti Chattopadhyay, and Jesse Thomason. Generating contextually-relevant navigation instructions for blind and low vision people. arXiv preprint arXiv:2407.08219, 2024. 1, 2, 3, 4 +[28] Juan Ramos et al. Using tfidf to determine word relevance in document queries. In Proceedings of the first instructional conference on machine learning, pages 29-48. Cite-seer, 2003. 6 +[29] Santiago Real and Alvaro Araujo. Navigation systems for the blind and visually impaired: Past work, challenges, and open problems. Sensors, 19(15):3404, 2019. 1 +[30] Machel Reid, Nikolay Savinov, Denis Teptyashin, Dmitry Lepikhin, Timothy Lillicrap, Jean-baptiste Alayrac, Radu Soricut, Angeliki Lazaridou, Orhan First, Julian Schrittweiser, et al. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context. arXiv preprint arXiv:2403.05530, 2024. 3 +[31] Wenhao Shi, Zhiqiang Hu, Yi Bin, Junhua Liu, Yang Yang, See-Kiong Ng, Lidong Bing, and Roy Ka-Wei Lee. Math-llava: Bootstrapping mathematical reasoning for multimodal large language models. arXiv preprint arXiv:2406.17294, 2024.3 +[32] Wu Tang, De-er Liu, Xiaoli Zhao, Zenghui Chen, and Chen Zhao. A dataset for the recognition of obstacles on blind sidewalk. Universal Access in the Information Society, 22 (1):69–82, 2023. 2, 4 +[33] Shishun Tian, Minghuo Zheng, Wenbin Zou, Xia Li, and Lu Zhang. Dynamic crosswalk scene understanding for the visually impaired. IEEE transactions on neural systems and rehabilitation engineering, 29:1478-1486, 2021. 2 +[34] Xiaoyu Tian, Junru Gu, Bailin Li, Yicheng Liu, Yang Wang, Zhiyong Zhao, Kun Zhan, Peng Jia, Xianpeng Lang, and Hang Zhao. Drivevlm: The convergence of autonomous driving and large vision-language models. arXiv preprint arXiv:2402.12289, 2024. 3 +[35] Hao Wang, Jiayou Qin, Ashish Bastola, Xiwen Chen, John Suchanek, Zihao Gong, and Abolfazl Razi. Visiongpt: Llm-assisted real-time anomaly detection for safe visual navigation. arXiv preprint arXiv:2403.12415, 2024. 1, 2, 4 +[36] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Yang Fan, Kai Dang, Mengfei Du, Xuancheng Ren, Rui Men, Dayiheng Liu, Chang Zhou, Jingren Zhou, and Junyang Lin. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 6, 7 +[37] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824-24837, 2022. 5 + +[38] Haiying Xia, Cong Yao, Yumei Tan, and Shuxiang Song. A dataset for the visually impaired walk on the road. Displays, 79:102486, 2023. 1, 2, 4 +[39] Jingyi Xie, Rui Yu, He Zhang, Sooyeon Lee, Syed Ma sum Billah, and John M Carroll. Emerging practices for large multimodal model (lmm) assistance for people with visual impairments: Implications for design. arXiv preprint arXiv:2407.08882, 2024. 1 +[40] An Yang, Baosong Yang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Zhou, Chengpeng Li, Chengyuan Li, Dayiheng Liu, Fei Huang, et al. Qwen2 technical report. arXiv preprint arXiv:2407.10671, 2024. 6 +[41] Bufang Yang, Lixing He, Kaiwei Liu, and Zhenyu Yan. Vassist: Adapting multi-modal large language models for users with visual impairments. arXiv preprint arXiv:2404.02508, 2024. 1, 2, 3 +[42] Zongming Yang, Liang Yang, Liren Kong, Ailin Wei, Jesse Leaman, Johnell Brooks, and Bing Li. Seeway: Vision-language assistive navigation for the visually impaired. In 2022 IEEE International Conference on Systems, Man, and Cybernetics (SMC), pages 52-58. IEEE, 2022. 1 +[43] Yuan Yao, Tianyu Yu, Ao Zhang, Chongyi Wang, Junbo Cui, Hongji Zhu, Tianchi Cai, Haoyu Li, Weilin Zhao, Zhihui He, et al. Minicpm-v: A gpt-4v level mllm on your phone. arXiv preprint arXiv:2408.01800, 2024. 3, 6, 7 +[44] Alex Young, Bei Chen, Chao Li, Chengen Huang, Ge Zhang, Guanwei Zhang, Heng Li, Jiangcheng Zhu, Jianqun Chen, Jing Chang, et al. Yi: Open foundation models by 01. ai. arXiv preprint arXiv:2403.04652, 2024. 6, 7 +[45] Jingyi Zhang, Jiaxing Huang, Sheng Jin, and Shijian Lu. Vision-language models for vision tasks: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 1, 2 +[46] Xingli Zhang, Lei Liang, Shenglu Zhao, and Zhihui Wang. Grfb-unet: A new multi-scale attention network with group receptive field block for tactile paving segmentation. Expert Systems with Applications, 238:122109, 2024. 2 +[47] Yuan Zhang, Chun-Kai Fan, Junpeng Ma, Wenzhao Zheng, Tao Huang, Kuan Cheng, Denis Gudovskiy, Tomoyuki Okuno, Yohei Nakata, Kurt Keutzer, et al. Sparsevm: Visual token sparsification for efficient vision-language model inference. arXiv preprint arXiv:2410.04417, 2024. 1 +[48] Yi Zhao, Yilin Zhang, Rong Xiang, Jing Li, and Hillming Li. Vialm: A survey and benchmark of visually impaired assistance with large models. arXiv preprint arXiv:2402.01735, 2024. 1, 2, 3 +[49] Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric Xing, et al. Judging llm-as-a-judge with mt-bench and chatbot arena. Advances in Neural Information Processing Systems, 36:46595-46623, 2023. 6 +[50] Xingyi Zhou, Rohit Girdhar, Armand Joulin, Philipp Krahenbuhl, and Ishan Misra. Detecting twenty-thousand classes using image-level supervision. In ECCV, 2022. 3, 6 +[51] Yichen Zhu, Minjie Zhu, Ning Liu, Zhicai Ou, Xiaofeng Mou, and Jian Tang. Llava-φ: Efficient multi-modal assistant with small language model. arXiv preprint arXiv:2401.02330, 2024. 3 \ No newline at end of file diff --git a/ICCV/2025/WalkVLM_ Aid Visually Impaired People Walking by Vision Language Model/images.zip b/ICCV/2025/WalkVLM_ Aid Visually Impaired People Walking by Vision Language Model/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..b4959e3b3b7ee93ce09fb18b7fa979a65df7d1bf --- /dev/null +++ b/ICCV/2025/WalkVLM_ Aid Visually Impaired People Walking by Vision Language Model/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e71bb114a31de7a920a9698c87f99523316e5600405ef8d17a502c998f11caf7 +size 511299 diff --git a/ICCV/2025/WalkVLM_ Aid Visually Impaired People Walking by Vision Language Model/layout.json b/ICCV/2025/WalkVLM_ Aid Visually Impaired People Walking by Vision Language Model/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..3cf69f4823360b10c679b9545cab3215d07d8ffa --- /dev/null +++ b/ICCV/2025/WalkVLM_ Aid Visually Impaired People Walking by Vision Language Model/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c3f76cf6e7dd1b95b8d4fccec4e7ff1d25172b503eceaa5dc6c4a018463eba58 +size 402160 diff --git a/ICCV/2025/WarpHE4D_ Dense 4D Head Map toward Full Head Reconstruction/f316d5e0-ad4b-4e9b-ba4d-1c0b8a170fec_content_list.json b/ICCV/2025/WarpHE4D_ Dense 4D Head Map toward Full Head Reconstruction/f316d5e0-ad4b-4e9b-ba4d-1c0b8a170fec_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..8f2fbf6485b80c2d0d3337874ddcaefa51112e81 --- /dev/null +++ b/ICCV/2025/WarpHE4D_ Dense 4D Head Map toward Full Head Reconstruction/f316d5e0-ad4b-4e9b-ba4d-1c0b8a170fec_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b3394f7107fb4962420432d72d50eba21710a05868861fca46129d3ab58c991e +size 98423 diff --git a/ICCV/2025/WarpHE4D_ Dense 4D Head Map toward Full Head Reconstruction/f316d5e0-ad4b-4e9b-ba4d-1c0b8a170fec_model.json b/ICCV/2025/WarpHE4D_ Dense 4D Head Map toward Full Head Reconstruction/f316d5e0-ad4b-4e9b-ba4d-1c0b8a170fec_model.json new file mode 100644 index 0000000000000000000000000000000000000000..69a27108045a76c24c9f2a2e8f566b2b3eff4280 --- /dev/null +++ b/ICCV/2025/WarpHE4D_ Dense 4D Head Map toward Full Head Reconstruction/f316d5e0-ad4b-4e9b-ba4d-1c0b8a170fec_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c76451f24d67faaf34e5b808f60afdfad03a6fa43805679f14923ccf35d57312 +size 117790 diff --git a/ICCV/2025/WarpHE4D_ Dense 4D Head Map toward Full Head Reconstruction/f316d5e0-ad4b-4e9b-ba4d-1c0b8a170fec_origin.pdf b/ICCV/2025/WarpHE4D_ Dense 4D Head Map toward Full Head Reconstruction/f316d5e0-ad4b-4e9b-ba4d-1c0b8a170fec_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..c0da50526dbf89286461b7f65bfa67765ffd35a2 --- /dev/null +++ b/ICCV/2025/WarpHE4D_ Dense 4D Head Map toward Full Head Reconstruction/f316d5e0-ad4b-4e9b-ba4d-1c0b8a170fec_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3ddbcfff46524a7828fb553d51f8703400522feed30fbef405be4cc6611fdc87 +size 9292853 diff --git a/ICCV/2025/WarpHE4D_ Dense 4D Head Map toward Full Head Reconstruction/full.md b/ICCV/2025/WarpHE4D_ Dense 4D Head Map toward Full Head Reconstruction/full.md new file mode 100644 index 0000000000000000000000000000000000000000..dee6b7a137a40141428cfe230d474078b6b8aee6 --- /dev/null +++ b/ICCV/2025/WarpHE4D_ Dense 4D Head Map toward Full Head Reconstruction/full.md @@ -0,0 +1,450 @@ +# WarpHE4D: Dense 4D Head Map toward Full Head Reconstruction + +Jongseob Yun $^{1,2}$ Yong-Hoon Kwon $^{2,3}$ Min-Gyu Park $^{2,3}$ Ju-Mi Kang $^{2,3}$ + +Min-Ho Lee2,3 Inho Chang2,3 Ju Hong Yoon2,3 Kuk-Jin Yoon1, $\dagger$ + +$^{1}$ Korea Advanced Institute of Science and Technology (KAIST) $^{2}$ Polygom + +$^{3}$ Korea Electronics Technology Institute (KETI) + +{jseob.y,kjyoon}@kaist.ac.kr，{yhkwon,mpark,amino7,yypeip,damonchang23,jhyoon}@keti.re.kr + +![](images/7477d579daecc5fae8d9371f56fc965a6ec703d3e42b6fd59491810923dbf390.jpg) +Figure 1. WarpHE4D jointly predicts pixel-aligned depth and UV coordinates, enabling the reconstruction of a colored pointcloud with UV mapping. This representation is well-suited for downstream tasks such as template mesh registration and texture map optimization. + +# Abstract + +We address the 3D head reconstruction problem and the facial correspondence search problem in a unified framework, named as WarpHE4D. The underlying idea is to establish correspondences between the facial image and the fixed UV texture map by exploiting powerful self-supervised visual representations, i.e., DINOv2. In other words, we predict UV coordinates for each pixel that maps the pixel to a point in the UV map. At the same time, we predict the nose-centered depth map leveraged by the facial correspondences. Note that our framework does not require fitting a template model, e.g., 3DMM, to the image, which directly regresses 4D vectors for each pixel. The experimental results show that our approach not only improves the accuracy of head geometry but also significantly improves the robustness under pose or viewpoint variations, particularly when the head is rotated more than 90 degrees. We believe our method can be a groundwork for photorealistic head avatar generation, even in uncalibrated camera settings. The code is available at github.com/jseobyun/WarpHE4D. + +# 1. Introduction + +The reconstruction of 3D human head geometry serves as the groundwork for follow-up research such as facial animation and AI-driven content generation. Recently, high-level + +attributes, semantic labels and UV coordinates, have gained significance in enabling advanced representations [36, 46] and manipulations [31]. Notably, 3DMM [18, 27] has been widely used as a geometric prior. However, we claim existing approaches still have two major challenges. First, statistical head models oversimplify non-facial regions, i.e., hairs, because of their diversity. Second, post-processing is often necessary to align resulting geometry with image pixels, relying on additional guides such as 2D facial landmarks or head pose. The alignment error introduced in this process restricts the utilization of geometry under perspective projection, especially in multi-view settings. + +In this paper, we revisit 3D head geometry reconstruction as a per-pixel regression problem to alleviate the aforementioned issues. Specifically, we define the human head as a 4D vector map representation, defined as the 4D head map, where each pixel in an image retains the UV coordinates along with nose-centered depth and confidence. Notably, this simple transition not only enables seamless integration of multiple images under perspective projection without alignment but also avoids cumbersome post-processing procedures. For example, by merging multi-view outputs with camera parameters, a full head point cloud can be generated. Since each point retains the UV coordinates of the 3DMM, it can be converted into a topology-unified mesh, as illustrated in Figure 1. Unlike previous method [39] that relies on parametric fitting, our approach enables non-rigid + +fitting beyond the 3DMM space. Moreover, our 4D head map directly maps image pixels to the UV coordinates, simplifying the texture map generation procedure. By leveraging the output UV coordinates along with the confidence maps, our method enables confidence-based texture map optimization when multi-view inputs are given, instead of a simple averaging approach. + +Additionally, we propose a novel framework built upon self-supervised visual features, serving as an alternative to explicit 3DMM priors. Our key observation, as shown in Figure 2, is that DINOv2 [33] consistently extracts head features regardless of head pose, occlusion, or even severe distortions in the unwrapped texture map. Similar evidence has been reported in [49], which analyzes the 3D consistency of DINOv2, and in [12, 30, 44, 50, 53], where DINOv2 has constructed 3D universal features. Assuming DINOv2 is a strong feature extractor for both semantic and geometric priors, we design our method with two encoders and two decoders, where both encoders include DINOv2. Specifically, each encoder constructs a feature pyramid, where the coarsest feature is derived from pre-trained DINOv2, while finer features are extracted from the learnable convolution layers. One encoder processes the input image, while another encodes a fixed head texture map as a guide. The decoders then align and warp features from the UV domain to the image domain to infer the final 4D head map in a coarse-to-fine manner. Leveraging self-supervised features, our method not only enhances robustness while reducing reliance on costly 3D data and 3DMM as shown in Fig. 7) but also outperforms the state-of-the-art across various tasks. + +In summary, our key contributions are as follows. + +1. We propose a novel dense 4D head map for 3D head reconstruction, enabling seamless integration under perspective projection and avoiding post-processing. +2. We present a framework based on self-supervised features, improving robustness and reducing reliance on 3DMM, resulting in superior performance across different domains. + +# 2. Related Work + +# 2.1. Self-Supervised Visual Features + +Self-supervised learning (SSL) has emerged as a powerful paradigm for learning robust visual representations without requiring extensive labeled data, significantly impacting 3D reconstruction tasks. Early foundational methods laid the groundwork for SSL. DINO [8] pioneered distillation-based SSL, demonstrating emergent properties in vision transformers, while iBOT [52] enhanced self-distillation learning by introducing online tokenization for more effective feature extraction. MAE [22] proposed reconstructing masked image patches through a self-supervised + +![](images/c675a1d0d352f9846b226fb783cd4c25b2f49451fd50ee9260fbbcc94d0a3142.jpg) + +![](images/23dc78e6f76637bb1b31fe338d917c632714126fa1745e9a357b1bd5bb05016d.jpg) +Figure 2. PCA visualization of DINOv2 features. The top row represents input images, while the bottom row shows the corresponding DINOv2 [33] features. These features are reduced to three dimensions using PCA and visualized as RGB. DINOv2 effectively extracts features that distinguish between heads and objects and maintains consistent head features even from the unwrapped FLAME texture map. Note that similar colors indicate high feature similarity. + +![](images/1ca34765d5629bfa991ffd1550c76183292549e37e73aeadca5ba31771550d23.jpg) + +![](images/89a5848abf3a520de53092d885ec8e7a6e8352911e766e9f8d710afe93f3d256.jpg) + +encoder-decoder framework, enabling scalable vision learners. Multi-MAE [3] highlighted the advantages of masked modeling by leveraging multi-modal inputs. I-JEPA [2] explored masking strategies within a joint-embedding predictive architecture, revealing further potential of masked modeling. DINOv2 [33] represents a significant breakthrough, delivering exceptional feature extraction for 3D-aware and multi-view tasks through large-scale self-supervised training. Studies such as [49] showcase DINOv2's multi-view equivariance, enhancing 3D correspondence with minimal fine-tuning, while [12, 30, 44, 49, 50, 53] highlight its ability to generate universal 3D features. For example, Diff3F [12] distilled 3D semantic features from DINOv2 representations for textured shapes, while TRELLIS [44] directly utilized a DINOv2-based feature volume as input to a shape VAE. These advancements underscore SSL's growing significance in 3D vision, offering a robust alternative to supervised priors like 3DMMs. Our work leverages DINOv2's 3D consistency and semantic depth to bypass explicit model priors, embracing this shift toward flexible, data-efficient learning. + +# 2.2. Dense Face Correspondences + +Dense face correspondence estimation seeks to establish detailed point-to-point mappings across facial surfaces, extending sparse landmarks to capture full 3D topology. This task is inherently challenging due to variations in facial geometry, occlusions, and topological differences, with approaches broadly categorized into 3D and 2D mappings. For 3D correspondence, non-rigid iterative closest point (ICP) methods between a 3D template and a scan have been the dominant approach. While Amberg et al. [1] introduced local affine registration, Li et al. [26] later improved upon this with a global optimization scheme incorporating + +confidence weights. Gilani et al. [20] and Pan et al. [34] used regional segmentation for dense matching, and Fan et al. [15] introduced an iterative seed-based method for dense registration, combining progressive augmentation with local Gaussian-weighted alignment. More recently, learning-based approaches such as NICP [32] and ETCH [25] have emerged, enabling data-driven dense registration. In 2D correspondence, methods establish mappings from image pixels to a 3D template mesh that acts as an intermediate representation. Early approaches focused on aligning a 3D template to the input image, as in Blanz et al. [7], and Patel et al. [35] later improved non-rigid alignment using thinplate spline warping. Recent advances, such as position map regression in PRNet [16] and iterative UV-to-image flow in FlowFace [39], directly estimate dense UV coordinates of the template in the pixel space. Our approach adopts the spirit of [39] but enhances the matching between UV and image domain by self-supervised features, i.e., DINOv2. Additionally, by jointly estimating depth, we extend the framework to facilitate direct 2D-to-3D correspondence without requiring template fitting. + +# 2.3. Face Reconstruction + +Face reconstruction has evolved significantly, driven by deep learning and multi-cue integration, aiming for high-fidelity 3D geometry and texture from limited inputs. Early methods [10, 37] regressed 3D face shapes from single images, jointly optimized geometry and alignment. Xu et al. [45] and 3DDFAv2 [21] enhanced shape recovery with deep portrait synthesis and landmark regression regularization. TokenFace [51] leveraged a transformer architecture and self-supervision scheme for precise reconstruction, and HRN [24] utilized progressive refinement networks. 3DDFAv3 [42] optimized real-time performance with geometric guidance from facial part segmentation. Recently, beyond the traditional perspective, Chandran et al. [9] proposed infinite-resolution landmarks for continuous 2D-to-3D mapping, moving beyond the limitations of template model space. However, full-head reconstruction still remains a challenge, as previous baselines, e.g., DECA [17] and EMOCA [11], often oversimplify non-facial regions such as ears and hair. Our WarpHE4D framework departs from these approaches by treating head reconstruction as a per-pixel regression problem and converts each pixel into a 4D vector. By converting visible head pixels into a pointcloud with UV coordinates, we aim for full-head reconstruction. + +# 3. The Proposed Method + +# 3.1. Dense 4D Head Map Representation + +We predict head geometry as a dense 4D vector map $\mathbf{W} \in \mathbb{R}^{h \times w \times 4}$ , which integrates the FLAME [27] UV coordi + +![](images/eec2e88d1defa4ce05a6af70111811cacbbf0ab5f4910222f7a2c6765cc46743.jpg) +Figure 3. An overview of WarpHE4D architecture. Two encoders that include DINOv2 extract descriptive features, and the coarse decoder predicts an initial head map from only DINOv2 features. By leveraging fine features, the fine decoder refines and upsamples the initial head map to a high-resolution head map. + +nates with a nose-centered depth, and confidence per pixel in the image where $h$ and $w$ denote the image height and width, respectively. $\mathbf{W}$ can be decomposed into three components, $\mathbf{W}_{\mathrm{UV}}$ , $\mathbf{W}_{\mathrm{D}}$ , and $\mathbf{W}_{\mathrm{C}}$ , corresponding to UV coordinates, the depth map, and the confidence map, respectively. The UV coordinates $\mathbf{W}_{UV} \in \mathbb{R}^{h \times w \times 2}$ transform the input image into the UV space, reversing the UV-to-image mapping described in [16]. The depth map $\mathbf{W}_D \in \mathbb{R}^{h \times w \times 1}$ is convertible to a 3D point cloud via back-projection and the confidence map $\mathbf{W}_C \in \mathbb{R}^{h \times w \times 1}$ contains the confidence value for each pixel, that is, the likelihood of a pixel belonging to head regions. For notational efficiency, we also define $\mathbf{W}_{UVD}$ which combines $\mathbf{W}_{UV}$ and $\mathbf{W}_D$ . As a result, the prediction of the dense 4D head map not only reconstructs the 3D head geometry of the visible head regions but also assigns UV coordinates, providing dense correspondences between the FLAME model and image pixels. In the rest of the paper, we refer to our representation as the 4D head map. This simple yet effective representation offers clear advantages in multi-view scenarios, as it can be directly used with camera parameters in a pixel-aligned manner. + +Nose-centered depth normalization of $\mathbf{W}_D$ It is important to normalize the range of depth values to be consistent regardless of the positions of the head. Popularly used approaches are scale invariant [14, 28, 48] and average depth-based [41] formulations. However, these approaches suffer from a non-uniform scale, even for the same subject, as they fundamentally rely on the min-max depth values of the head. For example, the same person with different hairstyles would be normalized differently depending on the viewpoint. We claim these non-uniform scaling approaches can degrade the performance of head geometry reconstruction. To alleviate the non-uniform scaling problem, we propose a nose-centered depth normalization defined as + +$$ +D _ {n o r m} = \left(\frac {D _ {r a w}}{d _ {n o s e}} - 1. 0\right) \times \frac {f o c a l}{w i d t h \times 0 . 5}, \tag {1} +$$ + +![](images/5787214d5a6570718879c2c63e029df7e8872a3bf4051d3b271e0e396462f5e8.jpg) +Figure 4. An illustration of nose-centered depth normalization. The raw depth of the head is normalized using camera projection to bring the nose tip onto the image plane (stage1), then normalized again so the image plane size is [-1, 1] after being shifted by -1 (stage2). Note that the normalized depth can be smaller than that of the normalized image plane when the head is in a side view. + +where $D_{norm}$ and $D_{raw}$ denote normalized and raw depth of head. $d_{nose}$ indicates the depth of the nose tip. The first term, $D_{raw}/d_{nose}$ , scales $D_{raw}$ making the nose tip touch the normalized image plane, e.g., $z = 1.0$ . Then, $D_{raw}/d_{nose}$ is normalized again with focal length and image width after being shifted by -1. Here, the nose tip is always at the zero-centered plane i.e., $z = 0.0$ . Furthermore, when the image is not square, we adjust it to a square shape by centering the principal point through padding and affine transformation, and update the camera parameters accordingly. Figure 4 describes the nose-centered normalization procedure, which firstly scales the depth map, denoted as Stage 1, and normalizes the scaled depth map, i.e., Stage 2. As a result, our normalization maintains a consistent scale per subject regardless of head size, hairstyles, and camera parameters while ensuring depth values proportional to head size in the image. Hence, the range of depth values becomes similar to the range of UV coordinates, making the dense head 4D prediction more effective. + +# 3.2. WarpHE4D + +Based on key observations as shown in Figure 2, our goal is to predict the 4D head map by leveraging the consistency of DINOv2 features extracted from images and UV texture maps. Our key insight is that image-to-UV mapping can be inferred by searching correspondences between image and UV texture features, similar to FlowFace [39]. Additionally, the 3D-awareness [12, 30, 44, 50, 53] of DINOv2 provides a strong foundation for nose-centered depth estimation. We employ a dense matching framework [13] between image and texture map as a baseline and extend its coarse-to-fine decoder to jointly handle the nose-centered depth, as shown in Figure 3. + +Image encoder $\mathcal{E}_{img}$ consists of two pre-trained networks, + +DINOv2 [33] and VGG19 [38], following [13]. DINOv2 focuses on capturing coarse features, while VGG19 remains trainable and is optimized for extracting fine-grained details. The image encoder $\mathcal{E}_{img}$ is defined as + +$$ +\mathbf {C} _ {i m g}, \mathbf {F} _ {i m g} = \mathcal {E} _ {i m g} (\mathbf {I}), \tag {2} +$$ + +where $\mathbf{C}_{img}$ and $\mathbf{F}_{img}$ denote the coarse and fine features from the DINOv2 and VGG19 networks, respectively, with $\mathbf{I}$ as the head input image. + +UV encoder $\mathcal{E}_{uv}$ extracts coarse and fine features from the UV head texture map. It shares the same structure as the image encoder but has distinct weights. Since the input texture map only serves as a visual guide for extracting DINOv2 features from the UV domain, we use the same head texture map. Specifically, we select and average samples from FFHQ-UV dataset [4], which follows the FLAME UV mapping defined in HiFi3D [5]. This approach is similar to UV positional embeddings [39], but offers more descriptive power by incorporating DINOv2 and fine VGG features. The UV encoder $\mathcal{E}_{uv}$ is defined as + +$$ +\mathbf {C} _ {u v}, \mathbf {F} _ {u v} = \mathcal {E} _ {u v} (\mathbf {T}), \tag {3} +$$ + +where $\mathcal{E}_{uv}$ takes the UV texture map $\mathbf{T}$ as an input and outputs $\mathbf{C}_{uv}$ and $\mathbf{F}_{uv}$ as the coarse and fine UV texture features, respectively. + +Coarse decoder $\mathcal{D}_{\mathcal{C}}$ utilizes vision transformer (ViT) and takes two coarse features $\mathbf{C}_{img}$ and $\mathbf{C}_{uv}$ . These features are concatenated to form the decoder input, which is then decoded into an initial 4D head map $\mathbf{W}^0$ . The $\mathcal{D}_{\mathcal{C}}$ predicts $\mathbf{W}_{UVD}^0$ by discretizing each dimension into $B = 64$ bins and producing the corresponding probability distributions, following the formulation introduced in [13]. The coarse decoder $\mathcal{D}_{\mathcal{C}}$ is defined as + +$$ +\mathbf {Q}, \mathbf {W} _ {C} ^ {0} = \mathcal {D} _ {\mathcal {C}} \left(C _ {i m g}, C _ {u v}\right), \tag {4} +$$ + +where $\mathbf{Q} \in \mathbb{R}^{\frac{h}{14} \times \frac{w}{14} \times 3B}$ indicates quantized bins of $\mathbf{W}_{UVD}^{0}$ with class probabilities and superscript 0 indicates the coarse level that is further refined with the fine decoder. Note that the factor of 14 arises from the patch size of DINOv2. The $3B$ channels are split into three groups of $B$ bins for $U, V$ , and $D$ maps, + +$$ +\mathbf {Q} = \left[ \mathbf {Q} _ {U}, \mathbf {Q} _ {V}, \mathbf {Q} _ {D} \right], \quad \mathbf {Q} _ {U}, \mathbf {Q} _ {V}, \mathbf {Q} _ {D} \in \mathbb {R} ^ {\frac {h}{1 4} \times \frac {w}{1 4} \times B}, \tag {5} +$$ + +where $\mathbf{Q}_U, \mathbf{Q}_V,$ and $\mathbf{Q}_D$ correspond to the quantized $B$ -bin groups for head map except confidence, respectively. Since the initial confidence map $\mathbf{W}_C^0$ is given from the coarse decoder, we compute the rest of the initial head map $\mathbf{W}_{UVD}^0$ , + +$$ +\mathbf {W} _ {U V D} ^ {0} = \sum_ {i \in \{U, V, D \}} \sum_ {j = 1} ^ {B} p _ {i} (j) \cdot q _ {i} (j), \tag {6} +$$ + +where $p_i(j)$ is the softmax probability of the $j$ -th bin of $Q_i$ and $q_i(j)$ is the corresponding bin center for $i \in \{U, V, D\}$ . Fine decoder $\mathcal{D}_F$ progressively refines initial head map $\mathbf{W}^0$ and increases the spatial resolution of $\mathbf{W}^0$ from $\left(\frac{h}{14}, \frac{w}{14}, 3\right)$ to $(h, w, 3)$ . We define the fine decoder as + +$$ +\mathbf {W} ^ {l + 1} = \mathcal {D} _ {F} ^ {l} (\mathcal {U} (\mathbf {W} ^ {l}), \mathcal {G} (\mathbf {W} _ {U V} ^ {l}, \mathbf {F} _ {u v} ^ {l + 1}), \mathbf {F} _ {i m g} ^ {l + 1}), \quad (7) +$$ + +where $\mathcal{U}$ and $\mathcal{G}$ indicate bilinear upsampling and grid sampling that warps UV features to the image domain, and $l$ indicates the level, $l\in \{0,1,2,3\}$ . The number of levels indicates the four upsampling convolution blocks, $\mathcal{D}_F^l$ , in the decoder. Each block takes the previous head map $\mathbf{W}^l$ , image features $\mathbf{F}_{img}^{l + 1}$ and UV features $\mathbf{F}_{uv}^{l + 1}$ , which are warped into the image domain using $\mathbf{W}_{UV}^l$ . It outputs a finer 4D head map $\mathbf{W}^{l + 1}$ , which includes aligned UV along with depth and confidence. + +# 3.3. Loss Functions + +The overall loss function is defined as follows, + +$$ +E = E _ {\text {c o a r s e}} + E _ {\text {f i n e}} + \lambda_ {k p} E _ {k p} + \lambda_ {c} E _ {\text {c o n f}}. \tag {8} +$$ + +Here, $E_{\text{coarse}}$ and $E_{\text{fine}}$ represent the primary regression losses for the coarse and fine predictions, and $E_{kp}$ and $E_{conf}$ denote additional keypoint and confidence losses, weighted by $\lambda_{kp}$ and $\lambda_c$ . Empirically, we set $\lambda_{kp}$ and $\lambda_c$ to 0.1 and 0.5, respectively. + +Coarse regression loss $E_{\text{coarse}}$ encourages the quantized probability bins from Eq. 5 to align with quantized ground truth probabilities, + +$$ +E _ {\text {c o a r s e}} = \sum_ {i \in \{U, V, D \}} \sum_ {j = 0} ^ {B} \operatorname {B C E} \left(p _ {i} (j), p _ {i} ^ {G T} (j)\right), \tag {9} +$$ + +where $p_i^{GT}(\cdot)$ represents the quantized ground truth probabilities of $j^{\mathrm{th}}$ bin of $Q_{i}$ , and $\mathrm{BCE}(\cdot ,\cdot)$ denotes the binary cross-entropy loss. + +Fine regression loss $E_{fine}$ minimizes the Charbonnier loss with $\alpha$ of 1.0 [6, 13], + +$$ +E _ {f i n e} = \sum_ {l \in \{1, 2, 3, 4 \}} \left(\left(\frac {\mathbf {W} _ {U V D} ^ {l} - \mathbf {W} _ {U V D} ^ {l , G T}}{c}\right) ^ {2} + 1\right) ^ {\frac {1}{2}}, \tag {10} +$$ + +where $\mathbf{W}_{UVD}^{l}$ and $\mathbf{W}_{UVD}^{l,GT}$ denote the predicted and ground truth head maps at each level, excluding the confidence channel. We set $c$ to 1e-3. + +Keypoint loss $E_{kp}$ enhances the accuracy of UV predictions $W_{UV}^{l}$ , particularly around the eyes and mouth regions. Since the supervision signal is derived from FLAME + +meshes registered to 3D scans, inherent noise exists in regions such as the eyes and mouth due to registration difficulties. Consequently, relying solely on regression losses degrades accuracy around the eyes and mouth regions. Therefore, we incorporate keypoint loss by leveraging a state-of-the-art 2D keypoint detector [23]. Specifically, we extract 68 facial keypoints, $\mathbf{p}_{2D}$ , from the input image and normalize them into [-1, 1]. We then sample UV coordinates at these locations and define $E_{kp}$ by minimizing the discrepancy to the predefined UV coordinates of the corresponding FLAME keypoints. + +$$ +E _ {k p} = \sum_ {l \in \{1, 2, 3, 4 \}} \left(\left(\frac {\mathcal {G} \left(\mathbf {p} _ {2 D} , \mathbf {W} _ {U V} ^ {l}\right) - \mathbf {u} ^ {k p}}{c}\right) ^ {2} + 1\right) ^ {\frac {1}{2}}, \tag {11} +$$ + +where grid sampling function $\mathcal{G}$ samples UV predictions corresponding to $\mathbf{p}_{2D}$ , and then, they are compared with ground truth UV coordinates of the keypoints $\mathbf{u}^{kp}$ . $c$ is set to 0.3. + +Confidence loss $E_{conf}$ penalizes predictions in non-head regions, + +$$ +E _ {c o n f} = \sum_ {l \in \{0, 1, 2, 3, 4 \}} \operatorname {B C E} \left(\mathbf {W} _ {C} ^ {l}, \mathbf {W} _ {C} ^ {l, G T}\right). \tag {12} +$$ + +Here, we compute the BCE loss between predicted and ground truth confidence maps. + +# 4. Experimental Results + +# 4.1. Training Dataset + +To construct a large-scale ground truth 4D head map, we gathered photorealistic 3D head scans and processed them into FLAME topology through registration. Specifically, we collected head scans from FaceScape [47] and NPHM [19], resulting in 16,869 and 10,425 FLAME topology meshes, respectively. We subdivided the FLAME topology twice after removing the eyeballs to preserve the details of the scans. Subsequently, we rendered these into a 4D head map by adjusting focal lengths, azimuth, and elevation angles. During rendering, background augmentation was applied by randomly selecting human-free images from the COCO dataset [29]. Note that Multiface [43] and FaceV-erse [40] are not included in our training dataset. + +# 4.2. Benchmarks + +FaceScape FaceScape benchmark is designed to evaluate 3D face reconstruction across diverse identities and expressions. We assess the geometric accuracy of our method, focusing on nose-centered depth quality using it. Notably, we only evaluate similar NoW challenge [37] qualitatively in supplementary material, as it handles only neutralized face, excluding expressions. + +![](images/0a1d7256f63fd5feb5152e1c696b028499dfd0bc3517f15dd346b75312a71222.jpg) + +![](images/1aed34ad2039a463bc0f9f9ad94bdc54b4862d66c5be78ecbec627d3edb6610f.jpg) +Figure 5. A visualization of evaluation protocols. We divide the head region into facial and non-facial regions (top row). For each region, we evaluate the results using the proposed UVAUC and 3DAUC by drawing the curves as the error thresholds increase(bottom row). + +Table 1. Quantitative evaluation on FaceScape benchmark measured by chamfer distance(mm). The numbers are the official reported metrics. + +
MethodFS-WildFS-Lab
0°-5°5°-30°30°-60°60°-90°Avg0°30°60°Avg
PRNet2.643.074.283.883.462.943.403.743.56
Deep3DFaceRec3.214.135.748.905.493.995.905.545.28
3DDFA v22.562.693.223.683.033.053.413.823.60
DECA2.922.652.914.813.323.304.144.204.68
Facescape(Learn)2.623.624.119.084.862.407.273.866.27
HRN----2.91---3.87
Ours2.482.592.633.582.821.391.451.661.44
+ +Multiface Since existing benchmarks only focus on geometric accuracy, we propose the Multiface [43] benchmark for comprehensive evaluation of the 4D head map accuracy of our method. The Multiface dataset provides calibrated images with topology-unified tracked meshes. We accurately registered FLAME to each tracked mesh using vertex correspondence and chamfer distance, then rendered them into 4D head map using the given cameras. Note that the list of selected categories is provided in the supplementary material, with a total number of samples is 54,215. + +FaceVerse We additionally propose FaceVerse [40] benchmark, built similarly to the Multiface Benchmark. Since the MultiFace benchmark includes only 10 among 13 subjects, it lacks subject diversity. Our method requires a larger number of subjects to ensure diverse non-facial regions for evaluation. Therefore, we created a synthetic benchmark using all FaceVerse 3D head scans with 110 subjects. Rendered 175,250 samples are generated with predefined camera parameters(see the suppl.). Notably, since FaceVerse does not provide an official scale, we assume the scans are in mm scale by normalizing the canthal distance to $35\mathrm{mm}$ . + +# 4.3. Evaluation Protocols + +Categorization We primarily separate face and non-facial regions to demonstrate that our method performs effectively on both for the Multiface benchmark, as shown in Fig- + +ure 5(top). The evaluation of facial region is comparable to baselines, while the non-facial regions highlight differences not covered by them. Additionally, we categorize overall-region performance across azimuth and elevation angles to demonstrate robustness to head pose variations for the FaceVerse benchmark. + +Metrics Intuitively, we evaluate the mean pixel UV errors (MPUVE) and mean point 3D distance errors (MP3DE) for each categorization. However, these simple metrics do not capture the coverage region (CR), which is a crucial aspect. These metrics only cover measurable region. Therefore, we propose AUC (area under curve)-based metrics for both UV and 3D errors. As shown in Figure 5(bottom), we plot the ratios of pixels below progressively increasing thresholds relative to the target region. We set the maximum thresholds to 0.1 for UV errors and $0.05\mathrm{m}$ for 3D errors, respectively. Then, we compute the AUC by normalizing the threshold axis to a maximum threshold of 1. These AUCs (UVAUC and 3DAUC) represent the percentage of the entire region within the maximum threshold, serving as a metric that encapsulates both accuracy and CR. + +# 4.4. Implementation details + +All input images are resized to $448 \times 448$ without any resolution augmentation, as most facial images are already tightly cropped. Following [13], we set the learning rate for the encoders to effective batch size $\times$ 1e-5, and for the decoders to effective batch size $\times$ 2e-4. We use the AdamW optimizer with a weight decay of 0.01. The model is trained for 600K iterations on four NVIDIA L40S GPUs in two stages: 500K iterations without keypoint supervision, followed by 100K iterations with keypoint loss. Additionally, we use training images with azimuth angles within $[-90^{\circ}, 90^{\circ}]$ . DINOv2 features show limited descriptive power when the face is completely occluded (e.g., pure back views), whereas the representation remains stable when the facial region is at least partially visible. We observed that this selection contributes to a slight performance gain. + +# 4.5. Quantitative Evaluation Results + +FaceScape As shown in Table 1, our method outperforms other baselines by a noticeable margin under both orthogonal (FS-Wild) and perspective (FS-Lab) projections. This indicates that our method is comparable to other baselines even for the facial region, without relying on a template model. Furthermore, it demonstrates minimal variation across different head poses. The full report of other metrics and categories is shown in the supplementary. + +Multiface In the facial region, our method still achieves comparable performance in terms of UVAUC and 3DAUC, with margins of -0.0209 and +0.0179, respectively, as shown in Table 2. Remarkably, while the baselines were evaluated after the alignment of Procrustes with the GT + +Table 2. Quantitative evaluation on Multiface benchmark categorized by region. + +
Eval. Region → Method ↓3D AlignFaceNon-Facial
UVAUC ↑3DAUC ↑CR ↑MPUVE ↓MP3DE ↓UVAUC ↑3DAUC ↑CR ↑MPUVE ↓MP3DE ↓
PRNet[16]✓0.67060.77100.79680.02380.00460.10180.11730.1164N/AN/A
3DDFA v2[21]✓0.69180.78980.82630.02240.00430.09390.10340.1049N/AN/A
3DDFA v3[42]✓0.69130.76930.83780.02030.00430.10020.11100.1165N/AN/A
HRN[24]✓0.68350.77560.83300.02130.00390.08390.09310.0953N/AN/A
DECA[17]✓0.80730.89360.98510.02020.00480.60220.63170.82680.03410.0120
EMOCA[11]✓0.80060.89160.97070.02090.00490.60070.63470.79200.03360.0119
OursX0.78640.91151.00000.02210.00440.68580.70911.00000.04630.0105
Ours(w/o kp loss)X0.76950.88461.00000.02400.00550.67740.68911.00000.04890.0204
Ours(w/ vgg)X0.52520.67561.00000.15890.01820.24010.39251.00000.43210.0404
Ours(w/ sapiens)X0.12820.13241.00000.24900.07760.02040.11751.00000.52100.1073
+ +Table 3. Quantitative evaluation on Faceverse benchmark categorized by pose angle. + +
Pose Angle → Method ↓Azim. (0°)Azim. (30°)Azim. (60°)Azim. (90°)Elev. (0°)Elav. (15°)Elev. (30°)
UVAUC ↑3DAUC ↑UVAUC ↑3DAUC ↑UVAUC ↑3DAUC ↑UVAUC ↑3DAUC ↑UVAUC ↑3DAUC ↑UVAUC ↑3DAUC ↑UVAUC ↑3DAUC ↑
PRNet0.47770.50170.46290.48140.49190.50820.44070.46640.48070.49760.47320.49290.45610.4793
3DDFA v20.49480.50370.48020.48750.53120.54300.51140.52630.51190.52080.50860.51870.49880.5113
3DDFA v30.51260.53980.49460.51950.52930.56410.49750.53210.51000.53900.50940.53960.50540.5378
HRN0.51460.54110.49110.51640.51920.54920.47960.49910.50190.52640.50260.52780.49560.5213
DECA0.63430.64590.65840.67020.72780.75540.73910.76630.69920.71710.70020.71860.69490.7191
EMOCA0.62010.64380.64540.66640.72210.75380.72940.75900.68990.71420.69000.71530.68390.7136
Ours0.85520.83670.86010.85500.87390.87300.87370.83450.87140.85720.86960.85480.86280.8459
+ +mesh, our method does not benefit from alignment, making the baseline results more favorable. This indicates that our method achieves both accurate surface parameterization and precise geometric reconstruction, even without relying on template priors and alignment. Furthermore, considering that MPUVE and MP3DE are also comparable, it is evident that the performance of our method in UVAUC and 3DAUC is not simply due to a larger CR. In other words, our approach covers a wider area while maintaining a performance level that outperforms existing methods. For the non-facial region, our method outperforms DECA [17] and EMOCA [11] by a notable margin of $+0.085$ and $+0.075$ in UVAUC and 3DAUC, respectively. This highlights its superior ability to capture details in areas where the template model tends to oversimplify. + +FaceVerse Even under large-scale subject diversity and pose-based categorization, our method performs consistently well across all angles, showing a minimum margin of 0.09 and a maximum margin of 0.23 in UVAUC, and a minimum margin of 0.06 and maximum margin of 0.19 in 3DAUC. As shown in Table 3, our method shows consistent performance even in extreme head poses without the need for alignment. Note that this table shows the performance across the entire region, while region-based evaluation, similar to the Multiface benchmark, is provided in the supplementary materials. + +# 4.6. Qualitative Evaluation Results + +Figure 6 shows that our method not only covers the full head region compared to template-model based approaches, but also exhibits lower UV and 3D distance errors. As men + +tioned above, other baselines are aligned with the GT mesh via Procrustes alignment. Additionally, our results on inthe-wild images are visualized in Figure 7. Our method robustly reconstructs the head from arbitrary head images, even including human-like head images, such as virtual and movie characters. Since the training dataset did not include such out-of-distribution data, this capability can be attributed to the framework built on DINOv2 features. In other words, this demonstrates that leveraging large-scale self-supervised features as a prior is not only more effective than using a template model prior but also provides greater flexibility. + +# 4.7. Ablation Studies + +As introduced in Sec. 1, our key observation is DINOv2's robust descriptive power on head images. Based on this, we assume DINOv2 features serve as semantic and geometric priors instead of template models. To validate this assumption, we performed two ablation studies: one without DINOv2 and the other using a human-centric foundation model. In particular, Sapiens [23] surpasses DINOv2 in terms of training data scale. However, as shown in Table 2, our method based on Sapiens performed even worse than the case without any self-supervised features. This implies that, unlike Sapiens, DINOv2 can extract semantic and geometric features without fine-tuning. Moreover, we demonstrate the effectiveness of our keypoint loss, with the results shown in Table 2 and Figure 8. Our keypoint loss not only bring a metric improvement in both UV and 3D domain but also it leads to noticeable qualitative enhancements, particularly around the eyes and mouth regions. + +![](images/8fda06477a526e064ac7efcc6ccb2942f6d049e83cdc44f9bc919f3934a2913c.jpg) +Image + +Mesh +(a) Ground truth + +![](images/5c07954bae653397ba50b1d74f17b74cb0f06c3e9c5b4998f8c6e21834856bd9.jpg) + +![](images/4369039028d35a22e537f67a54754db7c27126b26f5bf3804459aa4f47595faf.jpg) + +![](images/682b9d3a4be8137920a9088905669ed9ab6b2d84917a378020a88d2908eb6d5c.jpg) + +![](images/b3cdfde1a90794328b52a38309c1b53778cb28fa57a062834ed0bee309381713.jpg) +UV error +(b) HRN + +![](images/a09425db062484f66e5209ca74b618c3d9403baceb15e21008f28c1f117b1a39.jpg) + +![](images/9ae2485fa25e07cca03e6d597ae3983e8dc79d85efb8772879c54cd56e2846e4.jpg) + +![](images/b675fe549afab8103f9608c3221d4c17ae1a43b16eee9ec51f306e0ad073f9f4.jpg) + +![](images/d9dce621d9d8bac5415219c68dd0f5b5e48d44e2cb67f9aaee438f947bc56efd.jpg) +3D error + +![](images/3c848308d530a3cf9ce26f14a1aac4ce544fb31a21863feb1319833f2254b0a6.jpg) + +![](images/c6e87467d34e444046a0ac6ebf41db9251577007caf2db6466deebb67b535320.jpg) + +![](images/9ea298f882b871fa628d0f075226ad2d7e7886b281e652471783787c3028f438.jpg) + +![](images/5364bbbb0dbf86e0ac6ffe04dfe5ad1a3e6dd775f9af223ec29839f7e8e4c185.jpg) +UV error +(c) DECA + +![](images/6a30b6510c0993e90cecbd76dd271432a83235e65ae82655a25f41e45666cb29.jpg) + +![](images/a7423a00ccffb185aee481505db06ed505ed5d95e8147d28f71d1d39d80da1e1.jpg) + +![](images/ca4e353fb27344ae9149542174ff9f6ed5f1f5c8a2e76e1bc6b54751d9b2b071.jpg) + +![](images/f3c229a013aa0b3f80152199669c7d45b1608f12facc76f322175e5b6d80d2c4.jpg) +3D error + +![](images/b49873e816ec6a44584deab45c1ca76de530c0f0cd7ca55efd56880941c6171a.jpg) + +![](images/bf0ebb727f96d0c54d2cc8621a43b54bb8e582f7685e211366c3eca9b7c48829.jpg) + +![](images/94a308e54cd038580d82efed97ad585d9235060c1c3c0314547a61499f65931f.jpg) + +![](images/6677322fb0199205db0d08bd5dcf2a71ac13c526ec534bb25c5fc6e1d83a1a8e.jpg) +UV error +(d) EMOCA + +![](images/8ff4f08de36e8c268424363bb73e6acc2365bcc6779fc03433061d1b9fe84c7e.jpg) + +![](images/79c0dd268e7495c47927060442183b56c5a44fc0c588f658014741233c3bfced.jpg) + +![](images/c0033b318bdbbb9b7c979fe87b1451a75d8805b8c0204f4fe215daabcaf8cb78.jpg) + +![](images/63aff9a6af0811ebf690738e259b7b69a29abce4c40cb30b0a41488d976f9ea7.jpg) +3D error + +![](images/08af26c389170c1f219e2e2708f0af3ebe386bf5c5aae84e54fb6f83176fad7c.jpg) + +![](images/95178697660c4f2ac1d96adcfeb5f48fdcda7e15a2a2f8052e3deee94a2127dc.jpg) + +![](images/f2ddf6d4317450020dd9c20289eb32a2eabad04670d590d72c67a3cdfbafc8b9.jpg) + +![](images/6d554e4f85ebc1e10461915c93e1cc3f91e650d0864443a46f2166a7bf36b55f.jpg) +UV error + +![](images/afb76cc97480549b506845803cd5bf30c66cd476185182af853fcb80147027ed.jpg) + +![](images/7407374779297959aeec963b258aac8d75ed88b04f4b986fa77732ad0c77b820.jpg) + +![](images/3c7038d4afcf35096d9afac745328590c68a58c6b7367507824d8177ad33b01b.jpg) + +![](images/83dca76cd7394d91d579400775e9de4c5ee6c8577ea73df80857d68b1d0cda8c.jpg) +(e) Ours + +![](images/e25b843e3449e523cca154ee053af3d064a5b90b3d521c0035a613f654aab210.jpg) +Figure 6. Qualitative results of Multiface(row.1-2) and FaceVerse(row.3-4) benchmarks. (a) shows the ground truth image and mesh, while (b), (c), (d), and (e) alternately display the UV error and 3D error for HRN, DECA, EMOCA, and WarpHE4D, respectively. Our method demonstrates a wider region coverage while maintaining lower errors. +(a) Input +(b) Ours +(c) DECA +(d) EMOCA +Figure 7. Qualitative results on in-the-wild images. $W_{UVD}$ and $W_{C}$ outputs are visualized as (b) UV colored meshes and (e) masks. Notably, confidence reflects both prediction accuracy and semantic headness + +![](images/27d0a0a791c38b9f89417da7b7b4844cf9186f937df8e0535291ac18b6f76743.jpg) + +![](images/41d0374211a67a1be5e57022477206b4f8cd778f05d8d472044e02a206c9b931.jpg) + +![](images/b2765354c0593da260d3c9dda92ad129b23e0227e1d52992cdb4a30a01bfd2c7.jpg) + +![](images/12984e8598d3708fc66ecba2a18c084222a172e1c96681092374751eefedb6c5.jpg) + +![](images/0724f161847b7724d5564e8c78024a5df704ab113ff5e045fb8bed080ac6f206.jpg) + +![](images/95961f849ad079e7d8cc63f60551406dd5ca2d005ab53407a189266941e74d9f.jpg) +(e) $W_{C}$ + +![](images/2ba53db56c808325135fd4dfc8ab55cd0a2bd6fc5e542fe8aba6c1b3c846ea07.jpg) + +![](images/e47d65dda78e338ed318b5620513e5bbe501148daf3c8616b44e9f03efb21903.jpg) + +![](images/0b3c654884418ffa6da25818b3ef46bae75a404910e5a858a17901a24e139b07.jpg) + +![](images/56f5c17f41012605b277a4fceb8779b1cc6a35fe5712dc43774176bfbd523303.jpg) +Figure 8. Keypoint loss ablation results. (a) input image, (b) UV texture, and (c-d) UV results without/with keypoint loss are shown. For clear comparison, we visualize UV outputs by warping (b). The keypoint loss improves both UV accuracy and stability for the eyes and mouth region. + +![](images/8b4157a10ba083a85dad655c8a37d58bb4ffa699b60f0e76242401098f68e47d.jpg) + +![](images/d6e17ce390a791b25645e8e1745eb494b1bbe38afbc4f269a31b02db9b3492ba.jpg) +(f) PCD + +# 4.8. Computation Cost + +Our method operates at 325 GFLOPs with 121.95M parameters and requires 4.7GB of memory when using a batch size of 1. After training, the UV feature pyramid can be precomputed, reducing the computation to 254 GFLOPs, 111.36M parameters, and 3.5GB of memory. On an RTX + +![](images/d38c5f928dea9a67da30d6c65505d5160d717e1aaaeeb38ff2f7998d9ad39859.jpg) + +![](images/30f1f056a8d208b67d42f97135a2b490a522f5e70f4939b3139df29165e6745c.jpg) +(a) Input + +![](images/eacc240a0cbda9254c0c3be55733c0b72755f592789726b6cf9478a7cf779a9d.jpg) + +![](images/8c805dba6d18a12144dabe73f2efc7e67f09f9028a4ecb85adb664cb296f7579.jpg) +(b) UV texture + +![](images/2f394394ccab9041f6c0c092ef3385c64b3d35ae17e89c497ffdcceefa3ed80a.jpg) + +![](images/8db4d315f82ba31509cc9e890f069cf42dd8c47b53f482f7cd2938137c514fb0.jpg) +(c) w/o $E_{kp}$ + +![](images/be25fd729ddf85c5d0b191506baab7308b1f2160db7b7730623dd3e9a4da681c.jpg) + +![](images/58834e9bb1bd852b9bb16027adbb92baf983c848cfab0e7a3c401acb81005833.jpg) +(d) w/ $E_{kp}$ + +ADA5000 GPU, the model runs at 5 FPS in fp32 and 8FPS in fp16. + +# 5. Conclusion + +We have proposed a novel framework, WarpHE4D, which simultaneously reconstructs the visible head regions with geometry and high-level attributes (i.e. UV coordinates). This approach not only enables full head reconstruction under calibrated settings but also flexible post-processing tasks such as template registration and texture map optimization. One potential drawback is that the inherent difficulty of representing hair as a mesh, even with sufficient data, may lead to performance degradation across various hair types. In future work, we aim to explore new representations for non-facial regions. + +# Acknowledgments + +This work was supported by the National Research Foundation of Korea(NRF) grant funded by the Korea government(MSIT) (NRF2022R1A2B5B03002636) and by Institute for Information & Communications Technology Planning & Evaluation (IITP) grant funded by the Korea government(MSIT) (RS-2024-00398830, RS-2025-25441838, RS-2022-11220566). + +# References + +[1] Brian Amberg, Sami Romdhani, and Thomas Vetter. Optimal step nonrigid icip algorithms for surface registration. In 2007 IEEE conference on computer vision and pattern recognition, pages 1-8. IEEE, 2007. 2 +[2] Mahmoud Assran, Quentin Duval, Ishan Misra, Piotr Bojanowski, Pascal Vincent, Michael Rabbat, Yann LeCun, and Nicolas Ballas. Self-supervised learning from images with a joint-embedding predictive architecture. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15619–15629, 2023. 2 +[3] Roman Bachmann, David Mizrahi, Andrei Atanov, and Amir Zamir. Multimae: Multi-modal multi-task masked autoencoders. In European Conference on Computer Vision, pages 348-367. Springer, 2022. 2 +[4] Haoran Bai, Di Kang, Haoxian Zhang, Jinshan Pan, and Linchao Bao. Ffhq-uv: Normalized facial uv-texture dataset for 3d face reconstruction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 362–371, 2023. 4 +[5] Linchao Bao, Xiangkai Lin, Yajing Chen, Haoxian Zhang, Sheng Wang, Xuefei Zhe, Di Kang, Haozhi Huang, Xinwei Jiang, Jue Wang, Dong Yu, and Zhengyou Zhang. High-fidelity 3d digital human head creation from rgb-d selfies. ACM Transactions on Graphics, 2021. 4 +[6] Jonathan T Barron. A general and adaptive robust loss function. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4331-4339, 2019. 5 +[7] Volker Blanz and Thomas Vetter. Face recognition based on fitting a 3d morphable model. IEEE Transactions on pattern analysis and machine intelligence, 25(9):1063-1074, 2003. 3 +[8] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 2 +[9] Prashanth Chandran, Gaspard Zoss, Paulo Gotardo, and Derek Bradley. Infinite 3d landmarks: Improving continuous 2d facial landmark detection. In Computer Graphics Forum, page e15126. Wiley Online Library, 2024. 3 +[10] Vasileios Choutas, Georgios Pavlakos, Timo Bolkart, Dimitrios Tzionas, and Michael J Black. Monocular expressive body regression through body-driven attention. In European Conference on Computer Vision, pages 20-40. Springer, 2020. 3 + +[11] Radek Daneček, Michael J Black, and Timo Bolkart. Emoca: Emotion driven monocular face capture and animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20311-20322, 2022. 3, 7 +[12] Niladri Shekhar Dutt, Sanjeev Muralikrishnan, and Niloy J Mitra. Diffusion 3d features (diff3f): Decorating untextured shapes with distilled semantic features. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4494-4504, 2024. 2, 4 +[13] Johan Edstedt, Qiyu Sun, Georg Bokman, Marten Wadenback, and Michael Felsberg. Roma: Robust dense feature matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19790-19800, 2024. 4, 5, 6 +[14] David Eigen, Christian Puhrsch, and Rob Fergus. Depth map prediction from a single image using a multi-scale deep network. Advances in neural information processing systems, 27, 2014. 3 +[15] Zhenfeng Fan, Xiyuan Hu, Chen Chen, and Silong Peng. Boosting local shape matching for dense 3d face correspondence. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10944–10954, 2019. 3 +[16] Yao Feng, Fan Wu, Xiaohu Shao, Yanfeng Wang, and Xi Zhou. Joint 3d face reconstruction and dense alignment with position map regression network. In Proceedings of the European conference on computer vision (ECCV), pages 534-551, 2018. 3, 7 +[17] Yao Feng, Haiwen Feng, Michael J Black, and Timo Bolkart. Learning an animatable detailed 3d face model from in-the-wild images. ACM Transactions on Graphics (ToG), 40(4): 1-13, 2021. 3, 7 +[18] Thomas Gerig, Andreas Morel-Forster, Clemens Blumer, Bernhard Egger, Marcel Luthi, Sandro Schonborn, and Thomas Vetter. Morphable face models—an open framework. In 2018 13th IEEE international conference on automatic face & gesture recognition (FG 2018), pages 75–82. IEEE, 2018. 1 +[19] Simon Giebenhain, Tobias Kirschstein, Markos Georgopoulos, Martin Rünz, Lourdes Agapito, and Matthias Nießner. Learning neural parametric head models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21003-21012, 2023. 5 +[20] Syed Zulqarnain Gilani, Ajmal Mian, and Peter Eastwood. Deep, dense and accurate 3d face correspondence for generating population specific deformable models. Pattern Recognition, 69:238-250, 2017. 3 +[21] Jianzhu Guo, Xiangyu Zhu, Yang Yang, Fan Yang, Zhen Lei, and Stan Z Li. Towards fast, accurate and stable 3d dense face alignment. In Proceedings of the European Conference on Computer Vision (ECCV), 2020. 3, 7 +[22] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 2 + +[23] Rawal Khirodkar, Timur Bagautdinov, Julieta Martinez, Su Zhaoen, Austin James, Peter Selednik, Stuart Anderson, and Shunsuke Saito. Sapiens: Foundation for human vision models. In European Conference on Computer Vision, pages 206-228. Springer, 2024. 5, 7 +[24] Biwen Lei, Jianqiang Ren, Mengyang Feng, Miaomiao Cui, and Xuansong Xie. A hierarchical representation network for accurate and detailed face reconstruction from in-the-wild images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 394-403, 2023. 3, 7 +[25] Boqian Li, Haiwen Feng, Zeyu Cai, Michael J Black, and Yuliang Xiu. Etch: Generalizing body fitting to clothed humans via equivariant tightness. arXiv preprint arXiv:2503.10624, 2025. 3 +[26] Hao Li, Robert W Sumner, and Mark Pauly. Global correspondence optimization for non-rigid registration of depth scans. In Computer graphics forum, pages 1421-1430. Wiley Online Library, 2008. 2 +[27] Tianye Li, Timo Bolkart, Michael J Black, Hao Li, and Javier Romero. Learning a model of facial shape and expression from 4d scans. ACM Trans. Graph., 36(6):194-1, 2017. 1, 3 +[28] Zhengqi Li and Noah Snavely. Megadepth: Learning single-view depth prediction from internet photos. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2041-2050, 2018. 3 +[29] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer vision-ECCV 2014: 13th European conference, zurich, Switzerland, September 6-12, 2014, proceedings, part v 13, pages 740-755. Springer, 2014. 5 +[30] Minghua Liu, Ruoxi Shi, Linghao Chen, Zhuoyang Zhang, Chao Xu, Xinyue Wei, Hansheng Chen, Chong Zeng, Jiayuan Gu, and Hao Su. One-2-3-45++: Fast single image to 3d objects with consistent multi-view generation and 3d diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10072–10083, 2024. 2, 4 +[31] Shengjie Ma, Yanlin Weng, Tianjia Shao, and Kun Zhou. 3d gaussian blendshapes for head avatar animation. In ACM SIGGRAPH 2024 Conference Papers, pages 1-10, 2024. 1 +[32] Riccardo Marin, Enric Corona, and Gerard Pons-Moll. Nicp: Neural icip for 3d human registration at scale. In European Conference on Computer Vision, pages 265-285. Springer, 2024. 3 +[33] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 2, 4 +[34] Gang Pan, Xiaobo Zhang, Yueming Wang, Zhenfang Hu, Xiaoxiang Zheng, and Zhaohui Wu. Establishing point correspondence of 3d faces via sparse facial deformable model. IEEE Transactions on Image Processing, 22(11):4170-4181, 2013. 3 + +[35] Ankur Patel and William AP Smith. 3d morphable face models revisited. In 2009 IEEE conference on computer vision and pattern recognition, pages 1327-1334. IEEE, 2009. 3 +[36] Shenhan Qian, Tobias Kirschstein, Liam Schoneveld, Davide Davoli, Simon Giebenhain, and Matthias Nießner. Gaussianavatars: Photorealistic head avatars with rigged 3d gaussians. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20299-20309, 2024. 1 +[37] Soubhik Sanyal, Timo Bolkart, Haiwen Feng, and Michael Black. Learning to regress 3D face shape and expression from an image without 3D supervision. In Proceedings IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 7763-7772, 2019. 3, 5 +[38] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 4 +[39] Felix Taubner, Prashant Raina, Mathieu Tuli, Eu Wern Teh, Chul Lee, and Jinmiao Huang. 3d face tracking from 2d video through iterative dense uv to image flow. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1227-1237, 2024. 1, 3, 4 +[40] Lizhen Wang, Zhiyuan Chen, Tao Yu, Chenguang Ma, Liang Li, and Yebin Liu. Faceverse: a fine-grained and detail-controllable 3d face morphable model from a hybrid dataset. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 20333-20342, 2022. 5, 6 +[41] Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. Dust3r: Geometric 3d vision made easy. In CVPR, 2024. 3 +[42] Zidu Wang, Xiangyu Zhu, Tianshuo Zhang, Baiqin Wang, and Zhen Lei. 3d face reconstruction with the geometric guidance of facial part segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1672-1682, 2024. 3, 7 +[43] Cheng-hsin Wu, Ningyuan Zheng, Scott Ardisson, Rohan Bali, Danielle Belko, Eric Brockmeyer, Lucas Evans, Timothy Godisart, Hyowon Ha, Xuhua Huang, Alexander Hypes, Taylor Koska, Steven Krenn, Stephen Lombardi, Xiaomin Luo, Kevyn McPhail, Laura Millerschoen, Michal Perdoch, Mark Pitts, Alexander Richard, Jason Saragih, Junko Saragih, Takaaki Shiratori, Tomas Simon, Matt Stewart, Autumn Trimble, Xinshuo Weng, David Whitewolf, Chenglei Wu, Shouu-I Yu, and Yaser Sheikh. Multiface: A dataset for neural face rendering. In arXiv, 2022. 5, 6 +[44] Jianfeng Xiang, Zelong Lv, Sicheng Xu, Yu Deng, Ruicheng Wang, Bowen Zhang, Dong Chen, Xin Tong, and Jiaolong Yang. Structured 3d latents for scalable and versatile 3d generation. arXiv preprint arXiv:2412.01506, 2024. 2, 4 +[45] Sicheng Xu, Jiaolong Yang, Dong Chen, Fang Wen, Yu Deng, Yunde Jia, and Xin Tong. Deep 3d portrait from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7710-7720, 2020. 3 +[46] Yuelang Xu, Benwang Chen, Zhe Li, Hongwen Zhang, Lizhen Wang, Zerong Zheng, and Yebin Liu. Gaussian head + +avatar: Ultra high-fidelity head avatar via dynamic gaussians. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 1 +[47] Haotian Yang, Hao Zhu, Yanru Wang, Mingkai Huang, Qiu Shen, Ruigang Yang, and Xun Cao. Facescape: a large-scale high quality 3d face dataset and detailed riggable 3d face prediction. In Proceedings of the IEEE/cvf conference on computer vision and pattern recognition, pages 601-610, 2020. 5 +[48] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10371-10381, 2024. 3 +[49] Yang You, Yixin Li, Congyue Deng, Yue Wang, and Leonidas Guibas. Multiview equivariance improves 3d correspondence understanding with minimal feature finetuning. arXiv preprint arXiv:2411.19458, 2024. 2 +[50] Yuanwen Yue, Anurag Das, Francis Engelmann, Siyu Tang, and Jan Eric Lenssen. Improving 2d feature representations by 3d-aware fine-tuning. In European Conference on Computer Vision, pages 57-74. Springer, 2024. 2, 4 +[51] Tianke Zhang, Xuangeng Chu, Yunfei Liu, Lijian Lin, Zhendong Yang, Zhengzhuo Xu, Chengkun Cao, Fei Yu, Changyin Zhou, Chun Yuan, et al. Accurate 3d face reconstruction with facial component tokens. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9033-9042, 2023. 3 +[52] Jinghao Zhou, Chen Wei, Huiyu Wang, Wei Shen, Cihang Xie, Alan Yuille, and Tao Kong. ibot: Image bert pre-training with online tokenizer. arXiv preprint arXiv:2111.07832, 2021. 2 +[53] Zi-Xin Zou, Zhipeng Yu, Yuan-Chen Guo, Yangguang Li, Ding Liang, Yan-Pei Cao, and Song-Hai Zhang. Triplane meets gaussian splatting: Fast and generalizable single-view 3d reconstruction with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10324-10335, 2024. 2, 4 \ No newline at end of file diff --git a/ICCV/2025/WarpHE4D_ Dense 4D Head Map toward Full Head Reconstruction/images.zip b/ICCV/2025/WarpHE4D_ Dense 4D Head Map toward Full Head Reconstruction/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..52addf1ec56de67046b5a16a89474831d169234b --- /dev/null +++ b/ICCV/2025/WarpHE4D_ Dense 4D Head Map toward Full Head Reconstruction/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:245f3ece0d28abd40481c4c6216f1ff1b4e9b8c3186231adbc38733a391dc612 +size 665468 diff --git a/ICCV/2025/WarpHE4D_ Dense 4D Head Map toward Full Head Reconstruction/layout.json b/ICCV/2025/WarpHE4D_ Dense 4D Head Map toward Full Head Reconstruction/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..2fe18d2cd81fe0a2e4a1767b69972bdb6e198b47 --- /dev/null +++ b/ICCV/2025/WarpHE4D_ Dense 4D Head Map toward Full Head Reconstruction/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:004b09ac27c2c9dccd4a23656b9c268a53ab8ca6df44d057ec16b89c02ad9162 +size 535172 diff --git a/ICCV/2025/Wasserstein Style Distribution Analysis and Transform for Stylized Image Generation/b8746cda-704c-4a0a-9467-542e4e841b49_content_list.json b/ICCV/2025/Wasserstein Style Distribution Analysis and Transform for Stylized Image Generation/b8746cda-704c-4a0a-9467-542e4e841b49_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..b48d186c737278da209e041e12c7434c574fb946 --- /dev/null +++ b/ICCV/2025/Wasserstein Style Distribution Analysis and Transform for Stylized Image Generation/b8746cda-704c-4a0a-9467-542e4e841b49_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:56fa18568c4929caeed453cdb01139eb92cd1bd3e555fa5ff4425a57eb21cc99 +size 70208 diff --git a/ICCV/2025/Wasserstein Style Distribution Analysis and Transform for Stylized Image Generation/b8746cda-704c-4a0a-9467-542e4e841b49_model.json b/ICCV/2025/Wasserstein Style Distribution Analysis and Transform for Stylized Image Generation/b8746cda-704c-4a0a-9467-542e4e841b49_model.json new file mode 100644 index 0000000000000000000000000000000000000000..ea689364a6cf09d2c2aeec5a8ec90675cb6b1e22 --- /dev/null +++ b/ICCV/2025/Wasserstein Style Distribution Analysis and Transform for Stylized Image Generation/b8746cda-704c-4a0a-9467-542e4e841b49_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:216437ddbedbc8fb59085fd9a7316d7ba546140e8dbd171302fd6146aeb4438c +size 91006 diff --git a/ICCV/2025/Wasserstein Style Distribution Analysis and Transform for Stylized Image Generation/b8746cda-704c-4a0a-9467-542e4e841b49_origin.pdf b/ICCV/2025/Wasserstein Style Distribution Analysis and Transform for Stylized Image Generation/b8746cda-704c-4a0a-9467-542e4e841b49_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..d2b5e8d750402c15fee18c8ba7ff5bf3aa2d71b7 --- /dev/null +++ b/ICCV/2025/Wasserstein Style Distribution Analysis and Transform for Stylized Image Generation/b8746cda-704c-4a0a-9467-542e4e841b49_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ff240a021510b3e42930a9f29ae8242a6e25072ce43764e21b63097145bbc891 +size 17559204 diff --git a/ICCV/2025/Wasserstein Style Distribution Analysis and Transform for Stylized Image Generation/full.md b/ICCV/2025/Wasserstein Style Distribution Analysis and Transform for Stylized Image Generation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..b8cdc3c7b531d5084d1d5aa8ab97efb8b15505fa --- /dev/null +++ b/ICCV/2025/Wasserstein Style Distribution Analysis and Transform for Stylized Image Generation/full.md @@ -0,0 +1,307 @@ +# Wasserstein Style Distribution Analysis and Transform for Stylized Image Generation + +Xi Yu + +Xiang Gu* + +Zhihao Shi + +Jian Sun* + +School of Mathematics and Statistics, Xi'an Jiaotong University Xi'an, 710049, China + +ericayu@stu.xjtu.edu.cn, xianggu@xjtu.edu.cn, szhxjtu777@stu.xjtu.edu.cn, jiansun@xjtu.edu.cn + +![](images/726cb874909e8c0f849bed404b7b8d7ef36f0c6761f036f536010d89f687eac5.jpg) +Figure 1. Visualization of stylized images generated by our WSDT method through Wasserstein style distribution transform. Subfigures (a) and (b) show the results of WSDT built on SDXL [25] and SD3 [9] base models, respectively. + +![](images/d7664b419af2ff4c4b524b4702940b430de7a3bf4451010b36b05f28fbb76cd3.jpg) + +# Abstract + +Large-scale text-to-image diffusion models have achieved remarkable success in image generation, thereby driving the development of stylized image generation technologies. Recent studies introduce style information by empirically replacing specific features in attention blocks with style features. However, the relationship between features and style remains unclear. In this paper, we systematically analyze the relationship between features in attention blocks and style. By quantifying the distribution discrepancy induced by style variations using the Wasserstein distance, we find that features in self-attention blocks exhibit high sensitivity to style compared to features in cross-attention blocks. Our analysis provides valuable insights into the contribu + +tion of different features to style. Based on our findings, we propose a novel Wasserstein Style Distribution Transform (WSDT) method, which generates stylized images by transforming the distribution of style-sensitive features to align with that of style features. WSDT applies channel adaptive distribution transform to ensure that information not related to the style is not introduced. Our approach is simple yet efficient, optimization-free, and can be seamlessly integrated into attention-based text-to-image diffusion models. Extensive experiments demonstrate the effectiveness of our approach in stylized image generation tasks. + +# 1. Introduction + +Text-to-image diffusion models [2, 23, 25, 29, 30, 34] have made significant progress in the field of image generation, giving rise to numerous applications, such as stylized image + +generation [18, 32, 42, 46, 49]. Stylized image generation aims to generate images that are consistent with a reference style and aligned with the conditioned text prompt. + +Existing approaches for stylized image generation can be classified into two categories: optimization-based and optimization-free. Some optimization-based methods [3, 20, 26, 43, 45] rely on extensive dataset collection and pretraining of style encoders or adapters to boost the style representation capabilities of the model, incurring high costs during data collection and training. In addition, methods like StyleDrop [37] and InstaStyle[5] enable single style image optimization at test time, however, resulting in significantly increasing inference time. + +This work focuses on the optimization-free approach that generates style images using the off-the-shelf diffusion models without any optimization. Recent optimization-free methods [4, 7, 13, 17] control the stylized image generation process by modifying the internal features of the UNet [31]. There are two fundamental challenges to be tackled for such approach. 1) Since the diffusion model contains various features, which feature should be modified for stylized image generation? 2) Given the features, how to modify the features to generate the desired style? + +For the first challenge, previous methods [4, 7, 13, 17] attempt to perform feature modification in self or cross-attention blocks. However, as there are multiple types of features in attention blocks, such as input features, value features, etc., it is still unclear which features should be modified for stylized image generation task. Building upon previous work, in this paper, we propose to systematically analyze the sensitivity of various features in attention blocks to style. We quantify the discrepancy of feature distributions caused by style, based on Wasserstein distance [22]. We analyze three diffusion models and find that the features in self-attention block are more sensitive to style variations compared to those in cross-attention block, indicating their significant role in stylized image generation. Therefore, we suggest modifying these style-sensitive features in self-attention block for effective stylized image generation. + +To tackle the second challenge, we propose Wasserstein Style Distribution Transform (WSDT) method. Wasserstein distance motivates a feature distribution transform that relies on the mean and variance of the features. Specifically, based on the findings in the style sensitivity analysis, we apply the feature distribution transform to style-sensitive features in self-attention block. Considering that the significance of each channel to style is different, we further develop channel adaptive distribution transform to keep the channels less related to style unchanged, which helps prevent content leakage. WSDT is optimization-free and can be integrated into attention-based diffusion models. + +We conduct extensive experiments for evaluation. The results demonstrate that the images generated by our + +method not only achieve strong alignment with the style image but also faithfully follow the content prompts. Moreover, our approach is simple but efficient and can be easily applied to various stylized image generation applications. + +In summary, our main contributions are as follows: + +- We systematically analyze the relationship between features in attention blocks and style. We quantify the feature distribution discrepancy caused by style variations using Wasserstein distance to examine the sensitivity of features to style. We find that the features in self-attention are particularly sensitive to style. +- We propose WSDT method, which generates stylized images by transforming the distributions of style-sensitive features to that of style features. To avoid content leakage, we design a channel-adaptive distribution transform. +- Extensive experiments demonstrate the superiority of our method. Additionally, our method can be integrated into transformer-based diffusion models and other image generation methods, showcasing its strong scalability. + +# 2. Related Work + +# 2.1. Text-to-Image Diffusion Models + +Diffusion models [8, 9, 14, 25, 30] have attracted considerable attention in recent years. Text-to-image diffusion models leverage large-scale image-text datasets [35] and cutting-edge language models [28] to excel in text-to-image generation tasks. As a pioneer, GLIDE [23] processes input text into token sequences and employs a transformer [39] to map tokens into conditioning inputs for the diffusion model. Latent Diffusion [30], also known as the Stable Diffusion (SD), performs efficient diffusion process within a low-dimensional latent space using a pre-trained autoencoder. SDXL [25] employs UNet [31] architecture as the diffusion network, integrating cross-attention to inject text embeddings for conditioned image generation. SD3 [9] utilizes multimodal diffusion transformer [39] architecture to improve text understanding and image generation capabilities. In this paper, we focus on analyzing the impact of the SD models on stylization. + +# 2.2. Stylized Image Generation + +Stylized image generation aims to generate images with a specific style. Gatys et al. [11] propose Gram matrices to express style in CNNs. To enable real-time arbitrary style transfer, Huang and Belongie [16] introduce Adaptive Instance Normalization (AdaIN), which aligns the mean and variance of content features to those of style features. AdaAttN [21], StyleFormer [44], and StyTr² [6] extract features using VGG or transformer encoders, introducing style by injecting key and value via attention mechanism. + +Diffusion model-based methods employ text prompt and reference style image for stylized image generation. + +Optimization-based methods [3, 20, 26, 37, 43, 45] utilize style encoders to capture style information, requiring high training costs. For instance, ArtAdapter [3] and StyleTokenizer [20] employ large-scale datasets to train style encoders to encode style information into diffusion models. Some works [10, 24, 36] integrate LoRA [15] to learn the style of reference images, such as K-LoRA [24]. HyperDreamBooth [33] uses a hypernetwork to generate personalized weights from a single portrait, enabling stylized face generation via weights composition and fine-tuning. D2Styler [38] encodes content and style with VQ-GAN and models their joint latent space using a discrete diffusion model with AdaIN features as condition. Optimization-based methods either require pre-training additional modules or optimization at test time, which is time-consuming. + +Optimization-free methods can generate stylized images by manipulating UNet features during the denoised process. The method $\mathcal{Z}^*$ [7] generates stylized images by incorporating a cross-attention re-weighting strategy to fuse the features key and value from the cross-attention block of the reference images. StyleID [4], VSP [17] and StyleAligned [13] operate on self-attention blocks, generating images by integrating the key and value features from the self-attention blocks of the reference images during the generation process. InstantStyle [41] injects reference image features into specific cross-attention layers of the IP-Adapter [48] to incorporate style information. These methods empirically select features in self or cross attention blocks for complete or weighted feature replacement, which may lead to failures in stylized image generation. + +In this paper, we propose an efficient Wasserstein style distribution transform method that generates stylized images without optimization. We explore the sensitivity of features in attention blocks to style. By selecting the more effective features and applying channel adaptive distribution transform, we can generate images that align with the style while exhibiting less content leakage. + +# 3. Preliminaries + +Latent Diffusion Models. Given an image $x \in \mathbb{R}^{H \times W \times 3}$ , the LDM uses a pre-trained encoder to encode $x$ into the latent space, yielding $z \in \mathbb{R}^{h \times w \times c}$ . The network $\epsilon_{\theta}$ is trained in the latent space to predict the added noise $\epsilon$ . The training objective of LDM is formulated as follows: + +$$ +\mathcal {L} _ {L D M} = \mathbb {E} _ {z, \epsilon \sim \mathcal {N} (\mathbf {0}, \mathbf {I}), t, \mathcal {C}} \left[ \| \epsilon_ {\theta} (z _ {t}, t, \mathcal {C}) - \epsilon \| _ {2} ^ {2} \right], \tag {1} +$$ + +where $\mathcal{C}$ denotes the condition, $t\in [0,T]$ represents the diffusion timestep, $z_{t}$ is the noisy latent at timestep $t$ + +Attention Mechanism in LDM. Self-attention mechanism captures the internal relationships within image features, enabling self-aggregation of features, while cross-attention + +![](images/ca8e1557d37d550d660742bce1807baa7f3138ca2a8e20deeadf5e5bafcc02df.jpg) +Figure 2. Images generated by SDXL [25] conditioned on the prompts from constructed datasets. Prompts used in the first row are sourced from the content-only prompt set, while those in second and third rows are sourced from the content-style prompt set. + +integrates text features into the image features, providing guidance based on text prompts. The attention operation is defined as $\text{Attention}(Q, K, V) = \text{Softmax}\left(\frac{QK^T}{\sqrt{d}}\right)V$ , where $d$ denotes the dimension of key vectors. In self-attention, query $Q$ , key $K$ , and value $V$ are obtained by projecting image features through projection matrices. In cross-attention, $Q$ is derived from the projected image features, $K$ and $V$ are derived from projected text features. + +Wasserstein Distance. Wasserstein distance [40] measures the discrepancy between two distributions. Given distributions $\mathbb{P}$ and $\mathbb{Q}$ , the $d$ -Wasserstein distance is defined as: + +$$ +W (\mathbb {P}, \mathbb {Q}) := \left(\inf _ {\gamma \in \Pi (\mathbb {P}, \mathbb {Q})} \mathbb {E} _ {(X, Y) \sim \gamma} \left[ \| X - Y \| ^ {d} \right]\right) ^ {1 / d}, \tag {2} +$$ + +where $\Pi (\mathbb{P},\mathbb{Q})$ is the set of all joint distributions of $X$ and $Y$ marginally distributed as $X\sim \mathbb{P}$ and $Y\sim \mathbb{Q}$ . + +# 4. Method + +Previous studies [4, 12] have observed that the attention maps determine the spatial layout, the key and value features in attention blocks are used to adjust the generated content. However, the influence of these features on style remains unclear. Empirically selecting features often leads to suboptimal results. We aim to analyze the impact of features in attention blocks on style and design an effective feature transform method to better generate stylized images while mitigating content leakage. We design our method by answering the following two questions: + +- Feature Style Sensitivity. Which features in attention blocks are more sensitive to style? Our goal is to identify the features sensitive to style and make modifications on those features for effective stylized image generation. +- Distribution Transform for Stylized Image Generation. How can we transform the feature distribution of the + +![](images/0d2cd41414883053f8390185853f1772f524e325f49688ba1a5fb76b27786038.jpg) +(a) SDXL + +![](images/d422c250c53c3e0cf8f690738a0cae9370ae51070ed7aff26e97a441861431b3.jpg) +(b) SD V2.1 + +![](images/fc61cabebf907e1808b84b2b1cf85d85b6a834cab9a5ebdd939419157467cb5b.jpg) +(c) SD V1.4 +Figure 3. Comparison of style sensitivity across different features in attention blocks for three diffusion models (SDXL, SD V2.1, and SD V1.4). $F$ , $M$ , $V$ , and $O$ represents the input feature, attention map, value feature and output feature, respectively. The vertical axis represents the ratio $\frac{W_s^2}{W_s^2 + W_c^2}$ . + +generated image to match that of the style images without introducing content information of style images? Our goal is to generate stylized images while preventing content leakage from the style image. + +# 4.1. Feature Style Sensitivity Analysis + +To answer the first question, i.e., investigating the sensitivity of features to style, we construct a dataset containing paired images, which share the same content but with different styles in each pair. The idea is that if style change causes a significant variation in the feature distribution, then that feature contains meaningful style-related information. + +Datasets construction. To achieve this, we use ChatGPT to generate category and style descriptions to construct two sets of prompt, i.e., the content-only and content-style prompt sets. The prompt in content-only set is formatted as: $\langle object\rangle\langle action\rangle$ . For example, "A cat is lazily napping next to the window". This set includes 20 categories, each with 5 descriptions, totaling 100 descriptions. The content-style set combines 20 style prompts with content prompts in content-only set, resulting in 2000 content-style prompts with the formulation: $\langle object\rangle\langle action\rangle\langle style\rangle$ . For example, "A cat is lazily napping next to the window in animation style". Figure 2 shows example images generated by SDXL conditioned on the prompts from the constructed dataset. + +Measuring feature sensitivity to style. To quantify the sensitivity of features to style, we examine the Wasserstein distance between distributions of features obtained from the diffusion process conditioned on style changed prompts. The Wasserstein distance allows us to quantify the discrepancy of the feature distributions caused by style variations. Specifically, given $f, g \in \mathbb{R}^{N \times M}$ with $N$ and $M$ denoting the spatial and channel sizes, we assume $f \sim \mathbb{P}$ and $g \sim \mathbb{Q}$ . $\mathbb{P}$ and $\mathbb{Q}$ are modeled as Gaussian distributions with means + +![](images/e2b480a8532b6b552983e72d7187ceae745af3db2c25a2db7806a9153849cad0.jpg) +Figure 4. Results of distribution transform on different features. The top-left is the image generated without transform. The top-right is the reference style image. The second/third row displays images generated by transforming distributions of different features in self/cross-attention blocks. + +$\pmb{\mu}_{f}, \pmb{\mu}_{g} \in \mathbb{R}^{M}$ and covariances $\pmb{\Sigma}_{f}, \pmb{\Sigma}_{g} \in \mathbb{R}^{M \times M}$ . The square of 2-Wasserstein distance between $\mathbb{P}$ and $\mathbb{Q}$ is given by the following closed-set formulation: + +$$ +\begin{array}{l} W ^ {2} (\mathbb {P}, \mathbb {Q}) = \| \boldsymbol {\mu} _ {f} - \boldsymbol {\mu} _ {g} \| _ {2} ^ {2} \\ + \operatorname {T r} \left(\boldsymbol {\Sigma} _ {f} + \boldsymbol {\Sigma} _ {g} - 2 \left(\boldsymbol {\Sigma} _ {f} ^ {\frac {1}{2}} \boldsymbol {\Sigma} _ {g} \boldsymbol {\Sigma} _ {f} ^ {\frac {1}{2}}\right) ^ {\frac {1}{2}}\right). \tag {3} \\ \end{array} +$$ + +Intuitively, a larger discrepancy in the feature distributions indicates that the feature is more sensitive to style, and therefore exerts a greater influence on the style information. + +Analysis results. Using the prompts from the constructed datasets as conditions, we extract the input features $F$ , attention maps $M$ , value features $V$ , and output features $O$ from the self and cross-attention blocks during the diffusion process. According to Eq. (3), we compute the Wasserstein distance $W_{s}^{2} / W_{c}^{2}$ for the feature distributions obtained from different styles/categories in the prompt. For computational simplicity, we assume the covariance matrix is diagonal. To exclude the influence of feature scale, we calculate the ratio $\frac{W_{s}^{2}}{W_{s}^{2} + W_{c}^{2}}$ to measure style sensitivity. Ratio greater than 0.5 indicates that the feature distribution distance caused by style change is larger than that caused by category change, implying that the feature is more sensitive to style. A larger ratio signifies higher sensitivity to style. + +We select three base models for analysis to explore patterns in feature style sensitivity: SD V1.4 [30], SD V2.1 [30], and SDXL [25]. As shown in Figure 3, we observe a similar pattern across all models. The results reveal interesting insights. Features $F$ , $V$ , and $O$ in self-attention block exhibit significantly larger distribution differences caused by style changes compared to category changes. For the features in the cross-attention blocks, the distribution distances induced by style changes are smaller than that + +![](images/46a0406584cf7360840e114aa4ea35cde021f9744fa543659333c10d0bf42d0d.jpg) +Figure 5. Overview of our WSDT method. WSDT generates images that match the style of the reference style image by performing channel adaptive distribution transform on style-sensitive features. Specifically, we first extract style-sensitive features (yellow) and style features (purple) during the denoising process. Then we compute the channel-wise distribution distances for these features. We transform the feature distribution to align with that of the reference style image for channels with smaller distribution distances and keep the features unchanged for channels with larger distances (gray). We utilize a threshold to determine which channels should be transformed. + +induced by category changes, which suggests that cross-attention tends to be more sensitive to content. Additionally, attention maps in self and cross-attention blocks are more sensitive to category changes. This observation aligns with empirical findings that attention maps primarily govern the structural information of the generated images. Intuitively, style of images could be represented by color/lightness related to scaling/bias or high-frequency texture. While the content may reflect spatial feature relationships. The softmax of attention map is robust to feature global scaling and bias, and the following aggregation may reduce high-frequency information. Thus attention map may be less sensitive to style. Conversely, MLP-generated features are sensitive to scaling, bias, or high-frequency variations. + +To further validate our findings, we use the method discussed in Sect. 4.2 to transform the distribution of different features respectively during generation. We guide the generation process using text prompt describing only content. The results are shown in Figure 4. Without distribution transform, the generated image depicts a sport car in a natural scene. When we transform the distribution of features $F$ , $V$ , and $O$ in self-attention block, the generated image exhibits a style similar to the style image, even though the text prompt contains no style description. In contrast, transforming other features does not produce a stylized image. This further confirms that the features $F$ , $V$ , and $O$ in self-attention blocks are key to determining the style. + +Based on the analysis, we conclude that the input feature $F$ , value feature $V$ , and output feature $O$ in self-attention block are more sensitive to style. They significantly influence the style of the generated image. Therefore, we suggest transforming the distributions of $F$ , $V$ , and $O$ in self-attention block for effectively stylized image generation. + +# 4.2. Wasserstein Style Distribution Transform + +Now we address the second question: how to transform the feature distribution of the generated image to match that of the style image. From the perspective of optimal transport, if distributions $\mathbb{P}$ and $\mathbb{Q}$ are both Gaussian, the optimal transport plan from $\mathbb{P}$ to $\mathbb{Q}$ takes the following form: + +$$ +T (f) = \pmb {\mu} _ {g} + \pmb {\Sigma} _ {f} ^ {- \frac {1}{2}} (\pmb {\Sigma} _ {f} ^ {\frac {1}{2}} \pmb {\Sigma} _ {g} \pmb {\Sigma} _ {f} ^ {\frac {1}{2}}) ^ {\frac {1}{2}} \pmb {\Sigma} _ {f} ^ {- \frac {1}{2}} (f - \pmb {\mu} _ {f}). \qquad (4) +$$ + +We aim to use the optimal transport plan to transform the feature distribution of the generated image to match that of the style image. When covariance matrices are diagonal, this distribution transform degrades to the form used in AdaIN [16]: replacing the mean and variance of each channel in the generated image's feature with those in the style image's feature. This inspires us to further design a channel-wise Wasserstein distance to measure the distribution discrepancy of each channel, allowing us to adaptively select channels with stronger style significance for distribution transform. We avoid transforming features that are less related to style, which exhibit larger distribution distance. These features tend to contain style-irrelevant information, and transforming them may introduce content information from the style image, leading to content leakage issues. + +Assuming $\pmb{\mu} = \text{diag}(\mu_1, \mu_2, \dots, \mu_M)$ and $\pmb{\Sigma} = \text{diag}(\sigma_1, \sigma_2, \dots, \sigma_M)$ with $\mu_i$ and $\sigma_i$ denotes the mean and variance of the $i$ -th channel. The square of 2-Wasserstein distance for the $i$ -th channel can then be calculated by: + +$$ +W _ {i} ^ {2} = \left(\mu_ {f, i} - \mu_ {g, i}\right) ^ {2} + \left(\sigma_ {f, i} - \sigma_ {g, i}\right) ^ {2}. \tag {5} +$$ + +We employ a threshold $\tau$ to select channels with smaller distribution distances between style and generated features for transform, while keeping channels with larger distances + +![](images/f20e4e7f8a6d45470380e7bf672036940956181340b359ffbe70b87e01b2fcb1.jpg) +Figure 6. Qualitative comparison results on evaluation datasets. + +unchanged. The set of channels selected is defined as: + +$$ +\mathcal {A} = \left\{i \mid W _ {i} ^ {2} < \tau , i = 1, 2, \dots , M \right\}. \tag {6} +$$ + +A larger threshold $\tau$ causes more channels' features to be transformed to match the distribution of style feature, resulting in generated images that are more consistent with the reference style image. We discuss the influence of $\tau$ on the generated images in Sect. 5.3. + +# 4.3. Stylized Image Generation + +As shown in Figure 5, our method applies channel adaptive distribution transform for stylized image generation. Given a reference style image, we first apply the forward process to obtain the noisy latent for each timestep. During denoising process, we extract the features in self-attention blocks for distribution transform. Following the Wasserstein style distribution transform method proposed in Sect. 4.2, we apply channel-wise Wasserstein distance to adaptively select the more style-relevant channels and transform the distributions to align with that of style features. This ensures that the generated image matches the style of the reference image while mitigating content leakage. Based on the above methods, we show a simple yet efficient algorithm for stylized image generation in Algorithm 1. + +# Algorithm 1 Wasserstein style distribution transform. + +Input: $I$ : Reference style image; $\mathcal{C}$ : style prompt; $\mathcal{C}^*$ : target prompt. + +Output: $I^{*}$ : generated stylized image. + +1: $\{z_t\}_{t=0}^T \gets \text{DDPM-Forward}(I)$ ; +2: $z_T^*\sim \mathcal{N}(\mathbf{0},\mathbf{I})$ +3: for $t = T, T - 1, \dots, 1$ do +4: $F, V, O \gets \epsilon_{\theta}(z_t, t, \mathcal{C})$ ; +5: $F^{*},V^{*},O^{*}\gets \epsilon_{\theta}(z_{t}^{*},t,\mathcal{C}^{*})$ +6: Compute distribution distance for each channel; +7: Select channel set using Eq. (6); +8: Adaptively transform the feature distribution of $F^{*}, V^{*}, O^{*}$ to that of $F, V, O$ using Eq. (4); +9: Compute $z_{t - 1}^{*}$ using the transformed $F^{*},V^{*},O^{*}$ +10: end for +11: Return $I^{*} \gets z_{0}^{*}$ . + +# 5. Experiments + +# 5.1. Experiment Setup + +Evaluation Datasets. We conduct experiments on the evaluation datasets constructed by StyleDrop [37] and In + +Table 1. Quantitative comparison results on evaluation datasets. We use bold and underline to indicate the best and second-best results. + +
MethodVSP [17]StyleDrop [37]StyleAligned [13]DEADiff [26]InstaStyle [5]CSGO [45]WSDT
Content Score ↑0.2820.2540.2700.2430.2700.2590.285
Style Score ↑0.5250.4710.5380.4790.5290.4750.546
GM Metric ↓10.8719.6310.5322.0811.7312.035.52
User Preference ↑0.2460.0490.0800.0200.0880.0240.494
+ +![](images/1bb94c7164af7a128287d1e3aaa1f4d5755690ef0b4bc470f034d7e6d7cecbba.jpg) +Figure 7. Results for methods built on the SD3 base model. + +staStyle [5]. StyleDrop includes 24 style images and 30 text prompts, while InstaStyle contains 60 style images and uses 100 categories from the CIFAR-100 [19] as text prompts. The evaluation datasets encompass various styles and content. Our method was comprehensively evaluated across a wide range of style and content scenarios. To demonstrate the generality of our method, there is no data duplication between our evaluation and analysis datasets. + +Implementation Details. We use SDXL [25] as our base model. We perform all experiments on a single NVIDIA GeForce RTX 4090 GPU. We employ the DDIM sampler with 50 sampling steps and set the guidance scale to 7.5. We set the threshold $\tau$ to 0.2. Unless otherwise specified, other parameters are set to their default settings. + +Evaluation Metrics. We utilize CLIP [27] to quantitatively measure the content/style alignment between the generated images and the content prompts/reference images. Specifically, we define content score by calculating the cosine similarity between generated images and content prompts to evaluate content alignment. Style score is obtained by sub + +Table 2. The performance on evaluation datasets with varying values of hyperparameter $\tau$ . The results in the last column are obtained by applying distribution transform across all channels. + +
τ0.00.10.20.4All
Content Score ↑0.2980.2900.2850.2830.281
Style Score ↑0.4760.5340.5460.5460.547
+ +tracting the similarity between the generated images and the content prompts of the reference images, from that between the generated images and the reference images. We use GM Metric to further assess style alignment by computing the difference between the Gram matrices of the generated and style images. For user study, we show 40 participants 40 groups of images generated by different methods in random order. Participants are suggested to select images that perform best in both content and style alignment. We calculate the percentage of votes for each method for human preference comparison. + +# 5.2. Comparison Results + +We compare WSDT with state-of-the-art methods, including VSP [17], StyleDrop [37], StyleAligned [13], DEADiff [26], InstaStyle [5], and CSGO [45]. As shown in Figure 6, the images generated by VSP exhibit overall inferior quality with blurring artifacts. The images generated by StyleDrop, DEADiff, and CSGO significantly deviate from the style of the reference images. While StyleAligned and InstaStyle methods achieve a certain degree of style alignment, StyleAligned still exhibits noticeable deviations from the reference style. And InstaStyle suffers from content leakage, as the generated images incorporate content information from the reference images. In comparison, our approach generates high-quality stylized images that consistently match the reference style while avoiding content leakage. As shown in the quantitative results presented in Table 1, our method achieves the best performance in all metrics, outperforming the competing methods. Moreover, the last row of Table 1, which presents the user preference results, indicates that the participants exhibit a strong preference towards our method. + +Results in SD3 base model. Our method can be seamlessly integrated into attention-based diffusion models. To evaluate the performance of WSDT on the transformer-based dif + +![](images/a966098bd9ff57db817152e4588e209295dfb600648fcc791e84ec6c995e3c15.jpg) +Figure 8. Effect of hyper-parameters $\tau$ . The first column presents the text prompt alongside the reference image. Columns 2-5 display the results with different $\tau$ values. The results in last column are obtained by applying transform across all channels. + +![](images/d7695e30d65cdaa61802ea75bb161526f015e9b501c998b1e9ac7d47762d3bbf.jpg) +Figure 9. The first column presents the text prompt alongside the reference image. Images in second column are generated by SDXL base model. Images in third column are obtained by applying AdaIN on the noisy latents, and images in last column are generated by our method through distribution transform. + +fusion model, we implement WSDT on SD3 [9] and extend StyleAligned [13] and VSP [17] to SD3 for comparison. StyleAligned and VSP maintain style consistency by sharing or swapping the keys and values with those in the self-attention layers of style images. As shown in Figure 7, the images generated by StyleAligned and VSP exhibit inferior quality, which may be attributed to the direct incorporation of feature elements that disrupt the denoising process. In contrast, our method generates high-quality stylized images that highly consistent with the reference style, demonstrating the effectiveness and generalizability of our method. + +# 5.3. Ablation Study + +Effect of hyper-parameters $\tau$ . As mentioned in Sect. 4.2, we perform adaptive distribution transform by applying a threshold. The threshold $\tau$ determines which channels are transformed to match the distribution of the style image. In Figure 8, we show the effect of different values of $\tau$ on the generated images. When $\tau = 0.0$ , no channels are transformed, resulting in style difference between the generated and style images. As $\tau$ increases, more style sensitive channels are transformed, causing the style of the generated image to gradually align with the style image. When $\tau$ in + +creases beyond a certain range, a larger $\tau$ will cause channels that are less sensitive to style to be transformed, posing a risk of content leakage. As shown in Table 2, when $\tau$ exceeds 0.2, the increase in the style score from a larger $\tau$ is limited, and the content score is reduced. When distribution transform is applied to all channels, the content from reference image may leak into the generated images, as indicated by the red box in Figure 8. We suggest setting $\tau$ to 0.2 to achieve style alignment while mitigating content leakage. + +Comparison with method applying AdaIN on the noisy latents. Some works [1, 47] apply AdaIN [16] on the noisy latent for color consistency between the generated and the reference images. Differently, our method align the distributions of the features that are most sensitive to style within the internal features of the UNet. As illustrated in Figure 9, applying AdaIN on the noisy latent achieves only color alignment, whereas our method yields style alignment between the generated and the reference images. + +# 5.4. Limitations + +Our method has some limitations. While the proposed channel-adaptive distribution transform can alleviate content leakage from the reference style image to a certain extent, it does not completely eliminate this issue. We use a fixed threshold $\tau$ across all reference style images for simplification. However, different styles may require different thresholds to achieve the optimal balance between style fidelity and content preservation. A single threshold may be insufficient to accommodate the diversity of style characteristics, potentially leading to either excessive content leakage or inadequate stylization in some cases. In future work, we plan to investigate adaptive thresholding strategies that consider style-specific requirements to enhance the controllability and generalization of our approach. And we plan to further investigate the relationship between features and style in more advanced models. + +# 6. Conclusion + +In this paper, we propose WSDT, a simple yet efficient approach for stylized image generation. Based on Wasserstein distance, we analyze the sensitivity of features in attention blocks to style. We find that the features in self-attention block are highly sensitive to style. Therefore, we propose to apply feature distribution transform on these style sensitive features in self-attention block to generate stylized images effectively. Furthermore, to alleviate content leakage, we adaptively select channels for distribution transform based on channel-wise Wasserstein distance. Our method does not require optimization and can be easily integrated into attention-based diffusion models. Extensive experiments demonstrate the effectiveness of our method. + +# 7. Acknowledgement + +This work was supported by National Key R&D Program 2021YFA1003002, NSFC (12125104, 12426313, 623B2084), China National Postdoctoral Program for Innovative Talents (BX20240276), the Fundamental Research Funds for the Central Universities (xzy022025047). + +# References + +[1] Yuval Alaluf, Daniel Garibi, Or Patashnik, Hadar Averbuch-Elor, and Daniel Cohen-Or. Cross-image attention for zero-shot appearance transfer. In SIGGRAPH, 2024. 8 +[2] Huiwen Chang, Han Zhang, Jarred Barber, AJ Maschinot, Jose Lezama, Lu Jiang, Ming-Hsuan Yang, Kevin Murphy, William T Freeman, Michael Rubinstein, et al. Muse: Text-to-image generation via masked generative transformers. In ICML, 2023. 1 +[3] Dar-Yen Chen, Hamish Tennent, and Ching-Wen Hsu. Artadapter: Text-to-image style transfer using multi-level style encoder and explicit adaptation. In CVPR, 2024. 2, 3 +[4] Jiwoo Chung, Sangeek Hyun, and Jae-Pil Heo. Style injection in diffusion: A training-free approach for adapting largescale diffusion models for style transfer. In CVPR, 2024. 2, 3 +[5] Xing Cui, Zekun Li, Pei Pei Li, Huaibo Huang, Xuannan Liu, and Zhaofeng He. Instastyle: Inversion noise of a stylized image is secretly a style adviser. In ECCV, 2024. 2, 7 +[6] Yingying Deng, Fan Tang, Weiming Dong, Chongyang Ma, Xingjia Pan, Lei Wang, and Changsheng Xu. Stytr2: Image style transfer with transformers. In CVPR, 2022. 2 +[7] Yingying Deng, Xiangyu He, Fan Tang, and Weiming Dong. Z*: Zero-shot style transfer via attention reweighting. In CVPR, 2024. 2, 3 +[8] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. In NeurIPS, 2021. 2 +[9] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In ICML, 2024. 1, 2, 8 +[10] Yarden Frenkel, Yael Vinker, Ariel Shamir, and Daniel Cohen-Or. Implicit style-content separation using b-lora. arXiv preprint arXiv:2403.14572, 2024. 3 +[11] Leon A Gatys, Alexander S Ecker, and Matthias Bethge. Image style transfer using convolutional neural networks. In CVPR, 2016. 2 +[12] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. In ICLR, 2023. 3 +[13] Amir Hertz, Andrey Voynov, Shlomi Fruchter, and Daniel Cohen-Or. Style aligned image generation via shared attention. In CVPR, 2024. 2, 3, 7, 8 +[14] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 2 + +[15] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. In ICLR, 2022. 3 +[16] Xun Huang and Serge Belongie. Arbitrary style transfer in real-time with adaptive instance normalization. In ICCV, 2017. 2, 5, 8 +[17] Jaeseok Jeong, Junho Kim, Yunjey Choi, Gayoung Lee, and Youngjung Uh. Visual style prompting with swapping self-attention. arXiv preprint arXiv:2402.12974, 2024. 2, 3, 7, 8 +[18] Jaeseok Jeong, Mingi Kwon, and Youngjung Uh. Training-free style transfer emerges from h-space in diffusion models. In WACV, 2024. 2 +[19] A Krizhevsky. Learning multiple layers of features from tiny images. Master's thesis, University of Tront, 2009. 7 +[20] Wen Li, Muyuan Fang, Cheng Zou, Biao Gong, Ruobing Zheng, Meng Wang, Jingdong Chen, and Ming Yang. Style-tokensizer: Defining image style by a single instance for controlling diffusion models. In ECCV, 2024. 2, 3 +[21] Songhua Liu, Tianwei Lin, Dongliang He, Fu Li, Meiling Wang, Xin Li, Zhengxing Sun, Qian Li, and Errui Ding. Adaattn: Revisit attention mechanism in arbitrary neural style transfer. In ICCV, 2021. 2 +[22] Anton Mallasto and Aasa Feragen. Learning from uncertain curves: The 2-wasserstein metric for gaussian processes. In NeurIPS, 2017. 2 +[23] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 1, 2 +[24] Ziheng Ouyang, Zhen Li, and Qibin Hou. K-lora: Unlocking training-free fusion of any subject and style loras. arXiv preprint arXiv:2502.18461, 2025.3 +[25] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. In ICLR, 2024. 1, 2, 3, 4, 7 +[26] Tianhao Qi, Shancheng Fang, Yanze Wu, Hongtao Xie, Jiawei Liu, Lang Chen, Qian He, and Yongdong Zhang. Deadiff: An efficient stylization diffusion model with disentangled representations. In CVPR, 2024. 2, 3, 7 +[27] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 7 +[28] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. JMLR, 21(140):1-67, 2020. 2 +[29] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 1 + +[30] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 1, 2, 4 +[31] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI, 2015. 2 +[32] Litu Rout, Yujia Chen, Nataniel Ruiz, Abhishek Kumar, Constantine Caramanis, Sanjay Shakkottai, and Wen-Sheng Chu. Rb-modulation: Training-free personalization of diffusion models using stochastic optimal control. arXiv preprint arXiv:2405.17401, 2024. 2 +[33] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Wei Wei, Tingbo Hou, Yael Pritch, Neal Wadhwa, Michael Rubinstein, and Kfir Aberman. Hyperdream booth: Hypernetworks for fast personalization of text-to-image models. In CVPR, 2024. 3 +[34] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. In NeurIPS, 2022. 1 +[35] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. In NeurIPS, 2022. 2 +[36] Viraj Shah, Nataniel Ruiz, Forrester Cole, Erika Lu, Svetlana Lazebnik, Yuanzhen Li, and Varun Jampani. Ziplora: Any subject in any style by effectively merging loras. arXiv preprint arXiv:2311.13600, 2023. 3 +[37] Kihyuk Sohn, Nataniel Ruiz, Kimin Lee, Daniel Castro Chin, Irina Blok, Huiwen Chang, Jarred Barber, Lu Jiang, Glenn Entis, Yuanzhen Li, et al. Styledrop: Text-to-image generation in any style. In NeurIPS, 2023. 2, 3, 6, 7 +[38] Onkar Susladkar, Gayatri Deshmukh, Sparsh Mittal, and Parth Shastri. $\mathrm{D}^2$ styler: Advancing arbitrary style transfer with discrete diffusion methods. In ICPR, 2025. 3 +[39] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 2 +[40] Cedric Villani. Topics in optimal transportation.(books). OR/MS Today, 30(3):66-67, 2003. 3 +[41] Haofan Wang, Matteo Spinelli, Qixun Wang, Xu Bai, Zekui Qin, and Anthony Chen. Instantstyle: Free lunch towards style-preserving in text-to-image generation. arXiv preprint arXiv:2404.02733, 2024. 3 +[42] Zhizhong Wang, Lei Zhao, and Wei Xing. *Stylediffusion: Controllable disentangled style transfer via diffusion models*. In ICCV, 2023. 2 +[43] Zhouxia Wang, Xintao Wang, Liangbin Xie, Zhongang Qi, Ying Shan, Wenping Wang, and Ping Luo. Styleadapter: A unified stylized image generation model. IJCV, pages 1-18, 2024. 2, 3 +[44] Xiaolei Wu, Zhihao Hu, Lu Sheng, and Dong Xu. Styleformer: Real-time arbitrary style transfer via parametric style composition. In ICCV, 2021. 2 + +[45] Peng Xing, Haofan Wang, Yanpeng Sun, Qixun Wang, Xu Bai, Hao Ai, Renyuan Huang, and Zechao Li. Csgo: Content-style composition in text-to-image generation. arXiv preprint arXiv:2408.16766, 2024. 2, 3, 7 +[46] Serin Yang, Hyunmin Hwang, and Jong Chul Ye. Zero-shot contrastive loss for text-guided diffusion image style transfer. In ICCV, 2023. 2 +[47] Shuai Yang, Yifan Zhou, Ziwei Liu, and Chen Change Loy. Rerender a video: Zero-shot text-guided video-to-video translation. In SIGGRAPH, 2023. 8 +[48] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023. 3 +[49] Zhanjie Zhang, Quanwei Zhang, Wei Xing, Guangyuan Li, Lei Zhao, Jiakai Sun, Zehua Lan, Junsheng Luan, Yiling Huang, and Huaizhong Lin. Artbank: Artistic style transfer with pre-trained diffusion model and implicit style prompt bank. In AAAI, 2024. 2 \ No newline at end of file diff --git a/ICCV/2025/Wasserstein Style Distribution Analysis and Transform for Stylized Image Generation/images.zip b/ICCV/2025/Wasserstein Style Distribution Analysis and Transform for Stylized Image Generation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..d8d698d7f3a976370ef263d9a520f787faa2d3e8 --- /dev/null +++ b/ICCV/2025/Wasserstein Style Distribution Analysis and Transform for Stylized Image Generation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:60e65de3b9274d308c496245ca5dc3bc6d0fa694264f18356fc75abeff23ab81 +size 888169 diff --git a/ICCV/2025/Wasserstein Style Distribution Analysis and Transform for Stylized Image Generation/layout.json b/ICCV/2025/Wasserstein Style Distribution Analysis and Transform for Stylized Image Generation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..2a37e72ee4bb4b59c809cff87848b0a30e8eaf20 --- /dev/null +++ b/ICCV/2025/Wasserstein Style Distribution Analysis and Transform for Stylized Image Generation/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:70e74001552d73fd386174edecf4616ec83c298c9995c1b510119475faad2d6a +size 397049 diff --git a/ICCV/2025/Wave-MambaAD_ Wavelet-driven State Space Model for Multi-class Unsupervised Anomaly Detection/d167cc0d-f166-4fec-a4c9-a57721d19178_content_list.json b/ICCV/2025/Wave-MambaAD_ Wavelet-driven State Space Model for Multi-class Unsupervised Anomaly Detection/d167cc0d-f166-4fec-a4c9-a57721d19178_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..bbf27b4caad619a2b2e039e338c387434368cd27 --- /dev/null +++ b/ICCV/2025/Wave-MambaAD_ Wavelet-driven State Space Model for Multi-class Unsupervised Anomaly Detection/d167cc0d-f166-4fec-a4c9-a57721d19178_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:74dec17968520eba8c366e25acc518684fd97e911ffeea0c0292a498e12f9c5f +size 72721 diff --git a/ICCV/2025/Wave-MambaAD_ Wavelet-driven State Space Model for Multi-class Unsupervised Anomaly Detection/d167cc0d-f166-4fec-a4c9-a57721d19178_model.json b/ICCV/2025/Wave-MambaAD_ Wavelet-driven State Space Model for Multi-class Unsupervised Anomaly Detection/d167cc0d-f166-4fec-a4c9-a57721d19178_model.json new file mode 100644 index 0000000000000000000000000000000000000000..588743677ca292abfca74e2fdd8bcd743fda1416 --- /dev/null +++ b/ICCV/2025/Wave-MambaAD_ Wavelet-driven State Space Model for Multi-class Unsupervised Anomaly Detection/d167cc0d-f166-4fec-a4c9-a57721d19178_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:29f915d04650b2f61e53bcbea806188e4963bcd1b71f61ea5c7e67ecb1a677e9 +size 90169 diff --git a/ICCV/2025/Wave-MambaAD_ Wavelet-driven State Space Model for Multi-class Unsupervised Anomaly Detection/d167cc0d-f166-4fec-a4c9-a57721d19178_origin.pdf b/ICCV/2025/Wave-MambaAD_ Wavelet-driven State Space Model for Multi-class Unsupervised Anomaly Detection/d167cc0d-f166-4fec-a4c9-a57721d19178_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..3c92fa481274b4dd895f3f88b88ba9d3c67a6cda --- /dev/null +++ b/ICCV/2025/Wave-MambaAD_ Wavelet-driven State Space Model for Multi-class Unsupervised Anomaly Detection/d167cc0d-f166-4fec-a4c9-a57721d19178_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ebd8d0b3febe928572d77dfe866810d23cf660ac4f63417517f12cf55ceccbe5 +size 1329431 diff --git a/ICCV/2025/Wave-MambaAD_ Wavelet-driven State Space Model for Multi-class Unsupervised Anomaly Detection/full.md b/ICCV/2025/Wave-MambaAD_ Wavelet-driven State Space Model for Multi-class Unsupervised Anomaly Detection/full.md new file mode 100644 index 0000000000000000000000000000000000000000..c0cf4579d9d45960eba482d329f2ba28c8f54d04 --- /dev/null +++ b/ICCV/2025/Wave-MambaAD_ Wavelet-driven State Space Model for Multi-class Unsupervised Anomaly Detection/full.md @@ -0,0 +1,292 @@ +# Wave-MambaAD: Wavelet-driven State Space Model for Multi-class Unsupervised Anomaly Detection + +Qiao Zhang + +Mingwen Shao* + +Xinyuan Chen + +Xiang Lv + +Kai Xu + +Shandong Key Laboratory of Intelligent Oil & Gas Industrial Software, Qingdao Institute of Software, College of Computer Science and Technology, China University of Petroleum (East China) + +bz24070008@s.upc.edu.cn, smw278@126.com, b24070011@s.upc.edu.cn, lvxiang1997@126.com, + +s23070014@s.upc.edu.cn + +# Abstract + +The Mamba model excels in anomaly detection through efficient long-range dependency modeling and linear complexity. However, Mamba-based anomaly detectors still face two critical challenges: (1) insufficient modeling of diverse local features leading to inaccurate detection of subtle anomalies; (2) spatial-wise scanning mechanism disrupting the spatial continuity of large-scale anomalies, resulting in incomplete localization. To address these challenges, we propose Wave-MambaAD, a wavelet-driven state space model for unified subtle and large-scale anomaly detection. Firstly, to capture subtle anomalies, we design a high-frequency state space model that employs horizontal, vertical, and diagonal scanning mechanisms for processing directionally aligned high-frequency components, enabling precise anomaly detection through multidimensional feature extraction. Secondly, for comprehensive localization of large-scale anomalies, we propose a low-frequency state space model implementing channel-adaptive dynamic scanning mechanisms to maintain structural coherence in global contexts, which facilitates large-scale anomaly detection via adaptive feature integration. Finally, we develop a dynamic spatial enhancement block to improve anomalous feature representation by enhancing feature diversity through coordinated inter-channel communication and adaptive gating mechanisms. Comprehensive experiments on benchmark anomaly detection datasets show that WaveMambaAD achieves competitive performance at lower parameters and computational costs. + +# 1. Introduction + +Visual anomaly detection (AD) aims to identify abnormal areas in images that deviate from normal patterns. AD technology helps prevent potential risks and improve safety, + +![](images/1907a2f22587382c4b2921b8ab924a8c33d15f44c5defdc80ded0b67be970ac0.jpg) +Figure 1. Motivation and effectiveness of our method. Compared with the spatial-domain baseline (MambaAD), our approach (Wave-MambaAD) employs wavelet-domain modeling: directional scanning for high-frequency components to capture subtle anomalies, and channel-adaptive dynamic scanning for low-frequency components to fully localize large-scale anomalies. 'HF' and 'LF' denote the high- and low-frequency, respectively. + +and therefore has a wide range of applications in industrial defect detection [5], medical image diagnosis [44], video surveillance [7], autonomous driving [36], and agricultural spot detection [29, 41, 42], etc. + +Previous AD methods have predominantly relied on manual labeling and extensive annotated data, a paradigm proven impractical in real-world scenarios due to the inherent scarcity of labeled anomalies. This limitation has propelled the emergence of unsupervised anomaly detection (UAD) methods, which operate without labeled data and have attracted substantial research interest. Current UAD methods are broadly classified into three categories: augmentation-based, embedding-based, and reconstruction-based methods. Augmentation-based methods [24, 37] address anomaly sample scarcity through + +![](images/2cfd5e9ae7c852db6e8ae7c541935677f6580432d45c2eeae497c6620b13a78b.jpg) +Figure 2. Comparison with other SoTA methods regarding performance and efficiency. Our method can achieve competitive performance and inference speed. Here, 'mAD' denotes the average performance of the three datasets, and the circle size indicates the memory consumption during inference. + +data transformations and synthetic anomaly generation. However, suboptimal augmentation strategies may distort anomalous feature learning. Embedding-based methods [13, 18, 28] project normal data into low-dimensional manifolds for analysis, nevertheless, their effectiveness comes at the cost of demanding substantial computational resources. Reconstruction-based methods [6, 15, 35] have recently gained prominence owing to their robust performance and scalability. These methods model normal data distributions, identifying anomalies by quantifying reconstruction errors. For instance, RD4AD [6] establishes a single-class reconstruction framework via reverse distillation. However, it necessitated training individual models per class, leading to high computational costs. + +To overcome this limitation, UniAD [35] has pioneered a unified multi-class reconstruction approach that maintains competitive performance while drastically reducing resource consumption, thereby validating the feasibility of multi-class reconstruction in cost-sensitive environments. Currently, most multi-class reconstruction methods rely on convolutional neural networks (CNNs) [6], Transformers [39], GANs [20] or Diffusion Models [15] to construct models. However, CNNs struggle with long-range dependency modeling due to their localized receptive fields, Transformers suffer from quadratic computational complexity in self-attention operations, and GANs or Diffusion Models face challenges in convergence stability and training efficiency. These limitations impede their deployment in large-scale real-time applications. + +Recently, Mamba [10] has gained attention for its strong performance and linear computational efficiency. He et al. [14] proposed MambaAD, a Mamba-based anomaly de + +tectors that introduces a Hilbert curve-based scanning strategy in the spatial domain to enhance contextual modeling, achieving competitive performance. Nevertheless, current Mamba-based anomaly detectors (e.g., MambaAD) still face two limitations. First, the limited ability to model diverse local features leads to inaccurate capture of subtle anomalies. Second, the spatial-wise scanning mechanism disrupts the spatial continuity of large-scale anomalies, leading to incomplete localization. As indicated in Figure 1 (a), MambaAD struggles to locate subtle anomalies and large-scale anomalies. + +To tackle these limitations, we propose a wavelet-driven state space anomaly detector, dubbed Wave-MambaAD, that effectively localizes both subtle and large-scale anomalies (Figure 1(b)). Wave-MambaAD comprises a pretrained encoder and a novel Wavelet-Mamba Decoder. Concretely, subtle anomalies manifest as weak local feature variations; however, Mamba's global modeling strength limits its sensitivity to such localized patterns. In contrast, the wavelet transform decomposes image features into directional high-frequency components (horizontal, vertical, and diagonal), which contain fine-grained texture details crucial for detecting subtle anomalies. Accordingly, we present a high-frequency state space (HFSS) model that applies directional scanning mechanisms (horizontal, vertical, and diagonal) to capture subtle anomalies from the corresponding high frequency. + +Next, the low-frequency component not only reflects the global structure of image features but also captures uniform variations at large scales, both of which are crucial for detecting large-scale anomalies. Consequently, to leverage these properties, we propose a low-frequency state space (LFSS) model that employs a channel-adaptive dynamic scanning mechanism, thereby preserving global structural integrity and enhancing the detection of large-scale anomalies. + +Finally, to further enhance anomaly feature representations, we designed a dynamic spatial enhancement (DSE) block that promotes inter-channel information flow and leverages adaptive gating mechanisms to enrich spatial features. Notably, HFSS, LFSS and wavelet transform are combined to form a wavelet-based dual state space (WDSS) block. WDSS and DSE are combined to form the Wavelet-Mamba Module, which is the core of the Wavelet-Mamba Decoder. We propose the Wave-MambaAD achieves competitive performance, demonstrating faster inference speeds and lower memory consumption (as illustrated in Figure 2). + +Our main contributions can be summarized as: + +- Wave-MambaAD combines wavelet transform and state space model to detect both subtle and large-scale anomalies via frequency-aware processing, achieving competitive performance with low complexity. +- A high-frequency state space model with directional + +scanning is devised to capture subtle anomalies in multiple directions. + +- A low-frequency state-space model with channel-adaptive scanning preserves global structure for detecting large-scale anomalies. +- A dynamic spatial enhancement block is proposed to enhance anomaly features through inter-channel interaction and adaptive gating. + +# 2. Related Work + +# 2.1. Unsupervised Anomaly Detection + +Most UAD methods are broadly categorized into Augmentation-based, Embedding-based, and Reconstruction-based methods. Augmentation-based methods improve training by generating diverse anomaly data, like SimpleNet [24], which adds Gaussian noise to normal features, and DRAEM [38], which enhances training for anomalous features. However, these methods heavily rely on the design of augmentation strategies, and inappropriate augmentation may hinder the model's ability to learn true anomaly features. Meanwhile, embedding-based methods analyze normal sample embeddings to identify deviations, such as MDND [26] with Gaussian distributions, PatchCore [28] using memory banks, and CFLOW-AD [13] and PyramidFlow [18] with normalized flows. While effective, these methods require significant resources and expertise. Besides, reconstruction-based methods focus on self-trained encoders and decoders to reconstruct images, with commonly used reconstruction backbones including CNNs [27], Transformers [25], GANs [20], Diffusion Models [15], and Mamba [14]. However, generalisation of the model sometimes leads to inaccurate localization of anomalous regions. + +# 2.2. Multi-class Anomaly Detection + +Most AD methods are trained separately for each class, increasing time and memory overhead as the number of classes grows, and struggling with large intra-class diversity. Therefore, multi-class anomaly detection has become a research focus to address these challenges. For example, UniAD [35] introduced the multi-class paradigm with a unified framework, while DiAD [15] used a diffusion model with semantic guidance to ensure reconstructed image consistency. Furthermore, ViTAD [39] advanced the field with a pure Transformer architecture for better global and local anomaly detection. However, existing reconstruction-based methods—whether CNNs, Transformers, or diffusion models—have limitations: CNNs lack long-range modeling, Transformers are computationally expensive, and diffusion models are hard to optimize. Recently, MambaAD [14] addresses the above challenges by leveraging the advantages of Mamba's long-range modeling and linear complexity. + +# 2.3. State Space Models + +State space models (SSMs) [8, 10-12, 30] have gained attention for their performance in natural language processing. Although the structured state space sequence (S4) model [11] captures long-range dependencies through diagonal parameterization, it compresses all historical information, causing redundancy. Recently, Mamba [10] improves this by introducing an information selectivity mechanism, enabling efficient filtering of redundant data. Mamba's superior long-range modeling capabilities and linear computational efficiency have inspired extensive exploration of computer vision. For instance, VMamba [23] proposes a cross-scanning mechanism to address orientation sensitivity, sparking further research in fields such as medical imaging [22, 33, 34], remote sensing [3, 4, 32], and low-level vision [9, 19, 31, 43, 46]. In the field of AD, MambaAD [14] pioneered the application of Mamba to anomaly detection by designing a Hilbert-type scanning approach, achieving competitive performance. + +However, MambaAD, which relies on spatial domain modeling, suffers from false or missed detections when handling multi-orientation subtle anomalies, while fixed spatial scanning disrupts the structural continuity of large-scale anomalies. Instead, our wavelet domain-based detector addresses these limitations through dual mechanisms: (1) We design the directional scanning mechanism to process high-frequency components to capture subtle anomalies in specific directions; (2) We use a channel-adaptive dynamic scanning mechanism to process low-frequency components to maintain global spatial coherence for complete localization of large-scale anomalies. + +# 3. Method + +# 3.1. Preliminaries + +Recently, state space models such as S4 have emerged, inspired by linear time-invariant systems. In these models, a one-dimensional input function or sequence $x(t) \in \mathbb{R}$ is transformed into an output $y(t) \in \mathbb{R}$ via a hidden state $h(t) \in \mathbb{R}^N$ , where the hidden state dynamics are governed by a linear ordinary differential equation (ODE): + +$$ +\left\{ \begin{array}{l} h ^ {\prime} (t) = \mathbf {A} h (t) + \mathbf {B} x (t), \\ y (t) = \mathbf {C} h (t) + \mathbf {D} x (t). \end{array} \right. \tag {1} +$$ + +where $N$ denotes the state size, $\mathbf{A}\in \mathbb{R}^{N\times N}$ $\mathbf{B}\in \mathbb{R}^{N\times 1}$ $\mathbf{C}\in \mathbb{R}^{1\times N}$ , and $\mathbf{D}\in \mathbb{R}$ + +Following this, the continuous-time system was embedded within deep-learning models by introducing a timescale parameter $\Delta$ , which enables the discretization of the continuous parameters $\mathbf{A}$ and $\mathbf{B}$ using the zero-order hold (ZOH) technique. This conversion yields the discrete pa + +![](images/3882fdc60c10fe75d919b5c4fe30df36de561f5944b225554b589cec1a3ec048.jpg) + +![](images/d98070976c28577964028efa0e5b92fdb99d73967e96b3f74bff5e0f0eb863ff.jpg) +Figure 3. Overview of the proposed Wave-MambaAD. The framework adopts a pyramid autoencoder to reconstruct multi-scale features via the proposed Wavelet-Mamba Module. Each module comprises a wavelet-driven dual state space (WDSS) block for capturing both subtle and large-scale anomalies, as well as a dynamic spatial enhancement (DSE) block for refining the representations of anomalies. In WDSS, 'DWT' and 'IWT' refer to discrete wavelet transform and inverse wavelet transform, respectively. + +rameters $\overline{\mathbf{A}}$ and $\overline{\mathbf{B}}$ . + +$$ +\left\{ \begin{array}{l} \overline {{\mathbf {A}}} = e x p (\boldsymbol {\Delta} \mathbf {A}), \\ \overline {{\mathbf {B}}} = (\boldsymbol {\Delta} \mathbf {A}) ^ {- 1} (e x p (\boldsymbol {\Delta} \mathbf {A}) - \mathbf {I}) (\boldsymbol {\Delta} \mathbf {B}). \end{array} \right. \tag {2} +$$ + +After discretisation, Eq.1 can be represented as: + +$$ +\left\{ \begin{array}{l} h (t) = \overline {{\mathbf {A}}} h _ {t - 1} + \overline {{\mathbf {B}}} x _ {t}, \\ y (t) = \mathbf {C} h _ {t} + \mathbf {D} x (t). \end{array} \right. \tag {3} +$$ + +The final output can be obtained directly through full convolution computation. + +However, the aforementioned process parameters are static across different inputs, which restricts their adaptability. To overcome this limitation, Mamba incorporates a scanning mechanism alongside data-dependent, learnable parameters $\Delta$ , $\overline{\mathbf{B}}$ , and $\mathbf{C}$ , allowing the model to dynamically adjust its learning context in response to the specific characteristics of the input. + +# 3.2. Overview + +Figure 3 shows the Wave-MambaAD framework for multi-class anomaly detection. During training, a pretrained encoder (ResNet-34 [16]) extracts multi-scale features, which are fused in the Half-FPN bottleneck [21] and + +then passed to a Wave-Mamba Decoder for reconstruction. The final loss function is the sum of the mean squared error $(L_{m})$ computed across feature maps at three scales. During inference, the anomaly map is generated by summing the cosine similarities of features across scales. + +Within the Wavelet-Mamba Decoder, we introduce the Wavelet-Mamba Module, which consists of a wavelet-based dual state space (WDSS) block and a dynamic spatial enhancement (DSE) block. Specifically, the WDSS block comprises a high-frequency state space (HFSS) model and a low-frequency state space (LFSS) model. The HFSS model processes the high-frequency component to capture subtle anomalies, while the LFSS model processes the low-frequency component to fully localize large-scale anomalies. Finally, the DSE block is attached to the WDSS to further enhance the anomaly feature representation. + +# 3.3. Wavelet-based Dual State Space + +While Mamba-based anomaly detectors demonstrate promising performance, however, there are two limitations: (1) Inadequate modelling of diverse local features leads to inaccurate detection of subtle defects; and (2) the spatial-wise scanning tends to compromise the continuity of large- + +![](images/f97ffb116aab878bdceb42e9ed3c41a43197f1f15618dd651ae56b6e4464130d.jpg) +(a) High-Frequency State Space (HFSS) + +(b) Low-Frequency State Space (LFSS) +Figure 4. Overview of the proposed HFSS and LFSS. The HFSS employs directional selective scanning mechanisms to capture subtle anomalies in high-frequency features. LFSS utilizes channel-wise selective scanning mechanisms to preserve spatial structural integrity, effectively detecting large-scale anomalies. In (a), 'H', 'V', and 'D' are the horizontal, vertical, and diagonal high-frequency components. +![](images/58eb509b3f656d8a7ed300f87e534406073397416ec9bccc4df0f54c656cf05f.jpg) +DWC Depth-wise Convolution H/W-GAP Global Average Pooling along the High/ Wide SiLU C Concatenation + +scale anomalies causing incomplete detection. To address the above limitations, we propose a wavelet-based dual state space (WDSS) block that contains a high-frequency state space (HFSS) model and a low-frequency state space (LFSS) model. Concretely, HFSS captures subtle anomalies in the corresponding high-frequency subbands via a directional scanning mechanism (horizontal, vertical, and diagonal). Furthermore, LFSS maintains global spatial structural integrity and locates large-scale anomalies in their entirety through a channel-adaptive dynamic scanning mechanism. The detailed flow of WDSS is shown in Figure 3, and the whole process can be described as follows: + +$$ +\left\{ \begin{array}{l} F _ {h} ^ {i n}, F _ {l} ^ {i n} = \mathbf {D W T} (\mathbf {L N} (F _ {i n})), \\ F _ {h} ^ {\text {o u t}} = \mathbf {H F S S} \left(F _ {h} ^ {i n}\right), \\ F _ {l} ^ {\text {o u t}} = \mathbf {L F S S} \left(F _ {l} ^ {i n}\right), \\ F _ {w d s s} = \mathbf {I W T} \left(F _ {h} ^ {\text {o u t}}, F _ {l} ^ {\text {o u t}}\right) + F _ {i n}, \end{array} \right. \tag {4} +$$ + +where $\mathbf{LN}(\cdot)$ denotes layer normalization, $\mathbf{DWT}(\cdot)$ and $\mathbf{IWT}(\cdot)$ represent Haar wavelet transform and Haar wavelet inverse transform, respectively. The detailed for + +murations of $\mathbf{DWT}(\cdot)$ and $\mathbf{IWT}(\cdot)$ are provided in the supplementary materials. $\mathbf{HFSS}(\cdot)$ and $\mathbf{LFSS}(\cdot)$ denote the high-frequency state space model and the low-frequency state space model, respectively. $F_{h}$ denotes the high-frequency component, comprising horizontal, vertical, and diagonal high frequencies. $F_{l}$ represents the low-frequency component. $F_{wdss}$ is the output of WDSS. + +High-Frequency State Space. The HFSS is illustrated in Figure 4 (a). The HFSS retains the architectural framework of VSSM [23], with the difference that we design a directional (horizontal, vertical, diagonal) selective scanning mechanism specifically optimized for processing direction-aligned high-frequency components, enabling precise capture of subtle anomaly signatures. The process of HFSS can be represented as: + +$$ +\left\{ \begin{array}{l} F _ {h} ^ {1} = \mathbf {L N} (\mathbf {D S S} (\mathbf {S i L U} (\mathbf {D W C} (\mathbf {L i n e a r} (F _ {h} ^ {i n}))))), \\ F _ {h} ^ {2} = \mathbf {S i L U} (\mathbf {L i n e a r} (F _ {h})), \\ F _ {h} ^ {\text {o u t}} = \mathbf {L i n e a r} \left(F _ {h} ^ {1} \cdot F _ {h} ^ {2}\right), \end{array} \right. \tag {5} +$$ + +where $\mathbf{DWC}(\cdot)$ denotes depth-wise convolution. $\mathbf{DSS}(\cdot)$ + +represents the directional selective scan (DSS). The DSS pipeline first serializes directional high-frequency components (H, V, D) into 1D sequences, applies corresponding directional scanning strategies, then employs discrete state space modeling to capture long-range dependencies, followed by summation-based fusion and dimensional reconstruction to recover the original 2D structure. + +Low-Frequency State Space. The detailed architecture of LFSS is shown in Figure 4 (b). Specifically, we first apply depth-wise convolution to $F_{l}^{in}$ , subsequently conduct channel grouping. Each partitioned group is then processed through Mamba blocks with channel-adaptive scanning mechanisms to model long-range dependencies. Finally, the processed groups are concatenated and the output feature is collectively generated via residual connections. The overall process of LFSS can be represented as: + +$$ +\left\{ \begin{array}{l} F _ {l} ^ {1}, F _ {l} ^ {2}, \dots , F _ {l} ^ {n} = \mathbf {S p l i t} (\mathbf {D W C} (F _ {l} ^ {i n})), \\ F _ {l} ^ {1} = \mathbf {C M a m b a} (F _ {l} ^ {1}), \\ F _ {l} ^ {2} = \mathbf {C M a m b a} (F _ {l} ^ {2}), \\ \dots , \\ F _ {l} ^ {n} = \mathbf {C M a m b a} (F _ {l} ^ {n}), \\ F _ {l} ^ {\text {o u t}} = \operatorname {C o n v} _ {\mathbf {1} \times \mathbf {1}} \left(\left[ F _ {l} ^ {1}, F _ {l} ^ {2}, \dots , F _ {l} ^ {n} \right]\right) + F _ {l} ^ {\text {i n}}, \end{array} \right. \tag {6} +$$ + +where $\text{Split}(\cdot)$ , and $[\cdot]$ denote split and concatenation operations, respectively. CMamba(\cdot) denotes the designed Mamba block with channel-wise selective scanning (Channel Mamba). The pipeline of Channel Mamba is shown in Figure 4 (b). Specifically, we compress spatial features using the axis pooling operation and then model long-distance dependencies between channels via bidirectional channel SSM (Bi-CSSM). Axis pooling reduces computational costs while maintaining certain spatial features. The whole process can be represented as: + +$$ +\left\{ \begin{array}{l} F ^ {\prime} = \mathbf {C S S M} _ {\mathrm {B i}} \left(\mathbf {G A P} _ {\mathrm {H}} (I n p u t)\right) \cdot I n p u t, \\ O u t p u t = \mathbf {C S S M} _ {\mathrm {B i}} \left(\mathbf {G A P} _ {\mathrm {W}} \left(F ^ {\prime}\right)\right) \cdot I n p u t, \end{array} \right. \tag {7} +$$ + +where $\mathbf{GAP_H}(\cdot)$ and $\mathbf{GAP_W}(\cdot)$ denote the global average pooling along height and width, respectively. $\mathbf{CSSM_{Bi}}(\cdot)$ is the bidirectional channel SSM. + +# 3.4. Dynamic Spatial Enhancement + +While the WDSS architecture decouples high-frequency and low-frequency modelling in the wavelet domain through a state space model (allowing for the simultaneous detection of both subtle defects and large-scale anomalies), its frequency-centric representation exhibits inherent limitations in preserving spatial structure continuity. To bridge this representation gap, we propose a dynamic spatial enhancement (DSE) block—a lightweight yet effective component that enhances the spatial structure through coordinated inter-channel communication pathways and adap + +tive gating mechanisms, thereby improving anomaly feature representation. The detailed architecture of DSE is illustrated in Figure 3, and its process can be formally described by the following equation: + +$$ +\left\{ \begin{array}{l} F _ {d} = \mathbf {D W C o n v} (\mathbf {C S} (\mathbf {C o n v} _ {\mathbf {1} \times \mathbf {1}} (\mathbf {L N} (F _ {w d s s})))), \\ F _ {d} ^ {1}, F _ {d} ^ {2} = \mathbf {S p l i t} (F _ {d}), \\ F _ {d} ^ {1} = \mathbf {G E L U} (F _ {d} ^ {2}) \cdot F _ {d} ^ {1}, \\ F _ {d} ^ {2} = \mathbf {G E L U} (F _ {d} ^ {1}) \cdot F _ {d} ^ {2}, \\ F _ {o u t} = \mathbf {C o n v} _ {\mathbf {1} \times \mathbf {1}} (F _ {d} ^ {1} + F _ {d} ^ {2}) + F _ {w d s s}, \end{array} \right. \tag {8} +$$ + +where $\mathbf{CS}(\cdot)$ denotes the channel shuffle operation. The DSE complements spatial structure information and further enhances the representation of anomaly features. + +# 3.5. Anomaly Segmentation and Classification. + +Anomaly segmentation provides a pixel-level anomaly score map to locate anomalies. The multi-scale anomaly maps $A_{i}$ (i=1, 2, 3) from constrained features are summed to form the final anomaly map $A_{sum}$ . For anomaly classification, we perform average pooling on $A_{sum}$ and use its maximum value as the image-level score to determine if the image is anomalous. + +# 4. Experiments and Analysis + +# 4.1. Implementation Details + +Our framework adopts a pre-trained ResNet-34 [16] as the feature extractor and a designed Wavelet-Mamba Decoder as the student network, where the Wavelet-Mamba Module employs a $[1, 2, 2, 2]$ layer configuration. During both training and inference phases, input images are uniformly resized to $256^2$ resolution. The optimization process utilizes the Adam optimizer with an initial learning rate of 0.005 and weight decay of 0.0001, executed 100 training epochs on an NVIDIA RTX A6000. The loss function is the mean square error (MSE) at multiple scales. During inference, the anomaly map is generated by summing cosine similarities across scales. + +# 4.2. Datasets and Metrics + +To evaluate Wave-MambaAD, we conduct experiments on three anomaly detection datasets: MVTec-AD [1], VisA [47], and MPDD [17]. Following previous work [14], we evaluate the model's performance using the following metrics: Area Under the Receiver Operating Characteristic Curve (AU-ROC), Average Precision (AP) [38], F1-score-max (F1_max) [47], and Area Under the Per-Region-Overlap (AU-PRO) [2]. Furthermore, we compute the average of all metrics (mAD) to measure the overall performance of the model. A more detailed description is provided in the supplementary materials. + +
DatasetsMethodImage-levelPixel-levelmAD
AU-ROCAPF1_maxAU-ROCAPF1_maxAU-PRO
MVTec-AD [1]CFLOW-AD [13]91.696.793.495.745.948.688.380.0
SimpleNet [24]95.998.696.097.049.351.987.382.3
PyramidalFlow [18]70.285.585.580.022.322.047.559.0
RealNet [45]88.595.091.876.240.638.762.770.5
DiAD [15]97.299.096.596.852.655.590.784.0
MambaAD [14]97.899.397.397.455.157.693.485.4
Wave-MambaAD98.299.397.497.555.958.393.185.6
VisA [47]CFLOW-AD [13]86.588.884.997.733.937.286.873.7
SimpleNet [24]88.891.285.497.135.038.483.274.1
PyramidalFlow [18]58.266.374.477.07.29.642.847.9
RealNet [45]80.885.178.774.025.729.945.759.9
DiAD [15]86.888.385.196.026.133.075.270.1
MambaAD [14]94.394.589.498.539.444.091.078.3
Wave-MambaAD94.494.890.198.640.143.891.479.0
MPDD [17]CFLOW-AD [13]75.780.181.796.826.328.089.568.3
SimpleNet [24]88.492.087.996.532.034.689.074.3
PyramidalFlow [18]73.677.079.494.121.117.877.262.9
RealNet [45]85.190.288.383.336.139.668.170.1
DiAD [15]68.377.980.190.410.913.166.158.1
MambaAD [14]88.793.290.897.533.638.192.376.3
Wave-MambaAD92.594.991.098.041.542.293.879.1
+ +Table 1. Quantitative comparison with other cutting-edge multi-class anomaly detection methods. The best results are denoted using bold and the second-best results are underlined. + +
MethodParams(M)FLOPs(G)Memory(MB)SpeedmAD
CFLOW-AD237.028.7189230.974.0
SimpleNet72.817.749464.976.9
PyramidalFlow34.3962.0283619.256.6
RealNet591.0115.0379443.566.8
DiAD1331.0451.5203061.870.7
MambaAD25.78.3148452.680.0
Ours22.37.5114258.881.2
+ +Table 2. Comparison of efficiency with SoTA methods. 'Memory' denotes the memory consumption during inference. The best results are shown in bold. + +# 4.3. Comparison with SoTAs on AD Datasets + +We compare the proposed method with SoTAs anomaly detection methods, including MambaAD [14], DiAD [15], RealNet [45], PyramidalFlow [18], SimpleNet [24], and CFLOW-AD [13]. Notably, some of the results of the above methods are obtained from ADer [40]. + +Quantitative Comparison. As shown in Table 1, our method achieves better performance on the MVtec-AD dataset across all metrics except AU-PRO, outperforming the current SoTA method (MambaAD). On the challenging VisA dataset, our method slightly underperforms MambaAD in pixel-level F1_max, yet achieves the best performance in all other metrics, with an overall improvement of 0.7 in the average metric (mAD). Notably, our method performs better on the real industrial dataset MPDD, which also proves the robustness of our method. Additionally, we present a comparison of the proposed Wave-MambaAD with other methods in terms of parameters, complexity, memory consumption, and inference speed in Table 2. Compared to other SoTA methods, our method has fewer parameters and complexity, faster inference speed, and less + +memory consumption. More quantitative comparisons are in the supplementary materials. + +Qualitative Comparison. Figure 5 presents a qualitative comparison between the proposed method and other SoTA methods. Our method demonstrates superior accuracy in detecting subtle defects compared to the current SoTA method (MambaAD). Specifically, as illustrated by the examples of 'Pill', 'Pub', and 'Capsules' in Figure 5, our method achieves more precise defect localization, whereas MambaAD tends to mislocalize anomalies. Furthermore, for large-scale defects such as 'Cable' and 'Total_rust', our method provides a more complete and accurate localization of anomalous region. + +# 4.4. Ablation Studies + +Table 3 presents a series of ablation experiments conducted on the MVTec-AD dataset to validate the effectiveness of the proposed components. + +Effectiveness of Wavelet-Mamba Module. As shown in Table 3 (a) and (f), replacing the Wavelet-Mamba module with HSS from MambaAD resulted in a performance decline, which confirms the effectiveness of our method and highlights the advantages of combining Mamba with wavelet transform for anomaly detection. + +Effectiveness of High-Frequency State Space model. As illustrated in (b) and (f) of Table 3, removing the HFSS results in a performance decline, validating its effectiveness. Furthermore, we evaluate the importance of directional scanning in HFSS, as shown in (d) and (f) of Table 3. Replacing the directional scanning (horizontal, vertical, and diagonal) with uniform scanning in a single direction (horizontal or vertical) leads to a degradation in model performance, which further demonstrates the effectiveness of us + +![](images/33b9164bef3adfa12d28a394cf67f18f57c4497c61db233fc1f119c2f61a26ec.jpg) +Figure 5. Qualitative visualization of pixel-level anomaly segmentation on anomaly datasets. Our method can accurately and completely locate anomalies. + +
SettingsImage-levelPixel-levelmAD
AU-ROCAPF1_maxAU-ROCAPF1_maxAU-PRO
(a)WM→HSS97.598.896.396.752.556.391.384.2
(b)w/o HFSS97.799.096.696.952.055.691.484.2
w/o LFSS98.099.196.996.953.256.691.684.6
(c)DSE→MLP97.999.196.897.054.557.391.884.9
(d)HFSS(Horizontal scan)97.999.297.197.254.857.992.785.0
HFSS(Vertical scan)97.899.196.897.154.156.891.885.3
HFSS(w/o Diagonal scan)98.199.397.497.355.458.092.885.5
(e)LFSS(Spatial-wise)98.099.096.997.154.057.891.784.9
(f)All98.299.397.497.555.958.393.185.6
+ +Table 3. Ablation studies of the proposed components. 'WM' denotes the Wavelet-Mamba Module. 'HSS' denotes the Hybrid State Space Block in MambaAD[14]. $\rightarrow$ is substitution. The best results are shown in bold. + +ing a directional scanning strategy for high-frequency. + +Effectiveness of Low-Frequency State Space model. By comparing (b) and (g) in Table 3, the model's overall performance decreases by 1.0 without the LFSS, confirming its effectiveness. Additionally, we replace the channel scanning in LFSS with spatial scanning, as in (e) in Table 3, and the model's performance decreases, indicating that the channelwise scanning is more effective for large-scale anomalies. + +Effectiveness of Dynamic Space Enhancement block. To validate the effectiveness of the DSE, we replace it with a multi-layer perception (MLP) network. As shown in (c) and (f) of Table 3, this results in a 0.7 drop in overall performance, highlighting the importance of supplementing spatial structure information extracted in the frequency domain. + +# 5. Conclusion + +In this work, we propose a wavelet-driven state space anomaly detector (Wave-MambaAD) to tackle the multiclass anomaly detection task. Our approach integrates three innovative components: (1) a high-frequency state space model with directional scanning to capture subtle defects; (2) a low-frequency state space model with channel-adaptive dynamic scanning to preserve global structure for accurate localization of large-scale anomalies; and (3) a dynamic spatial enhancement block that strengthens anomaly representations through inter-channel interaction and adaptive gating. Experiments show that Wave-MambaAD is a competitive anomaly detector. + +# Acknowledgements + +This work was supported by the National Key Research and Development Program of China (Grant No. 2021YFA1000102), the National Natural Science Foundation of China (Grant Nos. 62376285 and 61673396), and Natural Science Foundation of Shandong Province, China (Grant No. ZR2022MF260). + +# References + +[1] Paul Bergmann, Michael Fauser, David Sattlegger, and Carsten Steger. Mvtec ad-a comprehensive real-world dataset for unsupervised anomaly detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9592-9600, 2019. 6, 7 +[2] Paul Bergmann, Michael Fauser, David Sattlegger, and Carsten Steger. Uninformed students: Student-teacher anomaly detection with discriminative latent embeddings. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4183–4192, 2020. 6 +[3] Hongruixuan Chen, Jian Song, Chengxi Han, Junshi Xia, and Naoto Yokoya. Changemamba: Remote sensing change detection with spatio-temporal state space model. arXiv preprint arXiv:2404.03425, 2024. 3 +[4] Keyan Chen, Bowen Chen, Chenyang Liu, Wenyuan Li, Zhengxia Zou, and Zhenwei Shi. Rsmamba: Remote sensing image classification with state space model. IEEE Geoscience and Remote Sensing Letters, 2024. 3 +[5] Yajun Chen, Yuanyuan Ding, Fan Zhao, Erhu Zhang, Zhangnan Wu, and Linhao Shao. Surface defect detection methods for industrial products: A review. Applied Sciences, 11(16): 7657, 2021. 1 +[6] Hanqiu Deng and Xingyu Li. Anomaly detection via reverse distillation from one-class embedding. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9737-9746, 2022. 2 +[7] Omar Elharrouss, Noor Almaadeed, and Somaya Al-Maadeed. A review of video surveillance systems. Journal of Visual Communication and Image Representation, 77: 103116, 2021. 1 +[8] Daniel Y Fu, Tri Dao, Khaled K Saab, Armin W Thomas, Atri Rudra, and Christopher Ré. Hungry hungry hippos: Towards language modeling with state space models. arXiv preprint arXiv:2212.14052, 2022. 3 +[9] Hu Gao, Bowen Ma, Ying Zhang, Jingfan Yang, Jing Yang, and Depeng Dang. Learning enriched features via selective state spaces model for efficient image deblurring. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 710-718, 2024. 3 +[10] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023. 2, 3 +[11] Albert Gu, Karan Goel, and Christopher Ré. Efficiently modeling long sequences with structured state spaces. arXiv preprint arXiv:2111.00396, 2021. 3 +[12] Albert Gu, Isys Johnson, Karan Goel, Khaled Saab, Tri Dao, Atri Rudra, and Christopher Ré. Combining recurrent, + +convolutional, and continuous-time models with linear state space layers. Advances in neural information processing systems, 34:572-585, 2021. 3 +[13] Denis Gudovskiy, Shun Ishizaka, and Kazuki Kozuka. Cflow-ad: Real-time unsupervised anomaly detection with localization via conditional normalizing flows. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 98-107, 2022. 2, 3, 7 +[14] Haoyang He, Yuhu Bai, Jiangning Zhang, Qingdong He, Hongxu Chen, Zhenye Gan, Chengjie Wang, Xiangtai Li, Guanzhong Tian, and Lei Xie. Mambaad: Exploring state space models for multi-class unsupervised anomaly detection. arXiv preprint arXiv:2404.06564, 2024. 2, 3, 6, 7, 8 +[15] Haoyang He, Jiangning Zhang, Hongxu Chen, Xuhai Chen, Zhishan Li, Xu Chen, Yabiao Wang, Chengjie Wang, and Lei Xie. A diffusion-based framework for multi-class anomaly detection. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 8472-8480, 2024. 2, 3, 7 +[16] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 4, 6 +[17] Stepan Jezek, Martin Jonak, Radim Burget, Pavel Dvorak, and Milos Skotak. Deep learning-based defect detection of metal parts: evaluating current methods in complex conditions. In 2021 13th International congress on ultra modern telecommunications and control systems and workshops (ICUMT), pages 66-71. IEEE, 2021. 6, 7 +[18] Jiarui Lei, Xiaobo Hu, Yue Wang, and Dong Liu. Pyramidflow: High-resolution defect contrastive localization using pyramid normalizing flow. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14143-14152, 2023. 2, 3, 7 +[19] Xiaoyan Lei, Wenlong ZHang, and Weifeng Cao. Dvmsr: Distillated vision mamba for efficient super-resolution. arXiv preprint arXiv:2405.03008, 2024. 3 +[20] Yufei Liang, Jiangning Zhang, Shiwei Zhao, Runze Wu, Yong Liu, and Shuwen Pan. Omni-frequency channel-selection representations for unsupervised anomaly detection. IEEE Transactions on Image Processing, 2023. 2, 3 +[21] Tsung-Yi Lin, Piotr Dólar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2117-2125, 2017. 4 +[22] Jiarun Liu, Hao Yang, Hong-Yu Zhou, Yan Xi, Lequan Yu, Cheng Li, Yong Liang, Guangming Shi, Yizhou Yu, Shaoting Zhang, et al. Swin-umamba: Mamba-based unet with imagenet-based pretraining. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 615–625. Springer, 2024. 3 +[23] Yue Liu, Yunjie Tian, Yuzhong Zhao, Hongtian Yu, Lingxi Xie, Yaowei Wang, Qixiang Ye, and Yunfan Liu. Vmamba: Visual state space model. arXiv preprint arXiv:2401.10166, 2024. 3, 5 +[24] Zhikang Liu, Yiming Zhou, Yuansheng Xu, and Zilei Wang. Simplenet: A simple network for image anomaly detection + +and localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20402-20411, 2023. 1, 3, 7 +[25] Jonathan Parnay and Keng Chai. Inpainting transformer for anomaly detection. In International Conference on Image Analysis and Processing, pages 394-406. Springer, 2022. 3 +[26] Oliver Rippel, Patrick Mertens, and Dorit Merhof. Modeling the distribution of normal data in pre-trained deep features for anomaly detection. In 2020 25th International Conference on Pattern Recognition (ICPR), pages 6726-6733. IEEE, 2021. 3 +[27] Nicolae-Cătălin Ristea, Neelu Madan, Radu Tudor Ionescu, Kamal Nasrollahi, Fahad Shahbaz Khan, Thomas B Moeslund, and Mubarak Shah. Self-supervised predictive convolutional attentive block for anomaly detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13576-13586, 2022. 3 +[28] Karsten Roth, Latha Pemula, Joaquin Zepeda, Bernhard Schölkopf, Thomas Brox, and Peter Gehler. Towards total recall in industrial anomaly detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14318-14328, 2022. 2, 3 +[29] Abdullah Ali Salamai, Nouran Ajabnoor, Waleed E Khalid, Mohammed Maqsood Ali, and Abdulaziz Ali Murayr. Lesion-aware visual transformer network for paddy diseases detection in precision agriculture. European Journal of Agronomy, 148:126884, 2023. 1 +[30] Jimmy TH Smith, Andrew Warrington, and Scott W Linderman. Simplified state space layers for sequence modeling. arXiv preprint arXiv:2208.04933, 2022. 3 +[31] Jiangwei Weng, Zhiqiang Yan, Ying Tai, Jianjun Qian, Jian Yang, and Jun Li. Mamballie: Implicit retinax-aware low light enhancement with global-then-local state space. arXiv preprint arXiv:2405.16105, 2024. 3 +[32] Yi Xiao, Qiangqiang Yuan, Kui Jiang, Yuzeng Chen, Qiang Zhang, and Chia-Wen Lin. Frequency-assisted mamba for remote sensing image super-resolution. arXiv preprint arXiv:2405.04964, 2024. 3 +[33] Zhaohu Xing, Tian Ye, Yijun Yang, Guang Liu, and Lei Zhu. Segmamba: Long-range sequential modeling mamba for 3d medical image segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 578-588. Springer, 2024. 3 +[34] Zhongxing Xu, Feilong Tang, Zhe Chen, Zheng Zhou, Weishan Wu, Yuyao Yang, Yu Liang, Jiyu Jiang, Xuyue Cai, and Jionglong Su. Polyp-mamba: Polyp segmentation with visual mamba. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 510-521. Springer, 2024. 3 +[35] Zhiyuan You, Lei Cui, Yujun Shen, Kai Yang, Xin Lu, Yu Zheng, and Xinyi Le. A unified model for multi-class anomaly detection. Advances in Neural Information Processing Systems, 35:4571-4584, 2022. 2, 3 +[36] Ekim Yurtsever, Jacob Lambert, Alexander Carballo, and Kazuya Takeda. A survey of autonomous driving: Common practices and emerging technologies. IEEE access, 8:58443-58469, 2020. 1 + +[37] Vitjan Zavrtanik, Matej Kristan, and Danijel Skocaj. Draem - a discriminatively trained reconstruction embedding for surface anomaly detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 8330-8339, 2021. 1 +[38] Vitjan Zavrtanik, Matej Kristan, and Danijel Skočaj. Draem-a discriminatively trained reconstruction embedding for surface anomaly detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 8330-8339, 2021. 3, 6 +[39] Jiangning Zhang, Xuhai Chen, Yabiao Wang, Chengjie Wang, Yong Liu, Xiangtai Li, Ming-Hsuan Yang, and Dacheng Tao. Exploring plain vit features for multi-class unsupervised visual anomaly detection. Available at SSRN 4866147, .2, 3 +[40] Jiangning Zhang, Haoyang He, Zhenye Gan, Qingdong He, Yabiao Wang, Chengjie Wang, Lei Xie, Yong Liu, et al. A comprehensive library for benchmarking multi-class visual anomaly detection. . 7 +[41] Qiao Zhang, Yanliang Ge, Cong Zhang, and Hongbo Bi. Tprnet: camouflaged object detection via transformer-induced progressive refinement network. The Visual Computer, 39 (10):4593-4607, 2023. 1 +[42] Qiao Zhang, Xiaoxiao Sun, Yurui Chen, Yanliang Ge, and Hongbo Bi. Attention-induced semantic and boundary interaction network for camouflaged object detection. Computer Vision and Image Understanding, 233:103719, 2023. 1 +[43] Qiao Zhang, Mingwen Shao, and Xinyuan Chen. Poolmamba: Pooling state space model for low-light image enhancement. Neurocomputing, 635:130005, 2025. 3 +[44] Shaoting Zhang and Dimitris Metaxas. On the challenges and perspectives of foundation models for medical image analysis. Medical image analysis, 91:102996, 2024. 1 +[45] Ximiao Zhang, Min Xu, and Xiuzhuang Zhou. Realnet: A feature selection network with realistic synthetic anomaly for anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16699-16708, 2024. 7 +[46] Zhuoran Zheng and Chen Wu. U-shaped vision mamba for single image dehazing. arXiv preprint arXiv:2402.04139, 2024.3 +[47] Yang Zou, Jongheon Jeong, Latha Pemula, Dongqing Zhang, and Onkar Dabeer. Spot-the-difference self-supervised pretraining for anomaly detection and segmentation. In European Conference on Computer Vision, pages 392-408. Springer, 2022. 6, 7 \ No newline at end of file diff --git a/ICCV/2025/Wave-MambaAD_ Wavelet-driven State Space Model for Multi-class Unsupervised Anomaly Detection/images.zip b/ICCV/2025/Wave-MambaAD_ Wavelet-driven State Space Model for Multi-class Unsupervised Anomaly Detection/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..dbd60e2314537611f8b9c776c3b8e3d4fe87460c --- /dev/null +++ b/ICCV/2025/Wave-MambaAD_ Wavelet-driven State Space Model for Multi-class Unsupervised Anomaly Detection/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:098d9894b10f6ff7c3b24e605807eeef29447ca04ab3e61ffc3ef0e5fc451b9c +size 743432 diff --git a/ICCV/2025/Wave-MambaAD_ Wavelet-driven State Space Model for Multi-class Unsupervised Anomaly Detection/layout.json b/ICCV/2025/Wave-MambaAD_ Wavelet-driven State Space Model for Multi-class Unsupervised Anomaly Detection/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..b6b62450737d81f860e94004093f3f9b2102475b --- /dev/null +++ b/ICCV/2025/Wave-MambaAD_ Wavelet-driven State Space Model for Multi-class Unsupervised Anomaly Detection/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:006d28a7c097a0b82959fd82ae4ae48f1811dfe48f2ed8c26863dcd9a7aa048c +size 313865 diff --git a/ICCV/2025/WaveMamba_ Wavelet-Driven Mamba Fusion for RGB-Infrared Object Detection/c16d775a-5b9c-4030-b5af-7f8546eac6be_content_list.json b/ICCV/2025/WaveMamba_ Wavelet-Driven Mamba Fusion for RGB-Infrared Object Detection/c16d775a-5b9c-4030-b5af-7f8546eac6be_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..c3f5aa82fd3d19ec985c849ae76464af90a16c2b --- /dev/null +++ b/ICCV/2025/WaveMamba_ Wavelet-Driven Mamba Fusion for RGB-Infrared Object Detection/c16d775a-5b9c-4030-b5af-7f8546eac6be_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:656fc12ea83e80cd99746c83c3f17ae227b03433ffa7f4e4980ee5cd50373fc3 +size 86451 diff --git a/ICCV/2025/WaveMamba_ Wavelet-Driven Mamba Fusion for RGB-Infrared Object Detection/c16d775a-5b9c-4030-b5af-7f8546eac6be_model.json b/ICCV/2025/WaveMamba_ Wavelet-Driven Mamba Fusion for RGB-Infrared Object Detection/c16d775a-5b9c-4030-b5af-7f8546eac6be_model.json new file mode 100644 index 0000000000000000000000000000000000000000..238372b1ce5362f44c6bfc285ec7416a864ce3d2 --- /dev/null +++ b/ICCV/2025/WaveMamba_ Wavelet-Driven Mamba Fusion for RGB-Infrared Object Detection/c16d775a-5b9c-4030-b5af-7f8546eac6be_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:94c1860daa7344fdb2f7c75045b2d77d7ec3f0fb6e5412de16a6ba7e7eafb612 +size 112573 diff --git a/ICCV/2025/WaveMamba_ Wavelet-Driven Mamba Fusion for RGB-Infrared Object Detection/c16d775a-5b9c-4030-b5af-7f8546eac6be_origin.pdf b/ICCV/2025/WaveMamba_ Wavelet-Driven Mamba Fusion for RGB-Infrared Object Detection/c16d775a-5b9c-4030-b5af-7f8546eac6be_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..eb00183673c4eadc4a0c385d886caf1a51822efb --- /dev/null +++ b/ICCV/2025/WaveMamba_ Wavelet-Driven Mamba Fusion for RGB-Infrared Object Detection/c16d775a-5b9c-4030-b5af-7f8546eac6be_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:adda454b5a0cbc4c5ced49d9f3311260bc1630901b941077b193fbecc955c243 +size 9599916 diff --git a/ICCV/2025/WaveMamba_ Wavelet-Driven Mamba Fusion for RGB-Infrared Object Detection/full.md b/ICCV/2025/WaveMamba_ Wavelet-Driven Mamba Fusion for RGB-Infrared Object Detection/full.md new file mode 100644 index 0000000000000000000000000000000000000000..054d4127285bc7e8cff2c75c92fb0033a7594f33 --- /dev/null +++ b/ICCV/2025/WaveMamba_ Wavelet-Driven Mamba Fusion for RGB-Infrared Object Detection/full.md @@ -0,0 +1,330 @@ +# WaveMamba: Wavelet-Driven Mamba Fusion for RGB-Infrared Object Detection + +Haodong Zhu $^{1*}$ Wenhao Dong $^{1*}$ Linlin Yang $^{2\dagger}$ Hong Li $^{1}$ Yuguang Yang $^{1}$ Yangyang Ren $^{1}$ Qingcheng Zhu $^{1}$ Zichao Feng $^{1}$ Changbai Li $^{1}$ Shaohui Lin $^{4}$ Runqi Wang $^{3}$ Xiaoyan Luo $^{1\dagger}$ Baochang Zhang $^{1}$ $^{1}$ Beihang University, China $^{2}$ Communication University of China, China $^{3}$ Beijing Jiaotong University, China $^{4}$ East China Normal University, China + +# Abstract + +Leveraging the complementary characteristics of visible (RGB) and infrared (IR) imagery offers significant potential for improving object detection. In this paper, we propose WaveMamba, a cross-modality fusion method that efficiently integrates the unique and complementary frequency features of RGB and IR decomposed by Discrete Wavelet Transform (DWT). An improved detection head incorporating the Inverse Discrete Wavelet Transform (IDWT) is also proposed to reduce information loss and produce the final detection results. The core of our approach is the introduction of WaveMamba Fusion Block (WMFB), which facilitates comprehensive fusion across low-/high-frequency sub-bands. Within WMFB, the Low-frequency Mamba Fusion Block (LMFB), built upon the Mamba framework, first performs initial low-frequency feature fusion with channel swapping, followed by deep fusion with an advanced gated attention mechanism for enhanced integration. High-frequency features are enhanced using a strategy that applies an "absolute maximum" fusion approach. These advancements lead to significant performance gains, with our method surpassing state-of-the-art approaches and achieving average mAP improvements of $4.5\%$ on four benchmarks. + +# 1. Introduction + +The fusion of multi-spectral features from visible (RGB) and infrared (IR) imagery has emerged as a pivotal technique in computer vision. This is especially applicable to scenarios in challenging environments where traditional + +![](images/b68d8e7ab48ea86752cba0e4eaab751d5c09643395b2df3e8c8a5b4fe77b9c62.jpg) + +![](images/5576820653a2c4bee6f43ebe119cc07382e42ea442705429addb56f0f85638a4.jpg) + +![](images/d5ea0acb076c774136549b2f7ad61243ea92bf6893835c859af4e07a4bf7a344.jpg) +Figure 1. (a) shows the averaged normalized information entropy of low- and high-frequency components from RGB and IR images in $M^3\mathrm{FD}$ val dataset using DWT. (b) depicts the RGB and IR features filtered by DWT in two directions, with low-frequency (L) and high-frequency (H) components obtained in each direction, resulting in four feature sub-bands: LL, LH, HL, and HH. Red and blue dashed boxes highlight the low-frequency and high-frequency components, respectively. The first row shows RGB and the second row shows IR components. (c) shows the result of fusion of each frequency. + +methods alone may be insufficient [21, 56, 63]. RGB-based methods often encounter limitations in low-light or adverse + +weather conditions, as RGB images tend to lose detail and contrast. Conversely, IR images excel in these conditions by capturing thermal signatures that are independent of visible light, thus offering complementary information [15, 36, 76]. However, the integration of these distinct modalities (i.e., RGB and IR) into a cohesive framework poses significant challenges, including the effective fusion of heterogeneous data and maintaining computational efficiency. + +Over the past decades, numerous approaches have been developed to integrate information from visible and infrared images for cross-modality object detection. These approaches can be typically classified into CNN-based [31, 49, 57, 62], Transformer-based [7, 23, 39, 72], and Mamba-based [10, 34, 84] methods. While CNN-based methods have demonstrated success, they often struggle to capture long-range dependencies, leading to a growing interest in Transformer-based techniques. Recently, the high computational complexity of Transformer models has further motivated the development of Mamba-based methods, offering a more efficient alternative. For instance, Gao et al. [10] introduced MambaST, an efficient cross-spectral spatial-temporal fusion pipeline that enhances pedestrian detection by employing a multi-head hierarchical patching and aggregation structure to fuse visible and thermal imagery. Therefore, we utilize the Mamba framework [33] as the foundation for our feature fusion due to its ability to balance efficiency and performance. + +While existing cross-modality object detection methods have shown promising performance, research on effectively using inherent features between modalities for complementary advantages to enhance performance remains insufficient. We endeavor to explore answers from the frequency domain. Specifically, as shown in Fig. 1(a), applying Discrete Wavelet Transform (DWT) to $M^3\mathrm{FD}$ validation set reveals that IR images exhibit higher normalized information entropy in the low-frequency sub-band, while RGB images exhibit greater entropy in the high-frequency subbands. The value of information entropy reflects the amount of information carried by the image. Generally speaking, the higher the information entropy, the richer the content of image is [58]. This statistical observation suggests that RGB images are more adept at capturing high-frequency details, in contrast to IR images, which predominantly encapsulate low-frequency signals. Furthermore, Fig. 1(b) illustrates the DWT outputs of P2-level RGB and IR features derived from the YOLOv8 backbone trained on each respective modality. The low- and high-frequency components are delineated by the red and blue dashed box, respectively. In the low-frequency region, IR features manifest with clearer shapes, while RGB features present more pronounced edges and contours in the high-frequency region, affirming their distinct frequency characteristics—an observation that aligns with prior studies [59]. Building on + +these insights, we aim to exploit the unique and complementary frequency characteristics of IR and RGB modalities to achieve enhanced feature fusion, thereby improving object detection performance. + +In this paper, we propose WaveMamba, as illustrated in Fig. 2, which leverages the DWT to map features into the wavelet domain, exploiting the complementary frequency characteristics of RGB and IR for efficient feature fusion. An improved detection head with Inverse DWT (IDWT) is then used to generate the final detection results, leading to improved detection accuracy. The main innovation of WaveMamba lies in WaveMamba Fusion Block (WMFB), which comprises two key components: the Low-frequency Mamba Fusion Block (LMFB) and the High-Frequency Enhancement (HFE) strategy. LMFB utilizes the exceptional low-frequency modeling capability of Mamba [35] to integrate low-frequency information. It comprises two modules: the Shallow Fusion Mamba (SFM) module, which facilitates lightweight cross-modality interaction by swapping portions of RGB and IR channels, and the Deep Fusion Mamba (DFM) module, which further refines the fusion process using a gated attention mechanism to filter out redundant information and improve the quality of the low-frequency feature fusion. Meanwhile, the HFE strategy enhances the selection and retention of critical details within high-frequency features by adopting an "absolute maximum" fusion approach. This combination of innovations significantly improves the fusion of features from both modalities, as shown in Fig. 1(c). The low-frequency features of RGB fusion are enriched, while the IR fusion features effectively suppress redundant noise in the originally low-frequency characteristics of IR, such as road shadows. Meanwhile, the fused high-frequency features become more detailed and pronounced. In summary, this paper's main contributions are: + +- We propose WaveMamba, which leverages the intrinsic advantageous frequency characteristics of different modalities to fuse features, facilitating sufficient feature interaction and enhancing the performance of RGB-IR object detection. +- WaveMamba Fusion Block efficiently integrates distinct frequency information, including a Low-frequency Mamba Fusion Block that combines low-frequency features and a High-Frequency Enhancement strategy to capture details without adding computational complexity. An improved YOLOv8 detection head utilizes IDWT for final detection results. +- Extensive experiments on four public datasets demonstrate that our approach achieves state-of-the-art performance in cross-modality object detection tasks, with an average mAP exceeding that of the second-best method by $4.5\%$ . + +# 2. Related Works + +Multi-modality Object Detection. Various multi-modality object detection methods have been proposed based on the complementary information from RGB-IR images. So far, these methods can be categorized into two main directions: pixel-based and feature-based. Pixel-based methods such as CDDFuse [82] and SuperFusion [45], involve fusing two images at the pixel level to obtain a single fused image, upon which object detection is then performed. Feature-based methods fuse the output from a certain stage of the two-stream detector, such as the early or middle stage features extracted by the backbone [2, 27, 29, 48, 57] or the results of the detector (i.e., late fusion) [5, 25, 28, 30]. However, the fusion modules designed in these methods are predominantly based on CNN or Transformer architectures. CNN's are constrained by their local receptive fields, while Transformer's face inefficiencies in model detection due to their complexity. Recently, the DMM framework, as described in [84], incorporates fusion modules designed to optimize and merge features from two modalities using Mamba, leading to successful outcomes in oriented object detection. Despite the excellent performance achieved by these methods, further enhancement in detection performance can be realized by utilizing distinct characteristics in the frequency domain between modalities. + +Wavelet Transform in Computer Vision. As wavelet transform can achieve lossless up- and down-sampling and obtain frequency domain information of images, it has been widely applied to various computer vision tasks [14, 47, 70, 86], resulting in performance improvements. WTConv [9] achieves a significant increase in the convolutional receptive field by integrating wavelet and inverse wavelet transforms into the convolution process. WF-Diff [77] attains cutting-edge performance in underwater image restoration by using wavelet transform to obtain the high- and low-frequency components of the image after preliminarily enhanced, and then sending them into two separate diffusion branches to adjust the high- and low-frequency information. DCFNet [60] realizes an improvement in the quality of fused images for the task of RGB and infrared image fusion. It incorporates wavelet transform and inverse wavelet transform as pre-processing steps for the Encoder and Decoder, respectively. However, these methods fail to distinguish different frequency domain characteristics of low- and high-frequency components, adopting either identical processing methods for both or without processing the high-frequency components altogether. Consequently, the extraction of frequency domain information is inadequate. + +Unlike the aforementioned methods, we utilize wavelet transform and its inverse transform in cross-modality object detection, designing distinct processing modules for high- and low-frequency components, resulting in significant improvements in detection efficiency and performance. + +# 3. Methods + +# 3.1. Preliminaries + +Discrete Wavelet Transform. Haar wavelet transform is a standard DWT due to its simplicity and efficiency, and we set it as the default. For an image $I$ , the one-level transform applies depth-wise convolution using low-pass $L = \frac{1}{\sqrt{2}} [1, 1]$ and high-pass $H = \frac{1}{\sqrt{2}} [1, -1]$ filters, followed by a down-sampling step that reduces the resolution by a factor of 2. Extending to two dimensions involves four filters, and we have sub-bands for $I$ as below: + +$$ +F _ {L}, F _ {H} = \mathrm {D W T} (I), \tag {1} +$$ + +where $\mathrm{DWT}(\cdot)$ denotes the Haar wavelet transform, subbands $F_{L} = \{F_{LL}\}$ captures low-frequency information and $F_{H} = \{F_{LH}, F_{HL}, F_{HH}\}$ captures high-frequency details and noise. + +2D Selective Scan. In linear time-invariant systems, state space models (SSMs) process 1D input signals through state variables and generate outputs using ODEs. Mamba [11] enhances this modeling approach, but its direct application to vision tasks is limited by the mismatch between 2D visual data and 1D language sequences. To address this, the 2D Selective Scan (SS2D) mechanism introduced in [33] expands image patches in four directions, creating independent sequences. Each feature sequence is processed using the Selective Scan Space State Sequential Model (S6) [11], and the sequences are then aggregated to reconstruct the 2D feature map. + +# 3.2. WaveMamba + +The architecture of our model, as depicted in Fig. 2, consists of a dual-stream feature extraction backbone augmented with the DWT, three WaveMamba Fusion Blocks, and an improved YOLOv8 detection head that incorporates the IDWT. + +Given the input RGB images $I_{\mathrm{RGB}}$ and infrared images $I_{\mathrm{IR}}$ , our model utilizes the first two layers of the dual-stream backbone to extract local features $f_{\mathrm{RGB}}^i$ and $f_{\mathrm{IR}}^i$ where the superscript $i \in \{1,2\}$ indicates the layer. $f_{\mathrm{RGB}}^2$ and $f_{\mathrm{IR}}^2$ undergo the Haar wavelet transformation and result in corresponding features in the wavelet domain: + +$$ +\begin{array}{l} F _ {L, \mathrm {R G B}} ^ {2}, F _ {H, \mathrm {R G B}} ^ {2} = \mathrm {D W T} \left(f _ {\mathrm {R G B}} ^ {2}\right), \tag {2} \\ F _ {L, \mathrm {I R}} ^ {2}, F _ {H, \mathrm {I R}} ^ {2} = \mathrm {D W T} \left(f _ {\mathrm {I R}} ^ {2}\right), \\ \end{array} +$$ + +where $F_{L,\mathrm{RGB}}^{2}$ and $F_{L,\mathrm{IR}}^{2}$ denote the low-frequency subbands, while $F_{H,\mathrm{RGB}}^{2}$ and $F_{H,\mathrm{IR}}^{2}$ represent the high-frequency sub-bands of the RGB and infrared modalities, respectively. With these frequency features combined with our proposed WMFB, we obtain fused features $F_{\mathrm{RGB}}^{3}$ and $F_{\mathrm{IR}}^{3}$ . Like [9], multi-level wavelet transformations are used + +![](images/d24f14168f414071fef3c94c34f496b33b6084e2159e3c0b142aca4d787c3959.jpg) +(a)Architecture of proposed WaveMamba + +![](images/01892816ce2dadbebf4eab3083d4f472eec561e725f6a11d721c4ee46f8f5549.jpg) +Figure 2. Illustrative diagram of the complete network and sub-modules. (a) shows the overall architecture of proposed WaveMamba network, which mainly includes the dual-stream feature extraction backbone improved by adding DWT, three WaveMamba Fusion Blocks and the improved YOLOv8 head augmented with IDWT. (b) shows the detailed structure of the WMFB, which includes the LMFB and the HFE strategy. (c), (d) and (e) show the detailed structure of the SFM module, DFM module and VSS block, respectively. + +in our framework to achieve more detailed multi-scale features. Specifically, DWT and WMFB are applied to the fused features at the $3^{rd}$ and $5^{th}$ layers. Finally, the fused low- and high-frequency sub-bands (i.e., all the output of WMFB) are fed into the improved YOLOv8 detection head, where IDWT is used for feature up-sampling, resulting in the final detection outputs. In the following sections, we introduce WMFB in subsec. 3.3 and the improved YOLOv8 head in subsec. 3.4. + +# 3.3. WaveMamba Fusion Block + +As shown in Fig. 1, we find that RGB images typically possess more high-frequency details, while IR images primarily capture low-frequency information. To effectively leverage the complementary frequency characteristics of both RGB and IR for feature fusion, we develop the WMFB. Specifically, WMFB consists of the Low-frequency Mamba Fusion Block and the High-Frequency Enhancement strategy for low- and high-frequency fusion, respectively. + +# 3.3.1. Low-frequency Mamba Fusion Block + +LMFB processes low-frequency features $F_{L,\mathrm{RGB}}^{i}$ , $F_{L,\mathrm{IR}}^{i}$ in $i^{th}$ layer to integrate low-frequency information from shallow to deep. In this case, LMFB designs the shallow fusion + +mamba module and deep fusion mamba module. + +SFM Module. This module is designed to enhance the interaction of low-frequency features across modalities for shallow feature fusion. It utilizes a channel swapping operation [13] in conjunction with VSS blocks [33]. We first integrate information from distinct channels to construct cross-modality feature correlations, thereby enriching the diversity of channel features and improving fusion performance. We begin with the channel swapping to derive new local features for low-frequency RGB and IR modalities, respectively. This can be formulated as follows: + +$$ +T _ {L, \mathrm {R G B}} ^ {i}, T _ {L, \mathrm {I R}} ^ {i} = \mathrm {C S} \left(F _ {L, \mathrm {R G B}} ^ {i}, F _ {L, \mathrm {I R}} ^ {i}\right), \tag {3} +$$ + +where $\mathrm{CS}(\cdot)$ denotes the channel swapping operation. This operation is implemented through channel splitting and concatenation to enhance the diversity of channel features. To further enhance the fused features, two VSS blocks are applied to $T_{L,\mathrm{RGB}}^i$ and $T_{L,\mathrm{IR}}^i$ separately as below: + +$$ +\tilde {F} _ {L, \mathrm {R G B}} ^ {i} = \mathrm {V S S} \left(T _ {L, \mathrm {R G B}} ^ {i}\right), \quad \tilde {F} _ {L, \mathrm {I R}} ^ {i} = \mathrm {V S S} \left(T _ {L, \mathrm {I R}} ^ {i}\right), \tag {4} +$$ + +where $\mathrm{VSS}(\cdot)$ refers to the VSS block, as shown in Fig. 2(e). The outputs $\tilde{F}_{L,\mathrm{RGB}}^{i}$ and $\tilde{F}_{L,\mathrm{IR}}^{i}$ represent the shallow fused features from the RGB and IR modalities, respectively. + +DFM Module. To further facilitate the interaction and deep fusion of low-frequency features between modalities, we introduce DFM based on the output of SFM as shown in Fig. 2(d). In DFM, the SFM outputs $(\tilde{F}_{L,\mathrm{RGB}}^{i},\tilde{F}_{L,\mathrm{IR}}^{i})$ alternate between primary and auxiliary roles. The primary modality is linearly embedded and split into two streams: one passes through a $3\times 3$ depth-wise convolution, SiLU activation, the core SS2D module, and layer normalization; the other undergoes only SiLU activation. The auxiliary modality follows the primary stream's first pathway. Given the complexity and diversity of features present across different modalities, it is crucial to selectively emphasize unique and complementary characteristics while minimizing redundancy. On this basis, we introduce a gated attention mechanism to effectively facilitate the learning of complementary features while suppressing redundant features. Specifically, this mechanism uses the output of the second information flow of the primary modality to regulate the outputs of both the first information flow of the primary modality and the auxiliary modality. By summing these regulated results, we obtain the final output of the DFM module for $i^{th}$ layer, formalized as follows: + +$$ +\hat {F} _ {L, \mathrm {R G B}} ^ {i}, \hat {F} _ {L, \mathrm {I R}} ^ {i} = \operatorname {D F M} \left(\tilde {F} _ {L, \mathrm {R G B}} ^ {i}, \tilde {F} _ {L, \mathrm {I R}} ^ {i}\right). \tag {5} +$$ + +# 3.3.2. High-Frequency Enhancement Strategy + +High-frequency sub-bands capture fine details in the original features, where larger absolute values indicate more significant information. To select and retain critical details, we employ a HFE strategy, which efficiently integrates high-frequency information by choosing pixels with larger absolute values. The HFE strategy for $i^{th}$ layer high-frequency sub-bands is defined as follows: + +$$ +\begin{array}{l} F _ {H} ^ {i} = \operatorname {M a s k} \left(\left| F _ {H, \mathrm {R G B}} ^ {i} \right| - \left| F _ {H, \mathrm {I R}} ^ {i} \right|\right) \odot F _ {H, \mathrm {R G B}} ^ {i} \tag {6} \\ + \operatorname {M a s k} \left(\left| F _ {H, I R} ^ {i} \right| - \left| F _ {H, R G B} ^ {i} \right|\right) \odot F _ {H, I R} ^ {i}, \\ \end{array} +$$ + +where $\odot$ is element-wise product operation. $\mathrm{Mask}(\cdot)$ generates an element-wise binary mask, where element is set to 1 if its corresponding input is greater than 0; otherwise, it is set to 0. Note that $F_{H}^{i} = \{F_{LH}^{i}, F_{HL}^{i}, F_{HH}^{i}\}$ based on Eq. 1. In this case, we process $F_{LH}^{i}, F_{HL}^{i}, F_{HH}^{i}$ independently and construct $F_{H}^{i}$ . Lastly, $F_{H}^{i}$ is the resulting merged high-frequency subbands. + +# 3.4. Improved YOLOv8 head + +Since our network architecture utilizes multilevel wavelet transformations to obtain multi-scale features, we intuitively design a corresponding YOLOv8 detection head structure using IDWT. Compared to traditional upsampling, IDWT reduces information loss and significantly improves the accuracy of detection results. + +As shown in Fig. 2(a), initially, we aggregate the fused low-frequency features from two modalities to get the final low-frequency feature. Subsequently, IDWT is utilized + +on this low-frequency feature alongside the fused high-frequency features. Aside from this modification, the rest in the detection head remain consistent with those in the original YOLOv8. Through these operations, we derive the ultimate detection results. + +# 4. Experiments + +# 4.1. Experimental Setups + +We evaluate our model on six datasets with $mAP_{50}$ and $mAP: M^3FD$ [32], DroneVehicle [44], LLVIP [17], FLIR-Aligned [75] in the main paper and VEIDA [37], KAIST [16] datasets in the supplementary. + +Moreover, for the FLIR-Aligned dataset, we also employ precision, recall, and F1 score as evaluation metrics. + +Our model and training framework are based on a modified dual-stream framework of the YOLO architecture by Ultralytics [50]. To verify the universality of our method, we implement our approach with three standard backbones: ResNet50, YOLOv5 and YOLOv8, following existing methods [39]. The loss function and training pipeline are identical to those used in standard YOLOv8 training [50]. More details can be found in the supplementary materials. + +# 4.2. Comparison with SOTA Results + +$\mathbf{M}^{3}\mathbf{F}\mathbf{D}$ Dataset. $M^{3}\mathrm{FD}$ is a multi-class object detection dataset in extreme weather conditions. The results on $M^{3}\mathrm{FD}$ are summarized in Table 1. Compared to existing state-of-the-art methods, our YOLOv8-based method achieves the highest performance, with improvements of $5.5\%$ and $5.1\%$ in $mAP_{50}$ and $mAP$ , respectively. Similarly, our method using ResNet50 and YOLOv5 as backbones also outperforms other methods using the same backbones. For performance in different categories, our method with different backbones consistently achieves top-three rankings in most cases. Specifically, our method outperforms the fourth-ranked method by $1.1\%$ , $6.0\%$ , and $8.5\%$ in the categories of cars, lamps, and people, respectively. Note that even without special design for rare classes, our method still achieves similar performance in these rare classes (e.g., "Bus", "Truck", "Motorcycle"). Compared to the single-modal YOLOv8l model, our fusion models achieve significant performance improvements, demonstrating the effectiveness of our method in integrating beneficial information from both the IR and RGB modalities. As our method effectively leverages the global structural information from the low-frequency sub-band of IR and the detailed information, such as edge textures of objects, from the high-frequency sub-bands of RGB, we demonstrate exceptional detection performance under extreme weather conditions. Furthermore, the consistent improvement across the three backbones indicates that our fusion approach is versatile and adaptable to various feature + +
MethodsBackbone\( {mAP}_{50} \)\( {mAP} \)CarBusLampPeopleTruckMotorcycle
(LOPET'24) LG+MDA [69]ResNet5066.440.283.084.437.762.575.554.6
(ECCV'24) DAMSDet [12]ResNet5080.252.9------
OursResNet5090.962.395.696.492.891.684.884.4
(CVPR'22) RFNet [65]YOLOv579.453.291.178.285.079.469.972.8
(CVPR'22) TarDAL [32]YOLOv580.554.194.881.387.181.568.769.3
(IJCAI'20) DIDFuse [80]YOLOv577.251.491.078.383.278.566.366.3
(IJCV'21) SDNet [74]YOLOv579.352.991.178.285.079.469.072.8
(MM'22) DetFusion [43]YOLOv580.853.892.583.087.880.871.469.4
(CVPR'23) CDDFuse [82]YOLOv580.553.691.682.486.081.171.271.1
(MM'23) IGNet [26]YOLOv581.554.592.882.486.981.672.173.0
(TCSVT'24) MMFN [67]YOLOv586.257.493.292.187.683.087.473.7
(CVPR'24) EMMA [83]YOLOv582.955.493.583.287.782.073.577.7
(TIM'24) KCDNet [51]YOLOv583.256.390.988.480.383.372.184.1
(TJRS'24) CRSIOD [53]YOLOv584.057.292.293.380.578.485.573.9
OursYOLOv591.563.895.495.493.291.887.885.7
YOLOv8l-IR [50]YOLOv879.553.190.090.963.082.985.964.6
YOLOv8l-RGB [50]YOLOv880.952.591.292.975.370.686.069.6
(JAS'22) SuperFusion [45]YOLOv783.556.091.093.270.083.785.877.4
(2024) TSJNet [19]YOLOv786.058.991.895.770.481.889.386.8
(YAC'24) RI-YOLO [4]YOLOv883.656.691.688.575.784.187.873.5
(Sensors'24) MRD-YOLO [40]YOLOv886.659.3------
OursYOLOv892.164.495.995.693.892.688.686.1
+ +Table 1. Comparison results with SOTA methods on $M^3$ FD dataset. The best results are highlighted in red. The second and third best results are highlighted in green and blue, respectively. + +
MethodsBackbone\( mAP_{50} \)\( mAP \)CarBusTruckFreight-carVan
(ECCV'22)TSFADet [73]ResNet5073.144.189.989.867.963.754.0
(TGRS'24) \( C^2 \)Former- \( S^2 \)ANet [72]ResNet5074.247.390.289.868.364.458.5
(GRSL'24) GLFNet [20]ResNet5071.454.890.388.072.753.652.6
(JST'24) FFODNet [55]ResNet5076.356.9-----
(2024) DMM [84]ResNet5077.255.890.488.777.863.066.0
(TGRS'24) LF-MDet [41]ResNet5071.851.3-----
OursResNet5079.359.994.690.279.668.064.1
(YAC'24) CAFN-IA [66]YOLOv569.356.189.190.862.057.347.1
(MTA'23) SLBAF-Net [6]YOLOv576.849.590.289.976.068.259.9
(CVPR'24) CSOM-ODAF [3]YOLOv577.150.190.189.875.668.261.8
(RSL'24) multimodal DINO [42]YOLOv572.550.389.588.875.454.354.3
OursYOLOv579.560.294.890.180.168.264.3
YOLOv8I-IR [50]YOLOv871.949.193.491.969.353.751.1
YOLOv8I-RGB [50]YOLOv870.248.692.591.768.847.850.2
(Sensors'23) Dual-YOLO [1]YOLOv771.955.295.991.669.755.946.6
(AEORS'24) YOLOFIV [52]YOLOv864.753.195.991.664.234.637.3
(Sensors'24) IV-YOLO [49]YOLOv874.656.897.294.365.463.153.0
(RS'24) DAAB-FFPN [54]YOLOv875.256.3-----
(TGRS'24) CRSIOD [53]YOLOv873.250.895.692.271.750.555.8
(IJAEOG'24) CMA [18]YOLOv876.850.495.893.175.959.859.4
OursYOLOv879.860.595.090.680.468.564.5
+ +Table 2. Comparison results with SOTA methods on Drone Vehicle dataset. The best results are highlighted in red. The second and third best results are highlighted in green and blue, respectively. + +extraction methods, thus providing stable performance improvements. + +DroneVehicle Dataset. DroneVehicle is a challenging remote sensing dataset with densely annotated images and small targets (e.g., "car", "bus", "van"). As shown in Table 2, our method achieves top-three rankings in both $mAP_{50}$ and $mAP$ metrics, surpassing the fourth-place method by $2.6\%$ and $3.6\%$ , respectively. Unlike other methods, which leverages domain-specific designs for remote sensing, our approach remains general yet still secures top-three positions in the categories of "truck" and "freight-car". Thanks to our unique low- and high-frequency fusion, + +our method excels at capturing edge details and achieves superior overall performance, highlighting its ability to enhance object detection across diverse scenarios without specialized remote sensing design. + +LLVIP Dataset. LLVIP is a low-light dataset for pedestrian detection. As presented in Table 3, we compare the performance of our approach on LLVIP dataset against state-of-the-art single-modal and fusion models. Our method achieves superior performance when using the same backbone. In particular, our YOLOv8-based method achieves remarkable state-of-the-art results, with $98.3\%$ in $mAP_{50}$ and $66.0\%$ in $mAP$ . + +
MethodsBackbone\( {mAP}_{50} \)\( {mAP} \)
(CVPRW'23) CSAA [2]ResNet5094.354.2
(2024) RSDet [79]ResNet5095.861.3
(TIV'24) MMSANet [24]ResNet5096.462.5
(CVPR'24) Text-IF [71]Transformer94.160.2
OursResNet5097.365.0
(IF'23) DIVFusion [46]YOLOv589.852.0
(PR'24) ICAFusion [39]YOLOv595.260.1
OursYOLOv597.665.2
YOLOv8l-IR [50]YOLOv895.262.1
YOLOv8l-RGB [50]YOLOv891.954.0
(ICHMS'24) MMDDet [68]YOLOv796.561.5
(RS'24) ACDF-YOLO [8]YOLOv896.561.3
(ICHMS'24) MSFF [68]YOLOv896.861.9
(Sensors'24) FAWDet [78]YOLOv897.162.1
OursYOLOv898.366.0
+ +Table 3. Comparison results with SOTA methods on LLVIP dataset. The best results are highlighted in red. The second and third best results are highlighted in green and blue, respectively. + +![](images/1e24c8d8007dd2625aba245790a7a2dcc8e5bd41761bb4ba40c1c7b3f6e62996.jpg) +RGB Input + +![](images/c7d278730c1ff10aacee262e83863dae5dff4cdf430ba2fcdbf9a4591359465a.jpg) +TarDAL + +![](images/65ef068a71b50413dff67eba23f20e05e3f602419d703d290c8d7ff79ea968e7.jpg) +SuperFusion + +![](images/1ea5345fbe66b8917d1ce8ce2ce5bcf5fd47611300d9277ad01bdfb5f9a1638b.jpg) +WaveMamba + +![](images/8b6e553595036407e3d33f9f884d1fd0e19ea7b797042c51e1b71d5dd9c8da67.jpg) +RGB Input + +![](images/119e2ef9a5ff2e2cb5e219b9bc90a64ddf2d9a02ccf5cd004eab71e1b1d100b3.jpg) +SLBAF-Net + +![](images/2dd7da8109d9f1ad7ca3c24cdfcc8f7a8ea617c60591d4aa4af3409eff8986af.jpg) +(a) $\mathbf{M}^3\mathbf{FD}$ +C²Former-S²ANet +(b) DroneVehicle + +![](images/d36df6d884c5f6c170214d4ac9ceb521b2189f388b39f490bce7ce765672b92a.jpg) +WaveMamba +Figure 3. Heatmap visualization of several cross-modality object detection methods on $M^3\mathrm{FD}$ and DroneVehicle. + +FLIR-Aligned Dataset. To demonstrate the advantages of our approach, we compare our methods with state-of-the-art methods with more metrics. As shown in Table 4, our method outperforms other fusion methods using the same backbone, achieving superior accuracy, while also reducing the size of the parameters and enhancing the speed of inference. For example, compared to CrossFormer, our method surpasses $5.8\%$ in precision, $6.9\%$ in recall and $6.3\%$ in the F1 score, and reduces the model size by 294.4 million parameters and decreases the inference time by 35.9 milliseconds. MFPT and ESSFN exhibit a similar trend. This benefit is due to our fusion design and the linear time complexity of Mamba. This further highlights the effectiveness of our approach in optimizing computational resources and improving detection performance. + +# 4.3. Visualization + +Heatmaps. We visualize the heatmap of our model's first inverse wavelet transform layer based on Grad-CAM [38] and compare it with other state-of-the-art methods [6, 32, + +45, 72]. As shown in Fig. 3, our model effectively prioritizes the target by utilizing wavelet transform and feature fusion, minimizing unnecessary attention on the background. In contrast, others tend to focus on larger areas, including backgrounds, which renders small targets indistinguishable. + +Detection Results. We visualize our detection results and compare them with other state-of-the-art methods [6, 32, 45, 72]. As shown in Fig. 4, our method reduces the number of missed or false detections in challenging scenarios (i.e., adverse weather, low lighting, heavy occlusion, or dense detection conditions), achieving the superior detection performance. + +# 4.4. Ablation Study + +To verify the efficacy of each module and different fusion strategies, we perform ablation studies using the YOLOv8 backbone on the $M^3\mathrm{FD}$ dataset. More ablation studies such as the position and number of our WMFB are provided in the supplementary materials. + +Effects of the Improved Head. We compare our improved head with the original YOLOv8 head in Table 5. As we can see, utilizing our improved head results in a $1.1\%$ improvement in $mAP_{50}$ and $mAP$ , coupled with a decrease of $7.6\mathrm{M}$ in the number of parameters. The achievement can be attributed to the use of DWT and IDWT rather than down- and up-sampling. The use of DWT and IDWT preserves more detailed information, reduces information loss, and enhances the capture and reconstruction of features at multiple scales, resulting in improved overall model performance [9, 61, 70]. + +Effects of the SFM and DFM Modules. In Table 6, we ablate the impact of SFM and DFM Modules by removing these two modules separately. Removing SFM leads to a decrease in performance by $1.5\%$ in $mAP_{50}$ and $2.1\%$ in $mAP$ . Similarly, removing DFM results in a drop of $1.9\%$ in $mAP_{50}$ and $2.2\%$ in $mAP$ . This verifies the importance of SFM and DFM. Their fusion design, which progresses from shallow to deep levels, has exhibited a positive effect on the integration of features across various modalities. + +Effects of Different Fusion Strategies. To verify the effectiveness of our proposed fusion strategies, we evaluate them against a standard fusion strategy Avg [81] in Table 7. Here, ("A", "B") indicates using strategy "A" and "B" to fuse high-frequency and low-frequency components, respectively. Starting with a simple baseline (Avg for fusion without DWT) and then progressively incorporating DWT along with different fusion strategies. The baseline lacks DWT and is incompatible with our improved YOLOv8 head, thus all experiments use the original head for fair comparison. By comparing the 1st and 2nd rows, we demonstrate the effectiveness of RGB-IR fusion in the wavelet domain, and further comparisons (2nd to 5th rows) show that our final combination (5th row) + +
MethodsBackbonePrecisionRecallF1\( {mAP}_{50} \)\( {mAP} \)ParametersInference time (ms)
(TITS'23) MFPT [85]ResNet5079.372.976.080.041.9200.0M80.0
OursResNet5084.277.980.986.547.9193.2M53.2
(PRL'24) CrossFormer [22]YOLOv578.172.875.479.342.1340.0M80.0
OursYOLOv583.979.781.785.847.545.6M44.1
YOLOv8-IR [50]YOLOv875.165.369.972.938.343.7M22
YOLOv8-RGB [50]YOLOv871.662.266.666.328.243.7M22
(GRSL'24) ESSFN [64]YOLOv881.473.577.280.842.380.2M47.0
OursYOLOv884.280.982.588.448.169.1M40.0
+ +![](images/9717349b3d80bfd5fe09225b058861175bbbb2bc994565e32c1f6a73e02f28e9.jpg) +Figure 4. Detection results' visualization of several cross-modality object detection methods on $M^3\mathrm{FD}$ and DroneVehicle. Wherein, (a)-(d) present the results of $M^3\mathrm{FD}$ dataset, and (e)-(h) present the results of DroneVehicle dataset. The targets encircled by red ellipses are false positives, while those encircled by green ellipses are missed detections. Please zoom in for more details. + +Table 4. Comparison results with SOTA methods on FLIR-Aligned dataset. The best results are highlighted in bold. + +
Methods\( {mAP}_{50} \)\( {mAP} \)Parameters
OriginYOLOv8 head91.063.376.7M
ImprovedYOLOv8 head92.164.469.1M
+ +Table 5. Effects of the improved head on $M^3\mathrm{FD}$ dataset. + +
Methods\( {mAP}_{50} \)\( {mAP} \)Parameters
w/o SFM module90.662.363.4M
w/o DFM module90.262.264.1M
ours92.164.469.1M
+ +achieves the best performance, effectively preserving both low-frequency spatial information and high-frequency detail characteristics of objects. + +# 5. Conclusion + +In this paper, we propose WaveMamba, a method that utilizes the unique and complementary frequency characteristics of different modalities to integrate features, prompting + +Table 6. Effects of the SFM and DFM on ${M}^{3}\mathrm{{FD}}$ dataset. + +
Methods\( {mAP}_{50} \)\( {mAP} \)Parameters
baseline83.255.168.9M
(Avg, Avg)86.658.266.3M
(Avg, LMFB)90.662.676.7M
(HFE, Avg)90.462.366.3M
(HFE, LMFB)91.063.376.7M
+ +Table 7. Effects of different fusion strategies in WaveMamba Fusion Blocks on $M^3$ FD dataset. + +extensive feature interaction and improving the effectiveness of RGB-IR object detection. The main innovation of WaveMamba lies in WMFB, which incorporates an LMFB to progressively integrate low-frequency features from shallow to deep layers, alongside a HFE strategy for efficient integration of high-frequency information. Extensive experiments on four distinct types of public RGB-IR datasets demonstrate that our method achieves state-of-the-art results while balancing parameters and efficiency. As a result, this work establishes a new baseline for RGB-IR object detection, setting a precedent for future research in the field. + +# Acknowledgements + +The work was supported by the National Key Research and Development Program of China (Grant No. 2023YFC3306401). This research was also supported by the Zhejiang Provincial Natural Science Foundation of China (Grant No. LD24F020007), Beijing Natural Science Foundation (Grant No. L223024 and L244043), National Natural Science Foundation of China (Grant No. 62076016, 62406298 and 62176068), "One Thousand Plan" projects in Jiangxi Province Jxsg2023102268, Beijing Municipal Science & Technology Commission, Administrative Commission of Zhongguancun Science Park Grant No.Z231100005923035, Taiyuan City "Double hundred Research action" 2024TYJB0127, the Fundamental Research Funds for the Central Universities (CUC25QT17). + +# References + +[1] Chun Bao, Jie Cao, Qun Hao, Yang Cheng, Yaqian Ning, and Tianhua Zhao. Dual-yolo architecture from infrared and visible images for object detection. Sensors, 2023. 6 +[2] Yue Cao, Junchi Bin, Jozsef Hamari, Erik Blasch, and Zheng Liu. Multimodal object detection by channel switching and spatial attention. In CVPRW, 2023. 3, 7 +[3] Chen Chen, Jiahao Qi, Xingyue Liu, Kangcheng Bin, Ruigang Fu, Xikun Hu, and Ping Zhong. Weakly misalignment-free adaptive feature alignment for uavs-based multimodal object detection. In CVPR, 2024. 6 +[4] Yishuo Chen, Boran Wang, Wenbin Zhu, and Jing Yuan. Rgb-ir yolo combining modality-specific reconstruction and information integration. In YAC, 2024. 6 +[5] Yi-Ting Chen, Jinghao Shi, Zelin Ye, Christoph Mertz, Deva Ramanan, and Shu Kong. Multimodal object detection via probabilistic ensembling. In ECCV, 2022. 3 +[6] Xiaolong Cheng, Keke Geng, Ziwei Wang, Jinhu Wang, Yuxiao Sun, and Pengbo Ding. Slbaf-net: Super-lightweight bimodal adaptive fusion network for uav detection in low recognition environment. Multimedia Tools Appl., 2023. 6, 7 +[7] Qingyun Fang, Dapeng Han, and Zhaokui Wang. Cross-modality fusion transformer for multispectral object detection. arXiv:2111.00273, 2021. 2 +[8] Xuan Fei, Mengyao Guo, Yan Li, Renping Yu, and Le Sun. Acdf-yolo: Attentive and cross-differential fusion network for multimodal remote. sens. object detection. Remote. Sens., 2024. 7 +[9] Shahaf E. Finder, Roy Amoyal, Eran Treister, and Oren Freifeld. Wavelet convolutions for large receptive fields. arXiv:2407.05848, 2024. 3, 7 +[10] Xiangbo Gao, Asiegbu Miracle Kanu-Asiegbu, and Xiaoxiao Du. Mambast: A plug-and-play cross-spectral spatial-temporal fuser for efficient pedestrian detection. arXiv:2408.01037, 2024. 2 +[11] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv:2312.00752, 2023. 3 + +[12] Junjie Guo, Chenqiang Gao, Fangcen Liu, Deyu Meng, and Xinbo Gao. Damsdet: Dynamic adaptive multispectral detection transformer with competitive query selection and adaptive feature fusion. In ECCV, pages 464-481, 2024. 6 +[13] Xuanhua He, Ke Cao, Jie Zhang, Keyu Yan, Yingying Wang, Rui Li, Chengjun Xie, Danfeng Hong, and Man Zhou. Panmamba: Effective pan-sharpening with state space model. CoRR'24, 2024. 4 +[14] Lida Huang, Gang Liu, Yan Wang, Hongyong Yuan, and Tao Chen. Fire detection in video surveillances using convolutional neural networks and wavelet transform. EAAI, 110: 104737, 2022. 3 +[15] Lian Huang, Zongju Peng, Fen Chen, Shaosheng Dai, Ziqiang He, and Kesheng Liu. Cross-modality interaction for few-shot multispectral object detection with semantic knowledge. Neural Networks, 2024. 2 +[16] Soonmin Hwang, Jaesik Park, Namil Kim, Yukyung Choi, and In So Kweon. Multispectral pedestrian detection: Benchmark dataset and baseline. In CVPR, pages 1037-1045, 2015. 5 +[17] Xinyu Jia, Chuang Zhu, Minzhen Li, Wenqi Tang, and Wenli Zhou. Llvip: A visible-infrared paired dataset for low-light vision. In ICCVW, 2021. 5 +[18] Chenchen Jiang, Huazhong Ren, Hong Yang, Hongtao Huo, Pengfei Zhu, Zhaoyuan Yao, Jing Li, Min Sun, and Shihao Yang. M2fnet: Multi-modal fusion network for object detection from visible and thermal infrared images. Int. J. Appl. Earth Obs. Geoinf., 2024. 6 +[19] Yuchan Jie, Yushen Xu, Xiaosong Li, and Haishu Tan. Tsjnet: A multi-modality target and semantic awareness joint-driven image fusion network. arXiv:2402.01212, 2024. 6 +[20] Xudong Kang, Hui Yin, and Puhong Duan. Global-local feature fusion network for visible-infrared vehicle detection. IEEE Geosci. Remote Sens. Lett., 2024. 6 +[21] Jung Uk Kim, Sungjune Park, and Yong Man Ro. Uncertainty-guided cross-modal learning for robust multispectral pedestrian detection. TCSVT, 2022. 1 +[22] Seungik Lee, Jaehyeong Park, and Jinsun Park. Crossformer: Cross-guided attention for multi-modal object detection. Pattern Recognit. Lett., 2024. 8 +[23] Wei-Yu Lee, Ljubomir Jovanov, and Wilfried Philips. Cross-modality attention and multimodal fusion transformer for pedestrian detection. In ECCVW, 2023. 2 +[24] Ang Li, Ziwei Wang, Fanxun Wang, Zhichao Liu, Guodong Yin, Ruiqi Fang, and Keke Geng. A novel semantic information perception architecture for extreme targets detection in complex traffic scenarios. IEEE Trans. Intell. Veh., 2024. 7 +[25] Chengyang Li, Dan Song, Ruofeng Tong, and Min Tang. Illumination-aware faster r-cnn for robust multispectral pedestrian detection. PR, 2019. 3 +[26] Jiawei Li, Jiansheng Chen, Jinyuan Liu, and Huimin Ma. Learning a graph neural network with cross modality interaction for image fusion. In ACMMM, 2023. 6 +[27] Qing Li, Changqing Zhang, Qinghua Hu, Huazhu Fu, and Pengfei Zhu. Confidence-aware fusion using dampstershafer theory for multispectral pedestrian detection. IEEE Trans. Multimed., 2023. 3 + +[28] Qing Li, Changqing Zhang, Qinghua Hu, Pengfei Zhu, Huazhu Fu, and Lei Chen. Stabilizing multispectral pedestrian detection with evidential hybrid fusion. TCSVT, 2024. 3 +[29] Ruimin Li, Jiajun Xiang, Feixiang Sun, Ye Yuan, Longwu Yuan, and Shuiping Gou. Multiscale cross-modal homogeneity enhancement and confidence-aware fusion for multispectral pedestrian detection. IEEE Trans. Multimed., 2024. 3 +[30] Xiangyang Li, Shiguo Chen, Chunna Tian, Heng Zhou, and Zhenxi Zhang. M2fnet: Mask-guided multi-level fusion for rgb-t pedestrian detection. IEEE Trans. Multimed., 2024. 3 +[31] Jingjing Liu, Shaoting Zhang, Shu Wang, and Dimitris N. Metaxas. Multispectral deep neural networks for pedestrian detection. In BMVC, 2016. 2 +[32] Jinyuan Liu, Xin Fan, Zhanbo Huang, Guanyao Wu, Risheng Liu, Wei Zhong, and Zhongxuan Luo. Target-aware dual adversarial learning and a multi-scenario multi-modality benchmark to fuse infrared and visible for object detection. In CVPR, 2022. 5, 6, 7 +[33] Yue Liu, Yunjie Tian, Yuzhong Zhao, Hongtian Yu, Lingxi Xie, Yaowei Wang, Qixiang Ye, and Yunfan Liu. Vmamba: Visual state space model. arXiv preprint arXiv:2401.10166, 2024. 2, 3, 4 +[34] Haolong Ma, Hui Li, Chunyang Cheng, Gaoang Wang, Xiaoning Song, and Xiaojun Wu. S4fusion: Saliency-aware selective state space model for infrared visible image fusion, 2024. 2 +[35] Xiaowen Ma, Zhenliang Ni, and Xinghao Chen. Tinyvim: Frequency decoupling for tiny hybrid vision mamba. arXiv preprint arXiv:2411.17473, 2024. 2 +[36] Fang Qingyun and Wang Zhaokui. Cross-modality attentive feature fusion for object detection in multispectral remote. sens. imagery. PR, 2022. 2 +[37] Sebastien Razakarivony and Frederic Jurie. Vehicle detection in aerial imagery: A small target detection benchmark. JVCIR, 34:187-203, 2016. 5 +[38] Ramprasaath R. Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In ICCV, pages 618-626, 2017. 7 +[39] Jifeng Shen, Yifei Chen, Yue Liu, Xin Zuo, Heng Fan, and Wankou Yang. Icafusion: Iterative cross-attention guided feature fusion for multispectral object detection. PR, 2024. 2, 5, 7 +[40] Chaoyue Sun, Yajun Chen, Xiaoyang Qiu, Rongzhen Li, and Longxiang You. Mrd-yolo: A multispectral object detection algorithm for complex road scenes. Sensors, 2024. 6 +[41] Xu Sun, Yinhui Yu, and Qing Cheng. Low-rank multimodal remote. sens. object detection with frequency filtering experts. IEEE Trans. Geosci. Remote Sens., 2024. 6 +[42] Xu Sun, Yinhui Yu, and Qing Cheng. Adaptive multimodal feature fusion with frequency domain gate for remote. sens. object detection. Remote Sens. Lett., 2024. 6 +[43] Yiming Sun, Bing Cao, Pengfei Zhu, and Qinghua Hu. Det-fusion: A detection-driven infrared and visible image fusion network. In ACMMM, page 4003-4011, 2022. 6 + +[44] Yiming Sun, Bing Cao, Pengfei Zhu, and Qinghua Hu. Drone-based rgb-infrared cross-modality vehicle detection via uncertainty-aware learning. TCSVT, 2022. 5 +[45] Linfeng Tang, Yuxin Deng, Yong Ma, Jun Huang, and Jiayi Ma. Superfusion: A versatile image registration and fusion network with semantic awareness. IEEE/CAA J. Autom. Sin., 2022. 3, 6, 7 +[46] Linfeng Tang, Xinyu Xiang, Hao Zhang, Meiqi Gong, and Jiayi Ma. Divfusion: Darkness-free infrared and visible image fusion. Inf. Fusion, 2023. 7 +[47] Chunwei Tian, Menghua Zheng, Wangmeng Zuo, Bob Zhang, Yanning Zhang, and David Zhang. Multi-stage image denoising with the wavelet transform. PR, 134:109050, 2023. 3 +[48] Chao Tian, Zukun Zhou, Yuqing Huang, Gaojun Li, and Zhenyu He. Cross-modality proposal-guided feature mining for unregistered rgb-thermal pedestrian detection. IEEE Trans. Multimed., 2024. 3 +[49] Dan Tian, Xin Yan, Dong Zhou, Chen Wang, and Wenshuai Zhang. Iv-yolo: A lightweight dual-branch object detection network. Sensors, 2024. 2, 6 +[50] Rejin Varghese and Sambath M. Yolov8: A novel object detection algorithm with enhanced performance and robustness. In ADICS, 2024. 5, 6, 7, 8 +[51] Haoyu Wang, Shiyuan Qu, Zhenzhuang Qiao, and Xiaomin Liu. Kcdnet: Multimodal object detection in modal information imbalance scenes. IEEE Trans. Instrum. Meas., 2024. 6 +[52] Huiying Wang, Chunping Wang, Qiang Fu, Binqiang Si, Dongdong Zhang, Renke Kou, Ying Yu, and Changfeng Feng. Yolofiv: Object detection algorithm for around-the-clock aerial remote. sens. images by fusing infrared and visible features. IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens., 2024. 6 +[53] Huiying Wang, Chunping Wang, Qiang Fu, Dongdong Zhang, Renke Kou, Ying Yu, and Jian Song. Cross-modal oriented object detection of uav aerial images based on image feature. IEEE Trans. Geosci. Remote Sens., 2024. 6 +[54] Jinpeng Wang, Nan Su, Chunhui Zhao, Yiming Yan, and Shou Feng. Multi-modal object detection method based on dual-branch asymmetric attention backbone and feature fusion pyramid network. Remote. Sens., 2024. 6 +[55] Jinpeng Wang, Congan Xu, Chunhui Zhao, Long Gao, Junfeng Wu, Yiming Yan, Shou Feng, and Nan Su. Multimodal object detection of uav remote sensing based on joint representation optimization and specific information enhancement. J-STARS, 17:12364-12373, 2024. 6 +[56] Kunpeng Wang, Zhengzheng Tu, Chenglong Li, Cheng Zhang, and Bin Luo. Learning adaptive fusion bank for multi-modal salient object detection. TCSVT, 2024. 1 +[57] Qingwang Wang, Yongke Chi, Tao Shen, Jian Song, Zifeng Zhang, and Yan Zhu. Improving rgb-infrared object detection by reducing cross-modality redundancy. Remote. Sens., 14, 2022. 2, 3 +[58] Zhiyu Wang, Jiayan Zhuang, Ningyuan Xu, Sichao Ye, Jiangjian Xiao, and Chengbin Peng. Image restoration quality assessment based on regional differential information entropy. arxiv2107.03642, 2022. 2 + +[59] Dan Wu, Mina Han, Yang Yang, Shan Zhao, Yujing Rao, Hao Li, Lin Xing, Chengjiang Zhou, and Haicheng Bai. Dcfusion: A dual-frequency cross-enhanced fusion network for infrared and visible image fusion. IEEE Trans. Instrum. Meas., 2023. 2 +[60] Dan Wu, Yanzhi Wang, Haoran Wang, Fei Wang, and Guowang Gao. Dcfnet: Infrared and visible image fusion network based on discrete wavelet transform and convolutional neural network. Sensors, 2024. 3 +[61] Guangrun Xiao, Xiaobo Wang, and Dezheng Liu. Wavelet transformation of functional data for hyperspectral image classification. In 2019 International Conference on Internet of Things (iThings) and IEEE Green Computing and Communications (GreenCom) and IEEE Cyber, Physical and Social Computing (CPSCom) and IEEE Smart Data (SmartData), pages 403-409, 2019. 7 +[62] Yumin Xie, Langwen Zhang, Xiaoyuan Yu, and Wei Xie. YOLO-MS: multispectral object detection via feature interaction and self-attention guided fusion. IEEE Trans. Cogn. Dev. Syst., 2023. 2 +[63] Zhengxuan Xie, Feng Shao, Gang Chen, Hangwei Chen, Qiuping Jiang, Xiangchao Meng, and Yo-Sung Ho. Cross-modality double bidirectional interaction and fusion network for RGB-T salient object detection. TCSVT, 2023. 1 +[64] Fengxiang Xu, Tingfa Xu, Lang Hong, Peiran Peng, Jiaxin Guo, and Jianan Li. Enhanced spectral-spatial fusion network for multispectral object detection in ground-aerial images. IEEE Geoscience and Remote. Sens. Letters, 2024. 8 +[65] Han Xu, Jiayi Ma, Jiteng Yuan, Zhuliang Le, and Wei Liu. Rfnet: Unsupervised network for mutually reinforcing multimodal image registration and fusion. In CVPR, 2022. 6 +[66] Junwei Xu, Bo Mo, Jie Zhao, Chunbo Zhao, Yimeng Tao, and Shuo Han. Cross-modal adaptive fusion object detection based on illumination-awareness. In YAC, 2024. 6 +[67] Fan Yang, Binbin Liang, Wei Li, and Jianwei Zhang. Multidimensional fusion network for multispectral object detection. TCSVT, 2024. 6 +[68] Huanyu Yang, Weiwei Bian, Jun Wang, Yuming Bo, and Ying Mi. A dual-modality pedestrian detection method based on multi-scale feature fusion. In ICHMS, 2024. 7 +[69] Jinqi Yang, Xin Yang, Yizhao Liao, Jinxiang Huang, Hongyu He, Erfan Zhang, Ya Zhou, and Yong Song. Multispectral sample augmentation and illumination guidance for RGB-T object detection by mm detection framework. In LOPET, 2024. 6 +[70] Ting Yao, Yingwei Pan, Yehao Li, Chong-Wah Ngo, and Tao Mei. Wave-vit: Unifying wavelet and transformers for visual representation learning. In ECCV, 2022. 3, 7 +[71] Xunpeng Yi, Han Xu, Hao Zhang, Linfeng Tang, and Jiayi Ma. Text-if: Leveraging semantic text guidance for degradation-aware and interactive image fusion. In CVPR, 2024. 7 +[72] Maoxun Yuan and Xingxing Wei. C²former: Calibrated and complementary transformer for rgb-infrared object detection. IEEE Trans. Geosci. Remote Sens., 2024. 2, 6, 7 +[73] Maoxun Yuan, Yinyan Wang, and Xingxing Wei. Translation, scale and rotation: Cross-modal alignment meets rgb-infrared vehicle detection. In ECCV, 2022. 6 + +[74] Hao Zhang and Jiayi Ma. Sdnet: A versatile squeeze-and-decomposition network for real-time image fusion. IJCV, 2021. 6 +[75] Heng Zhang, Élisa Fromont, Sébastien Lefèvre, and Bruno Avignon. Multispectral fusion for object detection with cyclic fuse-and-refine blocks. In ICIP, 2020. 5 +[76] Lu Zhang, Zhiyong Liu, Shifeng Zhang, Xu Yang, Hong Qiao, Kaizhu Huang, and Amir Hussain. Cross-modality interactive attention network for multispectral pedestrian detection. Inf. Fusion, 2019. 2 +[77] Chen Zhao, Weiling Cai, Chenyu Dong, and Chengwei Hu. Wavelet-based fourier information interaction with frequency diffusion adjustment for underwater image restoration. In CVPR, 2024. 3 +[78] Pujie Zhao, Xia Ye, and Ziang Du. Object detection in multispectral remote. sens. images based on cross-modal cross-attention. Sensors, 2024. 7 +[79] Tianyi Zhao, Maoxun Yuan, Feng Jiang, Nan Wang, and Xingxing Wei. Removal then selection: A coarse-to-fine fusion perspective for rgb-infrared object detection. arXiv:2401.10731, 2024. 7 +[80] Zixiang Zhao, Shuang Xu, Chunxia Zhang, Junmin Liu, Jiangshe Zhang, and Pengfei Li. Didfuse: Deep image decomposition for infrared and visible image fusion. In IJCAI, 2020. 6 +[81] Zixiang Zhao, Shuang Xu, Jiangshe Zhang, Chengyang Liang, Chunxia Zhang, and Junmin Liu. Efficient and model-based infrared and visible image fusion via algorithm unrolling. TCSVT, 2022. 7 +[82] Zixiang Zhao, Haowen Bai, Jiangshe Zhang, Yulun Zhang, Shuang Xu, Zudi Lin, Radu Timofte, and Luc Van Gool. Cddfuse: Correlation-driven dual-branch feature decomposition for multi-modality image fusion. In CVPR, 2023. 3, 6 +[83] Zixiang Zhao, Haowen Bai, Jiangshe Zhang, Yulun Zhang, Kai Zhang, Shuang Xu, Dongdong Chen, Radu Timofte, and Luc Van Gool. Equivariant multi-modality image fusion. In CVPR, 2024. 6 +[84] Minghang Zhou, Tianyu Li, Chaofan Qiao, Dongyu Xie, Guoqing Wang, Ningjuan Ruan, Lin Mei, and Yang Yang. Dmm: Disparity-guided multispectral mamba for oriented object detection in remote. sens. arXiv:2407.08132, 2024. 2, 3, 6 +[85] Yaohui Zhu, Xiaoyu Sun, Miao Wang, and Hua Huang. Multi-modal feature pyramid transformer for rgb-infrared object detection. IEEE Trans. Intell. Transport. Syst., 24(9): 9984-9995, 2023. 8 +[86] Wenbin Zou, Mingchao Jiang, Yunchen Zhang, Liang Chen, Zhiyong Lu, and Yi Wu. Sdwnet: A straight dilated network with wavelet transformation for image deblurring. In ICCVW, 2021. 3 \ No newline at end of file diff --git a/ICCV/2025/WaveMamba_ Wavelet-Driven Mamba Fusion for RGB-Infrared Object Detection/images.zip b/ICCV/2025/WaveMamba_ Wavelet-Driven Mamba Fusion for RGB-Infrared Object Detection/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..2cbb9437089f67f9f40ef410a5b33d03746d1d38 --- /dev/null +++ b/ICCV/2025/WaveMamba_ Wavelet-Driven Mamba Fusion for RGB-Infrared Object Detection/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f6e6ad6783889509a236034c227593e4fcb12a7de3aa8706a7a3dd0712a3139e +size 791665 diff --git a/ICCV/2025/WaveMamba_ Wavelet-Driven Mamba Fusion for RGB-Infrared Object Detection/layout.json b/ICCV/2025/WaveMamba_ Wavelet-Driven Mamba Fusion for RGB-Infrared Object Detection/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..af5422373190d1d2ed36bde5b8b8467a197d0955 --- /dev/null +++ b/ICCV/2025/WaveMamba_ Wavelet-Driven Mamba Fusion for RGB-Infrared Object Detection/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:161f12266a243a743376a02a26e55504c101712b1374628ea142c3de8ddab1bc +size 453232 diff --git a/ICCV/2025/Wavelet Policy_ Lifting Scheme for Policy Learning in Long-Horizon Tasks/61340c8d-ce4e-42e4-9ab7-a22c1a9e1133_content_list.json b/ICCV/2025/Wavelet Policy_ Lifting Scheme for Policy Learning in Long-Horizon Tasks/61340c8d-ce4e-42e4-9ab7-a22c1a9e1133_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..6b92b116484b1358b7092eb283e12fd52d68c546 --- /dev/null +++ b/ICCV/2025/Wavelet Policy_ Lifting Scheme for Policy Learning in Long-Horizon Tasks/61340c8d-ce4e-42e4-9ab7-a22c1a9e1133_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8386e891892eb8f4300e73a191a5c73258cc481c42f69d1b0f3a24caf810c3b7 +size 79393 diff --git a/ICCV/2025/Wavelet Policy_ Lifting Scheme for Policy Learning in Long-Horizon Tasks/61340c8d-ce4e-42e4-9ab7-a22c1a9e1133_model.json b/ICCV/2025/Wavelet Policy_ Lifting Scheme for Policy Learning in Long-Horizon Tasks/61340c8d-ce4e-42e4-9ab7-a22c1a9e1133_model.json new file mode 100644 index 0000000000000000000000000000000000000000..6569b017919522653e087d8acb1ce8ebb775f00a --- /dev/null +++ b/ICCV/2025/Wavelet Policy_ Lifting Scheme for Policy Learning in Long-Horizon Tasks/61340c8d-ce4e-42e4-9ab7-a22c1a9e1133_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6b3a6574100a5b61abc5efd30982a46e71ec0ccbc461594636e92e7c8e6aa48c +size 100473 diff --git a/ICCV/2025/Wavelet Policy_ Lifting Scheme for Policy Learning in Long-Horizon Tasks/61340c8d-ce4e-42e4-9ab7-a22c1a9e1133_origin.pdf b/ICCV/2025/Wavelet Policy_ Lifting Scheme for Policy Learning in Long-Horizon Tasks/61340c8d-ce4e-42e4-9ab7-a22c1a9e1133_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..c8b6c6b34b71fac1cb3f3aee3f7e44bd86ea23a4 --- /dev/null +++ b/ICCV/2025/Wavelet Policy_ Lifting Scheme for Policy Learning in Long-Horizon Tasks/61340c8d-ce4e-42e4-9ab7-a22c1a9e1133_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ffdaab44a644de17c19480149813b53398f7861953948b3d460ee11488d5b099 +size 3450269 diff --git a/ICCV/2025/Wavelet Policy_ Lifting Scheme for Policy Learning in Long-Horizon Tasks/full.md b/ICCV/2025/Wavelet Policy_ Lifting Scheme for Policy Learning in Long-Horizon Tasks/full.md new file mode 100644 index 0000000000000000000000000000000000000000..7067991f3e0f0af587589ffde6fe51fdaa695cc3 --- /dev/null +++ b/ICCV/2025/Wavelet Policy_ Lifting Scheme for Policy Learning in Long-Horizon Tasks/full.md @@ -0,0 +1,311 @@ +# Wavelet Policy: Lifting Scheme for Policy Learning in Long-Horizon Tasks + +Hao Huang\*1,3, Shuaihang Yuan\*1,2,3, Geeta Chandra Raju Bethala1,3, Congcong Wen1,3 Anthony Tzes2,3, Yi Fang $\dagger$ 1,2,3 + +$^{1}$ Embodied AI and Robotics (AIR) Lab, NYUAD $^{2}$ NYUAD Center for Artificial Intelligence and Robotics (CAIR) $^{3}$ New York University Abu Dhabi + +# Abstract + +Policy learning focuses on devising strategies for agents in embodied artificial intelligence systems to perform optimal actions based on their perceived states. One of the key challenges in policy learning involves handling complex, long-horizon tasks that require managing extensive sequences of actions and observations with multiple modes. Wavelet analysis offers significant advantages in signal processing, notably in decomposing signals at multiple scales to capture both global trends and fine-grained details. In this work, we introduce a novel wavelet policy learning framework that utilizes wavelet transformations to enhance policy learning. Our approach leverages learnable multi-scale wavelet decomposition to facilitate detailed observation analysis and robust action planning over extended sequences. We detail the design and implementation of our wavelet policy, which incorporates lifting schemes for effective multi-resolution analysis and action generation. This framework is evaluated across multiple complex scenarios, including robotic manipulation, self-driving, and multi-robot collaboration, demonstrating the effectiveness of our method in improving the precision and reliability of the learned policy. Our project is available at https://hhuang-code.github.io/wavelet_policy/. + +# 1. Introduction + +Policy learning in embodied artificial intelligence aims to generate optimal actions for an agent to achieve specific goals based on the observed states. This field has evolved from foundational methods such as behavior cloning [22, 53] and reinforcement learning [27, 40] to more complex approaches like hierarchical reinforcement learning [18, 30] and inverse reinforcement learning [19, 50]. Recent trends + +![](images/7a80d1e372f701918703c87194e1ba76c5019bfdbdd5f6d2f8d6c6f8315297b2.jpg) +Figure 1. A kitchen robot arm (left) conducts a series of actions, and we decompose the action sequences (blue) of the arm's $8^{th}$ (middle) and $9^{th}$ (right) joints into four frequencies ( $f-1$ to $f-4$ , red to orange) using Haar wavelet. Here $f-1$ represents high-frequency components of the original actions, while $f-4$ represents the low-frequency components of the actions. The action modes (trends) become clear in the low-frequency components without fluctuations (noises). + +![](images/210a39b2a8ffc984a1a9f1182e7c7eac7f9d5d770c9cbc2e7fdbfddc1f1eb656.jpg) + +![](images/966dcd579fdb405429ecfa4d21045ef7fc6244310199ea66d23484700a53dbae.jpg) + +in policy learning emphasize handling complex, multi-stage tasks that require high-precision operation and long-term decision-making in environments with multi-modal action distributions and high-dimensional action spaces [7, 16, 47]. Techniques such as implicit policy learning [15] which generates agent behavior through an implicit energy-based model [29] and hierarchical visual policy learning [57] tailored for long-horizon tasks in cluttered scenes, have illustrated the breadth of policy learning being developed. + +Despite the advances mentioned above, policy learning for embodied agents still faces significant challenges, particularly when dealing with complex, long-horizon tasks that require handling long observation and action sequences [6, 46]. In long-horizon scenarios, an agent must maintain consistent behavior across multiple steps, effectively managing dependencies over time, which can lead to an accumulation of errors if not properly handled. Furthermore, many real-world tasks exhibit multi-modal action patterns, i.e., there are often multiple valid action sequences to achieve the same goal [33, 47]. This diversity in valid actions creates more difficulties for the learning pro + +cess, which must generalize across different potential action plans without overfitting to a specific one. Lastly, certain applications require high precision in operations [14, 38], as small inaccuracies in control or planning can lead to significant issues, especially in delicate environments like medical diagnosis or industrial automation. Therefore, ensuring the learned policies to generate reliable, precise actions under complex conditions is one of the most challenging yet important aspects of policy learning. + +Wavelet analysis [9] is a powerful tool in signal processing by decomposing signals into multiple scales, capturing both global trends and localized details in time and frequency domains. Inspired by such characteristics of wavelet, we propose wavelet policy to learn optimal actions based on an agent's observations from a signal processing perspective. To be more specific, multi-scale decomposition produces the representations of an input signal at different time scales, which means that wavelet transform breaks down a complex long-term sequence into different frequency components, providing a high-level overview of the overall trend (i.e., low-frequency components) of the sequence as well as capturing detailed changes (i.e., high-frequency components) simultaneously. By analyzing global and local characteristics of the sequence, wavelet policy maintains consistency over long horizons and also focuses on critical changes at finer time resolutions. Furthermore, by decomposing the input signals into different frequency components and isolating low-frequency components from high-frequency components which usually contain noise, wavelet policy can enhance the visibility of distinct action modes embedded in the data. In addition, wavelet policy could also contribute to high precision by providing localized detail, as the high-frequency components obtained from wavelet decomposition often represent rapid changes in the data. + +More importantly, learning actions in a coarse-to-fine manner by adding details gradually not only lets the policy learn actions with more accuracy but also eases the learning process. As illustrated in Figure 1, the sequence of images in the left panel depicts a robot arm executing a sequence of actions in a kitchen scene [16]. The middle and right panels show the decomposition of actions for the arm's $8^{th}$ and $9^{th}$ joints, respectively. Specifically, we break down each action sequence (blue) using Haar wavelet at four different scales, labeled as $f-1$ to $f-4$ (red to orange), representing the transition from finer to coarser components. We notice that the curves of both joints at the coarsest level $f-4$ reveal several distinct modes and are smoother and less fluctuating than those at the finer levels. This indicates a dominant, stable trend in the action of the joints, possibly representing a key movement or trajectory segment necessary for successful task completion. Moreover, to generate the fine-grained actions, we can start by generating the smoothest actions from + +$f$ -4 scale and then add details gradually to the smoothest action sequence, which conceptually shares similarities with residual connection [20]. + +However, another two problems arise when directly integrating wavelet analysis into wavelet policy. First, manually selecting a certain type of wavelet, e.g., Haar, Daubechies, or Morlet, is heuristic, which could be less accurate. Second, the traditional wavelet analysis is a non-learnable process that lacks flexibility and generalizability to different action prediction tasks. To resolve these two problems, we design our wavelet policy based on lifting scheme [51], a technique for both designing wavelets and performing discrete wavelet transform. Our wavelet policy can be learned in an end-to-end manner, inheriting the advantages from both wavelet analysis and neural networks. We conduct experimental evaluations of our wavelet policy across five different simulation environments containing different complex tasks, including but not limited to self-driving simulations [13], sequential task completion in kitchen environments [16], multi-robot collaboration [36], etc. Our contributions are summarized as follows: + +1. To the best of our knowledge, it is the first attempt to introduce wavelet analysis to policy learning from a signal processing perspective. +2. We design a learnable wavelet policy network inspired by the lifting scheme and provide some key design insights, along with one instantiation of this network. +3. We evaluate our wavelet policy on five benchmarks of long-horizon tasks and achieve superior or comparable performance to the compared baselines. + +# 2. Related Work + +Imitation learning. Learning from expert demonstrations is an effective way to enable an agent to acquire the ability to interact with environments. OTR [34] assigns rewards to offline trajectories by using optimal transport to align unlabeled trajectories with expert demonstrations, generating an interpretable reward signal. PIRLNav [41] proposes a two-stage navigation policy learning approach that combines behavior cloning pretraining on human demonstrations with reinforcement learning fine-tuning. Diffusion policy [7] leverages a conditional denoising diffusion process [23] to model multi-modal action distributions via a learned gradient field, and [62] extends diffusion policy to 3D cases where the denoising process is trained on 3D point cloud scenes. BESO [42] introduces a policy representation for goal-conditioned imitation learning using a score-based model [28], enabling efficient goal-specified behavior generation. SDP [58] proposes a multi-task learning approach that uses Mixture of Experts [5] within a Transformer-based diffusion model to activate specific experts for each task. EMMA [60] learns to adapt a multi-modal agent in a visual world by distilling corrective feedback from a large + +language model. IBC [15] proposes implicit behavioral cloning with energy-based models [3, 29] for supervised policy learning, surpassing traditional explicit models. + +Learning for long-horizon tasks. Learning for long-horizon tasks focuses on developing robust policies to enable effective planning and task execution over extended temporal horizons. RPL [16] proposes a two-phase relay policy learning algorithm that uses a hierarchical, goal-conditioned approach with a high-level policy to set sub-goals and a low-level policy to act over shorter subgoal intervals. MIMICPLAY [56] first uses human play data to train a goal-conditioned latent planner that generates high-level task guidance and then utilizes limited teleoperation data to train a robot policy that follows the latent plans from the pre-trained planner. SPIRE [64] delegates complex actions to human teleoperators to collect demonstrations, which are then used to train an imitation learning policy. UVD [63] identifies visual subgoals in long-horizon manipulation tasks by detecting phase shifts in pre-trained visual embedding spaces to facilitate goal-based reward shaping. GSC [37] trains individual skill diffusion models and combines them linearly at test time to solve unseen long-horizon tasks. Hejna et al. [21] leverage language as auxiliary supervision to train agents with an instruction prediction loss, enhancing performance in complex, long-horizon planning tasks with limited data. Similarly, IGE-LLMs [54] uses large language models as an intrinsic reward source to guide exploration in sparse, long-horizon manipulation tasks. DHRL [31] decouples the horizons of high-level and low-level policies and adopts a graph to bridge the difference, enabling longer temporal abstraction and faster training. Behavior Transformer [47] adapts Transformer [55] with action discretization, allowing effective multi-modal continuous action prediction in long-horizon tasks. + +Wavelet for vision. Wavelet decomposes input signals to multi-scale frequencies and has been widely used in vision tasks such as image denoising, 3D shape representation, generative modeling, etc. MWDCNN [52] proposes a multi-stage image denoising convolution that incorporates a cascaded wavelet transform for effective noise suppression and detail recovery. WIRE [45] introduces a novel activation function based on the complex Gabor wavelet, enabling highly accurate and robust implicit neural representations. AWT-Net [24] proposes a multi-scale wavelet-based approach for 3D shape representation learning by hierarchically decomposing 3D shapes into frequency sub-bands using a lifting scheme [51]. WSGM [17] accelerates score-based generative models [48] by factorizing the data distribution into conditional probabilities of wavelet coefficients across scales, allowing efficient image synthesis. Hui et al. [25] propose a denoising diffusion process for 3D shape generation in the wavelet domain with a compact representation using coarse and detail coefficients to model shapes. + +![](images/91949283d2b559741443b649303bcdd397c5b0441ff459d226bdac52b5066825.jpg) +Figure 2. Lifting scheme shows the analysis stage (left) where the input $x$ is split and processed by predict $(P)$ and update $(U)$ functions to produce detail $d$ and approximation $s$ components, and the synthesis stage (right) where inverse functions $(\hat{U}, \hat{P})$ and merging operation reconstruct the original signal as $\hat{x}$ . + +Phung et al. [39] propose to decompose and adaptively process low- and high-frequency components at both pixel and feature levels for fast image generation. WaveNeRF [59] introduces wavelet-based radian fields for high-quality and generalizable novel view synthesis by integrating wavelet frequency decomposition into multi-view stereo images. Wave-ViT [61] integrates wavelet transforms within self-attention, enhancing self-attention outputs through inverse wavelet transforms to improve efficiency and accuracy in different vision tasks. + +# 3. Method + +In this section, we first briefly review wavelet and lifting scheme in Section 3.1, and then propose our wavelet policy in Section 3.2. Next, we describe some key designs in our wavelet policy in Section 3.3, and end with the loss functions for optimizing our policy in Section 3.4. + +# 3.1. Wavelet and Lifting Scheme + +Wavelet analysis [9] is a mathematical tool used to represent signals at multiple resolutions by decomposing the signals into high- and low-frequency components, which has proven essential for numerous applications in signal processing [35, 43, 49]. The classical wavelet transform involves two core functions: scaling function (or low-pass filter) $\phi(x)$ and wavelet function (or high-pass filter) $\psi(x)$ . The wavelet transform projects a signal $f(x)$ onto a set of scaled and translated scaling and wavelet functions: + +$$ +f (x) = \sum_ {k} c _ {j _ {0}, k} \phi_ {j _ {0}, k} (x) + \sum_ {j = j _ {0}} \sum_ {k} d _ {j, k} \psi_ {j, k} (x), \tag {1} +$$ + +where $c_{j_0,k}$ are approximation coefficients, $d_{j,k}$ are detail coefficients, and $j$ and $k$ denote scale and translation parameters, respectively. This decomposition, however, relies on convolution operations with filters derived from $\phi(t)$ and $\psi(t)$ , which is computationally intensive. + +The lifting scheme [10, 51] is an alternative, computationally efficient way to compute wavelet transforms. This + +![](images/83f1ff4ab622d82261d92822bb166cdb67c4c62d3e529caf3765c27c3dcab922.jpg) +Figure 3. An illustration of wavelet policy network with a multi-scale lifting scheme, where an observation sequence (left) is processed through splitting, prediction (P), and update (U) steps recursively, with converters and fusers recombining components to generate a corresponding action sequence (right). Blue lines and arrows indicate low-frequency (coarse-scale) components while red ones indicate high-frequency (fine-scale) components. Here, we only plot two scales for illustration. + +scheme consists of three main steps, as shown in the left panel of Figure 2: + +- Split: Divide the signal $x[n]$ into even and odd samples, i.e., $x_{\mathrm{even}} = x[2n]$ and $x_{\mathrm{odd}} = x[2n + 1]$ . +- Predict: Use the even samples to predict the odd samples, capturing high-frequency details: + +$$ +d = x _ {\text {o d d}} - P \left(x _ {\text {e v e n}}\right). \tag {2} +$$ + +- Update: Use the predicted details to update the even samples, capturing low-frequency approximation: + +$$ +s = x _ {\text {e v e n}} + U (d). \tag {3} +$$ + +Note that $P$ and $U$ are the predict and update functions, respectively. For instance, if we set $P(x) = x$ and $U(x) = \frac{x}{2}$ , we can recover the Haar wavelet transform. The lifting scheme maintains the core characteristics of classical wavelets while providing improved computational efficiency. Another key advantage of the lifting scheme is that we can design $P$ and $U$ as adaptive or learnable functions, as adopted in [8, 24, 44], to increase the flexibility of the transforms. However, these previous methods only focus on the analysis (or decomposition) stage while ignoring the synthesis (or reconstruction) stage, as shown in the right panel of Figure 2. The synthesis stage also contains three main steps, as symmetric to the analysis stage: + +- Inverse Update: Use details to reconstruct the even samples by removing the influence of $d$ from $s$ : + +$$ +\hat {x} _ {\text {e v e n}} = s - \hat {U} (d). \tag {4} +$$ + +- Inverse Predict: Approximate the odd samples based on the even samples: + +$$ +\hat {x} _ {\text {o d d}} = d + \hat {P} (\hat {x} _ {\text {e v e n}}). \tag {5} +$$ + +- Merge: Interleave the reconstructed even and odd samples to recover the original signal, i.e., $\hat{x}[2n] = \hat{x}_{\text{even}}$ and $\hat{x}[2n + 1] = \hat{x}_{\text{odd}}$ . + +Provided that there are no numerical inaccuracies or data loss during computation, the input signal can be always perfectly reconstructed with $P = \hat{P}$ and $U = \hat{U}$ . + +# 3.2. Wavelet Policy + +We design a wavelet policy rooted in the lifting scheme described in Section 3.1 to transform an observation sequence $S = \{s_t, \dots, s_{t+N}\}$ into a corresponding action sequence $A = \{a_t, \dots, a_{t+N}\}$ where $N$ is the window size (or context/history size), as shown in Figure 3. Given an observation sequence $S$ , a splitter decomposes the observation sequence into two streams $S_e^1$ and $S_o^1$ where the superscript $l = \{1,2,\dots,L\}$ denotes the level (or scale), and these streams are further processed through a learnable $P-Net$ and a learnable $U-Net$ , resulting in $S_s^1$ and $S_d^1$ streams. Distinct from the single scale scheme illustrated in Figure 2, we further forward $S_s^1$ into another splitter to repeat this process. Then, we adopt a series of converters to convert $S_s^L$ and $\{S_d^l\}$ into $A_s^L$ and $\{A_d^l\}$ . The converted two streams are undergone through another series of learnable $\hat{U}-Net$ and $\hat{P}-Net$ , and recombined into an action sequence by multiple fusers, which are also learnable networks. + +Formally, we consider each analysis block defined as: + +$$ +S _ {d} = S _ {o} - \mathcal {P} \left(S _ {e}, \left\{W _ {i} \right\}\right), S _ {s} = S _ {e} + \mathcal {U} \left(S _ {d}, \left\{W _ {i} \right\}\right), \tag {6} +$$ + +where the $W_{i}$ represents network weights to be learned. Similarly, a synthesis block is defined as: + +$$ +A _ {e} = A _ {s} - \hat {\mathcal {U}} (A _ {d}), \{W _ {j} \}) , A _ {o} = A _ {d} + \hat {\mathcal {P}} (A _ {e}, \{W _ {j} \}). \tag {7} +$$ + +The plus/minus operations in Eq. 6 and Eq. 7 are elementwise addition and subtraction. We omit superscript for conciseness. Instead of the traditional merge operation based on positional interleaving described in Section 3.1, we employ a learnable fuser to mix the two streams $A_{e}$ and $A_{o}$ at each scale. Note that the form of $\mathcal{P}, \mathcal{U}, \hat{\mathcal{P}}$ , and $\hat{\mathcal{U}}$ is flexible, provided that the following requirements are satisfied. + +# 3.3. Key Designs + +Issues of vanilla lifting architecture. First, temporal causality refers to the principle that current and future states/observations (or actions) are only causally influenced by past and present states, which is important in policy learning. However, the scheme expressed in Eq. 2 to Eq. 5 does not necessarily always satisfy the causality requirement, provided no constraint on $\mathcal{P}$ , $\mathcal{U}$ , $\hat{\mathcal{P}}$ and $\hat{\mathcal{U}}$ . Second, due to the even-odd split operation in each analysis block, for a lifting structure with $L$ scales, the input sequence at least has to be $2^{L}$ in length, if no padding is considered at each scale. This issue weakens the flexibility of the structure in processing input sequences with variable lengths. Third, if we adopt redundant lifting scheme, a variant of the basic lifting scheme described in Section 3.1 that modifies the traditional approach by not splitting the input sequence strictly into even and odd samples, the merge operation will be problematic, as it doubles the length of the input. + +Solutions. To resolve the issues mentioned above, we first instantiate $\mathcal{P},\mathcal{U},\hat{\mathcal{P}}$ and $\hat{\mathcal{U}}$ as causal convolution [12] with dilation, which is fundamentally used in temporal signal processing. It ensures that the output at time $t$ is computed only from inputs at times $t$ and earlier, preventing the model from violating the temporal order where future inputs influence past and present outputs, as shown in the top-left panel in Figure 4. For the second issue, we adopt the redundant lifting scheme, i.e., we instantiate the splitters as Transformers with self-attention and copy the output to both "even" and "odd" streams. However, if we examine the Eq. 3 to Eq. 5, it is worth noting that the input elements that are temporally spaced one apart from each other are processed together within $\mathcal{P},\mathcal{U},\hat{\mathcal{P}}$ and $\hat{\mathcal{U}}$ . Therefore, we introduce dilation into the causal convolution. Another benefit is that it allows the network to integrate input information over wider time intervals without increasing the number of parameters or the computational complexity. To address the third issue, instead of applying interleaving, we adopt a Transformer with cross-attention as the fuser to recombine the two streams together, as shown in the right panel of Figure 4. We set $Q = A_s^l$ and $K = V = A_d^l$ , such that it is consistent with the concept of reconstructing a signal from coarse to fine by adding high-frequency (details) to low-frequency (approximation) hierarchically. + +Converter. As shown in the middle of Figure 3, we introduce another (learnable) subnetwork, dubbed a converter, + +![](images/d37eef118c52342f79d726e390a00c4c5e28dea11de6155034a83e175306e771.jpg) +Figure 4. Illustration of: causal dilated convolution (top-left) to instantiate $\mathcal{P}$ , $\mathcal{U}$ , $\hat{\mathcal{P}}$ and $\hat{\mathcal{U}}$ ; causal moving average (bottom-left); Transformer-based structure of fusers (right). + +for each stream. The role of the converter is to explicitly convert the streams from observation space to action space. The concrete instantiations of the converter are flexible, e.g., causal convolution, attention with causal masks, and multi-layer perceptrons, provided they satisfy the temporal causality requirement. We can also instantiate the converter as an identity transformation, assuming the conversion between the two spaces is implicit. In our experiment, we adopt causal convolution as a converter. + +# 3.4. Loss Functions + +Task-specific terms. Given that we have access to the ground-truth action sequence from expert demonstrations, the task-specific loss is straightforward, such as cross-entropy loss or mean squared error loss. Here, we denoted as $\mathcal{L}_{\mathrm{task}}$ , and its concrete form is determined by the tasks. + +Wavelet-related terms. Different from the conventional manually designed $P$ , $U$ , $\hat{P}$ , and $\hat{U}$ that can achieve frequency decomposition effects, these learnable networks need optimization to form a valid lifting scheme. Specifically, the approximation stream primarily captures the low-frequency components, which represent the coarse information or the general trend of the data, and thus, we enforce each even/approximation stream to have the same local average as its corresponding stream in the upper scale: + +$$ +\mathcal {L} _ {\text {a p p r o x}} = \sum_ {i = t} ^ {t + N} \operatorname {S m o o t h} _ {L _ {1}} \left(\mathcal {C} \left(A _ {s} ^ {l}\right) - A _ {s} ^ {l + 1}\right), \tag {8} +$$ + +where $\mathcal{C}(\cdot)$ denotes causal moving average, as shown in the left-bottom panel in Figure 4. The detail stream should, on the one hand, preserve high-frequency components, and on the other hand, not overpower the approximation stream. Following [24], we adopt the loss as: + +$$ +\mathcal {L} _ {\text {d e t a i l}} = \sum_ {i = t} ^ {t + N} \operatorname {S m o o t h} _ {L _ {1}} \left(A _ {d} ^ {l}\right). \tag {9} +$$ + +The total loss for optimization is the sum of all the three above loss terms: + +$$ +\mathcal {L} = \mathcal {L} _ {\text {a p p r o x}} + \alpha \mathcal {L} _ {\text {a p p r o x}} + \beta \mathcal {L} _ {\text {d e t a i l}}, \tag {10} +$$ + +where $\alpha$ and $\beta$ balance the weights of the last two terms. + +# 4. Experiments + +# 4.1. Environments and Datasets + +CARLA self-driving. CARLA [13], build on Unreal Engine [2], creates a visually realistic simulated driving environment. The agent's action space is two-dimensional, allowing for acceleration/braking and left/right steering. The observation space consists of $224 \times 224$ RGB images captured from the car's perspective. In total, one hundred demonstration episodes navigate a building block in two opposite directions. Following [47], we use a frozen ResNet-18 encoder [20], pretrained on ImageNet [11], as the visual encoder to encode the observations into real-value vectors. + +Franka kitchen. To showcase the challenge of completing long-horizon tasks, we employ the Relay Kitchen Environment [16], where a Franka robot [1] operates within a simulated kitchen setting. The kitchen contains seven interactive objects: a microwave, a kettle, a sliding cabinet, a hinged cabinet, a light switch, and two burner knobs. This environment contains a dataset of 566 demonstrations gathered from human participants using VR headsets to complete a series of at least four object-interaction tasks [16, 47] with these objects. + +Push-T. Push-T is derived from IBC [15], aiming at pushing a T-shaped block toward a designated target using a circular end-effector. The task introduces variability through randomized starting positions for both the T-block and the end-effector. Successfully completing the task demands precise manipulation of the block by leveraging intricate, contact-heavy object dynamics via point contacts. Evaluations were conducted using 9 keypoints in a 2D space, extracted from the ground-truth pose of the T-block, alongside proprioceptive feedback for the end-effector's position. + +Transport. We select a difficult multi-robot collaboration task, i.e., Transport, from Robomimic [36], a robotic manipulation simulated environment for imitation learning and offline reinforcement learning evaluation. One variant of this task contains teleoperated demonstrations from a skilled operator (PH), while another one includes mixed demonstrations performed by both skilled and less skilled operators (MH). We present performance outcomes based on the state information inputs. + +D3IL. The D3IL simulation environment [26] is designed to evaluate policy learning algorithms on tasks that require diverse behaviors, multi-object manipulation, and closed-loop sensory feedback. The environment consists of five tasks: Avoiding, Aligning, Pushing, Sorting, and Stacking, + +each with distinct challenges such as obstacle avoidance, object alignment, multi-object sorting, and dexterous stacking. These tasks are designed to incorporate multiple valid policies, ensuring that models must learn multi-modal behavior rather than a single deterministic policy. + +# 4.2. Implementation + +We implement two variants of our wavelet policy: one integrated into Behavior Transformer (BeT) [47], and one integrated into Diffusion Policy with Transformer [7]. For the former one, we replace MinGPT [4] with our wavelet network while keeping all other components unchanged, ensuring a fair comparison. For the latter one, we replace the series of Transformer-based [55] decoders with our wavelet network while still using the Transformer-based encoders to encode the input to high-dimensional features. We tune the hyperparameters $\alpha$ and $\beta$ , and set them to 0.1 in most cases. All experiments are conducted with three different random seeds, and the average values and standard deviations of each metric are reported. + +# 4.3. Evaluation Results + +CARLA self-driving and Franka kitchen. Using the same metrics as [47], in the CARLA task, success is evaluated by the probability that the car reaches its target destination, and evaluations are conducted over 100 rollouts. For the Kitchen environment, performance is measured by the probability that the model completes up to $N$ tasks (from T1 to T5), i.e., interacts with up to $N$ objects successfully, within a limit of 280 timesteps. In total, 1,000 rollouts are evaluated for the Kitchen environment. + +The quantitative results are shown in Table 1. Our approach surpasses or closely matches the baselines BeT and VQ-BeT in all tasks in both CARLA and the Franka kitchen environment. In CARLA, we achieved a higher success rate $(0.847 \pm 0.090)$ compared to $0.832 \pm 0.167$ for BeT and $0.839 \pm 0.125$ for VQ-BeT, indicating a more reliable ability to reach the target destination across 100 rollouts. In the Kitchen tasks, performance is very similar for T1 to T3, but our method shows a noticeable improvement for T4 and T5, which suggests that our approach excels at handling more complex or long-horizon interactions. The reduced standard deviations in several cases (e.g., CARLA and T1 in Kitchen) also indicate stronger consistency of our results. + +Qualitative visualizations on the two tasks are shown in Figure 5. In the CARLA benchmark ( $1^{st}$ and $2^{nd}$ rows), the car, indicated by a red bounding box, successfully navigates through a simulated environment by driving to the right in the $1^{st}$ row and driving to the left in the $2^{nd}$ row. The Kitchen task ( $5^{th}$ and $6^{th}$ rows), features a Franka robotic arm performing complex interactions with kitchen appliances. The arm successfully completes tasks like manipulating the microwave, kettle, bottom knob, and hinge, + +
CARLA SuccessKitchen
T1T2T3T4T5
BeT [47]0.832±0.1670.948±0.0340.773±0.0650.562±0.0950.275±0.0660.027±0.023
VQ-BeT [32]0.839±0.1250.950±0.0210.775±0.0460.559±0.1050.306±0.0570.029±0.020
Ours0.847±0.0900.953±0.0200.775±0.0570.563±0.0630.339±0.0710.041±0.027
+ +showing the model's competence in handling multi-step, object-specific interactions. + +Multi-modal analysis. Following the protocols proposed in [47], for the CARLA task, we assess the likelihood of the car turning left versus right at intersections, excluding any out-of-distribution rollouts. For the Franka Kitchen task, we measure the empirical entropy of task sequences generated by the model, treating these sequences as strings. The results are shown in Table 2 in which the values closest to those observed in the demonstrations are highlighted. As shown in Table 2, our approach more closely recovers the multi-modal behavior distribution exhibited in the demonstrations for both CARLA and the Kitchen setting. In CARLA, the relative frequency of taking left versus right turns is closer to the 50-50 split observed in the demonstration data. In the Kitchen environment, our model's generated task sequences also achieve a higher entropy than BeT and VQ-BeT, bringing us slightly nearer to the 2.96 entropy level from the demonstrations. These results suggest that our approach produces behavior distributions better aligned to the diverse, multi-modal nature of the training data. + +Table 1. Quantitative results on CARLA and Franka kitchen tasks. + +
CARLAKitchen Entropy
LeftRight
BeT [47]0.293±0.1530.699±0.1542.506±0.078
VQ-BeT [32]0.315±0.1460.674±0.1452.508±0.075
Demonstrations0.500.502.96
Ours0.337±0.1260.662±0.1612.511±0.069
+ +Push-T and Transport. As reported in Table 3, our DP-Wavelet consistently outperforms DP-Transformer, with particularly notable gains in Transport-mh, suggesting it is more robust to variability in demonstration quality. The results imply that wavelet-based diffusion policies can effectively handle both uniform and mixed-skill data, leading to stronger overall performance. It also verifies that our wavelet policy can more effectively capture and utilize both local and global features relevant to pushing, positioning, and collaboration tasks. + +D3IL. As summarized in Table 4, our method offers consistent gains in success rate across all six tasks when compared + +Table 2. Statistics of multi-modal behavior modes learned from different methods on CARLA and Franka kitchen tasks. + +
Push-TTransport-phTransport-mh
DP-Transformer [7]0.942±0.0140.820±0.0550.440±0.068
DP-Wavelet (Ours)0.958±0.0160.835±-0.0620.497±0.085
+ +Table 3. Quantitative results of success rate on Push-T and Transport tasks. 'DP' denotes Diffusion Policy. + +to BeT, while maintaining or slightly increasing the policy entropy in most cases. In the Avoiding, Aligning, and Pushing tasks, our approach achieves notably higher success rates (e.g., $0.567 \pm 0.040$ vs. $0.463 \pm 0.163$ for Avoiding) alongside entropy values close to those of BeT, indicating more reliable task execution without collapsing into a single, repetitive strategy. In the Sorting-2 task, both methods display very low entropy, reflecting that the task can be solved effectively by repeating the same or similar actions. For the Pushing task, our method also exhibits higher noticeable success rates and marginally higher entropy, suggesting it can better handle the multi-step nature of stacking while still exploring diverse strategies. These findings imply that our wavelet policy attains higher overall performance while preserving behavioral diversity, consistent with the design goal of matching the distribution of behaviors in the data collection process. + +# 5. Ablation Study + +Effect of causal convolution. As discussed above in Section 3.3, causal dilated convolution is a key design factor. In Table 5, we replace causal convolution with conventional non-causal convolution in all $\mathcal{P}$ , $\mathcal{U}$ , $\hat{\mathcal{P}}$ , and $\hat{\mathcal{U}}$ , and notice a substantial drop in performance across all Kitchen tasks. This underscores the importance of preserving causal order in policy learning to properly handle sequential interactions during manipulation. + +Effect of learnable wavelet. We replace all the learnable networks of $\mathcal{P}$ , $\mathcal{U}$ , $\hat{\mathcal{P}}$ and $\hat{\mathcal{U}}$ shown in Figure 3 with two types of traditional, non-learnable wavelet, i.e., Haar and Daubechies 2 (DB2) wavelets, and the results are shown in Table 6. Note that we still keep Splitters and Convertors as learnable. The results suggest that using a flexible, data-driven wavelet representation enables the model to adaptively capture critical task features, whereas fixed wavelets + +![](images/0ca36611a045a8ebefd0342800d55d78ae57a5457375822275d0c8fdb437b726.jpg) +Figure 5. Visualization of agent trajectories in two benchmarks. The car (highlighted by a red box) in CARLA drives right $(1^{th}$ row) and left $(2^{nd}$ row). A Franka Arm accomplishes the task of Microwave - Kettle - Bottom knob - Hinge $(5^{th}$ row) and another task of Kettle - Light switch - Slide $(6^{th}$ row). Note that the last figure in the last row shows that the Arm tries to open the hinge, but fails. + +
AvoidingAligningPushing
Success RateEntropySuccess RateEntropySuccess RateEntropy
BeT [47]0.793±0.3320.837±0.1110.575±0.0500.458±0.0950.578±0.0640.774±0.043
IBC [15]0.523±0.0400.769±0.0360.633±0.0580.283±0.1010.535±0.0120.784±0.036
Ours0.820±0.2770.889±0.0310.660±0.0410.543±0.1090.778±0.0270.814±0.009
Sorting-2Stacking-1Stacking-2
Success RateEntropySuccess RateEntropySuccess RateEntropy
BeT [47]0.693±0.0190.001±0.0000.188±0.0960.293±0.0610.031±0.0100.078±0.005
IBC [15]0.456±0.0300.332±0.0480.192±0.1030.285±0.0530.025±0.0540.076±0.129
Ours0.723±0.1010.001±0.0000.197±0.0890.270±0.0770.041±0.0270.085±0.082
+ +Table 4. Quantitative results of six tasks in D3IL simulation environment. Entropy is used to qualify the diversity of behaviors learned by the methods. The higher value, i.e., approaching 1.0, indicates more diverse modes of behavior. + +
Kitchen
T1T2T3T4T5
Conv0.494±0.0990.259±0.0910.112±0.0570.033±0.0220.002±0.004
Causal0.953±0.0200.775±0.0570.563±0.0630.339±0.0710.041±0.027
+ +may not align as effectively with the nuances of the underlying state transitions. + +# 6. Conclusion + +We introduce a novel wavelet policy learning framework designed to enhance the precision and reliability of robotic tasks by utilizing multi-scale wavelet transformations. This + +Table 5. Comparison with conventional non-causal convolution. + +
Kitchen
T1T2T3T4T5
Haar0.884±0.0550.668±0.2170.535±0.0960.265±0.0780.029±0.022
DB20.831±0.0930.666±0.1140.473±0.1100.225±0.0790.029±0.036
Ours0.953±0.0200.775±0.0570.563±0.0630.339±0.0710.041±0.027
+ +Table 6. Comparison with non-learnable Haar, DB2 wavelets. + +method leverages a lifting scheme to decompose complex, long-term sequences of actions and observations, allowing for a nuanced understanding of behaviors. We evaluate our method on five simulation environments across different tasks, and the experimental results are promising and demonstrate its potential to broader application scenarios. + +# References + +[1] Franka Robotics. https://www.franka.de/.6 +[2] Unreal Engine. https://www.unrealengine.com/. 6 +[3] M Arbel, L Zhou, and A Gretton. Generalized energy based models. In International Conference on Learning Representations, 2021. 3 +[4] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in Neural Information Processing Systems, 33:1877-1901, 2020. 6 +[5] Weilin Cai, Juyong Jiang, Fan Wang, Jing Tang, Sunghun Kim, and Jiayi Huang. A survey on mixture of experts. arXiv preprint arXiv:2407.06204, 2024. 2 +[6] Yuanpei Chen, Chen Wang, Li Fei-Fei, and Karen Liu. Sequential dexterity: Chaining dexterous policies for long-horizon manipulation. In Annual Conference on Robot Learning, 2023. 1 +[7] Cheng Chi, Zhenjia Xu, Siyuan Feng, Eric Cousineau, Yilun Du, Benjamin Burchfiel, Russ Tedrake, and Shuran Song. Diffusion policy: Visuomotor policy learning via action diffusion. International Journal of Robotics Research, 2023. 1, 2, 6, 7 +[8] Roger L Claypoole, Geoffrey M Davis, Wim Sweldens, and Richard G Baraniuk. Nonlinear wavelet transforms for image coding via lifting. IEEE Transactions on Image Processing, 12(12):1449-1459, 2003. 4 +[9] Ingrid Daubechies and Bruce J Bates. Ten lectures on wavelets, 1993. 2, 3 +[10] Ingrid Daubechies and Wim Sweldens. Factoring wavelet transforms into lifting steps. Journal of Fourier Analysis and Applications, 4:247-269, 1998. 3 +[11] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 248–255. IEEE, 2009. 6 +[12] Sander Dieleman, Heiga Zen, Karen Simonyan, Oriol Vinyals, Alex Graves, Nal Kalchbrenner, Andrew Senior, Koray Kavukcuoglu, et al. Wavenet: A generative model for raw audio. arXiv preprint arXiv:1609.03499, 12, 2016.5 +[13] Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. Carla: An open urban driving simulator. In Annual Conference on Robot Learning, pages 1-16. PMLR, 2017. 2, 6 +[14] Michael Drolet, Simon Stepputtis, Siva Kailas, Ajinkya Jain, Jan Peters, Stefan Schaal, and Heni Ben Amor. A comparison of imitation learning algorithms for bimanual manipulation. Robotics and Automation Letters, 2024. 2 +[15] Pete Florence, Corey Lynch, Andy Zeng, Oscar A Ramirez, Ayzaan Wahid, Laura Downs, Adrian Wong, Johnny Lee, Igor Mordatch, and Jonathan Tompson. Implicit behavioral cloning. In Annual Conference on Robot Learning, pages 158-168. PMLR, 2022. 1, 3, 6, 8 +[16] Abhishek Gupta, Vikash Kumar, Corey Lynch, Sergey Levine, and Karol Hausman. Relay policy learning: Solving + +long-horizon tasks via imitation and reinforcement learning. In Annual Conference on Robot Learning, pages 1025-1037. PMLR, 2020. 1, 2, 3, 6 +[17] Florentin Guth, Simon Coste, Valentin De Bortoli, and Stephane Mallat. Wavelet score-based generative modeling. Advances in Neural Information Processing Systems, 35:478-491, 2022. 3 +[18] Danijar Hafner, Kuang-Huei Lee, Ian Fischer, and Pieter Abbeel. Deep hierarchical planning from pixels. Advances in Neural Information Processing Systems, 35:26091-26104, 2022. 1 +[19] Siddhant Haldar, Vaibhav Mathur, Denis Yarats, and Lerrel Pinto. Watch and match: Supercharging imitation with regularized optimal transport. In Annual Conference on Robot Learning, pages 32-43. PMLR, 2023. 1 +[20] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. 2, 6 +[21] Joey Hejna, Pieter Abbeel, and Lerrel Pinto. Improving long-horizon imitation through instruction prediction. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 7857-7865, 2023. 3 +[22] Jonathan Ho and Stefano Ermon. Generative adversarial imitation learning. Advances in Neural Information Processing Systems, 29, 2016. 1 +[23] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 2 +[24] Hao Huang and Yi Fang. Adaptive wavelet transformer network for 3d shape representation learning. In International Conference on Learning Representations, 2022. 3, 4, 5 +[25] Ka-Hei Hui, Ruihui Li, Jingyu Hu, and Chi-Wing Fu. Neural wavelet-domain diffusion for 3d shape generation. In SIGGRAPH Asia Conference, pages 1–9, 2022. 3 +[26] Xiaogang Jia, Denis Blessing, Xinkai Jiang, Moritz Reuss, Atalay Donat, Rudolf Lioutikov, and Gerhard Neumann. Towards diverse behaviors: A benchmark for imitation learning with human demonstrations. In International Conference on Learning Representations, 2024. 6 +[27] Lukasz Kaiser, Mohammad Babaeizadeh, Piotr Miłos, Błajej Osiński, Roy H Campbell, Konrad Czechowski, Dumitru Erhan, Chelsea Finn, Piotr Kozakowski, Sergey Levine, et al. Model based reinforcement learning for atari. In International Conference on Learning Representations, 2020. 1 +[28] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. Advances in Neural Information Processing Systems, 35:26565-26577, 2022. 2 +[29] Yann LeCun, Sumit Chopra, Raia Hadsell, M Ranzato, Fujie Huang, et al. A tutorial on energy-based learning. Predicting Structured Data, 1(0), 2006. 1, 3 +[30] Kyowoon Lee, Seongun Kim, and Jaesik Choi. Adaptive and explainable deployment of navigation skills via hierarchical deep reinforcement learning. In IEEE International Conference on Robotics and Automation, pages 1673-1679. IEEE, 2023. 1 + +[31] Seungjae Lee, Jigang Kim, Inkyu Jang, and H Jin Kim. Dhrl: a graph-based approach for long-horizon and sparse hierarchical reinforcement learning. Advances in Neural Information Processing Systems, 35:13668-13678, 2022. 3 +[32] Seungjae Lee, Yibin Wang, Haritheja Etukuru, H Jin Kim, Nur Muhammad Mahi Shafiullah, and Lerrel Pinto. Behavior generation with latent actions. In International Conference on Machine Learning, pages 26991-27008. PMLR, 2024. 7 +[33] Seungjae Lee, Yibin Wang, Haritheja Etukuru, H Jin Kim, Nur Muhammad Mahi Shafiullah, and Lerrel Pinto. Behavior generation with latent actions. In International Conference on Machine Learning, 2024. 1 +[34] Yicheng Luo, Samuel Cohen, Edward Grefenstette, Marc Peter Deisenroth, et al. Optimal transport for offline imitation learning. In International Conference on Learning Representations, 2023. 2 +[35] Stephane Mallat. A wavelet tour of signal processing, 1999. 3 +[36] Ajay Mandlekar, Danfei Xu, Josiah Wong, Soroush Nasiriyany, Chen Wang, Rohun Kulkarni, Li Fei-Fei, Silvio Savarese, Yuke Zhu, and Roberto Martin-Martin. What matters in learning from offline human demonstrations for robot manipulation. In Annual Conference on Robot Learning, pages 1678-1690. PMLR, 2022. 2, 6 +[37] Utkarsh Aashu Mishra, Shangjie Xue, Yongxin Chen, and Danfei Xu. Generative skill chaining: Long-horizon skill planning with diffusion models. In Annual Conference on Robot Learning, pages 2905-2925. PMLR, 2023. 3 +[38] Andrew S Morgan, Bowen Wen, Junchi Liang, Abdeslam Boullarias, Aaron M Dollar, and Kostas Bekris. Vision-driven compliant manipulation for reliable, high-precision assembly tasks. In Robotics: Science and Systems. MIT Press Journals, 2021. 2 +[39] Hao Phung, Quan Dao, and Anh Tran. Wavelet diffusion models are fast and scalable image generators. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10199-10208, 2023. 3 +[40] Lerrel Pinto, James Davidson, Rahul Sukthankar, and Abhinav Gupta. Robust adversarial reinforcement learning. In International Conference on Machine Learning, pages 2817-2826. PMLR, 2017. 1 +[41] Ram Ramrakhya, Dhruv Batra, Erik Wijmans, and Abhishek Das. Pirlnav: Pretraining with imitation and rl finetuning for objectnav. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17896-17906, 2023. 2 +[42] Moritz Reuss, Maximilian Li, Xiaogang Jia, and Rudolf Lioutikov. Goal-conditioned imitation learning using score-based diffusion policies. In Robotics: Science and Systems. MIT Press Journals, 2023. 2 +[43] Olivier Rioul and Martin Vetterli. Wavelets and signal processing. IEEE Signal Processing Magazine, 8(4):14-38, 1991. 3 +[44] Maria Ximena Bastidas Rodriguez, Adrien Gruson, Luisa Polania, Shin Fujieda, Flavio Prieto, Kohei Takayama, and Toshiya Hachisuka. Deep adaptive wavelet network. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3111-3119, 2020. 4 + +[45] Vishwanath Saragadam, Daniel LeJeune, Jasper Tan, Guha Balakrishnan, Ashok Veeraraghavan, and Richard G Baraniuk. Wire: Wavelet implicit neural representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18507-18516, 2023. 3 +[46] Pierre Sermanet, Corey Lynch, et al. Broadly-exploring, local-policy trees for long-horizon task planning. In Annual Conference on Robot Learning, 2021. 1 +[47] Nur Muhammad Shafiullah, Zichen Cui, Ariuntuya Arty Altanzaya, and Lerrel Pinto. Behavior transformers: Cloning $k$ modes with one stone. Advances in Neural Information Processing Systems, 35:22955-22968, 2022. 1, 3, 6, 7, 8 +[48] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In International Conference on Learning Representations, 2021. 3 +[49] D Sundararajan. Discrete wavelet transform: a signal processing approach. John Wiley & Sons, 2016. 3 +[50] Gokul Swamy, David Wu, Sanjiban Choudhury, Drew Bagnell, and Steven Wu. Inverse reinforcement learning without reinforcement learning. In International Conference on Machine Learning, pages 33299-33318. PMLR, 2023. 1 +[51] Wim Sweldens. The lifting scheme: A construction of second generation wavelets. SIAM Journal on Mathematical Analysis, 29(2):511-546, 1998. 2, 3 +[52] Chunwei Tian, Menghua Zheng, Wangmeng Zuo, Bob Zhang, Yanning Zhang, and David Zhang. Multi-stage image denoising with the wavelet transform. Pattern Recognition, 134:109050, 2023. 3 +[53] Faraz Torabi, Garrett Warnell, and Peter Stone. Behavioral cloning from observation. In Proceedings of the International Joint Conference on Artificial Intelligence, pages 4950-4957, 2018. 1 +[54] Eleftherios Triantafyllidis, Filippos Christianos, and Zhibin Li. Intrinsic language-guided exploration for complex long-horizon robotic manipulation tasks. In IEEE International Conference on Robotics and Automation, pages 7493-7500. IEEE, 2024. 3 +[55] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 3, 6 +[56] Chen Wang, Linxi Fan, Jiankai Sun, Ruohan Zhang, Li Fei-Fei, Danfei Xu, Yuke Zhu, and Anima Anandkumar. Mimicplay: Long-horizon imitation learning by watching human play. In Annual Conference on Robot Learning, pages 201–221. PMLR, 2023. 3 +[57] Hecheng Wang, Lizhe Qi, Bin Fang, and Yunquan Sun. Hierarchical visual policy learning for long-horizon robot manipulation in densely cluttered scenes. arXiv preprint arXiv:2312.02697, 2023. 1 +[58] Yixiao Wang, Yifei Zhang, Mingxiao Huo, Thomas Tian, Xiang Zhang, Yichen Xie, Chenfeng Xu, Pengliang Ji, Wei Zhan, Mingyu Ding, et al. Sparse diffusion policy: A sparse, reusable, and flexible policy for robot learning. In Annual Conference on Robot Learning, 2024. 2 + +[59] Muyu Xu, Fangneng Zhan, Jiahui Zhang, Yingchen Yu, Xiaogin Zhang, Christian Theobalt, Ling Shao, and Shijian Lu. Wavenerf: Wavelet-based generalizable neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18195-18204, 2023. 3 +[60] Yijun Yang, Tianyi Zhou, Kanxue Li, Dapeng Tao, Lusong Li, Li Shen, Xiaodong He, Jing Jiang, and Yuhui Shi. Embodied multi-modal agent trained by an llm from a parallel textworld. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26275-26285, 2024. 2 +[61] Ting Yao, Yingwei Pan, Yehao Li, Chong-Wah Ngo, and Tao Mei. Wave-vit: Unifying wavelet and transformers for visual representation learning. In European Conference on Computer Vision, pages 328-345. Springer, 2022. 3 +[62] Yanjie Ze, Gu Zhang, Kangning Zhang, Chenyuan Hu, Muhan Wang, and Huazhe Xu. 3d diffusion policy: Generalizable visuomotor policy learning via simple 3d representations. In IEEE International Conference on Robotics and Automation Workshop, 2024. 2 +[63] Zichen Zhang, Yunshuang Li, Osbert Bastani, Abhishek Gupta, Dinesh Jayaraman, Yecheng Jason Ma, and Luca Weihs. Universal visual decomposer: Long-horizon manipulation made easy. In IEEE International Conference on Robotics and Automation, pages 6973-6980. IEEE, 2024. 3 +[64] Zihan Zhou, Animesh Garg, Dieter Fox, Caelan Reed Garrett, and Ajay Mandlekar. Spire: Synergistic planning, imitation, and reinforcement learning for long-horizon manipulation. In Annual Conference on Robot Learning, 2024. 3 \ No newline at end of file diff --git a/ICCV/2025/Wavelet Policy_ Lifting Scheme for Policy Learning in Long-Horizon Tasks/images.zip b/ICCV/2025/Wavelet Policy_ Lifting Scheme for Policy Learning in Long-Horizon Tasks/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..7a7e2952a1313d77d817decc22efc9ddb5c11d60 --- /dev/null +++ b/ICCV/2025/Wavelet Policy_ Lifting Scheme for Policy Learning in Long-Horizon Tasks/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d6faa809616f596a268ce7a8330d80a82646eb799f9353a4bf510429d91dffa7 +size 497729 diff --git a/ICCV/2025/Wavelet Policy_ Lifting Scheme for Policy Learning in Long-Horizon Tasks/layout.json b/ICCV/2025/Wavelet Policy_ Lifting Scheme for Policy Learning in Long-Horizon Tasks/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..6dd4cf5020a8fa5643dc8aa508a9953a3c1b877d --- /dev/null +++ b/ICCV/2025/Wavelet Policy_ Lifting Scheme for Policy Learning in Long-Horizon Tasks/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:363695413c5f2d7d548fbef88363a1f5a4e9ce474815623065ddf57bb6e354f5 +size 423054 diff --git a/ICCV/2025/Weakly Supervised Visible-Infrared Person Re-Identification via Heterogeneous Expert Collaborative Consistency Learning/1b97471f-d27a-4c34-a569-593af9bedc9c_content_list.json b/ICCV/2025/Weakly Supervised Visible-Infrared Person Re-Identification via Heterogeneous Expert Collaborative Consistency Learning/1b97471f-d27a-4c34-a569-593af9bedc9c_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..6cc9c2761e0b0f99e702dff8c90160aae1d62e64 --- /dev/null +++ b/ICCV/2025/Weakly Supervised Visible-Infrared Person Re-Identification via Heterogeneous Expert Collaborative Consistency Learning/1b97471f-d27a-4c34-a569-593af9bedc9c_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ca11fb06f46feeadb253a60120556be6d225f101b8b8b8d1689bd0f427c93f8c +size 90079 diff --git a/ICCV/2025/Weakly Supervised Visible-Infrared Person Re-Identification via Heterogeneous Expert Collaborative Consistency Learning/1b97471f-d27a-4c34-a569-593af9bedc9c_model.json b/ICCV/2025/Weakly Supervised Visible-Infrared Person Re-Identification via Heterogeneous Expert Collaborative Consistency Learning/1b97471f-d27a-4c34-a569-593af9bedc9c_model.json new file mode 100644 index 0000000000000000000000000000000000000000..650df9f1273122368f0c1c77c7d46285ade7fd6a --- /dev/null +++ b/ICCV/2025/Weakly Supervised Visible-Infrared Person Re-Identification via Heterogeneous Expert Collaborative Consistency Learning/1b97471f-d27a-4c34-a569-593af9bedc9c_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9923e4c208b95af06ced28dcf37018ab35e6487d28a9f0fd7b6e9f88d4f086f4 +size 110667 diff --git a/ICCV/2025/Weakly Supervised Visible-Infrared Person Re-Identification via Heterogeneous Expert Collaborative Consistency Learning/1b97471f-d27a-4c34-a569-593af9bedc9c_origin.pdf b/ICCV/2025/Weakly Supervised Visible-Infrared Person Re-Identification via Heterogeneous Expert Collaborative Consistency Learning/1b97471f-d27a-4c34-a569-593af9bedc9c_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..369f2b82c3de3e2eb6d53f9ff4b6c9c539c741da --- /dev/null +++ b/ICCV/2025/Weakly Supervised Visible-Infrared Person Re-Identification via Heterogeneous Expert Collaborative Consistency Learning/1b97471f-d27a-4c34-a569-593af9bedc9c_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:edd9eea98249f9685e6c8db133cb119582a8bd06f11d5a0cbb6d41ac5fa89c56 +size 3393665 diff --git a/ICCV/2025/Weakly Supervised Visible-Infrared Person Re-Identification via Heterogeneous Expert Collaborative Consistency Learning/full.md b/ICCV/2025/Weakly Supervised Visible-Infrared Person Re-Identification via Heterogeneous Expert Collaborative Consistency Learning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..f2a0e0fc1d539deb7ce66b9b54d365990cc6799b --- /dev/null +++ b/ICCV/2025/Weakly Supervised Visible-Infrared Person Re-Identification via Heterogeneous Expert Collaborative Consistency Learning/full.md @@ -0,0 +1,352 @@ +# Weakly Supervised Visible-Infrared Person Re-Identification via Heterogeneous Expert Collaborative Consistency Learning + +Yafei Zhang $^{1}$ , Lingqi Kong $^{1}$ , Huafeng Li $^{1*}$ , Jie Wen $^{2}$ $^{1}$ Faculty of Information Engineering and Automation, Kunming University of Science and Technology + $^{2}$ School of Computer Science and Technology, Harbin Institute of Technology, Shenzhen + +zyfeimail@163.com, lingqikong2333@gmail.com, hfchina99@163.com, jiewen-pr@126.com + +# Abstract + +To reduce the reliance of visible-infrared person re-identification (ReID) models on labeled cross-modal samples, this paper explores a weakly supervised cross-modal person ReID method that uses only single-modal sample identity labels, addressing scenarios where cross-modal identity labels are unavailable. To mitigate the impact of missing cross-modal labels on model performance, we propose a heterogeneous expert collaborative consistency learning framework, designed to establish robust cross-modal identity correspondences in a weakly supervised manner. This framework leverages labeled data from each modality to independently train dedicated classification experts. To associate cross-modal samples, these classification experts act as heterogeneous predictors, predicting the identities of samples from the other modality. To improve prediction accuracy, we design a cross-modal relationship fusion mechanism that effectively integrates predictions from different experts. Under the implicit supervision provided by cross-modal identity correspondences, collaborative and consistent learning among the experts is encouraged, significantly enhancing the model's ability to extract modality-invariant features and improve cross-modal identity recognition. Experimental results on two challenging datasets validate the effectiveness of the proposed method. Code is available at https://github.com/KongLingqi2333/WSL-VIREID. + +# 1. Introduction + +Visible-Infrared Person Re-Identification (VIREID), a core task in cross-modal person re-identification (ReID), plays a pivotal role in intelligent security surveillance and smart city construction. This task involves matching pedestrian images from one modality (infrared or visible) with those from the other to accurately identify target pedestrians. It effectively overcomes the limitations of RGB-based ReID + +![](images/7512ab08228a082ad05047a1963a980a8e5c23275a362483ee220a0cab76de90.jpg) +Figure 1. VIReID faces two key challenges: asynchronous visible-infrared (VIS-IR) camera operation causes missing cross-modal identity correspondences, while large appearance differences raise annotation costs beyond those in single-modality scenarios. + +in nighttime and low-light conditions. However, due to the inherent differences in imaging principles between visible and infrared images, significant visual discrepancies exist between them. These modality-specific differences not only make annotating cross-modal datasets significantly more challenging and costly but also pose formidable challenges for cross-modal pedestrian retrieval. + +Existing approaches typically fall into three learning paradigms: fully supervised, semi-supervised, and unsupervised. Fully supervised approaches [10, 27, 28, 54] rely on accurately labeled cross-modal identity samples for model training. Semi-supervised approaches [30, 33, 56] leverage a combination of partially labeled and unlabeled data, thereby reducing the reliance on labeled samples. These approaches typically label samples from a single modality to alleviate the challenges of cross-modal labeling. Unsupervised approaches [4, 36, 43, 50] operate under the assumption that both intra-modal and inter-modal samples are unlabeled. Effective training requires accurate and reliable pseudo-label prediction to enable supervised learning. However, existing pseudo-label prediction methods are sensitive to the quantity of cross-modal positive samples and modality discrepancies, often leading to predictions with + +many noisy labels, which ultimately hinder model performance enhancement. + +In response to the aforementioned issues, this paper proposes a novel weakly supervised paradigm for VIREID. This paradigm relies solely on intra-modal identity labels, eliminating the need for cross-modal identity correspondence annotations, thereby substantially mitigating the challenges and costs associated with cross-modal labeling. As illustrated in Figure 1, this design is driven by two key challenges: First, infrared and visible cameras often operate at different times, making the direct establishment of cross-modal pedestrian correspondences challenging in real-world scenarios; second, the two modalities differ significantly in spectral characteristics and visual appearance, rendering manual cross-modal identity annotation both time-consuming and labor-intensive. In contrast, acquiring identity annotations within a single modality is comparatively easier. Therefore, by fully exploiting intramodal identity labels, this paper indirectly infers cross-modal identity correspondences. This strategy not only significantly reduces the cost of labeling but also effectively reduces the accumulation of label noise observed in existing methods. + +To establish robust cross-modal identity correspondences in a weakly supervised environment, this paper proposes a heterogeneous expert collaborative consistency learning framework. This framework leverages labeled samples within each modality to independently train dedicated classification experts. These classification experts serve as heterogeneous predictors, estimating pedestrian identities for cross-modal samples. Specifically, they predict and output the identities of cross-modal samples. To enhance prediction accuracy, we propose a cross-modal relationship fusion mechanism that integrates the predictions from both experts, forming accurate inter-modal identity correspondences. Guided by these correspondences, the model gradually enhances its modality-invariant feature representation through iterative optimization and collaborative consistency learning among experts. Throughout this process, the experts and the model mutually reinforce each other, effectively tackling the challenges of training without cross-modal identity correspondences. The main innovative contributions of this paper are as follows: + +- We explore VIREID under a weakly supervised setting for the first time and propose an effective mechanism for predicting cross-modal sample relationships. Our method requires annotations only for single-modal samples, eliminating the need for explicit cross-modal correspondence annotations. This avoids labor-intensive process of cross-modal annotation, significantly reducing annotation costs. +- We design a heterogeneous expert collaborative consistency learning framework. This framework constructs reliable cross-modal identity correspondences by integrat- + +ing predictions from experts, enabling supervised refinement of the ReID model. Furthermore, by dynamically guiding experts to make consistent identity predictions within and across modalities, our framework achieves collaborative and mutually reinforcing training between experts and the model. + +- We conduct extensive experimental validation on the SYSU-MM01, and LLCM datasets. Experimental results consistently validate the effectiveness of our method, achieving performance comparable to certain fully supervised approaches. + +# 2. Related Work + +# 2.1. Supervised VIREd + +The primary challenge in supervised VIREID is to overcome cross-modal feature mismatch caused by modality discrepancies. To address this issue, researchers have proposed various effective methods, which can be categorized into three types based on their strategies for handling modality differences [14]: modality-shared representation learning (MSRL) methods [2, 27, 47, 48], modality-specific information compensation (MSIC) methods [38, 53], and external information-assisted (EIA) methods [15, 17, 46]. MSRL methods aim to extract common features from both VIS-IR pedestrian images to mitigate the modality gap. Specifically, these methods can be classified into three categories: feature mapping [9, 29, 35, 54], which aligns the representations of IR-VIS images in a shared feature space; feature disentanglement [5, 8, 13, 16, 24], which often leverages adversarial learning to separate modality-invariant features while preserving identity-related information; and metric learning [6, 10, 21, 44, 45], which imposes identity consistency constraints to directly learn representations that are both identity-discriminative and modality-invariant. However, some of these methods may inadvertently lead to a loss of modality-specific identity-related information, ultimately weakening identity discrimination. + +To mitigate this issue, MSIC methods [34, 38, 53] generate missing modality information so that cross-modal samples of the same identity share consistent representations. Meanwhile, EIA methods leverage external cues (e.g., pedestrian poses [1, 23] and masks [7, 18, 55]) to enhance cross-modal consistency and reduce the impact of modality differences on identification. In a supervised setting, the success of these methods often depends on largescale labeled training data. However, due to modality differences, labeling extensive cross-modal datasets requires significant human resources. Thus, developing more cost-effective training paradigms is crucial. + +# 2.2. Semi-supervised VIREID + +To reduce dependence on cross-modal labeled samples, researchers have proposed semi-supervised VIREID methods. + +![](images/1724f3ab7cfdc2d1fa3465647cc1d7f48019438a28352bb6f07a4f4b5b6ecace.jpg) +Figure 2. Overview of the proposed method. HEL first performs intra-modality training to obtain experts for each modality, generating cross-modality prediction results for CRE. CRE then establishes robust cross-modality correspondences by fusing predictions from multiple experts. CCL utilizes the correspondences provided by CRE to facilitate CMCL and CLAE. + +These methods typically leverage labeled visible samples and unlabeled infrared samples for training. However, a key challenge lies in effectively establishing correspondences between samples from different modalities. To address this, the OTLA [33] method introduces an optimal transport strategy, modeling label correspondences as an optimal transport problem. However, OTLA may still introduce label noise in predictions. To mitigate this issue, DPIS [30] and MUCG [56] reweight the loss function based on pseudo-label confidence, reducing the adverse impact of pseudo-label noise on model performance. Nevertheless, the absence of infrared sample labels in semi-supervised methods increases the risk of label noise in cross-modal samples, thereby limiting further performance gains. + +# 2.3. Unsupervised VIREd + +To eliminate reliance on labeled samples, unsupervised VIREID methods have been proposed. These methods [20, 39] generate pseudo-labels via clustering within the same modality and establish cross-modal correspondences using similarity-based matching. However, cross-modal pseudo-label noise limits the performance of these methods. To address this, researchers have explored various strategies. Latent sample mining [4, 37, 43, 50] expands the search space for pseudo-label correspondences, reducing mismatches. Auxiliary information methods [3, 25, 36] use intermediate modality information to bridge and enhance cross-modal matching. Multi-level difference optimization [19, 26, 31, 40] decomposes cross-modal differences into intra- and inter-modality components, gradually improving representation consistency and reducing label noise. Cross + +modal relationship integration [12, 32, 41, 42, 49] refines pseudo-labels by aligning multiple correspondences. Despite these efforts, intra-modality label noise still affects inter-modality matching, which limits performance gains. This study explores VIRelID under a weakly supervised setting, addressing cross-modal labeling challenges in real-world scenarios while balancing annotation cost and model effectiveness. + +# 3. Methodology + +# 3.1. Problem Definition and Method Overview + +Given a dataset $\mathcal{X} = \{\mathcal{X}^v,\mathcal{X}^r\}$ comprising visible and infrared person images, where $\mathcal{X}^v = \{x_i^v\}_{i = 1}^{N^v}$ and $\mathcal{X}^r = \{x_i^r\}_{i = 1}^{N^r}$ denote the sets of $N^v$ visible and $N^r$ infrared person images, respectively. The corresponding label sets are given by $\mathcal{Y}^v = \{\pmb {y}_i^v\}_{i = 1}^{N^v}$ and $\mathcal{Y}^r = \{\pmb {y}_i^r\}_{i = 1}^{N^r}$ , where $\pmb {y}_i^v\in \mathbb{R}^{1\times C^v}$ and $\pmb {y}_i^r\in \mathbb{R}^{1\times C^r}$ represent the labels associated with samples $x_{i}^{v}$ and $x_{i}^{r}$ , respectively. Under the weakly supervised setting considered in this paper, label information is available within each modality, whereas cross-modal label correspondences remain unknown. + +To address this issue, we propose the method illustrated in Figure 2, which consists of heterogeneous expert learning (HEL), cross-modal relationship establishment (CRE), and collaborative consistency learning (CCL). During training, single-modality classifiers serve as heterogeneous experts to learn cross-modal consistency in feature representation. Cross-modal relationships are established by comparing the consistency of predictions made by these experts. Based on the correspondences, prototypes of the same identity across modalities are selected for collaborative learning among the + +experts. Meanwhile, leveraging these inter-modal correspondences, the encoder and shared classifier are further updated, enabling the network to extract features that maintain cross-modal consistency for the same identity. + +# 3.2. Heterogeneous Expert Learning + +To establish identity correspondences between cross-modal images, we propose a cross-modality judgment approach. Specifically, to fully exploit identity-discriminative features within each modality, we construct separate pedestrian identity classifiers, denoted as $W^v$ and $W^r$ , for visible and infrared images, respectively. Each classifier is trained under supervision from identities within its respective modality, enabling the feature encoders to learn identity-discriminative representations. As shown in Figure 2, both feature encoders, $E^v$ and $E^r$ , adopt ResNet-50 [11] as the backbone. The first layers of each encoder serve as a modality-specific encoder with unshared parameters, while the remaining convolutional layers are shared to enhance cross-modal feature consistency. + +Given the training samples $x_{i}^{v}$ and $x_{i}^{r}$ , we pass them through the encoders $E^{v}$ and $E^{r}$ to obtain the feature representations $f_{i}^{v}$ and $f_{i}^{r}$ . To enhance their discriminative capability, we employ a combination of cross-entropy loss and weighted regularization triplet loss [48] to update $E^{v}$ and $E^{r}$ . The cross-entropy loss is defined as: + +$$ +\mathcal {L} _ {i d} ^ {e x p} = - \sum_ {t \in \{v, r \}} \frac {1}{n ^ {t}} \sum_ {i = 1} ^ {n ^ {t}} \boldsymbol {y} _ {i} ^ {t} \log \boldsymbol {p} _ {i} ^ {t}, \tag {1} +$$ + +where $n^t$ denotes the number of samples of modality $t$ in a mini-batch, and $\pmb{p}_i^t = \pmb{W}^t(\pmb{f}_i^t) \in \mathbb{R}^{C_t \times 1}$ represents the identity prediction vector produced by the classifier $\pmb{W}^t$ (for $t = v, r$ ) given the features $\pmb{f}_i^v$ and $\pmb{f}_i^r$ . + +The weighted regularization triplet loss is defined as: + +$$ +\mathcal {L} _ {w r t} ^ {i n t r} = \sum_ {t \in \{v, r \}} \frac {1}{n ^ {t}} \sum_ {i = 1} ^ {n ^ {t}} \log \left(1 + \exp \left(\sum_ {j} w _ {i j} ^ {p} d _ {i j} ^ {p} - \sum_ {k} w _ {i k} ^ {n} d _ {i k} ^ {n}\right)\right) \tag {2} +$$ + +where + +$$ +w _ {i j} ^ {p} = \frac {\exp \left(d _ {i j} ^ {p}\right)}{\sum_ {d _ {i j} \in \mathcal {P} _ {i}} \exp \left(d _ {i j} ^ {p}\right)}, \quad w _ {i k} ^ {n} = \frac {\exp \left(- d _ {i k} ^ {n}\right)}{\sum_ {d _ {i k} \in \mathcal {N} _ {i}} \exp \left(- d _ {i k} ^ {n}\right)}, \tag {3} +$$ + +with $\mathcal{P}_i$ and $\mathcal{N}_i$ denoting the sets of positive and negative samples, respectively, for the anchor sample $x_i^t$ within the same modality in a mini-batch. Here, $j$ and $k$ index the positive and negative samples, while $d_{ij}^p = \| f_i^t -f_j^t\| _2$ and $d_{ik}^{n} = \| f_{i}^{t} - f_{k}^{t}\|_{2}$ represent the Euclidean distances between positive and negative pairs, respectively. The classifiers $W^{v}$ and $W^{r}$ , trained by minimizing the loss functions in Eqs. (1) and (2), serve as experts in identity discrimination for their respective modalities. Since $W^{v}$ and $W^{r}$ are trained on distinct modalities and focus on different + +identity-related information in pedestrian samples, we refer to them as heterogeneous experts. + +# 3.3. Establishment of Cross-Modal Relationships + +Although the encoders trained with the constraints in Eq.(1) and Eq.(2) extract features with certain discriminative power, the absence of cross-modality sample pairs poses challenges for $W^{v}$ and $W^{r}$ in cross-identification. To address this issue, it is essential to establish correspondences between modalities. To this end, we propose a cross-modality identity relationship establishment mechanism. This mechanism mitigates the limitations of individual experts by integrating their predictions to construct reliable cross-modality identity correspondences. Specifically, after training the heterogeneous experts using the loss functions in Eq.(1) and Eq.(2), these experts are employed to predict the identities of all samples from the other modality. In the classification vector, the score at each position represents the probability that a given sample belongs to the corresponding identity in the other modality. + +Similar to [39], we employ the Count Priority Selection method to obtain the expert decision matrix $M^{t\rightarrow \bar{t}}$ where $t,\bar{t}\in \{v,r\} ,t\neq \bar{t}$ . The notation $t\to \bar{t}$ denotes the prediction results of $\bar{t}$ -modality experts on $t$ -modality samples, and $M^{t\rightarrow \bar{t}}\in \mathbb{R}^{C^t\times C^\bar{t}}$ encodes the identity correspondences between the two modalities. For example, the element at the $i$ -th row and $j$ -th column of $M^{v\rightarrow r}$ represents the correspondence between the $i$ -th visible pedestrian identity and the $j$ -th infrared pedestrian identity. If the two identities correspond, the respective element in $M^{v\rightarrow r}$ is set to 1; otherwise, it is 0. While relying solely on identical predictions across experts as pseudo-labels ensures high reliability, it may overlook many potential identity correspondences. To mitigate this issue, we classify all cross-modal identity correspondences into three categories based on the consistency of expert predictions: consistent matches, unique matches, and conflicting matches. + +Based on the definition of consistent matching, the consistent identity relationship matrix $M_{c}$ is obtained as follows: + +$$ +\boldsymbol {M} _ {c} = \boldsymbol {M} ^ {t \rightarrow \bar {t}} \odot (\boldsymbol {M} ^ {\bar {t} \rightarrow t}) ^ {\mathrm {T}} \tag {4} +$$ + +where $\odot$ denotes the Hadamard product. When $M_{c}(i,j) = 1$ , it indicates that the experts unanimously agree that the $i$ -th identity in the $t$ modality corresponds to the $j$ -th identity in the $\bar{t}$ modality. The single identity relationship matrix for one expert is defined as: + +$$ +\begin{array}{l} \boldsymbol {M} _ {s} ^ {t \rightarrow \bar {t}} = \boldsymbol {A} ^ {\mathrm {T}} \odot \boldsymbol {M} ^ {t \rightarrow \bar {t}}, \quad \boldsymbol {A} (i, j) = r _ {i} \cdot c _ {j}, \\ r _ {i} = \mathbf {1} _ {\{0 \}} \left(\sum_ {j = 1} ^ {C ^ {t}} M ^ {\bar {t} \rightarrow t} (i, j)\right), \tag {5} \\ c _ {j} = \mathbf {1} _ {\{0 \}} (\sum_ {i = 1} ^ {C ^ {\bar {t}}} M ^ {\bar {t} \rightarrow t} (i, j)), \\ \end{array} +$$ + +where $\mathbf{1}_{\{0\}}(\cdot)$ is an indicator function that returns 1 if the input equals 0 and 0 otherwise. Let $M_{s} = M_{s}^{v\rightarrow r} + (M_{s}^{r\rightarrow v})^{\mathrm{T}}$ . The contradictory identity relationship matrix is then given by: + +$$ +M _ {w} = M ^ {v \rightarrow r} + \left(M ^ {r \rightarrow v}\right) ^ {\mathrm {T}} - 2 M _ {c} - M _ {s} \tag {6} +$$ + +The matrices $M_{c}$ , $M_{s}$ , and $M_{w}$ effectively represent the cross-modal identity correspondences. Consequently, they are used to guide $E^{v}$ and $E^{r}$ to extract consistent features for cross-modal samples of the same identity. + +# 3.4. Collaborative Consistency Learning + +Cross-modal Consistency Learning. Due to the challenge of unknown pedestrian identities across modal samples, it is not feasible to utilize identity relationships between cross-modal samples to guide the training of encoders, limiting the model's cross-modal matching performance. To address this issue, we propose a cross-modal consistency learning method based on the identity relationships established by $M_{c}$ , $M_{s}$ , and $M_{w}$ among samples across modalities. In this method, for samples with established identity relationships in $M_{c}$ , $M_{s}$ , and $M_{w}$ , we apply a cross-modal cross-entropy loss as a constraint: + +$$ +\mathcal {L} _ {i d} ^ {s t r o} = - \frac {1}{n ^ {s}} \sum_ {i = 1} ^ {n ^ {s}} \hat {\boldsymbol {q}} _ {i} ^ {v} \log \left(\boldsymbol {W} ^ {c} \left(\boldsymbol {f} _ {i} ^ {v}\right)\right) - \frac {1}{n ^ {r}} \sum_ {i = 1} ^ {n ^ {r}} \boldsymbol {q} _ {i} ^ {r} \log \left(\boldsymbol {W} ^ {c} \left(\boldsymbol {f} _ {i} ^ {r}\right)\right) \tag {7} +$$ + +where $n^r$ denotes the number of infrared samples in a minibatch, and $n^s$ denotes the number of visible samples in the mini-batch that have corresponding identities with the infrared modality samples in the training set. $\hat{\pmb{q}}_i^v$ represents the one-hot vector of the infrared modality identity corresponding to the $i$ -th visible sample in $M_c$ or $M_s$ -i.e., the pseudo-label for the visible sample. $\pmb{W}^{c}(\pmb{f}_{i}^{t})$ represents the classification result of $\pmb{f}_i^t$ by the shared classifier $\pmb{W}^c$ . + +For samples in $M_w$ with contradictory identity correspondences, we employ a relaxed identity discrimination constraint during model training to mitigate the impact of cross-modal label noise: + +$$ +\mathcal {L} _ {i d} ^ {\text {w e a k}} = - \frac {1}{n _ {w} ^ {v}} \sum_ {i = 1} ^ {n _ {w} ^ {v}} \boldsymbol {m} _ {i} \log \left(1 - \boldsymbol {W} ^ {c} \left(\boldsymbol {f} _ {i} ^ {v}\right) + \epsilon\right) \tag {8} +$$ + +$$ +m _ {i l} = \left\{ \begin{array}{l l} 1, & \text {i f} p _ {i l} ^ {v} < \min _ {k \in K} \left(p _ {i k} ^ {v}\right), \\ 0, & \text {o t h e r w i s e} \end{array} \right. \tag {9} +$$ + +where $K = \{k \mid M_w(j, k) = 1\}$ , $j$ is the intra-modal identity corresponding to sample $x_i^v$ , and $\epsilon = 10^{-10}$ is used to prevent numerical overflow. $p_{il}^v$ represents the probability score at the $l$ -th position in the classification result $W^c(\pmb{f}_i^v)$ for sample $x_i^v$ . $m_{il}$ is the $l$ -th element of the vector $m_i$ . This loss function effectively leverages samples with uncertain cross-modal identities for model training. Additionally, during model training, we adopt a cross-modal weighted + +regularization triplet loss, $\mathcal{L}_{wrt}^{cros}$ . Unlike $\mathcal{P}_i$ in $\mathcal{L}_{wrt}^{intr}$ , $\mathcal{P}_i$ in $\mathcal{L}_{wrt}^{cros}$ represents the set of positive samples corresponding to the $i$ -th anchor sample $x_i^t$ in a mini-batch, which includes both intra-modal positive samples and cross-modal positive samples obtained from $M_c$ . + +Collaborative Learning Among Experts. Since experts within each modality are trained on uni-modal samples, their ability to discriminate cross-modal identities is limited. Consequently, when heterogeneous experts predict pedestrian identities across modalities, they may produce inconsistent results for challenging samples, reducing the accuracy of $M_{c}$ , $M_{s}$ , and $M_{w}$ . To address this issue, we propose a collaborative consistency learning method among heterogeneous experts. Specifically, we guide the experts to generate consistent predictions for cross-modal positive samples, introducing a degree of homogeneity to enhance cross-modal consistency. However, within a mini-batch, cross-modal positive samples of the same identity may be absent. Even when present, they may not fully capture the diversity of pedestrian appearances. To mitigate this, we construct prototypes $\mathcal{P}^t = [\mathcal{P}_1^t, \mathcal{P}_2^t, \dots, \mathcal{P}_{C^t}^t]^\top \in \mathbb{R}^{C^t \times d}$ for pedestrian identities in each modality, where $d$ denotes the feature dimension. During training, we initialize the prototypes using the mean feature of each identity and update them as follows: + +$$ +\mathcal {P} _ {i} ^ {t} \leftarrow \lambda \mathcal {P} _ {i} ^ {t} + (1 - \lambda) \bar {\boldsymbol {f}} _ {i} ^ {t} \tag {10} +$$ + +where $\mathcal{P}_i^t$ represents the prototype feature of the $i$ -th identity in modality $t$ , $\lambda \in [0,1]$ is the momentum update coefficient, and $\bar{f}_i^t$ denotes the mean feature of the samples associated with identity $i$ within a mini-batch. + +Assume that $\pmb{p}_i^{v\rightarrow v} = \pmb{W}^v (\pmb{f}_i^v)$ represents the classification result of $W^v$ for $\pmb{f}_i^v$ , and $\pmb{p}_i^{r\rightarrow v} = \pmb{W}^v (\mathcal{P}_{\hat{y}_i^r}^r)$ represents the classification result of $W^v$ for $\mathcal{P}_{\hat{y}_i^r}^r$ , where $\hat{y}_i^r$ represents the identity of $x_i^v$ in $M_c$ . To promote consistency in the decision-making process of experts across different modalities, we propose a collaborative consistency loss for $W^v$ : + +$$ +\mathcal {L} _ {h o m o} ^ {v} = \frac {1}{n ^ {c} \times C ^ {v}} \sum_ {i = 1} ^ {n ^ {c}} \| \boldsymbol {p} _ {i} ^ {v \rightarrow v} - \boldsymbol {p} _ {i} ^ {r \rightarrow v} \| _ {2} ^ {2} \tag {11} +$$ + +where $n^c$ denotes the number of visible samples in a minibatch that have identity relationships in $M_c$ . Similarly, we define the collaborative consistency loss $\mathcal{L}_{homo}^r$ for $W^r$ . It is important to note that the cross-modal identity relationships predicted by experts may not always be accurate, and directly enforcing a closer match between the two modalities could negatively impact model training. + +Typically, when an expert's prediction confidence is high, the probability assigned to the predicted identity is large, whereas when the prediction is uncertain, this probability tends to be smaller. Based on this observation, we use a weight derived from the information entropy $H$ of the prediction results to dynamically adjust the constraint strength + +in the expert homogeneity loss. The more reliable an expert's prediction, the weaker the constraining effect of the loss. Specifically, the total collaborative consistency loss is formulated as follows: + +$$ +\mathcal {L} _ {h o m o} = w _ {v} \mathcal {L} _ {h o m o} ^ {v} + w _ {r} \mathcal {L} _ {h o m o} ^ {r} \tag {12} +$$ + +where + +$$ +w _ {v} = \frac {H \left(\boldsymbol {p} ^ {r \rightarrow v}\right)}{H \left(\boldsymbol {p} ^ {r \rightarrow v}\right) + H \left(\boldsymbol {p} ^ {v \rightarrow r}\right)}, w _ {r} = \frac {H \left(\boldsymbol {p} ^ {v \rightarrow r}\right)}{H \left(\boldsymbol {p} ^ {r \rightarrow v}\right) + H \left(\boldsymbol {p} ^ {v \rightarrow r}\right)} \tag {13} +$$ + +Additionally, to prevent experts from forgetting their original knowledge, we continue training them within their respective modalities by minimizing the loss in Eq. (1). + +# 3.5. Model Optimization + +In the expert construction phase, the overall optimization objective is formulated as: + +$$ +\mathcal {L} _ {\text {p h a s e 1}} = \mathcal {L} _ {i d} ^ {\text {e x p}} + \lambda_ {1} \mathcal {L} _ {\text {w r t}} ^ {\text {i n t r}} \tag {14} +$$ + +where $\lambda_{1}$ is a hyperparameter used to balance the model. + +In collaborative consistency learning phase, the overall optimization objective is: + +$$ +\mathcal {L} _ {\text {p h a s e 2}} = \mathcal {L} _ {i d} ^ {e x p} + \mathcal {L} _ {i d} ^ {\text {s t r o}} + \mathcal {L} _ {\text {h o m o}} + \lambda_ {1} \mathcal {L} _ {\text {w r t}} ^ {\text {c r o s}} + \lambda_ {2} \mathcal {L} _ {i d} ^ {\text {w e a k}} \tag {15} +$$ + +where $\mathcal{L}_{wrt}^{cros}$ is the weighted regularization triplet loss between cross-modal samples, as defined in Eq. (2). During collaborative consistency learning, cross-modal consistency learning and expert collaborative learning reinforce each other, enabling the network to extract cross-modal consistent features and effectively addressing the absence of cross-modal sample labels. Simultaneously, heterogeneous experts provide more reliable identity predictions during the homogenization process, significantly improving the accuracy of $M_{c}$ , $M_{s}$ , and $M_{w}$ . + +# 4. Experiments + +# 4.1. Datasets and Evaluation Protocols + +To evaluate the proposed method, we assess its performance on two public VirID datasets: SYSU-MM01 [35] and LLCM [54]. + +SYSU-MM01, the first large-scale VIREID dataset, contains images captured by four visible cameras and two near-infrared cameras across various indoor and outdoor scenes. It includes 491 pedestrian identities, with a total of 287,628 visible images and 15,792 infrared images. The training set consists of 22,258 visible images and 11,909 infrared images from 395 identities, while the test set covers 96 identities. We evaluate performance under both all-search and indoor-search modes. + +LLCM, one of the most challenging large-scale VIReID datasets, comprises images captured by nine visible cameras and nine infrared cameras deployed in low-light environments. It contains 46,767 images from 1,064 identities, + +including 16,946 visible and 13,975 infrared images. The training set includes samples from 713 identities, while the test set consists of 8,680 visible and 7,166 infrared images from 351 identities. We assess performance in both visible-to-infrared and infrared-to-visible retrieval modes. + +Evaluation Protocols. Following standard practice, we use Cumulative Matching Characteristic (CMC) [45], mean Average Precision (mAP), and mean Inverse Negative Penalty (mINP) [48] for evaluation. + +# 4.2. Implementation Details + +The proposed method is implemented in PyTorch, with experiments conducted on a single RTX 4090 GPU. ResNet-50 serves as the encoder, and all input images are resized to $288 \times 144$ . During training, data augmentation techniques, including style augmentation [57], random cropping, horizontal flipping, and random erasing, are applied. After completing heterogeneous expert learning, collaborative consistency learning is performed for 120 epochs. The initial learning rate is set to $3 \times 10^{-4}$ for the encoder and $6 \times 10^{-4}$ for the experts and shared classifier. A warmup strategy is applied for the first 10 epochs to gradually increase the learning rate, which is then decayed to $10\%$ at the 30th and 70th epochs. The hyper-parameter $\lambda$ for prototypes is set to 0.8, while $\lambda_{1}$ and $\lambda_{2}$ are both set to 0.25. + +# 4.3. Comparison with State-of-the-Art Methods + +To validate the effectiveness of our method, we compare it with state-of-the-art approaches in supervised VIREID (SVIREID), semi-supervised VIREID (SSVIREID), and unsupervised VIREID (USVIREID). The experimental results on the SYSU-MM01 and LLCM datasets are presented in Tables 1 and 2, with the best performance highlighted in bold and the second-best underlined. + +Comparison with Supervised Methods. In a weakly supervised setting with missing cross-modal annotations, our method achieves performance on the SYSU-MM01 and LLCM datasets that is comparable to or even surpasses some SVIREID methods. This demonstrates its ability to effectively handle missing cross-modal annotations and learn cross-modal invariant features. + +Comparison with Semi-Supervised Methods. As shown in Tables 1 and 2, our method significantly outperforms existing SSVIREID approaches. As previously noted, pseudolabel noise caused by missing infrared annotations hinders the performance of semi-supervised methods, especially on the challenging SYSU-MM01 and LLCM datasets. Specifically, on SYSU-MM01 (all search), our method improves Rank-1 accuracy and mAP by $12.0\%$ and $11.0\%$ , respectively, compared to DPIS. It also significantly outperforms MAUM-50 and MAUM-100, which are trained with 50 and 100 infrared identities, respectively. On LLCM (VIS to IR), our method achieves improvements of $11.1\%$ and $10.5\%$ in + +Table 1. Comparison of CMC (%) , mAP (%) and mINP (%) performances with the state-of-the-art methods on SYSU-MM01 dataset. + +
SYSU-MM01 SettingsAll SearchIndoor Search
TypeMethodsVenueRank-1Rank-10Rank-20mAPmINPRank-1Rank-10Rank-20mAPmINP
SVIRelIDAGW[48]TPAMI'2147.584.492.147.735.354.291.195.663.059.2
LbA[27]ICCV'2155.4--54.1-58.5--66.3-
CAJ[47]ICCV'2169.995.798.566.953.676.397.999.580.476.8
DEEN[54]CVPR'2374.797.699.271.8-80.399.099.883.3-
MUN [51]ICCV'2376.297.8-73.8-79.498.1-82.1-
CSDN [52]TMM'2576.797.299.173.0-84.599.299.686.8-
SSVIRelIDMAUM-50 [22]CVPR'2228.8--36.1------
MAUM-100 [22]CVPR'2238.5--39.2------
OTLA [33]ECCV'2248.2--43.9-47.4--56.8-
DPIS [30]ICCV'2358.4--55.6-63.0--70.0-
USVIRelIDADCA [39]MM'2245.585.393.242.728.350.689.796.259.155.2
CCLNet [3]MM'2354.088.895.050.2-56.791.197.265.1-
GUR [40]ICCV'2363.5--61.647.971.1--76.272.6
MMM [31]ECCV'2461.693.398.057.9-64.495.098.270.4-
DLM [49]TPAMI'2562.293.497.458.443.767.395.999.272.768.9
PCAL [43]TIFS'2557.991.096.452.936.960.193.397.366.762.1
WSVIRelIDOurs-70.495.898.866.652.776.597.599.380.276.6
+ +Table 2. Comparison of Rank-1 $(\%)$ and mAP $(\%)$ performances with the state-of-the-art methods on LLCM dataset, where \* denotes the results obtained using publicly available code for the experiment. + +
LLCM SettingsVIS to IRIR to VIS
TypeMethodsVenueRank-1mAPRank-1mAP
SVIRelIDAGW [48]TPAMI'2151.555.343.651.8
LbA[27]ICCV'2150.855.643.853.1
CAJ [47]ICCV'2156.559.848.856.6
DEEN [54]CVPR'2362.565.854.962.9
CSDN [52]TMM'2563.766.555.863.5
SSVIRelIDOTLA [33]ECCV'2244.248.236.242.2
USVIRelIDADCA*[39]MM'2242.546.938.444.4
PGM*[37]CVPR'2344.949.039.445.3
WSVIRelIDOurs-55.358.747.353.3
+ +Rank-1 accuracy and mAP, respectively, over OTLA. These results validate its effectiveness in addressing the challenge of missing cross-modal labels. + +Comparison with Unsupervised Methods. Compared to the USVIReID methods, our approach achieves substantial performance gains with minimal annotation costs. Specifically, on SYSU-MM01 (all search), it surpasses GUR by $6.9\%$ and $5.0\%$ in Rank-1 accuracy and mAP, respectively. On LLCM (VIS to IR), it outperforms PGM by $10.4\%$ and $9.7\%$ in Rank-1 accuracy and mAP, respectively. These improvements stem from the collaborative promotion mechanism, which effectively establishes relationships between cross-modal samples and enhances supervision. + +# 4.4. Ablation Study + +The proposed method consists of three main components: HEL, CRE, and CCL. The model trained solely with HEL serves as the Baseline. HEL enables the model to acquire fundamental cross-modal recognition capabilities, providing a foundation for constructing cross-modal correspondences based on prediction results. CRE is designed to enhance cross-modal correspondences for CMCL and must be + +evaluated in conjunction with CMCL. To assess its impact, we compare consistency prediction using heterogeneous experts as cross-modal correspondences with CRE+CMCL results. CCL comprises two parts: CMCL and CLAE. In the ablation study, CMCL without CRE prediction results is denoted as B+CMCL\CRE, while incorporating CRE is denoted as B+CRE+CMCL. The full model is referred to as B+CRE+CMCL+CLAE. Table 3 presents the experimental results for different components. + +Effectiveness of HEL. HEL is a core component of the Baseline model. On the SYSU-MM01 dataset, the Baseline achieves Rank-1 and mAP scores of $47.8\%$ and $47.2\%$ , respectively, in all-search mode, and $60.8\%$ and $67.7\%$ , respectively, in indoor-search mode. These results indicate that a model trained solely within modalities already possesses some cross-modal retrieval capability, providing a strong foundation for further optimization. However, intramodal training is limited to optimizing within each modality and does not establish correspondences between cross-modal samples. Therefore, relying solely on intra-modal training is insufficient to address the challenges posed by the absence of inter-modal correspondences in VI-ReID. + +Effectiveness of CMCL. Using only expert-provided consistency predictions for cross-modal consistency learning, B+CMCL\CRE improves Rank-1 and mAP by $18.9\%$ and $15.6\%$ , respectively, over the Baseline in the all-search setting. In the indoor-search setting, Rank-1 and mAP improve by $12.6\%$ and $9.9\%$ , respectively. These results demonstrate that CMCL effectively reduces modality differences and enhances cross-modal retrieval. By leveraging CMCL, the model learns more consistent feature representations across modalities, leading to improved retrieval performance. + +Effectiveness of CRE. By introducing the cross-modal relation establishment strategy, B+CRE+CMCL improves Rank-1 and mAP by $1.3\%$ and $1.8\%$ , respectively, over B+CMCL\CRE in the all-search setting. In the indoor + +Table 3. Ablation study on SYSU-MM01. "B" denotes the Baseline method. + +
SettingsAll SearchIndoor Search
Rank-1mAPRank-1mAP
B47.847.260.867.7
B+CMCL\CRE66.762.873.477.6
B+CRE+CMCL68.064.674.078.3
B+CRE+CMCL+CLAE70.466.676.580.2
+ +search setting, Rank-1 and mAP improve by $0.6\%$ and $0.7\%$ , respectively. This improvement arises because the relation establishment strategy not only integrates consistent predictions from experts but also leverages inconsistent ones. Consequently, it facilitates more comprehensive cross-modal identity correspondence and effectively mitigates relation-matching errors caused by the limitations of individual experts. + +Table 4. Further analysis of CRE effectiveness on SYSU-MM01. + +
SettingsAll SearchIndoor Search
Rank-1mAPRank-1mAP
B+CMCL(Mr→v)61.057.569.875.1
B+CMCL(Mv→r)67.764.473.777.8
B+CMCL\CRE66.762.873.477.6
B+CRE+CMCL68.064.674.078.3
+ +Effectiveness of CLAE. After introducing CLAE, $\mathrm{B + CRE + CMCL + }$ CLAE improves Rank-1 and mAP by $2.4\%$ and $2.0\%$ , respectively, over $\mathrm{B + CRE + CMCL}$ in the all-search setting. In the indoor-search setting, Rank-1 and mAP improve by $2.5\%$ and $1.9\%$ , respectively. These results indicate that CLAE further enhances identity prediction accuracy for cross-modal samples. By leveraging CLAE, the two heterogeneous experts progressively reduce modality discrepancies, enhance collaboration, and generate more consistent cross-modal predictions. Additional experimental results are in the supplementary material. + +# 4.5. Further Analysis of CRE + +To further analyze the role of the proposed CRE, we utilize the cross-modal relationships established by $M^{r \to v}$ and $M^{v \to r}$ in Eq. (6) to guide model training. The models trained with these relationships are denoted as B+CMCL( $M^{r \to v}$ ) and B+CMCL( $M^{v \to r}$ ), respectively. As shown in Table 4, B+CMCL( $M^{r \to v}$ ) exhibits significantly lower recognition performance compared to B+CMCL( $M^{v \to r}$ ). This is mainly because infrared images lack discriminative color information, limiting the ability of visible expert to recognize infrared images, whereas infrared expert do not face this issue and thus perform better. Since the recognition abilities of different experts are complementary, the B+CMCL model trained solely on consistent predictions among experts fails to fully exploit this complementarity. The CRE method effectively addresses + +this issue, and with its assistance, the B+CRE+CMCL model achieves the best recognition performance, further validating the effectiveness of CRE. + +![](images/8605a61247c90b1458ae7a4765376763363716d8e2e6da9e75979e73f6267949.jpg) + +![](images/163d80aaff586a0ff4f309e6e18dbf318dd757ebca8a781d438c94df15ff337b.jpg) + +![](images/a78bb0218df61d269b12dd3894f0d81381fb4a11bff1cef86f25609f5552df8b.jpg) +(a) Baseline + +![](images/1082872120da8f4720e7e51b7d92ba1d48415eace1830c0e2d9271ca0e194e6f.jpg) +(b) Baseline+CRE+CMCL+CLAE +Figure 3. Visualization comparison of feature distributions between baseline and our method on SYSU-MM01: intra- and interclass distances of cross-modal sample features in the first row, and t-SNE mapped 2D feature space visualization in the second row, with 10 randomly selected identities and 10 samples each. + +# 4.6. Visualization Analysis + +We further validate our method by visualizing feature distributions. Figure 3 (top) compares the intra-class and inter-class distances of features from cross-modal samples between the baseline and our method. Our method significantly increases the gap between intra-class and inter-class distances, effectively reducing the distance between same-identity cross-modal samples and minimizing modality discrepancies. Figure 3 (bottom) shows the t-SNE visualization of features in 2D, with circular points for visible images, triangular points for infrared images, and colors representing pedestrian identities. Compared to the baseline, our method clusters same-identity features more effectively while maintaining separation between different identities, demonstrating its ability to mitigate modality discrepancies. + +# 5. Conclusion + +This paper proposes a weakly supervised cross-modal reID method that establishes identity correspondences without labeled cross-modal identity pairs. By leveraging a heterogeneous expert collaborative consistency learning framework, our method effectively integrates predictions from modality-specific classification experts through a cross-modal relationship fusion mechanism. Experimental results on two challenging datasets validate its effectiveness in enhancing cross-modal identity recognition and extracting modality-invariant representations. This work provides a promising direction for reducing reliance on fully labeled cross-modal datasets, with potential for further improvements in expert collaboration and adaptive learning. + +# Acknowledgments + +# References + +[1] Cuiqun Chen, Mang Ye, Meibin Qi, Jingjing Wu, Jianguo Jiang, and Chia-Wen Lin. Structure-aware positional transformer for visible-infrared person re-identification. IEEE Transactions on Image Processing, 31:2352-2364, 2022. 2 +[2] Kunfeng Chen, Zhisong Pan, Jiabao Wang, Shanshan Jiao, Zhicheng Zeng, and Zhuang Miao. Joint feature learning network for visible-infrared person re-identification. In Chinese Conference on Pattern Recognition and Computer Vision (PRCV), pages 652–663. Springer, 2020. 2 +[3] Zhong Chen, Zhizhong Zhang, Xin Tan, Yanyun Qu, and Yuan Xie. Unveiling the power of clip in unsupervised visible-infrared person re-identification. In Proceedings of the 31st ACM International Conference on Multimedia (ACMMM), pages 3667–3675, 2023. 3, 7 +[4] De Cheng, Lingfeng He, Nannan Wang, Shizhou Zhang, Zhen Wang, and Xinbo Gao. Efficient bilateral cross-modality cluster matching for unsupervised visible-infrared person reid. In Proceedings of the 31st ACM International Conference on Multimedia (ACMMM), pages 1325-1333, 2023. 1, 3 +[5] Seokeon Choi, Sumin Lee, Youngeun Kim, Taekyung Kim, and Changick Kim. Hi-cmd: Hierarchical cross-modality disentanglement for visible-infrared person re-identification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10257–10266, 2020. 2 +[6] Pingyang Dai, Rongrong Ji, Haibin Wang, Qiong Wu, and Yuyu Huang. Cross-modality person re-identification with generative adversarial training. In Proceedings of the Twenty-Seventh International Joint Conference on Artificial Intelligence (IJCAI), pages 677–683, 2018. 2 +[7] Jiawei Feng, Ancong Wu, and Wei-Shi Zheng. Shape-erased feature learning for visible-infrared person re-identification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 22752–22761, 2023. 2 +[8] Zhanxiang Feng, Jianhuang Lai, and Xiaohua Xie. Learning modality-specific representations for visible-infrared person re-identification. IEEE Transactions on Image Processing, 29:579–590, 2020. 2 +[9] Chaoyou Fu, Yibo Hu, Xiang Wu, Hailin Shi, Tao Mei, and Ran He. Cm-nas: Cross-modality neural architecture search for visible-infrared person re-identification. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 11823–11832, 2021. 2 +[10] Yi Hao, Nannan Wang, Jie Li, and Xinbo Gao. Hsme: Hypersphere manifold embedding for visible thermal person re-identification. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), pages 8385–8392, 2019. 1, 2 + +This work was supported in part by the National Natural Science Foundation of China (62276120, 62161015, 61966021), and the Yunnan Fundamental Research Projects (202401AS070106, 202301AV070004, 202501AS070123). +[11] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 770-778, 2016. 4 +[12] Lingfeng He, De Cheng, Nannan Wang, and Xinbo Gao. Exploring homogeneous and heterogeneous consistent label associations for unsupervised visible-infrared person reid. International Journal of Computer Vision, 133:3129-3148, 2025. 3 +[13] Weipeng Hu, Bohong Liu, Haitang Zeng, Yanke Hou, and Haifeng Hu. Adversarial decoupling and modality-invariant representation learning for visible-infrared person re-identification. IEEE Transactions on Circuits and Systems for Video Technology, 32(8):5095-5109, 2022. 2 +[14] Nianchang Huang, Jianan Liu, Yunqi Miao, Qiang Zhang, and Jungong Han. Deep learning for visible-infrared cross-modality person re-identification: A comprehensive review. Information Fusion, 91:396–411, 2023. 2 +[15] Yan Huang, Qiang Wu, Jingsong Xu, Yi Zhong, Peng Zhang, and Zhaoxiang Zhang. Alleviating modality bias training for infrared-visible person re-identification. IEEE Transactions on Multimedia, 24:1570–1582, 2022. 2 +[16] Kajal Kansal, A. V. Subramanyam, Zheng Wang, and Shin'ichi Satoh. Sdl: Spectrum-disentangled representation learning for visible-infrared person re-identification. IEEE Transactions on Circuits and Systems for Video Technology, 30(10):3422-3432, 2020. 2 +[17] Diangang Li, Xing Wei, Xiaopeng Hong, and Yihong Gong. Infrared-visible cross-modal person re-identification with an $\mathbf{x}$ modality. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), pages 4610-4617, 2020. 2 +[18] Shuang Li, Jiaxu Leng, Ji Gan, Mengjingcheng Mo, and Xinbo Gao. Shape-centered representation learning for visible-infrared person re-identification. Pattern Recognition, 167:111756, 2025. 2 +[19] Zhiyong Li, Haojie Liu, Xiantao Peng, and Wei Jiang. Inter-intra modality knowledge learning and clustering noise alleviation for unsupervised visible-infrared person re-identification. IEEE Transactions on Knowledge and Data Engineering, 36(8):3934-3947, 2024. 3 +[20] Wenqi Liang, Guangcong Wang, Jianhuang Lai, and Xiaohua Xie. Homogeneous-to-heterogeneous: Unsupervised learning for rgb-infrared person re-identification. IEEE Transactions on Image Processing, 30:6392-6407, 2021. 3 +[21] Haijun Liu, Xiaoheng Tan, and Xichuan Zhou. Parameter sharing exploration and hetero-center triplet loss for visible-thermal person re-identification. IEEE Transactions on Multimedia, 23:4414-4425, 2021. 2 +[22] Jialun Liu, Yifan Sun, Feng Zhu, Hongbin Pei, Yi Yang, and Wenhui Li. Learning memory-augmented unidirectional metrics for cross-modality person re-identification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19366-19375, 2022. 7 +[23] Min Liu, Yeqing Sun, Xueping Wang, Yuan Bian, Zhu Zhang, and Yaonan Wang. Pose-guided modality-invariant feature alignment for visible-infrared object re + +identification. IEEE Transactions on Instrumentation and Measurement, 73:1-10, 2024. 2 +[24] Zefeng Lu, Ronghao Lin, and Haifeng Hu. Disentangling modality and posture factors: Memory-attention and orthogonal decomposition for visible-infrared person re-identification. IEEE Transactions on Neural Networks and Learning Systems, 36(3):5494–5508, 2025. 2 +[25] Zhiqi Pang, Chunyu Wang, Honghu Pan, Lingling Zhao, Junjie Wang, and Maozu Guo. Mimr: Modality-invariance modeling and refinement for unsupervised visible-infrared person re-identification. Knowledge-Based Systems, 285: 111350, 2024. 3 +[26] Zhiqi Pang, Chunyu Wang, Lingling Zhao, Yang Liu, and Gaurav Sharma. Cross-modality hierarchical clustering and refinement for unsupervised visible-infrared person re-identification. IEEE Transactions on Circuits and Systems for Video Technology, 34(4):2706-2718, 2024. 3 +[27] Hyunjong Park, Sanghoon Lee, Junghyup Lee, and Bumsub Ham. Learning by aligning: Visible-infrared person re-identification using cross-modal correspondences. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 12046–12055, 2021. 1, 2, 7 +[28] Nan Pu, Wei Chen, Yu Liu, Erwin M. Bakker, and Michael S. Lew. Dual gaussian-based variational subspace disentanglement for visible-infrared person re-identification. In Proceedings of the 28th ACM International Conference on Multimedia (ACMMM), pages 2149–2158, 2020. 1 +[29] Liuxiang Qiu, Si Chen, Yan Yan, Jing-Hao Xue, Da-Han Wang, and Shunzhi Zhu. High-order structure based middle-feature learning for visible-infrared person re-identification. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), pages 4596–4604, 2024. 2 +[30] Jiangming Shi, Yachao Zhang, Xiangbo Yin, Yuan Xie, Zhizhong Zhang, Jianping Fan, Zhongchao Shi, and Yanyun Qu. Dual pseudo-labels interactive self-training for semi-supervised visible-infrared person re-identification. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 11218-11228, 2023. 1, 3, 7 +[31] Jiangming Shi, Xiangbo Yin, Yeyun Chen, Yachao Zhang, Zhizhong Zhang, Yuan Xie, and Yanyun Qu. Multimemory matching for unsupervised visible-infrared person re-identification. In European Conference on Computer Vision (ECCV), pages 456–474. Springer, 2024. 3, 7 +[32] Xiao Teng, Xingyu Shen, Kele Xu, and Long Lan. Enhancing unsupervised visible-infrared person re-identification with bidirectional-consistency gradual matching. In Proceedings of the 32nd ACM International Conference on Multimedia (ACMMM), pages 9856–9865, 2024. 3 +[33] Jiangming Wang, Zhizhong Zhang, Mingang Chen, Yi Zhang, Cong Wang, Bin Sheng, Yanyun Qu, and Yuan Xie. Optimal transport for label-efficient visible-infrared person re-identification. In European Conference on Computer Vision (ECCV), pages 93–109. Springer, 2022. 1, 3, 7 +[34] Zhixiang Wang, Zheng Wang, Yinqiang Zheng, Yung-Yu Chuang, and Shin'ichi Satoh. Learning to reduce dual-level discrepancy for infrared-visible person re-identification. In + +Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 618-626, 2019. 2 +[35] Ancong Wu, Wei-Shi Zheng, Hong-Xing Yu, Shaogang Gong, and Jianhuang Lai. Rgb-infrared cross-modality person re-identification. In Proceedings of the IEEE international Conference on Computer Vision (ICCV), pages 5380-5389, 2017. 2, 6 +[36] Ancong Wu, Chengzhi Lin, and Wei-Shi Zheng. Asymmetric mutual learning for unsupervised transferable visible-infrared re-identification. IEEE Transactions on Circuits and Systems for Video Technology, 34(11):10524–10537, 2024. 1, 3 +[37] Zesen Wu and Mang Ye. Unsupervised visible-infrared person re-identification via progressive graph matching and alternate learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9548–9558, 2023. 3, 7 +[38] Ruida Xi, Nianchang Huang, Changzhou Lai, Qiang Zhang, and Jungong Han. Fmcnet+: Feature-level modality compensation for visible-infrared person re-identification. IEEE Transactions on Neural Networks and Learning Systems, 36 (7):13247-13261, 2025. 2 +[39] Bin Yang, Mang Ye, Jun Chen, and Zesen Wu. Augmented dual-contrastive aggregation learning for unsupervised visible-infrared person re-identification. In Proceedings of the 30th ACM International Conference on Multimedia (ACMMM), pages 2843–2851, 2022. 3, 4, 7 +[40] Bin Yang, Jun Chen, and Mang Ye. Towards grand unified representation learning for unsupervised visible-infrared person re-identification. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 11069–11079, 2023. 3, 7 +[41] Bin Yang, Jun Chen, Cuiqun Chen, and Mang Ye. Dual consistency-constrained learning for unsupervised visible-infrared person re-identification. IEEE Transactions on Information Forensics and Security, 19:1767-1779, 2024. 3 +[42] Bin Yang, Jun Chen, and Mang Ye. Shallow-deep collaborative learning for unsupervised visible-infrared person re-identification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16870–16879, 2024. 3 +[43] Yiming Yang, Weipeng Hu, and Haifeng Hu. Progressive cross-modal association learning for unsupervised visible-infrared person re-identification. IEEE Transactions on Information Forensics and Security, 20:1290–1304, 2025. 1, 3, 7 +[44] Mang Ye, Xiangyuan Lan, Jiawei Li, and Pong Yuen. Hierarchical discriminative learning for visible thermal person re-identification. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), pages 7501-7508, 2018. 2 +[45] Mang Ye, Zheng Wang, Xiangyuan Lan, and Pong C. Yuen. Visible thermal person re-identification via dual-constrained top-ranking. In Proceedings of the Twenty-Seventh International Joint Conference on Artificial Intelligence, (IJCAI), pages 1092–1099, 2018. 2, 6 +[46] Mang Ye, Jianbing Shen, and Ling Shao. Visible-infrared person re-identification via homogeneous augmented tri + +modal learning. IEEE Transactions on Information Forensics and Security, 16:728-739, 2020. 2 +[47] Mang Ye, Weijian Ruan, Bo Du, and Mike Zheng Shou. Channel augmented joint learning for visible-infrared recognition. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 13547-13556, 2021. 2, 7 +[48] Mang Ye, Jianbing Shen, Gaojie Lin, Tao Xiang, Ling Shao, and Steven C. H. Hoi. Deep learning for person re-identification: A survey and outlook. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(6):2872-2893, 2022. 2, 4, 6, 7 +[49] Mang Ye, Zesen Wu, and Bo Du. Dual-level matching with outlier filtering for unsupervised visible-infrared person re-identification. IEEE Transactions on Pattern Analysis and Machine Intelligence, 47(5):3815-3829, 2025. 3, 7 +[50] Xiangbo Yin, Jiangming Shi, Yachao Zhang, Yang Lu, Zhizhong Zhang, Yuan Xie, and Yanyun Qu. Robust pseudo-label learning with neighbor relation for unsupervised visible-infrared person re-identification. In Proceedings of the 32nd ACM International Conference on Multimedia (ACMMM), pages 2242–2251, 2024. 1, 3 +[51] Hao Yu, Xu Cheng, Wei Peng, Weihao Liu, and Guoying Zhao. Modality unifying network for visible-infrared person re-identification. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 11185–11195, 2023. 7 +[52] Xiaoyan Yu, Neng Dong, Liehuang Zhu, Hao Peng, and Dapeng Tao. Clip-driven semantic discovery network for visible-infrared person re-identification. IEEE Transactions on Multimedia, 27:4137–4150, 2025. 7 +[53] Qiang Zhang, Changzhou Lai, Jianan Liu, Nianchang Huang, and Jungong Han. Fmcnet: Feature-level modality compensation for visible-infrared person re-identification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7349–7358, 2022. 2 +[54] Yukang Zhang and Hanzi Wang. Diverse embedding expansion network and low-light cross-modality benchmark for visible-infrared person re-identification. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2153–2162, 2023. 1, 2, 6, 7 +[55] Zhiwei Zhao, Bin Liu, Qi Chu, Yan Lu, and Nenghai Yu. Joint color-irrelevant consistency learning and identity-aware modality adaptation for visible-infrared cross modality person re-identification. In Proceedings of the AAAI conference on artificial intelligence (AAAI), pages 3520–3528, 2021. 2 +[56] Xiying Zheng, Yukang Zhang, Yang Lu, and Hanzi Wang. Semi-supervised visible-infrared person re-identification via modality unification and confidence guidance. In Proceedings of the 32nd ACM International Conference on Multimedia (ACMMM), pages 5761-5770, 2024. 1, 3 +[57] Chuhao Zhou, Jinxing Li, Huafeng Li, Guangming Lu, Yong Xu, and Min Zhang. Video-based visible-infrared person re-identification via style disturbance defense and dual interaction. In Proceedings of the 31st ACM International Conference on Multimedia (ACMMM), page 46–55, 2023. 6 \ No newline at end of file diff --git a/ICCV/2025/Weakly Supervised Visible-Infrared Person Re-Identification via Heterogeneous Expert Collaborative Consistency Learning/images.zip b/ICCV/2025/Weakly Supervised Visible-Infrared Person Re-Identification via Heterogeneous Expert Collaborative Consistency Learning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..9aefd5147a172d5e0536305dd0ebbddfc5a05df0 --- /dev/null +++ b/ICCV/2025/Weakly Supervised Visible-Infrared Person Re-Identification via Heterogeneous Expert Collaborative Consistency Learning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6edb45f365493e3b44b4375d6e1bd2413647b7cca1360638afd436024036b7d5 +size 453735 diff --git a/ICCV/2025/Weakly Supervised Visible-Infrared Person Re-Identification via Heterogeneous Expert Collaborative Consistency Learning/layout.json b/ICCV/2025/Weakly Supervised Visible-Infrared Person Re-Identification via Heterogeneous Expert Collaborative Consistency Learning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..6b9f230dbbb0ed2710e109c2ef6d325418d0f72d --- /dev/null +++ b/ICCV/2025/Weakly Supervised Visible-Infrared Person Re-Identification via Heterogeneous Expert Collaborative Consistency Learning/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6e66612d9aaa76cacb5895457452394ee71b37eb64ac407658240272dadc0b05 +size 487366 diff --git a/ICCV/2025/Weakly-Supervised Learning of Dense Functional Correspondences/3dc2108f-7d0a-4c28-b604-10f9ccf91942_content_list.json b/ICCV/2025/Weakly-Supervised Learning of Dense Functional Correspondences/3dc2108f-7d0a-4c28-b604-10f9ccf91942_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..bb4f357456c418a899d252234532bc4d7c2ff01f --- /dev/null +++ b/ICCV/2025/Weakly-Supervised Learning of Dense Functional Correspondences/3dc2108f-7d0a-4c28-b604-10f9ccf91942_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:44196ec05b38f234f9b2f202077715aa5d4490ea6d30b5a8f9a5d680e6705ba2 +size 94267 diff --git a/ICCV/2025/Weakly-Supervised Learning of Dense Functional Correspondences/3dc2108f-7d0a-4c28-b604-10f9ccf91942_model.json b/ICCV/2025/Weakly-Supervised Learning of Dense Functional Correspondences/3dc2108f-7d0a-4c28-b604-10f9ccf91942_model.json new file mode 100644 index 0000000000000000000000000000000000000000..eeeb897e60b66b4274c1d82b2dae01473a7dd02e --- /dev/null +++ b/ICCV/2025/Weakly-Supervised Learning of Dense Functional Correspondences/3dc2108f-7d0a-4c28-b604-10f9ccf91942_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:eb11ce3fd008ce37669572fbce2a433d2ad1456cb95d92a4b50f5bc8a083d700 +size 120861 diff --git a/ICCV/2025/Weakly-Supervised Learning of Dense Functional Correspondences/3dc2108f-7d0a-4c28-b604-10f9ccf91942_origin.pdf b/ICCV/2025/Weakly-Supervised Learning of Dense Functional Correspondences/3dc2108f-7d0a-4c28-b604-10f9ccf91942_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..cf060f919c25a45a2c1699ddb5314cc594da82df --- /dev/null +++ b/ICCV/2025/Weakly-Supervised Learning of Dense Functional Correspondences/3dc2108f-7d0a-4c28-b604-10f9ccf91942_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9f48341bb2d83a12e9caef243d738d6706e54d93eff9455d926ae221d30c13bd +size 17729765 diff --git a/ICCV/2025/Weakly-Supervised Learning of Dense Functional Correspondences/full.md b/ICCV/2025/Weakly-Supervised Learning of Dense Functional Correspondences/full.md new file mode 100644 index 0000000000000000000000000000000000000000..4e3ac14f59ed301e30e84d4cd8ebc0a301b781c8 --- /dev/null +++ b/ICCV/2025/Weakly-Supervised Learning of Dense Functional Correspondences/full.md @@ -0,0 +1,324 @@ +# Weakly-Supervised Learning of Dense Functional Correspondences + +Stefan Stojanov* Linan Zhao* Yunzhi Zhang Daniel L. K. Yamins Jiajun Wu Stanford University + +# Abstract + +Establishing dense correspondences across image pairs is essential for tasks such as shape reconstruction and robot manipulation. In the challenging setting of matching across different categories, the function of an object, i.e., the effect that an object can cause on other objects, can guide how correspondences should be established. This is because object parts that enable specific functions often share similarities in shape and appearance. We derive the definition of dense functional correspondence based on this observation and propose a weakly-supervised learning paradigm to tackle the prediction task. The main insight behind our approach is that we can leverage vision-language models to pseudo-label multi-view images to obtain functional parts. We then integrate this with dense contrastive learning from pixel correspondences to distill both functional and spatial knowledge into a new model that can establish dense functional correspondence. Further, we curate synthetic and real evaluation datasets as task benchmarks. Our results demonstrate the advantages of our approach over baseline solutions consisting of off-the-shelf self-supervised image representations and grounded vision language models.1 + +# 1. Introduction + +Finding pixel correspondence across image pairs is fundamental for object understanding and is critical for applications like shape reconstruction [39, 43, 46, 72], editing [19], and object manipulation in robotics [17, 28, 29, 55]. This task requires reasoning beyond visual similarity in local appearance, geometry, and texture across images. It also involves structural similarity, e.g., the part-whole relationships of objects and their part components, and semantic similarity, e.g., the functional properties of parts of objects. + +These aspects of similarity are essential for learning efficient generalizable systems for downstream applications. For example, in imitation learning in robotics, human demonstrations are a scarce and valuable data source. + +Given a demonstration with an object, such as pouring with a kettle, establishing dense functional correspondence with another object that supports this function, e.g., a bottle, enables the efficient transfer of the demonstration. + +It becomes harder to find dense correspondence when the input images shift from being two views of the same object to different objects from the same category, and finally to objects from distinct categories, as the visual similarity becomes less apparent. This work focuses on the most challenging scenario with objects from different categories. We aim to establish dense pixel-level correspondence between pairs of images containing objects with parts whose shape enables the execution of similar functions. Specifically, by "function", we refer to the effect one object can have on another object or substance, e.g., the function "cut-with" for a knife and a spatula or "hang-onto" for objects with hooks. + +Practically, training and evaluation for this task are challenging due to the lack of labeled data. Supervised training at scale is infeasible because manual dense correspondence labeling is intractable, emphasizing the need for a self- or weakly-supervised approach. For evaluation, while datasets exist for dense within-category correspondence [34, 68, 77] and sparse functional keypoint correspondence across categories [37], there is still no established task or dataset for dense correspondence across categories. In this work, we make progress toward addressing both the challenges of training and evaluation. + +The key insight behind our training approach is that the capabilities of self-supervised image representations like DINOv2 [54] or Stable Diffusion [62] and vision language models (VLMs) [24, 78] are complementary but individually insufficient for solving this task. On the one hand, surprisingly accurate dense correspondences can be established using image features from pre-trained self-supervised models. This works well when the input images contain visually similar object instances from the same category, e.g., two cats or two cars [87]. However, the accuracy decreases for the more generic scenario when objects come from distinct categories. On the other hand, VLMs can detect the bounding boxes of object parts with similar functions in a zero-shot manner [24, 78] but cannot perform fine-grained reasoning about correspondences across objects. + +![](images/88fc4ee6528672022af4dbc16438937a712ccc9058bf78721be49c18d35ff742.jpg) +Figure 1. Dense Functional Correspondence refers to establishing dense correspondences across object instances based on function similarity (e.g., "pour-with"). This task is especially challenging when objects have visually different but functionally similar parts, requiring both semantic understanding, i.e., identifying which parts can perform the same function, and structural understanding, i.e., establishing dense correspondence across the parts at a surface-point-level based on functionally equivalent alignment. We propose a method to learn such correspondences with little human supervision, leveraging automated data curation and annotation, and dense contrastive learning. + +![](images/dcc435af69fc214bc6ac4c1bb243171581a0ef0f6c6589ed54ad2402c96e7e50.jpg) +Learned Dense Functional Correspondence + +![](images/d4cb75b14626157b631c9d8428bbcca63f739ce191489380b986960426f677ec.jpg) + +We distill the strengths of each approach into a new model using a scalable technique that requires minimal human supervision. Specifically, we first obtain multi-view-consistent pseudo labels of functionally relevant regions of 3D object assets [7] using an off-the-shelf grounded VLM [78]. We then combine these labels with multi-view correspondences [17, 64] in a contrastive learning framework building on pre-trained DINOv2 [54] feature extractor. For evaluation, we define the dense functional 2D correspondence task and develop an annotation procedure based on aligning 3D object pairs in functionally equivalent poses. + +In sum, we define the task of dense functional correspondence as a means for investigating cross-category dense correspondence. We then curate synthetic and real-world evaluation datasets for this task. We further propose a scalable, weakly-supervised method leveraging vision foundation models, which empirically outperforms baselines. + +# 2. Related Work + +Learning representations to establish dense functional correspondence requires fine-grained structural and semantic visual reasoning about objects. The most relevant prior works come from the object-level correspondences and affordance learning research domains. We also review recent work on vision foundation models and VLMs, focusing on works relating to fine-grained object understanding. + +Learning Correspondences. For this work, it is relevant to categorize correspondence learning methods based on their degrees of generalization. For generalizing across geometric scene transforms, works on multi-view correspondences aim to match different views of the same scene [27, 63, 65], whereas optical flow techniques match consecu + +tive video frames [25, 26, 69]. For generalizing within categories, NOCS-style representations [34, 76, 77] enable dense matching across instances of a category, whereas learning sparse keypoints [47, 66] enables sparse matching based on pre-defined semantic keypoint taxonomies. For generalizing across categories, Lai et al. [37] propose matching based on object function by learning five keypoints per object function category. The main drawback of keypoint-based correspondences is the requirement for a keypoint taxonomy, which by definition limits such techniques' capability to capture nuanced similarities across highly dissimilar objects (e.g., a bottle and a kettle). Through our dense functional correspondence formulation, we overcome the limitation of keypoint definitions and enable higher precision in downstream applications. + +Learning Affordances. In his seminal work [18], James J. Gibson defines affordances as objects' "opportunities for interaction." Various object affordance formalisms have been developed in computer vision and robotics, such as estimating grasps [2, 16, 48, 49], and localizing affordance regions in 2D [5, 14, 44, 50-52] and 3D [10, 20, 83] through bounding boxes and segments [14, 50-52], heatmaps [10, 44, 50] or keypoints [58, 73, 81]. Early works adopt a fully supervised learning paradigm [2, 14, 50], while more recent works aim to use less supervision by learning from human object interaction videos [51], egocentric videos [41] or unlabeled exocentric images [40, 44]. Our work has two key distinctions: First, affordance heatmaps or segments identify object regions or parts in individual images. They do not allow for fine-grained spatial correspondence across object parts in different images (e.g., can identify the blades of two knives but cannot find correspondences for pixels be + +tween the tips or edges of the blades). Second, our focus is on object function – the effect an object can cause on something else, rather than the broader concept of affordance, which emphasizes potential interactions with a specific object instance (e.g., striking with a hammer vs. holding). Last, our goal is to learn dense functional correspondence in a weakly-supervised manner, without relying on human annotations of ground-truth correspondences. + +Vision Foundation Models. Recent developments in large-scale language [11, 60, 71] and image [59, 86] pretraining have led to the development of vision-language models (VLMs) capable of strong zero-shot performance through vision-question answering [42, 78], which have been adapted to reasoning about functional affordances and grasping in robotics [13, 24, 57, 85]. Powerful correspondence representations have been found to emerge [1, 67] in DINO [3, 54] and Stable Diffusion [62], which have led to direct applications in low-shot affordances [29] and object manipulation [12, 29, 35, 55, 56]. In this work, we leverage the complementary characteristics of VLMs and self-supervised image models to go beyond their individual capabilities for dense functional correspondence. + +# 3. Dense Functional Correspondence + +Distinct object categories with similar functionality, e.g., a "kettle" and a "bottle" which can both pour liquid, may have different visual shapes and appearances as well as distinct part organizations. However, individual parts that serve the specific functionality of interest, e.g., the spout of a kettle and the mouth of a bottle in this example, have a higher resemblance with each other than at the overall object level. Such consistency is a consequence of how form follows function - object parts that fulfill a specific function tend to remain consistent across objects, even if other parts vary greatly in shapes and arrangements. The part-level consistency provides a crucial ground from which we can derive the definition of functional correspondence (Section 3.1) and develop a corresponding evaluation data curation pipeline to benchmark this task (Section 3.2). + +# 3.1. Problem Definition + +We refer to the effect that an object causes on other objects or substances as an "object function." This concept has been widely studied in model generalization in visual computing [36, 37, 70, 89] and the development of categorization in humans [31, 38, 79]. Examples are shown in Figure 1, e.g., "pour with." When executing a function with an object, such as pouring with a kettle, the functional part (the spout) follows a specific 3D trajectory. To replicate this function with a different object, e.g., a bottle, the neck of the bottle would be aligned with the spout and follow the same trajectory. This illustrates how the same object function is fulfilled with different objects via aligning function- + +![](images/2ec2e165f1c4f1810d156d9484985b8923d5096c0b8419eeec810a4dcd98b703.jpg) +Figure 2. Annotation Pipeline (Evaluation Only). Given a 3D object pair (left) and a function ("pour-with"), we annotate the functional alignment of two objects by aligning the functional parts in 3D (middle). Afterward, we derive dense 2D correspondences (right) based on 3D distances of corresponding object surface points, with matching pixels shown in the same color. + +ally equivalent parts, which is central to robotic applications with imitation learning approaches [23, 55, 90]. + +The above observations lead us to define dense functional correspondence through 3D object alignment based on functionally equivalent parts. Specifically, given two objects (e.g., a kettle and a bottle) and an object function (e.g., "pour with"), the objects are aligned if and only if the parts that fulfill this function (e.g., the kettle spout and the bottle neck) are spatially close to each other. The alignment induces an image-space distance: for any pair of pixels on the functional parts of two objects, the pixels are in functional correspondence if their respective surface points are close in 3D when the objects are aligned. Since this distance is defined at the pixel level, it is inherently dense. + +Formally, the input consists of an object function $\mathcal{F}$ and an image pair $(I_1, I_2)$ , where each image is a view of a 3D object $O_1$ and $O_2$ . Let $\pi^{-1}: I \to O$ represent the backprojection function that maps an image pixel to the corresponding 3D object surface point. We define $M(O; \mathcal{F})$ as the functional part of object $O$ responsible for executing $\mathcal{F}$ and let $M(I; \mathcal{F})$ be its projected 2D mask in the image. In our setup, the functional parts of both objects, $M(O_1; \mathcal{F})$ and $M(O_2; \mathcal{F})$ , are assumed to be aligned in 3D such that they follow the same trajectory when performing $\mathcal{F}$ . We therefore define dense functional correspondence as a mapping $f(I_1, I_2; \mathcal{F}): M(I_1; \mathcal{F}) \to M(I_2; \mathcal{F})$ that minimizes $\sum_{p \in M(I_1; \mathcal{F})} ||\pi^{-1}(p) - \pi^{-1}(f(p))||_2$ . This ensures that pixel pairs in functional correspondence are from spatially close locations in 3D when the objects are aligned. + +# 3.2. Evaluation Dataset Curation + +The problem definition in Section 3.1 provides a guiding principle to obtain ground truth annotations for dense functional correspondence in image pairs by aligning objects in $3D$ . We introduce the annotation procedure and use it to construct both synthetic and real-world evaluation datasets for quantitative evaluation in Section 5. + +![](images/5bdbaca1b0a2938c0c406a8c6798b18633ff003586e18eb2ffa4205a035b9b8c.jpg) +Figure 3. Training Data Curation via VLM Pseudo Labeling. Given a large unstructured dataset like Objaverse [7], we leverage off-the-shelf VLMs to curate and label the functional parts. Specifically, GPT-4 [53] generates category-specific functional part prompts, and CogVLM [78] produces bounding box proposals for multi-view image renderings, which are aggregated onto a 3D point cloud. The point cloud is post-processed to produce pixel-level functional part labels for training. + +Annotation Procedure. To obtain ground-truth functional correspondence for an image pair, we assume each object is rendered from a known 3D asset. By aligning the two assets in 3D, we derive dense pixel correspondences between the images. This procedure eliminates the need for manual dense 2D labeling, enabling large-scale evaluation. An overview is shown in Figure 2. + +Specifically, given two 3D meshes of objects supporting the same function, we first align them based on their functional parts and annotate a 3D bounding box around each functional part. Then, for a pair of rendered images, we unproject pixels from the functional parts onto the object surfaces and compute 3D distances between these points to perform minimum-cost matching. Pixels corresponding to visible surface points that are spatially close in 3D are matched. A detailed description of the annotation procedure is provided in the Supplement. + +Synthetic Evaluation Dataset. We use the 3D assets from Objaverse [7], a large collection of diverse 3D models, to obtain a synthetic evaluation dataset. We hand-label 3D annotations for 950 pairs of assets from Objaverse spanning 24 functions, selected for quality and diversity. See Section 4.1 for how the assets and functions are selected. As such, $85\%$ of the ground-truth pairs contain across-category correspondences. From these 3D annotations, we derive $1,800+$ unique 2D image pairs rendered from the 3D assets, with ground truth dense functional correspondences. + +Real Evaluation Dataset. Setting up a real-world benchmark is crucial for measuring model performance on real images. Thus, we utilize the HANDAL dataset [20], which contains images and 3D reconstructions of real-world objects. After manually fixing the geometry of the 3D scanned assets, e.g., the missing concavities of mugs, pots and pans, we label 190 asset pairs spanning 13 functions. This results in a real evaluation dataset of $500+$ unique 2D real image pairs with ground-truth functional correspondence labels. + +# 4. Approach + +Our goal is to develop a scalable learning framework for dense functional correspondences without relying on human-labeled ground truth. Since this task requires both semantic and structural knowledge, we distill from off-the-shelf VLMs to obtain pseudo-labeled training data (Sec. 4.1), which is further combined with dense spatial correspondences from synthetic data in a contrastive learning framework (Sec. 4.2). This approach enables the model to generalize to real-world data, as we will show in Sec. 5. + +# 4.1. Dataset of Pseudo-labeled Functional Parts + +A dataset for learning dense functional correspondences at scale requires a diverse source of object images, a diverse taxonomy of functions and associated functional parts, and a low-cost, reliable means for part labeling. + +Image Data. Our approach requires a large and diverse multi-view image dataset where functional parts are visible. Existing multi-view object datasets [61, 80, 84, 88] are suboptimal because they have few desired objects like tools and utensils, the objects are in canonical poses that may not reveal functional parts, or are placed in cluttered contexts where occlusions often occur. To overcome this, we render high-quality images from the Objaverse [7] dataset using ray-tracing and HDRI environments [21] in Blender [6], obtaining arbitrary amounts of diverse multi-view data. + +Object and Function Taxonomy. To curate relevant object assets for our training dataset, we prompt GPT-4 [53] for common functions and refer to object functions studied in [37, 49]. Then, we prompt GPT-4 to generate a comprehensive list of object categories for each function. After dedduplication and manual filtering, our taxonomy has 24 functions and 160 object categories. + +Object Asset Selection. To retrieve relevant assets from the noisy-labeled Objaverse dataset based on the list of object + +categories, we utilize asset captions from Caption3D [45]. We use Llama 3.1 [15] to summarize the captions into category names and use Llama word embeddings to match the summaries to our category list. Finally, we prompt Llama to verify these matches. To ensure diversity, we cap each category at 200 assets. To ensure quality, we manually filter the retrieved assets to obtain 8,285 assets in total, $80\%$ of which are used for training. Details about prompting, filtering, and the taxonomy are included in the Supplement. + +Functional Part Pseudo-Labeling. Labeling data at scale using large pre-trained models has been shown as an effective approach for achieving high performance with minimal human effort [75, 82]. The key elements for success are a sufficiently accurate pre-trained model and a low-cost and reliable procedure for rejecting low-quality labels. Grounded VLMs [8, 42, 78] have shown remarkable capabilities for zero-shot prompt-based object detection. We, therefore, use the 17B grounded CogVLM [78] model, which has state-of-the-art referring expression detection performance. For an overview of the pseudolabeling pipeline, see Figure 3. Given our list of object categories and functions, we prompt GPT-4 to obtain the names and appearance descriptions of functional parts to serve as prompts for CogVLM, which we then manually filtered and deduplicated. Because functional part names can be different across categories (e.g., the spout of a kettle vs. the neck of a bottle), we generate these functional part lists separately for each category. We empirically found that prompting CogVLM with part names and appearance descriptions significantly improves the bounding box predictions. + +Given a set of rendered views for an object and a functional part text prompt, we generate bounding box predictions with CogVLM [78], which vary due to sampling in VLM inference. The accuracy of the bounding boxes also depends on viewpoint because of part pose and visibility. To aggregate these possibly noisy labels and obtain a final part label, we sample a dense point cloud on the surface of the object, and accumulate the 2D labels across views onto the 3D points. We post-process these labeled point clouds to generate 2D masks for views rendered for training. + +This dataset curation and pseudo-labeling procedure allows us to generate a large dataset of functional part segmentation labels with relatively little human effort, which was mostly necessary for prompt engineering and quality control. In this work, we apply this approach on the $\approx 600\mathrm{K}$ labeled meshes from Caption3D, but it is straightforward to scale up to the millions of meshes in Objaverse-XL [9]. + +# 4.2. Learning Dense Functional Correspondence + +To learn dense functional correspondence, we train a feature embedding that captures both the high-level function semantics and the structural similarity between functional parts. For instance, given a bottle and a kettle for the func- + +![](images/b7cb85f4e4b9e68e90c340734a7f5dcd2888414cb882b9720ca5f3907725477c.jpg) +Figure 4. Local Functional Feature Extraction. To obtain dense functionally conditioned features, we apply an MLP on top of a function text embedding and the spatial DINO features. The MLP is trained with both functional and spatial contrastive losses. + +tion "pour-with," the features for the neck of the bottle and the spout of the kettle should be similar. Moreover, the mouth of the bottle and the tip of the kettle spout should be in correspondence, as well as the bottom of the bottle's neck and the bottom of the kettle's spout. To achieve this, we train a function-conditioned network on top of frozen DINOv2 [54] and CLIP [59] (illustrated in Figure 4), that is applied at the local feature level. Because of significant developments in object segmentation [30, 33] and our focus on object-level understanding, we assume that the input images consist of segmented objects. + +Function-Conditioned MLP. Given an image and a function, we first extract the image features from the last three blocks of DINOv2 and the function conditioning from CLIP text embeddings. We average the DINOv2 features from each block using learned weights into a single feature grid, and use bilinear interpolation to obtain a feature vector for each pixel location. Then, we concatenate the image feature with the CLIP embedding of the function and pass it through a 3-layer MLP, which produces the final feature at each pixel location. This network can be thought of as a function-conditioned version of the final projection layer used in contrastive learning [4, 22]. We parameterize our model as $g_{\theta}(p|I,\mathcal{F})$ , which outputs the normalized feature of pixel $p$ on image $I$ conditioned on the function $\mathcal{F}$ . + +We also investigate the option of adding an extra fully connected layer that maps the output feature vector to a prediction for the functional part mask. This allows us to obtain a binary functional part mask at inference time. + +Functional Part Contrastive Learning. To distill the knowledge of functional part semantics from the VLM, we use contrastive learning based on the pseudo-labeled functional part masks. The parts from two objects that can be used to perform the same function should share a more similar embedding space. Specifically, given two images, $I_{1}$ and $I_{2}$ of objects that can perform the same function $\mathcal{F}$ , let the functional part segments be $P_{1}^{+}$ and $P_{2}^{+}$ . Then, define the rest of the objects' pixels as $P_{1}^{-}$ and $P_{2}^{-}$ . Learning correspondence requires the pixels in $P_{1}^{+}$ to be similar to the ones in $P_{2}^{+}$ but different from the ones in $P_{2}^{-}$ . In addition, to encourage the model to focus on the functionally relevant + +regions of objects, we add a term that pushes the features of $P_{1}^{-}$ away from that of $P_{2}^{-}$ . + +Let $\mathrm{sim}(x,y|I_1,I_2,\bar{\mathcal{F}}) = g(x|I_1,\mathcal{F})\cdot g(y|I_2,\mathcal{F})$ represent the feature similarity between pixel $x$ on image $I_{1}$ and pixel $y$ on image $I_{2}$ when conditioned on function $\mathcal{F}$ . For brevity, we short-hand it to $\mathrm{sim}(x,y)$ below. The infoNCE loss [74] for the function-part contrastive learning given $(p_1^+,p_1^-,p_2^+,p_2^-)\in (P_1^+,P_1^-,P_2^+,P_2^-)$ is thus + +$$ +\mathcal {L} _ {\text {f u n c}} = - \log \frac {e ^ {\sin \left(p _ {1} ^ {+} , p _ {2} ^ {+}\right) / \tau}}{e ^ {\sin \left(p _ {1} ^ {+} , p _ {2} ^ {+}\right) / \tau} + e ^ {\sin \left(p _ {1} ^ {+} , p _ {2} ^ {-}\right) / \tau} + e ^ {\sin \left(p _ {1} ^ {-} , p _ {2} ^ {-}\right) / \tau}} \tag {1} +$$ + +for temperature $\tau$ + +When the model predicts functional part masks, we add a binary cross-entropy loss $\mathcal{L}_{\mathrm{mask}}$ to compare the predicted mask with the pseudo-labeled functional part segment. + +Part Structure via Multi-view Contrastive Learning. If we train the embedding with only the functional part contrastive loss, we inevitably run into mode collapse issues. That is, the whole spout of the kettle would have the same features regardless of the pixel's spatial location. To preserve the structural information, we apply dense contrastive learning from multi-view correspondences. + +Given two views of an object, we can find corresponding pixels that project to the same location in 3D space. We require a view-invariant feature embedding - a pixel should have high similarity with its corresponding pixel on the other image but remain different from all the other pixels. This encourages the model to learn the structural information of the object, to not collapse the embedding space, and to encode the object part consistently across different views. This multi-view contrastive objective only applies to two images of the same asset. However, because the underlying DINOv2 embedding space enables generalization for visually similar regions, the trained feature embedding can retain information about the structural similarities between functional parts across categories. + +Formally, let $q$ be a pixel in the first view $I$ , $q_{+}^{\prime}$ be a pixel in the second view $I^{\prime}$ that corresponds to the same location in 3D as $q$ , and any other pixel on $I^{\prime}$ be denoted as $q_{-}^{\prime}$ . The multiview contrastive objective is + +$$ +\mathcal {L} _ {\text {s p a t i a l}} = - \log \frac {e ^ {\sin \left(q , q _ {+} ^ {\prime}\right) / \tau}}{e ^ {\sin \left(q , q _ {+} ^ {\prime}\right) / \tau} + e ^ {\sin \left(q , q _ {-} ^ {\prime}\right) / \tau}}. \tag {2} +$$ + +Combining the terms, we obtain the final loss + +$$ +\mathcal {L} = \mathcal {L} _ {\text {f u n c}} + \lambda_ {\text {s p a t i a l}} \mathcal {L} _ {\text {s p a t i a l}} + \lambda_ {\text {m a s k}} \mathcal {L} _ {\text {m a s k}}. \tag {3} +$$ + +# 4.3. Implementation Details + +We use DINOv2-B as the backbone and an image size of 224. The MLP projector has 3 layers with 1024 hidden dimensions each. We use the Adam [32] optimizer with + +![](images/0c19194f28f7a3715b7fd95d3bc4b864d7e108bfa0a0ea971037a044e5e6abfc.jpg) +Figure 5. Training Objectives. To ensure functional part similarity in the learned feature space, we use a part-level contrastive objective to distill functional part semantics from VLMs (left). The spatial contrastive loss (right) serves a complementary role and prevents the model from collapsing predictions for different regions of a part, e.g., the top and bottom of a kettle spout. + +![](images/f89dc34c1e442a63d86c507233ffc928a52969d953db2adaefb5395f3d22e792.jpg) + +default hyperparameters, a batch size of 50 image pairs, 128 positive and negative sampled points on each image, and a learning rate of $1 \times 10^{-4}$ . In addition, we use a weight of $\lambda_{\mathrm{spatial}} = 10$ for the spatial loss and a weight of $\lambda_{\mathrm{mask}} = 1$ for the mask loss. We use random-color background augmentation during training following [17]. A sensitivity analysis of loss weights and a breakdown of computational costs are provided in the Supplement. + +# 5. Experiments + +In this section, we benchmark our approach in Sec. 4 and several baseline solutions on the dense functional correspondence task. Since our problem formulation in Sec. 3.1 requires a function as input and focuses on matches within functional parts, it differs significantly from existing benchmarks on semantic correspondence [47, 66]. As such, we leverage the evaluation datasets from Sec. 3.2. + +# 5.1. Metrics + +We evaluate dense functional correspondence from two different aspects: correspondence label transfer, which assesses the precision with which the model can transfer one functional part to another, and correspondence discovery, which assesses the model's ability to identify relevant functional correspondences without any reference input labels. + +Correspondence Label Transfer. To evaluate the precision of the correspondences that can be found using the learned features, we use normalized pixel distance (Normalized Dist) and percentage of correct keypoints (PCK). + +Specifically, let the ground-truth correspondences between images $I_1, I_2$ given the function $\mathcal{F}$ be $\{p_1^1, p_1^2, \dots, p_1^k\}, \{p_2^1, p_2^2, \dots, p_2^k\}$ . For each pixel $p_1^i$ on image $I_1$ , we can find its most similar match $p_2^{j(i)}$ on $I_2$ using feature similarity. The normalized distance metric is simply the mean of $\| p_2^{j(i)} - p_2^i \|_2$ normalized by the image size, and PCK@k pixels is the mean of $\mathbb{1}_{||p_2^{j(i)} - p_2^i ||_2 < k}$ . + +Correspondence Discovery. In addition to label transfer, models should discover the relevant set of functional correspondences on its own, without assuming a priori that the + +
ModelCorrespondence Label TransferCorrespondence Discovery
Normalized Dist (↓)PCK@23p (↑)PCK@10p (↑)Best F1@23p (↑)Best F1@10p (↑)AP@23p (↑)AP@10p (↑)
Synthetic Evaluation Dataset
Chance0.3100.1650.0460.4160.1760.2560.093
DINO [54]0.2120.3810.1480.5780.2810.3810.130
SD [87]0.2680.2980.1260.4790.2310.2670.097
SD-DINO [87]0.2270.3760.1610.5630.3010.3410.144
CogVLM [78] + DINO0.1800.4160.1580.6780.3330.5560.188
ManipVQA-P [24] + DINO0.2230.3460.1300.5750.2690.4180.134
ManipVQA-F [24] + DINO0.2720.2590.0930.5280.2440.3200.097
Ours (functional only)0.2280.2870.0940.5750.2330.4410.112
Ours (spatial only)0.2040.4700.2270.6100.3690.4120.211
Ours (full without mask loss)0.1700.4860.2270.7680.4700.6850.338
Ours (full with mask loss)0.1720.4800.2230.7740.4710.6840.330
Real Evaluation Dataset
Chance0.3130.1700.0450.4170.1670.2480.087
DINO [54]0.2060.4080.1590.5890.2940.3820.138
SD [87]0.2590.3090.1270.5030.2380.2850.101
SD-DINO [87]0.2200.3850.1630.5770.3010.3430.142
CogVLM [78] + DINO0.1720.4400.1690.6950.3500.5610.198
ManipVQA-P [24] + DINO0.2040.3980.1530.6000.2950.4200.148
ManipVQA-F [24] + DINO0.2560.3090.1140.5750.2810.3680.126
Ours (functional only)0.2000.3360.1150.6520.2830.5320.148
Ours (spatial only)0.2030.4720.2280.7080.3530.3820.182
Ours (full without mask loss)0.1520.5160.2490.7750.4760.6910.344
Ours (full with mask loss)0.1530.5010.2350.8080.5020.7300.360
+ +Table 1. Quantitative Evaluation on the synthetic and real evaluation datasets. The simplest baselines, self-supervised features from Stable Diffusion and DINo2, perform relatively poorly. Adding semantic knowledge from predicted functional part labels from VLMs can offer slight improvement. Our approach, combining the strengths of both self-supervised features and VLMs, achieves the best performance. + +relevant pixels on one image have been given. This capability is essential for potential downstream applications such as object alignment in robot object manipulation. + +First, since we assume that the input images are segmented, let $M_1, M_2$ be the object masks for images $I_1, I_2$ . For every pixel $p_1^i \in M_1$ , we find its most similar match $p_2^{j(i)}$ on $I_2$ and find the backward match of $p_2^{j(i)}$ on $I_1$ , denoted as $q_1^i$ . As such, $||p_1^i - q_1^i||_2$ captures the level of cycle-consistency of the match. We therefore construct a score $s = (1 - ||p_1^i - q_1^i||_2) \cdot \sin(p_1^i, p_2^{j(i)})$ to rank each pair of $(p_1^i, p_2^{j(i)})$ , using both similarity and cycle consistency. + +Then, we consider the top $t\%$ of all pairs as "discovered" and compare them with the ground-truth. A discovered pair $(x_{1}, x_{2})$ is equivalent to a ground-truth pair $(y_{1}, y_{2})$ if both end points are within $k$ pixels of the ground truth. Increasing $t$ results in higher recall but potentially lower precision: the number of discovered ground-truth correspondences monotonically increases while the percentage of correct correspondence tends to decrease. Sweeping $t$ produces a precision-recall curve, from which we can calculate the best F1 score (at $k$ pixels) and the average precision (AP) (at $k$ pixels). Formally, Best F1 = $\max_t \frac{2 \times \text{Precision}_t \times \text{Recall}_t}{\text{Precision}_t + \text{Recall}_t}$ and $\text{AP} = \sum_t (\text{Recall}_t - \text{Recall}_{t-1}) \text{Precision}_t$ . + +# 5.2. Baselines + +We describe several baseline methods below. + +Self-Supervised Features. Powerful correspondences emerge in the feature space of large pre-trained vision foundation models, as reviewed in Sec. 2. We use features extracted from DINOv2 [54], Stable Diffusion [62, 87], and fused features of the two [87] as baselines. We use feature-level similarity between pixel pairs to find correspondences. + +Self-Supervised Features and VLM Grounding. Since our task requires both semantic and structural reasoning based on the function prompt, these baselines chain a VLM that grounds functional parts with a pre-trained model that provides structural priors. Given an image pair, we use functional part bounding boxes generated by the VLM for each image, and then use self-supervised features to find correspondences within these part labels. This approach can benefit both label transfer and discovery because the functional part prediction adds a constraint on the space of possible matches, making it easier to find accurate matches. We consider two VLMs as the functional part grounding modules to be combined with off-the-shelf DINOv2 features: + +- CogVLM [78], which outputs bounding boxes based on prompts of the functional part. + +![](images/fb0cef783aac65f07a93d8b74c45485c275e88a0d65a19276dca95825ac5a0da.jpg) +Figure 6. Correspondence Discovery Comparisons. We observe that our approach more reliably retrieves the functionally relevant correspondences than off-the-shelf DINov2. The top 10 highest-ranked matches are shown. + +![](images/98c5af78c95e7bd7c81e959ca3b7ea52f812b330b8a33ab847783455df8b50cb.jpg) + +- ManipVQA [24], an affordance-grounding model that outputs bounding boxes conditioned on actions. We use the 7B model in our experiments. We also prompt ManipVQA in two ways, one with the functional part name and the other with the function itself because the model is finetuned for robotic tasks. We refer to these as ManipVQA-P and ManipVQA-F, respectively. + +# 5.3. Quantitative Comparisons + +Results in Table 1 evaluate the performance of our method and baseline solutions on the synthetic and real evaluation datasets introduced in Sec. 3.2. Results show that our model trained on fully synthetic data can generalize to real images. + +Compared to baseline solutions that solely use self-supervised features, our full model - trained with both functional and spatial contrastive loss - consistently outperforms. These metrics demonstrate that the pseudolabel quality is sufficient for learning meaningful functional correspondences. Additionally, given that the evaluation dataset predominantly includes cross-category pairs, Table 1 illustrates that self-supervised features struggle with cross-category generalization. Further evidence is provided in the Supplement, where we present a detailed breakdown of metrics for both within- and across-category pairs. + +Compared to baselines using VLM grounding, even with CogVLM bounding boxes as additional functional part information, off-the-shelf DINOv2 features underperform relative to our full model. The margin is generally smaller, which highlights the importance of understanding the context of the function. On the other hand, ManipVQA outputs less accurate bounding boxes, which is reflected in the metrics. In particular, prompting with the part instead of the function is significantly better, which shows the difficulty of zero-shot affordance grounding given a function name. Note also that running CogVLM inference is roughly 50 times slower than our model and running ManipVQA inference is roughly 1000 times slower than our model. + +Ablations. We ablate the role of the functional and spatial contrastive loss in Table 1. The model trained solely with functional loss performs poorly in both label transfer and + +correspondence discovery. The model trained solely with spatial loss is better but still falls short compared to the full model due to its lack of functional information. Finally, models with and without mask loss share similar performances. The model with mask loss does outperform the model without it in all metrics for correspondence discovery on the real evaluation dataset, which represents the least constrained and most realistic case. This optional mask prediction module can learn functional part masks with minimal additional cost. + +# 5.4. Qualitative Results + +We present results for correspondence label transfer in Figure 1 and correspondence discovery in Figure 6. Our model predictions not only capture object parts specific to the input function, but also preserve the structural relation among parts. Figure 6 shows top 10 matches according to the score from Section 5.1 separated by 5 pixels each. DINOv2 features are not function-aware and result in inaccurate matching, especially in cross-category image pairs. In comparison, our model produces dense matches between functional parts from different object categories with high spatial precision, e.g., matching the rim of a saucepan with the rim of a jug. Overall, our model demonstrates a deep understanding of functional and structural information of objects, which produces high-quality dense functional correspondences. + +# 6. Conclusion + +We have introduced the problem of dense functional correspondence, where input images contain objects with similar functionality but possibly come from distinct object categories. We have proposed a principled approach to obtain dense 2D functional correspondences from 3D object alignments and curated datasets for comprehensive evaluations. To tackle the task, we have presented a weakly-supervised framework that distills semantic information from vision-language models, while learning structural information through tuning self-supervised features with a multi-view contrastive loss. Our model outperforms a set of baselines in both synthetic and real-world benchmarks. + +Acknowledgments. This work is in part supported by NSF RI #2211258 and #2338203, NSF CAREER #1844724, NSF NCS-FR #2123963, ONR YIP N00014-24-1-2117, ONR MURI N00014-22-1-2740, ONR N00014-20-1-2589, ONR MURI N00014-21-1-2801, ONR MURI N00014-24-1-2748, and Simons Foundation grant SFI-AN-NC-GB-Culmination-00002986-05. We also thank the Stanford HAI for their support with computing resources. + +# References + +[1] Shir Amir, Yossi Gandelsman, Shai Bagon, and Tali Dekel. Deep vit features as dense visual descriptors. arXiv preprint arXiv:2112.05814, 2021. 3 +[2] Samarth Brahmbhatt, Cusuh Ham, Charles C Kemp, and James Hays. Contactdb: Analyzing and predicting grasp contact via thermal imaging. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8709-8719, 2019. 2 +[3] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 3 +[4] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 5 +[5] Ching-Yao Chuang, Jiaman Li, Antonio Torralba, and Sanja Fidler. Learning to act properly: Predicting and explaining affordances from images. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 975–983, 2018. 2 +[6] Blender Online Community. Blender - a 3D modelling and rendering package. Blender Foundation, Stichting Blender Foundation, Amsterdam, 2018. 4 +[7] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13142-13153, 2023. 2, 4 +[8] Matt Deitke, Christopher Clark, Sangho Lee, Rohun Tripathi, Yue Yang, Jae Sung Park, Mohammadreza Salehi, Niklas Muennighoff, Kyle Lo, Luca Soldaini, et al. Molmo and pixmo: Open weights and open data for state-of-the-art multimodal models. arXiv preprint arXiv:2409.17146, 2024. 5 +[9] Matt Deitke, Ruoshi Liu, Matthew Wallingford, Huong Ngo, Oscar Michel, Aditya Kusupati, Alan Fan, Christian Laforte, Vikram Voleti, Samir Yitzhak Gadre, et al. Objverse-xl: A universe of $10\mathrm{m} + 3\mathrm{d}$ objects. Advances in Neural Information Processing Systems, 36, 2024. 5 +[10] Shengheng Deng, Xun Xu, Chaozheng Wu, Ke Chen, and Kui Jia. 3d affordancenet: A benchmark for visual object affordance understanding. In proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1778-1787, 2021. 2 + +[11] Jacob Devlin. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018.3 +[12] Norman Di Palo and Edward Johns. On the effectiveness of retrieval, alignment, and replay in manipulation. IEEE Robotics and Automation Letters, 2024. 3 +[13] Kairui Ding, Boyuan Chen, Ruihai Wu, Yuyang Li, Zongzheng Zhang, Huan-ang Gao, Siqi Li, Guyue Zhou, Yixin Zhu, Hao Dong, et al. Preafford: Universal affordance-based pre-grasping for diverse objects and environments. arXiv preprint arXiv:2404.03634, 2024. 3 +[14] Thanh-Toan Do, Anh Nguyen, and Ian Reid. Affordancenet: An end-to-end deep learning approach for object affordance detection. In 2018 IEEE international conference on robotics and automation (ICRA), pages 5882-5889. IEEE, 2018. 2 +[15] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 5 +[16] Hao-Shu Fang, Chenxi Wang, Minghao Gou, and Cewu Lu. Graspnet-1billion: A large-scale benchmark for general object grasping. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11444–11453, 2020. 2 +[17] Peter R Florence, Lucas Manuelli, and Russ Tedrake. Dense object nets: Learning dense visual object descriptors by and for robotic manipulation. In Conference on Robot Learning, pages 373-385. PMLR, 2018. 1, 2, 6 +[18] James J Gibson. The ecological approach to visual perception: Classic edition. Taylor & Francis, 1979. 2 +[19] Yuchao Gu, Yipin Zhou, Bichen Wu, Licheng Yu, Jia-Wei Liu, Rui Zhao, Jay Zhangjie Wu, David Junhao Zhang, Mike Zheng Shou, and Kevin Tang. Videoswap: Customized video subject swapping with interactive semantic point correspondence. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7621–7630, 2024. 1 +[20] Andrew Guo, Bowen Wen, Jianhe Yuan, Jonathan Tremblay, Stephen Tyree, Jeffrey Smith, and Stan Birchfield. Handal: A dataset of real-world manipulable object categories with pose annotations, affordances, and reconstructions. In 2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 11428-11435. IEEE, 2023. 2, 4 +[21] HDRI Haven. Hdri haven. https://hdri-haven.com, 2024. Accessed: 2024-9. 4 +[22] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 5 +[23] Nick Heppert, Max Argus, Tim Welschehold, Thomas Brox, and Abhinav Valada. Ditto: Demonstration imitation by trajectory transformation. In 2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2024. 3 + +[24] Siyuan Huang, Iaroslav Ponomarenko, Zhengkai Jiang, Xiaqi Li, Xiaobin Hu, Peng Gao, Hongsheng Li, and Hao Dong. Manipvqa: Injecting robotic affordance and physically grounded information into multi-modal large language models. arXiv preprint arXiv:2403.11289, 2024. 1, 3, 7, 8 +[25] Zhaoyang Huang, Xiaoyu Shi, Chao Zhang, Qiang Wang, Ka Chun Cheung, Hongwei Qin, Jifeng Dai, and Hongsheng Li. Flowformer: A transformer architecture for optical flow. In European conference on computer vision, pages 668-685. Springer, 2022. 2 +[26] Eddy Ilg, Nikolaus Mayer, Tonmoy Saikia, Margret Keuper, Alexey Dosovitskiy, and Thomas Brox. Flownet 2.0: Evolution of optical flow estimation with deep networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2462-2470, 2017. 2 +[27] Wei Jiang, Eduard Trulls, Jan Hosang, Andrea Tagliasacchi, and Kwang Moo Yi. Cotr: Correspondence transformer for matching across images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6207-6217, 2021. 2 +[28] Zhenyu Jiang, Hanwen Jiang, and Yuke Zhu. Doduo: Learning dense visual correspondence from unsupervised semantic-aware flow. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 12420-12427. IEEE, 2024. 1 +[29] Yuanchen Ju, Kaizhe Hu, Guowei Zhang, Gu Zhang, Mingrun Jiang, and Huazhe Xu. Robo-abc: Affordance generalization beyond categories via semantic correspondence for robot manipulation. arXiv preprint arXiv:2401.07487, 2024. 1, 3 +[30] Lei Ke, Mingqiao Ye, Martin Danelljan, Yu-Wing Tai, Chi-Keung Tang, Fisher Yu, et al. Segment anything in high quality. Advances in Neural Information Processing Systems, 36, 2024. 5 +[31] Deborah G Kemler Nelson, Rachel Russell, Nell Duke, and Kate Jones. Two-year-olds will name artifacts by their functions. Child development, 71(5):1271-1288, 2000. 3 +[32] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization, 2017. 6 +[33] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 5 +[34] Akshay Krishnan, Abhijit Kundu, Kevis-Kokitsi Maninis, James Hays, and Matthew Brown. Omninocs: A unified noscs dataset and model for 3d lifting of 2d objects. arXiv preprint arXiv:2407.08711, 2024. 1, 2 +[35] Yuxuan Kuang, Junjie Ye, Haoran Geng, Jiageng Mao, Congyue Deng, Leonidas Guibas, He Wang, and Yue Wang. Ram: Retrieval-based affordance transfer for generalizable zero-shot robotic manipulation. arXiv preprint arXiv:2407.04689, 2024. 3 +[36] Hamid Laga, Michela Mortara, and Michela Spagnuolo. Geometry and context for semantic correspondences and functionality recognition in man-made 3d shapes. ACM Transactions on Graphics (TOG), 32(5):1-16, 2013. 3 + +[37] Zihang Lai, Senthil Purushwalkam, and Abhinav Gupta. The functional correspondence problem. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15772-15781, 2021. 1, 2, 3, 4 +[38] Barbara Landau, Linda Smith, and Susan Jones. Object shape, object function, and object name. Journal of memory and language, 38(1):1-27, 1998. 3 +[39] Yixing Lao, Xiaogang Xu, Xihui Liu, Hengshuang Zhao, et al. Corresnerf: Image correspondence priors for neural radiance fields. Advances in Neural Information Processing Systems, 36:40504-40520, 2023. 1 +[40] Gen Li, Varun Jampani, Deqing Sun, and Laura Sevilla-Lara. Locate: Localize and transfer object parts for weakly supervised affordance grounding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10922-10931, 2023. 2 +[41] Gen Li, Nikolaos Tsagkas, Jifei Song, Ruaridh MonWilliams, Sethu Vijayakumar, Kun Shao, and Laura Sevilla-Lara. Learning precise affordances from egocentric videos for robotic manipulation. arXiv preprint arXiv:2408.10123, 2024. 2 +[42] Ziyi Lin, Chris Liu, Renrui Zhang, Peng Gao, Longtian Qiu, Han Xiao, Han Qiu, Chen Lin, Wenqi Shao, Keqin Chen, et al. Sphinx: The joint mixing of weights, tasks, and visual embeddings for multi-modal large language models. arXiv preprint arXiv:2311.07575, 2023. 3, 5 +[43] Philipp Lindenberger, Paul-Edouard Sarlin, Viktor Larsson, and Marc Pollefeys. Pixel-perfect structure-from-motion with featuremetric refinement. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5987-5997, 2021. 1 +[44] Hongchen Luo, Wei Zhai, Jing Zhang, Yang Cao, and Dacheng Tao. Learning affordance grounding from exocentric images. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2252-2261, 2022. 2 +[45] Tiange Luo, Chris Rockwell, Honglak Lee, and Justin Johnson. Scalable 3d captioning with pretrained models. Advances in Neural Information Processing Systems, 36, 2024. 5 +[46] Jinjie Mai, Wenxuan Zhu, Sara Rojas, Jesus Zarzar, Abdullah Hamdi, Guocheng Qian, Bing Li, Silvio Giancola, and Bernard Ghanem. Tracknerf: Bundle adjusting nef from sparse and noisy views via feature tracks. arXiv preprint arXiv:2408.10739, 2024. 1 +[47] Juhong Min, Jongmin Lee, Jean Ponce, and Minsu Cho. Spair-71k: A large-scale benchmark for semantic correspondence. arXiv preprint arXiv:1908.10543, 2019. 2, 6 +[48] Arsalan Mousavian, Clemens Eppner, and Dieter Fox. 6-dof grapnet: Variational grasp generation for object manipulation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2901-2910, 2019. 2 +[49] Adithyavairavan Murali, Weiyu Liu, Kenneth Marino, Sonia Chernova, and Abhinav Gupta. Same object, different grasps: Data and semantic knowledge for task-oriented grasping. In Conference on robot learning, pages 1540-1557. PMLR, 2021. 2, 4 + +[50] Austin Myers, Ching L Teo, Cornelia Fermüller, and Yiannis Aloimonos. Affordance detection of tool parts from geometric features. In 2015 IEEE International Conference on Robotics and Automation (ICRA), pages 1374-1381. IEEE, 2015. 2 +[51] Tushar Nagarajan, Christoph Feichtenhofer, and Kristen Grauman. Grounded human-object interaction hotspots from video. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8688-8697, 2019. 2 +[52] Anh Nguyen, Dimitrios Kanoulas, Darwin G. Caldwell, and Nikos G. Tsagarakis. Object-based affordances detection with convolutional neural networks and dense conditional random fields. In 2017 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 5908-5915, 2017. 2 +[53] OpenAI, Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, Red Avila, Igor Babuschkin, Suchir Balaji, Valerie Balcom, Paul Baltescu, Haiming Bao, Mohammad Bavarian, Jeff Belgium, Irwan Bello, Jake Berdine, Gabriel BernadettShapiro, Christopher Berner, Lenny Bogdonoff, Oleg Boiko, Madelaine Boyd, Anna-Luisa Brakman, Greg Brockman, Tim Brooks, Miles Brundage, Kevin Button, Trevor Cai, Rosie Campbell, Andrew Cann, Brittany Carey, Chelsea Carlson, Rory Carmichael, Brooke Chan, Che Chang, Fotis Chantzis, Derek Chen, Sully Chen, Ruby Chen, Jason Chen, Mark Chen, Ben Chess, Chester Cho, Casey Chu, Hyung Won Chung, Dave Cummings, Jeremiah Currier, Yunxing Dai, Cory Decareaux, Thomas Degry, Noah Deutsch, Damien Deville, Arkar Dhar, David Dohan, Steve Dowling, Sheila Dunning, Adrien Ecoffet, Atty Eleti, Tyna Eloundou, David Farhi, Liam Fedus, Niko Felix, Simon Posada Fishman, Juston Forte, Isabella Fulford, Leo Gao, Elie Georges, Christian Gibson,Vik Goel, Tarun Gogineni, Gabriel Goh, Rapha Gontijo-Lopes, Jonathan Gordon Morgan Grafstein, Scott Gray, Ryan Greene, Joshua Gross, Shixiang Shane Gu, Yufei Guo, Chris Hallacy, Jesse Han, Jeff Harris, Yuchen He, Mike Heaton, Johannes Heidecke, Chris Hesse, Alan Hickey, Wade Hickey, Peter Hoeschele, Brandon Houghton, Kenny Hsu, Shengli Hu, Xin HuJoost Huizinga, Shantanu Jain, Shawn Jain, Joanne JangAngela Jiang,Roger Jiang,Haozhun Jin,Denny JinShino Jomoto Billie Jonn Heewoo Jun Tomer Kaftan Lukasz Kaiser Ali Kamali Ingmar Kanitscheider Nitish Shirish Keskar Tabarak Khan Logan Kilpatrick,Jong Wook KimChristina Kim,Yongjik Kim Jan Hendrik Kirchner Jamie Kiros Matt Knight Daniel Kokotajlo,Lukasz KondraciukAndrew Kondrich Aris Konstantinidis Kyle Kosic,Gretchen Krueger,Vishal Kuo,Michael Lampe,Ikai LanTeddy Lee Jan Leike,Jade Leung,Daniel LevyChak Ming LiRachel LimMolly Lin Stephanie LinMateusz LitwinTheresa Lopez,Ryan Lowe Patricia LueAnna Makanju,Kim MalfaciniSam Manning,Todor Markov,Yaniv Markovsky Bianca Martin Katie Mayer Andrew MayneBob McGrewScott Mayer McKinney Christine McLeaveyPaul McMillan Jake McNeil David Medina Aalok Mehta Jacob Menick Luke Metz Andrey Mishchenko Pamela Mishkin + +Vinnie Monaco, Evan Morikawa, Daniel Mossing, Tong Mu, Mira Murati, Oleg Murk, David Mély, Ashvin Nair, Reichiro Nakano, Rajeev Nayak, Arvind Neelakantan, Richard Ngo, Hyeonwoo Noh, Long Ouyang, Cullen O'Keefe, Jakub Pachocki, Alex Paino, Joe Palermo, Ashley Pantuliano, Giambattista Parascandolo, Joel Parish, Emy Parparita, Alex Passos, Mikhail Pavlov, Andrew Peng, Adam Perelman, Felipe de Avila Belbute Peres, Michael Petrov, Henrique Ponde de Oliveira Pinto, Michael, Pokorny, Michelle Pokrass, Vitchyr H. Pong, Tolly Powell, Alethea Power, Boris Power, Elizabeth Proehl, Raul Puri, Alec Radford, Jack Rae, Aditya Ramesh, Cameron Raymond, Francis Real, Kendra Rimbach, Carl Ross, Bob Rotsted, Henri Roussez, Nick Ryder, Mario Saltarelli, Ted Sanders, Shibani Santurkar, Girish Sastry, Heather Schmidt, David Schnurr, John Schulman, Daniel Selsam, Kyla Sheppard, Toki Sherbakov, Jessica Shieh, Sarah Shoker, Pranav Shyam, Szymon Sidor, Eric Sigler, Maddie Simens, Jordan Sitkin, Katarina Slama, Ian Sohl, Benjamin Sokolowsky, Yang Song, Natalie Staudacher, Felipe Petroski Such, Natalie Summers, Ilya Sutskever, Jie Tang, Nikolas Tezak, Madeleine B. Thompson, Phil Tillet, Amin Tootoonchian, Elizabeth Tseng, Preston Tuggle, Nick Turley, Jerry Tworek, Juan Felipe Cerón Uribe, Andrea Vallone, Arun Vijayvergiya, Chelsea Voss, Carroll Wainwright, Justin Jay Wang, Alvin Wang, Ben Wang, Jonathan Ward, Jason Wei, CJ Weinmann, Akila Welihinda, Peter Welinder, Jiayi Weng, Lilian Weng, Matt Wiethoff, Dave Willner, Clemens Winter, Samuel Wolrich, Hannah Wong, Lauren Workman, Sherwin Wu, Jeff Wu, Michael Wu, Kai Xiao Tao Xu Sarah Yoo Kevin Yu Qiming Yuan Wojciech Zaremba Rowan Zellers Chong Zhang Marvin Zhang Shengjia Zhao Tianhao Zheng Juntang Zhuang William Zhuk and Barret Zoph. Gpt-4 technical report 2024. 4 +[54] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 1, 2, 3, 5, 7 +[55] Norman Di Palo and Edward Johns. Dinobot: Robot manipulation via retrieval and alignment with vision foundation models. In IEEE International Conference on Robotics and Automation (ICRA), 2024. 1, 3 +[56] Georgios Papagiannis and Edward Johns. Miles: Making imitation learning easy with self-supervision. In Proceedings of the Conference on Robot Learning (CoRL), 2024. 3 +[57] Shengyi Qian, Weifeng Chen, Min Bai, Xiong Zhou, Zhuowen Tu, and Li Erran Li. Affordancelm: Grounding affordance from vision language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7587-7597, 2024. 3 +[58] Zengyi Qin, Kuan Fang, Yuke Zhu, Li Fei-Fei, and Silvio Savarese. Keto: Learning keypoint representations for tool manipulation. In 2020 IEEE International Conference on Robotics and Automation (ICRA), pages 7278-7285. IEEE, 2020. 2 +[59] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning + +transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 3, 5 +[60] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of Machine Learning Research, 21(140):1-67, 2020. 3 +[61] Jeremy Reizenstein, Roman Shapovalov, Philipp Henzler, Luca Sbordone, Patrick Labatut, and David Novotny. Common objects in 3d: Large-scale learning and evaluation of real-life 3d category reconstruction. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10901-10911, 2021. 4 +[62] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 3, 7 +[63] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superglue: Learning feature matching with graph neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4938–4947, 2020. 2 +[64] Stefan Stojanov, Anh Thai, Zixuan Huang, and James M Rehg. Learning dense object descriptors from multiple views for low-shot category generalization. Advances in Neural Information Processing Systems, 35:12566-12580, 2022. 2 +[65] Jiaming Sun, Zehong Shen, Yuang Wang, Hujun Bao, and Xiaowei Zhou. Loftr: Detector-free local feature matching with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8922-8931, 2021. 2 +[66] Yixuan Sun, Yiwen Huang, Haijing Guo, Yuzhou Zhao, Run-min Wu, Yizhou Yu, Weifeng Ge, and Wenqiang Zhang. Misc210k: A large-scale dataset for multi-instance semantic correspondence. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7121-7130, 2023. 2, 6 +[67] Luming Tang, Menglin Jia, Qianqian Wang, Cheng Perng Phoo, and Bharath Hariharan. Emergent correspondence from image diffusion. Advances in Neural Information Processing Systems, 36:1363-1389, 2023. 3 +[68] Tatsunori Taniai, Sudipta N Sinha, and Yoichi Sato. Joint recovery of dense correspondence and cosegmentation in two images. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4246-4255, 2016. 1 +[69] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16, pages 402-419. Springer, 2020. 2 +[70] Skye Thompson, Leslie Pack Kaelbling, and Tomas Lozano-Perez. Shape-based transfer of generic skills. In 2021 IEEE International Conference on Robotics and Automation (ICRA), pages 5996-6002. IEEE, 2021. 3 +[71] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste + +Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. +Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 3 +[72] Prune Truong, Marie-Julie Rakotosaona, Fabian Manhardt, and Federico Tombari. Sparf: Neural radiance fields from sparse and noisy poses. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4190-4200, 2023. 1 +[73] Dylan Turpin, Liquan Wang, Stavros Tsogkas, Sven Dickinson, and Animesh Garg. Gift: Generalizable interaction-aware functional tool affordances without labels. arXiv preprint arXiv:2106.14973, 2021. 2 +[74] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding, 2019. 6 +[75] Raviteja Vemulapalli, Hadi Pouransari, Fartash Faghri, Sachin Mehta, Mehrdad Farajtabar, Mohammad Rastegari, and Oncel Tuzel. Knowledge transfer from vision foundation models for efficient training of small task-specific models. In Forty-first International Conference on Machine Learning, 2024. 5 +[76] Boyan Wan, Yifei Shi, and Kai Xu. Socs: Semantically-aware object coordinate space for category-level 6d object pose estimation under large shape variations. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14065-14074, 2023. 2 +[77] He Wang, Srinath Sridhar, Jingwei Huang, Julien Valentin, Shuran Song, and Leonidas J Guibas. Normalized object coordinate space for category-level 6d object pose and size estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2642-2651, 2019. 1, 2 +[78] Weihan Wang, Qingsong Lv, Wenmeng Yu, Wenyi Hong, Ji Qi, Yan Wang, Junhui Ji, Zhuoyi Yang, Lei Zhao, Xixuan Song, et al. Cogvlm: Visual expert for pretrained language models. arXiv preprint arXiv:2311.03079, 2023. 1, 2, 3, 4, 5, 7 +[79] Elizabeth A Ware and Amy E Booth. Form follows function: Learning about function helps children learn about shape. Cognitive Development, 25(2):124-137, 2010. 3 +[80] Tong Wu, Jiarui Zhang, Xiao Fu, Yuxin Wang, Liang Pan Jiawei Ren, Wayne Wu, Lei Yang, Jiaqi Wang, Chen Qian, Dahua Lin, and Ziwei Liu. Omniobject3d: Large-vocabulary 3d object dataset for realistic perception, reconstruction and generation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 4 +[81] Ruinian Xu, Fu-Jen Chu, Chao Tang, Weiyu Liu, and Patricio A Vela. An affordance keypoint detection network for robot manipulation. IEEE Robotics and Automation Letters, 6(2):2870-2877, 2021. 2 +[82] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10371-10381, 2024. 5 +[83] Yuhang Yang, Wei Zhai, Hongchen Luo, Yang Cao, Jiebo Luo, and Zheng-Jun Zha. Grounding 3d object affordance + +from 2d interactions in images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10905-10915, 2023. 2 +[84] Xianggang Yu, Mutian Xu, Yidan Zhang, Haolin Liu, Chongjie Ye, Yushuang Wu, Zizheng Yan, Chenming Zhu, Zhangyang Xiong, Tianyou Liang, et al. Mvimgnet: A large-scale dataset of multi-view images. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9150-9161, 2023. 4 +[85] Wentao Yuan, Jiafei Duan, Valts Blukis, Wilbert Pumacay, Ranjay Krishna, Adithyavairavan Murali, Arsalan Mousavian, and Dieter Fox. Robopoint: A vision-language model for spatial affordance prediction for robotics. arXiv preprint arXiv:2406.10721, 2024. 3 +[86] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11975-11986, 2023. 3 +[87] Junyi Zhang, Charles Herrmann, Junhwa Hur, Luisa Polania Cabrera, Varun Jampani, Deqing Sun, and Ming-Hsuan Yang. A tale of two features: Stable diffusion complements dino for zero-shot semantic correspondence, 2023. 1, 7 +[88] Chenchen Zhu, Fanyi Xiao, Andres Alvarado, Yasmine Babaei, Jiabo Hu, Hichem El-Mohri, Sean Culatana, Roshan Sumbaly, and Zhicheng Yan. Egoobjects: A large-scale egocentric dataset for fine-grained object understanding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20110-20120, 2023. 4 +[89] Yixin Zhu, Yibiao Zhao, and Song Chun Zhu. Understanding tools: Task-oriented object modeling, learning and recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2855-2864, 2015. 3 +[90] Yifeng Zhu, Arisrei Lim, Peter Stone, and Yuke Zhu. Vision-based manipulation from single human video with open-world object graphs. arXiv preprint arXiv:2405.20321, 2024. 3 \ No newline at end of file diff --git a/ICCV/2025/Weakly-Supervised Learning of Dense Functional Correspondences/images.zip b/ICCV/2025/Weakly-Supervised Learning of Dense Functional Correspondences/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..9f5a86b4e120ce6d9f7d44ab803ca10ef6cafc80 --- /dev/null +++ b/ICCV/2025/Weakly-Supervised Learning of Dense Functional Correspondences/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c2bfc552d897395999b0aea02acde948183f8a2c94616a0b799bef5602587b25 +size 475596 diff --git a/ICCV/2025/Weakly-Supervised Learning of Dense Functional Correspondences/layout.json b/ICCV/2025/Weakly-Supervised Learning of Dense Functional Correspondences/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..1d36a5c02fb170af4a943ed895297d96a3505a2b --- /dev/null +++ b/ICCV/2025/Weakly-Supervised Learning of Dense Functional Correspondences/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:eccfa4c967611ae8036d560617bfac691784e01239fa76300088afafb6ffd1cf +size 435108 diff --git a/ICCV/2025/WeaveSeg_ Iterative Contrast-weaving and Spectral Feature-refining for Nuclei Instance Segmentation/4176a988-4d3a-45b0-ab15-eb8f0b2fa50e_content_list.json b/ICCV/2025/WeaveSeg_ Iterative Contrast-weaving and Spectral Feature-refining for Nuclei Instance Segmentation/4176a988-4d3a-45b0-ab15-eb8f0b2fa50e_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..4e0bb25469951c4a7ca80e67131c6f458d465114 --- /dev/null +++ b/ICCV/2025/WeaveSeg_ Iterative Contrast-weaving and Spectral Feature-refining for Nuclei Instance Segmentation/4176a988-4d3a-45b0-ab15-eb8f0b2fa50e_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a396a19ebc1b186ab570f407ba22d0504ce5c07a2ec3131fba21fde2f6ae4b42 +size 82396 diff --git a/ICCV/2025/WeaveSeg_ Iterative Contrast-weaving and Spectral Feature-refining for Nuclei Instance Segmentation/4176a988-4d3a-45b0-ab15-eb8f0b2fa50e_model.json b/ICCV/2025/WeaveSeg_ Iterative Contrast-weaving and Spectral Feature-refining for Nuclei Instance Segmentation/4176a988-4d3a-45b0-ab15-eb8f0b2fa50e_model.json new file mode 100644 index 0000000000000000000000000000000000000000..c3e02543da3dd5f911700f3782545b7178411f7c --- /dev/null +++ b/ICCV/2025/WeaveSeg_ Iterative Contrast-weaving and Spectral Feature-refining for Nuclei Instance Segmentation/4176a988-4d3a-45b0-ab15-eb8f0b2fa50e_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e5927264dce70d8d804511390dae88d3f771e1834ffed063d62c299ad0725f39 +size 103218 diff --git a/ICCV/2025/WeaveSeg_ Iterative Contrast-weaving and Spectral Feature-refining for Nuclei Instance Segmentation/4176a988-4d3a-45b0-ab15-eb8f0b2fa50e_origin.pdf b/ICCV/2025/WeaveSeg_ Iterative Contrast-weaving and Spectral Feature-refining for Nuclei Instance Segmentation/4176a988-4d3a-45b0-ab15-eb8f0b2fa50e_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..bd1359e97bf22a060b568969ef6ffb5be2f93f0a --- /dev/null +++ b/ICCV/2025/WeaveSeg_ Iterative Contrast-weaving and Spectral Feature-refining for Nuclei Instance Segmentation/4176a988-4d3a-45b0-ab15-eb8f0b2fa50e_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b6025ed4e6506588f108f1858baae001fed05b2482c3d481a62860824b89a5b2 +size 3153490 diff --git a/ICCV/2025/WeaveSeg_ Iterative Contrast-weaving and Spectral Feature-refining for Nuclei Instance Segmentation/full.md b/ICCV/2025/WeaveSeg_ Iterative Contrast-weaving and Spectral Feature-refining for Nuclei Instance Segmentation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..7935049e8211a77042c839cfea2bcb5638d029d5 --- /dev/null +++ b/ICCV/2025/WeaveSeg_ Iterative Contrast-weaving and Spectral Feature-refining for Nuclei Instance Segmentation/full.md @@ -0,0 +1,372 @@ +# WeaveSeg: Iterative Contrast-weaving and Spectral Feature-refining for Nuclei Instance Segmentation + +Jiajia Li $^{1}$ Huisi Wu $^{1*}$ Jing Qin $^{2}$ + +1 College of Computer Science and Software Engineering, Shenzhen University + +2 Centre for Smart Health, School of Nursing, The Hong Kong Polytechnic University + +2400101028@mails.szu.edu.cn, hswu@szu.edu.cn, harry.qin@polyu.edu.hk + +# Abstract + +Automatic and accurate nuclei instance segmentation in histopathology images is a fundamental task in computational pathology. It is also a very challenging task due to complex nuclei morphologies, ambiguous boundaries, and staining variations. Existing methods often struggle to precisely delineate overlapping nuclei and handle class imbalance. We introduce WeaveSeg, a novel deep learning framework that synergistically integrates two key innovations: an adaptive spectral refinement (ASR) module to enhance high-frequency boundary details often blurred by standard convolutions, and an iterative contrast-weaving (ICW) module. Guided by a specialized uncertainty-aware loss, this module leverages contrastive attention and a novel local self-adaptive mechanism to progressively resolve ambiguous instances. Extensive experiments on MoNuSeg, CoNSEp, and CPM17 demonstrate WeaveSeg's SOTA performance over existing models. The code is available at https://github.com/jj-sterne/WeaveSeg. + +# 1. Introduction + +Accurate segmentation of cell nuclei within histopathological images is a fundamental task in computational pathology, crucial for advancing cancer diagnosis, treatment planning, and prognosis [8, 15, 25, 33]. The morphological characteristics of nuclei, including size, shape, texture, and spatial arrangement, provide critical information for histopathologists to evaluate tumor malignancy and guide clinical decisions [29, 35, 42]. Automating this process through deep learning models has significant potential to enhance diagnostic precision and efficiency, and thus improve treatment outcomes [18, 34, 37, 40]. Despite advances in deep learning-based medical image segmentation, the precise segmentation of nuclei in histopathological images remains a formidable challenge [21]. + +![](images/e2a172f48b4e3626b57f85648ea2f71558c2d44218517f437b42e8637c1975eb.jpg) +(a) + +![](images/b81e1e7d1f8afd1eeeeb6f08f694bb76dce62b6d9a71609b20b232655ce1603c.jpg) +(b) + +![](images/624c7d77769d4ecfbcce01e58d1dcdcc37d27ba4368dc8ccb1be117c6121cdc5.jpg) +(c) +Figure 1. Challenges in Nuclei segmentation. (a) a representative colorectal adenocarcinoma image; (b) ambiguous boundaries between nuclei and surrounding stroma; (c) morphological heterogeneity of nuclei; and (d) densely packed and overlapping nuclei + +![](images/84937899889289995c09b1b276e664f3d75fa12793186a4899eced2852db554a.jpg) +(d) + +As illustrated in Figure 1, several inherent complexities within histopathology images impede accurate nuclei boundary delineation. First, because of weak staining, inflammation, or obscuring cellular structures, there exist many ambiguous boundaries between nuclei and the surrounding stroma (Figure 1 (b)), making it quite difficult to accurately identify these boundaries. Second, histopathological nuclei exhibit substantial morphological heterogeneity (Figure 1 (c)); variations in shape, size, and texture across cell types and disease states complicate the definition of universally applicable features. Third, nuclei frequently appear in densely packed clusters (Figure 1(d)), leading to significant overlap and occlusion, hindering the separation of individual nuclei, especially with indistinct boundaries. Furthermore, staining intensity and color balance variations across samples, along with potential class imbalance (rarity of certain nuclei types), present additional challenges. + +While existing deep learning methods for nuclei segmentation have shown promise, they often fall short of comprehensively addressing these challenges. Encoder-decoder architectures, exemplified by U-Net [31] and its variants [23, 45], while effective for multi-scale feature fusion, tend to blur fine boundary details due to repeated downsampling and upsampling. Instance segmentation approaches [11], are usually confounded by severe nuclei overlap and class imbalance, resulting in missed detections or false positives. Furthermore, boundary-based methods [3, 44], while directly leveraging boundary informa + +tion, are vulnerable to noise and indistinct boundaries. Distance-based methods [8, 12, 25], designed to separate touching nuclei, still exhibit limitations when confronted with highly irregular nuclei shapes. Topology-aware approaches [24, 38], despite enhancing shape representation, do not directly address boundary ambiguity. Attention and recurrent mechanisms [7, 13, 41], aimed at improving feature learning, lack specific focus on challenges like nuclei overlap or staining variations. + +A critical limitation of conventional models is their reliance on spatial-only convolutions. While effective for general feature extraction, repeated downsampling and upsampling operations inherently blur the subtle yet vital high-frequency details. Unlike in natural images, histopathological nuclei depend on these high-frequency cues for accurate delineation. Existing frequency-based methods [4, 9, 30], often address global characteristics and fail to capture the localized, boundary-specific spectral information crucial for separating instances. + +To comprehensively address above limitations, in this paper, we present WeaveSeg, a novel deep learning model for nuclei segmentation that integrates an iterative contrast enhancement strategy with an adaptive activation mechanism, and a frequency-aware detail adaptation across multiple scales. The core of the WeaveSeg is the iterative contrast-weaving (ICW) module, which goes far beyond traditional simple feature fusion. ICW sufficiently exploits the intrinsic contrast between nuclei, surrounding tissue, and ambiguous boundary regions, iteratively disentangling and refining feature representations using a unique "weaving" process. This process, combined with an adaptive activation mechanism (AGLU) [1], allows the network to progressively resolve uncertainties and enhance the discriminability of nuclei features. Complementing ICW, the adaptive spectral detail refinement (ASR) module explicitly addresses frequency-domain inconsistencies that often arise from feature fusion in encoder-decoder architectures. The ASR leverages adaptive filtering to enhance both semantic consistency within nuclei and boundary sharpness. Our main contributions can be summarized as follows. + +- We propose WeaveSeg, a framework that synergistically integrates spectral and spatial refinement. It explicitly targets blurred boundaries and overlapping instances by coupling ASR with ICW. +- We introduce two innovative components: the ASR module for adaptive, boundary-aware frequency adaptation, and the ICW module, which leverages contrastive attention and is guided by a specialized uncertainty-aware loss for feature disentanglement. +- Extensive experiments on publicly available histopathological nuclei segmentation datasets demonstrate that WeaveSeg achieves state-of-the-art performance. + +# 2. Related Works + +# 2.1. Nuclei Instance Segmentation + +Early deep learning methods primarily relied on encoder-decoder architectures like U-Net [31] and its variants [27, 45]. While effective for multi-scale feature fusion, they often suffer from boundary blurring due to repeated down-sampling and upsampling, particularly affecting the delineation of small and irregularly shaped nuclei. To address boundary blurring, boundary-based methods [3, 44] utilize contour information, but their performance is highly sensitive to image noise and inherent indistinctness of nuclei boundaries in histopathology images. Distance-based methods [8, 12, 25] improve separation of touching nuclei, but struggle with highly irregular shapes and often requires precise shape priors. Topology-aware approaches [24, 38] offer improvements in shape representation by incorporating topological constraints, but their ability to resolve fundamental issues like ambiguous boundaries and significant morphological variations remains limited. Attention mechanisms [13] and recurrent processing [7] are helpful in enhancing feature discriminability, but often lack explicit mechanisms to directly tackle boundary ambiguity and severe nuclei overlap, crucial factors in accurate instance segmentation. + +In contrast, WeaveSeg's iterative refinement with contrastive attention and uncertainty loss is proposed to comprehensively address these multifaceted challenges. + +# 2.2. Frequency-Domain Feature Enhancement + +Frequency-domain analysis offers a complementary perspective in the realm of medical image analysis. Qin et al. introduced FcaNet [30], which pioneers the application of channel attention mechanisms within the frequency domain. Subsequently, Guo et al. [9] explored the utility of adaptive frequency filters for capturing long-range dependencies in feature representations. More recently, Li et al. [20] leveraged a high-frequency prior to enhance superresolution performance. While these methodologies effectively demonstrate the inherent benefits of incorporating frequency information, they predominantly concentrate on global frequency characteristics. + +In contrast to these approaches, the ASR module of WeaveSeg is specifically designed to focus on localized, boundary-aware frequency information. By adaptively fusing frequency components at a granular, local level, WeaveSeg aims to address the critical challenge of discerning subtle textural and structural disparities between nuclei and the surrounding stroma. + +# 2.3. Contrastive and Decoupling Learning + +Contrastive learning methodologies [5, 10, 18] have garnered considerable attention and demonstrated substantial + +Figure 2. Overview of the WeaveSeg framework. The architecture features three main components: foreground, background, and uncertainty regions, and incorporates two key modules: iterative contrast-weaving (ICW) and adaptive spectral detail refinement (ASR). +![](images/234d841d2ff7f6bc2387dffd2560ac153d8e39e6bfb72506879f784b906056ac.jpg) +Iterative skip connections +loss skip connections +functional connectivity +$\bigoplus$ pixel-by-pixel addition + +promise in facilitating the learning of highly discriminative feature representations. Nevertheless, the majority of existing applications of contrastive learning within medical image segmentation [2, 19, 39] predominantly employ global or pixel-level contrastive strategies. + +Distinctly, the ICW module of WeaveSeg uniquely integrates localized, boundary-aware contrastive learning principles. While decoupling methodologies have proven effective in enhancing feature representation capabilities, they frequently lack dedicated mechanisms to effectively manage uncertainty inherent in ambiguous regions. WeaveSeg directly addresses this critical limitation through the incorporation of a specialized uncertainty-aware loss function, explicitly designed for the sophisticated modeling of ambiguity associated with nuclei boundaries and overlaps. + +# 3. Methodology + +# 3.1. Overall architecture + +WeaveSeg (Figure 2) is an encoder-decoder architecture for precise nuclei instance segmentation in histopathology images. The encoder extracts multi-scale feature hierarchies from input imagery. Subsequently, the decoder processes these features via two pathways: frequency-aware refinement and contrast enhancement. Notably, both pathways converge within the iterative contrast-weaving (ICW) module for core feature integration. + +The adaptive spectral detail refinement (ASR) module (Section 3.2) lies within the first pathway. ASR selectively enhances high-frequency components associated with nu + +clei boundaries and fine textures, while preserving low-frequency information crucial for overall nuclei structure, providing a refined feature representation. The second pathway decouples the encoder features into foreground (nuclei), background (surrounding tissue), and uncertain region components, facilitating targeted contrast enhancement in ambiguous boundary areas. + +The iterative contrast-weaving module (Section 3.3) is central to WeaveSeg as the core processing unit. ICW iteratively refines features via foreground-background contrast, guided by ASR's frequency-adapted features. Critically, a closed-loop iteration feeds intermediate ICW outputs back to ASR. This ASR-ICW interplay(ASR refines spectra and ICW enhances contrast) ultimately resolves ambiguities, generating a precise nuclei instance segmentation map as the final ICW output. + +# 3.2. Adaptive Spectral Detail Refinement (ASR) + +Standard upsampling and feature fusion methods often struggle to preserve the high-frequency details crucial for accurate nuclei delineation, consequently leading to blurred boundaries. To mitigate this issue, we introduce the ASR module, illustrated in Figure 3(a), which directly integrates frequency-aware processing into the feature fusion and upsampling pipeline. ASR is designed to promote both intranuclei semantic consistency and sharp inter-nuclei boundaries. + +ASR takes as input a low-resolution, high-level feature map, $\mathbf{F}_{\mathrm{high}} \in \mathbb{R}^{C \times H \times W}$ , from a deeper decoder layer, and a high-resolution, low-level feature map, $\mathbf{F}_{\mathrm{low}} \in$ + +![](images/16715452e40ab7019411b63b72922630ce1e7180dc3259e21b8aa6e978f0e2f1.jpg) +(a) ASR + +![](images/282f7355283562f5f22d9ffa18fd0d3813bb92b91fe14076473edaaa7e7319a0.jpg) +(b) ICM +Figure 3. Detailed architecture of the proposed modules. (a) ASR enhances features by adaptively fusing frequency components and employing Local Similarity Guided Sampling for high-fidelity upsampling. (b) ICW iteratively refines features using contrastive attention, dense connections, and an adaptive gated linear unit (AGLU) for modulation. + +$\mathbb{R}^{C\times 2H\times 2W}$ , from a shallower decoder layer. It produces a refined output, $\mathbf{F}_{\mathrm{out}}\in \mathbb{R}^{C\times 2H\times 2W}$ , enriched with high-frequency details. + +The ASR module adaptively generates both low- and high-frequency feature components through distinct branches. For low-frequency enhancement, aiming to promote intra-nuclei consistency, ASR learns a spatially-varying low-pass filter, $\mathbf{W}_{\mathrm{LP}} \in \mathbb{R}^{K^2 \times 2H \times 2W}$ . This learnable filter, implicitly modulated by a Hamming window due to localized convolution, allows for adaptive smoothing based on feature content. A low-pass filtered version of $\mathbf{F}_{\mathrm{high}}$ , denoted $\mathbf{F}_{\mathrm{high}}^{\mathrm{LP}}$ , is then obtained via depthwise convolution ( $\ast$ ) and pixel shuffling: + +$$ +\mathbf {F} _ {\text {h i g h}} ^ {\mathrm {L P}} = \operatorname {P i x e l S h u f f l e} \left(\mathbf {W} _ {\mathrm {L P}} * \mathbf {F} _ {\text {h i g h}}\right) \tag {1} +$$ + +Complementing this, to effectively recover fine details, ASR employs Local Similarity-Guided Resampling (LSGR). LSGR leverages a local cosine similarity map, $\mathbf{S} \in \mathbb{R}^{8 \times 2H \times 2W}$ , computed from the combined features $\mathbf{F}_{\mathrm{comb}}$ . This similarity map guides the prediction of spatial offsets, $\mathbf{O} \in \mathbb{R}^{2G \times 2H \times 2W}$ , enabling content-aware resampling that preserves high-frequency details: + +$$ +\mathbf {O} = \operatorname {C o n v} _ {3 \times 3} ([ \mathbf {F} _ {\text {c o m b}}, \mathbf {S} ]) \tag {2} +$$ + +These offsets resample $\mathbf{F}_{\mathrm{high}}^{\mathrm{LP}}$ to produce a refined high-level map, $\mathbf{F}_{\mathrm{high}}^{\mathrm{R}}$ : + +$$ +\mathbf {F} _ {\text {h i g h}} ^ {\mathrm {R}} (i, j) = \mathbf {F} _ {\text {h i g h}} ^ {\mathrm {L P}} (i + u, j + v) \tag {3} +$$ + +In parallel, a high-frequency enhancement branch, mirroring the low-frequency branch's structure, operates on $\mathbf{F}_{\mathrm{low}}$ to generate a high-frequency-enhanced map, $\mathbf{F}_{\mathrm{low}}^{\mathrm{HF}}$ , adaptively generating high-pass filters analogous to the low-pass filter generation. + +Finally, element-wise addition fuses the resampled high-level map, $\mathbf{F}_{\mathrm{high}}^{\mathrm{R}}$ , and the high-frequency-enhanced low-level map, $\mathbf{F}_{\mathrm{low}}^{\mathrm{HF}}$ : + +$$ +\mathbf {F} _ {\text {o u t}} = \mathbf {F} _ {\text {h i g h}} ^ {\mathrm {R}} + \mathbf {F} _ {\text {l o w}} ^ {\mathrm {H F}} \tag {4} +$$ + +Through this frequency-modulated and adaptively-refined fusion process, ASR effectively produces feature maps that exhibit both semantic coherence within nuclei and sharp, precise boundaries, which are essential for accurate instance segmentation performance. + +# 3.3. Iterative Contrast-Weaving Module (ICW) + +Complementing the per-scale feature enhancement of the ASR module, the ICW is designed to focus on contextual relationships and progressive refinement. ICW transcends simple feature fusion by employing a novel "weaving" process. This process dynamically integrates local contrastive attention and adaptive gating within a feedback loop to iteratively enhance feature discriminability, a critical factor for precise nuclei segmentation, especially in the presence of ambiguous boundaries and class imbalances. Specifically, it mitigates the foreground-background imbalance by employing a contrastive attention mechanism that amplifies nuclei-specific features while suppressing background noise, thus enhancing the nuclei-background distinction. The architecture of the ICW module is detailed in Figure 3(b). + +ICW operates through an iterative refinement process. Each iteration $t$ refines the feature representation by taking as input frequency-adapted features, $\mathbf{F}_{\mathrm{ASR}}^{(t - 1)}\in$ $\mathbb{R}^{C\times H\times W}$ , from the ASR module, and semantic context maps, $\mathbf{S}^{\mathrm{f}},\mathbf{S}^{\mathrm{b}},\mathbf{S}^{\mathrm{u}}\in \mathbb{R}^{H\times W}$ . These context maps, representing foreground, background, and uncertain regions respectively. The initial input, $\mathbf{F}_{\mathrm{ASR}}^{(0)}$ , is the output $\mathbf{F}_{\mathrm{out}}$ from the ASR module. + +Each ICW iteration performs three core operations, conceptualized as a refinement step towards improved instance separation. First, a dense block, DenseBlock(\cdot), enhances inter-layer feature propagation, yielding an intermediate feature map $\mathbf{F}_{\mathrm{dense}}^{(t)}$ : + +$$ +\mathbf {F} _ {\text {d e n s e}} ^ {(t)} = \operatorname {D e n s e B l o c k} \left(\mathbf {F} _ {\mathrm {A S R}} ^ {(t - 1)}\right) \tag {5} +$$ + +This dense connectivity facilitates feature reuse and mitigates the vanishing gradient problem, promoting robust feature learning. + +Next, a contrastive attention (CA) mechanism selectively attends to features based on local context within a $K \times K$ neighborhood $\mathcal{N}(x,y)$ around each spatial location $(x,y)$ . Attention weights, $\alpha_{(x,y)}^{(t)}$ , are computed using downsampled foreground $(\mathbf{S}^{\mathrm{f}\downarrow})$ and background $(\mathbf{S}^{\mathrm{b}\downarrow})$ context maps, thus incorporating a broader contextual window to guide attention. The context-aware representation, $\mathbf{V}_{\mathrm{CA}}^{(t)}(x,y)$ , is then formulated as: + +$$ +\mathbf {V} _ {\mathrm {C A}} ^ {(t)} (x, y) = \sum_ {(i, j) \in \mathcal {N} (x, y)} \alpha^ {(t)} (x, y, i, j) \cdot \mathbf {V} ^ {(t)} (i, j) \tag {6} +$$ + +where $\mathbf{V}^{(t)}$ represents value features obtained through a linear transformation: + +$$ +\mathbf {V} ^ {(t)} = \mathbf {W} _ {v} \mathbf {F} _ {\text {d e n s e}} ^ {(t)} \tag {7} +$$ + +and the attention weight $\alpha_{\mathrm{(x,y)}}^{(t)}$ is defined as: + +$$ +\alpha_ {(\mathrm {x}, \mathrm {y})} ^ {(t)} = \frac {\exp \left(\mathbf {w} _ {f} ^ {T} \mathbf {S} _ {(\mathrm {x} , \mathrm {y})} ^ {\mathrm {f} \downarrow} - \mathbf {w} _ {b} ^ {T} \mathbf {S} _ {(\mathrm {x} , \mathrm {y})} ^ {\mathrm {b} \downarrow}\right)}{\sum_ {(m , n) \in \mathcal {N} (x , y)} \exp \left(\mathbf {w} _ {f} ^ {T} \mathbf {S} _ {(\mathrm {m} , \mathrm {n})} ^ {\mathrm {f} \downarrow} - \mathbf {w} _ {b} ^ {T} \mathbf {S} _ {(\mathrm {m} , \mathrm {n})} ^ {\mathrm {b} \downarrow}\right)} \tag {8} +$$ + +with learnable weight vectors $\mathbf{w}_f$ and $\mathbf{w}_b$ . This local contrastive learning mechanism, distinct from global selfattention, drives the network to learn features that are discriminative for the foreground (nuclei) relative to the background. It effectively amplifies features that contribute to instance separation while suppressing irrelevant ones, resulting in a contrast-enhanced map, $\mathbf{F}_{\mathrm{CA}}^{(t)}$ . + +Subsequently, an adaptive gated linear unit (AGLU) modulates the influence of $\mathbf{F}_{\mathrm{CA}}^{(t)}$ , functioning as a dynamic gate to control information flow: + +$$ +\mathbf {F} _ {\text {g a t e d}} ^ {(t)} = \mathbf {F} _ {\mathrm {C A}} ^ {(t)} \odot \sigma (\kappa \cdot \mathbf {F} _ {\mathrm {C A}} ^ {(t)} + \lambda) \tag {9} +$$ + +where $\odot$ denotes element-wise multiplication, $\sigma$ is the sigmoid function, and $\kappa$ and $\lambda$ are learnable parameters. This adaptivity ensures that feature modulation is neither too aggressive (suppressing useful information) nor too passive (allowing noise to propagate), dynamically balancing contrast enhancement with feature fidelity across diverse image characteristics and learning stages. + +Finally, a residual connection integrates the gated features $\mathbf{F}_{\mathrm{gated}}^{(t)}$ with the input features $\mathbf{F}_{\mathrm{ASR}}^{(t - 1)}$ , generating the refined features $\mathbf{F}_{\mathrm{ASR}}^{(t)}$ for the subsequent iteration: + +$$ +\mathbf {F} _ {\mathrm {A S R}} ^ {(t)} = \mathbf {F} _ {\mathrm {A S R}} ^ {(t - 1)} + \mathbf {F} _ {\text {g a t e d}} ^ {(t)} \tag {10} +$$ + +The iteratively refined features from the final ICW iteration are then utilized to derive the ultimate segmentation output. + +![](images/f9ca4d0fc5b29dd2813568d6e6669906dcb3f2359f5d72e760c6200ebfad5648.jpg) +(a) + +![](images/20fdfc8b1a3814d960feb5482f42f26570049695b1e58362f15c5d271e541c4a.jpg) +(b) + +![](images/edeb9b3c438be3bcaadffdf2241380742a872028a7fca2994c282e31d728a328.jpg) +(c) + +![](images/9c8e6b98100fe6dd71b3260067157abab213be44ffb6f5b96ee28e55c92738ec.jpg) +(d) + +![](images/4624a461a343dfb80723d73929c69590da078df95e40dae92fe1fa310dedec74.jpg) +(e) + +![](images/d0d05a303ab1fb63d94f9463e0009055c414f8ecd31ede2a16b8cc1a03ae967d.jpg) +(f) + +![](images/dadcd1dd0bbc0cbab7969d8c6c177d4ece8c6b5a18dda64db049c72ae98d2624.jpg) +(g) +Figure 4. Illustration of the uncertainty map generation process. (a) Ground truth mask; (b) Grayscale ground truth mask, delineating nuclei interior (white), boundary (gray), and background (black) regions; (c) Gradient magnitude map; (d) Centroid plot; (e) Mask erosion, marking nuclei interior (blue); (f) Mask dilation, expanding nuclei region (red); (g) Uncertainty behavior map; (h) Final uncertainty map. + +![](images/fc0bf60981a499c7d3b8aef23d57e8da229f8434d269b975faa1ecb5d94d1cb3.jpg) +(h) + +Through this iterative "weaving" process—integrating dense propagation, local contrastive attention, and adaptive gating within a closed-loop system with ASR. WeaveSeg progressively resolves ambiguities, enhances boundaries, and achieves improved segmentation accuracy. Each iteration facilitates a refinement of the feature representation, iteratively optimizing towards improved instance separation in the feature space. This iterative refinement strategy, leveraging contrastive learning and adaptive gating, holds potential for broader application in other image segmentation tasks demanding robust handling of complex contextual dependencies and feature ambiguities. + +# 3.4. Loss Function + +WeaveSeg's training process is guided by a comprehensive multi-task loss function, $\mathcal{L}_{\mathrm{Total}}$ , meticulously designed to promote accurate instance segmentation, semantic consistency, and robust handling of ambiguous regions inherent in histopathology images. This loss function is composed of several distinct terms: binary cross-entropy (BCE) losses for the primary segmentation mask $(\mathcal{L}_{\mathrm{mask}})$ and auxiliary foreground, background predictions $(\mathcal{L}_{\mathrm{fg}},\mathcal{L}_{\mathrm{bg}})$ , a trilateral decomposition-inspired loss $(\mathcal{L}_{\mathrm{tri}})$ , and a novel uncertainty loss $(\mathcal{L}_{\mathrm{uncert}})$ . The $\mathcal{L}_{\mathrm{uncert}}$ term is specifically introduced to explicitly address ambiguous regions, particularly prevalent at nuclei boundaries and overlaps. + +The uncertainty loss, $\mathcal{L}_{\mathrm{uncert}}$ , serves as a key component for robustly handling ambiguous regions. It operates on the predicted uncertainty map, $\mathbf{S}^{\mathrm{u}} \in \mathbb{R}^{H \times W}$ , with the objective of encouraging the model to express high uncertainty in areas where the ground truth segmentation is inherently ambiguous (typically boundaries and overlaps) and conversely, low uncertainty in regions with clear foreground (nuclei interiors) or background. To facilitate this, a target uncertainty map, $\mathbf{U}^* \in \mathbb{R}^{H \times W}$ , is generated, with its generation + +![](images/6c2fb6281d605a86efa9a8b47109c7451b558b4834f9fd9677fc33498c3a2d9a.jpg) +(a) + +![](images/6e42902e07372045d401c17982030daef37f141d2fb68adffba21677ecf297b0.jpg) +(b) + +![](images/f3c847757f744bf3b37ec75929643afa3ef6e724f8d16386f8be3d6384d950cf.jpg) +Figure 5. Qualitative comparison of edge prediction performance. (a) Input image; (b) Baseline model; (c) Baseline + ICW, showing improved boundary delineation; (d) Baseline + ICW + ASR (WeaveSeg), demonstrating the accurate edge prediction. + +![](images/bc46315bf5fd89b82be99fdbb0ae8a0094a4aefaeb533c55905202e625e9360a.jpg) +(d) + +process illustrated in Figure 4. The uncertainty loss is then defined as the mean squared error (MSE) between the predicted uncertainty map, $\mathbf{S}^{\mathrm{u}}$ , and the target uncertainty map, $\mathbf{U}^{*}$ : + +$$ +\mathcal {L} _ {\text {u n c e r t}} = \frac {1}{N} \sum_ {i = 1} ^ {N} \left(\mathbf {S} ^ {\mathrm {u} (i)} - \mathbf {U} ^ {* (i)}\right) ^ {2} \tag {11} +$$ + +where $N$ denotes the total number of pixels, and $(i)$ indexes a specific pixel location. This loss term effectively encourages the model to refrain from making over-confident, yet potentially erroneous, predictions in ambiguous regions. Consequently, it guides the ICW module to prioritize the learning of discriminative features while explicitly acknowledging and accommodating boundary uncertainty. Further details regarding the target uncertainty map generation are provided in the supplementary material. + +The final multi-task loss, $\mathcal{L}_{\mathrm{Total}}$ , is formulated as a weighted sum of these individual loss terms: + +$$ +\begin{array}{l} \mathcal {L} _ {\text {T o t a l}} = \lambda_ {\text {m a s k}} \mathcal {L} _ {\text {m a s k}} + \lambda_ {\mathrm {f g}} \mathcal {L} _ {\mathrm {f g}} + \lambda_ {\mathrm {b g}} \mathcal {L} _ {\mathrm {b g}} \\ + \lambda_ {\text {t r i}} \mathcal {L} _ {\text {t r i}} + \lambda_ {\text {u n c e r t}} \mathcal {L} _ {\text {u n c e r t}} \tag {12} \\ \end{array} +$$ + +The loss weights $(\lambda_{\mathrm{mask}} = 1.0, \lambda_{\mathrm{fg}} = 0.5, \lambda_{\mathrm{bg}} = 0.5, \lambda_{\mathrm{tri}} = 0.3,$ and $\lambda_{\mathrm{uncert}} = 1.0)$ , were empirically determined via validation tuning. + +# 4. Experiments + +# 4.1. Datasets and Evaluation Metrics + +We rigorously evaluated WeaveSeg on three publicly available histopathology benchmarks: CoNSeP [8], MoNuSeg [17] and the CPM17 [36] dataset. These datasets feature diverse cell types and challenging conditions, such as significant nuclei overlap and staining variations, making them suitable for comprehensive evaluation. Following standard protocols, we used the official data splits for training and testing. Segmentation performance was quantitatively assessed using four widely-accepted metrics: F1-score (F1), Dice coefficient (Dice), Hausdorff distance (HausD), Panoptic Quality (PQ) score and Aggregated Jaccard Index (AJI) [17, 26]. + +
ModelsF1DiceHausDAJI
Baseline0.7250.82827.3810.515
w/o ICW0.7380.83526.9330.529
w/o ASR0.7460.84426.3540.547
Full WeaveSeg0.7590.85225.8350.563
+ +Table 1. Ablation study on the CoNSeP dataset, evaluating the impact of different components. "w/o" denotes removing the corresponding component. + +
Iter.F1DiceHausDAJI
00.8590.8237.1410.582
10.8650.8306.7540.622
20.8720.8386.3120.637
30.8740.8416.0420.648
40.8680.8366.4530.635
+ +Table 2. Performance comparison across different iterations of the ICW module on the MoNuSeg dataset. The baseline refers to the initial framework without ICW or ASR (Iter. = 0). + +# 4.2. Implementation Details + +WeaveSeg was implemented using PyTorch 1.8 [28] and CUDA 10.1. Training was conducted on a workstation with dual NVIDIA RTX 2080Ti GPUs. Network weights were initialized using ImageNet pre-trained weights [6], leveraging transfer learning methodologies [43]. The Adam optimizer [16] was employed with an initial learning rate of $10^{-4}$ , which was reduced by a factor of 0.1 every 25 epochs, reaching a final learning rate of $10^{-7}$ . A batch size of 32 was utilized, and training was performed for 100 epochs. The network processed $256 \times 256$ RGB image patches, normalized to the range [0, 1]. Standard data augmentation techniques, including random color transformations, horizontal flips, rotations, elastic transformations, and cropping, were applied throughout training. As depicted in Figure 8, the $\kappa$ and $\lambda$ parameters within the AGLU module were adaptively adjusted throughout the training process. + +# 4.3. Ablation Studies + +To rigorously validate the individual contributions of the ASR and ICW modules, ablation studies were performed on the CoNSeP validation set, utilizing a U-Net variant as the baseline model. Table 1 presents the ablation study results. The complete WeaveSeg model consistently outperforms both the baseline and the ablation variants (ablated of ASR or ICW modules) across all evaluation metrics, unequivocally demonstrating the synergistic effect of these two proposed modules. + +As detailed in Table 1, the removal of the ICW module leads to a more pronounced performance degradation compared to the removal of the ASR module. Specifically, the Dice coefficient decreases by 0.017 without ICW, whereas it decreases by 0.008 without ASR. A similar trend + +![](images/bf0c73cfec7db174f157f1c6e6f1d7c54b9d71a805a27ddf6cdf9514347312e7.jpg) +Figure 6. Visualization of semantic segmentation results with different methods. Different colors represent different types. + +
LmaskLfg & LbgLtriLuncertMoNuSegCoNSeP
F1DiceHausDAJIF1DiceHausDAJI
✓0.8610.8216.9620.5880.7330.81630.5270.505
✓✓0.8680.8286.7940.6030.7480.83428.7590.536
✓✓✓0.8770.8326.7510.6240.7640.83927.6830.542
✓✓✓✓0.8740.8366.4350.6440.7560.84426.3540.547
+ +Table 3. Ablation study on the impact of different loss function components within the ICW module. + +is observed for the instance-level AJI metric. This underscores the critical role of the iterative contrast enhancement provided by the ICW module in achieving accurate segmentation and instance discrimination, particularly in challenging scenarios involving overlapping nuclei and ambiguous boundaries. The ICW module's capacity for refining features based on contextual relationships and addressing class imbalance appears crucial for attaining optimal performance. Figure 5 visually corroborates the enhanced edge prediction performance resulting from the progressive integration of ICW and ASR. + +Furthermore, we investigated the individual impact of different loss components within the ICW module itself (Table 3), isolating their contributions by excluding the ASR module. Progressive incorporation of the foreground/background losses $(\mathcal{L}_{\mathrm{fg}}\& \mathcal{L}_{\mathrm{bg}})$ , the trilateral decomposition loss $(\mathcal{L}_{\mathrm{tri}})$ , and the novel uncertainty loss $(\mathcal{L}_{\mathrm{uncert}})$ ,in addition to the baseline mask loss $(\mathcal{L}_{\mathrm{mask}})$ -reveals a consistent performance improvement trend. On MoNuSeg, the incorporation of $\mathcal{L}_{\mathrm{fg}}$ and $\mathcal{L}_{\mathrm{bg}}$ yields improvements in Dice and AJI. Incorporating $\mathcal{L}_{\mathrm{tri}}$ further enhances F1-score. Cru + +cially, adding $\mathcal{L}_{\mathrm{uncert}}$ results in the highest Dice, HausD, and AJI values, albeit with a slight decrease in F1-score. This indicates that $\mathcal{L}_{\mathrm{uncert}}$ effectively prioritizes precise boundary delineation and instance separation, which is paramount for accurate nuclei segmentation. Consistent trends are observed on CoNSeP, with progressive enhancements in Dice, HausD, and AJI corresponding to the inclusion of each successive loss component. Table 2 performance peaks at the 3rd iteration, demonstrating the effectiveness of the iterative refinement process. This empirically validates our choice of 3 iterations as the optimal configuration for balancing performance and computational efficiency. + +# 4.4. Comparative Experiments + +We conducted a comprehensive comparison of WeaveSeg against several recent state-of-the-art methods, including Hover-Net [8], CD-Net [12], Mudslide [38] and CellVIT-B [14]. As presented in Table 4, our proposed WeaveSeg demonstrates superior performance across all three benchmark datasets. On MoNuSeg, WeaveSeg achieves an AJI of 0.648 and a PQ of 0.644, outperforming all competitors. + +![](images/5275b3b4e32c18a4abbdac296c332813dedd970a8f98a8c821de94c9a41ebbba.jpg) +Figure 7. Dice score comparison on MoNuSeg and CoNSeP datasets. WeaveSeg (Our Net) achieves superior performance compared to other methods. + +
MethodMoNuSegCoNSePCMP17
AJIPQAJIPQAJIPQ
U-Net [31]0.5970.5810.5010.4920.6550.578
Mask-RCNN [11]0.5460.5090.4940.4600.684-
DCAN [3]0.5250.4920.289-0.5610.545
DIST [25]0.5590.4430.5020.3980.6160.504
Hover-Net [8]0.6180.5970.5360.5320.7050.697
CD-Net [12]0.6370.6190.5450.5240.729-
TopoSeg [24]0.6410.625--0.7430.705
MedSAM [22]0.6220.6180.5170.5050.7070.694
CellViT-B [14]0.6390.6540.5470.5390.7370.713
NucSeg-B [32]0.6340.6260.5420.5330.7310.726
Mudslide [38]0.6420.6370.5470.5420.7400.714
Our Net0.6480.6440.5630.5480.7510.733
+ +Table 4. Performance comparisons on MoNuSeg, CoNSEP and CPM17. AJI and PQ (%) are reported. '-' indicates value not available in original SOTA paper. + +Similarly, on CoNSeP, it sets a new SOTA with an AJI of 0.563 and a PQ of 0.548. Notably, our method also excels on the CPM17 dataset, achieving the highest AJI (0.751) and PQ (0.733) scores. These results consistently underscore WeaveSeg's effectiveness and its robust capability to handle diverse and challenging nuclei segmentation scenarios. The qualitative results in Figure 6 further visualize our model's proficiency in accurately segmenting challenging instances with overlapping and densely packed nuclei compared to other methods. + +Further details on ablation studies and comparative experiments can be found in the Supplementary Material. + +# 4.5. Limitation + +Although WeaveSeg effectively addresses numerous scenarios featuring ambiguous boundaries and overlapping nuclei, it still encounters limitations in particularly challenging instances involving severely clustered nuclei with nearly + +![](images/b8af96cbe3e24cc9b9ca3d64aa87496e688acc456f6b66ed8ea0ea75f2a3debf.jpg) +Figure 8. Visualization of hyperparameter adaptation during training. The parameters $\kappa$ and $\lambda$ in the AGLU module are dynamically adjusted to optimize feature modulation. + +![](images/c85f2ebf2597b05a28b3e4be7eed6a7786335ccb6fb1aac2151f76b527345e5c.jpg) +(a) + +![](images/2f22d68f654c0fa47c98f1d820d03e8ce3ea7ec8ff2f39a9521e37a323ccaa19.jpg) +(b) + +![](images/fe0de5342051030c84f0bf403830720564a35b1bfa2c32bdf4cb42300110b69a.jpg) +(c) +Figure 9. Illustration of challenging cases where Weaveseg shows limitations. (a) Ground truth and (b) failure case on MoNuSeg, where it has limitation; (c) ground truth, and (d) failure case on CoNSEP. + +![](images/451ab23b14b8508e8ff2c1fd5bd654aedb0bd6077aab1bd8703034bfce2e66bc.jpg) +(d) + +indistinguishable boundaries, often further complicated by staining variations. Figure 9 illustrates representative failure cases. Furthermore, while our method has been validated on three diverse datasets, its generalization capability could be further strengthened by evaluation on a broader spectrum of tissue types and staining protocols. Future research will focus on integrating more sophisticated context modeling mechanisms to handle severe clustering. + +# 5. Conclusion + +In summary, WeaveSeg introduces a novel framework for nuclei instance segmentation in histopathology images, effectively addressing persistent challenges of ambiguous boundaries, complex morphologies, and nuclei overlap. Achieving state-of-the-art performance, WeaveSeg synergistically integrates frequency-aware feature adaptation via the ASR module, iterative contrast enhancement through the ICW module, and a dedicated uncertainty loss function. Empirical validation on MoNuSeg and CoNSeP datasets, substantiated by ablation studies, demonstrates significant performance gains over existing methodologies, underscoring the contribution of iterative refinement. While WeaveSeg exhibits robust performance, future efforts will enhance resilience to staining variations in clustered nuclei. + +# Acknowledgments + +# References + +[1] Konstantinos P. Alexandridis, Jianfei Deng, Anh Nguyen, Ser-Nam Lim, and Timothy M. Hospedales. Adaptive parametric activation. In European Conference on Computer Vision, pages 455-476, Cham, 2024. Springer Nature Switzerland. 2 +[2] Krishna Chaitanya, Neeraj Karani, Sriram Konjeti, Ardeshir Katouzian, and Deblood Sheet. Contrastive learning of global and local features for medical image segmentation with limited annotations. In Medical Image Computing and Computer Assisted Intervention – MICCAI 2020, pages 586–596. Springer, 2020. 3 +[3] Hao Chen, Xiaojuan Qi, Lequan Yu, and Pheng-Ann Heng. Dcan: Deep contour-aware networks for object instance segmentation from histology images. Medical Image Analysis, 36:135-146, 2016. 1, 2, 8 +[4] Lei Chen, Yuxi Fu, Li Gu, et al. Frequency-aware feature fusion for dense image prediction. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 2 +[5] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey E. Hinton. A simple framework for contrastive learning of visual representations. In International Conference on Machine Learning, ICML 2020, pages 1597-1607. PMLR, 2020. 2 +[6] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 6 +[7] Prerna Gehlot, Dinesh Kumar Jha, Nidhi Agrawal, Kshitij Gupta, Sumit Kumar, Poonam Johri, and Poonam Tiwari. Ednfc-net: an encoder decoder-based nested feature cascade network for biomedical image segmentation. Multimedia Tools and Applications, pages 1-25, 2024. Online First. 2 +[8] Simon Graham, Quoc Dang Vu, Syed EA Raza, Atif Azam, Yee Wah Tsang, Jin Tae Kwak, and Nasir Rajpoot. Hovernet: Simultaneous segmentation and classification of nuclei in multi-tissue histology images. Medical image analysis, 58:101563, 2019. 1, 2, 6, 7, 8 +[9] Ming-Hui Guo, Cheng-Ze Lu, Qing Hou, Zhen-Ning Liu, Ming-Ming Cheng, and Shi-Min Hu. Adaptive frequency filters as efficient global token mixers. In CVPR, pages 11645-11654, 2023. 2 +[10] Raia Hadsell, Sumit Chopra, and Yann LeCun. Dimensionality reduction by learning an invariant mapping. In CVPR (2), pages 1735-1742. IEEE Computer Society, 2006. 2 +[11] Kaiming He, Georgia Gkioxari, Piotr Dollar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 1, 8 + +This work was supported partly by the National Natural Science Foundation of China (No.62273241), and the Natural Science Foundation of Guangdong Province, China (No.2024A1515011946). +[12] Xin He, Sheng-Kun Zhou, Yi Zhou, Zheng Chen, Yefeng Wang, Pheng-Ann Heng, Danny Z. Chen, and Jian Wu. Cdnet: Centripetal directional network for nuclear instance segmentation. In 2021 IEEE 18th International Symposium on Biomedical Imaging (ISBI), pages 1255–1259. IEEE, 2021. 2, 7, 8 +[13] Xiao He, Sheng Zhao, Hang Zhang, Yutong Chen, Bo Zhou, Pheng-Ann Heng, and Sheng-Kun Zhou. HAN-Net: a hybrid attention nested U-shaped network for pathological image segmentation. IEEE Transactions on Medical Imaging, 2024. Early Access. 2 +[14] F. Hörst, M. Rempe, L. Heine, and et al. Cellvit: Vision transformers for precise cell segmentation and classification. Medical Image Analysis, 94:103143, 2024. 7, 8 +[15] Lijing Hou, Ayush Agarwal, Dimitris Samaras, Tahsin M Kurc, Rajarsi R Gupta, and Joel H Saltz. Unsupervised histopathology image synthesis. IEEE transactions on medical imaging, 39(2):382-394, 2020. 1 +[16] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6 +[17] Nishant Kumar, Rakesh Verma, Sachin Sharma, Saurabh Bhargava, Ali Vahadane, and Anant Sethi. A dataset and a technique for generalized nuclear segmentation for computational pathology. IEEE Transactions on Medical Imaging, 36(7):1550-1560, 2017. 6 +[18] Yann LeCun, Yoshua Bengio, and Geoffrey Hinton. Deep learning. nature, 521(7553):436-444, 2015. 1, 2 +[19] Meng Lei, Huisi Wu, Xiaohua Lv, et al. Condseg: A general medical image segmentation framework via contrast-driven feature enhancement. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 4571-4579, 2025. 3 +[20] Shangchen Li, Xinyi Shi, Qiang Ye, Xin Yang, Xun Lin, Xiaopeng Xie, Wenming Cao, Mingming Yang, Reynold Cheng, and Wangmeng Zuo. Feature modulation transformer: Cross-refinement of global representation via high-frequency prior for image super-resolution. In The Twelfth International Conference on Learning Representations, ICLR 2024, 2024. 2 +[21] Dongnan Liu, Donghao Zhang, Yang Song, Heng Huang, and Weidong Cai. Panoptic feature fusion net: A novel instance segmentation paradigm for biomedical and biological images. IEEE Transactions on Image Processing, 30:2045-2059, 2021. 1 +[22] Jican Ma, Yang He, Fan Li, Li Wang, Hao Wang, Kan Li, Weixin Liu, Jie Zhang, Li Xu, Xuan Han, et al. Segment anything in medical images. Nature Communications, 15(1): 654, 2024. 8 +[23] Fausto Milletari, Nassir Navab, and Seyed-Ahmad Ahmadi. V-net: Fully convolutional neural networks for volumetric medical image segmentation. In 2016 fourth international conference on 3D vision (3DV), pages 565-571. IEEE, 2016. 1 +[24] Anna Mosinska, Florian Dubost, Jonathan Saltz, Peter J. Wild, Katja A. Frei, Florian Wempe, Maximilian Baust, Nasir Rajpoot, and Dinesh Jha. Toposeg: Topology-aware nuclear instance segmentation. In Medical Image Computing + +and Computer Assisted Intervention - MICCAI 2023, pages 303-313. Springer, 2023. 2, 8 +[25] Peter Naylor, Mathieu Laé, Frédéric Reyal, and Thomas Walter. Segmentation of nuclei in histopathology images by deep regression of the distance map. IEEE transactions on medical imaging, 38(2):448-459, 2018. 1, 2, 8 +[26] Davy Neven, Bert De Brabandere, Marc Proesmans, and Luc Van Gool. Instance segmentation by jointly optimizing spatial embeddings and clustering bandwidth. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8837-8845, 2019. 6 +[27] Ozan Oktay, Joachim Schlemper, Loic Le Folgoc, Matthew Lee, Martin Heinrich, Kazunari Misawa, Kensaku Mori, Steven McDonagh, Nathalie Y. Hammerla, Bernhard Kainz, and et al. Attention u-net: Learning where to look for the pancreas. Medical Image Analysis, 61:101640, 2020. arXiv preprint arXiv:1804.03999. 2 +[28] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killea, Zeming Lin, Yuqing Yang, Gregory DeVito, Zachary Raison, Alykhan Tejani, Sasank Chilimuntala, Benoit Prabhat, et al. Pytorch: An imperative style, high-performance deep learning library. In Advances in neural information processing systems, pages 8026-8037, 2019. 6 +[29] H. Pinckaers, W. Bulten, J. van der Laak, and G. Litjens. Detection of prostate cancer in whole-slide images through end-to-end training with image-level labels. IEEE Transactions on Medical Imaging, 40(7):1817-1826, 2021. 1 +[30] Zheng Qin, Pengyi Zhang, Feifei Wu, and Xiao Li. Fcanet: Frequency channel attention networks. In ICCV, pages 783-792, 2021. 2 +[31] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18, pages 234-241. Springer International Publishing, 2015. 1, 2, 8 +[32] Z. Shui, Y. Zhang, K. Yao, and et al. Unleashing the power of prompt-driven nucleus instance segmentation. In European Conference on Computer Vision, pages 288-304. Springer Nature Switzerland, Cham, 2024. 8 +[33] Kriangsak Sirinukunwattana, Syed EA Raza, Weilin Ya, Quanfeng Ba, Chi-Liang Chen, Fan Xing, Josien PW Pluim, Pheng-Ann Heng, Ursula Sanchez, Elodie Bruni, et al. Gland segmentation in colon histology images: the glas challenge contest. Medical image analysis, 35:489-502, 2017. 1 +[34] Eric Upschulte, Stefan Harmeling, Katrin Amunts, and Timo Dickscheid. Uncertainty-aware contour proposal networks for cell segmentation in multi-modality high-resolution microscopy images. In *Competitions in Neural Information Processing Systems*, pages 1-12. PMLR, 2023. 1 +[35] Mitko Veta, Yi-Ju Heng, Nikolaos Stathonikos, Babak Ehteshami Bejnordi, Florentin Beca, Thomas Wollmann, Kay Rohr, Muhammad Arsalan Shah, Dayong Wang, Michel Rousson, et al. Predicting breast tumor proliferation from whole-slide images: the tupac16 challenge. Medical image analysis, 54:111-121, 2019. 1 + +[36] Q. D. Vu, S. Graham, T. Kurc, and et al. Methods for segmentation and classification of digital microscopy tissue images. Frontiers in bioengineering and biotechnology, 7:53, 2019. 6 +[37] Hong Wang, Huisi Wu, and Jing Qin. Incremental nuclei segmentation from histopathological images via future-class awareness and compatibility-inspired distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11408-11417, 2024. 1 +[38] J. Wang. Mudslide: A universal nuclear instance segmentation method. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11673-11682, 2024. 2, 7, 8 +[39] Zikun Wang, Liang Zheng, Yichen Li, Shenglin Lu, Bowen Zhang, Zongyuan Zhou, Ruimao Zhang, Xiaopeng Li, Jian Yang, Xiaoguang Li, and et al. Self-supervised nuclei segmentation via local and global contrastive learning. arXiv preprint arXiv:2403.08474, 2024. 3 +[40] Huisi Wu, Zhaoze Wang, Youyi Song, Lin Yang, and Jing Qin. Cross-patch dense contrastive learning for semisupervised segmentation of cellular nuclei in histopathologic images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11666-11675, 2022. 1 +[41] Ziyuan Xiang, Wen Luo, Zhiqiang Li, and Jun Su. Bionet: Bi-directional o-shape network for nuclei instance segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 612-621. Springer, 2024. 2 +[42] Fan Xing and Lin Yang. Robust nucleus/cell detection and segmentation in digital pathology and microscopy images: A comprehensive review. IEEE reviews in biomedical engineering, 9:234-263, 2016. 1 +[43] Jason Yosinski, Jeff Clune, Yoshua Bengio, and Hod Lipson. How transferable are features in deep neural networks? In Advances in neural information processing systems, pages 3320-3328, 2014. 6 +[44] Bo Zhou, Hao Chen, Jian Yang, Wei Wang, Xiaojuan Qi, and Pheng-Ann Heng. Cia-net: Robust nuclei instance segmentation with contour-aware information aggregation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 326–336. Springer, 2020. 1, 2 +[45] Zongwei Zhou, Md Mahfuzur Rahman Siddiquee, Nima Tajbakhsh, and Jianming Liang. Unet++: A nested u-net architecture for medical image segmentation. In Deep learning in medical image analysis and multimodal learning for clinical decision support: 4th international workshop, DLMIA 2018, and 8th international workshop, ML-CDS 2018, held in conjunction with MICCAI 2018, Granada, Spain, September 20, 2018, proceedings 4, pages 3-11. Springer International Publishing, 2018. 1, 2 \ No newline at end of file diff --git a/ICCV/2025/WeaveSeg_ Iterative Contrast-weaving and Spectral Feature-refining for Nuclei Instance Segmentation/images.zip b/ICCV/2025/WeaveSeg_ Iterative Contrast-weaving and Spectral Feature-refining for Nuclei Instance Segmentation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..16a8bba9efdc2c499ca3388b456be9c6229013ef --- /dev/null +++ b/ICCV/2025/WeaveSeg_ Iterative Contrast-weaving and Spectral Feature-refining for Nuclei Instance Segmentation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:143017029a59dc1fafc921014213e9786f12115de0b6dbacbe8d4441c4fb7cc0 +size 877692 diff --git a/ICCV/2025/WeaveSeg_ Iterative Contrast-weaving and Spectral Feature-refining for Nuclei Instance Segmentation/layout.json b/ICCV/2025/WeaveSeg_ Iterative Contrast-weaving and Spectral Feature-refining for Nuclei Instance Segmentation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..db726dbc3c158fa37ef879efe85f6eccbb5d35c6 --- /dev/null +++ b/ICCV/2025/WeaveSeg_ Iterative Contrast-weaving and Spectral Feature-refining for Nuclei Instance Segmentation/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7d83a04254ee7c73e2bf258ac64d649ba5a86b216517fbeb8ff835389fc716fb +size 423240 diff --git a/ICCV/2025/Web Artifact Attacks Disrupt Vision Language Models/793c8111-1366-4e5c-94bf-770223d04fae_content_list.json b/ICCV/2025/Web Artifact Attacks Disrupt Vision Language Models/793c8111-1366-4e5c-94bf-770223d04fae_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..c6a65e54d8bf86f9c55117ea71ddb8c15738ac4d --- /dev/null +++ b/ICCV/2025/Web Artifact Attacks Disrupt Vision Language Models/793c8111-1366-4e5c-94bf-770223d04fae_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a57aeda6e245218d7e6d48863b37514b703b777cacbbd4f91dbb6c9de1393df5 +size 76263 diff --git a/ICCV/2025/Web Artifact Attacks Disrupt Vision Language Models/793c8111-1366-4e5c-94bf-770223d04fae_model.json b/ICCV/2025/Web Artifact Attacks Disrupt Vision Language Models/793c8111-1366-4e5c-94bf-770223d04fae_model.json new file mode 100644 index 0000000000000000000000000000000000000000..759822386cf259441142eb66b166dfb0c518ba1e --- /dev/null +++ b/ICCV/2025/Web Artifact Attacks Disrupt Vision Language Models/793c8111-1366-4e5c-94bf-770223d04fae_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f6f668ca2552e7e00496c65b5b5eb7b8653479a0a13ff253f323ed001f89368b +size 92131 diff --git a/ICCV/2025/Web Artifact Attacks Disrupt Vision Language Models/793c8111-1366-4e5c-94bf-770223d04fae_origin.pdf b/ICCV/2025/Web Artifact Attacks Disrupt Vision Language Models/793c8111-1366-4e5c-94bf-770223d04fae_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..a426782bcaa488737a1094c731bf62e8455d3f93 --- /dev/null +++ b/ICCV/2025/Web Artifact Attacks Disrupt Vision Language Models/793c8111-1366-4e5c-94bf-770223d04fae_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bfa55fb9038d45ef41d558d5c4f5740184f5ccc6321c152244030e6947a7b401 +size 6129422 diff --git a/ICCV/2025/Web Artifact Attacks Disrupt Vision Language Models/full.md b/ICCV/2025/Web Artifact Attacks Disrupt Vision Language Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..0282eb028be786864cd36cbf42c92231fd6d764f --- /dev/null +++ b/ICCV/2025/Web Artifact Attacks Disrupt Vision Language Models/full.md @@ -0,0 +1,297 @@ +# Web Artifact Attacks Disrupt Vision Language Models + +Maan Qraitem, Piotr Teterwak, Kate Saenko, Bryan A. Plummer +Boston University + +{mqraitem, piotrt, saenko, bplum}@bu.edu + +# Abstract + +Vision-language models (VLMs) (e.g., CLIP, LLaVA) are trained on large-scale, lightly curated web datasets, leading them to learn unintended correlations between semantic concepts and unrelated visual signals. These associations degrade model accuracy by causing predictions to rely on incidental patterns rather than genuine visual understanding. Prior work has weaponized these correlations as an attack vector to manipulate model predictions, such as inserting a deceiving class text onto the image in a "typographic" attack. These attacks succeed due to VLMs' text-heavy bias—a result of captions that echo visible words rather than describing content. However, this attack has focused solely on text that matches the target class exactly, overlooking a broader range of correlations, including non-matching text and graphical symbols, which arise from the abundance of branding content in web-scale data. To address this gap, we introduce "artifact-based" attacks: a novel class of manipulations that mislead models using both non-matching text and graphical elements. Unlike typographic attacks, these artifacts are not predefined, making them simultaneously harder to defend against and more challenging to find. We address this by framing artifact attacks as a search problem and demonstrate their effectiveness across five datasets, with some artifacts reinforcing each other to reach $100\%$ attack success rates. These attacks transfer across models with up to $90\%$ effectiveness, making it possible to attack unseen models. To defend against these attacks, we extend prior work's artifact aware prompting to the graphical setting. We see a moderate reduction of success rates of up to $15\%$ relative to standard prompts, suggesting a promising direction for enhancing model robustness. Code: https://github.com/mqraitem/Web-Artifact-Attacks + +# 1. Introduction + +Vision-language models (VLMs), such as CLIP [27] and LLaVA [19], are trained on large-scale, lightly curated web datasets [30, 32]. Since web captions often describe what is present rather than what is important, models learn spurious + +
Prior work AttackOur Web Artifact Attacks
(a) Matching Text(b) Non-Matching Text(c) Graphics wo/ text(d) Graphics w/ text
SeniorSentonPENSION +OFFERED
Senior +VL Model + Senior PersonSenton +VL Model + Senior PersonSenior +VL Model + Senior PersonSenior +VL Model + Senior Person
+ +Figure 1. We propose a novel class of Web Artifact Attacks that expand the attack surface of Vision-Language Models beyond prior typographic attacks. For a sample task of predicting a person's age, (a) illustrates Typographic Attacks from prior work [3] which manipulate predictions using class-matching text. In contrast, our proposed web artifact attacks, consisting of (b) non-matching text, (c) standalone graphics, and (d) graphics with embedded text, can also mislead the model. + +patterns—unintended correlations that arise from frequent but unrelated co-occurrences [1, 12, 14, 16, 18, 23, 33, 34, 37]. For example, prior work found that Chinese characters frequently co-occur with the word "carton" in ImageNet [16], leading the model to misclassify images containing Chinese characters as a "carton," regardless of its actual content. Moreover, these correlations can be weaponized by attackers to manipulate model predictions. Fig. 1(a) shows a well-documented example, typographic attacks [3, 26], where inserting the text of the wrong category into an image tricks the model into predicting that category. This attack takes advantage of the model's over-reliance on text, a bias reinforced by pretraining datasets in which captions frequently parrot text visible in the images [18]. + +A shortcoming of prior work is that they focused only on explicit class-matching text, overlooking a broader set of vulnerabilities, including non-matching text and graphical artifacts such as logos. These elements, which are common in web-scale datasets due to branding and advertisements, may introduce unintended visual cues that models may rely on for classification. To address this gap, in Fig. 1, we introduce Web Artifact Attacks, a broader class of attacks that exploit + +![](images/e7d50464df49f9ac7f2711e28432b6ba3db7756edb629d2bbd3336dfb3189f1e.jpg) +Figure 2. Overview of the Web Artifacts Search and Selection. The process involves two steps: Step 1: We filter an image-text dataset using CLIP-based retrieval and a set of artifact prompts. Retrieved artifacts are categorized into three groups: graphics with text ( $\mathcal{A}^{\mathrm{graphics-text}}$ ), graphics without text ( $\mathcal{A}^{\mathrm{graphics-no-text}}$ ), and unrelated text ( $\mathcal{A}^{\mathrm{text}}$ ) using OCR-based text extraction (Sec. 3.1.1). Step 2: artifacts are applied to opposing-class images (e.g., "Smiling" images when targeting "Frowning"), and their effectiveness is evaluated based on their ability to mislead model predictions. Artifacts are ranked according to their attack success rate, resulting in a refined set of high-impact artifacts for each category ( $\mathcal{A}_{c*}^{\mathrm{graphics-text}}$ , $\mathcal{A}_{c*}^{\mathrm{graphics-no-text}}$ , and $\mathcal{A}_{c*}^{\mathrm{text}}$ ) (Sec. 3.1.2). + +![](images/a5ff35904cc984bb63fc4e98b3524be5c98c03db68af9b276e03558af8e8d0c9.jpg) + +both non-matching text (Fig. 1(b)) and graphics (Fig. 1(c) and (d)) to mislead VLM predictions. We categorize graphical artifacts into two types: those with embedded text, which may reinforce textual biases, and those without text, isolating the impact of purely visual cues. + +Unlike typographic attacks [3, 26], where the adversary inserts class-matching text to deliberately exploit the model's reliance on words, Web Artifact Attacks pose a greater threat because the defender does not know in advance the attacking artifacts. Thus, it is not obvious how to identify effective attacking artifacts or defend against them. To find attacks, we frame Web Artifact Attacks as a search problem, scanning image-text datasets [6, 29] to uncover text and graphics that can distort model predictions. The simplicity of these attacks makes them a notable security concern for deployed VLMs, as they only require dataset queries using similarity retrieval, making them highly accessible. + +To that end, our search process consists of three stages: (1) Fig. 2 Green : Artifact Retrieval, where we extract candidate artifacts by scanning image-text datasets for images with text or graphics (2) Fig. 2 Purple : Effect Estimation, where we apply artifacts to images from opposing classes and evaluate their influence on model predictions, and (3) placement optimization, where we refine the positioning of artifacts to maximize attack success. + +We evaluate the resulting attack artifacts across different VLM training objectives, visual encoder architectures, and pretraining dataset curation. Our findings demonstrate that Web Artifact Attacks are highly effective across various VLM training paradigms, including contrastive learning [27], sigmoid loss [36], and Large Language Model (LLM) fusion + +with visual features [19, 38]. Furthermore, these vulnerabilities persist regardless of the visual encoder size (ViT-B-32, ViT-B-16, ViT-L-14), and attempts at stronger pretraining dataset curation (DataComp [11]). + +A key property of these attacks is high transferability—artifacts identified for one model retain up to $90\%$ effectiveness when applied to another, indicating shared failure modes across vision language models. Additionally, attacks remain effective even when artifacts are small or transparent, revealing persistent vulnerabilities even in cases where artifacts are harder to detect. Furthermore, combining different artifact types significantly amplifies attack success, with text and graphical symbols reinforcing each other to reach nearly $100\%$ misclassification rates. While text-based artifacts and graphics with embedded text are the most effective, graphics without text provide a stealthier attack vector, making them harder to detect while still influencing predictions. + +Finally, we explore mitigation strategies by building on prior work [8], which showed that incorporating typographic attacks into the prompt can reduce their effectiveness. For example, if an image of a youth is attacked with the text "senior," the prompt "an image of a senior with the word 'youth' written on it" can weaken the attack. However, unlike typographic attacks, our attacks include graphical elements that cannot be easily incorporated into text. To address this, we propose augmenting the prompt with descriptions of graphical artifacts obtained through a captioning model. For LLM-based VLMs, we prompt the model to identify and describe these artifacts in its response. This approach reduces success rates by up to $15\%$ relative to standard prompts, suggesting a promising direction for enhancing model robustness. + +To summarize our contributions: + +- We expose Web Artifact Attacks, which extend beyond typographic attacks by leveraging non-matching text, standalone graphics, and graphics with embedded text to mislead VLM predictions. +- We develop a scalable attack pipeline to search, evaluate, and optimize artifacts from pretraining datasets, leveraging their natural abundance in branding and advertisements. +- We show that our attacks are highly transferable, effective even when artifacts are small or transparent, and significantly stronger when combining multiple artifacts, reaching up to $100\%$ success rates. +- We explore mitigation strategies that incorporate descriptions of the artifacts into the VLM(s) prompt which successfully mitigate some of their effect. + +# 2. Related Work + +Bias due to Artifacts. Prior work [16] has shown that models, from ResNet-50 [13] to large-scale VLMs [27, 36], can mistakenly associate Chinese characters watermark with the carton class in Imagenet [9]. Bykov et al. [5] further demonstrated that multilingual text artifacts (e.g., Arabic, Latin, and Hindi) influence model predictions in unintended ways. Unlike these studies, which primarily observe such correlations, our work weaponizes these artifacts as an attack vector. Additionally, while prior work focused on manually discovered watermarks, we develop an automated mining pipeline that extracts a diverse set of non-matching text, graphical symbols with and without text at scale. + +Typographic Attacks. Prior work [3, 8, 22, 26] has demonstrated that VLMs, including CLIP [27] and LLaVA [19], are highly susceptible to typographic attacks, where inserting class-matching text into an image directly manipulates predictions. These attacks exploit VLMs' over-reliance on textual cues, a bias reinforced by pretraining datasets in which captions often echo the visible text within images [18]. However, typographic attacks are inherently predefined, as they rely on explicitly matching class names, making them relatively easy to detect. In contrast, we show that non-matching text, graphical symbols, and mixed artifacts—elements not known a priori—can also mislead VLMs, enabling more flexible and covert attack strategies. + +Relation to Adversarial Attacks. Adversarial attacks [7, 10, 28, 35, 39] learn an imperceptible noise, that disrupts the model visual recognition capabilities when added to the model. More recently, adversarial techniques have been weaponized for jailbreaking multimodal models to bypass safety mechanisms and generating restricted outputs [4, 24, 31]. While our attacks share a similar end goal to adversarial attacks, namely, disrupting model behavior, adversarial and artifact attacks are fundamentally different from each other. Our attacks originate as a result of spurious correlation in uncurated pretraining datasets, while adversar + +ial attacks result from an artificial noise uniquely designed through the use of model intrinsics (e.g., gradients) to disrupt the fundamental mechanics of neural nets. Therefore, to develop such attacks, an attacker needs a sophisticated understanding of the intrinsic mechanics of neural nets such as e.g., Gradient [28, 35, 39], Meta Learning [10] or Attention maps [7] to name a few. However, the attacker, in our setup, only needs to know how to query the model. + +# 3. Web Artifact Attacks Against Vision-Language Models + +Vision-language models (VLMs) [19, 27, 36] are trained on large, uncurated web-scale datasets, leading them to form unintended correlations between visual elements and semantic categories [16, 18, 23, 33, 34]. Prior work has demonstrated that typographic attacks, which insert explicit class-matching text into images, exploit these correlations to manipulate model predictions [3, 26]. However, typographic attacks rely on direct textual matches, making them predictable and relatively easy to detect. + +We propose Web Artifact Attacks, a new vector of unpredictable class of attacks that leverage artifacts from imagetext datasets [6, 29]. Unlike typographic attacks, these artifacts—such as non-matching text (text that contains partial overlaps, phonetic similarities, or visually similar characters but does not explicitly spell out the target class), graphical symbols—do not need to explicitly match the target class yet can still mislead model predictions. + +Formally, our artifact-based attack introduces an artifact $a$ correlated with a target class $c^*$ into an image $x$ , causing the model $f_{\theta}$ to misclassify it as $c^*$ : + +$$ +\arg \max _ {c} f _ {\theta} (x) = \hat {c} \quad \Rightarrow \quad \arg \max _ {c} f _ {\theta} (x \oplus a) = c ^ {*}, +$$ + +where $\oplus$ denotes artifact insertion and $\hat{c}$ is the correct class. + +Since these artifacts are not known a priori, they are harder to defend against while also making it challenging to determine in advance which artifacts will be effective. Therefore, we frame finding our attack artifacts as a search problem: systematically identifying artifacts within large-scale image-text datasets that can trigger misclassification. + +To that end, Sec. 3.1 describes our proposed the search pipeline, Sec. 3.2 analyzes the extracted artifacts to reveal broader insights into model vulnerabilities, Sec. 3.3 introduces an evaluation metric that accounts for occlusion effects, and finally, Sec. 5 proposes a mitigation strategy. + +# 3.1. Attack Pipeline + +This section describes our search pipeline to systematically find attack artifacts. First, in Sec. 3.1.1, we extract artifacts from pretraining datasets by filtering images containing either text or graphical elements, categorizing them into non-matching text, graphics with text, and graphics without text. + +Next, in Sec. 3.1.2, we evaluate how these artifacts influence model predictions by applying them to opposing-class images and selecting those that consistently increase misclassification rates. Finally, in Sec. 3.1.3, we optimize artifact placement to maximize attack effectiveness. + +# 3.1.1. Extracting Artifacts from a Pretraining Dataset + +To search for our artifacts, we leverage the fact that graphics (e.g. a logo of a company) frequently appear in web-scale datasets as standalone images due to the abundance of branding content. Using a CLIP-based retrieval mechanism, we apply broad text prompts $\mathcal{P}$ (e.g., "a photo of a graphic") to extract images from an image-text dataset $\mathcal{D}$ (e.g., CC12M [6]). We then rank the retrieved images by their similarity scores and retain the top $1\%$ of the highest-scoring images, ensuring that the selected artifacts closely match the intended query while filtering out irrelevant content. To verify the quality of the retrieved artifacts, we conduct a human study by randomly sampling 1,000 images from the extracted set and manually assessing their noise level, defined as the percentage of non-graphic or irrelevant images. We find that the noise level remains below $2\%$ , indicating that our retrieval process effectively isolates graphical artifacts while minimizing unrelated content. + +Next, as shown in Fig. 2, we categorize artifacts into three types: Green: unrelated text, graphics with embedded text, and graphics without text. To achieve this categorization, we apply OCR-based text detection (We use off-the-shelf models from EasyOCR [2]) to identify artifacts containing text and separate them from purely graphical artifacts. Additionally, to ensure that the extracted text does not merely replicate typographic attacks from prior work [3, 26], we filter out any text that explicitly matches class names in the downstream dataset. This structured approach allows us to systematically evaluate how VLMs respond to textual and graphical cues, both independently and in combination, providing deeper insight into their learned vulnerabilities. + +# 3.1.2. Finding Artifacts That Influence a Target Class + +After obtaining our set of artifacts in Sec. 3.1.1, the next step is to determine which ones are most effective at misleading model predictions. Not all artifacts significantly influence classification—some may be ignored by the model, while others may reinforce correct predictions. To systematically identify high-impact artifacts, we evaluate their ability to shift model predictions toward a target class. + +To that end, as Fig. 2 Purple illustrates, we apply each artifact to opposing-class images (e.g. if the target class is "car images" then we use every image that is not a car) and measure its effectiveness in shifting predictions toward the target class. If an artifact consistently causes images from unrelated classes to be classified as the target class, it reveals a strong learned correlation in the model that can be exploited for adversarial purposes [25]. Formally, given a downstream + +dataset $\mathcal{D}_{\mathrm{downstream}}$ , we define a subset of opposing-class images as: + +$$ +\mathcal {D} _ {\neg c ^ {*}} ^ {\text {t r a i n}} \subset \left\{x \mid x \in \mathcal {D} _ {\text {d o w n s t r e a m}}, y \neq c ^ {*} \right\}, +$$ + +where $y$ is the ground-truth label. As we show in our experiments, $\mathcal{D}_{\neg c^*}^{\mathrm{train}}$ can be as small as 10 samples per class while still uncovering the most effective artifacts. We then apply each artifact $a$ from the three previously defined categories to this subset. + +To ensure efficient evaluation, artifacts are initially placed at random locations. This allows us to identify artifacts that demonstrate adversarial effects across a variety of placements, rather than optimizing for a single position prematurely. By deferring location-specific optimization, we reduce the computational complexity from $\mathcal{O}(|\mathcal{A}||\mathcal{D}_{\neg c^*}^{\mathrm{train}}||\mathcal{L}|)$ to $\mathcal{O}(|\mathcal{A}||\mathcal{D}_{\neg c^*}^{\mathrm{train}}|)$ during this stage (where $\mathcal{L}$ refers to the set of target locations), ensuring that we focus on selecting artifacts with strong generalizable effects before fine-tuning their positioning. + +Artifact impact score $a$ is defined per-artifact as the proportion of modified images classified as the target class: + +$$ +P _ {c ^ {*}} (a) = \frac {1}{n} \sum_ {i = 1} ^ {n} \mathbf {1} \left[ f _ {\theta} \left(x _ {i} ^ {a}\right) = c ^ {*} \right], +$$ + +where higher values indicate a stronger influence on misclassification. We compute this score for each artifact in our three categories: text artifacts $(\mathcal{A}^{\mathrm{text}})$ , graphics with embedded text $(\mathcal{A}^{\mathrm{graphics-text}})$ , and graphics without text $(\mathcal{A}^{\mathrm{graphics-no-text}})$ , as defined in Sec. 3.1.1. For each artifact category, we rank artifacts by $P_{c^{*}}(a)$ in descending order and select the top $s$ most effective ones. This results in the sets $\mathcal{A}_{c^{*}}^{\mathrm{text}}, \mathcal{A}_{c^{*}}^{\mathrm{graphics-text}}$ , and $\mathcal{A}_{c^{*}}^{\mathrm{graphics-no-text}}$ , where $c^{*}$ denotes the target class. + +# 3.1.3. Optimizing Artifact Placement + +While Sec. 3.1.2 identified artifacts that are consistently effective across different placements, their positioning within an image can further enhance misclassification rates. + +To address this, for each artifact $a \in \mathcal{A}_{c^*}^{\mathrm{top - text}} \cup \mathcal{A}_{c^*}^{\mathrm{top - graphics - text}} \cup \mathcal{A}_{c^*}^{\mathrm{top - graphics - no - text}}$ , we evaluate a set of candidate locations $\mathcal{L}$ to determine the most effective positioning. For each artifact, we pick the location that obtains the highest success rate. This step fine-tunes the effect of already strong artifacts, maximizing misclassification rates. Refer to the Supp. B for an ablation of location effect. In short, we find that the top border of the image (especially top middle) to be the most effective. + +# 3.2. Insights from the Artifacts + +After extracting and ranking artifacts based on their influence on model predictions, we examine the highest-impact artifacts for different target classes. Fig. 3 presents examples of these artifacts, categorized into graphics without text ( $A_{c^*}^{graphics-no-text}$ ), graphics with embedded text + +
A graphics-no-textA graphics-textA text C*
Brown HairBrownx JACUAR CHAIRCHAIR DARKcRown DARK VS
JapanASIA center JAMAN GLOVO大関開スロー暗間の外katsu JpnP J^P^N
SeniorTENSIO SHS senors Retirement Serviceseniors Retirement Serviceseniors Retirement Service
+ +Figure 3. Examples of Web Artifacts split into: graphics without text ( $A_{c^*}^{\text{graphics-no-text}}$ ), graphics with embedded text ( $A_{c^*}^{\text{graphics-text}}$ ), and unrelated text ( $A_{c^*}^{\text{ext}}$ ). Each row corresponds to a different target class (Brown Hair, Japan, Adult), that models have learned to associate with the artifacts. Notably, text artifacts need not match the class exactly, while graphical symbols can represent indirect but learned associations. These findings highlight the diverse range of artifacts that can manipulate model predictions. Refer to Supp. A for more examples. + +$(\mathcal{A}_{c^*}^{\mathrm{graphics-text}})$ , and unrelated text artifacts $(\mathcal{A}_{c^*}^{\mathrm{text}})$ . These examples reveal distinct patterns in how different artifact types induce misclassification. + +Non-textual graphics often share visual attributes loosely related to the target class. For instance, brown-colored logos and animal symbols appear frequently for "Brown Hair," while "Japan" is associated with red-and-white designs reminiscent of its national flag. Similarly, the "Senior" class is linked to medical symbols and retirement-related imagery, suggesting that models rely on broad visual associations learned from pretraining data. + +Graphics with embedded text frequently contain contextually related words, even when the complete phrase is irrelevant. Examples include "Brown" from unrelated brand names appearing in the "Brown Hair" category and "Japan Cloud" for the "Japan" category. This demonstrates that models are highly sensitive to the presence of text fragments within graphical elements, further reinforcing the text-heavy biases introduced during pretraining. + +Text-only artifacts, we observe frequent misspellings, phonetic approximations, and visually similar character patterns rather than exact class matches. Words like "senors" for "Senior" or "J^P^N" for "Japan" illustrate how models generalize text-based associations beyond strict class labels, making these artifacts more difficult to detect than conventional typographic attacks. + +Refer to Supplementary A for more examples. + +# 3.3. Attack Success Evaluation + +Prior evaluations of typographic attacks [3, 8, 22, 26] measure attack success by checking whether inserting class-matching text misleads VLMs. However, they overlook occlusion effects—misclassification may occur due to the added text obscuring important image features. To ensure that attack success is solely due to the artifact rather than occlusion, we introduce an evaluation framework that isolates its effect while preserving the image's classifiability. + +Given an input image $x$ and an artifact $a$ placed at location $l$ , we define the attack-modified image as $x^a = x \oplus a$ , where + +$\oplus$ denotes artifact insertion. To verify that misclassification stems from the artifact itself, we introduce a masking-based validation step. We generate a masked version of the image, $x^{-l}$ , where the region $l$ is occluded using a mask $M_{l}$ . + +To ensure that the image remains classifiable after artifact removal, we retain only samples where $x^{-l}$ is still correctly classified as its original class $c$ . Let $f_{\theta}$ be the model's classifier, and let $c$ be the ground-truth class of $x$ . We discard any image where $f_{\theta}(x^{-l}) \neq c$ , ensuring that attack success cannot be attributed to the removal of critical visual features. + +The Attack Success Rate (ASR) is then defined as the proportion of cases where the model predicts the target class $c^*$ for $x^a$ but maintains correct classification for $x^{-l}$ : + +$$ +A S R = \frac {1}{N} \sum_ {i = 1} ^ {N} \mathbf {1} \left[ f _ {\theta} \left(x _ {i} ^ {a}\right) = c ^ {*} \wedge f _ {\theta} \left(x _ {i} ^ {- l}\right) = c \right], +$$ + +where $\mathbf{1}[\cdot ]$ is the indicator function, and $N$ is the number of samples passing the masking validation step. This stricter evaluation ensures that attack success is genuinely due to the artifact rather than incidental occlusion effects. To ensure that the attack generalizes beyond the images used during search, the final ASR is evaluated on a separate test set disjoint from the training set $\mathcal{D}_{\neg c^*}^{\mathrm{train}}$ used for artifact mining. + +# 4. Experiments + +Datasets. We use CC12M [6] as our primary dataset for searching and extracting artifacts. Our extraction process yields 87k artifacts, but this approach is dataset-agnostic and can be applied to any large-scale image-text dataset, such as LAION [29, 30]. To evaluate the effectiveness of these artifacts, we conduct experiments across five distinct tasks spanning both human-related attributes and object/geography classification. For human-related tasks, we predict Gender (Male/Female) and Age (Child, Teen, Adult, Elderly) using FairFace [15], as well as Smiling using CelebA [20]. For non-human classification tasks, we predict aircraft models using FGVC Aircraft [21] and countries using Country211 [27]. For each downstream dataset, we use only 32 images + +![](images/932cd8349a4d49f510eb846342609cc4450f411c025301b5a848619c8c49892e.jpg) +(a) Vision/Language Fusion + +![](images/fd60ce8589d383c6969e2893a7a52ada22f72b98e4a84e653b70d960839d93b0.jpg) +(b) Pretraining Dataset +Figure 4. Breakdown of Artifact Attacks Performance. Our Web Artifact attacks are highly effective across different configurations: namely (a) Vision/Language Fusion, (b) Pretraining Dataset, and (c) Vision Encoder Architecture. Refer to Sec. 4.1 for further discussion.. + +![](images/6d80fdeff901efdb44bcbb8ead05ff54fa0a1f8c182bb0e7d5c2c0055ff32792.jpg) +(c) Vision Encoder Architecture + +per class to estimate the Artifact Success Rate (ASR), which as we demonstrate in Sec. 4.1, is sufficient to estimate ASR. Metrics. We use the success rate as outlined in Sec. 3.3. In short, given a target class $c^*$ , we measure attack effectiveness by evaluating whether the model misclassifies an image after the artifact that correlates with $c^*$ is introduced while ensuring that the image remains correctly classified when the artifact is masked. We also report the change of the confidence score in target class $c^*$ after pasting the artifact. + +Models. We evaluate the robustness of VLMs to Web Artifact Attacks by varying vision-language fusion strategy, vision encoder architecture, and pretraining dataset size and curation. For fusion, we test contrastive learning (CLIP) [27], sigmoid-based alignment (SigLIP) [36], and LLM-based fusion (LLaVA) [19]. To assess the impact of architecture size, we evaluate ViT-B-32, ViT-B-16, ViT-L-14 while we fix the pretraining dataset to LAION 2B [29]. Finally, we compare models pretrained on LAION 400M, LAION-2B [29], and DataComp [11] while fixing the pretraining architecture to ViT-B-32 and ViT-B-16 to examine whether dataset curation affects attack vulnerability. + +Artifact Setting. We fix the artifact size to 10th of the image size. Refer to Supplementary. D for a visual example and ablation of artifact size and transparency. + +# 4.1. Results + +We benchmark the effect of Web Artifact Attacks on Vision-Language Models (VLMs) across different vision-language fusion methods, model architectures, and pretraining datasets. We make the following observations: + +Artifact Attacks are Effective across Diverse Model Configurations. As seen in Fig. 4, Web Artifact Attacks consistently achieve high success rates across different vision-language fusion strategies, pretraining datasets, and model architectures, reaching up to $90\%$ success. Comparing the artifacts, text-based attacks are the most effective, followed by graphics with embedded text, and finally graphics without text, suggesting that models are most vulnerable to explicit textual cues but still susceptible to purely visual artifacts. + +CLIP is most vulnerable to Artifact Attacks. In Fig. 4(a), models trained with standard contrastive learning (OpenAI + +CLIP) exhibit the highest attack success rates across all artifact types, particularly for text-based artifacts, reinforcing CLIP's strong reliance on textual cues. The LLM-based fusion model (LLaVA) shows a moderate reduction in vulnerability, likely due to its alignment on a smaller, curated dataset such as COCO [17], which likely have much fewer graphics and logos. Similarly, SigLIP, which employs a different training objective and a pretraining dataset, achieves moderate reduction in vulnerability too on par with LLaVA. + +Pretraining Dataset Curation has Minimal Effect on Success Rate. Fig. 4(b) shows that models pretrained DataComp [11] exhibit a slightly improved robustness to artifact attacks than uncurated Laion400M and LAION 2B. Despite DataComp's filtering, it does not majorly reduce susceptibility, suggesting that its filtering strategy may be ill-suited for mitigating spurious correlations. DataComp primarily optimizes for image quality and semantic similarity to high-quality benchmarks, rather than explicitly removing textual artifacts or graphical elements that reinforce spurious correlations. As a result, while the dataset curation improves image-text alignment for downstream tasks, it fails to eliminate the kind of incidental patterns that make models vulnerable to Web Artifact Attacks. This highlights a critical limitation: filtering strategies that focus on dataset relevance may not address Web Artifact Attacks. + +Larger Architecture is More Vulnerable to Textual Attacks. As seen in Fig. 4(c), larger models, such as ViT-L-14, exhibit higher attack success rates on text-based attacks. This can be attributed to their increased capacity for optical character recognition, enabling them to more effectively detect and leverage textual elements in images. As a result, they become more susceptible to attacks that exploit text-based spurious correlations present in web-scale datasets. + +Artifacts Based Attacks vs Prior Work's Typographic Attacks. Fig. 5 presents the average performance of our web-scale Web Artifact Attacks compared to Typographic Attacks [3]. The results demonstrate that our text-based attacks, derived from mined logo datasets, are more effective than typographic attacks, which are typically limited to using text from the opposing class. This suggests that optimizing attacks beyond simple class-matching text—by leveraging + +![](images/ce56c51593c20ae91241ae5e66a36730d7dd83426c1e49e17f6f1f8fcf8eedbd.jpg) +Figure 5. Comparison to Prior Work Attacks. Success rate of Web Artifact attacks vs prior work Typographic Attacks [3]. + +![](images/db074160a51fc4fd7f544380abed1852ff67f05f0284b5f0e6cde374d8de8af5.jpg) +Figure 6. Confidence in Target Class when different types of artifacts—graphics with text (orange), graphics without text (green), and text (red)—are introduced, compared to the baseline with no artifacts (blue). Results demonstrate that artifacts significantly increase model confidence in the target class. + +![](images/97405e31b9eaa37ed4a6d457c408227bf14559e527fec54b09c54ba18b84c49b.jpg) +Figure 7. Success Rate of Combining artifacts. Results indicate that combining artifact types leads to higher attack success. + +![](images/37aed1471800a34a4a79b696ef34197343b95ceae30c7f3b0e25045cbd6bac16.jpg) +Figure 8. Sample Size vs Success Rate. Success rate of artifact-based attacks vs the number of samples used from the downstream dataset to estimate their effect. Results indicate that a relatively small sample size is sufficient to assess attack effectiveness. + +a broader range of textual configurations—can significantly enhance their effectiveness. In contrast, graphical attacks tend to be less effective overall. However, they open up a novel attack surface that, as discussed in Section 5, proves more challenging to defend against. + +Artifacts Increase Model Confidence in the Target Class. + +Fig. 6 shows that the introduction of textual and graphical + +artifacts significantly increases the model's confidence in the target class compared to the no-artifact baseline. Text-based artifacts lead to the highest confidence scores, followed by graphics with embedded text and graphics without text. This increased confidence correlates with the attack success rates observed in prior experiments, suggesting that artifacts not only influence predictions but also enhance model certainty in incorrect classifications. + +The Compounding Effect of Different Types of Artifacts Fig. 7 presents the attack success rate across different artifact types and their combinations. The results highlight a clear synergistic effect when multiple artifacts are used together, significantly increasing misclassification rates compared to individual artifacts. While text-only artifacts remain highly effective, their impact is amplified when combined with graphical symbols. The highest success rates are observed in $(\mathrm{A} + \mathrm{C})$ and $(\mathrm{A} + \mathrm{B})$ cases, where text is paired with either graphical symbols with or without text, reaching near $100\%$ attack success in some instances. This suggests that models do not treat textual and graphical cues independently but rather reinforce their reliance on both when they co-occur, exacerbating spurious correlations learned during pretraining. Additionally, graphical artifacts without text (B) alone exhibit moderate attack success, but when combined with text $(\mathrm{B} + \mathrm{C})$ , their effectiveness increases substantially. + +A Small Sample Size is Enough for Estimating Artifact Effect. A key question for our attacks is how many samples are needed per class to reliably estimate an artifact's attack effect. Fig. 8 shows that across all artifact types, the attack success rate stabilizes quickly as the number of samples increases, indicating that only a small sample size is required to obtain a reliable estimate. This result is significant as it suggests that evaluating Web Artifact Attacks does not require extensive sampling, which greatly reduces computational overhead. By using only a fraction of the dataset, we can efficiently assess the impact of artifacts while preserving computational resources. This finding is particularly important for large-scale evaluations, where repeatedly testing artifacts on full datasets would be prohibitively expensive. + +Web Artifact Attacks are Highly Transferable. Fig. 9 illustrates that Web Artifact Attacks are highly transferable across different models. Across all artifact types, success rates remain consistently high when applying artifacts mined from one model to another. The most significant drop in success rate occurs when transferring attacks to a model trained with SigLIP [36], which employs a different training objective. However, even in this case, the attack success rates remain substantial, indicating that spurious correlations exploited by these attacks persist across diverse model architectures and training paradigms. This suggests that mitigating these vulnerabilities will require more than simply altering the training objective. + +![](images/a2d8d6caead21c5f29efdcb5a0b7074c1d4233cc972957225f072144ae39bb35.jpg) +Figure 9. Transferability of Web Artifact Attacks Across Different Vision Encoder Architectures. The heatmap shows the success rate (\%) of artifacts search on a specific model (y-axis) and evaluated on another model (x-axis). Higher success rates indicate greater transferability of spurious correlations. Results indicate that artifacts effect is highly transferable across architectures. + +![](images/8e91bae502443af0e9ab544b11d6348c96a9c9ad4a25047b0ab59c8dc3c73cd4.jpg) + +![](images/9331c1b3740301919ff3acb630277eea7fbb6264b7045912811fdac45c1962e8.jpg) + +![](images/b1b3ac3fb7dd3d079d95cf50ae0a146194e51dccd73fadbe562dd213407d65bf.jpg) +Figure 10. Artifact Aware Prompting. Results indicate that the attack effect can be mitigated with more informative prompts. + +# 5. Mitigation w/ Artifact Aware Prompting + +Recent work [8] has uncovered an important insight: VLMs can be guided to adjust their attention toward specific visual elements when provided with more informative prompts. This finding was leveraged as a mitigation strategy for typographic attacks by explicitly incorporating the attack text into the prompt. For example, assume the an image of a dog is attacked with the text of "a cat," then rather than using "a photo of a dog," the authors use: "a photo of a dog with cat written on it." While this approach is effective for textual artifacts, it does not generalize to graphical artifacts, which lack an explicit textual form and cannot be directly incorporated into the language input. + +To address this limitation, we propose a mitigation approach inspired by [8] for the two types of models. For Contrastive VLM(s): we use structured descriptions of graphical symbols to make these artifacts explicit to the model. Specifically, we employ a captioning model to transform the graphics into a textual description which then are appended to the input prompt, ensuring that the model processes them explicitly. Refer to Supp. E for an example. For Large VLM(s): following [8], we leverage the generative capabilities for LVLMs like LLaVA [19] to integrate artifact awareness directly. We prompt LLaVA to describe any graphical or textual artifacts present in the image, ensuring that these features are explicitly considered during inference. Note that our approach does not require model retraining or architectural modifications. The only overhead is the cost of + +executing the captioning model. + +Results. Fig. 10 illustrates the effectiveness of our approach in reducing attack success rates. Across all artifact types, artifact-aware prompting consistently lowers the success rate, indicating that explicitly guiding the model to consider artifacts can diminish their impact. However, this defense is less effective against our attacks compared to prior typographic attacks [3], with the largest performance gap observed for graphical attacks. This highlights the increased difficulty posed by our attacks and suggests promising directions for future research in developing more robust defenses. + +# 6. Conclusion + +In this work, we introduced Web Artifact Attacks, a novel vector of attacks that make use of non-matching text and graphical artifacts—both with and without embedded text—to mislead model predictions, revealing vulnerabilities beyond traditional typographic attacks. Our results show that these attacks not only increase model confidence in incorrect classifications but also persist across different model architectures, vision-language fusion strategies, and pretraining datasets. Moreover, we showed how our search process for attack artifacts is efficient only requiring a handful of samples to approximate attack effect. While mitigation strategies such as artifact-aware prompting reduce attack success rates, they are insufficient to eliminate the threat, particularly for non-textual graphical artifacts. + +Future Work. Our findings highlight the need for more robust defenses against our attacks. A promising direction is dataset curation that extends beyond simple CLIP-based filtering; developing more informative captioning strategies that better contextualize images containing logos and graphics could help models learn more meaningful associations rather than relying on spurious correlations. Additionally, while our artifact discovery process is effective, it could be further optimized for efficiency. Future improvements could explore faster, more scalable methods for identifying artifacts in large-scale datasets, reducing the computational cost of attack evaluations without sacrificing accuracy. + +# References + +[1] Sandhini Agarwal, Gretchen Krueger, Jack Clark, Alec Radford, Jong Wook Kim, and Miles Brundage. Evaluating clip: towards characterization of broader capabilities and downstream implications. arXiv preprint arXiv:2108.02818, 2021. 1 +[2] Jaded AI. EasyOCR: Ready-to-useOCR with $80+$ supported languages and growing. https://github.com/JadedAI/EasyOCR, 2020.4 +[3] Hiroki Azuma and Yusuke Matsui. Defense-prefix for preventing typographic attacks on clip. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3644-3653, 2023. 1, 2, 3, 4, 5, 6, 7, 8 +[4] Eugene Bagdasaryan, Tsung-Yin Hsieh, Ben Nassi, and Vitaly Shmatikov. Abusing images and sounds for indirect instruction injection in multi-modal llms. arXiv preprint arXiv:2307.10490, 2023. 3 +[5] Kirill Bykov, Klaus-Robert Müller, and Marina M-C Höhne. Mark my words: Dangers of watermarked images in imagenet. In European Conference on Artificial Intelligence, pages 426-434. Springer, 2023. 3 +[6] Soravit Changpinyo, Piyush Sharma, Nan Ding, and Radu Soricut. Conceptual 12m: Pushing web-scale image-text pretraining to recognize long-tail visual concepts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3558-3568, 2021. 2, 3, 4, 5 +[7] Sizhe Chen, Zhengbao He, Chengjin Sun, Jie Yang, and Xiaolin Huang. Universal adversarial attack on attention and the resulting dataset damagenet. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(4):2188-2197, 2020. 3 +[8] Hao Cheng, Erjia Xiao, Jindong Gu, Le Yang, Jinhao Duan, Jize Zhang, Jiahang Cao, Kaidi Xu, and Renjing Xu. Unveiling typographic deceptions: Insights of the typographic vulnerability in large vision-language model. ECCV, 2024. 2, 3, 5, 8 +[9] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 3 +[10] Jiawei Du, Hu Zhang, Joey Tianyi Zhou, Yi Yang, and Jiashi Feng. Query-efficient meta attack to deep neural networks. arXiv preprint arXiv:1906.02398, 2019. 3 +[11] Samir Yitzhak Gadre, Gabriel Ilharco, Alex Fang, Jonathan Hayase, Georgios Smyrnis, Thao Nguyen, Ryan Marten, Mitchell Wortsman, Dhruba Ghosh, Jieyu Zhang, et al. Datacomp: In search of the next generation of multimodal datasets. Advances in Neural Information Processing Systems, 36: 27092-27112, 2023. 2, 6 + +Acknowledgments This material is based upon work supported, in part, by DARPA under agreement number HR00112020054. Any opinions, findings, and conclusions or recommendations expressed in this material are those of the author(s) and do not necessarily reflect the views of the supporting agencies. +[12] Melissa Hall, Laura Gustafson, Aaron Adcock, Ishan Misra, and Candace Ross. Vision-language models performing zero-shot tasks exhibit gender-based disparities. arXiv preprint arXiv:2301.11100, 2023. 1 +[13] Kaiming He, X. Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 770-778, 2016. 3 +[14] Sepehr Janghorbani and Gerard De Melo. Multimodal bias: Introducing a framework for stereotypical bias assessment beyond gender and race in vision language models. arXiv preprint arXiv:2303.12734, 2023. 1 +[15] Kimmo Karkkainen and Jungseock Joo. Fairface: Face attribute dataset for balanced race, gender, and age for bias measurement and mitigation. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 1548-1558, 2021. 5 +[16] Zhiheng Li, Ivan Evtimov, Albert Gordo, Caner Hazirbas, Tal Hassner, Cristian Canton Ferrer, Chenliang Xu, and Mark Ibrahim. A whac-a-mole dilemma: Shortcuts come in multiples where mitigating one amplifies others. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20071-20082, 2023. 1, 3 +[17] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer vision-ECCV 2014: 13th European conference, zurich, Switzerland, September 6-12, 2014, proceedings, part v 13, pages 740-755. Springer, 2014. 6 +[18] Yiqi Lin, Conghui He, Alex Jinpeng Wang, Bin Wang, Weijia Li, and Mike Zheng Shou. Parrot captions teach clip to spot text. In European Conference on Computer Vision, pages 368-385. Springer, 2024. 1, 3 +[19] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2023. 1, 2, 3, 6, 8 +[20] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 3730-3738, 2015. 5 +[21] Subhransu Maji, Esa Rahtu, Juho Kannala, Matthew Blaschko, and Andrea Vedaldi. Fine-grained visual classification of aircraft. arXiv preprint arXiv:1306.5151, 2013. 5 +[22] Joanna Materzyńska, Antonio Torralba, and David Bau. Disentangling visual and written concepts in clip. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16410-16419, 2022. 3, 5 +[23] Yannic Neuhaus, Maximilian Augustin, Valentyn Boreiko, and Matthias Hein. Spurious features everywhere - large-scale detection of harmful spurious features in imagenet. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 20235-20246, 2023. 1, 3 +[24] Xiangyu Qi, Kaixuan Huang, Ashwinee Panda, Peter Henderson, Mengdi Wang, and Prateek Mittal. Visual adversarial examples jailbreak aligned large language models. In Proceedings of the AAAI conference on artificial intelligence, pages 21527-21536, 2024. 3 + +[25] Maan Qraitem, Kate Saenko, and Bryan A Plummer. Bias mimicking: A simple sampling approach for bias mitigation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20311-20320, 2023. 4 +[26] Maan Qraitem, Nazia Tasnim, Piotr Teterwak, Kate Saenko, and Bryan A Plummer. Vision-llms can fool themselves with self-generated typographic attacks. arXiv preprint arXiv:2402.00626, 2024. 1, 2, 3, 4, 5 +[27] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 2, 3, 5, 6 +[28] Jérôme Rony, Eric Granger, Marco Pedersoli, and Ismail Ben Ayed. Augmented lagrangian adversarial attacks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7738-7747, 2021. 3 +[29] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021. 2, 3, 5, 6 +[30] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278–25294, 2022. 1, 5 +[31] Erfan Shayegani, Yue Dong, and Nael Abu-Ghazaleh. Jailbreak in pieces: Compositional adversarial attacks on multimodal language models. arXiv preprint arXiv:2307.14539, 2023. 3 +[32] Bart Thomee, David A Shamma, Gerald Friedland, Benjamin Elizalde, Karl Ni, Douglas Poland, Damian Borth, and Li-Jia Li. Yfcc100m: The new data in multimedia research. Communications of the ACM, 59(2):64-73, 2016. 1 +[33] Maya Varma, Jean-Benoit Delbrouck, Zhihong Chen, Akshay Chaudhari, and Curtis Langlotz. Ravl: Discovering and mitigating spurious correlations in fine-tuned vision-language models. arXiv preprint arXiv:2411.04097, 2024. 1, 3 +[34] Qizhou Wang, Yong Lin, Yongqiang Chen, Ludwig Schmidt, Bo Han, and Tong Zhang. A sober look at the robustness of clips to spurious features. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 1, 3 +[35] Zheng Yuan, Jie Zhang, Yunpei Jia, Chuanqi Tan, Tao Xue, and Shiguang Shan. Meta gradient adversarial attack. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7748-7757, 2021. 3 +[36] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11975-11986, 2023. 2, 3, 6, 7 +[37] Kankan Zhou, Yibin LAI, and Jing Jiang. Vlstereoset: A study of stereotypical bias in pre-trained vision-language models. In Association for Computational Linguistics, 2022. 1 + +[38] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 2 +[39] Mingkang Zhu, Tianlong Chen, and Zhangyang Wang. Sparse and imperceptible adversarial attack via a homotopy algorithm. In International Conference on Machine Learning, pages 12868-12877. PMLR, 2021. 3 \ No newline at end of file diff --git a/ICCV/2025/Web Artifact Attacks Disrupt Vision Language Models/images.zip b/ICCV/2025/Web Artifact Attacks Disrupt Vision Language Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..72dde1f01117876a8b35a8960f1cccfd9bd80e44 --- /dev/null +++ b/ICCV/2025/Web Artifact Attacks Disrupt Vision Language Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:db917aeaca8c8385b6a932a35103636d1f3a658381e2d977e2e955ab49ee1d51 +size 371521 diff --git a/ICCV/2025/Web Artifact Attacks Disrupt Vision Language Models/layout.json b/ICCV/2025/Web Artifact Attacks Disrupt Vision Language Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..bafb1270728e05d38c951f018afefaf012c885d4 --- /dev/null +++ b/ICCV/2025/Web Artifact Attacks Disrupt Vision Language Models/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:94cdebae986ffa5e362aa50bbe086b7e7536c6f5b23c2132d53cd94f37203263 +size 347442 diff --git a/ICCV/2025/What Changed and What Could Have Changed_ State-Change Counterfactuals for Procedure-Aware Video Representation Learning/403d5bfc-2ad7-4b27-944d-e6bc63e5bcd5_content_list.json b/ICCV/2025/What Changed and What Could Have Changed_ State-Change Counterfactuals for Procedure-Aware Video Representation Learning/403d5bfc-2ad7-4b27-944d-e6bc63e5bcd5_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..d3fe8185d8129f578ff6f967c3d2a45ea051bbbc --- /dev/null +++ b/ICCV/2025/What Changed and What Could Have Changed_ State-Change Counterfactuals for Procedure-Aware Video Representation Learning/403d5bfc-2ad7-4b27-944d-e6bc63e5bcd5_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2148ede29fcf30228385a872a946a217986aefbf2d29e0a003437836533c0ccf +size 94792 diff --git a/ICCV/2025/What Changed and What Could Have Changed_ State-Change Counterfactuals for Procedure-Aware Video Representation Learning/403d5bfc-2ad7-4b27-944d-e6bc63e5bcd5_model.json b/ICCV/2025/What Changed and What Could Have Changed_ State-Change Counterfactuals for Procedure-Aware Video Representation Learning/403d5bfc-2ad7-4b27-944d-e6bc63e5bcd5_model.json new file mode 100644 index 0000000000000000000000000000000000000000..16cb307f559233fa4e674935a79e8bad030454ca --- /dev/null +++ b/ICCV/2025/What Changed and What Could Have Changed_ State-Change Counterfactuals for Procedure-Aware Video Representation Learning/403d5bfc-2ad7-4b27-944d-e6bc63e5bcd5_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dc05c1a4c23f9619e51ff10a99f114ba35c2fb49b02e02ed93f867383e44c17d +size 125605 diff --git a/ICCV/2025/What Changed and What Could Have Changed_ State-Change Counterfactuals for Procedure-Aware Video Representation Learning/403d5bfc-2ad7-4b27-944d-e6bc63e5bcd5_origin.pdf b/ICCV/2025/What Changed and What Could Have Changed_ State-Change Counterfactuals for Procedure-Aware Video Representation Learning/403d5bfc-2ad7-4b27-944d-e6bc63e5bcd5_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..50a417b804cecffb7bdb8b6922441b3773c0e804 --- /dev/null +++ b/ICCV/2025/What Changed and What Could Have Changed_ State-Change Counterfactuals for Procedure-Aware Video Representation Learning/403d5bfc-2ad7-4b27-944d-e6bc63e5bcd5_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e3fc2520f0f55f6a063f860a26d91aeb37e1bf31cd90082967033089d19e66bd +size 4143378 diff --git a/ICCV/2025/What Changed and What Could Have Changed_ State-Change Counterfactuals for Procedure-Aware Video Representation Learning/full.md b/ICCV/2025/What Changed and What Could Have Changed_ State-Change Counterfactuals for Procedure-Aware Video Representation Learning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..f85e52f6a89e850a6b09163d36b5c67b749f5ca9 --- /dev/null +++ b/ICCV/2025/What Changed and What Could Have Changed_ State-Change Counterfactuals for Procedure-Aware Video Representation Learning/full.md @@ -0,0 +1,377 @@ +# What Changed and What Could Have Changed? State-Change Counterfactuals for Procedure-Aware Video Representation Learning + +Chi-Hsi Kung* +Indiana University +kung@iu.edu + +Yi-Ting Chen† +National Yang-Ming Chiao-Tung University +ychen@cs.nycu.edu.tw + +Frangil Ramirez* +Indiana University +fraramir@iu.edu + +David Crandall† +Indiana University +djcran@iu.edu + +Juhyung Ha +Indiana University +yuhha@iu.edu + +Yi-Hsuan Tsai† +Atmanity Inc. +wasidennis@gmail.com + +![](images/c0c74c84c07197df15920065fd11316860eff48da80130e418fdc9f9bfe4d1c9.jpg) +Figure 1. Illustration of "what change" (actual action-induced transformations) and "what could have changed" (hypothetical deviations) for both clip- and video-level procedures. (a) The clip narration "Scoop oatmeal with a spoon and drop it into a bowl of water" changes the scene's Before and After states at the short-term clip-level. (b) Illustrates the hypothesized deviation, i.e., State-change Counterfactual, that could have happened and changed the scene differently at the clip level. (c) Illustrates the counterfactuals for video-level procedures, which depicts what could happen due to possible previous mistakes, for example, Missing-step Counterfactual and Misordered Counterfactual. + +# Abstract + +Understanding a procedural activity requires modeling both how action steps transform the scene, and how evolving scene transformations can influence the sequence of action steps, even those that are accidental or erroneous. Existing work has studied procedure-aware video representations by modeling the temporal order of actions, but has not explicitly learned the state changes (scene transformations). In this work, we study procedure-aware video representations learning by incorporating state-change descriptions generated by Large Language Models (LLMs) as supervision signals for video encoders. Moreover, we gener + +ate state-change counterfactuals that simulate hypothesized failure outcomes, allowing models to learn by imagining unseen "What if" scenarios. This counterfactual reasoning facilitates the model's ability to understand the cause and effect of each step in an activity. We conduct extensive experiments on procedure-aware tasks, including temporal action segmentation, error detection, action phase classification, frame retrieval, multi-instance retrieval, and action recognition. Our results demonstrate the effectiveness of the proposed state-change descriptions and their counterfactuals, and achieve significant improvements on multiple tasks. Code is available at https://github.com/HCISLab/counterfactual-video-pretrain. + +# 1. Introduction + +Procedural activities, such as following a cooking recipe [20, 24, 32, 37] or assembling a piece of furniture [49, 64], consist of sequences of interconnected steps that often happen in a specific and logical order to achieve a desired outcome. Understanding these activities from video data is essential for a variety of applications, including video retrieval [15, 22, 49, 84], intelligent collaborative agents [4, 27, 33, 79, 93], and robot learning from human demonstration [26, 30, 57, 65, 87, 89, 96, 105]. In contrast to general video action recognition that focuses on a single step in a short clip [7, 18, 21, 34, 77, 80], procedure-aware video understanding requires capturing both the "what changed" (actual action-induced state transformations) [41, 72], as shown in Figure 1 (a), and "what could have changed" (hypothetical deviations) [25, 58, 62, 101], as shown in Figures 1 (b) and (c). These two capabilities are important for understanding long-form procedures because the outcome of each action step can affect subsequent steps. + +Procedure-aware video representation learning can enable various downstream long-form video tasks, such as temporal action segmentation [20, 32, 37, 64] and error detection [37, 64]. There have been many approaches proposed to learn procedure-aware representations, including learning spatiotemporal features [7, 18, 21, 77], using action labels as supervision [43, 50, 86, 102], incorporating temporal order of steps [28, 102], and consulting external activity procedure knowledge databases [104] such as WikiiHow [31]. However, these approaches often struggle to explicitly model procedural knowledge that relies on how scene states evolve over time and how they could have evolved if the precondition and/or result of an action step had differed. For instance, the procedure "Make Oatmeal" in Figure 1 (c) cannot be finished correctly if the state of the bowl remains "Full of oatmeal and without water" or "Full of water without oatmeal" because neither state meets the precondition of the later action step, "Heat with microwave." Capturing these subtle transitions requires state-change reasoning to understand how an action transforms the environment and objects, and counterfactual analysis to foresee alternative outcomes for the same action. + +In this paper, we propose a hierarchical procedure-aware video representation learning framework that leverages state-changes and counterfactuals generated by a Large Language Model (LLM) [17]. We use the actual state-change descriptions to temporally contrast features, and the counterfactuals as the negative exemplars in contrastive learning [10, 29, 54] to enhance the modeling of clip-level and video-level procedure transformations. Specifically, at the clip level, we first model before and after states to describe the change in the environment and/or objects resulting from an action. We also incorporate state-change counterfactuals that estimate a hypothesized state if the action + +step were to fail for any reason. For example, if the action were "make coffee," a possible result state counterfactual could be "cup shards scattered on the floor." We then use temporal contrastive learning to encourage the model to learn a representation space in which visual features of later frames in short clips and after states are close together, and in which the early frames, before state, and state counterfactual are far away, and vice versa. At the video level, we also generate missing-step and misordered counterfactuals that simulate the hypothesized scenarios by disturbing action steps within a long video. Then, we treat these video-level counterfactuals as hard negatives in video-language alignment, enabling a more comprehensive understanding of the entire procedure in a long-form video. + +To validate our approach, we incorporate state changes and their counterfactuals into the learning process and pretrain our model on Ego4D [23, 42], a large-scale daily activity dataset filtered to include clean and aligned short clips along with their corresponding clip-level narration text and video-level summary text. We evaluate the learned procedure-aware video representations in six key procedural and short-term video understanding tasks—error detection, temporal action segmentation, multi-instance retrieval, action recognition, action phase classification, and frame retrieval—and show that they significantly enhance performance compared to strong baselines. In addition, we conduct a comprehensive ablation study to highlight the impact of learning state changes and their counterfactuals, demonstrating their importance in enhancing procedural video understanding. To summarize, our contributions are: + +1. a novel approach to video representation learning that explicitly incorporates state changes and their counterfactuals to enhance procedure awareness; +2. a hierarchical learning framework that integrates state-change and counterfactual descriptions into frame-, clip-, and video-level feature alignment; and +3. experiments showing state-of-the-art performance on several procedure-aware video downstream tasks and comprehensive analysis and discussion. + +# 2. Related Work + +# 2.1. Video Procedure-Aware Representations + +Procedure understanding, which involves recognizing and localizing individual action steps and reasoning about their causal relationships, is a fundamental challenge in computer vision, and is important for applications such as step recognition [5, 13, 76], temporal action segmentation [11, 20, 32, 64], and error detection [37]. Early work proposed learning procedure-aware video representations via exploiting the temporal consistency within videos [5, 55, 82, 86, 98] and temporal order [28] as pre-text tasks. Some work incorporates task graphs as guidance to learn procedure-aware rep + +resentations, either by constructing graphs based on knowledge bases [104] or through learning [3, 6, 48, 52, 63]. Recently, several papers have attempted to align video features in datasets such as HowTo100M [49] or Ego4D [23] with text descriptions extracted through automatic speech recognition [48, 50], refined subtitles [43] with the external database WikiHow [31, 48], or manually labeled annotations [2, 42, 102]. These text descriptions offer rich context of action steps and temporal relations, opening the door for procedure-awareness learning. For example, Brprompt [39] generates text prompts from ground truth action labels to learn ordinal temporal features. PVRL [102] formulates future action steps as next-token prediction [14], enhancing temporal order between actions. HierVL [2] proposes hierarchical video-text alignment for both short clips and long videos. + +However, these papers fall short of explicitly learning state changes—the causal relationships between action and state—which is essential to procedure understanding [53, 72]. Moreover, they can overfit to the seen correctly executed actions, which can hinder both understanding activity procedures and recognizing deviations such as erroneous action steps. In this work, we use state-change descriptions and counterfactuals to force our model to learn about action transformations and hypothetical failure outcomes, under the hypothesis that this type of learning will facilitate better procedure awareness. + +# 2.2. State-Change Understanding + +The idea of visual state changes has been explored extensively in the computer vision community, sometimes using different terms such as transformations between states [41, 83] or state-modifying actions [53, 70, 71, 73]. For example, papers have studied object state change localization [1, 45, 60, 70, 91] and generation [73], and action recognition with state changes [19, 60, 83]. Procedure planning, which aims to generate effective plans for task completion based on start and goal visual states, also incorporates the concept of understanding state changes [9, 40, 47, 51, 53, 78]. We were also inspired by SCHEMA [53] to use LLMs to generate before- and after-state descriptions of actions. + +However, three key factors sets our work apart from this existing work. First, unlike existing work that mainly focuses on a specific task such as procedure planning [53], we aim to design a more general video representation learning framework for various downstream tasks. Second, prior work does not incorporate hypothetical outcomes, i.e., counterfactuals, which we believe are essential for procedure causal reasoning. Last but not least, prior work often only models state changes involving objects that are directly interacted with. In contrast, our generated descriptions can cover state changes of objects, humans, and the surrounding environment. We provide extensive examples in our supple + +mentary materials. + +# 2.3. Video Representations with Counterfactuals + +Counterfactual example reasoning has been shown to help models better understand cause-and-effect and predict outcomes [61, 92, 97], yielding strong generalization and robustness on tasks such as video question-answering [38, 95, 100] and compositional reasoning [35, 75]. Extensive research has studied creating counterfactual examples with image data augmentation [103], rule-based text transformations [67], random word substitutions [59], language models [16, 81, 88, 99], and synthetic image-text pairs generated by a simulator [8] or generative model [35, 36, 66]. However, existing work mainly focuses on creating counterfactuals for images or short videos and does not consider the procedural nature of activities. In this work, we propose state-change counterfactuals at both the clip and video levels, highlighting both action transformations and procedure evolution. + +# 3. Proposed Method + +We first present the text generation process for state changes and state-change counterfactuals. Then, we provide an overview of the clip-video hierarchical representation learning framework. Finally, we present our pretraining optimization that incorporates the generated texts to learn procedure-aware representations. + +# 3.1. State Change & Counterfactual Generation + +We use the Llama 3 [17] LLM to generate clip-level state descriptions (before state, after state and state-change counterfactuals) and to generate video-level summary counterfactuals (missing step and misordered step). + +Clip-Level State Changes and their Counterfactuals. We feed the narration annotated for each short clip in Ego4D [23, 42] into Llama and ask it to generate a state-change description based on that narration. Specifically, + +Generate [Before] describing the scene before the action is performed and [After] describing the scene changed by the action. + +For example, given the narration “C picks a bag of clothes from the floor,” where $C$ is the camera-wearer, Llama might generate [Before]: “The floor is cluttered with clothes” and [After]: “The bag of clothes is now in $C$ 's hand, with the surrounding area slightly rearranged.” Then, we generate multiple state-change counterfactuals (SC-CF) using another prompt with the same action context, + +Now, based on the state changes you generated, create 3 state-change + +![](images/3541bac16a4ca1475daafc11e3d2ff038f9068b98ba14ba8c020c236f000080b.jpg) +Figure 2. Illustration of our learning framework. (a) We train frame features by temporally contrasting neighboring and distant frames by incorporating Before, After, and State-change Counterfactuals (SC-CF). (b) We align clip features with narration features. (c) Clip features are aggregated with the aggregator (Agg) as video features. Then the video features are contrasted positively with summaries and negatively with Missing-step Counterfactuals (K-CF) and Misordered Counterfactuals (M-CF). Note that all text features are extracted with the frozen text encoder. + +![](images/f9a297a5bcb4529cf14231319ea761ad9656d3bdfef0d1ecae71c370142ac5c9.jpg) + +counterfactuals [SC-CF] depicting scenes with incomplete or incorrectly completed actions. + +The generated results [SC-CF] might be "Clothes remain scattered on the floor" or "A small pile of clothes sits amidst remaining clutter." + +Video-Level Counterfactual Generation. We feed all clip narrations in a long video and their video-level summaries annotated in Ego4D [23, 42] to Llama. We ask the LLM to generate summary counterfactuals for missing key steps [K-CF] in which some crucial action steps (clip narrations) are missing, + +Generate 10 distinct counterfactuals [K-CF] by taking out some key actions. + +and misordered [M-CF] counterfactuals in which the order of key action steps is permuted, + +Generate 10 distinct counterfactuals [M-CF] by perturbing the order of actions. + +In this way, we generate diverse counterfactuals that can facilitate models to learn from hypothetical and unseen scenarios. As an example, consider the video summary "Make omelet with rice" in Figure 2. The procedure consists of a sequence of actions, ["Crack eggs," "Whisk eggs," "Fry eggs," "Fry rice", ...]. The video-level state-change counterfactuals could be missing action step "Whisk + +eggs", which might lead to state "Clumpy eggs", while if one misordered the action step "Fry rice" by performing it too early, it could lead to state "Burnt rice." We provide more examples of the generated descriptions in the supplementary materials. + +# 3.2. Preliminary: Hierarchical Video Feature Learning + +We adopt HierVL [2], a framework that learns hierarchical video-language representations at two temporal scales, clip-level and video-level. The clip-level narrations and video-level summaries are annotated by Ego4D/EgoClip [23, 42]. Note that since HierVL trains both a visual and text encoder, it requires vision-to-text and text-to-vision alignment. Below we describe the former, and the latter is defined symmetrically. Formally, let a video be divided into $M$ short clips $\{v_{1},\ldots ,v_{M}\}$ with corresponding narration texts $\{t_1,\dots ,t_M\}$ , and let $S$ be the video-level summary for the entire video. HierVL's vision-to-text training objective consists of two parts: clip-level alignment and video-level alignment losses. + +Clip-Level Alignment (Child Loss) is a contrastive loss that encourages matched $(v_{i},t_{i})$ pairs to have high similarity relative to mismatched pairs. Specifically, we use a softmax-based contrastive objective, + +$$ +\mathcal {L} _ {v 2 t} = \frac {- 1}{| B |} \sum_ {i \in B} \log \frac {\sum_ {p \in P (i)} \exp \left(v _ {i} ^ {T} t _ {p} / \tau\right)}{\sum_ {j \in B} \exp \left(v _ {i} ^ {T} t _ {j} / \tau\right)}, \tag {1} +$$ + +where $B$ is the mini-batch, $v_{i}$ and $t_j$ are visual and textual embeddings, respectively, and $P(i)$ denotes the positive samples of the $i^{\mathrm{th}}$ video clip. HierVL [2] leverages a variant of the action-aware loss, EgoNCE [42], in order to mine hard positives and negatives; please see [2, 42] for details. The overall loss at the child level is thus $\mathcal{L}_{\mathrm{child}} = \mathcal{L}_{v2t} + \mathcal{L}_{t2v}$ , where the latter is defined symmetrically. + +Video-Level Alignment (Parent Loss) captures long-range context. Clip features $v_{1}, \ldots, v_{m}$ and text features $t_{1}, \ldots, t_{m}$ are aggregated into single video-level representations $V_{i}$ and $T_{i}$ , respectively, + +$$ +V _ {i} = \operatorname {A g g} (v _ {1}, \dots , v _ {m}), +$$ + +$$ +T _ {i} = \operatorname {A g g} \left(t _ {1}, \dots , t _ {m}\right), +$$ + +where $\mathrm{Agg}(\cdot)$ is an aggregator function implemented with a self-attention transformer. This parent loss seeks to align the video-level features to the text summary provided by Ego4D, + +$$ +\mathcal {L} _ {\text {p a r e n t}} ^ {S V} = \frac {- 1}{| B |} \sum_ {i \in B} \log \frac {\sum_ {p \in P (i)} \exp \left(V _ {i} ^ {T} S _ {p} / \tau\right)}{\sum_ {j \in B} \exp \left(V _ {i} ^ {T} S _ {j} / \tau\right)}, \tag {2} +$$ + +where the superscript $SV$ emphasizes that this is the vision-to-text alignment loss, and $\mathcal{L}_{\text{parent}}^{ST}$ is defined analogously for text-to-text alignment. The overall loss at the parent level is thus $\mathcal{L}_{\text{parent}} = \mathcal{L}_{\text{parent}}^{SV} + \mathcal{L}_{\text{parent}}^{ST}$ . + +# 3.3. Pretraining Objective: State Change & Counterfactual + +We use a hierarchical contrastive pre-training framework with three timescales: frame-, clip-, and video-level. + +Frame-Level Alignment. At the frame level, the model is supervised by our proposed before-state loss $\mathcal{L}_{\mathrm{before}}$ and after-state loss $\mathcal{L}_{\mathrm{after}}$ , each defined as [29], + +$$ +\mathcal {L} = \frac {1}{| B |} \sum_ {i \in B} \frac {- 1}{| P (i) |} \sum_ {z _ {p} \in P (i)} \log \frac {\exp \left(f _ {i} ^ {T} z _ {p} / \tau\right)}{\sum_ {z _ {n} \in N (i)} \exp \left(f _ {i} ^ {T} z _ {n} / \tau\right)}, \tag {3} +$$ + +where $B$ is the batch size, $f_{i}$ is the visual embedding of the $i^{\mathrm{th}}$ frame, $z_{j}$ is either a visual or text embedding, $\tau$ is a temperature hyperparameter, and $P(i)$ and $N(i)$ denote the positive and negative samples of the $i^{\mathrm{th}}$ frame, respectively. Given a set of $K = 4$ frames sub-sampled from a video clip, $L_{\mathrm{before}}$ aims to align earlier-in-time frames along with the before-state text embeddings, while pushing them apart from later-in-time frames, the after-state, and counterfactual text embeddings. That is, $P(0) = \{f_1,t_b\}$ and $N(0) = \{f_3,t_a,t_{cf}\}$ , where $t_b$ , $t_a$ , and $t_{cf}$ denote the text embeddings corresponding to the before state, after state, and the counterfactuals, respectively. On the other hand, $L_{\mathrm{after}}$ aims to align later-in-time frames to the after + +state while separating them from earlier-in-time frames, the before-state, and counterfactual text embeddings. In other words, $P(3) = \{f_2, t_a\}$ and $N(3) = \{f_0, t_b, t_{cf}\}$ . Additionally, other samples from the same mini-batch are included as negatives in the denominator of Eq. (3) in both losses; we omit them in the notation above for simplicity. + +Clip-Level Alignment. At the clip-level, we use the $L_{v2t}$ loss described in Section 3.2, which seeks to align the videoclip embeddings to their corresponding text narrations from EgoClip [42]. Note that since we do not train a text encoder, the symmetric $L_{t2v}$ loss is neglected here. For more details on this loss, see Section 3.2 and [2, 42]. + +The resulting loss for the first two scales is thus, + +$$ +\mathcal {L} _ {\text {c h i l d}} = \mathcal {L} _ {v 2 t} + \lambda \left(\mathcal {L} _ {\text {b e f o r e}} + \mathcal {L} _ {\text {a f t e r}}\right), \tag {4} +$$ + +where $\lambda$ is a hyperparameter controlling the strength of the state-change aware supervision. + +Video-Level Alignment. The goal of this loss is to align video-level visual embeddings to summary text embeddings and to enhance procedural awareness using video-level counterfactuals. We first obtain video-level visual embeddings from clip-level embeddings using a self-attention block, as described in Section 3.2. Then, using contrastive learning, each visual embedding is aligned to its corresponding summary text embedding and contrasted against text embeddings of the generated counterfactual K-CF and M-CF. The formulation is defined as, + +$$ +\mathcal {L} _ {\text {p a r e n t}} = - \sum_ {i \in B} \log \frac {\sum_ {p \in P (i)} \exp \left(V _ {i} ^ {T} S _ {p}\right)}{\sum_ {n \in N (i)} \exp \left(V _ {i} ^ {T} S _ {n}\right) + \exp \left(V _ {i} ^ {T} S _ {n , w} ^ {c f}\right)}, \tag {5} +$$ + +where $V_{i}$ is the aggregated video-level visual embedding, $S_{j}$ is a summary text embedding and $S_{j,w}^{cf}$ are Misordered and Missing-step counterfactual text embeddings where $w \in \{1,\dots ,W\}$ is the total number of counterfactuals used, and $P(i)$ and $N(i)$ denote the positive and negative samples of the $i^{\text{th}}$ video, respectively. We also use a temperature hyperparameter $\tau$ as in Eq. (3), but we omit it in the equation above for simplicity of notation. Note that consistent with [2], in this case, the summation over positive samples is located inside the logarithm since the sampling strategy of [42] is used. However, prior work [29] finds the summation outside the logarithm to be more effective and thus this is the de-facto choice of the frame-level loss in Eq. (3). + +# 4. Experiments + +# 4.1. Implementation Details + +Pretraining. We pretrain our model on EgoClip [42], a dataset derived from Ego4D [23], with clean clip-level sentences and the video-level summaries. We follow [2, 42] + +Table 1. Temporal action segmentation results on the GTEA and EgoPER datasets. Bold denotes best and underline is second best. Text encoder denotes the VLM model with a trainable text encoder. Following [44] we also present the average across all metrics (Avg). + +
MethodPretraining DataGTEAEgoPER
F1@10F1@25F1@50EditAccAvgF1@50EditAccAvg
Br-prompt [39]GTEA94.192.083.091.681.288.4----
I3D [7]Kinetics90.188.879.284.679.784.548.871.973.964.9
CLIP [56]WIT [74]+Text encoder88.586.277.687.175.683.044.271.270.862.1
MIL-NCE [50]HowTo100M67.961.344.667.958.360.047.369.173.663.3
PVRL [102]HowTo100M85.282.672.281.171.278.545.673.273.464.1
HierVL [2]Ego4D+Text encoder90.488.581.286.778.585.152.673.077.367.6
OursEgo4D89.889.181.686.880.085.554.474.179.069.2
+ +to use the same training split that has 3.8M clip narrations and 120K long-term video summaries for pretraining. We pretrain our model on 8 NVIDIA L40S GPUs, with batch size 18 on each GPU for 7 epochs. We follow HierVL [2] and train models on clip- and video-level alignment alternatively. Specifically, we perform video-level alignment as in Eq. (5) after every 5 mini-batches of clip-level alignment as in Eq. (4). The frame-level loss is computed during the clip-level training iteration. We set the learning rate and weight decay to 1e-5 and 1e-4, respectively. Note that each short clip is sub-sampled to 4 frames during training [2, 42]. + +Text Generation. To generate all state-change descriptions and state-change counterfactuals, we use Llama 3.1 8B [17] for efficiency. We incorporate off-the-shelf text-encoder FLAVA [69] to pre-extract features of the annotated sentences and summaries and our generated text descriptions. + +Temporal Action Segmentation. We evaluate representations on temporal action segmentation, in which the goal is to output frame-wise action labels for untrimmed videos. This requires modeling procedures and recognizing the start and end of state transformations. We evaluate various representations by adopting the popular temporal action segmentation model ASFormer [94] that takes video features as input. We follow prior work [39] for the same training strategy and evaluation protocol and metrics including F1 scores, Edit distance, and frame-wise Accuracy, on the cooking procedure datasets, Georgia Tech Egocentric Activities (GTEA) [20] and EgoPER [37] Dataset. GTEA consists of 28 egocentric instructional videos capturing daily kitchen activities at 15 frames per second. EgoPER contains 368 videos spanning 28 hours of cooking scenarios. + +Error Detection. We evaluate representations on unsupervised error detection on the EgoPER dataset [37] where models are trained with error-free videos and output labels for each frame on whether it is an error. The error categories include step omission, step addition, step modification, step slip, and step correction. We follow [37], and use the EgoPED model [37] and equip it with differ + +ent video representations. Specifically, the training process of EgoPED consists of two stages: action segmentation and prototype learning. In stage 1, the temporal action segmentation model, ASFormer, is trained to predict frame-wise action labels on the error-free set by taking pre-extracted video features as input. In stage 2, the model learns prototype features for each class of action step, e.g., "Take bowl from microwave," with contrastive step prototype learning [37, 54]. For each class of activity, multiple prototypes are extracted for error detection at inference. Note that each activity is trained independently, such as Make Coffee. During the inference, EgoPED calculates the similarity score of observed frame features and any learned prototypes. A threshold is further used to determine whether the observed test frame is erroneous. + +Action Retrieval & Recognition. To demonstrate the applicability of the learned representations in core video understanding tasks, we evaluate zero-shot multi-instance retrieval on Epic-Kitchens (EK) [12] and zero-shot action recognition on Charades-Ego (CE) [68]. Following prior work [2, 42], both tasks are implemented via text-to-vision alignment by selecting the textual label with the highest similarity score to the given visual sample. Note that [42] reports overfitting when applying the model trained on Ego4D to Charades-Ego, and thus uses a task-specific checkpoint. We follow [2] and report the results of our pre-training checkpoint (denoted as PT ckpt in their work) rather than using task-specific checkpoints. + +Action Phase Classification & Frame Retrieval. To assess fine-grained and short-term procedure awareness, we test representations on action phase classification and zero-shot frame retrieval on the Align-Ego-Exo (AE2) dataset [90]. The AE2 dataset contains egocentric and exocentric videos of actions Break Eggs, Pour Milk, Pour Liquid, and Tennis Forehand. Each action is divided into two to four phases [90]. We follow [90] and train an SVM to predict the per-frame action phase labels for the classification task and use nearest neighbors for the zero-shot retrieval + +Table 2. Error detection results on the EgoPER dataset. Bold indicates best and underline is second best. HTM denotes the HowTo100M dataset [49]. Text denotes the VLM model with a trainable text encoder. + +
QuesadillaOatmealPinwheelCoffeeTeaAll
MethodPretraining DataEDAAUCEDAAUCEDAAUCEDAAUCEDAAUCEDAAUC
Random-19.950.011.850.015.750.08.2050.017.050.014.550.0
HF2-VAD [46]-34.562.625.462.329.152.710.059.636.662.127.159.9
S3R [85]Kinetics+I3D52.651.847.861.650.552.416.351.047.857.943.054.9
I3D [7]Kinetics62.765.651.465.159.655.055.358.356.066.057.062.0
CLIP [56]WIT [74]+Text77.667.269.667.566.959.568.569.075.657.771.664.2
MIL-NCE [50]HTM77.359.869.861.565.753.068.067.969.861.570.160.7
PVRL [102]HTM75.770.071.253.565.565.567.565.476.465.271.363.9
HierVL [2]Ego4D+Text77.970.270.866.465.258.167.469.875.166.471.366.2
OursEgo4D78.963.771.646.168.353.868.368.076.670.972.760.5
+ +task. Note that we merge the validation and test sets for more robust results. + +Baselines. We evaluate the popular I3D and CLIP feature [7, 56] commonly used in long-form video tasks and procedure-aware representations with publicly available pretrained model weights, including MIL-NCE [50], Bridge-prompt [39], and PVRL [102]. In addition, we evaluate the VLM HierVL [2] with only its video encoder. + +# 4.2. Temporal Action Segmentation Results + +Table 1 presents results on temporal action segmentation. The first line of the table shows Br-prompt [39] which is pre-trained on the target dataset (GTEA) and thus serves as an upper bound on performance. In our results, we found that MIL-NCE performs significantly worse than others, especially on GTEA. We assume this is because MIL-NCE is trained with low-quality automatic-speech-recognition (ASR) text and imprecise alignment between video and ASR sentences, unlike PVRL [2] which leverages pseudo-labels generated by CLIP [56]. We further compare our model with the VLM HierVL. Even though HierVL is trained with both video and text encoders, which require considerably more computational resources, our model outperforms HierVL in most metrics on both datasets. Furthermore, the performance gap against all non-VLM models is larger. This highlights the effectiveness of our proposed state-change descriptions and their counterfactuals on long-term procedure understanding. + +# 4.3. Error Detection Results + +Table 2 presents results of the error detection benchmark on EgoPER [37]. We first include results reported in EgoPER as a reference in the top group (rows 1-3) instead of a direct comparison. In the remaining groups (rows 4-9), we report the performance of representation learning approaches, tested using the same downstream architecture [94]. We ob + +Table 3. Multi-Instance Retrieval (MIR) on the Epic-Kitchens (EK) dataset [12] and Zero-shot Action Recognition on the Charades-Ego (CE) dataset [68]. + +
MethodMIR (EK)Action Rec. (CE)
mAPnDCGmAP
CLIP [56]8.415.320.5
MIL-NCE [50]5.810.36.9
PVRL [102]6.111.47.9
Ours15.722.624.8
+ +serve that all procedure-aware representations outperform general visual representations, such as I3D and CLIP, highlighting the importance of procedure awareness in the context of error detection. Furthermore, our proposed method surpasses the state-of-the-art in EDA [37], even outperforming the VLM HierVL. + +Moreover, we achieve competitive performance on the AUC metric. Note that although we follow prior work [37] to include frame-wise metric AUC, this metric may not be effective in measuring the performance and the procedure awareness of representations; as discussed in EgoPER [37], a heuristic random method can achieve competitive results in AUC but fails to localize erroneous segments according to EDA. This is also true when comparing I3D and earlier methods, such as $\mathrm{HF}^2$ -VAD [46] and S3R [85]. We hope our extensive experimental results can inspire future work to study more effective evaluation metrics for error detection. In summary, our representation learning method achieves state-of-the-art performance in EDA and competitive performance in AUC, highlighting the strong effectiveness of counterfactual reasoning in procedure awareness. + +# 4.4. Action Retrieval & Recognition Results + +Table 3 presents benchmark results on action retrieval and recognition. We found existing procedure-aware represent + +Table 4. Action Phase Classification and zero-shot Frame Retrieval results on the Align-Ego-Exo dataset [90]. + +
MethodClassification (F1)Retrieval (mAP)
ego+exoegoego+exoego
CLIP [56]59.561.664.468.0
MIL-NCE [50]53.054.259.563.0
PVRL [102]59.462.761.666.6
Ours61.364.864.970.3
+ +Table 5. Ablation study involving: Before & After states, State-change Counterfactuals (SC-CF), Missing-step (K-CF) and Mis-ordered (M-CF) on GTEA for temporal action segmentation and EgoPED for error detection. Note that we calculate EDA by averaging all categories of activity. Metrics: Acc. and EDA; mean ± sd of 5 and 10 runs, respectively. + +
IDBefore & After SC-CF K-CF M-CFGTEA (AS)EgoPER (ED)
1 (Weak HierVL)77.3±.671.2±.5
2✓78.4±.472.1±.6
3✓✓79.1±.271.5±.8
4✓✓✓78.7±.772.2±.4
5✓✓✓78.8±.872.2±.5
Ours✓✓✓✓79.6±.472.3±.5
+ +tations perform significantly worse on short-term action recognition tasks, while our representations trained with both short- and long-term state-changes and counterfactuals work well on popular video recognition tasks and datasets. + +# 4.5. Action Phase Classification & Retrieval Results + +Table 4 shows that our model outperforms others on average across actions on both the ego+exo and ego settings, demonstrating its effectiveness in short-term procedures and strong view-point generalization. Interestingly, CLIP [56] demonstrates strong performance. We conjecture this is due to the focus on short-term action in its image-to-text alignment. Full results with individual actions are available in the supplementary material. + +# 4.6. Ablation Study + +In Table 5, we study the effectiveness of state changes (Before and After), state-change counterfactuals (SC-CF), and video-level state-change counterfactuals (Missing-step CF and Misordered-step CF) by ablating each of them and evaluating on temporal action segmentation with GTEA and error detection with EgoPER. Note that the first row (ID 1) can be seen as Weak HierVL as it is trained with only the video encoder. In addition, since Weak HierVL only uses narrations from other videos in a batch as negatives, it can be seen as a random description approach. All ablations (including Ours) are pre-trained with a batch size of 12 on each GPU. + +Temporal Action Segmentation. One interesting insight + +on GTEA is that using only clip-level state changes and their counterfactuals (ID 3) achieves competitive performance, outperforming most other ablated models. Also, using only one type of video-level counterfactual (ID 4 and 5) may mislead the model to overfit to a certain type of counterfactual, resulting in lower performance. By considering all the proposed states and counterfactuals together, we show that they complement each other, thus achieving the best performance. + +Error Detection. The results on EgoPER show that our proposed state changes and their counterfactuals can improve performance by combining them all together, verifying the effectiveness of recognizing erroneous labels. In addition, we observe interesting results showing that with only clip-level counterfactuals (ID 3), the model performs slightly worse than its baseline (ID 2). We conjecture this is because the model overfits to short-term procedures and overlooks long-term activity procedures and their counterfactual reasoning. + +# 5. Conclusions + +In this work, we present a novel procedure-aware video representation learning framework that first incorporates state-change descriptions and state-change counterfactuals in clip-level alignment, enhancing causal reasoning of action transformations. Then, it uses video-level counterfactuals that perturb the local actions and create hypothesized scenarios to facilitate the understanding of activity procedures. Our learned representations demonstrate strong effectiveness in terms of procedure awareness and achieve state-of-the-art results on several benchmarks. For future work, a promising direction may be to generate multimodal counterfactual examples for procedure-aware activities, such as synthetic long-form videos showcasing mistakes and the corresponding corrections. + +# 6. Acknowledgment + +Authors CK, FR, JH, and DC were supported in part by the National Science Foundation under award DRL-2112635 to the AI Institute for Engaged Learning and by the IU Luddy Artificial Intelligence Center. YC was supported in part by the National Science and Technology Council under grants 113-2628-E-A49 -022 - and 114-2628-E-A49 -007 -, by the Higher Education Sprout Project of National Yang Ming Chiao Tung University, and by the Ministry of Education through the Yushan Fellow Program Administrative Support Grant. Computation resources were supported in part by the Lilly Endowment, Inc. through its support to the IU Pervasive Technology Institute. + +# References + +[1] Jean-Baptiste Alayrac, Ivan Laptev, Josef Sivic, and Simon Lacoste-Julien. Joint discovery of object states and manipulation actions. In Proceedings of the IEEE International Conference on Computer Vision, pages 2127-2136, 2017. 3 +[2] Kumar Ashutosh, Rohit Girdhar, Lorenzo Torresani, and Kristen Grauman. Hiervl: Learning hierarchical videolanguage embeddings. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23066-23078, 2023. 3, 4, 5, 6, 7 +[3] Kumar Ashutosh, Santhosh Kumar Ramakrishnan, Triantafyllos Afouras, and Kristen Grauman. Video-mined task graphs for keystep recognition in instructional videos. Advances in Neural Information Processing Systems, 36, 2024. 3 +[4] Rohan Banerjee, Rajat Kumar Jenamani, Sidharth Vasudev, Amal Nanavati, Katherine Dimitropoulou, Sarah Dean, and Tapomayukh Bhattacharjee. To ask or not to ask: Human-in-the-loop contextual bandits with applications in robot-assisted feeding. arXiv preprint arXiv:2405.06908, 2024. 2 +[5] Siddhant Bansal, Chetan Arora, and CV Jawahar. My view is the best view: Procedure learning from egocentric videos. In European Conference on Computer Vision, pages 657-675. Springer, 2022. 2 +[6] Siddhant Bansal, Chetan Arora, and CV Jawahar. United we stand, divided we fall: Unitygraph for unsupervised procedure learning from videos. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 6509-6519, 2024. 3 +[7] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6299-6308, 2017. 2, 6, 7 +[8] Paola Cascante-Bonilla, Khaled Shehada, James Seale Smith, Sivan Doveh, Donghyun Kim, Rameswar Panda, Gul Varol, Aude Oliva, Vicente Ordonez, Rogerio Feris, et al. Going beyond nouns with vision & language models using synthetic data. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20155-20165, 2023. 3 +[9] Chien-Yi Chang, De-An Huang, Danfei Xu, Ehsan Adeli, Li Fei-Fei, and Juan Carlos Niebles. Procedure planning in instructional videos. In European Conference on Computer Vision, pages 334–350. Springer, 2020. 3 +[10] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PmLR, 2020. 2 +[11] Zihao Chen, Hsuanyu Wu, Chi-Hsi Kung, Yi-Ting Chen, and Yan-Tsung Peng. Atars: An aerial traffic atomic activity recognition and temporal segmentation dataset. arXiv preprint arXiv:2503.18553, 2025. 2 +[12] Dima Damen, Hazel Doughty, Giovanni Maria Farinella, Sanja Fidler, Antonino Furnari, Evangelos Kazakos, Davide Moltisanti, Jonathan Munro, Toby Perrett, Will Price, + +and Michael Wray. Scaling egocentric vision: The epic-kitchens dataset. In European Conference on Computer Vision (ECCV), 2018. 6, 7 +[13] Dima Damen, Hazel Doughty, Giovanni Maria Farinella, Sanja Fidler, Antonino Furnari, Evangelos Kazakos, Davide Moltisanti, Jonathan Munro, Toby Perrett, Will Price, et al. The epic-kitchen dataset: Collection, challenges and baselines. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(11):4125-4141, 2020. 2 +[14] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 conference of the North American chapter of the association for computational linguistics: human language technologies, volume 1 (long and short papers), pages 4171–4186, 2019. 3 +[15] Jianfeng Dong, Xirong Li, Chaoxi Xu, Shouling Ji, Yuan He, Gang Yang, and Xun Wang. Dual encoding for zeroexample video retrieval. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9346-9355, 2019. 2 +[16] Sivan Doveh, Assaf Arbelle, Sivan Harary, Eli Schwartz, Roei Herzig, Raja Giryes, Rogerio Feris, Rameswar Panda, Shimon Ullman, and Leonid Karlinsky. Teaching structured vision & language concepts to vision & language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2657-2668, 2023. 3 +[17] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 2, 3, 6 +[18] Haoqi Fan, Bo Xiong, Karttikeya Mangalam, Yanghao Li, Zhicheng Yan, Jitendra Malik, and Christoph Feichtenhofer. Multiscale vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6824-6835, 2021. 2 +[19] Alireza Fathi and James M Rehg. Modeling actions through state changes. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2579-2586, 2013. 3 +[20] Alireza Fathi, Xiaofeng Ren, and James M. Rehg. Learning to recognize objects in egocentric activities. In CVPR 2011, pages 3281-3288, 2011. 2, 6 +[21] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6202-6211, 2019. 2 +[22] Valentin Gabeur, Chen Sun, Karteek Alahari, and Cordelia Schmid. Multi-modal transformer for video retrieval. In European Conference on Computer Vision, pages 214-229. Springer, 2020. 2 +[23] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, et al. Ego4d: Around the world in 3,000 hours of egocentric video. In Proceedings of the IEEE/CVF Conference on Computer Vi + +sion and Pattern Recognition, pages 18995-19012, 2022. 2,3,4,5 +[24] Kristen Grauman, Andrew Westbury, Lorenzo Torresani, Kris Kitani, Jitendra Malik, Triantafyllos Afouras, Kumar Ashutosh, Vijay Baiyya, Siddhant Bansal, Bikram Boote, et al. Ego-exo4d: Understanding skilled human activity from first-and third-person perspectives. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19383-19400, 2024. 2 +[25] Paul L. Harris, Tim German, and Patrick Mills. Children's use of counterfactual thinking in causal reasoning. Cognition, 61(3):233-259, 1996. 2 +[26] Vidhi Jain, Maria Attarian, Nikhil J Joshi, Ayzaan Wahid, Danny Driess, Quan Vuong, Pannag R Sanketi, Pierre Sermanet, Stefan Welker, Christine Chan, et al. Vid2robot: End-to-end video-conditioned policy learning with cross-attention transformers. arXiv preprint arXiv:2403.12943, 2024. 2 +[27] Rajat Kumar Jenamani, Daniel Stabile, Ziang Liu, Abrar Anwar, Katherine Dimitropoulou, and Tapomayukh Bhattacharjee. Feel the bite: Robot-assisted inside-mouth bite transfer using robust mouth perception and physical interaction-aware control. In Proceedings of the 2024 ACM/IEEE International Conference on Human-Robot Interaction, pages 313-322, 2024. 2 +[28] Simon Jenni, Givi Meishvili, and Paolo Favaro. Video representation learning by recognizing temporal transformations. In European Conference on Computer Vision, pages 425-442. Springer, 2020. 2 +[29] Prannay Khosla, Piotr Teterwak, Chen Wang, Aaron Sarna, Yonglong Tian, Phillip Isola, Aaron Maschinot, Ce Liu, and Dilip Krishnan. Supervised contrastive learning. In Proceedings of the 34th International Conference on Neural Information Processing Systems, Red Hook, NY, USA, 2020. Curran Associates Inc. 2, 5 +[30] Hema S. Koppula and Ashutosh Saxena. Anticipating human activities using object affordances for reactive robotic response. IEEE Transactions on Pattern Analysis and Machine Intelligence, 38(1):14-29, 2016. 2 +[31] Mahnaz Koupae and William Yang Wang. Wikihow: A large scale text summarization dataset. arXiv preprint arXiv:1810.09305, 2018. 2, 3 +[32] Hilde Kuehne, Ali Arslan, and Thomas Serre. The language of actions: Recovering the syntax and semantics of goal-directed human activities. In 2014 IEEE Conference on Computer Vision and Pattern Recognition, pages 780-787, 2014. 2 +[33] Chi-Hsi Kung, Chieh-Chi Yang, Pang-Yuan Pao, Shu-Wei Lu, Pin-Lun Chen, Hsin-Cheng Lu, and Yi-Ting Chen. Riskbench: A scenario-based benchmark for risk identification. arXiv preprint arXiv:2312.01659, 2023. 2 +[34] Chi-Hsi Kung, Shu-Wei Lu, Yi-Hsuan Tsai, and Yi-Ting Chen. Action-slot: Visual action-centric representations for multi-label atomic activity recognition in traffic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18451-18461, 2024. 2 + +[35] Chengen Lai, Shengli Song, Sitong Yan, and Guangneng Hu. Improving vision and language concepts understanding with multimodal counterfactual samples. In European Conference on Computer Vision, pages 174-191. Springer, 2024. 3 +[36] Tiep Le, Vasudev Lal, and Phillip Howard. Cocounterfactuals: Automatically constructed counterfactual examples for image-text pairs. Advances in Neural Information Processing Systems, 36:71195-71221, 2023. 3 +[37] Shih-Po Lee, Zijia Lu, Zekun Zhang, Minh Hoai, and Ehsan Elhamifar. Error detection in egocentric procedural task videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18655-18666, 2024. 2, 6, 7 +[38] Jiangtong Li, Li Niu, and Liqing Zhang. From representation to reasoning: Towards both evidence and commonsense reasoning for video question-answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 21273–21282, 2022. 3 +[39] Muheng Li, Lei Chen, Yueqi Duan, Zhilan Hu, Jianjiang Feng, Jie Zhou, and Jiwen Lu. Bridge-prompt: Towards ordinal action understanding in instructional videos. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 19880-19889, 2022. 3, 6, 7 +[40] Zhiheng Li, Wenjia Geng, Muheng Li, Lei Chen, Yansong Tang, Jiwen Lu, and Jie Zhou. Skip-plan: Procedure planning in instructional videos via condensed action space learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10297-10306, 2023. 3 +[41] Chen Liang, Wenguan Wang, Tianfei Zhou, and Yi Yang. Visual abductive reasoning. In IEEE/CVF International Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 3 +[42] Kevin Qinghong Lin, Alex Jinpeng Wang, Mattia Soldan, Michael Wray, Rui Yan, Eric Zhongcong Xu, Difei Gao, Rongcheng Tu, Wenzhe Zhao, Weijie Kong, et al. Egocentric video-language pretraining. arXiv preprint arXiv:2206.01670, 2022. 2, 3, 4, 5, 6 +[43] Xudong Lin, Fabio Petroni, Gedas Bertasius, Marcus Rohrbach, Shih-Fu Chang, and Lorenzo Torresani. Learning to recognize procedural activities with distant supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13853-13863, 2022. 2, 3 +[44] Daochang Liu, Qiyue Li, Anh-Dung Dinh, Tingting Jiang, Mubarak Shah, and Chang Xu. Diffusion action segmentation. In 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 10105-10115, 2023. 6 +[45] Yang Liu, Ping Wei, and Song-Chun Zhu. Jointly recognizing object fluents and tasks in egocentric videos. In Proceedings of the IEEE International Conference on Computer Vision, pages 2924-2932, 2017. 3 +[46] Zhian Liu, Yongwei Nie, Chengjiang Long, Qing Zhang, and Guiqing Li. A hybrid video anomaly detection framework via memory-augmented flow reconstruction and flow-guided frame prediction. In Proceedings of the IEEE/CVF + +international conference on computer vision, pages 13588-13597, 2021. 7 +[47] Yujie Lu, Weixi Feng, Wanrong Zhu, Wenda Xu, Xin Eric Wang, Miguel Eckstein, and William Yang Wang. Neurosymbolic procedural planning with commonsense prompting. arXiv preprint arXiv:2206.02928, 2022. 3 +[48] Effrosyni Mavroudi, Triantafyllos Afouras, and Lorenzo Torresani. Learning to ground instructional articles in videos through narrations. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15201-15213, 2023. 3 +[49] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. Howto100m: Learning a text-video embedding by watching hundred million narrated video clips. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2630-2640, 2019. 2, 3, 7 +[50] Antoine Miech, Jean-Baptiste Alayrac, Lucas Smaira, Ivan Laptev, Josef Sivic, and Andrew Zisserman. End-to-end learning of visual representations from uncurated instructional videos. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9879-9889, 2020. 2, 3, 6, 7, 8 +[51] Kumaranage Ravindu Yasas Nagasinghe, Honglu Zhou, Malitha Gunawardhana, Martin Renqiang Min, Daniel Harari, and Muhammad Haris Khan. Why not use your textbook? knowledge-enhanced procedure planning of instructional videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18816-18826, 2024. 3 +[52] Medhini Narasimhan, Licheng Yu, Sean Bell, Ning Zhang, and Trevor Darrell. Learning and verification of task structure in instructional videos. arXiv preprint arXiv:2303.13519, 2023. 3 +[53] Yulei Niu, Wenliang Guo, Long Chen, Xudong Lin, and Shih-Fu Chang. Schema: State changes matter for procedure planning in instructional videos. arXiv preprint arXiv:2403.01599, 2024. 3 +[54] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 2, 6 +[55] Zhiwu Qing, Shiwei Zhang, Ziyuan Huang, Yi Xu, Xiang Wang, Mingqian Tang, Changxin Gao, Rong Jin, and Nong Sang. Learning from untrimmed videos: Self-supervised video representation learning with hierarchical consistency. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13821-13831, 2022. 2 +[56] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PmLR, 2021. 6, 7, 8 +[57] Juntao Ren, Priya Sundaresan, Dorsa Sadigh, Sanjiban Choudhury, and Jeannette Bohg. Motion tracks: A unified representation for human-robot transfer in few-shot imitation learning. arXiv preprint arXiv:2501.06994, 2025. 2 + +[58] Neal J Roese. Counterfactual thinking. Psychological bulletin, 121(1):133, 1997. 2 +[59] Karsten Roth, Jae Myung Kim, A Koepke, Oriol Vinyals, Cordelia Schmid, and Zeynep Akata. Waffling around for performance: Visual classification with random words and broad concepts. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15746-15757, 2023. 3 +[60] Nirat Saini, Hanyu Wang, Archana Swaminathan, Vinoj Jayasundara, Bo He, Kamal Gupta, and Abhinav Shrivastava. Chop & learn: Recognizing and generating object-state compositions. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20247-20258, 2023. 3 +[61] Bernhard Schölkopf, Francesco Locatello, Stefan Bauer, Nan Rosemary Ke, Nal Kalchbrenner, Anirudh Goyal, and Yoshua Bengio. Toward causal representation learning. Proceedings of the IEEE, 109(5):612-634, 2021. 3 +[62] Peter Schulam and Suchi Saria. Reliable decision support using counterfactual models. Advances in neural information processing systems, 30, 2017. 2 +[63] Luigi Seminara, Giovanni Maria Farinella, and Antonino Furnari. Differentiable task graph learning: Procedural activity representation and online mistake detection from egocentric videos. arXiv preprint arXiv:2406.01486, 2024. 3 +[64] Fadime Sener, Dibyadip Chatterjee, Daniel Shelepop, Kun He, Dipika Singhania, Robert Wang, and Angela Yao. Assembly101: A large-scale multi-view video dataset for understanding procedural activities. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21096-21106, 2022. 2 +[65] Lin Shao, Toki Migimatsu, Qiang Zhang, Karen Yang, and Jeannette Bohg. Concept2robot: Learning manipulation concepts from instructions and human demonstrations. The International Journal of Robotics Research, 40(12-14): 1419-1434, 2021. 2 +[66] Cheng Shi and Sibei Yang. Logoprompt: Synthetic text images can be good visual prompts for vision-language models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2932-2941, 2023. 3 +[67] Haoyue Shi, Jiayuan Mao, Tete Xiao, Yuning Jiang, and Jian Sun. Learning visually-grounded semantics from contrastive adversarial samples. arXiv preprint arXiv:1806.10348, 2018. 3 +[68] Gunnar A. Sigurdsson, Abhinav Gupta, Cordelia Schmid, Ali Farhadi, and Karteek Alahari. Charades-ego: A large-scale dataset of paired third and first person videos, 2018. 6, 7 +[69] Amanpreet Singh, Ronghang Hu, Vedanuj Goswami, Guillaume Couairon, Wojciech Galuba, Marcus Rohrbach, and Douwe Kiela. Flava: A foundational language and vision alignment model. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15638-15650, 2022. 6 +[70] Tomáš Souček, Jean-Baptiste Alayrac, Antoine Miech, Ivan Laptev, and Josef Sivic. Look for the change: Learning object states and state-modifying actions from untrimmed + +web videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13956-13966, 2022. 3 +[71] Tomáš Souček, Jean-Baptiste Alayrac, Antoine Miech, Ivan Laptev, and Josef Sivic. Multi-task learning of object state changes from uncurated videos. arXiv preprint arXiv:2211.13500, 2022. 3 +[72] Tomás Souček, Dima Damen, Michael Wray, Ivan Laptev, and Josef Sivic. Genhowto: Learning to generate actions and state transformations from instructional videos. Proceedings of the IEEE/CVF Conference781 on Computer Vision and Pattern Recognition, 2024. 2, 3 +[73] Tomáš Souček, Jean-Baptiste Alayrac, Antoine Miech, Ivan Laptev, and Josef Sivic. Multi-task learning of object states and state-modifying actions from web videos. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 3 +[74] Krishna Srinivasan, Karthik Raman, Jiecao Chen, Michael Bendersky, and Marc Najork. Wit: Wikipedia-based image text dataset for multimodal multilingual machine learning. In Proceedings of the 44th international ACM SIGIR conference on research and development in information retrieval, pages 2443-2449, 2021. 6, 7 +[75] Pengzhan Sun, Bo Wu, Xunsong Li, Wen Li, Lixin Duan, and Chuang Gan. Counterfactual debiasing inference for compositional action recognition. In Proceedings of the 29th ACM International Conference on Multimedia, pages 3220-3228, 2021. 3 +[76] Yansong Tang, Dajun Ding, Yongming Rao, Yu Zheng, Danyang Zhang, Lili Zhao, Jiwen Lu, and Jie Zhou. Coin: A large-scale dataset for comprehensive instructional video analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1207-1216, 2019. 2 +[77] Zhan Tong, Yibing Song, Jue Wang, and Limin Wang. Videomae: Masked autoencoders are data-efficient learners for self-supervised video pre-training. Advances in neural information processing systems, 35:10078-10093, 2022. 2 +[78] An-Lan Wang, Kun-Yu Lin, Jia-Run Du, Jingke Meng, and Wei-Shi Zheng. Event-guided procedure planning from instructional videos with text supervision. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13565–13575, 2023. 3 +[79] Huaxiaoyue Wang, Kushal Kedia, Juntao Ren, Rahma Abdullah, Atiksh Bhardwaj, Angela Chao, Kelly Y Chen, Nathaniel Chin, Prithwish Dan, Xinyi Fan, et al. Mosaic: A modular system for assistive and interactive cooking. arXiv preprint arXiv:2402.18796, 2024. 2 +[80] Mengmeng Wang, Jiazheng Xing, and Yong Liu. Actionclip: A new paradigm for video action recognition. arXiv preprint arXiv:2109.08472, 2021. 2 +[81] Weihan Wang, Zhen Yang, Bin Xu, Juanzi Li, and Yankui Sun. *Vilta: Enhancing vision-language pre-training through textual augmentation.* In *Proceedings of the IEEE/CVF International Conference on Computer Vision*, pages 3158–3169, 2023. 3 +[82] Xiaolong Wang and Abhinav Gupta. Unsupervised learning of visual representations using videos. In Proceedings + +of the IEEE international conference on computer vision, pages 2794-2802, 2015. 2 +[83] Xiaolong Wang, Ali Farhadi, and Abhinav Gupta. Actions~ transformations. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 2658-2667, 2016. 3 +[84] Michael Wray, Hazel Doughty, and Dima Damen. On semantic similarity in video retrieval. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3650-3660, 2021. 2 +[85] Jhih-Ciang Wu, He-Yen Hsieh, Ding-Jie Chen, Chiou-Shann Fuh, and Tyng-Luh Liu. Self-supervised sparse representation for video anomaly detection. In European Conference on Computer Vision, pages 729-745. Springer, 2022. 7 +[86] Fanyi Xiao, Kaustav Kundu, Joseph Tighe, and Davide Modolo. Hierarchical self-supervised representation learning for movie understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9727-9736, 2022. 2 +[87] Tete Xiao, Ilija Radosavovic, Trevor Darrell, and Jitendra Malik. Masked visual pre-training for motor control. arXiv preprint arXiv:2203.06173, 2022. 2 +[88] Chen-Wei Xie, Siyang Sun, Xiong Xiong, Yun Zheng, Deli Zhao, and Jingren Zhou. Ra-clip: Retrieval augmented contrastive language-image pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19265-19274, 2023. 3 +[89] Mengda Xu, Zhenjia Xu, Cheng Chi, Manuela Veloso, and Shuran Song. Xskill: Cross embodiment skill discovery. In Conference on robot learning, pages 3536-3555. PMLR, 2023. 2 +[90] Zihui Xue and Kristen Grauman. Learning fine-grained view-invariant representations from unpaired ego-exo videos via temporal alignment. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 6, 8 +[91] Zihui Xue, Kumar Ashutosh, and Kristen Grauman. Learning object state changes in videos: An open-world perspective. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18493-18503, 2024. 3 +[92] Mengyue Yang, Furui Liu, Zhitang Chen, Xinwei Shen, Jianye Hao, and Jun Wang. Causalvae: Disentangled representation learning via neural structural causal models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9593-9602, 2021. 3 +[93] Ruolin Ye, Yifei Hu, Yuhan Anjelica Bian, Luke Kulm, and Tapomayukh Bhattacharjee. Morpheus: a multimodal one-armed robot-assisted peeling system with human users in-the-loop. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 9540–9547. IEEE, 2024. 2 +[94] Fangqiu Yi, Hongyu Wen, and Tingting Jiang. Asformer: Transformer for action segmentation. In The British Machine Vision Conference (BMVC), 2021. 6, 7 +[95] Kexin Yi*, Chuang Gan*, Yunzhu Li, Pushmeet Kohli, Jiajun Wu, Antonio Torralba, and Joshua B. Tenenbaum. + +Clevrer: Collision events for video representation and reasoning. In International Conference on Learning Representations, 2020. 3 +[96] Kevin Zakka, Andy Zeng, Pete Florence, Jonathan Tompson, Jeannette Bohg, and Debidatta Dwibedi. Xirl: Cross-embodiment inverse reinforcement learning. In Conference on Robot Learning, pages 537–546. PMLR, 2022. 2 +[97] Chuanqi Zang, Hanqing Wang, Mingtao Pei, and Wei Liang. Discovering the real association: Multimodal causal reasoning in video question answering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19027-19036, 2023. 3 +[98] Can Zhang, Tianyu Yang, Junwu Weng, Meng Cao, Jue Wang, and Yuexian Zou. Unsupervised pre-training for temporal action localization tasks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14031-14041, 2022. 2 +[99] Jianrui Zhang, Mu Cai, Tengyang Xie, and Yong Jae Lee. Countercurate: Enhancing physical and semantic visio-linguistic compositional reasoning via counterfactual examples. arXiv preprint arXiv:2402.13254, 2024. 3 +[100] Letian Zhang, Xiaotong Zhai, Zhongkai Zhao, Yongshuo Zong, Xin Wen, and Bingchen Zhao. What if the tv was off? examining counterfactual reasoning abilities of multimodal language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21853-21862, 2024. 3 +[101] Tianyu Zhang, Weiqing Min, Jiahao Yang, Tao Liu, Shuqiang Jiang, and Yong Rui. What if we could not see? counterfactual analysis for egocentric action anticipation. In IJCAI, pages 1316-1322, 2021. 2 +[102] Yiwu Zhong, Licheng Yu, Yang Bai, Shangwen Li, Xueting Yan, and Yin Li. Learning procedure-aware video representation from instructional videos and their narrations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14825-14835, 2023. 2, 3, 6, 7, 8 +[103] Zhun Zhong, Liang Zheng, Guoliang Kang, Shaozi Li, and Yi Yang. Random erasing data augmentation. In Proceedings of the AAAI conference on artificial intelligence, pages 13001-13008, 2020. 3 +[104] Honglu Zhou, Roberto Martin-Martin, Mubbasir Kapadia, Silvio Savarese, and Juan Carlos Niebles. Procedure-aware pretraining for instructional video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10727-10738, 2023. 2, 3 +[105] Yifeng Zhu, Arisrei Lim, Peter Stone, and Yuke Zhu. Vision-based manipulation from single human video with open-world object graphs. arXiv preprint arXiv:2405.20321, 2024. 2 \ No newline at end of file diff --git a/ICCV/2025/What Changed and What Could Have Changed_ State-Change Counterfactuals for Procedure-Aware Video Representation Learning/images.zip b/ICCV/2025/What Changed and What Could Have Changed_ State-Change Counterfactuals for Procedure-Aware Video Representation Learning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..0241d713fe09cc3462caba21b91120743260d14b --- /dev/null +++ b/ICCV/2025/What Changed and What Could Have Changed_ State-Change Counterfactuals for Procedure-Aware Video Representation Learning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a4d49e578121642e71c72bb37be0eaa7985326e259a2211095efef47160b4278 +size 479820 diff --git a/ICCV/2025/What Changed and What Could Have Changed_ State-Change Counterfactuals for Procedure-Aware Video Representation Learning/layout.json b/ICCV/2025/What Changed and What Could Have Changed_ State-Change Counterfactuals for Procedure-Aware Video Representation Learning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..c6ceff02d4b41eb8347aa4d0cc8a3bf41995fcbb --- /dev/null +++ b/ICCV/2025/What Changed and What Could Have Changed_ State-Change Counterfactuals for Procedure-Aware Video Representation Learning/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:56ba1aaa25ef1890822c6c5589b9d22b6fcc9973025f43533b4124185d782ca4 +size 435588 diff --git a/ICCV/2025/What Changed_ Detecting and Evaluating Instruction-Guided Image Edits with Multimodal Large Language Models/c0a4e339-69da-4cf2-9ffd-f439404dd16a_content_list.json b/ICCV/2025/What Changed_ Detecting and Evaluating Instruction-Guided Image Edits with Multimodal Large Language Models/c0a4e339-69da-4cf2-9ffd-f439404dd16a_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..6cbd22e918b86f9ef77f1df2f5c71e842b486a0b --- /dev/null +++ b/ICCV/2025/What Changed_ Detecting and Evaluating Instruction-Guided Image Edits with Multimodal Large Language Models/c0a4e339-69da-4cf2-9ffd-f439404dd16a_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f652e9665dc76a0d2db180fc5ad311ff243c5506fecc4c2eabce14e8cb48af1d +size 91172 diff --git a/ICCV/2025/What Changed_ Detecting and Evaluating Instruction-Guided Image Edits with Multimodal Large Language Models/c0a4e339-69da-4cf2-9ffd-f439404dd16a_model.json b/ICCV/2025/What Changed_ Detecting and Evaluating Instruction-Guided Image Edits with Multimodal Large Language Models/c0a4e339-69da-4cf2-9ffd-f439404dd16a_model.json new file mode 100644 index 0000000000000000000000000000000000000000..31a14c2409f45ff80284d5d1844084e0854fb445 --- /dev/null +++ b/ICCV/2025/What Changed_ Detecting and Evaluating Instruction-Guided Image Edits with Multimodal Large Language Models/c0a4e339-69da-4cf2-9ffd-f439404dd16a_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4244cf89822d6a06396b54e971218f22a54513fba626dc82bd80429495eebee7 +size 112588 diff --git a/ICCV/2025/What Changed_ Detecting and Evaluating Instruction-Guided Image Edits with Multimodal Large Language Models/c0a4e339-69da-4cf2-9ffd-f439404dd16a_origin.pdf b/ICCV/2025/What Changed_ Detecting and Evaluating Instruction-Guided Image Edits with Multimodal Large Language Models/c0a4e339-69da-4cf2-9ffd-f439404dd16a_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..ec6bf256e9935471a136b74bdc7298e6dd130fe6 --- /dev/null +++ b/ICCV/2025/What Changed_ Detecting and Evaluating Instruction-Guided Image Edits with Multimodal Large Language Models/c0a4e339-69da-4cf2-9ffd-f439404dd16a_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:34d0d7668e6aa40235e00ef0df99129b1549466fe354652e9dedbc042fca5c96 +size 1507021 diff --git a/ICCV/2025/What Changed_ Detecting and Evaluating Instruction-Guided Image Edits with Multimodal Large Language Models/full.md b/ICCV/2025/What Changed_ Detecting and Evaluating Instruction-Guided Image Edits with Multimodal Large Language Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..114ea3fcfd53c794125a0d6e1a3b74d3eaf8e557 --- /dev/null +++ b/ICCV/2025/What Changed_ Detecting and Evaluating Instruction-Guided Image Edits with Multimodal Large Language Models/full.md @@ -0,0 +1,375 @@ +# What Changed? Detecting and Evaluating Instruction-Guided Image Edits with Multimodal Large Language Models + +Lorenzo Baraldi\*2, Davide Bucciarelli\*1,2, Federico Betti\*3, Marcella Cornia1, Lorenzo Baraldi1, Nicu Sebe3, Rita Cucchiara1,4 + +1University of Modena and Reggio Emilia, Italy 2University of Pisa, Italy 3University of Trento, Italy 4IIT-CNR, Italy + +1{name.surname}@unimore.it,2{name.surname}@phd.unipi.it,3{name.surname}@unitn.it + +![](images/842b10635bc5ed4ccb9805704ba13e005c7ed007625c0f4464dd3426aeaeb220.jpg) +Editing Prompt: + +![](images/5a67f921f0ef05605a85c35500550062b1c0cbde8ac22777f3e47bde5121a195.jpg) +Change the color of the knife to blue +Detected Differences + +![](images/be323ced61f58f39ac764c1492820519ece0239b1ac4bae578d97f0ab726f9e9.jpg) +EDIT: pocketknife changed to blue pocketknife +Editing Prompt: + +![](images/e58eb83b8859b44d0d4ad05310c69f3f8dbbabf9e514a98100205232d58f6b79.jpg) +Remove the man from the image. +Detected Differences +REMOVE: person + +![](images/d0574b049b18276c8f1d0a9e13fb9f70932278c9860456c836e10f79bccfd096.jpg) +Editing Prompt: + +![](images/ea70fc4920bb02c6b08f500fb1cb13ee9fc13be37479d3fe6737a83973b2cfd4.jpg) +Remove the cup from the image +Detected Differences + +![](images/c36098269f3c337b98c0677b0da0523c8f14a4d02cecc9dddd5a449e32990772.jpg) +REMOVE:cup +Editing Prompt: + +![](images/ddb4deb1cb1b499e37bed15b2ec7a447b4c1f2b74b5c3604f8a6954230c15203.jpg) +Add a tie to the man +Detected Differences +ADD: necktie + +![](images/e95449ecaf1ad7bb1ae3a6db6d8d7a0aa732454a4ddd6d68a557140064e2d00c.jpg) +Editing Prompt: + +![](images/72a12db11654cafa0e816ba861a06a0d6a1afb99a33c702845ba73b61a4a8766.jpg) +Replace the cherub with a monkey +Detected Differences +Replace the ball with a tennis ball + +![](images/f4a8a54602d98c64e2f104ba6d414e5cbce899db40845d153d84b5f24bdc48b2.jpg) +EDIT: statue changed to monkeys; ADD: monkey +Editing Prompt: + +![](images/78d10b0a3ea6980a7bced3732c79d33a602d789cc99b6fe4d7bf7f8656706683.jpg) +Let the elephant be white +Detected Differences +EDIT: elephant changed to white elephant +Make the glass vase blue + +![](images/9255849964eb9bcba332f245f7103e52ae33e14bc89f8e29f7c8fbb8535a384c.jpg) +Editing Prompt: + +![](images/138871ca4b5762d6a56acfafd83b46c8f710c19729fa78044093dda8a82d529e.jpg) +Detected Differences + +![](images/38bc18c80cfe43ab864f35e5553885eee1648e845bc87f925876470cdd3147cc.jpg) +ADD: tennis ball +Editing Prompt: + +![](images/52401d67ae0d5be31211c7de67d43acc24062fdfc26cdd27f10ce5d2a5fe8ed2.jpg) +Detected Differences +EDIT: vase holder to [...] EDIT: flowerpot to [...] +Figure 1. Qualitative examples from DICE. Our approach detects differences between an original image and an edited one, identifying the involved objects and the type of edit. Further, DICE evaluates each difference to determine its coherence with the editing prompt. + +# Abstract + +Instruction-based image editing models offer increased personalization opportunities in generative tasks. However, properly evaluating their results is challenging, and most of the existing metrics lag in terms of alignment with human judgment and explainability. To tackle these issues, we introduce DICE (Difference Coherence Estimator), a model designed to detect localized differences between the original and the edited image and to assess their relevance to the given modification request. DICE consists of two key components: a difference detector and a coherence estimator, both built on an autoregressive Multimodal Large Language Model (MLLM) and trained using a strategy that leverages self-supervision, distillation from inpainting networks, and full supervision. Through extensive experiments, we evaluate each stage of our pipeline, comparing different MLLMs within the proposed framework. We demonstrate that DICE effectively identifies coherent edits, effectively evaluating images generated by different editing models with a strong correlation with human judgment. We publicly release our source code, models, and data at project page. + +# 1. Introduction + +The field of Generative AI has recently gained significant attention in both industry and research, driven by the development of models capable of generating content that closely follows the distribution of natural language [1, 14, 21] and real images [15, 39, 43]. Thanks to the availability of cross-modal operators and architectures, multimodal generative approaches, which can seamlessly integrate both the vision and language modalities, are reaching increasing levels of accuracy [6, 32, 49, 52]. These advancements are driving the development of new applications and tasks while enhancing user control over the generation process. + +Among these proposals, instruction-based image editing models [4, 18, 25, 54] enable the modification of an input image based on a free-form textual instruction from the user. This extends traditional text-to-image models, shifting the focus from generating an image from scratch to modifying an existing one while adhering to a given prompt. While + +such models promise to offer increased personalization levels, the understanding of the results they generate (and thus their proper evaluation) clearly becomes more challenging than in traditional text-to-image contexts. + +Prior work has approached the evaluation of image editing models by developing annotated benchmarks with ground-truth edited images [45, 46, 54] or by proposing metrics based on pre-trained backbones such as CLIP [40] and DINO [8]. Other methods leverage pre-trained large language models to assess the quality of image edits [3, 27, 35]. While these proposals might offer a first viable solution to assess the performance of image editing models, they still do not reach a sufficient alignment with human evaluation. Further, being based on either global embedding vectors or replies from pre-trained (and often private) models, explaining their evaluations becomes cumbersome. + +Drawing inspiration from these considerations, we take a novel approach and propose a learnable model for understanding localized image edits, which can provide increased alignment with human preferences and more easily explainable results. Although instruction-guided image editing encompasses a broad range of possible transformations, this work focuses on object-level edits, motivated by the observation that precise, localized modifications remain a significant challenge for generative models and thus require specialized evaluation protocols. Accordingly, our model first identifies object-level differences between the original and the edited images and subsequently evaluates the coherence of each detected modification relative to the user's input instruction. We term our pipeline as DICE, short for DDifference Coherence Estimator. Sample results from our model are shown in Fig. 1. + +Maintaining the goal of creating a fully explainable pipeline, we develop two distinct models, one for each stage, both built on an autoregressive Multimodal Large Language Model (MLLM) [6]. While both models leverage the understanding of RoIs, the first primarily focuses on their prediction, whereas the second emphasizes their interpretation to assess coherence with the given prompt. In particular, we first propose a difference detector that, given both images, estimates a set of localized differences that describe the change in content from one image to the other. The model also predicts the category of detected modifications, as either addition, removal, or editing of an object. A second model, based on the same architecture, is instead trained to predict the coherence of an object-level modification with respect to the modification request made by the user and provide a textual rationale for each decision made. + +We experimentally assess the appropriateness of our strategy by carefully investigating the performance of both our difference detector and our coherence estimator, in comparison with existing solutions, and when employing different MLLMs. Further, we conduct a dedicated user study + +to measure the alignment of the evaluations predicted by our proposal in terms of model ranking and evaluation. Experimental results demonstrate that our proposal achieves increased ranking capabilities while benefiting from an explainable-by-default approach. Further, when integrated into existing metrics, it acts as the basis for building model evaluation metrics with increased human alignment. + +# 2. Related Work + +Image Editing Models. Recent developments in Generative Adversarial Networks (GANs) [20, 28] and diffusion models [13, 15, 39, 43] have driven the rise of AI-generated content. However, users increasingly demand not only the creation of new visual data but also the modification of existing images, altering properties such as style [17, 28] and content [4, 16]. Within this domain, tasks vary based on the input used for conditioning. Image inpainting [10, 47], for instance, uses user-designed masks to guide edits in specific regions. In contrast, our work focuses on instruction-based image editing, where the model relies solely on textual instructions and the original image. + +Within instruction-based editing, traditional approaches rely on diffusion models fine-tuned for instruction following. A fundamental contribution in this domain is InstructPix2Pix [4], which is trained on fully synthetic triplets composed of an input image, an editing instruction, and the edited image. However, since reliance on synthetic data may introduce noise, MagicBrush [54] refines the InstructPix2Pix framework using a high-quality, curated dataset. Differently, InstructDiffusion [18] is trained to perform a wide variety of tasks, including image editing and grounding. Further, HIVE [55] proposes a reward-based training that incorporates human feedback to enhance its instruction following capabilities. Differently, MGIE [16] employs a frozen MLLM to generate concise edit instructions that guide the diffusion model during image editing, whereas SmartEdit [26] captures the interactions between image and text via MLLM, improving prompt understanding. + +Image Editing Evaluation and Benchmarks. Instruction-based image editing models have traditionally been evaluated along two dimensions: prompt adherence and background preservation. The former assesses the extent to which the output aligns with the given prompt instruction, whereas the latter evaluates the preservation of the unmodified contextual information. While benchmarks featuring annotated ground-truth edited images are available [45, 46, 54], our work prioritizes a reference-free evaluation that eliminates the need for human-defined labels. In this setting, background preservation is typically evaluated by using either CLIP [40] or DINO [8] to compute cosine similarities between the visual tokens extracted from the source and edited images [4, 16, 55]. Likewise, prompt adherence is usually quantified using CLIP-based similarities + +![](images/045d73d7b52c430a5f82a7d38a0bf5d4b237ae9020db63577b5ae35a3e3c8eac.jpg) +Figure 2. Illustration of DICE. We employ an MLLM and fine-tune it for two different tasks. In the first stage (difference detection), the MLLM is trained to detect semantic differences between the original image and the edited one. In the second stage (coherence estimation), the MLLM is instructed to analyze and assess the coherence of each detected difference with respect to the given user prompt. + +![](images/04d35b84060bd9aaeaa32b446e0fd8409f7ff09c5b431468b8d6db42d4884331.jpg) + +between the target caption and the edited image [44, 54]. + +With the recent proposal of MLLMs, these methods have been employed for prompt adherence evaluation. For instance, in HQ-Edit [27], GPT-4V [1] is used to assess coherence and alignment metrics. Additionally, I²E Bench [35] incorporates human annotation for certain queries, subsequently leveraging GPT-4V for question answering. However, reliance on GPT-based evaluations introduces potential variability and reproducibility issues, as model updates may alter the final scores. In contrast, we propose an interpretable solution based on an open-source MLLM, thereby increasing both the rationale behind our scoring mechanism and the reproducibility of the results. + +Multimodal Large Language Models. With the introduction of LLM, the research community has put great effort into connecting visual modalities with textual reasoning [5, 11, 32]. Building on these efforts, MLLMs have also been used to address visual grounding tasks. Among these, Shikra [9] leverages a novel training paradigm to give the MLLM the capacity to reason with bounding boxes, making it able to precisely locate objects and relations in the image. Similarly, LISA [29] and GROUNDHOG [56] perform segmentation incorporating pixel-level representations for fine-grained visual understanding. + +Compared to existing MLLMs for visual grounding, our model leverages multiple images to detect semantic differences, introducing a novel setting that requires an MLLM capable of multi-image understanding. In this context, Idefics3 [30] is able to comprehend multiple visual inputs interleaved with text, using a novel image encoding pipeline. Differently, Qwen [49] extends classical multimodal capabilities to video and multi-image understanding thanks to a novel encoding strategy and positional embedding technique for the visual inputs. Finally, mPLUG-Owl3 [52] introduces a novel attention-based strategy to ef + +ficiently encode long image sequences. Building on these multi-image models, we extend the capabilities of MLLMs by introducing a novel framework for multi-image detection, enabling the identification of semantic differences across multiple images within a unified system. + +# 3. Proposed Method + +# 3.1. Preliminaries + +The objective of our approach is to assess the quality of outputs generated by instruction-based image editing models [4, 16, 55]. Formally, given an original image $\pmb{x}$ and a textual instruction $\pmb{t}$ , an image editing method $f$ generates a new image $e = f(\pmb{x}, t)$ , which incorporates the modifications specified by the textual prompt $\pmb{t}$ , while preserving the contextual integrity of $\pmb{x}$ . + +While the modifications introduced by image editing models can be diverse - owing to the expressiveness of textual instructions and the flexibility of generative models - we focus specifically on object-level modifications. By means of this choice, we consider transformations that are clearly localizable inside the input scene. In particular, we assume that the original image contains a finite set of objects, denoted as $\{o_i\}_{i=1}^n$ , and that the textual instruction $\pmb{t}$ specifies an operation to be applied to it. Given a correct application of the specified operation, the expected edited image, denoted as $\tilde{\pmb{e}}$ , serves as the ground-truth for evaluating the intended modification. + +Object-level Modifications. Without loss of generality, we assume that the alterations required in the textual prompt can be classified as either the addition, removal, or modification of objects inside the input image. Specifically, we consider three categories of alterations, namely: + +1. Addition of an object. A new object instance $o_{n+1}$ is introduced into the scene, yielding a new image repre + +sentation. Treating images as a set of objects, and with a slight abuse of notation, this operation results in a new image that can be expressed as $\tilde{e} = (\bigcup_{i=1}^{n} o_i) \cup \{o_{n+1}\}$ . + +2. Removal of an object. The operation removes an object $o_j$ which was present in the original image, which results in $\tilde{e} = \left( \bigcup_{i=1}^{j-1} o_i \cup \bigcup_{i=j+1}^{n} o_i \right)$ . +3. Edit of an object. An existing object $o_j$ is altered in at least one visual aspect. Let $\tilde{o}_j$ denote the modified object; the updated image is then represented as $\tilde{e} = (\bigcup_{i=1}^{n} o_i \setminus \{o_j\}) \cup \{\tilde{o}_j\}$ . + +# 3.2. Overview of the Approach + +An image editing network can produce an image $e$ that does not accurately respect all the object-level operations required to produce $\tilde{e}$ . To quantify this discrepancy, we propose DDifference Coherence Estimator (DICE): a two-step approach in which (i) we detect the object-level differences between the edited and the original image and then (ii) we evaluate the coherence of each difference with the textual prompt to determine whether a discrepancy reflects the intended modifications or constitutes an unwanted alteration. An overview of the two stages of DICE is depicted in Fig. 2. + +Difference Detection. In this stage, we extract the set of objects that are present in the original image $x$ but absent or changed in the predicted image $e$ or vice versa. We denote this collection of differences as semantic difference, which is formally defined as + +$$ +\Delta (\boldsymbol {x}, \boldsymbol {e}) = \left\{o \mid o \in \{\boldsymbol {x} \setminus \boldsymbol {e} \} \cup \{\boldsymbol {e} \setminus \boldsymbol {x} \} \right\}. \tag {1} +$$ + +The aforementioned set of differences is estimated through a learnable difference detector $D(\boldsymbol{x}, \boldsymbol{e})$ that, given the original and modified image, predicts a set of object-level differences which serves as an approximation of $\Delta(\boldsymbol{x}, \boldsymbol{e})$ , expressed in terms of RoIs. Notably, this detection process is executed independently of the textual prompt $t$ to ensure that the outcome is not biased by the intended edit. + +Coherence Estimation. In the second step, we evaluate each detected difference individually to assess its coherence with the textual prompt. Specifically, a coherence estimator operates on each detected object-level change $o_i \in D(\boldsymbol{x}, \boldsymbol{e})$ , determining whether the visual modifications align with the intended edit $\tilde{\boldsymbol{e}}$ . Formally, + +$$ +C (\boldsymbol {x}, \boldsymbol {e}, \boldsymbol {t}, o _ {i}) = \left\{ \begin{array}{l l} 1 & \text {i f} o _ {i} \in \tilde {\boldsymbol {e}}, \\ 0 & \text {o t h e r w i s e .} \end{array} \right. \tag {2} +$$ + +This process enables the differentiation between intended and unintended edits, thereby facilitating a comprehensive and interpretable evaluation of the task. + +# 3.3. Detecting Differences + +To build the difference detector and the coherence estimator, we design two novel architectures building upon the + +general structure of autoregressive Multimodal Large Language Models (MLLMs), which offer a sufficiently general framework for encoding images, prompts, and RoIs. + +For difference detection, we train an MLLM $D(\mathbf{x}, \mathbf{e})$ to predict object-level differences between two images in terms of (i) the type of modification (i.e., addition, removal, edit), (ii) the RoI of the modified region, and (iii) the subjects involved in the modification. Formally, the difference detector operates as: + +$$ +D (\boldsymbol {x}, \boldsymbol {e}) = \left[ \left(c _ {i}, S _ {i}, b _ {i}\right) \right] _ {i = 1} ^ {k}, \tag {3} +$$ + +where $c_{i}$ denotes a command from $\{\mathrm{ADD},\mathrm{EDIT},\mathrm{REMOVE}\}$ , $S_{i}$ describes the modified object in free text, and $b_{i} \in \mathbb{R}^{4}$ defines the bounding box coordinates in the format $(x_{\min},y_{\min},x_{\max},y_{\max})$ . + +Each tuple $(c_i, S_i, b_i)$ is fully specified in the textual domain, and the model is optimized to output structured text representations of modifications, formatted as: "COMMAND: object, [bb1, bb2, bb3, bb4]". This structured output ensures precise localization and identification of changes. + +To estimate the confidence of a predicted RoI, we compute the probability of the predicted command $c_{i}$ relative to the total probability mass of all possible commands. This confidence measure aligns with conventional object detection confidence estimation techniques [7, 19]. + +Training Procedure Overview. Our training pipeline consists of two phases: (i) a pre-training stage leveraging real image pairs and (ii) a fine-tuning stage using images modified with inpainting models. Notably, the model is intentionally not trained on instruction-based paired edited images due to two key considerations. First, such training may lead to overfitting to specific editing architectures while underperforming on unseen methods. Second, instruction-edited image pairs would necessitate extensive human annotations for commands and RoIs, introducing significant cost and time constraints. + +Stage 1: Learning to Compare Similar Images. In this stage, we instruct the difference detector to identify object-level differences when considering pairs of similar real images. Specifically, the MLLM is trained to predict objects that are not mutually present in both images, leveraging an existing object detection dataset. + +In detail, we employ images and object-level annotations from the LVIS dataset [22]. We identify pairs of similar images by computing the cosine similarity in the DINOv2 [38] embedding space, and retaining only those with a cosine similarity greater than 0.6, ensuring high visual similarity. In addition, we make sure that the aforementioned pairs contain at least one common object class and differ by fewer than 15 object classes. This filtering results in a total of $118\mathrm{k}$ image pairs. When encoding image pairs in + +this stage, objects present in the first image but absent in the second are labeled as REMOVE, while objects appearing in the second image but not in the first are labeled as ADD. Objects exhibiting an intersection-over-union above a predefined threshold across both images are classified as EDIT. Further details regarding dataset construction are provided in the supplementary material. + +Stage 2: Learning to Detect Inpainted Areas. In the second stage, we refine the model by bridging the gap between pre-training on real image pairs and application to edited images. To accomplish this, we select images from the LVIS dataset and sample non-overlapping annotated objects. Starting from LVIS images, we employ inpainting to generate original and edited image pairs. For each object, an operation is randomly selected among $\{\mathrm{ADD},\mathrm{EDIT},\mathrm{REMOVE}\}$ . Inpainting is then performed using LaMa [47]: for the added objects, the corresponding region is inpainted on the original image, while for removed objects, the inpaint is performed on the edited one. For the edit operation, $30\%$ of the edited objects undergo a color change, while the remaining $70\%$ are substituted with a different object. We rely on GPT-4o to generate target objects suitable for substitution by prompting it with a caption of the original image as well as the original object. To perform the edits, the Kandinsky 2.2 inpaint model [42] is employed to inpaint selected regions. This process leads to a training set of 97k elements and a test set of 19k. + +# 3.4. Estimating the Coherence + +Given the original and edited image, along with a detected difference, the coherence estimator produces a binary decision (i.e., Yes/No) and a textual rationale indicating whether the modification is coherent with the intended edit. To induce the model to focus on a specific difference, the bounding box of the detected difference is visually depicted on the original and edited image before they are fed to the coherence estimator. Further, we also include the type of modification $c_{i}$ and its textual description $S_{i}$ as input to the model. + +To train the model, we annotated 116 samples from the EmuEdit [44] test set, with the corresponding edited images generated using InstructDiffusion [18]. For each sample, we manually annotated both the observed differences and a binary coherence indicator. In addition, we employed GPT-4o to generate a rationale for each annotation, which was subsequently refined through human review to eliminate any inaccuracies. While the command $c$ and the subjects $S$ were provided in textual format, the bounding boxes were directly depicted on the images using a color-coded scheme: red for additions, green for edits, and blue for removals. Specifically, additions are superimposed on the edited image, whereas the bounding boxes corresponding to removals and edits are applied solely to the original image. + +# 4. Experimental Results + +# 4.1. Implementation and Training Details + +In our experiments, we consider three recent MLLMs designed to support multi-image input tasks: Idefics3-8B [30], Qwen2-VL-7B [49], and mPLUG-Owl3-7B [52]. Following the training steps described in Sec. 3.3 and Sec. 3.4, we fine-tune each model using QLoRA [12]. Specifically, the vision encoder and projector of each model remain frozen, while the language model is quantized and augmented with LoRA [24] adapters. LoRA is applied to the self-attention and MLP layers across all 32 Transformer blocks. Further details regarding model configuration and training are provided in the supplementary material. + +# 4.2. Evaluating Detected Differences + +To evaluate the performance of our detection and coherence pipeline, we construct a new dataset, termed as DICE-D, specifically designed for the task. This dataset comprises 800 edited images extracted from the I²EBench [35] benchmark. We select images and prompts from a subset of the edit categories that involve object-based modifications, in particular: color alteration, counting (only if involving a single object change), object removal, and object replacement. Each pair of images is manually annotated with detected differences represented as bounding boxes, editing commands (i.e., ADD, REMOVE, EDIT), and textual descriptions. Also, we assign a coherence label for each bounding box to indicate its alignment with the prompt. + +Difference Detection Results. We first evaluate the ability of the difference detection model to detect modifications between the original and edited images. Following object detection literature [7, 31], we compute the mean average precision (AP) between predicted and labeled bounding boxes from DICE-D. Specifically, we average results at 10 IoU thresholds from 0.5 to 0.95. Further, we include $\mathrm{AP}_{50}$ (average precision at $50\%$ IoU), $\mathrm{AP}_{75}$ (average precision at $75\%$ IoU), $\mathrm{AP}_{M}$ (for medium-sized objects), and $\mathrm{AP}_{L}$ (for large-sized objects), ensuring a comprehensive evaluation across different object scales and IoU thresholds. + +Results are reported in Table 1, where we consider both class-agnostic detection and class-aware detection and compare DICE against other MLLMs and open-vocabulary detectors, OWLv2 [37] and Grounding DINO [33], for which the experimental setup is detailed in the supplementary material. In class-agnostic detection, all predictions are treated as belonging to the same category, to evaluate only the bounding box location. Vice versa, in class-aware detection, both the bounding box location and the editing command are considered, evaluating the accuracy of both the detection and the associated command. As shown, using Idefics as the base model consistently achieves the best performance across all metrics, with an increase of 17.1 and 20.0 in AP + +
ConfidenceTrainingClass-agnostic DetectionClass-aware Detection
Stage 1Stage 2AP\(AP_{50}\)\(AP_{75}\)\(AP_M\)\(AP_L\)AP\(AP_{50}\)\(AP_{75}\)\(AP_M\)\(AP_L\)\(AP_ADD\)\(APREM\)\(APEDIT\)
Grounding Models
OWLv2 [37]✓--16.425.217.312.421.49.814.010.79.110.612.37.19.8
Grounding DINO [33]✓--12.817.613.49.315.26.88.67.46.17.84.88.66.9
Alternative MLLMs
GlaMM [41]---5.08.34.82.06.32.74.12.81.13.53.91.32.9
GPT-4 [1]---0.51.90.20.00.80.31.10.100.51.71.90.3
Qwen2-VL-7B [49]-✓✓2.05.31.30.33.41.63.31.20.12.40.10.13.2
Qwen2-VL-7B [49]✓✓✓2.35.61.80.13.51.64.11.101.60.10.13.2
mPLUG-Owl3-7B [52]-✓✓3.810.32.40.55.12.56.71.70.33.31.11.55.1
mPLUG-Owl3-7B [52]✓✓✓5.213.73.20.57.34.410.63.10.35.50.17.34.8
Ablation Studies
Idefics3-8B [30]--✓15.130.113.48.518.69.418.19.35.511.07.37.513.4
Idefics3-8B [30]-✓✓16.730.716.010.420.310.617.910.96.612.57.47.716.5
Idefics3-8B [30]✓rand✓✓16.831.216.210.320.59.716.510.06.711.27.07.914.2
Idefics3-8B [30]✓✓-2.95.12.81.94.21.42.21.40.91.82.00.41.8
Idefics3-8B [30]✓-✓18.236.816.09.822.712.222.212.17.315.37.213.415.8
DICE (Ours)
Idefics3-8B [30]✓✓✓22.340.121.913.536.815.524.816.511.415.419.614.816.4
+ +Table 1. Performance comparison of various MLLMs in the difference detection stage of our pipeline, evaluated under both class-agnostic and class-aware settings. Results are presented in terms of AP metrics across various training configurations. + +
ConfidenceCoherence over GT AreasCoherence over Detected Areas
AccuracyAP\( AP_{50} \)\( AP_{75} \)
Alternative MLLMs
Qwen2-VL-7B [49]-76.61.43.80.6
Qwen2-VL-7B [49]✓76.61.84.90.9
mPLUG-Owl3-7B [52]-85.93.810.02.4
mPLUG-Owl3-7B [52]✓85.94.511.32.9
DICE (Ours)
Idefics3-8B [30]-85.412.020.312.0
Idefics3-8B [30]✓85.415.526.016.1
+ +Table 2. Performance comparison of various MLLMs in the coherence estimation stage of our pipeline. Coherence accuracy is measured using ground-truth differences, while coherence over detected areas employs the output of the first stage of our pipeline and is evaluated using AP metrics. + +in a class-agnostic setting compared to Qwen and mPLUG. Similarly, when considering the command labels, Idefics outperforms Qwen and mPLUG by 11.8 and 13.9. This superior performance may be attributed to the distinctive image encoding pipeline employed by Idefics, in which each image is resized to a larger dimension and segmented into a grid of crops, each encoded independently. Although this approach increases the amount of visual information processed, the grid-based encoding simultaneously introduces a localization factor to the captured visual details that can be leveraged in a task of detection. Further, the inferior performance of OWLv2 [37] and Grounding DINO [33] underscores that difference detection is fundamentally distinct from standard object detection. + +When validating the contribution of the key components of our pipeline, it is worth noting that the initial pre + +training stage significantly contributes to the final performance. Specifically, training Idefics in both stages results in improvements of 1.6 and 1.2 in class-agnostic and class-aware AP, respectively, compared to training solely on the second stage dataset. Notably, even training exclusively on the first stage enhances class-agnostic detection compared to Qwen (i.e., +0.4 AP). This observation implies that predicting missing objects in pairs of similar real images not only facilitates the learning of prediction syntax but also enables the transfer of relevant knowledge to the difference detection task in edited images. + +Moreover, incorporating the extracted confidence score enhances difference localization results, with improvements of 5.5 and 3.1 AP observed for Idefics-based models trained on both stages and solely on the second stage respectively, compared to using a fixed confidence value of 1. Additionally, using random confidence values results in worse performance, further indicating that the uncertainty in the command correlates with the actual prediction confidence. + +Coherence Estimation Results. The coherence estimation model assesses whether each detected difference is correctly aligned with the prompt using a binary prediction (i.e., Yes/No). Moreover, the coherence step incorporates explicit model reasoning [51], improving the interpretability of the results. Under this setting, models are evaluated on DICE-D, in terms of coherence over ground-truth areas and coherence over detected areas. To measure the coherence accuracy, ground-truth differences are fed into the coherence estimator, and results are evaluated against the manually annotated coherence. Instead, for the coherence over detected areas, the performance of DICE is measured through AP, considering the coherence label of the differ- + +
Editing ModelsCorrect EditsUnwanted EditsNo Visual Change
DICE (%) ↑Humans (1-5) ↑DICE (%) ↓Humans (1-5) ↑DICE (%) ↓Humans (%) ↓
HIVE [55]11.02.040.53.154.540.5
InstructPix2Pix [4]15.52.332.13.344.022.0
MGIE [16]23.02.721.63.811.510.0
MagicBrush [54]24.52.923.13.98.55.5
InstructDiffusion [18]30.03.019.14.013.510.5
+ +Table 3. Benchmark comparison of model rankings generated by DICE and those derived from the user study. The first two columns contrast average human ratings with scores obtained using DICE. The final column compares the percentage of unchanged images in the user study - cases with maximal background preservation and minimal prompt adherence - to the corresponding one identified by DICE. + +ence as the ground-truth category of the bounding box. This is done by extracting differences with the detector and subsequently predicting the coherence of each difference. + +Results are summarized in Table 2, evaluating various MLLMs as coherence estimators. Idefics consistently achieves the highest overall performance on AP, surpassing mPLUG and Qwen by 11.0 and 12.7 points, respectively, while maintaining comparable coherence accuracy. Additionally, incorporating confidence from the difference detection stage into the AP computation leads to performance improvements across all evaluated MLLMs. This highlights the importance of confidence extraction in the difference detection phase and motivates further research on aligning token probability with confidence estimation in object detection. Based on these results, Idefics3-8B is selected as the base MLLM for DICE in all subsequent experiments. + +# 4.3. Evaluation of Image Editing Models + +User Study Details. We conduct a user study to evaluate our proposed framework within the context of instruction-based image editing. In detail, we collect 200 prompts along with their corresponding original images, $50\%$ from MagicBrush [54] and $50\%$ from $\mathrm{I}^2\mathrm{EBench}$ [35]. For MagicBrush, we use GPT-4o to filter out prompts that do not involve object modification. Differently, for the $\mathrm{I}^2\mathrm{EBench}$ subset, we employ the same selection criteria as in the construction of DICE-D (cf. Sec. 4.2). We select five state-of-the-art generators and generate 200 edited images per generator. Specifically, our model selection includes HIVE [55], InstructPix2Pix [4], MGIE [16], MagicBrush [54], and InstructDiffusion [18]. To assess the quality of the generated edits, we asked 10 human annotators to evaluate each sample on two dimensions: prompt adherence and background preservation. Prompt adherence measures how well the modifications in the edited image align with the instructions given in the prompt. Differently, background preservation assesses how well the elements that were not intended to be modified remained unchanged. Ratings for each dimension were recorded on a 5-point Likert scale. We refer the user to the supplementary materials for more detailed information about the user study. + +Editing Model Ranking. To evaluate the performance of + +editing models, we establish three axes to rank the considered generators using DICE. Correct edits measure the percentage of images in which at least one detected difference is classified as coherent. Conversely, unwanted edits quantify the percentage of the total image area containing differences marked as non-coherent. Lastly, no visual change represents the percentage of images where DICE does not detect any visual alteration, potentially due to insufficient prompt clarity, lack of understanding, or inability of the model to execute the requested modification. Notably, correct and unwanted edits can be measured with the average prompt adherence and background preservation ratings from the user study. Similarly, from the user study, no visual change is assessed by identifying images rated 5 in background preservation and 1 in prompt following. + +In Table 3, we compare the rankings generated by our evaluation with those obtained from the user study. As it can be seen, model rankings from DICE align with human evaluations. In particular, InstructDiffusion and MagicBrush emerge as the top-performing models both according to DICE and human ratings, excelling in performing correct modifications while effectively preserving the background. Likewise, the ranking order for no visual change remains consistent across DICE and the user study, further validating the reliability of our approach. + +To qualitatively validate the predictions of DICE, Fig. 3 presents editing samples showcasing difference detection and coherence predictions across various prompts and models. Notably, DICE effectively identifies image differences and evaluates their coherence with the editing prompt. + +Correlation with Human Judgment. To further validate the alignment between the detected differences and human judgment, we integrate DICE predictions within CLIP-based evaluation metrics measuring image similarity (CLIP-I) for background preservation and image-text similarity (CLIP-T) for prompt adherence [23]. In particular, CLIP-I measures the similarity between the original and the edited image. To compute CLIP-T, instead, we caption the original image using an MLLM (i.e., Idefics3-8B in our experiments) and feed the extracted caption along with the editing prompt to GPT-4o to produce a caption of the ideal edited image (i.e., target caption). Through CLIP-T we then + +![](images/209d5eb675b0167181156d157b90fa386149700c6ad5252071fdfe627076ad1d.jpg) + +![](images/24947d319aaca58ab4bd9c09e78033799e54ac516ab3be43e18a9971a9a1c96e.jpg) + +![](images/f9d9d00229af79c243b4b642953b307c25ec4a5ad7626c111df03bd29e0fd416.jpg) + +![](images/935c900756e9cf35e61dd6e9d299f8df79847c5a47f606357ecf79ad7cbcb409.jpg) + +![](images/d2f20385e0642169f4a30ed6173de5228404f70845664e4e9c38c0740aaf1cb6.jpg) + +![](images/7b8d2038534801945ee4e5023cd6d7140658340e3aeaabb6092adf4238143c34.jpg) + +![](images/cd5253196282cc3193d71cb098ab027b8b5733bf40799a27aa99d89f8f383ab8.jpg) + +![](images/a735a842230ba751e1a10a5a3b0144d3f802fe979cf7879877d39149b06c8bc1.jpg) + +![](images/cb97cf0e1302fd1f58f335c89eedca6de0d00679facbd4f4b0b7aaa2b1401ac2.jpg) + +![](images/3e237d534f74dcee0cc625299b9b51b6f21db55976a885f374c0d39b42661c73.jpg) + +![](images/7a3520364aa786456d05ce08c067d9e9dbc57107fc4f90a8cd775cd217dee2d9.jpg) + +![](images/bc8836f0a24bf7c70c0df8ea02b90d28d1cd5129715a6472019a973ae92e7ace.jpg) + +![](images/319525bb1c5af5ca9becc0bb7ed59ab026beffbf918f9576c107ff12f21b746c.jpg) + +![](images/636fdef9a067e513413a0fbe85da579d08dbbfc1c7a08bdc288a3ffd1a9f8302.jpg) + +![](images/b99d8dd027372d17f29027b7285655d4c6d9c5b9893015befd92c572652934b8.jpg) + +![](images/ff8ae53bcb523a255fec8f8d09b818edfa1f881c8de3378ab89fece95b26da39.jpg) + +![](images/086d403b38c6cc0bc510aadc32da1938098434cf617d25a85445f68e1d6b0583.jpg) + +![](images/4a2dddfdd9073ddf49009a7a94136b4e263219b73126c49ebe8fbeb3b8ee12b1.jpg) + +![](images/fccfe449c91d78723c0c12b17c56e637d5407f0155a975e5925c1c6ccaf24770.jpg) +Figure 3. Qualitative samples of DICE applied on images edited by MGIE [16] and InstructDiffusion [18] models. + +![](images/f1b41cc0e9fe6e8971b64fc8295218a341e9e3722b64e91ed8850abaf81c4c24.jpg) + +![](images/89313b6b6616ca9347dadc58a9fde1d7a6574800e1e119e53aec0e0f746b32e0.jpg) + +![](images/fcb9bf40a54c4a67e6b4c9243ff1c761cbbea0eb629e72a6ae155d93c2e0a72f.jpg) + +![](images/9e495affc153c6d1758280db4bed6e0d1a2103110ca9c810224204767e24b8e1.jpg) + +![](images/e6c203ecaed62ae6b831758da6a8863d96487dc7a643e6c3cae83948891caba3.jpg) + +
ρρsτ
Background Preservation
CLIP-I51.142.933.3
w/ patch on random areas34.225.619.4
w/ patch on all detected differences32.715.611.6
w/DICE (patch on coherent differences)54.545.435.1
Prompt Adherence
CLIP-T21.523.117.1
w/ patch on random areas17.819.514.3
w/ patch on all detected differences13.314.610.7
w/DICE (patch on non-coherent differences)24.626.619.6
+ +Table 4. Correlation analysis showing improved alignment with human judgment when integrating DICE in CLIP-I and CLIP-T. + +measure the CLIP similarity between the target caption and the actual edited image. To integrate DICE in CLIP-I and CLIP-T, we modify both the original and edited images by selectively masking specific regions, taking into account DICE predictions. Specifically, for background preservation, we mask coherent differences, whereas, for prompt adherence assessment, non-coherent differences are patched. + +For this experiment, we compare the correlation scores of standard CLIP-I and CLIP-T metrics with those obtained by masking original and edited images using DICE. Results are shown in Table 4, where we measure Spearman's $\rho$ , Pearson's $\rho_{s}$ , and Kendall's $\tau$ correlation coefficients between automated evaluation scores and human judgment. We also compare when randomly masking areas from the differences detected by DICE, or when masking all detected differences, including both coherent and noncoherent changes. As it can be seen, DICE effectively enhances the correlation of both CLIP-I and CLIP-T metrics with human ratings. Specifically, for background preservation, CLIP-I exhibits the strongest correlation with hu + +man judgments when masking only coherent differences, achieving a Pearson's score of 54.5. This is a significant improvement over the baseline score of 51.1 (no masking), and substantially higher than the scores obtained with random masking or masking all detected differences. This result suggests that excluding correctly modified regions allows CLIP-I to better focus on the actual background, producing a metric that more accurately aligns with human perception. + +For prompt adherence, CLIP-T shows better correlation when we patch non-coherent differences, achieving a Pearson's score of 24.6, compared to 21.5 without masking. Instead, masking all differences leads to a drop to 13.3, likely due to indiscriminate removal of both erroneous and correct edits. This outlines that selectively removing only incorrect modifications helps CLIP-T focus on prompt-relevant changes, thus improving alignment with human evaluations. + +Overall, these results demonstrate that DICE enhances the reliability of CLIP-based metrics, making them more aligned with human perception of editing quality. In particular, distinguishing between coherent and non-coherent changes is crucial for obtaining meaningful evaluations. + +# 5. Conclusions + +In this work, we introduced DICE: a novel approach to evaluate instruction-guided image editing by detecting and assessing object-level differences between original and edited images. Our method leverages MLLMs to detect semantic differences between images and determine their coherence with the given editing instructions. This dual-step process allows for a structured and interpretable evaluation, addressing both the accuracy of the detected modifications and their semantic alignment with user intent. + +# Acknowledgments + +We acknowledge the CINECA award under the ISCRA initiative, for the availability of high-performance computing resources. This work has been supported by the EU Horizon projects "ELSA - European Lighthouse on Safe and Secure AI" (No. 101070617), "ELIAS - European Lighthouse of AI for Sustainability" (No. 101120237) and "ELLIOT - European Large Open Multi-Modal Foundation Models For Robust Generalization On Arbitrary Data Streams" (No. 101214398), and by the PNRRM4C2 project "FAIR - Future Artificial Intelligence Research" funded by the EU - NextGenerationEU. + +# References + +[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. GPT-4 Technical Report. arXiv preprint arXiv:2303.08774, 2023. 1, 3, 6 +[2] Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, et al. Qwen2.5-VL Technical Report. arXiv preprint arXiv:2502.13923, 2025. 2 +[3] Federico Betti, Jacopo Staiano, Lorenzo Baraldi, Lorenzo Baraldi, Rita Cucchiara, and Nicu Sebe. Let's ViCE! Mimicking Human Cognitive Behavior in Image Generation Evaluation. In ACM Multimedia, 2023. 2 +[4] Tim Brooks, Aleksander Holynski, and Alexei A Efros. InstructPix2Pix: Learning to Follow Image Editing Instructions. In CVPR, 2023. 1, 2, 3, 7 +[5] Davide Bucciarelli, Nicholas Moratelli, Marcella Cornia, Lorenzo Baraldi, and Rita Cucchiara. Personalizing Multimodal Large Language Models for Image Captioning: An Experimental Analysis. In ECCV, 2024. 3 +[6] Davide Caffagni, Federico Cocchi, Luca Barsellotti, Nicholas Moratelli, Sara Sarto, Lorenzo Baraldi, Lorenzo Baraldi, Marcella Cornia, and Rita Cucchiara. The Revolution of Multimodal Large Language Models: A Survey. In ACL Findings, 2024. 1, 2 +[7] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-End Object Detection with Transformers. In ECCV, 2020. 4, 5 +[8] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging Properties in Self-Supervised Vision Transformers. In ICCV, 2021. 2 +[9] Keqin Chen, Zhao Zhang, Weili Zeng, Richong Zhang, Feng Zhu, and Rui Zhao. Shikra: Unleashing Multimodal LLM's Referential Dialogue Magic. arXiv preprint arXiv:2306.15195, 2023. 3 +[10] Mang Tik Chiu, Yuqian Zhou, Lingzhi Zhang, Zhe Lin, Connelly Barnes, Sohrab Amirghodsi, Eli Shechtman, and Humphrey Shi. Brush2Prompt: Contextual Prompt Generator for Object Inpainting. In CVPR, 2024. 2 + +[11] Federico Cocchi, Nicholas Moratelli, Davide Caffagni, Sara Sarto, Lorenzo Baraldi, Marcella Cornia, and Rita Cucchiara. LLaVA-MORE: A Comparative Study of LLMs and Visual Backbones for Enhanced Visual Instruction Tuning. arXiv preprint arXiv:2503.15621, 2025. 3 +[12] Tim Dettmers, Artidoro Pagnoni, Ari Holtzman, and Luke Zettlemoyer. QLoRA: Efficient Finetuning of Quantized LLMs. In NeurIPS, 2023. 5 +[13] Prafulla Dhariwal and Alexander Nichol. Diffusion Models Beat Gans on Image Synthesis. In NeurIPS, 2021. 2 +[14] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, et al. The Llama 3 Herd of Models. arXiv preprint arXiv:2407.21783, 2024. 1 +[15] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, et al. Scaling Rectified Flow Transformers for High-Resolution Image Synthesis. In ICML, 2024. 1, 2 +[16] Tsu-Jui Fu, Wenze Hu, Xianzhi Du, William Yang Wang, Yinfei Yang, and Zhe Gan. Guiding Instruction-based Image Editing via Multimodal Large Language Models. In ICLR, 2024. 2, 3, 7, 8 +[17] Leon A Gatys, Alexander S Ecker, and Matthias Bethge. Image style transfer using convolutional neural networks. In CVPR, 2016. 2 +[18] Zigang Geng, Binxin Yang, Tiankai Hang, Chen Li, Shuyang Gu, Ting Zhang, Jianmin Bao, Zheng Zhang, Houqiang Li, Han Hu, et al. InstructDiffusion: A Generalist Modeling Interface for Vision Tasks. In CVPR, 2024. 1, 2, 5, 7, 8 +[19] Ross Girshick. Fast R-CNN. In ICCV, 2015. 4 +[20] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative Adversarial Nets. In NeurIPS, 2014. 2 +[21] Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, et al. DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning. arXiv preprint arXiv:2501.12948, 2025. 1 +[22] Agrim Gupta, Piotr Dollar, and Ross Girshick. LVIS: A Dataset for Large Vocabulary Instance Segmentation. In CVPR, 2019. 4, 1 +[23] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. CLIPScore: A Reference-free Evaluation Metric for Image Captioning. In EMNLP, 2021. 7 +[24] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, et al. LoRA: Low-Rank Adaptation of Large Language Models. In ICLR, 2022. 5 +[25] Yi Huang, Jiancheng Huang, Yifan Liu, Mingfu Yan, Jiaxi Lv, Jianzhuang Liu, Wei Xiong, He Zhang, Shifeng Chen, and Liangliang Cao. Diffusion Model-Based Image Editing: A Survey. arXiv preprint arXiv:2402.17525, 2024. 1 +[26] Yuzhou Huang, Liangbin Xie, Xintao Wang, Ziyang Yuan, Xiaodong Cun, Yixiao Ge, Jiantao Zhou, Chao Dong, Rui Huang, Ruimao Zhang, et al. SmartEdit: Exploring Complex Instruction-based Image Editing with Multimodal Large Language Models. In CVPR, 2024. 2 + +[27] Mude Hui, Siwei Yang, Bingchen Zhao, Yichun Shi, Heng Wang, Peng Wang, Yuyin Zhou, and Cihang Xie. HQ-Edit: A High-Quality Dataset for Instruction-based Image Editing. arXiv preprint arXiv:2404.09990, 2024. 2, 3 +[28] Tero Karras, Samuli Laine, and Timo Aila. A Style-Based Generator Architecture for Generative Adversarial Networks. In CVPR, 2019. 2 +[29] Xin Lai, Zhuotao Tian, Yukang Chen, Yanwei Li, Yuhui Yuan, Shu Liu, and Jiaya Jia. LISA: Reasoning Segmentation via Large Language Model. In CVPR, 2024. 3 +[30] Hugo Laurençon, Andres Marafioti, Victor Sanh, and Léo Tronchon. Building and better understanding vision-language models: insights and future directions. In NeurIPS Workshops, 2024. 3, 5, 6 +[31] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft COCO: Common Objects in Context. In ECCV, 2014. 5 +[32] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual Instruction Tuning. In NeurIPS, 2023. 1, 3 +[33] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Qing Jiang, Chunyuan Li, et al. Grounding DINO: Marrying DINO with Grounded Pre-Training for Open-Set Object Detection. In ECCV, 2024. 5, 6, 1 +[34] Ilya Loshchilov and Frank Hutter. Decoupled Weight Decay Regularization. In ICLR, 2019. 1 +[35] Yiwei Ma, Jiayi Ji, Ke Ye, Weihuang Lin, Zhibin Wang, Yonghan Zheng, Qiang Zhou, Xiaoshuai Sun, and Rongrong Ji. I2EBench: A Comprehensive Benchmark for Instruction-based Image Editing. In NeurIPS, 2024. 2, 3, 5, 7 +[36] Matthias Minderer, Alexey Gritsenko, Austin Stone, Maxim Neumann, Dirk Weissenborn, Alexey Dosovitskiy, Aravindh Mahendran, Anurag Arnab, Mostafa Dehghani, Zhuoran Shen, et al. Simple Open-Vocabulary Object Detection with Vision Transformers. In ECCV, 2022. 1 +[37] Matthias Minderer, Alexey Gritsenko, and Neil Houlsby. Scaling Open-Vocabulary Object Detection. In NeurIPS, 2023. 5, 6, 1 +[38] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. DINOv2: Learning Robust Visual Features without Supervision. arXiv preprint arXiv:2304.07193, 2023. 4 +[39] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. SDXL: Improving Latent Diffusion Models for High-Resolution Image Synthesis. arXiv preprint arXiv:2307.01952, 2023. 1, 2 +[40] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, et al. Learning Transferable Visual Models From Natural Language Supervision. In ICML, 2021. 2 +[41] Hanoona Rasheed, Muhammad Maaz, Sahal Shaji, Abdelrahman Shaker, Salman Khan, Hisham Cholakkal, Rao M Anwer, Eric Xing, Ming-Hsuan Yang, and Fahad S Khan. GLaMM: Pixel Grounding Large Multimodal Model. In CVPR, 2024. 6, 2 + +[42] Anton Razzhigaev, Arseniy Shakhmatov, Anastasia Maltseva, Vladimir Arkhipkin, Igor Pavlov, Ilya Ryabov, Angelina Kuts, Alexander Panchenko, et al. Kandinsky: an Improved Text-to-Image Synthesis with Image Prior and Latent Diffusion. arXiv preprint arXiv:2310.03502, 2023. 5, 1 +[43] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-Resolution image Synthesis with Latent Diffusion Models. In CVPR, 2022. 1, 2 +[44] Shelly Sheynin, Adam Polyak, Uriel Singer, Yuval Kirstain, Amit Zohar, Oron Ashual, Devi Parikh, and Yaniv Taigman. Emu Edit: Precise Image Editing via Recognition and Generation Tasks. In CVPR, 2024. 3, 5 +[45] Jing Shi, Ning Xu, Trung Bui, Franck Dernoncourt, Zheng Wen, and Chenliang Xu. A benchmark and baseline for language-driven image editing. In ACCV, 2020. 2 +[46] Jing Shi, Ning Xu, Yihang Xu, Trung Bui, Franck Dernoncourt, and Chenliang Xu. Learning by Planning: Language-Guided Global Image Editing. In CVPR, 2021. 2 +[47] Roman Suvorov, Elizaveta Logacheva, Anton Mashikhin, Anastasia Remizova, Arsenii Ashukha, Aleksei Silvestrov, Naejin Kong, Harshith Goka, Kiwoong Park, and Victor Lempitsky. Resolution-robust Large Mask Inpainting with Fourier Convolutions. In WACV, 2022. 2, 5 +[48] Gemma Team, Aishwarya Kamath, Johan Ferret, Shreya Pathak, Nino Vieillard, Ramona Merhej, Sarah Perrin, Tatiana Matejovicova, Alexandre Rame, et al. Gemma 3 Technical Report. arXiv preprint arXiv:2503.19786, 2025. 2 +[49] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, et al. Qwen2-VL: Enhancing Vision-Language Model's Perception of the World at Any Resolution. arXiv preprint arXiv:2409.12191, 2024. 1, 3, 5, 6 +[50] Sheng-Yu Wang, Oliver Wang, Richard Zhang, Andrew Owens, and Alexei A Efros. CNN-generated images are surprisingly easy to spot... for now. In CVPR, 2020. 1 +[51] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-Thought Prompting Elicit Insights Reasoning in Large Language Models. In NeurIPS, 2022. 6 +[52] Jiabo Ye, Haiyang Xu, Haowei Liu, Anwen Hu, Ming Yan, Qi Qian, Ji Zhang, Fei Huang, and Jingren Zhou. mPLUG-Owl3: Towards Long Image-Sequence Understanding in Multi-Modal Large Language Models. arXiv preprint arXiv:2408.04840, 2024. 1, 3, 5, 6 +[53] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid Loss for Language Image Pre-Training. In ICCV, 2023. 1 +[54] Kai Zhang, Lingbo Mo, Wenhu Chen, Huan Sun, and Yu Su. MagicBrush: A Manually Annotated Dataset for Instruction-Guided Image Editing. In NeurIPS, 2023. 1, 2, 3, 7 +[55] Shu Zhang, Xinyi Yang, Yihao Feng, Can Qin, Chia-Chih Chen, Ning Yu, Zeyuan Chen, Huan Wang, Silvio Savarese, Stefano Ermon, et al. HIVE: Harnessing Human Feedback for Instructional Visual Editing. In CVPR, 2024. 2, 3, 7 +[56] Yichi Zhang, Ziqiao Ma, Xiaofeng Gao, Suhaila Shakiah, Qiaozi Gao, and Joyce Chai. GROUNDHOG: Grounding Large Language Models to Holistic Segmentation. In CVPR, 2024. 3 \ No newline at end of file diff --git a/ICCV/2025/What Changed_ Detecting and Evaluating Instruction-Guided Image Edits with Multimodal Large Language Models/images.zip b/ICCV/2025/What Changed_ Detecting and Evaluating Instruction-Guided Image Edits with Multimodal Large Language Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..a84ef512a47dbd6f323f8712263c6002041940ba --- /dev/null +++ b/ICCV/2025/What Changed_ Detecting and Evaluating Instruction-Guided Image Edits with Multimodal Large Language Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:727178d8d1470293a182e3f41dc06d8a2dbb538071520cb4bd716bf00107ff4c +size 602443 diff --git a/ICCV/2025/What Changed_ Detecting and Evaluating Instruction-Guided Image Edits with Multimodal Large Language Models/layout.json b/ICCV/2025/What Changed_ Detecting and Evaluating Instruction-Guided Image Edits with Multimodal Large Language Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..471d5540929a6ed4bb2e911da45ed16dec76cbb6 --- /dev/null +++ b/ICCV/2025/What Changed_ Detecting and Evaluating Instruction-Guided Image Edits with Multimodal Large Language Models/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7976216dccf07d4588fab55dad33689ebf1bc20addac73bc0e2da08d206beef7 +size 437763 diff --git a/ICCV/2025/What If_ Understanding Motion Through Sparse Interactions/c33d9dcc-fde3-4776-81ca-8cda367e6981_content_list.json b/ICCV/2025/What If_ Understanding Motion Through Sparse Interactions/c33d9dcc-fde3-4776-81ca-8cda367e6981_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..d8ace3935bb17123418e07689182f9231247f316 --- /dev/null +++ b/ICCV/2025/What If_ Understanding Motion Through Sparse Interactions/c33d9dcc-fde3-4776-81ca-8cda367e6981_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1ab11d8cac6750fc1cb1f9ed1222761f0d928d06f534453eb18e4f0fd43c2f5c +size 80321 diff --git a/ICCV/2025/What If_ Understanding Motion Through Sparse Interactions/c33d9dcc-fde3-4776-81ca-8cda367e6981_model.json b/ICCV/2025/What If_ Understanding Motion Through Sparse Interactions/c33d9dcc-fde3-4776-81ca-8cda367e6981_model.json new file mode 100644 index 0000000000000000000000000000000000000000..5fcbac4cfc17ef7b319cb890b5971904d94b9be4 --- /dev/null +++ b/ICCV/2025/What If_ Understanding Motion Through Sparse Interactions/c33d9dcc-fde3-4776-81ca-8cda367e6981_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:52b99a4cf516734a423880f7952b6dc2ad481da6e83880ca32a477a5f53a8b59 +size 99381 diff --git a/ICCV/2025/What If_ Understanding Motion Through Sparse Interactions/c33d9dcc-fde3-4776-81ca-8cda367e6981_origin.pdf b/ICCV/2025/What If_ Understanding Motion Through Sparse Interactions/c33d9dcc-fde3-4776-81ca-8cda367e6981_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..5ae8ce90315f7a19eddeacfb2c72ae90d419d5b2 --- /dev/null +++ b/ICCV/2025/What If_ Understanding Motion Through Sparse Interactions/c33d9dcc-fde3-4776-81ca-8cda367e6981_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8c3c4b172e2df688f4e30191872aa82d1f84a8609a857c49d45295141cb1cd68 +size 28008887 diff --git a/ICCV/2025/What If_ Understanding Motion Through Sparse Interactions/full.md b/ICCV/2025/What If_ Understanding Motion Through Sparse Interactions/full.md new file mode 100644 index 0000000000000000000000000000000000000000..8f049567a41c9e14e25c0a6fa823e51eefd56db1 --- /dev/null +++ b/ICCV/2025/What If_ Understanding Motion Through Sparse Interactions/full.md @@ -0,0 +1,270 @@ +# What If: Understanding Motion Through Sparse Interactions + +Stefan Andreas Baumann* Nick Stracke* Timy Phan* Björn Ommer + +CompVis @ LMU Munich Munich Center for Machine Learning (MCML) + +# Abstract + +Understanding the dynamics of a physical scene involves reasoning about the diverse ways it can potentially change, especially as a result of local interactions. We present the Flow Poke Transformer (FPT), a novel framework for directly predicting the distribution of local motion, conditioned on sparse interactions termed "pokes". Unlike traditional methods that typically only enable dense sampling of a single realization of scene dynamics, FPT provides an interpretable directly accessible representation of multi-modal scene motion, its dependency on physical interactions and the inherent uncertainties of scene dynamics. + +We also evaluate our model on several downstream tasks to enable comparisons with prior methods and highlight the flexibility of our approach. On dense face motion generation, our generic pre-trained model surpasses specialized baselines. FPT can be fine-tuned in strongly out-of-distribution tasks such as synthetic datasets to enable significant improvements over in-domain methods in articulated object motion estimation. Additionally, predicting explicit motion distributions directly enables our method to achieve competitive performance on tasks like moving part segmentation from pokes which further demonstrates the versatility of our FPT. Code and models are publicly available at compvis.github.io/flow-poke-transformer. + +# 1. Introduction + +A key feat of human visual intelligence is motion understanding, our ability to understand and predict the various ways the world around us could potentially change at a given point in time (see Fig. 1). Our cortex is not creating a mental video, focusing on how the colors of individual pixels change. Rather, we are constantly making predictions about the various ways individual objects or parts thereof could potentially move and deform [30]. We do not perceive the future as an unambiguously deterministic sequence of events but as a vast space of possibilities. + +![](images/5083bff78608a3b70f485484f6c18c7310b7469960cd15e23450d64dcd788180.jpg) +Figure 1. What If: Our Flow Poke Transformer directly models the uncertainty of the world by predicting distributions of how objects $(\times)$ may move conditioned on some input movements (pokes, $\rightarrow$ ). We see that whether the hand (below paw) or the paw (above hand) moves downwards directly influences the other's movement. Left: the paw pushing the hand down, will force the hand downwards, resulting in a unimodal distribution. Right: the hand moving down results in two modes, the paw following along or staying put. + +It is natural to focus selectively on parts of a scene, infer how they might evolve, and reason about the underlying physical properties and interactions that drive change. This selective, probabilistic, and multimodal reasoning is rooted in the perceived inherent stochastic nature of the world. It is governed by the stochastic physical properties of complex systems and further compounded by the presence of agents with a complex, inaccessible internal state, lead by free will or other, from the outside often unapproachable causes. + +This inherent uncertainty makes dense, deterministic predictions of future motion both impractical and ill-posed to represent real-world dynamics. The prediction of pixel + +perfect and long-term sequences [28] requires models to commit to one trajectory and, in so doing, ignore the rich multimodality of real-world outcomes. At best, such a prediction biases the scenario towards a single plausible future; at worst, it produces photorealistic frames that show limited understanding of physical processes, interactions, and constraints. In many situations, like autonomous agent systems, robotics, and automated planning, the ability to predict and process multiple possible outcomes of a given situation is more valuable than the naive assumption that events play out according to a single trajectory. + +To address these issues, we propose a framework for representing the distribution of possible motions of parts of a scene. To control the degree of uncertainty, a human observer can interact with or perturb a scene with local "pokes", by nudging an object or applying a force. By repeating similar interactions, the multimodal nature of potential outcomes can be observed. Similarly, we allow conditioning the motion distribution on such sparse, localized pokes. Compared to traditional dense approaches, our method operates at a higher level of abstraction, predicting localized distributions of motion rather than committing to a single outcome. This approach aligns more closely with real-world dynamics where uncertainty and multimodality are intrinsic and actionable insights often emerge from reasoning about sparse, local changes rather than exhaustive dense predictions. These include aspects like the inherent interpretability of explicitly predicted distributions, such as identifying modes and quantifying uncertainty directly. + +By avoiding dense (video) prediction, our model reframes motion prediction as a problem of capturing potential dynamics, directly predicting motion distributions. For instance, a poke applied to an unstable stack of blocks might cause it to topple in multiple ways, remain stable, or shift slightly without collapsing. We capture this variability, avoiding the pitfalls of dense video models that have to commit to one specific sample in the set of potential outcomes. This also addresses the impracticality of dense or long-term predictions, where compounding uncertainty renders dense outputs increasingly arbitrary. + +Applications of our proposed model include (sparse) interactive simulation, where pokes guide scene exploration and the multimodal distributions of possible motions are directly captured, moving part segmentation, but also classic sampling of dense motion predictions. As opposed to optical flow estimation and tracking, where the future is given via a future frame, from which motion is estimated, we predict what future motion might be from only a single frame. + +Overall, we present a step toward a more efficient, flexible, and detailed understanding of scene dynamics. We focus on the vast distribution of what could happen—optionally conditioned on sparse interactions instead of rendering specific futures. Our framework is not only efficient and scalable, + +but also conceptually aligned with the inherent uncertainties our human perception and reasoning is facing when dealing with our changing environment. + +Our main contributions are as follows: + +- Multimodal Distribution Prediction: we directly predict full distributions of potential motion instead of just enabling sampling from them, providing increased flexibility in applications over previous approaches, such as directly estimating uncertainties. +- Sparse Kinematics Modeling: our method reasons about sparse, local motion distributions across the scene. This balances efficiency with expressive power by focusing computational resources where they matter most. +- Generalizability: our method can learn a generic motion understanding from unstructured web videos, generalizing effectively to diverse, open-world data. +- Efficiency: our approach of sparsely modeling interactions enables sparse predictions with our method in 25ms and throughputs of more than 160k parallel predictions per second on a single modern GPU which is promising for real-time applications. + +# 2. Related Work + +Estimation of plausible motion for a given image or scene has been approached in various ways over the years. What makes this task particularly challenging is that it requires the model to have a physical understanding of how objects move in general, how they can be manipulated, and how they relate to each other. Many approaches directly predict a video from still images which makes it harder to access and leverage the underlying motion understanding of the model. Other approaches first predict a dense flow map which they use to later warp the images. However, more complex scenes can have multiple instantiations of realistic motion depending on the given conditioning, which we aim to model directly. In the following, we review various methods which have been studied in the literature. + +Motion-based Editing A field that has recently gained attention is image editing using diffusion models by providing a set of pokes that indicate how specific parts of the image should move. [31] takes a GAN [13] generated image and warps it using motion supervision based on user-provided pokes. InstantDrag [36] on the other hand first predicts dense optical flow using a GAN and uses that as conditioning for a diffusion model to generate the final warped image. + +Similar approaches have been used in video editing [5, 23, 45] that extend base models with ControlNets [54] or LoRAs [16] to condition the model on the desired motion. The goal is then to move entire objects according to a specific poke by selecting the object with a bounding box or entity representation [46]. Unlike our method, this general + +direction neglects an understanding of physical and realistic motion in exchange for precise adherence to the poke guidance and realistic inpainting of occluded regions. + +Motion Generation Various works learn to hallucinate motion for static images [2, 11, 25, 33, 35, 43]. MoVideo [25] and Motion-I2V [35] use diffusion models to predict dense flow sequences given a start frame and use them to synthesize videos. Motion-I2V specifically allows conditioning on sparse movement information using "motion drags", similar to pokes, which is an improvement upon DragNUWA [50] that directly synthesizes dense RGB video from drags. The latter makes the actual motion prediction substantially less accessible because it needs to be estimated with an additional model like RAFT [40] or COTR [18], a property also shared by other methods [20, 21]. [43] predicts discrete bins of optical flow for static images with a classification loss. This enables them to model multiple flow fields for a single image. Im2Flow [11] predicts a single realization of continuous optical flow for an image and combines that with the image to boost action classification performance. + +Learning how objects move and behave together can also be used as a general pretext task to build physical scene understanding. [2] introduced the concept of pokes as sparse motion conditioning to indicate how the poked object should move and directly synthesize dense RGB videos on limited-domain datasets. This is an unspecified problem as movement information is only available for a small number of poked pixels and the model needs to learn how the remainder of the scene moves. DragAPart [21] and the follow-up work PuppetMaster [20] focus on modeling the movement of individual parts of objects for a closed-domain, synthetic dataset (part-level motion). While these works focus on building a more fine-grained physical understanding, they directly predict the result of the $\mathrm{poke(s)}$ in RGB space. This makes the underlying motion representation harder to access and requires e.g. optical flow estimation between frames. Additionally, they do not provide any uncertainty estimation in the form of an underlying motion distribution, but simply render a single possible sample of the result space. + +Other approaches focus on directly predicting a specific physical representation of motion. Generative Image Dynamics [24] learns oscillatory dynamics as commonly found in nature using Fourier-based motion representations. PhysDreamer [56] and PhysGaussian [49] extend the work of [24] from 2D to 3D scenes. While these approaches work well in their limited domains, they lack the flexibility to model the vast, often non-oscillatory motion space of the real world and are thus limited in the amount of general motion understanding they can obtain though training. + +Generative Models in Computer Vision Generative models have recently become a cornerstone in computer vision, + +as they allow modeling tasks through full conditional distributions $p(\mathbf{y}|\mathbf{x})$ instead of reducing predictions to single-point estimates, such as the expectation $\mathbb{E}[\mathbf{y}|\mathbf{x}]$ often used in discriminative models. Major paradigms include GANs [13], diffusion models [15, 37], and autoregressive (AR) models [42]. GANs and diffusion models enable sampling from the modeled distribution but provide limited direct insight into its structure. Diffusion models, in particular, have demonstrated scalability to general data distributions and billions of parameters [32], whereas GANs are typically constrained to closed-set distributions. AR models, extensively applied in NLP [42] and increasingly adopted for vision tasks [3, 9, 51], directly model probability mass functions (PMFs) for discrete distributions and scale well to hundreds of billions of parameters [8]. However, the discrete nature of PMFs limits their applicability to real-valued problems, which are prevalent in vision tasks. Recent advances such as GIVT [41] and [22] have extended AR transformers to continuous-valued outputs while retaining the scalability of AR transformer models [10]. Specifically, [22] employ a diffusion model to sample from autoregressively predicted feature vectors, while GIVT directly parameterizes distributions as Gaussian Mixture Models (GMMs) with diagonal covariances, enabling direct access to the probability density function (PDF) for downstream applications. Our implementation builds upon the latter, while extending it to non-diagonal covariances to enable accurate modeling of motion distributions. + +# 3. Method + +# 3.1. Problem Setting + +Given an image $\mathcal{I}$ , we aim to model the movement of all visible points in the image and their interdependencies. To this end, our goal is to model the conditional distribution $p(\mathbf{f}(\mathbf{q})|\mathcal{P},\mathcal{I})$ of the movement $\mathbf{f}(\cdot)\in \mathbb{R}^2$ of arbitrary query points $\mathbf{q}\in \mathbb{R}^2$ in the image conditioned on a set of $N_{p}$ pokes $\mathcal{P} = \{(\mathbf{p}_i,\mathbf{f}(\mathbf{p}_i))\}_{i = 1}^{N_p}$ , each specifying the movement $\mathbf{f}(\mathbf{p}_i)$ at locations $\mathbf{p}_i\in \mathbb{R}^2$ . Explicit conditioning on movement information given at specific points is crucial to enable the exploration of interactions and controlling movement prediction in the scene. Here, the movement $\mathbf{f}(\cdot)$ of a point describes its change in position from the current time $t$ to a future time $t + \Delta t$ , also referred to as forward flow. + +# 3.2. Flow Poke Transformer + +To model the movement distribution $p_{\theta}(\mathbf{f}(\mathbf{q})|\mathcal{P},\mathcal{I})$ , we use a transformer-based architecture, denoted as $p_{\theta}$ . Transformers [42] are especially well-suited for this task, as they are well-capable of working with sparse sequences due to token interactions only being implemented via the attention mechanism. We show a high-level overview in Fig. 2. We view each poke $(\mathbf{p}_i,\mathbf{f}(\mathbf{p}_i))\in \mathcal{P}$ and each query point $\mathbf{q}_j\in \mathcal{Q}$ as + +![](images/59bedf2d043261e39cb25f3b4e228a0bb4cd00d76ec8754aa8bf7030f00dfe4d.jpg) +Figure 2. High-level Model Architecture Overview. Given an image $\mathcal{I}$ , a set of given pokes $\mathcal{P}$ (visualized as arrows $\rightarrow$ ), and query positions $\mathbf{q}(\times)$ , our model directly predicts an explicit distribution of the movement at each query position. The flow poke transformer cross-attends to features from a jointly trained image encoder to incorporate visual information. Crucially, our architecture represents movement at individual points $\mathbf{q}$ (enabling sparse & off-grid motion processing) and directly predicts continuous, multimodal output distributions. + +individual tokens. Each poke's movement $\mathbf{f}(\mathbf{p})$ is encoded at the input using a Fourier embedding, while query tokens are set to a learned embedding. Positional encoding is implemented using relative positional embeddings [38], allowing positions to be set with arbitrary precision without needing to conform to any grid. This is important, as it enables training the model with high-quality but sparse and off-grid flow obtained via optical tracking. During self-attention, queries only attend to themselves and pokes, not to other queries. This enables evaluating the distribution $p_{\theta}(\mathbf{f}(\mathbf{q})|\mathcal{P},\mathcal{I})$ for multiple queries $\mathbf{q}_j$ in parallel, which is crucial for efficient dense flow predictions. The image $\mathcal{I}$ is encoded separately using a vision transformer, resulting in a set of spatial encoded image tokens $\mathcal{E}(\mathcal{I})$ . The poke and query tokens then cross-attend to the image tokens, with spatial information again encoded using relative positional embeddings [38]. + +To obtain the movement distribution $p_{\theta}(\mathbf{f}(\mathbf{q})|\mathcal{P},\mathcal{I})$ a projection head at the transformer's output directly predicts a Gaussian Mixture Model (GMM), enabling real-valued distributional predictions following GIVT [41]. The distribution being directly accessible in this manner enables a range of additional capabilities, such as directly capturing multi-modal distributions in a single forward pass or enabling the fine-grained quantification of uncertainty. Unlike [41], we parametrize each component $n$ using a full covariance matrix $\Sigma^{(n)}\in \mathbb{R}^{2\times 2}$ instead of a purely diagonal one, greatly increasing the prediction's degrees of freedom. The positive semi-definiteness of the covariance matrix is ensured by the model predicting a lower triangular matrix $\mathbf{L}^{(n)}\in \mathbb{R}^{2\times 2}$ with a positive diagonal (by soft-clipping to a lower threshold), from which the covariance matrix is computed as $\Sigma^{(n)} = \mathbf{L}^{(n)}(\mathbf{L}^{(n)})^{\top}$ . Overall, this results in the predicted $N$ -component GMM + +$$ +p _ {\theta} = \sum_ {n = 1} ^ {N} \pi^ {(n)} \cdot \mathcal {N} \left(\boldsymbol {\mu} ^ {(n)}, \boldsymbol {\Sigma} ^ {(n)}\right), \tag {1} +$$ + +with component mixture coefficients $\pi^{(n)}$ and means $\pmb{\mu}^{(n)}$ . + +A specific challenge with learning motion understanding from open-world web videos is that camera movement can dominate the overall motion distribution of a frame. Only training on videos with static cameras is not viable, as it would limit potential training data too much. We address this by replacing the typical normalization layers in the transformer with adaptive normalization layers [17], using which we condition the model on whether the camera is static. This allows us to learn motion prediction on general videos. + +Training Objective We directly train our model to minimize the negative log-likelihood (NLL) of a ground truth flow $\mathbf{f}(\mathbf{q})$ of the random query point $\mathbf{q}$ , conditioned on a random set of flow pokes $\mathcal{P}$ + +$$ +\begin{array}{l} \mathcal {L} (\mathbf {f} (\mathbf {q}), \mathcal {P}, \mathcal {I}; \theta) = - \log p _ {\theta} (\mathbf {f} (\mathbf {q}) | \mathcal {P}, \mathcal {I}) \\ = - \log \left(\sum_ {n = 1} ^ {N} \pi^ {(n)} \mathcal {N} (\mathbf {f} (\mathbf {q}) | \boldsymbol {\mu} _ {\theta} ^ {(n)} (\mathcal {P}, \mathcal {I}), \boldsymbol {\Sigma} _ {\theta} ^ {(n)} (\mathcal {P}, \mathcal {I}))\right). \tag {2} \\ \end{array} +$$ + +Specifically, we compute the loss for image $\mathcal{I}$ conditioned on random sets of pokes $\mathcal{P}^{(i)}$ of length $|\mathcal{P}^{(0)}| = 0,\ldots ,|\mathcal{P}^{(N_p)}| = N_p$ . Predicting the flow distribution at $N_{q}$ different random query positions $\mathbf{q}$ per set of pokes. To enable efficient training, we introduce a variation of teacher forcing [39], typically used to train autoregressive transformers [42]. We select the set of random pokes such that $|\mathcal{P}^{(0)}| \subset |\mathcal{P}^{(1)}| \subset \ldots \subset |\mathcal{P}^{(N_p)}|$ . We then use a causal attention mask on the poke tokens and let the queries for each set of pokes individually attend to all the pokes in their respective set. We call the resulting attention pattern query-causal attention (see Fig. A for visualizations). As opposed to training with independent sets of pokes and full self-attention for all sets of pokes and queries, this reduces the computational complexity from $\mathcal{O}(N_p^2 \cdot N_q^2)$ to $\mathcal{O}(N_p^2 + N_p \cdot N_q)$ for the same number of trained predictions. Since $N_{p}$ can be large during training, this substantially improves performance and enables efficient training. + +# 3.3. Downstream Applications + +Our method's primary goal is to enable efficient and interpretable modeling of multimodal movement distributions of different parts in scenes. We achieve this by modeling the conditional distribution of the movement of query points $\mathbf{q}$ in the image $\mathcal{I}$ given the movement of any number of pokes $\mathcal{P}$ to condition on. As our model directly makes probability density functions for each query's movement accessible and captures its multi-modality (c.f., Fig. 3), the distribution of potential movements can be directly interpreted. Besides modeling this relation of movement of different parts in a scene, this also enables other direct and indirect downstream tasks and applications, which we describe in this section. + +Dense Motion Prediction The conditional distribution our method learns to model can also be used to predict a dense grid of queries $\mathcal{Q}$ . This prediction can be done both purely conditioned on the reference image or given reference pokes $\mathcal{P}$ . This motion can be obtained by predicting the pointwise flow distributions in parallel or via autoregressive sampling. To sample from the joint distribution $p(\mathbf{f}(\mathcal{Q})|\mathcal{P},\mathcal{I})$ that models the precise interactions between all points in the image, we employ autoregressive sampling. Iteratively, we predict the flow distribution for a random query $\mathbf{q}_i\in \mathcal{Q}$ , sample a flow instance from the conditional distribution $\mathbf{f}(\mathbf{q}_i)\sim p_\theta (\mathbf{f}(\mathbf{q}_i)|\mathcal{P},\mathcal{I})$ , and add the query to the set of pokes $\mathcal{P}\gets \mathcal{P}\cup \{\mathbf{q}_i\}$ . This results in individual coherent samples $\mathbf{F}_{\mathrm{sample}}\sim p_{\theta}(\mathbf{f}(\mathcal{Q})|\mathcal{P},\mathcal{I})$ from the distribution of possible dense flows. We show qualitative examples of sampled dense flow in Fig. 6. For parallel sampling, we compute the mean dense flow $\mathbf{F}_{\mathrm{mean}} = \mathbb{E}\left[\mathbf{f}(\mathcal{Q})|\mathcal{P},\mathcal{I};\theta \right]$ in a pointwise manner for all queries $\mathbf{q}_i\in \mathcal{Q}$ in parallel. This provides efficient, high-quality flow predictions (see, e.g., appendix Fig. F), but also results in mode averaging. + +Segmenting Moving Parts Segmenting parts that move together is a task introduced in [21] and is useful for various applications such as predicting affordances. Given a poke $(\mathbf{p},\mathbf{f}(\mathbf{p}))$ , the aim is to segment the image into parts that would move in response to it. Unlike [21], we do not need to rely on involved methods for extracting and comparing internal feature activations of our model for this task. Instead, our model enables direct quantification of the effect a movement $\mathbf{f}(\mathbf{p})$ of a point $\mathbf{p}$ has on another point $\mathbf{q}$ by measuring the relative entropy between the conditional and unconditional distribution, i.e., how much conditioning on $\mathbf{p}$ changes the movement distribution of $\mathbf{q}$ . This is done using the Kullback-Leibler (KL) divergence + +$$ +D _ {\mathrm {K L}} \left(p _ {\theta} (\mathbf {f} (\mathbf {q}) | (\mathbf {p}, \mathbf {f} (\mathbf {p})), \mathcal {I}) \parallel p _ {\theta} (\mathbf {f} (\mathbf {q}) | \mathcal {I})\right). \tag {3} +$$ + +Specifically, if the movements of $\mathbf{q}$ and $\mathbf{p}$ are independent, the conditional distribution $p(\mathbf{f}(\mathbf{q}) | (\mathbf{p}, \mathbf{f}(\mathbf{p})), \mathcal{I})$ is equal to + +the marginal distribution $p(\mathbf{f}(\mathbf{q})|\mathcal{I})$ , and thus, the KL divergence in Eq. (3) is zero. Otherwise, it quantifies the change in movement distribution, and, thus, the motion interdependencies of different parts of the scene. We efficiently approximate the KL divergence using the matched bound approximation [12]. This can then be computed over all points $\mathbf{q}$ in the image $\mathcal{I}$ in parallel and directly quantifies the effect the given movement of $\mathbf{p}$ has on each point $\mathbf{q}$ . + +# 4. Experiments + +# 4.1. Dataset and Implementation Details + +For general pretraining, we train on a random 3.8M video clip subset of WebVid [1]. The wide variety of concepts present in WebVid enables our model to learn a general representation for motion instead of being limited to a specific domain, such as face-only videos. We train our model with flow from optical tracks using CoTracker3 [19] for a random 48-frame interval from each clip using a uniform $48^2$ grid from the respective start frames. + +The image encoder and the poke transformer are ViT-Base transformers [7] for a total parameter count of 220M. We use RoPE [4, 38] both for self-attention between flow tokens and for cross-attention to image tokens. We initialize the image encoder with DINOv2-R [6, 29] to make training more efficient but keep the weights unlocked. Jointly training the full model is essential, as DINOv2 does not have good instance segmentation capabilities (see Sec. B for additional details), which are essential for our task. We pass images to the vision encoder at a resolution of $448^{2}$ to obtain a $32^{2}$ grid of visual embedding tokens. The model is trained in bfloat16 precision for 800k steps using AdamW [26] with a learning rate of 5e-5, at a global batch size of 32 images, which is increased to 128 after 250k steps. Per image, we sample sets of random pokes of sizes 0, 1, ..., 128, and compute losses on $N_{q} = 15$ random query points per set of pokes. This results in a global batch size of 61,440 queries. Overall, training our model takes 7 days on 2 Nvidia H200s. Also see appendix Secs. A and B for additional details and ablations and Sec. D for an extension to 3D. + +Without inference optimizations such as quantization or 8-bit inference, a single conditional movement distribution prediction for query in an image can be obtained in less than $25\mathrm{ms}$ of delay on a single H200. This makes our model applicable for real-time applications. Throughput (with parallel predictions) is about $160\mathrm{k}$ predictions per second per image, thanks to our query-causal attention implementation. + +We generally evaluate predicted motion at a resolution of $64^2$ unless specified otherwise. We primarily rely on endpoint error $\mathrm{EPE} = \| \hat{\mathbf{f}} (\mathbf{q}) - \mathbf{f}_{\mathrm{GT}}(\mathbf{q})\| _2$ , which measures the difference between the true motion $\mathbf{f}_{\mathrm{GT}}(\mathbf{q})$ and the predicted motion $\hat{\mathbf{f}} (\mathbf{q})$ . Additionally, we also compute the percentage of correct keypoints PCK $= \mathbb{E}[\| \hat{\mathbf{f}} (\mathbf{q}) - \mathbf{f}_{\mathrm{GT}}(\mathbf{q})\|_2 < \alpha ]$ + +![](images/55013812ec36fd0d7eb88e2f2a8f74b15db819c2c7a5eca6924b5489a47592fb.jpg) +Figure 3. Multimodal Motion Distribution Prediction. We condition on one or multiple pokes $(\rightarrow)$ and then query the motion distribution of specific points $(\times)$ . Our model's predictions capture the multi-modal nature of motion and exhibit understanding of interactions, such as only lifting the cup by its handle not necessarily causing the whole cup to move upwards, while grabbing it at stable points does. It also demonstrates prior understanding from scenes, such as a car in an intersection being more likely to move forwards than backward and cars in traffic likely moving together. + +![](images/67fd2cc23735c39409bad2e912b00f8f8a6c3e9635859bf467d9a3a282d4fd63.jpg) + +with $\alpha = 1\mathrm{px}$ unless specified otherwise. + +# 4.2. Evaluation of FPT's Key Abilities + +FPT's Ability to Predict Movement Distributions We observe our model's predicted distributions $p_{\theta}(\mathbf{f}(\mathbf{q})|\mathcal{P},\mathcal{I})$ given an image $\mathcal{I}$ of a scene and conditioned on a sparse set of pokes $\mathcal{P}$ in Fig. 3 (see appendix Fig. E for additional samples). Qualitatively, our model exhibits an understanding of physical phenomena and interactions, predicting realistic movement distributions for the given pokes in the context of the respective scenes. Most importantly, it captures the multimodality of potential movements in different circumstances and their variability/uncertainty. Our approach is trained in an open-world setting, being not limited to individual object categories, but, nevertheless, captures fine-grained details of specific objects' potential motion. + +One can also draw samples from the joint distribution of movement of the whole scene $p_{\theta}(\mathbf{f}(\mathcal{Q})|\mathcal{P},\mathcal{I})$ by autoregressive sampling. We show examples of such unconditional motion generations in Fig. 6. They successfully show diverse but realistic global scene motion. + +FPT predicts meaningful multimodal motion distributions. One important property that differentiates FPT from common motion modeling approaches is that it directly predicts the multimodality of possible future motions. For these multimodal predictions to be valuable, they should cover the diverse modes of possible motion and have meaningful predicted confidences $\pi^{(n)}$ . We analyze these properties in + +![](images/f06c3edbc6ffba507aadda72a9e00c0dafc225983caaff8a288eca3d58f5e21d.jpg) +Figure 4. Predicted Mode Analysis. Values are computed at a resolution of $64^2$ . (a) Diversity of predicted modes is high, with mode variation covering a large fraction of poke magnitude. (b) One mode typically has a substantially higher confidence than others, which increases with given poke count. The mode closest to the ground truth consistently has a higher-than-average confidence. (c) More confident modes are more accurate as measured by PCK. + +![](images/8f2a753782d89dd347359875c10c8c33115783fe83d7c3ba25943c6ef13c7930.jpg) + +![](images/cb23d93c8f694040aa9d6f87bf4cbe6c48fcd422c7d0abc0d38dabb2e1fae1eb.jpg) + +![](images/f76c3a067d91d54b6982b84da292e034722dbfd08e7784d47555fb4cc9bed235.jpg) +Figure 5. Uncertainty Calibration. We find that the motion prediction error measured by EPE strongly correlates with the predicted uncertainty (Pearson $\rho = 0.64$ ). This relationship holds for low & high numbers of given pokes. + +![](images/eab46c69a022251e7dc51d2b3bbf5856be3efd5c7db217a697ef9f6c646c63c3.jpg) + +Fig. 4. Generally, we find the modes to be highly diverse (Fig. 4a, cf. Figs. 3 and E), with them covering substantially different movements. As expected, the multimodal predictions reduce to primarily unimodal predictions when enough conditioning information is available to reduce the stochastic uncertainty of the future and discern one clear correct mode (Fig. 4b). Importantly, the confidence of the mode closest to the ground truth motion is consistently substantially higher than the average. This indicates that the model's confidence predictions are meaningful. Similarly, analyzing the modes' accuracy (Fig. 4c) shows that the model assigns higher confidences to modes more likely to be correct. Still, secondary and tertiary predicted modes are also meaningful, as indicated by the accuracy of the mode closest to the ground truth exceeding that of the most confident one. + +FPT's predicted distributions accurately model uncertainty. We evaluate the predictive quality of our model's predicted uncertainty w.r.t. true prediction error in Fig. 5. Specifically, we investigate the relation between the predicted distribution's standard deviation $\mathrm{Std}[\mathbf{f}(\mathbf{q})|\mathcal{P},\mathcal{I};\theta]$ and the motion estimation error as measured with the endpoint error (EPE). We find that the predicted motion's error strongly correlates to the predicted uncertainty. This capability is independent of the approach to derive the single motion prediction from the predicted distribution. Sampling from the predicted distribution (Pearson $\rho = 0.66$ ), using its mean $(\rho = 0.64)$ , or using the most confident mode $(\rho = 0.62)$ all lead to high predictive accuracy of the true prediction error compared to the ground truth. + +
MethodTrained On1 Poke2 Pokes5 Pokes10 Pokes100 Pokes
EPE ↓PCK ↑LPIPS ↓EPE ↓PCK ↑LPIPS ↓EPE ↓PCK ↑LPIPS ↓EPE ↓PCK ↑LPIPS ↓EPE ↓PCK ↑LPIPS ↓
InstantDrag [36]Faces9.240.1930.189.120.1960.178.820.1970.178.390.1980.167.290.2120.15
Motion-I2V [35]Generic (Zero-Shot)29.080.0290.3527.400.0310.3424.220.0300.3220.900.0480.30n/an/an/a
OursGeneric (Zero-Shot)7.640.1500.166.870.1540.155.320.1670.134.200.1830.122.510.2640.10
+ +Table 1. Face Motion Generation Evaluation. We evaluate the accuracy of predicted motion on TalkingHead-1KH [44] given a starting frame and one or more pokes (partially) defining the head movement. Our method performs substantially better in a zero-shot comparison to Motion-I2V, which was also trained in a generic setting. Compared to InstantDrag, which was trained for specifically this setting, our method achieves a substantially better endpoint error (EPE) but slightly worse PCK for low poke counts, highlighting our model's capability to perform competitively with purpose-trained methods while being generic. It can also make more efficient use of the available information, achieving greater accuracy gains from additional pokes compared to other methods. When using the predicted motions to warp the source + +image, our method consistently outperforms others. + +![](images/984fd4fb138039167211a3a52032d25c4a5e95bfa873dc1d092bb0de44445a7b.jpg) +Figure 6. Unconditional AR Motion Sampling. We show samples of generated flow without prior motion conditioning on pokes. Our model can generate a wide variety of realistic motions. + +# 4.3. Comparisons on Motion Prediction + +To enable quantitative comparisons of our model's motion understanding with existing methods, we evaluate on predicting dense flow from sparse flow pokes, given a starting frame. We compare against baseline methods in the setting they have been trained on to enable fair comparisons. To evaluate against DragAPart [21] and PuppetMaster [20], which only generate images/videos based on drags, we extract flow from the source to the generated image using RAFT [40]. + +Face Motion Estimation We compare against Instant-Drag [36], which was trained on CelebV-Text [52] in their evaluation setting - poke-conditioned motion prediction on aligned faces on TalkingHead-1KH [44]. The qualitative results (see Fig. 7) show that our model tends to predict more accurate and localized motion. This can also be observed when visualizing the motion by warping the image. For quantitative evaluations, we extract chunks of length 0.8s (following [36]) and use CoTracker3 [19] at a grid size of $128^2$ to obtain the target motion from the start to the end frame. Then, we condition on $N \in \{1, 2, 5, 10, 100\}$ pokes $\mathcal{P}$ , where the first poke is chosen to be the one with the largest flow magnitude, and the others are sampled randomly, and compare the dense predicted motion to the target + +![](images/f3722cff1074019a65dcf0dbcf4f12a58c3afb399026325437cf5408ab6ec1c2.jpg) +Figure 7. Fine-grained Face Motion Control. We show fine-grained zero-shot poking results on faces and compare against InstantDrag [36], which was trained for this task. We further visualize the predicted motion as warps using I-D's face warping model. + +downsampled by a factor of two. We also compare quantitatively with [35], which was trained generically. We compare favorably to both methods in EPE independent of the number of given pokes, indicating further that our predicted flow is more precise (Tab. 1). The PCK is slightly worse than the non-generically trained InstantDrag for low poke counts but catches up with more conditioning. When using the respective method's generated motion to perform image warping using InstantDrag's warping stage, our method consistently outperforms both others, as measured by the LPIPS [55] distance to the ground truth images. + +Articulated Objects DragAPart [21] and PuppetMaster [20] were explicitly trained on part-level object motion on a synthetic dataset, Drag-A-Move [21] (DAM). To enable comparisons with them, we evaluate on the test set of DAM, which contains synthetic images of furniture with one or more pokes $\mathcal{P}$ that define the movement of one or more articulated parts of that object. It also provides ground truth dense flow at a resolution of $512^{2}$ , which we compare against. As DAM is significantly dissimilar from our train set, we evaluate our model both in a zero-shot and fine-tuned setting. For the fine-tuned setting, we fine-tune our model on DAM for 30k steps at a batch size of 128 with an exponentially decaying learning rate of 5e-7 that halves every 10k steps. Our quantitative evaluation (see Tab. 2, for additional qualitative samples see appendix Fig. C) shows + +
MethodTrained On(a) Motion Est. (b) Moving Part Segm.
EPE ↓PCK ↑mIoU ↑
Motion-I2V [35]Generic (Zero-Shot)33.270.0430.073†
DragAPart [21]Objects (DAM)9.690.5140.273†
PuppetMaster [20]Objects (DAM + OAHQ)9.620.4720.112
OursGeneric (Zero-Shot)12.740.1910.287
Ours (fine-tuned)Generic → DAM3.570.8340.572
+ +taken from original publication, our evaluation yields 0.228. +Table 2. Articulated Object Motion Estimation. We compare motion (flow) estimation and moving part segmentation performance on Drag-A-Move [21] (DAM). On zero-shot motion estimation, our model substantially outperforms the other zero-shot method M-I2V, while not being much worse than specifically trained methods. When adapted, our method significantly outperforms previous approaches. In moving part segmentation, even our generic model outperforms other, in-domain models. + +that, in a zero-shot setting, our model's predicted motion is substantially more accurate than Motion-I2V [35] but less accurate than the specifically trained DragAPart and PuppetMaster. This demonstrates that our model generalizes to out-of-distribution data comparatively well, but, expectedly, falls short of models explicitly trained for this OOD domain. Once fine-tuned, our model outperforms even the purpose-made DragAPart and PuppetMaster by a wide margin. + +# 4.4. Segmenting Moving Parts + +We perform moving part segmentation with our method by thresholding the KL divergence between the pointwise unconditional motion distribution and the pointwise motion distribution conditioned on a specific poke (Eq. (3)). We show qualitative results in Fig. 9 and compare quantitatively against other methods, similarly thresholding the flow magnitude in Tab. 2b. Here, we find that our method, especially when finetuned in-domain, outperforms DragAPart [21], which introduced this benchmark, and the other methods by a wide margin. A large part of this gain can be attributed to our divergence-based score, which leverages FPT's unique property of directly predicting distributions. Without it, our method achieves an mIoU of 0.415 - still outperforming the previous state-of-the-art by a wide margin, but less so. Our method is also robust to threshold choice - halving/doubling it leads to mIoUs of 0.54/0.52 respectively. + +![](images/c6373ee761f2166fb72f58ed9c08cec6fa72a0263e7d340551b1cc5e62a91ffb.jpg) +Figure 8. Common Failure Cases. Our model, generically pretrained on primarily realistic videos, does not generalize well to cartoons, causing parts of the background to be moved together with objects. Additionally, our model sometimes, but not consistently, jointly predicts the movement of shadows together with objects, which can be problematic for downstream use cases. + +![](images/76485b253f507f61dfe8b51204701e49d359675f963c610ab451891dc8cd769a.jpg) +(a) We directly replicate Fig. 7 from DragAPart [21] with our method. Our method provides spatially continuous predictions and makes fewer critical mistakes like segmenting the furniture body with the drawer (top right). + +Figure 9. Segmenting Moving Parts. We show qualitative results for moving part segmentation, as introduced in [21], both on the Drag-A-Move dataset (a) and in a generic, open-set setting (b). +![](images/e8698c394699ad998a8d2acece797369219610a7fbe369d10cd3f00e90fd4750.jpg) +(b) Open-set moving part dependency visualization. The degree to which the movement of each part is influenced by the poke $(\rightarrow)$ is visualized as a heatmap, where brighter color means a higher degree of influence. + +# 5. Conclusion + +We introduce the Flow Poke Transformer (FPT), a novel framework for motion understanding that captures real-world dynamics' multi-modal and stochastic nature through interpretable distributions of local motion, conditioned on targeted interactions (pokes). Contrary to previous motion prediction approaches, FPT directly models the probabilistic distribution of possible outcomes, providing insights into the effects of physical interactions and inherent uncertainties. + +Our evaluations demonstrate FPT's versatility across different domains and its generalization capabilities. Despite being designed and trained for sparse, general-purpose motion understanding, it also offers competitive performance in established tasks such as dense motion generation on faces or articulated objects. Importantly, while valuable for comparison, these evaluations do not fully reflect our method's primary strength – its ability to provide directly interpretable and useable predictions of motion distributions in interactive environments, bridging the gap between physical plausibility and efficiency. Furthermore, capabilities such as moving part segmentation directly emerge from our method's design. Overall, these results show our method's strength as a versatile, interpretable, general-purpose motion model. We envision this work as a foundation for more probabilistic, generalizable, and actionable approaches to motion understanding, paving the way for deeper insights into complex physical dynamics and future advancements in handling ambiguous and extreme out-of-distribution scenarios (Fig. 8). + +# Acknowledgment + +We would like to thank Mahdi M. Kalayeh for helpful discussions and feedback. We would further like to thank Kim-Louis Barwig, Enrico Shippole, Ming Gui, Thomas Ressler-Antal, Olga Grebenkova, Paul Hofman, Owen Vincent, and the anonymous reviewers. This work was supported in part by a research grant from Netflix. The authors thank Netflix for its support. This project was also supported by the Federal Ministry for Economic Affairs and Energy within the project "NXT GEN AI METHODS - Generative Methoden für Perzeption, Prädiktion und Planung", the project "GeniusRobot" (01IS24083) funded by the Federal Ministry of Research, Technology and Space (BMFTR), the bidt project KLIMA-MEMES, and the Horizon Europe project ELLIOT (GA No. 101214398). The authors gratefully acknowledge the Gauss Center for Supercomputing for providing compute through the NIC on JUWELS/JUPITER at JSC and the HPC resources supplied by the NHR @FAU Erlangen. + +# References + +[1] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In IEEE International Conference on Computer Vision, 2021. 5 +[2] Andreas Blattmann, Timo Milbich, Michael Dorkenwald, and Björn Ommer. ipoke: Poking a still image for controlled stochastic video synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14707-14717, 2021. 3 +[3] Chenjie Cao, Yuxin Hong, Xiang Li, Chengrong Wang, Chengming Xu, Yanwei Fu, and Xiangyang Xue. The image local autoregressive transformer. Advances in Neural Information Processing Systems, 34:18433-18445, 2021. 3 +[4] Katherine Crowson, Stefan Andreas Baumann, Alex Birch, Tanishq Mathew Abraham, Daniel Z Kaplan, and Enrico Shippole. Scalable high-resolution pixel-space image synthesis with hourglass diffusion transformers. In Proceedings of the 41st International Conference on Machine Learning, pages 9550-9575. PMLR, 2024. 5, 1 +[5] Zuozhuo Dai, Zhenghao Zhang, Yao Yao, Bingxue Qiu, Siyu Zhu, Long Qin, and Weizhi Wang. Fine-grained open domain image animation with motion guidance. arXiv preprint arXiv:2311.12886, 2023. 2 +[6] Timothee Darcet, Maxime Oquab, Julien Mairal, and Piotr Bojanowski. Vision transformers need registers, 2023. 5 +[7] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 5 +[8] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The + +llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 3 +[9] Patrick Esser, Robin Rombach, and Björn Ommer. Taming transformers for high-resolution image synthesis. 2021 IEEE. In CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 3 +[10] Lijie Fan, Tianhong Li, Siyang Qin, Yuanzhen Li, Chen Sun, Michael Rubinstein, Deqing Sun, Kaiming He, and Yonglong Tian. Fluid: Scaling autoregressive text-to-image generative models with continuous tokens. arXiv preprint arXiv:2410.13863, 2024. 3 +[11] Ruohan Gao, Bo Xiong, and Kristen Grauman. Im2flow: Motion hallucination from static images for action recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5937-5947, 2018. 3 +[12] Goldberger and Greenspan. An efficient image similarity measure based on approximations of kl-divergence between two gaussian mixtures. In Proceedings Ninth IEEE International conference on computer vision, pages 487-493. IEEE, 2003. 5 +[13] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2014. 2, 3 +[14] Dan Hendrycks and Kevin Gimpel. Bridging nonlinearities and stochastic regularizers with gaussian error linear units, 2017. 1 +[15] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 3 +[16] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 2 +[17] Xun Huang and Serge Belongie. Arbitrary style transfer in real-time with adaptive instance normalization. In Proceedings of the IEEE international conference on computer vision, pages 1501-1510, 2017. 4 +[18] Wei Jiang, Eduard Trulls, Jan Hosang, Andrea Tagliasacchi, and Kwang Moo Yi. Cotr: Correspondence transformer for matching across images. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6207-6217, 2021. 3 +[19] Nikita Karaev, Iurii Makarov, Jianyuan Wang, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. CoTracker3: Simpler and better point tracking by pseudolabelling real videos. 2024. 5, 7 +[20] Ruining Li, Chuanxia Zheng, Christian Rupprecht, and Andrea Vedaldi. Puppet-master: Scaling interactive video generation as a motion prior for part-level dynamics. arXiv preprint arXiv:2408.04631, 2024. 3, 7, 8 +[21] Ruining Li, Chuanxia Zheng, Christian Rupprecht, and Andrea Vedaldi. Dragapart: Learning a part-level motion prior for articulated objects. In Computer Vision - ECCV 2024, pages 165-183, Cham, 2025. Springer Nature Switzerland. 3, 5, 7, 8 + +[22] Tianhong Li, Yonglong Tian, He Li, Mingyang Deng, and Kaiming He. Autoregressive image generation without vector quantization. NeurIPS 2024, 2024. 3 +[23] Yaowei Li, Xintao Wang, Zhaoyang Zhang, Zhouxia Wang, Ziyang Yuan, Liangbin Xie, Yuexian Zou, and Ying Shan. Image conductor: Precision control for interactive video synthesis. arXiv preprint arXiv:2406.15339, 2024. 2 +[24] Zhengqi Li, Richard Tucker, Noah Snavely, and Aleksander Holynski. Generative image dynamics. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24142-24153, 2024. 3 +[25] Jingyun Liang, Yuchen Fan, Kai Zhang, Radu Timofte, Luc Van Gool, and Rakesh Ranjan. Movideo: Motion-aware video generation with diffusion models. In European Conference on Computer Vision, pages 0000-0000, 2024. 3 +[26] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2019. 5, 1 +[27] Kepan Nan, Rui Xie, Penghao Zhou, Tiehan Fan, Zhenheng Yang, Zhijie Chen, Xiang Li, Jian Yang, and Ying Tai. Openvid-1m: A large-scale high-quality dataset for text-to-video generation. arXiv preprint arXiv:2407.02371, 2024. 3 +[28] OpenAI. Sora, 2024. 2 +[29] Maxime Oquab, Timothée Darcet, Theo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Russell Howes, Po-Yao Huang, Hu Xu, Vasu Sharma, Shang-Wen Li, Wojciech Galuba, Mike Rabbat, Mido Assran, Nicolas Ballas, Gabriel Synnaeve, Ishan Misra, Herve Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. Dinov2: Learning robust visual features without supervision, 2023. 5, 2 +[30] Stephen E. Palmer. Vision Science: Photons to Phenomenology. The MIT Press, 1999. 1 +[31] Xingang Pan, Ayush Tewari, Thomas Leimkuhler, Lingjie Liu, Abhimitra Meka, and Christian Theobalt. Drag your gan: Interactive point-based manipulation on the generative image manifold. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-11, 2023. 2 +[32] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. In The Twelfth International Conference on Learning Representations, 2024. 3 +[33] Pol Rosello. Predicting future optical flow from static video frames. Retrieved on: Jul, 18:2, 2016. 3 +[34] Noam Shazeer. Glu variants improve transformer. arXiv preprint arXiv:2002.05202, 2020. 1 +[35] Xiaoyu Shi, Zhaoyang Huang, Fu-Yun Wang, Weikang Bian, Dasong Li, Yi Zhang, Manyuan Zhang, Ka Chun Cheung, Simon See, Hongwei Qin, et al. Motion-i2v: Consistent and controllable image-to-video generation with explicit motion modeling. SIGGRAPH 2024, 2024. 3, 7, 8, 2 +[36] Joonghyuk Shin, Daehyeon Choi, and Jaesik Park. Instantdrag: Improving interactivity in drag-based image editing. arXiv preprint arXiv:2409.08857, 2024. 2, 7 + +[37] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In International Conference on Learning Representations, 2021. 3 +[38] Jianlin Su, Yu Lu, Shengfeng Pan, Bo Wen, and Yunfeng Liu. Roformer: enhanced transformer with rotary position embedding. corr abs/2104.09864 (2021). arXiv preprint arXiv:2104.09864, 2021. 4, 5, 1 +[39] Ilya Sutskever, Oriol Vinyals, and Quoc V Le. Sequence to sequence learning with neural networks. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2014. 4 +[40] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16, pages 402-419. Springer, 2020. 3, 7 +[41] Michael Tschannen, Cian Eastwood, and Fabian Mentzer. Givt: Generative infinite-vocabulary transformers. In European Conference on Computer Vision, pages 292-309. Springer, 2024. 3, 4, 2 +[42] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 3, 4 +[43] Jacob Walker, Abhinav Gupta, and Martial Hebert. Dense optical flow prediction from a static image. In Proceedings of the IEEE international conference on computer vision, pages 2443-2451, 2015. 3 +[44] Ting-Chun Wang, Arun Mallya, and Ming-Yu Liu. One-shot free-view neural talking-head synthesis for video conferencing. In CVPR, 2021. 7 +[45] Zhouxia Wang, Ziyang Yuan, Xintao Wang, Yaowei Li, Tianshui Chen, Menghan Xia, Ping Luo, and Ying Shan. Motionctrl: A unified and flexible motion controller for video generation. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 2 +[46] Weijia Wu, Zhuang Li, Yuchao Gu, Rui Zhao, Yefei He, David Junhao Zhang, Mike Zheng Shou, Yan Li, Tingting Gao, and Di Zhang. Draganything: Motion control for anything using entity representation. In European Conference on Computer Vision, pages 331-348. Springer, 2025. 2 +[47] Yuxi Xiao, Jianyuan Wang, Nan Xue, Nikita Karaev, Iurii Makarov, Bingyi Kang, Xin Zhu, Hujun Bao, Yujun Shen, and Xiaowei Zhou. Spatialtrackerv2: 3d point tracking made easy. In ICCV, 2025. 3 +[48] Yuxi Xiao et al. Spatialtracker: Tracking any 2d pixels in 3d space. In CVPR, 2024. 3 +[49] Tianyi Xie, Zeshun Zong, Yuxing Qiu, Xuan Li, Yutao Feng, Yin Yang, and Chenfanfu Jiang. Physgaussian: Physics-integrated 3d gaussians for generative dynamics. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4389-4398, 2024. 3 +[50] Shengming Yin, Chenfei Wu, Jian Liang, Jie Shi, Houqiang Li, Gong Ming, and Nan Duan. Dragnuwa: Fine-grained control in video generation by integrating text, image, and trajectory. arXiv preprint arXiv:2308.08089, 2023. 3 + +[51] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, Ben Hutchinson, Wei Han, Zarana Parekh, Xin Li, Han Zhang, Jason Baldridge, and Yonghui Wu. Scaling autoregressive models for content-rich text-to-image generation. Transactions on Machine Learning Research, 2022. Featured Certification. 3 +[52] Jianhui Yu, Hao Zhu, Liming Jiang, Chen Change Loy, Weidong Cai, and Wayne Wu. Celebv-text: A large-scale facial text-video dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14805-14814, 2023. 7 +[53] Biao Zhang and Rico Sennrich. Root mean square layer normalization. Advances in Neural Information Processing Systems, 32, 2019. 1 +[54] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 2 +[55] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 7 +[56] Tianyuan Zhang, Hong-Xing Yu, Rundi Wu, Brandon Y Feng, Changxi Zheng, Noah Snavely, Jiajun Wu, and William T Freeman. Physdreamer: Physics-based interaction with 3d objects via video generation. In European Conference on Computer Vision, pages 388-406. Springer, 2025. 3 \ No newline at end of file diff --git a/ICCV/2025/What If_ Understanding Motion Through Sparse Interactions/images.zip b/ICCV/2025/What If_ Understanding Motion Through Sparse Interactions/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..e16645962536bfa870a3e02d8ec6a4a199456a18 --- /dev/null +++ b/ICCV/2025/What If_ Understanding Motion Through Sparse Interactions/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:36a5a0108bf60b8ef99b377ce94246ccd8bff162b3252bb1c461e7d057c0dd53 +size 471387 diff --git a/ICCV/2025/What If_ Understanding Motion Through Sparse Interactions/layout.json b/ICCV/2025/What If_ Understanding Motion Through Sparse Interactions/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..0ea25890e3645cb12ed13cbf8cb20944f82a9a79 --- /dev/null +++ b/ICCV/2025/What If_ Understanding Motion Through Sparse Interactions/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f44d311fe773bac3c415d74c6402c0d823086e6afe11f814dbd58865137d67a6 +size 390212 diff --git a/ICCV/2025/What Makes for Text to 360-degree Panorama Generation with Stable Diffusion_/cf8dae55-332f-44ad-840c-3f6b5a1ebfab_content_list.json b/ICCV/2025/What Makes for Text to 360-degree Panorama Generation with Stable Diffusion_/cf8dae55-332f-44ad-840c-3f6b5a1ebfab_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..eed1fd5ecef0335ad56e65cf82f5ac44d47a43fe --- /dev/null +++ b/ICCV/2025/What Makes for Text to 360-degree Panorama Generation with Stable Diffusion_/cf8dae55-332f-44ad-840c-3f6b5a1ebfab_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:195ad698c6ae9d260c63b1d9d97852244c2d5a6557642837061cbb0c5b9a248a +size 69129 diff --git a/ICCV/2025/What Makes for Text to 360-degree Panorama Generation with Stable Diffusion_/cf8dae55-332f-44ad-840c-3f6b5a1ebfab_model.json b/ICCV/2025/What Makes for Text to 360-degree Panorama Generation with Stable Diffusion_/cf8dae55-332f-44ad-840c-3f6b5a1ebfab_model.json new file mode 100644 index 0000000000000000000000000000000000000000..4e263bb6d3661ff3c817aa2c2b5a882d524486bf --- /dev/null +++ b/ICCV/2025/What Makes for Text to 360-degree Panorama Generation with Stable Diffusion_/cf8dae55-332f-44ad-840c-3f6b5a1ebfab_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:38853855132128f9e4e42245a2d48fd1c0b1283caa53a960f18f6a462dfc426e +size 89938 diff --git a/ICCV/2025/What Makes for Text to 360-degree Panorama Generation with Stable Diffusion_/cf8dae55-332f-44ad-840c-3f6b5a1ebfab_origin.pdf b/ICCV/2025/What Makes for Text to 360-degree Panorama Generation with Stable Diffusion_/cf8dae55-332f-44ad-840c-3f6b5a1ebfab_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..81d8ec4f2c02736f1e774049e45aa8e76dcbbc88 --- /dev/null +++ b/ICCV/2025/What Makes for Text to 360-degree Panorama Generation with Stable Diffusion_/cf8dae55-332f-44ad-840c-3f6b5a1ebfab_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c5574cb069ef270fe91e7f1307adbadb602e5e3d9c788a39a2d391bb0608f1d5 +size 9847356 diff --git a/ICCV/2025/What Makes for Text to 360-degree Panorama Generation with Stable Diffusion_/full.md b/ICCV/2025/What Makes for Text to 360-degree Panorama Generation with Stable Diffusion_/full.md new file mode 100644 index 0000000000000000000000000000000000000000..7d04657e05ac1b8d1aa6b2e03ede6d916e32b404 --- /dev/null +++ b/ICCV/2025/What Makes for Text to 360-degree Panorama Generation with Stable Diffusion_/full.md @@ -0,0 +1,274 @@ +# What Makes for Text to 360-degree Panorama Generation with Stable Diffusion? + +Jinhong Ni $^{1*}$ Chang-Bin Zhang $^{2}$ Qiang Zhang $^{3,4}$ Jing Zhang $^{1}$ $^{1}$ Australian National University + $^{2}$ The University of Hong Kong + $^{3}$ Beijing Innovation Center of Humanoid Robotics + $^{4}$ Hong Kong University of Science and Technology (Guangzhou) + +{jinhong.ni, jing.zhang}@anu.edu.au cbzhang@connect.hku.hk jony.zhang $@$ x-humanoid.com + +# Abstract + +Recent prosperity of text-to-image diffusion models, e.g. Stable Diffusion, has stimulated research to adapt them to 360-degree panorama generation. Prior work has demonstrated the feasibility of using conventional low-rank adaptation techniques on pre-trained diffusion models to generate panoramic images. However, the substantial domain gap between perspective and panoramic images raises questions about the underlying mechanisms enabling this empirical success. We hypothesize and examine that the trainable counterparts exhibit distinct behaviors when finetuned on panoramic data, and such an adaptation conceals some intrinsic mechanism to leverage the prior knowledge within the pre-trained diffusion models. Our analysis reveals the following: 1) the query and key matrices in the attention modules are responsible for common information that can be shared between the panoramic and perspective domains, thus are less relevant to panorama generation; and 2) the value and output weight matrices specialize in adapting pre-trained knowledge to the panoramic domain, playing a more critical role during fine-tuning for panorama generation. We empirically verify these insights by introducing a simple framework called UniPano, with the objective of establishing an elegant baseline for future research. UniPano not only outperforms existing methods but also significantly reduces memory usage and training time compared to prior dual-branch approaches, making it scalable for end-to-end panorama generation with higher resolution. The code is available1. + +# 1. Introduction + +Creating 360-degree panoramic images has gained substantial attention due to its significant potential [20, 49]. Despite the considerable advancement in text-to-image synthesis re + +cently [31-33], generating panoramas from text prompts remains challenging from the following aspect. Panoramic images encompass the entire surrounding view with a 360-degree horizontal and 180-degree vertical field of view, typically represented using equirectangular projection geometry. This results in distinctive features such as a $2:1$ aspect ratio and spherical distortion, setting them apart from standard square perspective images. On top of this, due to the high cost of capturing panoramic images in practice, the panoramic datasets are often relatively scarce, e.g. Matterport3D [5] contains 10,800 panoramic images. The lack of data complicates the training of generative models, as conventional perspective diffusion models [33] generally require billions of text-image pairs for training [36]. + +To mitigate data scarcity, the typical strategy is to fine-tune pre-trained diffusion models for downstream applications [17, 35, 53]. However, as stated in [16], the fundamental structural differences between panoramic and perspective images intuitively suggest that the embedded perspective knowledge within the pre-trained diffusion models may not be readily transferable. Aligning with this intuition, prior work [16, 18, 42] has proposed generating multiple perspective images according to predefined camera poses and stitching them into a panorama. Contrary to the aforementioned intuition, another line of work [52] has demonstrated that fine-tuning pre-trained diffusion models on limited panoramic data using conventional low-rank adaptation (LoRA) [14] still yields effective text-to-panorama generation results. This empirical success suggests the presence of some intrinsic mechanism that enables LoRA to effectively leverage prior knowledge from the pre-trained perspective diffusion models, thereby circumventing the structural differences. This motivates us to explore the following question: What exactly makes for fine-tuning Stable Diffusion for text-to-panorama generation? + +We base our analysis on the LoRA fine-tuning paradigm to study the behaviors and ideally functionalities of all trainable counterparts, particularly their impact on panorama generation, with the ultimate goal of elucidating the mecha + +"Amidst the ruins of an ancient civilization, deciphering hieroglyphics that tell the story of a lost world." + +![](images/9f92267af14d6333ca94e2bcd1e5b4c52f99f8adc561732a692bbff4301576d9.jpg) +"aliving room with a fireplace." +"ahomewithpoolandpatio." + +![](images/9c54ceacdd0a039fdb05382c39af07e0079ebf6033c7fb3afbca512f237092e2.jpg) +Figure 1. Our UniPano can synthesize realistic 360-degree panoramic images by fine-tuning Stable Diffusion. (Top) $1024 \times 2048$ panoramic images generated by UniPano. (Bottom) $512 \times 1024$ panoramic images generated by UniPano. + +![](images/b13feecea2e7768ca83a0cd935c880e2e0ec7508ae216b34cc16256a68fed317.jpg) + +nism that thrives in adapting perspective diffusion models for panorama generation. Our launching point is to isolate the trainable components within LoRA fine-tuning (i.e., $W_{\{q,k,v,o\}}$ , cf. Fig. 2) and examine their relevance for learning panoramic structures. Subsequently, we identify the underlying behaviors of each trainable component when they are tuned jointly. We draw two major empirical findings (cf. Sec. 3.2 for details): + +- $W_{\{q,k\}}$ in the attention blocks fail to learn the panoramic structures when they are trained in isolation, whereas $W_{\{v,o\}}$ both succeed in capturing such information. +- When $W_{\{q,k,v,o\}}$ are jointly trained, $W_{\{v,o\}}$ are responsible for learning panoramic-specific information (i.e., equirectangular structure), whereas $W_{\{q,k\}}$ learn shared knowledge across panoramic and perspective domains that are irrelevant to the panoramic structure. + +Our analysis reveals the following: After fine-tuning with panoramic images, we discover that the query and key + +within the cross-attention blocks capture less panoramic-specific information, namely, they function to 'preserve' or 'enhance' the pre-trained perspective knowledge; In contrast, the value and output weight matrices are responsible for adapting such perspective information into the panoramic domain. Based on the analysis, we believe that fine-tuning the query and key matrices is less relevant to panorama generation, whereas the representational capability of the value and output matrices should be emphasized. This yields our straightforward yet efficacious uni-branch solution, dubbed UniPano, targeting to serve as a simple baseline to foster future research. UniPano achieves state-of-the-art results on $512 \times 1024$ text-to-panorama generation while requiring notably less memory and training time compared to the current SoTA [52]. Thanks to this computational efficiency, UniPano can be scaled to generate panoramic images with even higher resolution in an end-to-end manner, as shown in Fig. 1. + +# 2. Related Work + +Diffusion Models. The recent breakthrough in diffusion models [13, 38, 39] has accelerated the inference process [22, 40, 54, 55] and significantly boosted the generation quality [8, 28]. The prosperity of large-scale pretrained diffusion models [31-33] has prompted various applications in different fields [6, 17, 29, 35], where the adaptation is usually achieved by parameter-efficient fine-tuning techniques such as low-rank adaptation (LoRA) [14], or via distillation [25]. This paper targets the former approach and attempts to demystify what makes for panorama generation by fine-tuning pre-trained diffusion models with LoRA. + +Panorama Generation. Existing works can be roughly divided into two categories, namely panorama outpainting and text-to-panorama generation. The former approach [3, 16, 24, 26, 44, 46] aims to complete a panoramic image based on a partial input image, exemplified by Cube-Diff [16] which proposes to jointly generate six faces of cubemap for panorama generation. Aligning with the prosperity of text-conditioned generation as in the perspective domain, text-to-panorama generation [4, 18, 21, 42, 45, 50-52] has gained attention recently. Among these works, [4, 18, 42, 51] generate a sequence of consistent perspective images and stitch them into a panorama. A separate branch of works fine-tunes the pre-trained text-to-image diffusion models to generate an equirectangular panoramic image in an end-to-end manner [45, 50, 52]. StitchDiffusion [45] fine-tunes pre-trained diffusion models with techniques ensuring panoramic continuity. DiffPano [50] includes multi-view panoramic awareness into the framework. PanoFree [21] stands apart from the aforementioned methods by employing a tuning-free approach to generate panoramas. PanFusion [52] introduces a dual-branch framework by simultaneously generating perspective and panoramic images and ensuring consistency through a cross-branch attention mechanism. Of particular importance to our study, [52] shows that LoRA fine-tuning on Stable Diffusion reports reasonable performance, and our work aims to investigate and elucidate the underlying factors contributing to this empirical success. + +# 3. What Makes for Panorama Generation? + +# 3.1. Preliminary + +Diffusion models [13, 38, 39] generate images by iteratively transforming the noise sampled from the prior distribution into the target data distribution, where each sampling step involves predicting the noise from the input noisy image. We defer details on diffusion models to the supplementary materials. Of particular relevance to our study, diffusion models often involve self- and cross-attention. Formally, + +![](images/41095657620aad125b51ea7129688a655f41a890a4c6c3caab1df52536abbacf.jpg) +Figure 2. Illustration of fine-tuning attention blocks within diffusion models with low-rank adaptation (LoRA). MHA, MatMul, $h$ denotes the multi-head attention, matrix multiplication, and the number of attention heads respectively. + +given an input latent $z_{t}$ and the corresponding condition $y^{2}$ , the attention computes: + +$$ +\operatorname {M H A} \left(z _ {t}, y\right) W _ {o}, +$$ + +where MHA denotes the multi-head attention, and $W_{o}$ represents the output weight; for notational simplicity, we write MHA in the single-head form + +$$ +\operatorname {s o f t m a x} \left(\frac {Q K ^ {\top}}{\sqrt {d _ {h}}}\right) V, +$$ + +with head channel dimension $d_h$ , $Q = z_t W_q$ , $K = y W_k$ , and $V = y W_v$ , where $W_{\{q, k, v, o\}}^3$ are the set of trainable weights. The community has shown that training these attention modules within text-to-image diffusion models using parameter efficient fine-tuning techniques (e.g. LoRA [14]) suffices to surrogate fine-tuning the entire models, exemplified by [10, 35]. We illustrate fine-tuning diffusion models with LoRA in Fig. 2. + +# 3.2. Motivation and Insights + +$360^{\circ}$ panoramic images capture a complete spherical view of the surroundings, involving a full 360-degree horizontal field of view and a vertical field of view of 180 degrees, typically stored in equirectangular format. These unique characteristics in view structures make legitimate panoramic images fundamentally different from the perspective ones, which at first glance, implies that perspective-related + +![](images/113ee7a1953a4d7d7a4217d8c62a444b1cc27e08616650aafc7ab6247fb49272.jpg) + +![](images/9c6eeea9b18110e2a566a93a63e5ab56591b7f64fc0a3164ba450d2f23426019.jpg) + +![](images/c40516b8f0cab289384f97c7e1820d33b9c87951a8260224ac3a63925dedca08.jpg) +(a) Training $W_{q}$ in isolation +(c) Training $W_{v}$ in isolation +Figure 3. Qualitative comparison for training $W_{\{q,k,v,o\}}$ in isolation separately. Training $W_{q}$ or $W_{k}$ in isolation fails to capture the spherical structure, as in (a) and (b); whereas training $W_{v}$ or $W_{o}$ in isolation successfully captures the spherical distortion of the panoramic images, as in (c) and (d). All visualizations are generated with the text prompt "a kitchen with stainless steel appliances". + +![](images/943b383336aef2b202fbdafa20f7dedc2148eca8f1f1c142b3d4a6c4ee1f8fd6.jpg) +(b) Training $W_{k}$ in isolation +(d) Training $W_{o}$ in isolation + +
Panorama20 Views8 Views
FAED↓FID↓FID↓FID↓
Wq10.8681.0930.6627.44
Wk13.2767.6327.0124.47
Wv8.6652.6017.0119.35
Wo9.3852.1720.3220.24
+ +Table 1. Quantitative comparison for training $W_{\{q,k,v,o\}}$ in isolation separately. Training only $W_v$ or $W_o$ reports considerably better FAED and FID than $W_q$ or $W_k$ . Details of reported metrics are in Sec. 4.1. + +knowledge within the pre-trained diffusion models may not be immediately relevant. Contrary to this intuition, previous works [52] have demonstrated the feasibility of directly fine-tuning pre-trained perspective diffusion models (e.g., Stable Diffusion) with LoRA (cf., Sec. 3.1) for text-to-panorama generation, given a relatively scarce set of panoramic data. Evinced by this empirical success, we speculate that such an adaptation to the panoramic domain has to conceal some intrinsic mechanism to leverage perspective knowledge within the pre-trained diffusion models. + +To reveal such a mechanism and elucidate what makes such perspective-based diffusion adaptation succeed in panorama generation, we start by decomposing the trainable components and training them in isolation to identify the behaviors – and ideally the functionalities – of each coun- + +terpart. Specifically, we fine-tune $W_{\{q,k,v,o\}}$ (cf., Fig. 2) separately with LoRA for panorama generation, and showcase the comparison both qualitatively in Fig. 3 and quantitatively in Tab. 1. We highlight the following. + +Observation 3.1. As shown in Fig. 3 (a) and (b), it is evident that training $W_{q}$ or $W_{k}$ in isolation notably fails to capture the spherical structure within panoramic images, while both $W_{v}$ and $W_{o}$ are capable of achieving such goals, as in Fig. 3 (c) and (d). The quantitative results in Tab. 1 also align with the qualitative observation, as training $W_{v}$ or $W_{o}$ in isolation leads to considerably better FAED and FID than $W_{q}$ or $W_{k}$ . + +Conclusion 3.2. The four trainable components $W_{\{q,k,v,o\}}$ within the cross-attention modules exhibit varying abilities to learn the spherical structures for successful adaptation to panorama generation. In particular, $W_{q}$ and $W_{k}$ fail to capture such distortions even when fitted purely on panoramic data, whereas $W_{v}$ and $W_{o}$ both successfully adapt the pre-trained diffusion model to panorama generation when trained in isolation. + +Remark 3.3. Conclusion 3.2 suggests that $W_{\{q,k\}}$ have limited capability to be adapted for panorama generation. As such, excluding them during fine-tuning shall not impact the model's capability to learn the panoramic structure. + +Our subsequent step is to investigate how each component contributes to adaptation when fine-tuned collec + +![](images/3757f25b7d0e7e8615cd2af6843e5c7792f36d1845f1a1b85c99887d9eff7e87.jpg) +Figure 4. Illustration of roles of $W_{\{q,k,v,o\}}$ when jointly fine-tuned. We first fine-tune $W_{\{q,k,v,o\}}$ jointly with LoRA on panoramic data, then optionally deactivate some LoRAs for inference with different purposes. (a) the panoramic image generated by fine-tuning $W_{\{q,k,v,o\}}$ jointly with LoRA, where we simplify the LoRA architecture (cf. Fig. 2); (b) diffusion models with panorama fine-tuned LoRAs can only generate distorted, panoramic-like images, are thus no longer capable of generating perspective images; (c) by excluding $W_{v}$ and $W_{o}$ LoRAs, diffusion models fine-tuned on panoramic data recover the ability to generate perspective images; (d) excluding $W_{q}$ and $W_{k}$ LoRAs does not affect the model's ability to generate panoramic images. + +tively. Knowing that $W_{q}$ and $W_{k}$ are unable to capture the spherical structure characteristic of panoramic images, we hypothesize that their role is limited to learning common knowledge across both the perspective and panoramic domains. To validate this argument, we design the following experiment: we first fine-tune $W_{\{q,k,v,o\}}$ with LoRA jointly on panoramic data, then optionally deactivate some LoRAs during inference to examine their relevance to panorama generation. The illustrations are provided in Fig. 4, from which we draw the observation and conclusion below. + +Observation 3.4. After fine-tuning the diffusion models with panoramic data (Fig. 4 (a)), the model when attempted to generate perspective images with all LoRAs applied can only generate panoramic-like images, as shown in Fig. 4 (b). If the LoRAs associated with $W_{\{v,o\}}$ are excluded during inference, the models with only the remaining $W_{\{q,k\}}$ LoRAs successfully recover the ability to generate valid perspective images, evidenced in Fig. 4 (c). On top of this, excluding $W_{\{q,k\}}$ LoRAs does not affect the model's ability to generate panoramic images, as shown in Fig. 4 (d). + +Conclusion 3.5. When all weights $W_{\{q,k,v,o\}}$ are trained jointly, LoRAs associated with $W_{q}$ and $W_{k}$ learn shared knowledge across panoramic and perspective domains that are irrelevant to panoramic structures, whereas LoRAs associated with $W_{v}$ and $W_{o}$ are responsible for learning the spherical structures of the panoramic images. + +Remark 3.6. Conclusion 3.5 differentiates the trainable components based on their roles. In particular, $W_{\{q,k\}}$ even when trained on panoramic images do not apply any spherical distortion to the generated images, such a finding also aligns with Conclusion 3.2 as they also struggle to learn spherical information when trained isolatedly. Contrarily, + +as $W_{\{v,o\}}$ are responsible for learning the panoramic structure, their capacity needs to be emphasized more during fine-tuning. + +To this end, we have explicated what exactly makes for panorama generation with pre-trained diffusion models. Specifically, after fine-tuning with panoramic images, we find that $W_{q}$ and $W_{k}$ within the cross-attention modules do not capture the spherical distortion within the panorama at all. They act as if they were tuned using perspective images, and their roles are likely to be 'preserving' or 'refining' the pre-trained perspective knowledge. In stark contrast, $W_{v}$ and $W_{o}$ are responsible for adapting the information - captured both in $W_{\{q,k\}}$ and within the pre-trained model itself - into the panoramic domain. Their roles are thus learning the spherical structure of the panorama, which is much more instrumental to the task. + +# 3.3. UniPano + +As a byproduct of our insights, we present a memory-efficient uni-branch fine-tuning framework for adapting pretrained text-to-image diffusion models to panorama generation, dubbed UniPano. The core idea of UniPano is based on Remarks 3.3 and 3.6: in essence, it freezes $W_{\{q,k\}}$ as they are associated with non-parametric-specific information and emphasizes $W_{\{v,o\}}$ as they are critical to capture equirectangular distortion within the panoramas. We highlight that the purpose of UniPano is to empirically verify our insights in Sec. 3.2, and with the aim to provide a simple yet effective baseline for future research. + +Design Choices of $W_{o}$ . We propose to increase the representational capacity of the corresponding modules to enhance model's learning ability for the spherical structure of + +
Panorama20 Views8 Views
FAED↓FID↓FID↓FID↓
Pano Only [52]7.9050.4020.1020.56
DA [48]9.7846.4116.5618.77
SE [15]8.7250.6717.7119.61
LoRA (r=8) [14]8.3448.5816.9419.42
LA [11]7.6550.3918.2419.41
MoE [9, 47]7.2148.8319.5020.05
+ +Table 2. Comparison across several designs for increasing the capacity of $W_{o}$ . Based on the panorama branch only baseline (Pano Only), we compare several plausible attempts to increase the capacity of the output layer ( $W_{o}$ ). Details of reported metrics are in Sec. 4.1. + +panoramic images. We differentiate $W_{v}$ and $W_{o}$ with the intuition that $W_{v}$ is associated with head-wise projection while $W_{o}$ interacts directly with the entire set of representations. Due to this reason, although both components are capable of learning the spherical structure, we choose to increase the capacity of $W_{o}$ because of its directness. We adopt and compare several common strategies to enhance the representational capability: + +- Larger LoRA ranks (LoRA $r = 8$ ) [14] is the most straightforward way to increase capability. This approach simply doubles the LoRA ranks of $W_{o}$ from 4 to 8. +- Local Window Attention (LA) [11, 30] constrains the receptive field of the attention operation to the neighboring pixels. We insert such an attention block before each $W_{o}$ LoRA. +- Deformable Attention (DA) [48] introduces a learnable offset and computes the attention based on the sampled features. We attempt to insert a deformable attention block prior to each $W_{o}$ LoRA. +- Squeeze and Excitation (SE) [15] adaptively recalibrates channel-wise feature responses to strengthen the representational power. We similarly insert a SE block before each $W_{o}$ LoRA. +- Mixture of Experts (MoE) [9, 47] involves computing weighted sum over several expert networks, where the weights are learned via a routing network. We replace each $W_{o}$ LoRA with a MoE module. We adopt the same MoE architecture as [9], except that each expert network is a LoRA, similar to the design in [47]. We apply the same auxiliary loss as [9] for routing load balancing. + +The comparison is detailed in Tab. 2, where we also provide a reference to the panorama branch only (Pano Only) baseline. To ensure a fair comparison with this baseline, we also fine-tune $W_{\{q,k\}}$ LoRAs in all compared strategies. We simply opt to use MoE to enhance the capacity of $W_{o}$ because of its superiority on FAED, while we highlight that such a choice may not be optimal and encourage future work to investigate further. + +# 4. Experiments + +# 4.1. Experimental Setup + +Dataset. Matterport3D dataset [5] is a scene understanding dataset with 10,800 panoramic images. We use the same captions as [52], which are generated by BLIP-2 [19] with a prompt of "a 360 - degree view of". We adopt the same data split as [42, 52], containing 9,820 and 1,092 pairs for training and evaluation respectively. + +Implementation Details. To facilitate a fair comparison, we strictly follow [42, 52] to train our model using AdamW optimizer [23] with a batch size of 4 and a learning rate of $2 \times 10^{-4}$ for 10 epochs, with identical cosine annealing learning rate scheduler. Following [52], we base our model on Stable Diffusion 2 base version. + +Evaluation Metrics. We follow previous works to evaluate the generated panoramic images in the panorama [7, 52] and perspective [42] domain. + +- Panorama. Following [7, 52], we report Fréchet Inception Distance (FID) and Inception Score (IS) to measure the quality and realism of the generated panoramas. In addition, we report the CLIP Score (CS) to evaluate text-image consistency. Since both FID and IS are based on InceptionNet [41] which is trained using perspective images only, we follow [52] to report a panoramic-customized metric Fréchet Auto-Encoder Distance (FAED) [26] for panorama evaluation. +- Perspective. We follow [52] to randomly sample 20 perspective views to simulate practical navigation on panoramas, and these views are evaluated based on FID and IS. Following [42], we also report FID, IS, and CS on 8 horizontally evenly spaced views. + +As stated in [52], IS evaluates the diversity of objects within the generated image, as such, lower IS does not necessarily reflect the quality and realism of images in case models do not tend to generate unexpected objects. Similarly, as in [52], higher CS may be due to the repetition of objects to strengthen text-image alignment. On top of these, we also note that the 20 randomly sampled views may capture the top or the bottom of the panoramas, which are often blurred even on real panoramic images, impacting the evaluation quality. For the above reasons, we emphasize FAED and FID while caring horizontal FID more than 20-view FID among all evaluation metrics. + +# 4.2. Main Results + +Compared Methods. We compare our uni-branch approach with several baseline methods. + +- MVDiffusion [42] trains a multi-view diffusion model to simultaneously generate 8 horizontal views, which can be stitched into a panorama. + +
MethodsPeak Mem.†(GB)Dur.†(hrs)Panorama20 ViewsHorizontal 8 Views
FAED↓FID↓IS↑CS↑FID↓IS↑FID↓IS↑CS↑
SD+LoRA [14, 33, 52]31.692.267.1951.694.4028.8319.326.9020.686.4824.77
MVDiffusion [42]26.66 (-15.9%)9.86------25.276.9026.34
Pano Only [52]31.81 (+0.4%)2.337.9050.404.5428.6720.107.0620.566.3724.85
PanFusion [52]60.12 (+89.7%)6.616.0446.474.3628.5817.046.8519.886.5024.98
UniPano (Ours)32.59 (+2.8%)3.435.9046.474.1628.3717.096.7417.746.0024.82
+ +Table 3. Comparison between SoTA methods on $512 \times 1024$ panorama generation. We quantitatively evaluate the panorama images based on Fréchet Auto-Encoder Distance (FAED) Fréchet Inception Distance (FID), Inception Score (IS), and CLIP Score (CS). We follow [52] to randomly sample 20 views from a panoramic image and [42] to horizontally sample 8 evenly spaced views to evaluate the quality of cropped perspective images. We report the peak allocated GPU memory (Peak Mem.) and time duration (Dur.) for 10-epoch training. All evaluated results are based on Stable Diffusion 2 base. †: results are reproduced with FP32 precision. + +![](images/369b32199015df8ae4ca0dc7fdcd264e505be5b39eadcc41d74f064679216de9.jpg) +Figure 5. Selected qualitative comparisons between UniPano (Ours) and PanFusion. We show the generated panoramic image for each text prompt and 4 randomly sampled horizontal perspective views below. We highlight notable artifacts such as non-perspective lines with red boxes. More qualitative comparisons can be found in the supplementary material. + +- $SD + LoRA$ [14, 33] is the baseline method which finetunes the Stable Diffusion model [33] with LoRA [14] on panoramic images. +- Pano Only [52] is another baseline method introduced in [52] which additionally includes circular padding on top of SD+LoRA to ensure loop consistency. +- PanFusion [52] is the SoTA solution to date, which adopts a dual-branch approach and adds a cross-attention module between panoramic and perspective branches to ensure consistency. + +Quantitative Comparison. We present the quantitative comparison in Tab. 3. Our UniPano achieves state-of-the-art FAED and horizontal FID, while being on par with PanFusion on FID and 20-view FID. We highlight that UniPano introduces minimal computational overhead, with an additional $2.8\%$ allocated GPU memory and about 1 additional hour of training compared to SD+LoRA baseline. In comparison, training PanFusion [52] almost double the allocated GPU memory $(+89.7\%)$ and almost triple the time + +required for training compared to SD+LoRA baseline. + +Qualitative Comparison. We showcase the qualitative comparison in Fig. 5. PanFusion may sometimes generate panoramic images with invalid equirectangular projection, evidenced by the notable artifacts (curvy and panoramic-like lines) in the regions of the perspective views highlighted with red boxes. With the boosted capacity of the panoramic-specific modules, UniPano faithfully generates panoramic images that follow equirectangular projection in the illustrated cases. More qualitative comparisons are deferred to the supplementary material. + +# 4.3. Ablation Study + +Different fine-tuning strategies. We compare several different fine-tuning strategies for trainable weights $W_{\{q,k,v,o\}}$ , in Tab. 4, with different combinations of the following: freezing, fine-tuning with LoRA [14], and with mixture of experts (MoE) [9, 37, 47]. The panorama branch only baseline (highlighted in gray) fine-tunes all + +
W{q,k}WvWoPanorama20 Views8 Views
FAED↓FID↓FID↓FID↓
LoRALoRALoRA7.9050.4020.1020.56
LoRALoRAMoE7.2148.8319.5020.05
*LoRALoRA7.9948.6219.2719.30
*LoRAMoE5.9046.4717.0917.74
+ +Table 4. Ablation study on fine-tuning strategies. We compare several fine-tuning strategies for weights $W_{\{q,k,v,o\}}$ in attention modules, with various combinations of freezing (*), LoRA fine-tuning [14], and MoE [37, 47]. + +
Panorama20 Views8 Views
nkFAED↓FID↓FID↓FID↓
226.7551.0118.2719.71
425.9046.4717.0917.74
826.1747.1020.1919.51
847.3147.6122.0521.12
+ +Table 5. Ablation study on the mixture of experts in $W_{o}$ . We experiment with different hyperparameters for the mixture of experts (MoE) by adjusting the number of experts $n$ , and selecting top- $k$ experts with the highest weighting. + +trainable components with LoRA and is thereby considered as the baseline. Switching to MoE for fine-tuning $W_{o}$ contributes to a notable improvement in FAED and FID metrics, demonstrating the benefits of increasing the capacity for panoramic-specific components. Based on the baseline, we further experiment with keeping $W_{\{q,k\}}$ frozen throughout fine-tuning, resulting in improved FID scores while the FAED remained comparable to the baseline. This result is in line with our analysis that $W_{\{q,k\}}$ relate to non-parametric-specific information. Merging these two strategies - i.e., freezing $W_{\{q,k\}}$ and fine-tuning $W_{o}$ with MoE - yields our state-of-the-art UniPano. + +Different settings for mixture of experts. We additionally ablate different sets of hyperparameters for the mixture of experts (MoE) in Tab. 5, namely the number of experts $n$ and the number of top- $k$ selected experts per token. Our first observation is that unlike many other MoE applications [9, 37, 47], the performance in our context saturates for a relatively small amount of experts ( $n = 4$ ), evidenced by the notable deterioration in FID-related metrics when scaling $n$ from 4 to 8. This is likely due to the relative simplicity of fine-tuning a pre-trained diffusion model, in comparison to the typical use cases of MoE i.e. training the entire model from scratch. Additionally, aligning with the previous work [9, 37, 47], we find that sparsity is critical, as increasing the number of used experts per token $k$ from 2 to 4 reduces the FAED and FID metrics rapidly. + +![](images/c094dd2f88b03eb050afaa2d15b5f20ec8748ddf0dd6d922250efd97e772ef61.jpg) +Figure 6. Failure cases of UniPano. Similar to PanFusion [52], UniPano sometimes generates scenes with invalid layouts, such as rooms without entrances. + +![](images/f2c27cdb35ab80c38014d618c7cd6165413ca28d38a475dd5e765e79deef2d52.jpg) + +# 4.4. Scaling to Higher Resolution + +Panoramic images store the entire 360-degree surrounding scenes within one equirectangular image, cropping perspective views from a panoramic image up to $512 \times 1024$ thus still leading to somewhat low-resolution images. This motivates the importance of scaling panorama generation to higher resolution, which is typically achieved with a separate super-resolution stage [7]. As a direct benefit of lowering the memory burden of PanFusion, UniPano can be readily scaled for higher-resolution panorama generation in an end-to-end manner. As Stable Diffusion 2 base is optimized for generating images up to $512 \times 512$ , directly adapting it to generate higher-resolution images leads to suboptimal results. We thus adopt the state-of-the-art Stable Diffusion 3 [8] which natively supports $1024 \times 1024$ image generation. We show UniPano based on Stable Diffusion 3 can generate realistic $1024 \times 2048$ panoramic images, and is robust to out-of-distribution prompts and extremely complex prompts, in Fig. 1. More high-resolution results and experimental details are deferred to the supplementary material. + +# 5. Conclusion + +We have elucidated the underlying mechanism that facilitates low-rank adaptation of pre-trained perspective diffusion models to panorama generation. Particularly, our analysis reveals that the query and key matrices $(W_{\{q,k\}})$ learn common semantic information that can be shared between the panoramic and perspective domains, whereas the value and output matrices $(W_{\{v,o\}})$ specialize in capturing the equirectangular structure of panoramic images. Based on these insights, we propose UniPano, which outperforms and reduces the memory and training burden compared to the previous dual-branch approach. + +Limitations. The primary focus of this paper is to investigate the underlying behaviors of the trainable components within LoRAs when adapting pre-trained perspective diffusion models to panorama generation. While our UniPano reports the state-of-the-art results, given the abundance of hyperparameters, the performance is still possibly far from optimal. Additionally, similar to the drawbacks of PanFusion [52], we find that UniPano sometimes generates scenes with invalid layouts, such as rooms without entrances, as shown in Fig. 6. + +Acknowledgment. This research was, in part, funded by the U.S. Government - DARPA TIAMAT HR00112490421. The views and conclusions contained in this document are those of the authors and should not be interpreted as representing the official policies, either expressed or implied, of the U.S. Government. + +# References + +[1] Chief architect $360^{\circ}$ panorama renderings. https://www.chiefarchitect.com/products/360-panorama-viewer. Accessed: 2025-03-03. 11 +[2] Online $360^{\circ}$ panorama viewer vr. https://renderstuff.com/tools/360-panorama-web-viewer. Accessed: 2025-03-03. 11 +[3] Naofumi Akimoto, Yuhi Matsuo, and Yoshimitsu Aoki. Diverse plausible 360-degree image outpainting for efficient 3dcg background creation. In CVPR, 2022. 3 +[4] Omer Bar-Tal, Lior Yariv, Yaron Lipman, and Tali Dekel. Multidiffusion: Fusing diffusion paths for controlled image generation. In ICML, 2023. 3 +[5] Angel Chang, Angela Dai, Thomas Funkhouser, Maciej Halber, Matthias Niessner, Manolis Savva, Shuran Song, Andy Zeng, and Yinda Zhang. Matterport3d: Learning from rgb-d data in indoor environments. In 3DV, 2017. 1, 6, 13 +[6] Shoufa Chen, Peize Sun, Yibing Song, and Ping Luo. Diffusiondet: Diffusion model for object detection. In ICCV, 2023. 3 +[7] Zhaoxi Chen, Guangcong Wang, and Ziwei Liu. Text2light: Zero-shot text-driven hdr panorama generation. TOG, 2022. 6, 8 +[8] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In ICML, 2024. 3, 8 +[9] Zhengcong Fei, Mingyuan Fan, Changqian Yu, Debang Li, and Junshi Huang. Scaling diffusion transformers to 16 billion parameters. arXiv preprint arXiv:2407.11633, 2024. 6, 7, 8 +[10] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. In ICLR, 2023. 3 +[11] Ali Hassani, Steven Walton, Jiachen Li, Shen Li, and Humphrey Shi. Neighborhood attention transformer. In CVPR, 2023. 6 +[12] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 11 +[13] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 3 +[14] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In ICLR, 2022. 1, 3, 6, 7, 8, 11, 13 +[15] Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. In CVPR, 2018. 6 + +[16] Nikolai Kalischek, Michael Oechsle, Fabian Manhardt, Philipp Henzler, Konrad Schindler, and Federico Tombari. Cubediff: Repurposing diffusion-based image models for panorama generation. In ICLR, 2025. 1, 3, 13 +[17] Bingxin Ke, Anton Obukhov, Shengyu Huang, Nando Metzger, Rodrigo Caye Daudt, and Konrad Schindler. Repurposing diffusion-based image generators for monocular depth estimation. In CVPR, 2024. 1, 3 +[18] Jialu Li and Mohit Bansal. Panogen: Text-conditioned panoramic environment generation for vision-and-language navigation. In NeurIPS, 2023. 1, 3 +[19] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, 2023. 6 +[20] Renjie Li, Panwang Pan, Bangbang Yang, Dejia Xu, Shijie Zhou, Xuanyang Zhang, Zeming Li, Achuta Kadambi, Zhangyang Wang, Zhengzhong Tu, et al. 4k4dgen: Panoramic 4d generation at 4k resolution. arXiv preprint arXiv:2406.13527, 2024. 1 +[21] Aoming Liu, Zhong Li, Zhang Chen, Nannan Li, Yi Xu, and Bryan A Plummer. Panofree: Tuning-free holistic multiview image generation with cross-view self-guidance. In ECCV, 2024. 3 +[22] Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow. In ICLR, 2023. 3 +[23] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In *ICLR*, 2019. 6 +[24] Zhuqiang Lu, Kun Hu, Chaoyue Wang, Lei Bai, and Zhiyong Wang. Autoregressive omni-aware outpainting for open-vocabulary 360-degree image generation. In AAAI, 2024. 3 +[25] Chenlin Meng, Robin Rombach, Ruiqi Gao, Diederik Kingma, Stefano Ermon, Jonathan Ho, and Tim Salimans. On distillation of guided diffusion models. In CVPR, 2023. 3 +[26] Changgyoon Oh, Wonjune Cho, Yujeong Chae, Daehee Park, Lin Wang, and Kuk-Jin Yoon. Bips: Bi-modal indoor panorama synthesis via residual depth-aided adversarial learning. In ECCV, 2022. 3, 6 +[27] William Peebles and Saining Xie. Scalable diffusion models with transformers. In ICCV, 2023. 11 +[28] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. In ICLR, 2024. 3 +[29] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. In ICLR, 2023. 3 +[30] Jack Rae and Ali Razavi. Do transformers need deep long-range memory? In ACL, 2020. 6 +[31] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In ICML, 2021. 1, 3 +[32] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. + +[33] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 1, 3, 7, 11, 13 +[34] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI, 2015. 11 +[35] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In CVPR, 2023. 1, 3 +[36] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. In NeurIPS, 2022. 1 +[37] Noam Shazeer, Azalia Mirhoseini, Krzysztof Maziarz, Andy Davis, Quoc Le, Geoffrey Hinton, and Jeff Dean. Outrageously large neural networks: The sparsely-gated mixture-of-experts layer. In ICLR, 2017. 7, 8 +[38] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, 2020. 3, 11 +[39] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In ICLR, 2021. 3 +[40] Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever. Consistency models. In ICML, 2023. 3 +[41] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In CVPR, 2016. 6 +[42] Shitao Tang, Fuyang Zhang, Jiacheng Chen, Peng Wang, and Yasutaka Furukawa. Mvdiffusion: Enabling holistic multiview image generation with correspondence-aware diffusion. In NeurIPS, 2023. 1, 3, 6, 7, 11 +[43] Patrick von Platen, Suraj Patil, Anton Lozhkov, Pedro Cuenca, Nathan Lambert, Kashif Rasul, Mishig Davaadorj, Dhruv Nair, Sayak Paul, William Berman, Yiyi Xu, Steven Liu, and Thomas Wolf. Diffusers: State-of-the-art diffusion models. https://github.com/huggingface/diffusers, 2022.11 +[44] Guangcong Wang, Yinuo Yang, Chen Change Loy, and Zwei Liu. Stylelight: Hdr panorama generation for lighting estimation and editing. In ECCV, 2022. 3 +[45] Hai Wang, Xiaoyu Xiang, Yuchen Fan, and Jing-Hao Xue. Customizing 360-degree pancreas through text-to-image diffusion models. In WACV, 2024. 3 +[46] Tianhao Wu, Chuanxia Zheng, and Tat-Jen Cham. Panodifusion: 360-degree panorama outpainting via diffusion. In ICLR, 2024. 3 +[47] Xun Wu, Shaohan Huang, and Furu Wei. Mixture of lora experts. In ICLR, 2024. 6, 7, 8 +[48] Zhuofan Xia, Xuran Pan, Shiji Song, Li Erran Li, and Gao Huang. Vision transformer with deformable attention. In CVPR, 2022. 6 +[49] Bangbang Yang, Wenqi Dong, Lin Ma, Wenbo Hu, Xiao Liu, Zhaopeng Cui, and Yuewen Ma. Dreamspace: Dreaming + +your room space with text-driven panoramic texture propagation. In IEEE VR, 2024. 1 +[50] Weicai Ye, Chenhao Ji, Zheng Chen, Junyao Gao, Xiaoshui Huang, Song-Hai Zhang, Wanli Ouyang, Tong He, Cairong Zhao, and Guofeng Zhang. Diffpano: Scalable and consistent text to panorama generation with spherical epipolar-aware diffusion. In NeurIPS, 2024. 3 +[51] Jason J Yu, Fereshteh Forghani, Konstantinos G Derpanis, and Marcus A Brubaker. Long-term photometric consistent novel view synthesis with diffusion models. In ICCV, 2023. 3 +[52] Cheng Zhang, Qianyi Wu, Camilo Cruz Gambardella, Xiaoshui Huang, Dinh Phung, Wanli Ouyang, and Jianfei Cai. Taming stable diffusion for text to $360^{\circ}$ panorama image generation. In CVPR, 2024. 1, 2, 3, 4, 6, 7, 8, 11, 13, 14, 15 +[53] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, 2023. 1 +[54] Wenliang Zhao, Lujia Bai, Yongming Rao, Jie Zhou, and Jiwen Lu. Unipc: A unified predictor-corrector framework for fast sampling of diffusion models. In NeurIPS, 2023. 3 +[55] Kaiwen Zheng, Cheng Lu, Jianfei Chen, and Jun Zhu. Dpm solver-v3: Improved diffusion ode solver with empirical model statistics. In NeurIPS, 2023. 3 \ No newline at end of file diff --git a/ICCV/2025/What Makes for Text to 360-degree Panorama Generation with Stable Diffusion_/images.zip b/ICCV/2025/What Makes for Text to 360-degree Panorama Generation with Stable Diffusion_/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..dc0898627debd676c8eee07cc5cf85e9ae08fe58 --- /dev/null +++ b/ICCV/2025/What Makes for Text to 360-degree Panorama Generation with Stable Diffusion_/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9a61cfc8c3d134142bc3125a53982c12afcdfd0c1ca066b0fefa5ff1b37a896a +size 768423 diff --git a/ICCV/2025/What Makes for Text to 360-degree Panorama Generation with Stable Diffusion_/layout.json b/ICCV/2025/What Makes for Text to 360-degree Panorama Generation with Stable Diffusion_/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..43fa5f9adfeb0548a47903d2791933341ebd8ff6 --- /dev/null +++ b/ICCV/2025/What Makes for Text to 360-degree Panorama Generation with Stable Diffusion_/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:af8c82a077ff07d23cf29951210abd24eba623a993d7d4ce125718c94b900777 +size 413892 diff --git a/ICCV/2025/What You Have is What You Track_ Adaptive and Robust Multimodal Tracking/c6b9c23e-b4e6-4935-b399-70cebfc453b2_content_list.json b/ICCV/2025/What You Have is What You Track_ Adaptive and Robust Multimodal Tracking/c6b9c23e-b4e6-4935-b399-70cebfc453b2_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..cf7da3542192fa13b78ab505d314fc34876c20e1 --- /dev/null +++ b/ICCV/2025/What You Have is What You Track_ Adaptive and Robust Multimodal Tracking/c6b9c23e-b4e6-4935-b399-70cebfc453b2_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7c107385a78e9f1091e1f86583b42d22ef911b11ecbe8680a45567861d0580d5 +size 81403 diff --git a/ICCV/2025/What You Have is What You Track_ Adaptive and Robust Multimodal Tracking/c6b9c23e-b4e6-4935-b399-70cebfc453b2_model.json b/ICCV/2025/What You Have is What You Track_ Adaptive and Robust Multimodal Tracking/c6b9c23e-b4e6-4935-b399-70cebfc453b2_model.json new file mode 100644 index 0000000000000000000000000000000000000000..b6972d0089b16a824b79790b8536aa0e3f6d10d1 --- /dev/null +++ b/ICCV/2025/What You Have is What You Track_ Adaptive and Robust Multimodal Tracking/c6b9c23e-b4e6-4935-b399-70cebfc453b2_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dafa78f757947ba9d5a8d263a6cba86e47386d1083e3d0d8c3ac8058de291dd1 +size 103945 diff --git a/ICCV/2025/What You Have is What You Track_ Adaptive and Robust Multimodal Tracking/c6b9c23e-b4e6-4935-b399-70cebfc453b2_origin.pdf b/ICCV/2025/What You Have is What You Track_ Adaptive and Robust Multimodal Tracking/c6b9c23e-b4e6-4935-b399-70cebfc453b2_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..2bca0ad8089fcc1c19c8cc2235851ba0382da32b --- /dev/null +++ b/ICCV/2025/What You Have is What You Track_ Adaptive and Robust Multimodal Tracking/c6b9c23e-b4e6-4935-b399-70cebfc453b2_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5b8b9f0d4d87c758b68f6a36901d78df4164b717189e9b2c4d9d1e9b2bba3982 +size 12773718 diff --git a/ICCV/2025/What You Have is What You Track_ Adaptive and Robust Multimodal Tracking/full.md b/ICCV/2025/What You Have is What You Track_ Adaptive and Robust Multimodal Tracking/full.md new file mode 100644 index 0000000000000000000000000000000000000000..ed1c6cdf824d4f41f2a813893e770940f758d17d --- /dev/null +++ b/ICCV/2025/What You Have is What You Track_ Adaptive and Robust Multimodal Tracking/full.md @@ -0,0 +1,342 @@ +# What You Have is What You Track: Adaptive and Robust Multimodal Tracking + +Yuedong Tan $^{1,2,3}$ Jiawei Shao $^{1*}$ Eduard Zamfir $^{2}$ Ruanjun Li $^{4}$ Zhaochong An $^{5}$ + +Chao Ma $^{6}$ Danda Paudel $^{3}$ Luc Van Gool $^{3}$ Radu Timofte $^{2}$ Zongwei Wu $^{2*}$ + +1 TeleAI, China Telecom 2 Computer Vision Lab, CAIDAS & IFI, University of Wurzburg 3 INSAIT, Sofia University + +$^{4}$ ShanghaiTech University $^{5}$ University of Copenhagen $^{6}$ AI Institute, Shanghai Jiao Tong University + +# Abstract + +Multimodal data is known to be helpful for visual tracking by improving robustness to appearance variations. However, sensor synchronization challenges often compromise data availability, particularly in video settings where shortages can be temporal. Despite its importance, this area remains underexplored. In this paper, we present the first comprehensive study on tracker performance with temporally incomplete multimodal data. Unsurprisingly, under such a circumstance, existing trackers exhibit significant performance degradation, as their rigid architectures lack the adaptability needed to effectively handle missing modalities. To address these limitations, we propose a flexible framework for robust multimodal tracking. We venture that a tracker should dynamically activate computational units based on missing data rates. This is achieved through a novel Heterogeneous Mixture-of-Experts fusion mechanism with adaptive complexity, coupled with a video-level masking strategy that ensures both temporal consistency and spatial completeness — critical for effective video tracking. Surprisingly, our model not only adapts to varying missing rates but also adjusts to scene complexity. Extensive experiments show that our model achieves SOTA performance across 9 benchmarks, excelling in both conventional complete and missing modality settings. The code and benchmark will be made publicly available at https://github.com/supertyd/FlexTrack. + +# 1. Introduction + +Visual object tracking aims to localize a target object in a video while handling dynamic appearance changes [3, 5, 15, 33, 39, 53]. To improve robustness against occlusions, illumination shifts, and environmental variations, multimodal tracking has emerged as a powerful solution [29, 50, 52]. Complementary information from multiple sensors (e.g., depth, thermal, and event) has shown the ef + +![](images/dfdd7d87b4bcef2e43129f17abf80d76801d64ff7ce03ecd4044f2828f2d06da.jpg) + +![](images/3cafda64ca248823cfa72b52703233171a8fee85dca0453e3a1ba5a18a67c81a.jpg) +(a) Sensor Synchronization. +(b) Prior Work. + +![](images/7cc95abd1101b29350aaed76ba153399e8742d8991b81726de36e70ac56aeab3.jpg) +(c) Heterogenous Experts. +Figure 1. Synchronization has long been a challenge in multisensor fusion. As illustrated in (a), differences in frame timing and exposure duration make it difficult for a multisensor system to obtain perfectly synchronized pairs, often resulting in missing data. To address this issue, existing trackers [36], as shown in (b), handle each missing-data scenario individually, employing tailored and individual strategies for different cases. However, this approach does not account for the complexity variations introduced by different types of missing data. In contrast, in (c), we take a more holistic approach, unifying both missing and full-modality scenarios within a single framework. Our method leverages Heterogeneous Experts with adaptive complexity, providing a personalized yet structurally unified solution for robust tracking. + +fectiveness of improving unimodal approaches under ideal conditions. + +However, real-world multimodal data is rarely perfect [37, 38, 66]. Sensors often suffer from synchronization failures and temporary dropouts. As illustrated in Fig. 1, the inherent differences in exposure time and frame rates make it extremely challenging to achieve perfect synchronization across multiple sensors. Existing tracking datasets [29, 31, 60] simplify this issue by considering only paired modalities, such as RGB and depth. However, this setting remains limited, as it does not generalize to real-world scenarios where multiple modalities may be intermittently unavailable within a given temporal window during video tracking. Despite its practical importance, such a missing + +Table 1. Comparison of SOTA multimodal tracking models. + +
MethodUnified Arch.Unified Param.Missing ModalFlexi-Modal Adaptive
IPT [36]XX✓X
ViPT [67]✓XXX
UnTrack [56]✓✓XX
Ours✓✓✓✓
+ +scenario has not been well-adjusted. + +In this paper, we conduct a comprehensive study on temporally missing modalities in multimodal tracking [6, 21, 67]. For each benchmark dataset, we introduce a missing-data variation that simulates real-world synchronization challenges by incorporating diverse missing patterns, excluding cases where all modalities are absent simultaneously. Benchmarking existing methods under this setting reveals that they struggle to maintain reliable performance, primarily due to their reliance on template-matching frameworks between a reference template and a search region [5, 18, 32, 33, 62, 67], which are not designed to handle missing modalities. A recent attempt [36] to address this issue introduces placeholder and tailored prompts for both complete and missing cases. However, its rigid design raises a fundamental question: Should a tracker maintain the same computational complexity regardless of whether modalities are missing or available? + +Intuitively, we argue that an effective tracker should dynamically adapt, adjusting the computational resources it uses based on the available information and the severity of missing data. To achieve this, we propose FlexTrack, a unified framework designed to robustly handle multimodal tracking in both full and missing modality scenarios. Technically, we introduce a novel heterogeneous Mixture-of-Experts(HMoE) mechanism. This approach leverages varying levels of complexity and offers a unique advantage by routing and activating the most relevant experts based on the current data. Additionally, we propose a novel video-level masking strategy that enforces temporal consistency at the clip level while preserving contextual completeness, thus facilitating more robust object tracking. + +Our tracker stands out from SOTA counterparts, as demonstrated by the holistic comparison in Tab. 1, making it more adaptable to tracking with available data. Extensive evaluations on both conventional benchmarks with full modalities and our newly created missing-data variations validate its effectiveness. Notably, on full-modality benchmarks, we outperform the current SOTA by $2.6\%$ , with the performance gap widening significantly under the missing-data setting. + +# 2. Related Work + +# 2.1. Multimodal Tracking + +Modern object tracking has advanced significantly with the adoption of Siamese networks and transformer architectures [5, 16, 53, 62]. Despite these advancements, track- + +ing remains constrained by the limitations of single-modal (e.g., RGB) inputs, which struggle with challenges like occlusion or low-light conditions [51]. Multimodal tracking [22, 29, 69], which integrates RGB data with additional modalities (e.g., depth, thermal, event), has proven effective, especially when temporal cues are exploited [20, 21, 56, 67]. However, these methods assume consistent availability of modality, which is unrealistic in real-world scenarios [36]. Our work addresses this gap by introducing a video-level masking strategy that explicitly handles missing modalities while retaining video-level temporal coherence, and a HMoE mechanism to switch different sizes of experts to handle different levels of modality missing. + +# 2.2. Missing Modalities + +Missing modalities [37, 38, 66] is a wildly researched the problem. Real-world multimodal systems often face the challenge of having to handle cases where certain data modalities may be missing or incomplete. Existing methods for addressing missing modalities in object tracking include prompt-based approaches and generative models such as GANs [17] or diffusion models [8, 24]. While effective, these methods often introduce additional computational overhead. Alternatively, some approaches employ distillation techniques [1, 48], utilizing full-modal models to distill models with missing modalities. However, this can result in reduced model flexibility. Another common strategy involves actively masking modalities during training to enhance the model's robustness to each modality [54]. This technique typically involves randomly masking different modalities to improve the model's resilience. However, existing models primarily focus on single-instance prediction tasks such as classification and detection. In contrast, our model explores a novel video-level masking strategy for object tracking, effectively leveraging temporal information to mitigate the impact of missing modalities. + +# 2.3. Mixture-of-Experts + +Mixture-of-Experts(MoE) models utilize multiple specialized expert networks to adaptively select and process inputs, thereby enhancing performance across various tasks. Notably, large-scale MoE implementations such as DeepSeek-MoE [9] and Llama-MoE [68] have demonstrated exceptional efficiency in performance. Beyond their prominence in large language models, MoE architectures have also been effectively applied in multimodal fusion, leveraging flexible expert switching to accommodate diverse scenarios. For instance, FuseMoE [19] and Flex-MoE [63] employ MoE to integrate different modalities and address issues arising from missing modalities. However, they all have the uniform design for the expert size. + +In contrast, our approach introduces a novel module by combining temporal information with experts of varying + +![](images/aa72c9308c505362ec2fbb6776dcb9fe7cde798315bb4596092216e12908424e.jpg) +Figure 2. Our framework: Conventional trackers operate under the strong assumption that all frames are always available, disregarding practical scenarios with missing modalities. To address this limitation, we introduce heterogeneous experts, which dynamically switch between different experts to adapt test-time complexity. To further strengthen the robust learning, we employ a video-level masking strategy during training, encouraging the model to capture temporal continuity despite missing data. + +sizes to address the challenges of adaptive fusion. This methodology enables more precise and context-aware processing of multimodal data, thereby enhancing the model's robustness and flexibility in handling diverse scenarios. + +# 3. Method + +# 3.1. Overview + +In this paper, we propose a unified framework for object tracking that can handle flexible combinations of modalities. As illustrated in Fig. 2, our framework comprises two primary components: a multimodal fusion module and a video-level masking strategy which is only used in the training stage. Our tracker accepts video clips and search regions from RGB and auxiliary modalities (depth, thermal, and event modalities, collectively referred to as $\mathbf{X}$ ) as input. + +To address the challenge of varying modality completeness in tracking, we propose the Heterogeneous Mixture-of-Experts Fusion (HMoE-Fuse) module. Unlike conventional approaches that statically allocate computational resources, we dynamically adjust the model's test-time complexity based on the availability of multimodal inputs. + +Additionally, to enhance temporal continuity learning under partial modality-missing conditions, we propose a multimodal video-level masking mechanism during training. In each time step, we keep at least one frame available to focus on learning spatio-temporal consistency. + +# 3.2. Heterogeneous Mixture-of-Experts Fusion + +As illustrated in Fig. 2, we first feed both the RGB and X frames into the video encoder, which produces several token representations: RGB search region tokens, denoted as $T_{\mathrm{rgb}}^{s}$ ; video clip tokens ranging from $T_{\mathrm{rgb}}^{c1}$ to $T_{\mathrm{rgb}}^{ci}$ ; and, for additional X, we obtain $T_{x}^{s}, T_{x}^{c1}$ to $T_{x}^{ci}$ , where $i \in \{1, \dots, N\}$ . + +First, we concatenate all tokens to form the sequence token $T_{v}$ . The core component of the HMoE-Fuse is a heterogeneous MoE, which differs from typical MoEs where each expert is constructed with a distinct architecture. Specifically, each expert in our model consists of a linear layer, and its hidden state dimension is set as $2^{d}$ , where $d \in \{2, \ldots, D - 1\}$ . + +Contrary to prior MoEs [9, 19], we route the entire video clips to experts instead of individual tokens or images. First, a gating function $G$ produces a weight distribution, where higher values indicate better input-expert alignment. The output tokens $T_{y}^{1}$ of the HMoE layer is the weighted sum of outputs from the top- $K$ activated experts: + +$$ +\begin{array}{l} g _ {n} = \left\{ \begin{array}{l l} \operatorname {S o f t m a x} (G (T _ {v})) _ {n}, & \text {i f} G (T _ {v}) _ {n} \in \operatorname {T o p - K} (G (T _ {v})) \\ 0, & \text {o t h e r w i s e} \end{array} \right. \\ T _ {y} ^ {1} = \sum_ {n = 1} ^ {M} g _ {n} E _ {n} \left(T _ {v}\right). \tag {1} \\ \end{array} +$$ + +where $g_{n}$ represents the gating values derived from the function $G(T_{v})$ , and $E_{n}(T_{v})$ is the output of the $n$ -th expert. Besides, the gating function ensures that only the top $K$ ex + +![](images/4bdfbb168037a899786ed3600f42a0f41b17781eb8e9b1b651596ebcb9bd8cc8.jpg) +Figure 3. Comparison of different masking strategies for multimodal settings. (a) Conventional random or MAE-like masking [55] can be easily extended to multimodal settings but disrupts both spatial completeness and temporal continuity in video frames. (b) Extending VideoMAE [45] to multimodal settings enhances temporal consistency by preserving token continuity within each modality. However, it applies the same spatial randomness across all frames, resulting in persistent spatial information loss. (c) We argue that both spatial completeness and temporal consistency are crucial for object tracking: spatial completeness provides a detailed understanding of appearance, while temporal consistency captures dynamic changes. To achieve this balance, we propose a structured, video-level masking strategy that ensures the full spatial representation remains consistently retrievable in available data across different timestamps. + +perts are activated, here we set $K = 2$ in our method. With consistent video frame input, our HMoE selects the most appropriate experts to deal with missing modality cases. Simultaneously, we process the tokens $T_{v}$ through two linear layers, yielding an intermediate transformation: + +$$ +T _ {y} ^ {2} = T _ {v} W _ {1} \left(T _ {v} W _ {2}\right) ^ {T}, \tag {2} +$$ + +where $W_{1}$ and $W_{2}$ are learned weight matrices that perform linear attention on the tokens. + +Next, we perform a further transformation on the tokens by combining the expert outputs and the intermediate result: + +$$ +T _ {y} ^ {3} = T _ {y} ^ {1} \left(T _ {y} ^ {2}\right) ^ {T}. \tag {3} +$$ + +This operation integrates the expert outputs with the transformed tokens $T_{y}^{2}$ . Finally, we apply a softmax operation and a linear transformation to further refine the output: + +$$ +T _ {y} ^ {4} = \operatorname {S o f t m a x} \left(T _ {v} W _ {3}\right) T _ {y} ^ {3}, \tag {4} +$$ + +and + +$$ +T _ {y} ^ {5} = T _ {y} ^ {4} W _ {4}, \tag {5} +$$ + +where $W_{3}$ and $W_{4}$ are additional learnable weight matrices. The softmax operation is applied to $T_{v}W_{3}$ to scale the transformed tokens, and $W_{4}$ further adjusts the resulting values. + +# 3.3. Multimodal Video-Level Masking Strategy + +During inference, we find that our multimodal tracker fail to handle the missing modality for its substantial dependence on each frame in the video clip. (For a correspondence experiment, please refer to Tab. 6.) + +While VideoMAE [45]'s tube masking strategy demonstrates effectiveness in multiple video tasks through video cube masking, its design fundamentally assumes access to subsequent frames for masked content inference. However, the assumption is invalid for real-time tracking as future frames are inaccessible. Specifically, models cannot + +Algorithm 1: Video-level masking +Input: RGB and X search region tokens $T_{rgb}^{s}, T_{x}^{s}$ , RGB and X video clip tokens $T_{rgb}^{ci}, T_{x}^{ci}$ , $i \in \{1, \dots, N\}$ , batch size $B$ , video clip mask ratio $\alpha$ . Output: Updated tokens with applied video-level masking. +1 for $b = 1$ to $B$ do +2 $M_{s} \gets$ random([[1, 1], [1, 1], [1, 1], [1, 0], [0, 1])]; +3 $T_{x}^{s} \gets M_{s}[0] \times T_{x}^{s}$ ; +4 $T_{rgb}^{s} \gets M_{s}[1] \times T_{rgb}^{s}$ ; +5 $p \sim \mathcal{U}(0, 1)$ ; // Randomly select $X$ from the interval [0, 1]; +6 if $p < \alpha$ then +7 for $i = 1$ to $N$ do +8 $M_{vi} \gets$ random([[1, 1], [1, 0], [0, 1])]; +9 $T_{x}^{ci} \gets M_{vi}[0] \times T_{x}^{ci}$ ; +10 $T_{rgb}^{ci} \gets M_{vi}[1] \times T_{rgb}^{ci}$ ; +11 end +12 end +13 end + +utilize future temporal information to compensate for current modality absence, which poses a critical challenge in streaming tracking scenarios. We identify a unique advantage in multimodal tracking: temporal continuity persists even under spatial dropout in a modality. While individual frames may suffer from partial modality absence (e.g., missing depth data in frame $t$ ), adjacent frames ( $t - \delta$ to $t + \delta$ ) inherently preserve temporal coherence through at least one persistent modality information. As shown in Fig. 3, we exploit this through a multimodal video-level masking strategy to enforce temporal modeling. + +We illustrate the proposed video-level masking strategy in Algorithm 1. During training, we first apply search re + +gion masking by randomly selecting one of five predefined modality patterns to partially mask the RGB/X tokens in the search region. After applying the search region mask, a probability $p$ is sampled from a uniform distribution over the interval [0, 1]. If the value of $p$ falls below the predetermined threshold of $\alpha$ , this signifies the conditions necessary for implementing video-level masking. Thus, the algorithm proceeds to generate individual clip-level masks $M_{vi}, i \in 1, \ldots, N$ for each of the video clips. + +These masks are applied to the corresponding tokens of both modalities, further refining the focus of the model on the most pertinent information. + +# 3.4. Head and Loss Function + +To align with current multimodal trackers [20, 21, 56, 67], we utilize both a classification head and a regression head to localize the object center and approximate its height and width. Consequently, we adopt the same loss functions as these trackers, specifically the classification loss $\mathcal{L}_{cls}$ , the L1 loss $\mathcal{L}_{l1}$ , and the GIoU loss $\mathcal{L}_{GIoU}$ . Beyond the standard tracking task losses, we integrate the importance loss $\mathcal{L}_{important}$ and balance loss $\mathcal{L}_{balance}$ from the Switch Transformer [14] to ensure load balancing among experts and prevent imbalance within a single expert. The auxiliary loss $\mathcal{L}_{aux}$ is defined as follows: + +$$ +\mathcal {L} _ {a u x} = \mathcal {L} _ {\text {b a l a n c e}} + \mathcal {L} _ {\text {i m p o r t a n t}}. \tag {6} +$$ + +The overall loss function is formulated as: + +$$ +\mathcal {L} _ {a l l} = \lambda_ {1} \mathcal {L} _ {a u x} + \lambda_ {2} \mathcal {L} _ {c l s} + \lambda_ {3} \mathcal {L} _ {l 1} + \lambda_ {4} \mathcal {L} _ {G I o U}, \tag {7} +$$ + +where $\mathcal{L}_{all}$ represents the total loss used in the training process of our method. $\lambda_{1},\lambda_{2},\lambda_{3},\lambda_{4}$ is the hyperparameter with default values $\lambda_1 = 1$ $\lambda_{2} = 5$ $\lambda_3 = 2$ , and $\lambda_4 = 1$ + +# 4. Experiment + +# 4.1. Implementation Detail + +Consistent with recent trackers [20-22], our tracker is initialized with a pretrained encoder [25, 44] trained on large-scale datasets [13, 23, 34, 40]. The model is trained on 8 4090 GPUs with a batch size of 32 using the AdamW optimizer, with a base learning rate of 3e-4. Training consists of 30 epochs, each containing 600,000 image pairs. The learning rate is reduced by a factor of 10 after 22 epochs. For RGB-Thermal, we use LasHeR [31] for training, VisEvent [51] for RGB-Event, and DepthTrack [60] for RGB-Depth. To create a unified framework, we jointly train these three datasets in a single training process. During inference, we disable masking operations to fully leverage all available modalities while retaining the robustness learned through training phase masking simulations. + +Table 2. SOTA comparisons on RGB-Event tracking dataset. + +
MethodVisEvent [51]
PAUC
FlexTrack81.464.1
STTrack [22]78.661.9
SUTrack [6]79.962.7
OneTracker [20]76.760.8
SDSTrack [21]76.759.7
UnTrack [56]75.558.9
ViPT [67]75.859.2
ProTrack [61]63.247.1
MCITrack [25]69.753.3
OSTrack [62]69.553.4
SiamRCNN_E [46]65.949.9
TransT_E [4]65.047.4
LTMU_E [10]65.545.9
PrDiMP_E [12]64.445.3
VITAL_E [43]64.941.5
MDNet_E [41]66.142.6
ATOM_E [11]60.841.2
SiamCar_E [59]59.942.0
SiamBAN_E [7]59.140.5
SiamMask_E [49]56.236.9
+ +# 4.2. Evaluations on complete dataset: + +To demonstrate the performance of our tracker, we evaluated our tracker in 5 complete datasets. Under a strong assumption that all the frames are aligned well in temporal and there is not any missing modality caused by communication disruption and so on. + +**Evaluations on complete RGB-Event dataset:** The VisEvent[51] test set contains 320 videos covering a variety of challenging attributes. Our trackers achieved new SOTA results, exceeding current SOTA trackers by $1.5\%$ and $1.3\%$ in P and area-under-the-curve (AUC), respectively. + +Evaluations on complete RGB-Thermal dataset: For the RGB-Thermal tracking task, in Tab. 2, our FlexTrack models achieve new SOTA performance on both the LasHeR [31] and RGBT234 [29] datasets. Specifically, on the LasHeR benchmark, our FlexTrack achieves an AUC score of $62.0\%$ , surpassing the previous best-performing model, STTrack [22], by $1.7\%$ . On the RGBT234 benchmark, FlexTrack sets new SOTA score for MSR at $69.9\%$ . + +Evaluations on complete RGB-Depth dataset: In comprehensive RGB-Depth datasets, our model has demonstrated exceptional performance in DepthTrack [60]. Specifically, we have achieved an F-score that surpasses the previous best tracker, SUTrack [6], by $1.9\%$ . Furthermore, despite training exclusively on the VisEvent [51], LasHeR [31], and VisEvent [51] datasets, our model exhibits strong out-of-distribution performance. In Tab. 4, experimental re + +Table 3. SOTA comparisons on RGB-Thermal tracking datasets. + +
MethodLasHeR [31]RGBT234 [29]
PAUCMPRMSR
FlexTrack77.362.092.769.9
STTrack [22]76.060.389.866.7
SUTrack [6]74.559.592.269.5
OneTracker [20]67.253.885.764.2
SDSTrack [21]66.553.184.862.5
UnTrack [56]64.651.384.262.5
ViPT [67]65.152.583.561.7
ProTrack [61]53.842.079.559.9
MCITrack [25]64.250.279.160.4
OSTrack [62]51.541.272.954.9
APFNet [57]50.036.282.757.9
CMPP [47]--82.357.5
JMMC [65]--79.057.3
CAT [30]45.031.480.456.1
FANet [70]44.130.978.755.3
MaCNet [64]--79.055.4
+ +sults indicate that our model also achieves impressive results on the VOT-RGBD22 benchmark. + +# 4.3. Evaluations on missing modality dataset + +IPT [36] is the first method to address scenarios involving missing modalities. Initially, IPT was trained and tested exclusively on RGB-Thermal datasets. Subsequently, to better align with real-world applications, IPT developed a new dataset derived from the original one. This dataset includes instances of random missing, switched missing, and prolonged missing modalities. Building upon this dataset design, we expanded our evaluation to encompass multiple modalities, including DepthTrack [60], VisEvent [51], LasHeR [31], and RGBT234 [29]. + +RGB-Event missing modality datasets: As illustrated in Tab. 5, our tracker achieves a P of $72.8\%$ , outperforming SUTrack [6] by $6.2\%$ . Thanks to the video-level masking strategy, our model can capture more temporal information. This enables it to benefit from both the missing spatial information and the complete temporal information, even when the modalities are missing. + +RGB-Thermal missing modality datasets: To evaluate the scenarios involving missing modalities on the RGB-Thermal benchmark, we select two datasets, namely LasHeR [31] and RGBT234 [29]. As shown in Tab. 5, IPT [36] is a tracker specifically designed for the RGB-Thermal missing modality. It requires additional prompt parameters for tracking and needs to switch among three branches to handle different inference scenarios, including RGB-only, Thermal-only, and RGB-Thermal. In contrast, our tracker does not introduce new parameters for miss + +Table 4. SOTA comparisons on RGB-Depth tracking datasets. + +
MethodDepthTrack [60]VOT-RGBD22 [28]
F-scoreRePrEAOAcc.Rob.
FlexTrack67.066.967.178.083.893.1
STTrack [22]63.363.463.277.682.593.7
SUTrack [6]65.165.764.576.582.891.8
OneTracker [20]60.960.460.772.781.987.2
SDSTrack [21]61.460.961.972.881.288.3
UnTrack [56]61.060.861.172.182.086.9
ViPT [67]59.459.659.272.181.587.1
ProTrack [61]57.857.358.365.180.180.2
MCITrack [25]53.854.952.774.382.390.6
SPT [69]53.854.952.765.179.885.1
SBT-RGBD [58]---70.880.986.4
OSTrack [62]52.952.253.667.680.383.3
DeT [60]53.250.656.065.776.084.5
DMTrack [28]---65.875.885.1
DDiMP [27]48.556.950.3---
ATCAIS [27]47.645.550.055.976.173.9
LTMU-B [10]46.041.751.2---
GLGS-D [27]45.336.958.4---
DAL [42]42.936.951.2---
LTDSEd [26]40.538.243.0---
Siam-LTD [27]37.634.241.8---
SiamM-Ds [26]33.626.446.3---
CA3DMS [35]22.322.821.8---
DiMP [2]---54.370.373.1
ATOM [11]---50.559.868.8
+ +ing modality tracking. Instead, it only incorporates a novel video-level masking strategy. As a result, our tracker outperforms IPT [36] by $2.4\%$ on the LasHeR [31] dataset and by $2.1\%$ on the RGBT234 [29] dataset in terms of P. + +RGB-Depth missing modality datasets: We utilize DepthTrack [60] to evaluate the missing modalities in RGB-Depth tracking. The DepthTrack [60] dataset presents various challenges, such as long-term tracking and similar object interference. In long-term videos, video information plays a crucial role in capturing the variations of the object to be tracked. However, once a modality is missing, some trackers fail to track the object due to the lack of long-term contextual information. However, with our masking strategy, our tracker can still benefit from the incomplete modalities in the past frames. Additionally, our tracker employs HMoE-Fuse to handle different cases of missing modalities, enhancing the tracker's flexibility in dealing with various missing modality scenarios. As shown in Tab. 5, our tracker significantly outperforms other trackers. + +# 4.4. Ablation Study + +In subsequent experiments, we conducted evaluations on LasHeR [31] and VisEvent [51] datasets along with their missing-modality variants, using P as the primary metric to assess different methods. + +Video-level masking strategy: To enhance the robustness + +Table 5. SOTA comparison on DepthTrackmiss [60] for RGB-Depth, LasHeRmiss[31] and RGBT234miss[29] for RGB-Thermal, and VisEventmiss[51] for RGB-Event. + +
MethodDepthTrackmiss [60]LasHeRmiss [31]RGBT234miss [29]VisEventmiss [51]
F-score(↑)Re(↑)Pr(↑)P(↑)AUC(↑)MPR(↑)MSR(↑)P(↑)AUC(↑)
FlexTrack57.856.159.665.152.384.162.672.855.0
STTrack [22]49.948.851.054.544.973.854.265.549.7
SUTrack [6]49.547.351.958.347.682.060.866.650.5
SeqTrackv2 [5]45.040.950.050.039.970.849.957.643.1
SDSTrack [21]46.742.052.752.543.167.048.862.646.9
IPT [36]---61.746.482.059.4--
ViPT [67]44.440.546.640.134.052.439.457.243.2
MCITrack [25]49.742.959.134.240.053.640.949.936.5
+ +Table 6. Performance comparison of different masking strategies. + +
LasHeRVisEventLasHeRmissVisEventmiss
Ours77.381.465.172.8
w/o77.181.561.366.3
Random77.080.562.470.8
VideoMAE74.378.262.871.0
+ +of our tracker, we propose a novel video-level masking strategy. To validate its effectiveness, we conducted experiments with and without the masking strategy, as shown in Table 6. For common multimodal datasets (e.g., LasHeR [31] and VisEvent [51]), the performance remains consistent when the masking strategy is removed. However, in scenarios with missing modalities (e.g., $\mathrm{LasHeR}_{miss}$ and $\mathrm{VisEvent}_{miss}$ ), our masking strategy significantly improves performance compared to the baseline without masking. Specifically, we achieve a Precision (P) of 65.1 on $\mathrm{LasHeR}_{miss}$ and 72.8 on $\mathrm{VisEvent}_{miss}$ , outperforming the no-masking variant by $3.8\%$ and $6.5\%$ , respectively. + +Additionally, we compared our strategy with widely used methods, including random masking and VideoMAE [45]. While these methods slightly outperform the no-masking baseline in missing modality scenarios, they lead to performance drops in complete datasets (e.g., LasHeR and VisEvent). For instance, random masking reduces the Precision on LasHeR from 77.3 to 77.0, and VideoMAE further degrades it to 74.3. This suggests that these methods disrupt the temporal continuity required for robust video tracking. In contrast, our video-level masking strategy preserves temporal context, ensuring stable performance across both complete and missing modality scenarios. + +HMoE-Fuse The HMoE-Fuse is a crucial component for the efficient fusion of diverse modalities. In Table 7, we present an ablation study examining the impact of various design choices on the performance of HMoE-Fuse. Specifically, "w/o gating" indicates the removal of the gating function, "w/o linear attn." refers to the absence of linear attention, and "w/o HMoE" denotes the removal of the HMoE + +Table 7. Ablation studies on HMoE-Fuse. + +
LasHeRVisEventLasHeRmissVisEventmiss
Ours77.381.465.172.8
w/o gating76.379.963.171.7
w/o linear attn.75.478.261.869.9
w/o HMoE74.279.360.567.5
DeepseekMoE75.081.262.168.7
+ +Table 8. Ablation study on varying complexity. + +
LasHeRVisEventLasHeRmissVisEventmiss
Ours77.381.465.172.8
51276.180.962.771.7
6476.480.163.870.6
477.080.864.171.4
+ +module entirely. The experimental results demonstrate that eliminating these components leads to a significant decline in performance, highlighting their importance. To further validate the effectiveness of our HMoE, we replaced it with DeepSeekMoE [9]. The results show that, for video-level tasks, HMoE-Fuse, which incorporates temporal-spatial information from video clips, outperforms DeepSeekMoE. This is attributed to the heterogeneous experts in HMoE-Fuse, which enables the tracker to have richer representations and better handle the complexities of video data. The superior performance of HMoE-Fuse underscores its advantage in leveraging diverse expert sizes to enhance tracking accuracy and robustness. + +In this paper, we employ the HMoE to achieve a more complex and nuanced representation. To validate the effectiveness of our proposed HMoE approach, we conducted experiments using both medium-sized and large-sized experts. As shown in Table 8, our HMoE consistently outperforms other configurations across various metrics. Compared to the homogeneous MoE models, the heterogeneous design of our MoE enables the model to achieve stronger expressive capabilities. This advantage is particularly evident under different missing rates and in challenging scenar- + +![](images/d96e7fff5ad057e3bc8ea09696314c08342f2428b01df3293bd1b43492f789b5.jpg) +Figure 4. We are interested in the following question: With a higher missing rate, would a model favor a lighter expert or a more complex expert? To investigate this, we analyze the impact of the missing modality rate on the same video clip, as shown in (a) and (b). When the missing rate increases (e.g., $50\%$ vs. $20\%$ ), we observe that the model adaptively selects more complex experts with larger capacities to compensate for information scarcity through deeper reasoning. This insight leads us to a second question: Does this behavior persist in a full-modality setting? To answer this, we conduct an analysis in (c) and (d), visualizing expert selection in a similar indoor scene under varying levels of difficulty—such as object size (e.g., a moving hand vs. a stationary cat). Our findings reveal that the model prioritizes high-capacity experts to enable fine-grained cross-modal fusion, effectively addressing the challenges posed by dynamic variations in the scene. The block size in the middle represents the hidden dimension of heterogeneous experts in the MoE architecture, where a larger dimension indicates higher complexity. + +ios, where our model demonstrates superior performance. + +Visualization and Analysis: To gain a deeper understanding of the behavior of our HMoE-Fuse module under varying degrees of missing modalities and full modalities, we conducted a visualization analysis of the selection of experts with different missing rates and different full modalities scenes. The results are presented in Fig. 4. + +From the visualizations, we observed that as the missing modality ratio increases, the model tends to select experts with larger hidden dimensions. This is primarily because, when a high proportion of modalities are missing, the model opts for larger experts to ensure precision through more complex reasoning. In contrast, with only a few modalities missing, the model exhibits higher confidence in its predictions and thus favors smaller experts for inference. + +A similar pattern is evident in the full-modality scenario under complex conditions. For instance, in Fig. 4 (c), the depth and feature differences between the two hands are minimal, making it challenging for the model to derive accurate tracking results. As a result, the model selects a larger expert to tackle the increased difficulty. Conversely, in Fig. 4 (d), the distinct features of the "cat" compared to its surroundings enable the model to easily pinpoint its location, allowing for the use of smaller experts during tracking. + +Takeaways: Our HMoE framework offers a unique approach to interpretable multimodal fusion by explicitly linking the missing modality problem to complexity modeling — where more modalities reduce complexity, and vice versa. We observe that, in simpler scenarios when more information is available, the model favors lighter experts, as the increased joint information entropy from the input en + +hances scene understanding. Our framework demonstrates that multimodality not only improves performance but also contributes to a more adaptive system. This insight is introduced and discussed for the first time in the context of multimodal video tracking, aligning with the fundamental principles of multimodal fusion. + +# 5. Conclusion and Future work + +In this work, we tackle the critical yet often understudied challenge of robust multimodal tracking under missing modality conditions. To address this, we propose a novel MoE design with heterogeneous experts, paired with a video-level mask training strategy. Additionally, our model can adaptively select experts of different scales during testing, depending on the extent of missing modalities. Experimental results show that our model excels across all leading benchmarks, with both complete and missing modalities. + +While FlexTrack enhances multimodal tracking robustness, our future work will focus on integrating generative models for modality reconstruction and exploring lightweight deployment. Nonetheless, this work highlights the potential of heterogeneous expert systems in addressing the inherent unpredictability of real-world multimodal data. + +Acknowledgements. The authors sincerely thank the reviewers and all members of the program committee for their tremendous efforts and incisive feedback. This research was supported in part by the Alexander von Humboldt Foundation and in part by the Ministry of Education and Science of Bulgaria (support for INSAIT, under the Bulgarian National Roadmap for Research Infrastructure). C. Ma was supported in part by NSFC (62376156, 62322113). + +# References + +[1] Tadas Baltrusaitis, Chaitanya Ahuja, and Louis-Philippe Morency. Multimodal machine learning: A survey and taxonomy. TPAMI, 41(2):423-443, 2018. 2 +[2] Goutam Bhat, Martin Danelljan, Luc Van Gool, and Radu Timofte. Learning discriminative model prediction for tracking. In ICCV, 2019. 6 +[3] Yidong Cai, Jie Liu, Jie Tang, and Gangshan Wu. Robust object modeling for visual tracking. In ICCV, 2023. 1 +[4] Xin Chen, Bin Yan, Jiawen Zhu, Dong Wang, Xiaoyun Yang, and Huchuan Lu. Transformer tracking. In CVPR, 2021. 5 +[5] Xin Chen, Houwen Peng, Dong Wang, Huchuan Lu, and Han Hu. Seqtrack: Sequence to sequence learning for visual object tracking. In CVPR, 2023. 1, 2, 7 +[6] Xin Chen, Ben Kang, Wanting Geng, Jiawen Zhu, Yi Liu, Dong Wang, and Huchuan Lu. Sutrack: Towards simple and unified single object tracking. arXiv preprint arXiv:2412.19138, 2024. 2, 5, 6, 7 +[7] Zedu Chen, Bineng Zhong, Guorong Li, Shengping Zhang, Rongrong Ji, Zhenjun Tang, and Xianxian Li. Siamban: Target-aware tracking with siamese box adaptive network. TPAMI, 45(4):5158-5173, 2022. 5 +[8] Florinel-Alin Croitoru, Vlad Hondru, Radu Tudor Ionescu, and Mubarak Shah. Diffusion models in vision: A survey. TPAMI, 45(9):10850-10869, 2023. 2 +[9] Damai Dai, Chengqi Deng, Chenggang Zhao, RX Xu, Huazuo Gao, Deli Chen, Jiashi Li, Wangding Zeng, Xingkai Yu, Y Wu, et al. Deepseekmoe: Towards ultimate expert specialization in mixture-of-experts language models. arXiv preprint arXiv:2401.06066, 2024. 2, 3, 7 +[10] Kenan Dai, Yunhua Zhang, Dong Wang, Jianhua Li, Huchuan Lu, and Xiaoyun Yang. High-performance long-term tracking with meta-updater. In CVPR, 2020. 5, 6 +[11] Martin Danelljan, Goutam Bhat, Fahad Shahbaz Khan, and Michael Felsberg. Atom: Accurate tracking by overlap maximization. In CVPR, 2019. 5, 6 +[12] Martin Danelljan, Luc Van Gool, and Radu Timofte. Probabilistic regression for visual tracking. In CVPR, 2020. 5 +[13] Heng Fan, Liting Lin, Fan Yang, Peng Chu, Ge Deng, Sijia Yu, Hexin Bai, Yong Xu, Chunyuan Liao, and Haibin Ling. Lasot: A high-quality benchmark for large-scale single object tracking. In CVPR, 2019. 5 +[14] William Fedus, Barret Zoph, and Noam Shazeer. Switch transformers: Scaling to trillion parameter models with simple and efficient sparsity. JMLR, 23(120):1-39, 2022. 5 +[15] Zhihong Fu, Zehua Fu, Qingjie Liu, Wenrui Cai, and Yun-hong Wang. Sparsett: Visual tracking with sparse transformers. In IJCAI, 2022. 1 +[16] Shenyuan Gao, Chunluan Zhou, Chao Ma, Xinggang Wang, and Junsong Yuan. Aiatrack: Attention in attention for transformer visual tracking. In ECCV. Springer, 2022. 2 +[17] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139-144, 2020. 2 +[18] Dongyan Guo, Jun Wang, Ying Cui, Zhenhua Wang, and Shengyong Chen. Siamcar: Siamese fully convolutional + +classification and regression for visual tracking. In CVPR, 2020. 2 +[19] Xing Han, Huy Nguyen, Carl Harris, Nhat Ho, and Suchi Saria. Fusemoe: Mixture-of-experts transformers for flexi-modal fusion. NeurIPS, 37:67850-67900, 2025. 2, 3 +[20] Lingyi Hong, Shilin Yan, Renrui Zhang, Wanyun Li, Xinyu Zhou, Pinxue Guo, Kaixun Jiang, Yiting Chen, Jinglun Li, Zhaoyu Chen, et al. Onetracker: Unifying visual object tracking with foundation models and efficient tuning. In CVPR, 2024. 2, 5, 6 +[21] Xiaojun Hou, Jiazheng Xing, Yijie Qian, Yaowei Guo, Shuo Xin, Junhao Chen, Kai Tang, Mengmeng Wang, Zhengkai Jiang, Liang Liu, et al. Sdstrack: Self-distillation symmetric adapter learning for multi-modal visual object tracking. In CVPR, 2024. 2, 5, 6, 7 +[22] Xiantao Hu, Ying Tai, Xu Zhao, Chen Zhao, Zhenyu Zhang, Jun Li, Bineng Zhong, and Jian Yang. Exploiting multimodal spatial-temporal patterns for video object tracking. arXiv preprint arXiv:2412.15691, 2024. 2, 5, 6, 7 +[23] Lianghua Huang, Xin Zhao, and Kaiqi Huang. Got-10k: A large high-diversity benchmark for generic object tracking in the wild. TPAMI, 43(5):1562-1577, 2019. 5 +[24] Lan Jiang, Ye Mao, Xiangfeng Wang, Xi Chen, and Chao Li. Cola-diff: Conditional latent diffusion model for multimodal mri synthesis. In MICCAI, pages 398-408. Springer, 2023. 2 +[25] Ben Kang, Xin Chen, Simiao Lai, Yang Liu, Yi Liu, and Dong Wang. Exploring enhanced contextual information for video-level object tracking. arXiv preprint arXiv:2412.11023, 2024. 5, 6, 7 +[26] Matej Kristan, Jiri Matas, Ales Leonardis, Michael Felsberg, Roman Pflugfelder, Joni-Kristian Kamarainen, Luka Cehovin Zajc, Ondrej Drbohlav, Alan Lukezic, Amanda Berg, et al. The seventh visual object tracking VOT2019 challenge results. In ICCVW, 2019. 6 +[27] Matej Kristan, Aleš Leonardis, Jiří Matas, Michael Felsberg, Roman Pflugfelder, Joni-Kristian Kämäräinen, Martin Danelljan, Luka Čehovin Zajc, Alan Lukežić, Ondrej Drbohlav, et al. The eighth visual object tracking VOT2020 challenge results. In ECCVW, 2020. 6 +[28] Matej Kristan, Aleš Leonardis, Jiří Matas, Michael Felsberg, Roman Pflugfelder, Joni-Kristian Kämäräinen, Hyung Jin Chang, Martin Danelljan, Luka Čehovin Zajc, Alan Lukežić, et al. The tenth visual object tracking vot2022 challenge results. In ECCVW, 2023. 6 +[29] Chenglong Li, Xinyan Liang, Yijuan Lu, Nan Zhao, and Jin Tang. Rgb-t object tracking: Benchmark and baseline. PR, 96:106977, 2019. 1, 2, 5, 6, 7 +[30] Chenglong Li, Lei Liu, Andong Lu, Qing Ji, and Jin Tang. Challenge-aware RGBT tracking. In ECCV, 2020. 6 +[31] Chenglong Li, Wanlin Xue, Yaqing Jia, Zhichen Qu, Bin Luo, Jin Tang, and Dengdi Sun. Lasher: A large-scale high-diversity benchmark for rgbt tracking. TIP, 31:392-404, 2021. 1, 5, 6, 7 +[32] Wei Li, Zengfu Hou, Jun Zhou, and Ran Tao. Siambag: Band attention grouping-based siamese object tracking network for hyperspectral videos. TGRS, 61:1-12, 2023. 2 + +[33] Liting Lin, Heng Fan, Zhipeng Zhang, Yong Xu, and Haibin Ling. Swintrack: A simple and strong baseline for transformer tracking. NeurIPS, 35:16743-16754, 2022. 1, 2 +[34] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 5 +[35] Ye Liu, Xiao-Yuan Jing, Jianhui Nie, Hao Gao, Jun Liu, and Guo-Ping Jiang. Context-aware three-dimensional mean-shift with occlusion handling for robust object tracking in RGB-D videos. TMM, pages 664-677, 2018. 6 +[36] Andong Lu, Chenglong Li, Jiacong Zhao, Jin Tang, and Bin Luo. Modality-missing rgbt tracking: Invertible prompt learning and high-quality benchmarks. IJCV, pages 1-21, 2024. 1, 2, 6, 7 +[37] Mengmeng Ma, Jian Ren, Long Zhao, Sergey Tulyakov, Cathy Wu, and Xi Peng. Smil: Multimodal learning with severely missing modality. In AAAI, pages 2302-2310, 2021. 1, 2 +[38] Mengmeng Ma, Jian Ren, Long Zhao, Davide Testuggine, and Xi Peng. Are multimodal transformers robust to missing modality? In CVPR, pages 18177-18186, 2022. 1, 2 +[39] Christoph Mayer, Martin Danelljan, Danda Pani Paudel, and Luc Van Gool. Learning target candidate association to keep track of what not to track. In ICCV, 2021. 1 +[40] Matthias Muller, Adel Bibi, Silvio Giancola, Salman Alsubaihi, and Bernard Ghanem. Trackingnet: A large-scale dataset and benchmark for object tracking in the wild. In ECCV, 2018. 5 +[41] Hyeonseob Nam and Bohyung Han. Learning multi-domain convolutional neural networks for visual tracking. In CVPR, 2016. 5 +[42] Yanlin Qian, Song Yan, Alan Lukežić, Matej Kristan, Joni-Kristian Kämäräinen, and Jiri Matas. Dal: A deep depth-aware long-term tracker. In ICPR, 2021. 6 +[43] Yibing Song, Chao Ma, Xiaohe Wu, Lijun Gong, Linchao Bao, Wangmeng Zuo, Chunhua Shen, Rynson WH Lau, and Ming-Hsuan Yang. Vital: Visual tracking via adversarial learning. In CVPR, 2018. 5 +[44] Yunjie Tian, Lingxi Xie, Jihao Qiu, Jianbin Jiao, Yaowei Wang, Qi Tian, and Qixiang Ye. Fast-itpn: Integrally pretrained transformer pyramid network with token migration. TPAMI, 2024. 5 +[45] Zhan Tong, Yibing Song, Jue Wang, and Limin Wang. Videomae: Masked autoencoders are data-efficient learners for self-supervised video pre-training. NeurIPS, 35:10078-10093, 2022. 4, 7 +[46] Paul Voigtlaender, Jonathon Luiten, Philip H. S. Torr, and Bastian Leibe. Siam R-CNN: Visual tracking by redetection. In CVPR, 2020. 5 +[47] Chaoqun Wang, Chunyan Xu, Zhen Cui, Ling Zhou, Tong Zhang, Xiaoya Zhang, and Jian Yang. Cross-modal patternpropagation for rgb-t tracking. In CVPR, 2020. 6 +[48] Hu Wang, Congbo Ma, Jianpeng Zhang, Yuan Zhang, Jodie Avery, Louise Hull, and Gustavo Carneiro. Learnable cross-modal knowledge distillation for multi-modal learning with missing modality. In MICCAI, pages 216-226. Springer, 2023. 2 + +[49] Qiang Wang, Li Zhang, Luca Bertinetto, Weiming Hu, and Philip H. S. Torr. Fast online object tracking and segmentation: A unifying approach. In CVPR, 2019. 5 +[50] Xiao Wang, Jianing Li, Lin Zhu, Zhipeng Zhang, Zhe Chen, Xin Li, Yaowei Wang, Yonghong Tian, and Feng Wu. Visevent: Reliable object tracking via collaboration of frame and event flows. IEEE Transactions on Cybernetics, 2023. 1 +[51] Xiao Wang, Jianing Li, Lin Zhu, Zhipeng Zhang, Zhe Chen, Xin Li, Yaowei Wang, Yonghong Tian, and Feng Wu. Visevent: Reliable object tracking via collaboration of frame and event flows. TCYB, pages 1-14, 2023. 2, 5, 6, 7 +[52] Xiao Wang, Shiao Wang, Chuanming Tang, Lin Zhu, Bo Jiang, Yonghong Tian, and Jin Tang. Event stream-based visual object tracking: A high-resolution benchmark dataset and a novel baseline. In CVPR, 2024. 1 +[53] Xing Wei, Yifan Bai, Yongchao Zheng, Dahu Shi, and Yihong Gong. Autoregressive visual tracking. In CVPR, 2023. 1, 2 +[54] Sangmin Woo, Sumin Lee, Yeonju Park, Muhammad Adi Nugroho, and Changick Kim. Towards good practices for missing modality robust action recognition. In AAAI, pages 2776-2784, 2023. 2 +[55] Qiangqiang Wu, Tianyu Yang, Ziquan Liu, Baoyuan Wu, Ying Shan, and Antoni B Chan. Dropmae: Masked autoencoders with spatial-attention dropout for tracking tasks. In CVPR, 2023. 4 +[56] Zongwei Wu, Jilai Zheng, Xiangxuan Ren, Florin-Alexandru Vasluianu, Chao Ma, Danda Pani Paudel, Luc Van Gool, and Radu Timofte. Single-model and any-modality for video object tracking. In CVPR, 2024. 2, 5, 6 +[57] Yun Xiao, Mengmeng Yang, Chenglong Li, Lei Liu, and Jin Tang. Attribute-based progressive fusion network for rgbt tracking. In AAAI, 2022. 6 +[58] Fei Xie, Chunyu Wang, Guangting Wang, Yue Cao, Wankou Yang, and Wenjun Zeng. Correlation-aware deep tracking. In CVPR, 2022. 6 +[59] Bin Yan, Houwen Peng, Jianlong Fu, Dong Wang, and Huchuan Lu. Learning spatio-temporal transformer for visual tracking. In ICCV, 2021. 5 +[60] Song Yan, Jinyu Yang, Jani Käpylä, Feng Zheng, Ales Leonardis, and Joni-Kristian Kämäräinen. Depthtrack: Unveiling the power of rgbd tracking. In ICCV, 2021. 1, 5, 6, 7 +[61] Jinyu Yang, Zhe Li, Feng Zheng, Ales Leonardis, and Jingkuan Song. Prompting for multi-modal tracking. In ACMM, 2022. 5, 6 +[62] Botao Ye, Hong Chang, Bingpeng Ma, Shiguang Shan, and Xilin Chen. Joint feature learning and relation modeling for tracking: A one-stream framework. In ECCV, 2022. 2, 5, 6 +[63] Sukwon Yun, Inyoung Choi, Jie Peng, Yangfan Wu, Jingxuan Bao, Qiyiwen Zhang, Jiayi Xin, Qi Long, and Tianlong Chen. Flex-moe: Modeling arbitrary modality combination via the flexible mixture-of-experts. In NeurIPS. 2 +[64] Hui Zhang, Lei Zhang, Li Zhuo, and Jing Zhang. Object tracking in RGB-T videos using modal-aware attention network and competitive learning. Sensors, page 393, 2020. 6 + +[65] Pengyu Zhang, Jie Zhao, Chunjuan Bo, Dong Wang, Huchuan Lu, and Xiaoyun Yang. Jointly modeling motion and appearance cues for robust RGB-T tracking. TIP, pages 3335-3347, 2021. 6 +[66] Jinming Zhao, Ruichen Li, and Qin Jin. Missing modality imagination network for emotion recognition with uncertain missing modalities. In ACL, pages 2608-2618, 2021. 1, 2 +[67] Jiawen Zhu, Simiao Lai, Xin Chen, Dong Wang, and Huchuan Lu. Visual prompt multi-modal tracking. In CVPR, 2023. 2, 5, 6, 7 +[68] Tong Zhu, Xiaoye Qu, Daize Dong, Jiacheng Ruan, Jingqi Tong, Conghui He, and Yu Cheng. Llama-moe: Building mixture-of-experts from llama with continual pre-training. arXiv preprint arXiv:2406.16554, 2024. 2 +[69] Xue-Feng Zhu, Tianyang Xu, Zhangyong Tang, Zucheng Wu, Haodong Liu, Xiao Yang, Xiao-Jun Wu, and Josef Kittler. Rgbd1k: A large-scale dataset and benchmark for rgb-d object tracking. In AAAI, 2023. 2, 6 +[70] Yabin Zhu, Chenglong Li, Jin Tang, and Bin Luo. Quality-aware feature aggregation network for robust RGBT tracking. TIV, pages 121-130, 2020. 6 \ No newline at end of file diff --git a/ICCV/2025/What You Have is What You Track_ Adaptive and Robust Multimodal Tracking/images.zip b/ICCV/2025/What You Have is What You Track_ Adaptive and Robust Multimodal Tracking/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..6a4a2dc6186acc6bcc113a8370bb7779f7f4f1c6 --- /dev/null +++ b/ICCV/2025/What You Have is What You Track_ Adaptive and Robust Multimodal Tracking/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ed9cd002692291cec36b816850dc966001fa20b37719ac070bba9c75794bac43 +size 694972 diff --git a/ICCV/2025/What You Have is What You Track_ Adaptive and Robust Multimodal Tracking/layout.json b/ICCV/2025/What You Have is What You Track_ Adaptive and Robust Multimodal Tracking/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..b4aa7952d0f0dc50e7f272d7784c9d79d0594885 --- /dev/null +++ b/ICCV/2025/What You Have is What You Track_ Adaptive and Robust Multimodal Tracking/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f14f74338096125073d2205c812b54cff6ef8e198a0fd48dbd8ffb26c6e365f6 +size 414242 diff --git a/ICCV/2025/What to Distill_ Fast Knowledge Distillation with Adaptive Sampling/5cc2db1b-d50c-4519-b11e-7cd7283b5e50_content_list.json b/ICCV/2025/What to Distill_ Fast Knowledge Distillation with Adaptive Sampling/5cc2db1b-d50c-4519-b11e-7cd7283b5e50_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..f030ffbef2034cbc4a72992bd98f0370c8c5e393 --- /dev/null +++ b/ICCV/2025/What to Distill_ Fast Knowledge Distillation with Adaptive Sampling/5cc2db1b-d50c-4519-b11e-7cd7283b5e50_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d904ec62f52e526619579ffa2d994da88b358d9d335dbb572841f8f73de66939 +size 75310 diff --git a/ICCV/2025/What to Distill_ Fast Knowledge Distillation with Adaptive Sampling/5cc2db1b-d50c-4519-b11e-7cd7283b5e50_model.json b/ICCV/2025/What to Distill_ Fast Knowledge Distillation with Adaptive Sampling/5cc2db1b-d50c-4519-b11e-7cd7283b5e50_model.json new file mode 100644 index 0000000000000000000000000000000000000000..dc9155b94a00311c17e0b169cef7275c706175a7 --- /dev/null +++ b/ICCV/2025/What to Distill_ Fast Knowledge Distillation with Adaptive Sampling/5cc2db1b-d50c-4519-b11e-7cd7283b5e50_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fe4799053d3c821b699d2b62fedbba4ad2ebbaab84be5c06bc66aa27869ff05b +size 91304 diff --git a/ICCV/2025/What to Distill_ Fast Knowledge Distillation with Adaptive Sampling/5cc2db1b-d50c-4519-b11e-7cd7283b5e50_origin.pdf b/ICCV/2025/What to Distill_ Fast Knowledge Distillation with Adaptive Sampling/5cc2db1b-d50c-4519-b11e-7cd7283b5e50_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..a371a70b4875a57cea152ca99284f108b1c7c3a8 --- /dev/null +++ b/ICCV/2025/What to Distill_ Fast Knowledge Distillation with Adaptive Sampling/5cc2db1b-d50c-4519-b11e-7cd7283b5e50_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c21fbe62fafaeccb40066ec2e1f81b395f0914baed1d56ff5842c185b5517ed6 +size 882990 diff --git a/ICCV/2025/What to Distill_ Fast Knowledge Distillation with Adaptive Sampling/full.md b/ICCV/2025/What to Distill_ Fast Knowledge Distillation with Adaptive Sampling/full.md new file mode 100644 index 0000000000000000000000000000000000000000..94af94a7acaca55ccb85ae831638262d1c73c20e --- /dev/null +++ b/ICCV/2025/What to Distill_ Fast Knowledge Distillation with Adaptive Sampling/full.md @@ -0,0 +1,338 @@ +# What to Distill? Fast Knowledge Distillation with Adaptive Sampling + +Byungchul Chae Kyung Hee University; SqueezeBits Inc. +chaechae7@khu.ac.kr + +Seonyeong Heo Kyung Hee University seonyeong.heo@khu.ac.kr + +# Abstract + +Knowledge Distillation (KD) has been established as an effective technique for reducing the resource requirements of models when tackling computer vision tasks. Prior work has studied how to distill the knowledge of a teacher model better, but it overlooks how data affects the distillation result. This work examines the impact of data in knowledge distillation from two perspectives: (i) quantity of knowledge and (ii) quality of knowledge. Our examination finds that faster knowledge distillation can be achieved by using data with a large amount of high-quality knowledge in distillation. Based on the findings, this work proposes an efficient adaptive sampling method called KDAS for faster knowledge distillation, which enhances the distillation efficiency by selecting and applying 'good' samples for the distillation. This work shows that our adaptive sampling methods can effectively accelerate the training efficiency of a student model when combined with existing KD methods. + +# 1. Introduction + +As deep learning has become more prevalent, reducing the resource requirement of deep learning models is considered essential for resource-constraint computing systems [19]. Among various model compression techniques, knowledge distillation (KD) has emerged as a pivotal technique for lowering the complexity of deep learning models. KD is to transfer the knowledge of a sophisticated teacher model to a simpler student model, enabling the student model to obtain comparable performance with fewer resources [5]. The widespread adoption of KD demonstrates its significance in modern machine learning applications. + +Previous work has proposed a variety of methods to maximize the knowledge transfer from teacher models to student models [28]. Although these methods may help student models achieve a performance level similar to that of teacher models, they do not sufficiently consider the role of data in the distillation process. As far as we discover in this work, the selection of data can largely influence the effectiveness and efficiency of KD, yet this dimension remains underexplored in previous work. + +![](images/b91b491f33d57df3a8aea30d000f7f57806d5800c0315bb19255ee6bf306c738.jpg) +Figure 1. Average training time (relative to vanilla KD [9]) vs. top-1 accuracy on CIFAR-100. We set ResNet110 as the teacher and ResNet32 as the student. + +This work first conducts an empirical analysis of the impact of training data on KD, specifically focusing on two key perspectives: quantity of knowledge and quality of knowledge. The underlying hypothesis of our investigation is that not all data contribute equally to the distillation process; some data may contain richer information that can enhance the learning of the student model. We examine the impact of data with the new metrics for the quantity and quality of knowledge, aiming to provide insights into how data should be applied for distillation efficiency. + +Based on our analysis, we propose a new adaptive sampling method called KDAS for accelerating the distillation process. The method dynamically selects and applies training samples, considering both the quantity and quality of knowledge with two techniques: quantity-based subsampling and quality-calibrated loss weighting. The proposed method can efficiently distill the knowledge of the teacher model by excluding samples that have little or negative impact on the distillation process. + +This work evaluates the adaptive sampling method in combination with standard and state-of-the-art KD methods. The experimental results show that the proposed method can accelerate the training of student models by at most $36\%$ without compromising on performance. Furthermore, in many cases, the method enables an even higher performance of student models. The results demonstrate that the method can enable a more efficient and effective development of resource-efficient models with KD. + +![](images/6ea488ae98569e317d412f56f5ae8abf73f9aa6d382062ca2cc1a0ae09d5b85b.jpg) +Figure 2. Overview of the proposed approach. We first analyze the impact of the quantity and quality of knowledge on knowledge distillation. Based on the analysis, we design an adaptive sampling method called KDAS that enables efficient knowledge distillation with two techniques: quantity-based sampling and quality-based loss weighting. + +![](images/84f0fc4f21bdaba65a874a956c213ce8b57fcc8af3649d902c8bc67398860537.jpg) + +The contributions of this work are: + +- Empirical analysis of the impact of data on knowledge distillation, especially from two perspectives, quantity and quality of knowledge +- Design of a new adaptive sampling method, which enables effective and efficient knowledge distillation based on the analysis +- Evaluation of the proposed adaptive sampling method in combination with state-of-the-art knowledge distillation methods + +# 2. Related Work + +General Knowledge Distillation. Knowledge distillation (KD) is a novel method to transfer the knowledge of a large teacher model to a small student model [9]. The basic hypothesis is that the 'dark knowledge' from the teacher model contains richer information such as inter-class correlation than the ground-truth labels. Then, the student model can achieve an improved generalization performance from the knowledge of the teacher model. + +Previous research on knowledge distillation mainly investigates how to maximize the knowledge transfer, which has evolved in three main branches: logit-based methods [9, 12, 17, 24, 33, 34], which utilize the output logits of the teacher model; feature-based methods [2, 3, 6, 8, 20, 25], which consider intermediate features for more profound learning; and relation-based methods [11, 13, 18, 26, 29], which focus on the correlations between samples. + +Efficient Knowledge Distillation. Although previous work has successfully demonstrated the performance improvement of the student model, their regularization and optimization techniques often result in slowing down the distillation process. That is, such techniques require a larger amount of resources to obtain a 'good' student model. With the increasing demand for smaller but effective models, it is also important to accelerate the distillation process. + +Recently, research on fast knowledge distillation has been conducted actively [21, 27, 30]. Previous work [21, 30] proposes to accelerate the distillation process by reusing the logits of the teacher model. ReLabel [30] stores the global label map of the teacher model in the storage and then reuses the label map for distillation considering the target region of interest. FKD [21] enables faster knowledge distillation by considering multiple regions of interest. + +# 3. Background + +Generally in KD, the student model learns from both hard and soft targets for given samples. The hard target indicates the ground-truth predictions for the samples, and the soft target indicates the output predictions of the teacher model. As the soft target provides more information on the reasoning behind the predictions of the teacher model (so-called 'dark knowledge'), the student model can learn better by imitating the predictions of the teacher model. + +The typical loss function for KD combines the hard target loss function $\mathcal{L}_H$ for the ground-truth label and the soft target loss function for the teacher model $\mathcal{L}_S$ . For the hard target, a general loss function is chosen for a specific task. + +$$ +\mathcal {L} _ {K D} = \mathcal {L} _ {H} \left(P _ {S}, P _ {G T}\right) + \mathcal {L} _ {S} \left(P _ {T}, P _ {S}\right) \tag {1} +$$ + +Kullback-Leibler (KL) divergence is a widely used metric for the soft target loss. KL divergence defines a statistical distance between two probability distributions. In other words, it measures how a probability distribution is different from the other probability distribution. Then, the KL divergence between the teacher and student models is used to measure how the predictions of the student model differs from the predictions of the teacher model. + +$$ +\mathcal {L} _ {K L} \left(P _ {T}, P _ {S}\right) = \sum_ {x \in \mathcal {X}} P _ {S} (x) \log \left(\frac {P _ {S} (x)}{P _ {T} (x)}\right) \tag {2} +$$ + +# 4. Distillation Efficiency Analysis + +To design the adaptive sampling method for faster knowledge distillation, we first conduct an empirical analysis of how the intrinsic characteristics of data influence the efficiency of knowledge distillation, defining two metrics: quantity and quality of knowledge. + +We use KL divergence to measure the quantity and quality of knowledge for a certain sample. Our assumptions are: + +- Quantity of Knowledge: If the KL divergence between the teacher and student models is large, it means that the teacher model has more knowledge to transfer into the student model. + +$$ +K L _ {T S} = D _ {K L} \left(P _ {S} \| P _ {T}\right) +$$ + +- Quality of Knowledge: If the KL divergence between the teacher model and the ground truth is too small or too large, it means that the teacher model has low-quality knowledge to the student model. + +$$ +K L _ {T G} = D _ {K L} \left(P _ {T} | | P _ {G}\right) +$$ + +In this section, we present our findings regarding the quantity and quality of knowledge with experimental results, obtained using the vanilla KL-based distillation method [9] with the CIFAR-100 dataset. + +# 4.1. Quantity of Knowledge + +One fundamental challenge in knowledge distillation is to identify which samples would give plenty of knowledge to transfer. We use the KL divergence between the teacher and student models as a quantitative proxy to measure the information density for each training sample. + +Finding 1 (Quantity Impact). KD can achieve comparable or even superior performance using only samples with high teacher-student KL divergence, compared to the performance achieved with all samples. + +Our first finding is that the student model can be efficiently trained with samples with a larger quantity of knowledge. That is, if the predictions of the teacher and student models largely differ for samples, the samples are more effective for distillation than the other samples. Therefore, it is not necessary to use all training samples in distillation, as samples with a smaller quantity of knowledge have little impact on the distillation result. + +The finding is rather intuitive, as a high KL divergence between teacher and student models $(\mathcal{L}_S = KL_{TS})$ leads to a high KD loss $(\mathcal{L}_{KD} = \mathcal{L}_H + \mathcal{L}_S)$ . This concept resembles hard example mining (HEM) [22]; however, our approach differs as we consider only soft target loss to focus more on distilling the knowledge of the teacher model rather than on the difficulty of the samples. + +![](images/8237bae70b78fc2c272bc25d3f065bc8e8b3d1b7e6da8260adb9d381bbee59b2.jpg) +Figure 3. Performance of the student model when training on samples chosen according to the teacher-student KL divergence (Teacher: VGG13, Student: VGG8). + +
RatioHigh KL
20%72.42 %
40%73.76 %
60%73.69 %
80%73.20 %
100%73.14 %
+ +Table 1. Accuracy of student model with different sampling metrics in distillation (Sampling ratio: $50\%$ ). + +
Teacher → StudentLKD = LH + LSLS
R32x4 → WRN1675.37 %75.55 %
VGG13 → VGG873.28 %73.67 %
R56 → R2070.34 %70.54 %
+ +Figure 3 illustrates the experimental results that support our first finding. In the experiments, we measure how the performance of the student model changes when using samples with high, mid, and low KL divergence between the teacher and student predictions. For every original batch, we take a $p$ percent of the samples according $KL_{TS}$ and use only the taken samples for distillation. + +The experimental results demonstrate that samples with high $KL_{TS}$ yields a superior performance than samples with mid or low $KL_{TS}$ . Interestingly, the high $KL_{TS}$ group often obtain a higher performance than the full group. The results show that we can improve the distillation efficiency by using the samples with the high quantity of knowledge. Moreover, they reveal that only $40\%$ of the entire set can achieve a better performance than the full set. + +Additionally, we compare the performance with different sampling metrics, (i) the entire KD loss and (ii) the teacher-student KL divergence, to analyze the distinction against HEM. HEM basically selects the hard samples based on the training loss $(= \mathcal{L}_{KD})$ . However, Table 1 show the samples chosen according to the KD loss obtain a lower performance than the samples according to the KL divergence. Then, it is more effective for KD to use the quantity of knowledge metric than the typical training loss metric. + +The first finding can also be understood in the view of information theory. KL divergence is also known as the relative entropy of $P$ relative to $Q$ : + +$$ +D _ {K L} (P | | Q) = H (P, Q) - H (Q) \tag {3} +$$ + +Then, a teacher-student KL reflects an information gap between teacher and student. Therefore, it is reasonable to use KL divergence to measure the quantity of knowledge. + +Figure 4. Performance of the student model with fixed and dynamic sampling ratios (Fixed sampling ratio: $50\%$ ). +![](images/ddf29df802f5a03f940fd0c3c6c42a7d7c366d4f119de2e96bf9f92fc6fb4af8.jpg) +Index: $\bullet$ ResNet32x4→ShuffleV2, $\diamond$ WRN-40-2→ResNet8x4, + +Figure 5. Distribution of KL divergence between teacher and student models across different training epochs (early, middle, late). The distribution gradually shifts toward lower values as training progresses, but maintains a long tail of informative samples that our method selectively utilizes. +Finding 2 (Curriculum Sampling). KD can achieve a superior performance by progressively changing the sampling ratio than using a fixed sampling ratio. +![](images/ce34b31b5343c84d7d54fcf0d3452b90f8bba59d2378dae5e032aae85d441666.jpg) +VGG13 $\rightarrow$ VGG8, $\triangle: \mathrm{ResNet}56 \rightarrow \mathrm{ResNet}20$ . + +Our second finding is that the student model can be more efficiently trained when decreasing the sampling ratio dynamically. That is, it is good to select more samples in the early stage of distillation because there are typically more samples with a large quantity of knowledge (i.e., high $KL_{TS}$ ). Note that our approach is different from the typical curriculum sampling [1], which to use easy samples in the early stage and use hard samples in the late stage. + +Figure 4 shows the performance of the student model with fixed and dynamic sampling ratios for various combinations of teacher and student models. For curriculum sampling, we gradually decrease the sampling ratio every training step. The results show that the student model achieves a better performance with dynamic sampling ratio, even with the same total number of samples used in distillation. + +We further investigate the best curriculum for the dynamic sampling ratio. Figure 4 illustrates how the performance of the student model changes with different initial and final sampling ratio. As far as in our experiments, the student model obtains the best performance when decreasing the sampling ratio from $60\%$ to $40\%$ for every combination of the teacher and student models. + +![](images/0366961a4df6706cc57160c754c040bb73a4365dcb253b5f23d9b1bd2d709efb.jpg) +Figure 6. Performance of the student model when training on samples chosen according to the teacher-ground truth KL divergence (Teacher: VGG13, Student: VGG8). + +
RatioMid KL
60%72.86 %
70%73.11 %
80%73.34 %
90%73.29 %
100%73.14 %
+ +# 4.2. Quality of Knowledge + +While the previous findings focus on the quantity of transferable knowledge, we also consider the quality of the knowledge, based on our hypothesis that high-quality knowledge is more crucial for efficient knowledge distillation. We use the KL divergence between the teacher and ground-truth distributions to quantify the quality of knowledge, similarly as the quantity of knowledge. + +Finding 3 (Quality Impact). KD can achieve comparable or even superior performance using only samples with mid teacher-ground truth KL divergence, compared to the performance achieved with all samples. + +The third finding is that the student model can be efficiently trained with samples with a higher quality of knowledge. We assume the quality of knowledge becomes low if the teacher-ground truth is too large or too small. If the teacher prediction is too close to the ground truth (i.e., too low $KL_{TG}$ ), the teacher may not give dark enough knowledge. If the teacher prediction is too far from the ground truth (i.e., too high $KL_{TG}$ ), the teacher may give mistaken knowledge to the student model. + +Figure 6 shows the performance of the student model when using samples with high, mid, and low KL divergence between the teacher prediction and the ground truth. The results reveal that the samples with mid $KL_{TG}$ values consistently obtain the best performance, sometime even higher than the baseline performance. That is, it is most efficient to use the mid $KL_{TG}$ group for distillation. + +This finding illuminates why knowledge distillation sometimes fails despite using high-performing teacher models - the issue may not be the absolute accuracy of the teacher, but rather how its probability distributions are calibrated relative to ground truth. The optimal knowledge transfer occurs in a specific regime where teacher distributions differ enough from hard labels to provide rich additional information, yet remain sufficiently aligned with ground truth to be reliable. + +![](images/ab2fef6928b593c2533d30af1295fba78bec51bd037517a9fbb63ed291044edd.jpg) +Figure 7. Performance comparison between vanilla KD and KD with our penalization approach across different pairs of teacher and student models (with the full dataset). The consistent performance improvements demonstrate the effectiveness of penalizing samples with suboptimal knowledge quality. + +# Finding 4 (Penalization). KD can achieve superior accuracy by penalizing samples in the loss function regarding their quality of knowledge. + +Our fourth finding is that the student model can be more efficiently trained by giving more weights to the samples with the high quality of knowledge. Through the experimental results, we found that unlike the quantity of knowledge, it is less trivial to find the best set of samples with the high quality of knowledge. Therefore, we take the quality of knowledge into consideration by penalizing samples according to their quality of knowledge. + +For penalization, we use loss weighting based on the teacher-ground truth KL divergence. If the teacher-ground truth KL divergence is lower than the lower bound $(\theta_{low})$ or higher than the upper bound $(\theta_{high})$ for a certain sample, we scale the KL loss for the sample by multiplying a factor obtained using the KL divergence. The formal equation for the loss weighting is presented in Section 5. + +Figure 7 illustrates that our penalization approach consistently outperforms the vanilla KD across different pairs of teacher-student models. The results demonstrate that applying appropriate penalties to samples based on their $KL_{TG}$ values enhances the overall distillation process. That is, with the same amount of training samples, the penalization enables more efficient knowledge distillation by effectively filtering out potentially misleading or redundant knowledge from the teacher model. + +In conclusion, our four findings address that an ideal distillation should not only select samples with the large quantity of knowledge but also calibrate the influence of samples based on the quality of knowledge. We exploit these findings on the quantity and quality of knowledge to design an adaptive sampling method to enhance the knowledge distillation process. + +# 5. Faster Knowledge Distillation + +Building on our findings in Section 4, we introduce KDAS, an efficient adaptive sampling method that can enhance the distillation efficiency while maintaining or even improving performance. This method regards the quantity and quality of knowledge with two techniques: quantity-based subsampling and quality-calibrated loss weighting. + +# 5.1. Quantity-based Subsampling + +To maximize the efficiency of knowledge transfer based on Findings 1 and 2, KDAS selects samples with the high quantity of knowledge based on teacher-student KL divergence $(KL_{TS})$ . For every $\tau$ epochs, KDAS selects the top- $K\%$ samples with the highest $KL_{TS}$ values from the original dataset: + +$$ +\mathcal {D} _ {s u b} ^ {(t)} = \{x \in \mathcal {D} _ {o r g} | K L _ {T S} (x) > P _ {K ^ {(t)}} (K L _ {T S} (\mathcal {D} _ {o r g}) \}, +$$ + +where $\mathcal{D}_{org}$ and $\mathcal{D}_{sub}^{(t)}$ denote the original and subsampled datasets for epoch $t$ and $P_{K(t)}$ indicates the $(100 - P_{K^{(t)}})$ -th percentile. The sampling ratio $K^{(t)}$ gradually decreases following $K^{(t)} = \max (K_{\min}, K_{\mathrm{init}} - \alpha \cdot t)$ over time. + +This quantity-based subsampling requires the calculation of $KL_{TS}$ with the forward process of the teacher model for every training sample. The overhead of computing $KL_{TS}$ at every epoch may not be negligible, especially when the teacher model is much larger than the student model. Therefore, to effectively reduce the distillation time, KDAS sporadically performs the subsampling every $\tau$ epoch, which can also be controlled according to a predefined schedule. + +In the evaluation, our experimental results show that KD can still achieve a good performance with the sporadic sample selection. + +# 5.2. Quality-calibrated Loss Weighting. + +To consider the quality of knowledge based on Findings 3 and 4, KDAS weights the contribution of each sample based on teacher-ground truth KL divergence $(KL_{TG})$ . KDAS defines adaptive percentile-based thresholds $\theta_{\mathrm{low}}$ and $\theta_{\mathrm{high}}$ and gives penalty to the samples whose $KL_{TG}$ values are lower than or higher than the thresholds: + +$$ +w (x) = \left\{ \begin{array}{l l} \max (\gamma , 1 - \lambda \cdot (\theta_ {\text {l o w}} - K L _ {T G} (x))), & \\ & \text {i f} K L _ {T G} (x) < \theta_ {\text {l o w}} \\ \max (\gamma , 1 - \lambda \cdot (K L _ {T G} (x) - \theta_ {\text {h i g h}})), & \\ & \text {i f} K L _ {T G} (x) > \theta_ {\text {h i g h}} \\ 1, & \text {o t h e r w i s e} \end{array} \right. +$$ + +where $\lambda$ is a scaling factor for controlling the intensity of penalty, and $\gamma$ is the minimum weight for ensuring all samples maintain a minimum contribution. + +Algorithm 1 KDAS: Adaptive Sampling for KD +Require: Teacher model $T$ , Student model $S$ , Dataset $\mathcal{D}$ , Number of epochs $t_{max}$ , Interval $\tau$ , Sampling ratios $K_{init}, K_{min}$ , Penalty factor $\gamma$ , Penalty intensity $\lambda$ +1: Initialize $S$ and set $\mathcal{D}_{sub} \gets \mathcal{D}_{org}$ +2: Compute $\theta_{low}$ and $\theta_{high}$ from $KL_{TG}$ distribution +3: for $t = 1$ to $t_{max}$ do +4: if $t \mod \tau = 0$ then ▷ Subsampling +5: Compute $KL_{TS}$ for all samples +6: Update $K^{(t)}$ based on current epoch +7: Select top- $K^{(t)}$ % samples based on $KL_{TS}$ +8: end if +9: for each batch $\mathcal{B} \in \mathcal{D}_{sub}$ do +10: Perform forward process for $T$ and $S$ with $\mathcal{B}$ +11: Compute weight $w(x)$ for each $x \in \mathcal{B}$ +12: Compute loss $\mathcal{L}_{\mathrm{KDAS}}$ and update $S$ +13: end for +14: end for +15: return $S$ + +This loss weighting approach penalizes samples with suboptimal knowledge quality while preserving a minimum contribution from all examples. To allow for gradual adaptation, KDAS adopts a dynamic scaling mechanism where the intensity of penalty increases progressively during training after an initial warmup period. + +Combining the above two methods, the final loss function is applied to the student model: + +$$ +\mathcal {L} _ {\mathcal {K D}} (x) = (1 - \alpha) \mathcal {L} _ {H} (y _ {s}, y _ {g}) + \alpha w (x) \mathcal {L} _ {S} (y _ {s}, y _ {t}). +$$ + +$$ +\mathcal {L} _ {K D A S} = \frac {1}{| \mathcal {D} _ {s u b} |} \sum_ {x \in \mathcal {D} _ {s u b}} \mathcal {L} _ {K D} (x). +$$ + +where $\mathcal{L}_S$ is the hard target loss with ground truth (e.g., cross-entropy loss), $\mathcal{L}_S$ is the soft target loss between teacher and student, and $\alpha$ is the balancing factor. + +# 5.3. Overall Distillation Process + +Algorithm 1 briefly summarizes the entire KD process with adaptive sampling. For quantity-based subsampling, it selects the samples using $KL_{TS}$ and updates the sampling ratio $K$ , gradually decreasing it over time to implement curriculum learning. This allows the student to initially learn from diverse examples rather than focusing on the most informative ones in the early stage. For quality-based calibration, the algorithm first determines adaptive thresholds from the $KL_{TG}$ distribution (Line 2). Each sample is then weighted based on how well its $\mathrm{KL}_{TG}$ value aligns with this threshold window; samples falling outside the range receive reduced weight in proportion to their deviation. + +# 6. Evaluation + +We conduct a comprehensive evaluation of the proposed KDAS method to validate its effectiveness across various knowledge distillation scenarios. Our experiments aim to answer the following key questions: + +(1) Can KDAS improve the performance of existing KD methods for various teacher-student architectures? +(2) Can KDAS accelerate the distillation process while maintaining or even improving performance? + +# 6.1. Experimental Setup + +Model Architectures. As teacher and student models, we utilize various architectures from the ResNet [7], WideResNet (WRN) [31], VGG [23], and MobileNet [10] families. These diverse architecture pairs allow us to evaluate the effectiveness of our method across different capacity gaps and network designs. Additionally, we apply our approach to several transformer-based architectures [4, 16] to demonstrate the generalizability of KDAS to modern vision models. For each experiment, we use larger networks as teachers and smaller networks as students. All experiments are repeated using three different random seeds. + +Baseline KD Methods. To demonstrate the general applicability of our method, we apply the proposed adaptive sampling method on top of three representative logit-based KD methods that operate on different principles: KD [9], DKD [33], and LogitSTD [24]. Vanilla KD distills the knowledge of the teacher model using the general KD loss as explained in Section 3. DKD improves KD by decoupling the target-class and non-class KL loss and balancing them with hyperparameters. LogitSTD standardize the logits of the teacher and student models to reflect the performance of the student model. + +Training Parameters. For CIFAR-100, we train all models for 240 epochs using SGD with momentum 0.9, weight decay $5 \times 10^{-4}$ , and employ a cosine annealing learning rate scheduler with an initial learning rate of 0.05 (except ShuffleNet, we set to 0.01) and a minimum learning rate of $1 \times 10^{-5}$ . We use a batch size of 64 for all models. For ImageNet, we train for 100 epochs with an initial learning rate of 0.1, also using a cosine annealing scheduler with the same minimum learning rate. We use a batch size of 256. For vision transformer models, we use the same parameter settings as previous work [14, 15]. + +For KDAS-related parameters, we dynamically adjust both sampling frequency and ratio during training. For vanilla KD [9], we progressively increase the subsampling interval $\tau$ from 1 to 10 epochs while decreasing the sampling ratio $K$ from $60\%$ to $40\%$ . For DKD [33] and Logit-STD [24], we use a more conservative strategy with $\tau$ . We lower the subsampling interval from 1 to 4 epochs and the sampling ratio from $80\%$ to $60\%$ (70% sampling on average) to balance between performance and training time. + +Table 2. Comparison of knowledge distillation performance (Top-1 Accuracy, %) on CIFAR-100 across diverse teacher-student configurations. Both accuracy improvements ( $\Delta$ Acc.) and training time reductions ( $\Delta$ T.T.) are reported in the table. The underlined results are trained with a 20-epoch warm-up period. + +
TeacherResNet32x4ResNet32x4ResNet32x4WRN-40-2VGG13WRN-40-2ResNet56ResNet110ResNet110
79.4279.4279.4275.6174.6475.6172.3474.3174.31
StudentSHN-V2WRN-16-2WRN-40-2ResNet8x4VGG8WRN-40-1ResNet20ResNet32ResNet20
71.8273.2675.6172.5070.3671.9869.0671.1469.06
KD [9]74.4574.977.773.9772.9873.5470.6673.0870.67
KD + Ours74.7875.0777.8976.1173.9173.6871.6673.3170.72
Δ Acc. (%)+0.33+0.17+0.19+2.14+0.93+0.14+1.00+0.23+0.05
Δ T.T. (%)-29.91-26.54-26.87-29.54-30.22-29.14-27.10-28.11-28.78
DKD [33]77.0775.778.4675.5674.6874.8171.9774.1171.06
DKD + Ours77.1776.4378.6575.7674.7574.9771.9774.1271.81
Δ Acc. (%)+0.10+0.73+0.19+0.20+0.07+0.16+0.00+0.01+0.75
Δ T.T. (%)-10.35-9.98-10.67-12.34-12.75-9.56-8.79-8.12-9.20
LogitSTD [24]75.5675.2677.9277.1174.3674.3771.4374.1771.48
LogitSTD + Ours76.1875.3479.177.574.6374.4971.6774.4371.58
Δ Acc. (%)+0.62+0.08+1.18+0.39+0.27+0.12+0.24+0.26+0.10
Δ T.T. (%)-10.54-10.01-11.88-12.16-12.55-9.67-9.04-8.99-9.11
+ +For quality-based loss weighting, we use percentile thresholds $\theta_{\mathrm{low}} = 20$ and $\theta_{\mathrm{high}} = 80$ to define our optimal transfer window based on the $\mathrm{KL}_{\mathrm{TG}}$ distribution. We set the penalty factor $\gamma = 0.5$ as the minimum weight and penalty intensity $\lambda = 1000$ . The temperature parameter is set to 4 for all KD methods unless otherwise specified. + +# 6.2. Distillation Efficiency + +Results with CIFAR-100. Table 2 presents the top-1 accuracy of various KD methods with and without our KDAS approach on CIFAR-100 across diverse teacher-student architecture pairs. KDAS consistently improves accuracy while reducing training time across all configurations. When applied to vanilla KD [9], our method achieves substantial accuracy gains with approximately $28\%$ faster training on average, with variation depending on teacher model size. Interestingly, it seems that the performance improvement is not meaningfully affected by whether the teacher and student models are in the same architecture family. In addition, there is no evident relationship between the accuracy improvement and training time reduction. + +For advanced methods such as DKD [33] and Logit-STD [24], KDAS also provides consistent performance enhancements with around $10\%$ training time reduction. As we apply the adaptive sampling method more conservatively, the training time reduction is smaller compared with the vanilla KD. As the methods modify the loss function or the original logits, KDAS may give a different impact. + +These results demonstrate that our approach is not only effective for the standard KD method but also complementary to the state-of-the-art KD methods, highlighting the general applicability of our adaptive sampling method. + +Table 3. Top-1 accuracy (\%) on the ImageNet validation dataset. Improvements over the corresponding baseline are shown in $\Delta$ . + +
TeacherResNet34ResNet50
73.3176.16
StudentResNet18MobileNetV1
69.7568.87
KD [9]71.0370.50
KD + Ours71.3570.86
Δ Acc., T.T. (%)+0.22 -36.23+0.36 -35.81
DKD [33]71.7072.05
DKD + Ours71.8172.49
Δ Acc., T.T. (%)+0.11 -15.01+0.44 -14.74
LogitSTD [24]71.4272.18
LogitSTD + Ours71.4872.21
Δ Acc., T.T. (%)+0.06 -15.27+0.03 -15.01
+ +Results with ImageNet. On the large-scale ImageNet dataset, as shown in Table 3, KDAS maintains its effectiveness for practical teacher-student combinations. For the both pairs, our approach improves the accuracy of the student models while reducing the distillation time by over $35\%$ . With the advanced methods, KDAS continues to enhance performance with approximately $15\%$ faster training. The training time reduction is larger because KDAS can more efficiently subsample the bigger dataset. + +These results show that our approach scales well to a larger dataset where the computational efficiency is particularly more valuable. + +Table 4. Top-1 accuracy (\%) of various vision transformer models on CIFAR-100 with ResNet56 as the teacher model. Improvements over the corresponding baseline are shown in $\Delta$ . + +
MetricDeiT-TiT2T-ViT7PiT-TiPVT-Ti
Size5M4M5M13M
Accuracy65.0869.3773.5869.22
AT [32]73.5174.0176.0374.66
LG [14]78.1578.3578.4877.07
AutoKD [15]78.5878.6278.5177.48
LogitSTD [24]78.5578.4378.7678.43
KD [9]73.2574.1575.4773.60
KD + Ours77.8176.3976.8776.31
Δ Acc., (%)+4.56+2.24+1.40+2.71
Δ T.T. (%)-25.51-25.42-24.98-25.16
+ +Table 5. Impact of different KDAS components on CIFAR-100 with VGG13 $\rightarrow$ VGG8. (CS: Curriculum Sampling; PN: Quality-based Penalization). + +
VanillaSubsamplingCSPNOpt.Acc.Training time
✓73.14100 %
✓✓73.82 +0.68
✓✓✓74.08 +0.94-12.45 %
✓✓✓✓74.12 +0.98
✓✓✓✓✓73.91 +0.77-30.22 %
+ +Results with Vision Transformers. To verify the effectiveness of KDAS on modern architectures, we apply the approach to various vision transformer models on CIFAR-100. Table 4 summarizes the accuracy and training time improvement with the vanilla KD, comparing with other knowledge distillation methods. These results highlight that KDAS effectively bridges the architectural gap between the CNN-based teacher (ResNet56) and the transformer-based students while providing accuracy gains with improved distillation efficiency. + +# 6.3. Ablation Study + +Table 5 presents a systematic analysis of the contribution of each KDAS component. The quantity-based subsampling alone yields a $0.68\%$ accuracy improvement while reducing training time by $12.45\%$ . Incorporating curriculum sampling enhances accuracy further by $0.94\%$ , demonstrating the value of adaptive sample prioritization. The quality-based penalization, which calibrates the knowledge transfer, leads to the highest accuracy of $74.12\%$ . The complete method with our optimization techniques offers additional computational efficiency at a marginal accuracy trade-off. This analysis reveals that the complementary nature of our components, whose integration produces a synergistic effect that enables superior performance with substantially reduced computational cost. + +![](images/c519854b7f601258df2d67a28b7012be97e118dc251e82a98d748a7eba995ba3.jpg) + +![](images/ee74d6a3f116262a4503fda76402152331feaf812ae8855407ba856b90cdf833.jpg) + +![](images/9fbdfc8c6435bdf2563af73fd432494d0bedf3c41bb28c6413a80f3c80b285b3.jpg) +Figure 8. Training and test performance metrics across epochs for samples grouped by teacher-student KL divergence levels. + +![](images/6c2e612c4d9b350104cdbcc3aded70b37d7771b705eb9e4b7feea69579196e4a.jpg) + +# 6.4. Impact of KL Divergence Levels + +Figure 8 illustrates distinctive learning dynamics across different KL divergence levels. High-KL samples demonstrate elevated initial loss and gradual convergence during training, yet yield superior test accuracy. This behavior stems from their rich information entropy, encoding complex decision boundaries that contribute significantly to generalization capacity. Conversely, Low-KL samples exhibit rapid training convergence but inferior test performance, indicating they capture simpler patterns that facilitate memorization but inadequately support robust generalization. This information-theoretical stratification validates our adaptive sampling strategy, which dynamically prioritizes samples offering optimal knowledge transfer. + +# 7. Conclusion + +This work conducts an empirical analysis of the impact of data in knowledge distillation from two perspectives: (i) quantity of knowledge and (ii) quality of knowledge. Our analysis finds that faster knowledge distillation can be achieved by using data with a large amount of high-quality knowledge. Based on the findings, this work proposes an efficient adaptive sampling method for faster knowledge distillation, which accelerates the distillation process by selecting and applying 'good' samples for knowledge distillation. This work shows that our adaptive sampling methods can effectively accelerate the training efficiency of a student model when combined with existing KD methods. + +# Acknowledgement + +We thank the anonymous reviewers for their valuable feedbacks. This work was partly supported by Institute of Information & Communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (No. RS-2025-02214497, No. RS-2025-02263669, and RS-2022-00155911). This work was also supported by the National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT) (No. RS-2023-00207816). + +# References + +[1] N. N. Author. Suppressed for anonymity, 2021. 4 +[2] Defang Chen, Jian-Ping Mei, Hailin Zhang, Can Wang, Yan Feng, and Chun Chen. Knowledge distillation with the reused teacher classifier. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11933-11942, 2022. 2 +[3] Pengguang Chen, Shu Liu, Hengshuang Zhao, and Jiaya Jia. Distilling knowledge via knowledge review. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5008-5017, 2021. 2 +[4] Li et al. Locality guidance for improving vision transformers on tiny datasets. In ECCV, 2022. 6 +[5] Jianping Gou, Baosheng Yu, Stephen J Maybank, and Dacheng Tao. Knowledge distillation: A survey. International Journal of Computer Vision, 129(6):1789-1819, 2021. 1 +[6] Ziyao Guo, Haonan Yan, Hui Li, and Xiaodong Lin. Class attention transfer based knowledge distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11868-11877, 2023. 2 +[7] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6 +[8] Byeongho Heo, Jeesoo Kim, Sangdoo Yun, Hyojin Park, No-jun Kwak, and Jin Young Choi. A comprehensive overhaul of feature distillation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1921-1930, 2019. 2 +[9] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network, 2015. 1, 2, 3, 6, 7, 8 +[10] Andrew G Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco Andreetto, and Hartwig Adam. Mobilenets: Efficient convolutional neural networks for mobile vision applications. arXiv preprint arXiv:1704.04861, 2017. 6 +[11] Tao Huang, Shan You, Fei Wang, Chen Qian, and Chang Xu. Knowledge distillation from a stronger teacher. arXiv preprint arXiv:2205.10536, 2022. 2 +[12] Ying Jin, Jiaqi Wang, and Dahua Lin. Multi-level logit distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24276-24285, 2023. 2 + +[13] Gang Li, Xiang Li, Yujie Wang, Shanshan Zhang, Yichao Wu, and Ding Liang. Knowledge distillation for object detection via rank mimicking and prediction-guided feature imitation. In Proceedings of the AAAI conference on artificial intelligence, pages 1306-1313, 2022. 2 +[14] Kehan Li, Runyi Yu, Zhennan Wang, Li Yuan, Guoli Song, and Jie Chen. Locality guidance for improving vision transformers on tiny datasets. In European Conference on Computer Vision, pages 110-127. Springer, 2022. 6, 8 +[15] Lujun Li, Peijie Dong, Zimian Wei, and Ya Yang. Automated knowledge distillation via monte carlo tree search. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17413-17424, 2023. 6, 8 +[16] Lujun Li, Peijie Dong, Zimian Wei, and Ya Yang. Automated knowledge distillation via monte carlo tree search. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17413-17424, 2023. 6 +[17] Zheng Li, Xiang Li, Lingfeng Yang, Borui Zhao, Renjie Song, Lei Luo, Jun Li, and Jian Yang. Curriculum temperature for knowledge distillation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1504-1512, 2023. 2 +[18] Wonpyo Park, Dongju Kim, Yan Lu, and Minsu Cho. Relational knowledge distillation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3967-3976, 2019. 2 +[19] Visal Rajapakse, Ishan Karunanayake, and Nadeem Ahmed. Intelligence at the extreme edge: A survey on reformable tinyml. ACM Computing Surveys, 55(13s):1-30, 2023. 1 +[20] Adriana Romero, Nicolas Ballas, Samira Ebrahimi Kahou, Antoine Chassang, Carlo Gatta, and Yoshua Bengio. Fitnets: Hints for thin deep nets. arXiv preprint arXiv:1412.6550, 2014. 2 +[21] Zhiqiang Shen and Eric Xing. A fast knowledge distillation framework for visual recognition. In European conference on computer vision, pages 673-690. Springer, 2022. 2 +[22] Abhinav Shrivastava, Abhinav Gupta, and Ross Girshick. Training region-based object detectors with online hard example mining. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 3 +[23] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 6 +[24] Shangquan Sun, Wenqi Ren, Jingzhi Li, Rui Wang, and Xiaochun Cao. Logit standardization in knowledge distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15731-15740, 2024. 2, 6, 7, 8 +[25] Yonglong Tian, Dilip Krishnan, and Phillip Isola. Contrastive representation distillation. arXiv preprint arXiv:1910.10699, 2019. 2 +[26] Frederick Tung and Greg Mori. Similarity-preserving knowledge distillation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1365-1374, 2019. 2 +[27] Chaofei Wang, Qisen Yang, Rui Huang, Shiji Song, and Gao Huang. Efficient knowledge distillation from model check + +points. Advances in Neural Information Processing Systems, 35:607-619, 2022. 2 +[28] Lin Wang and Kuk-Jin Yoon. Knowledge distillation and student-teacher learning for visual intelligence: A review and new outlooks. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(6):3048-3068, 2022. 1 +[29] Jing Yang, Brais Martinez, Adrian Bulat, Georgios Tzimiropoulos, et al. Knowledge distillation via softmax regression representation learning. International Conference on Learning Representations (ICLR), 2021. 2 +[30] Sangdoo Yun, Seong Joon Oh, Byeongho Heo, Dongyoon Han, Junsuk Choe, and Sanghyuk Chun. Re-labeling imageNet: from single to multi-labels, from global to localized labels. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2340-2350, 2021. 2 +[31] Sergey Zagoruyko and Nikos Komodakis. Wide residual networks. arXiv preprint arXiv:1605.07146, 2016. 6 +[32] Sergey Zagoruyko and Nikos Komodakis. Paying more attention to attention: Improving the performance of convolutional neural networks via attention transfer. In International Conference on Learning Representations, 2017. 8 +[33] Borui Zhao, Quan Cui, Renjie Song, Yiyu Qiu, and Jiajun Liang. Decoupled knowledge distillation. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 11953-11962, 2022. 2, 6, 7 +[34] Borui Zhao, Quan Cui, Renjie Song, and Jiajun Liang. Dot: A distillation-oriented trainer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6189-6198, 2023. 2 \ No newline at end of file diff --git a/ICCV/2025/What to Distill_ Fast Knowledge Distillation with Adaptive Sampling/images.zip b/ICCV/2025/What to Distill_ Fast Knowledge Distillation with Adaptive Sampling/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..e52c7e8a0febd739934e1f4b397da5693ebd3ffb --- /dev/null +++ b/ICCV/2025/What to Distill_ Fast Knowledge Distillation with Adaptive Sampling/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:13fb41af0078e9001177350ab7d1651c3b54db906307faab444d9985cd31bab8 +size 520236 diff --git a/ICCV/2025/What to Distill_ Fast Knowledge Distillation with Adaptive Sampling/layout.json b/ICCV/2025/What to Distill_ Fast Knowledge Distillation with Adaptive Sampling/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..fe63775c903071d0ac6d0bdf26b8a9417e899cc3 --- /dev/null +++ b/ICCV/2025/What to Distill_ Fast Knowledge Distillation with Adaptive Sampling/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b2a0177eb7c57a049fb4e112fecebfccdddaae696423dca9fa4aeecd844a7618 +size 397968 diff --git a/ICCV/2025/What we need is explicit controllability_ Training 3D gaze estimator using only facial images/130f2ff4-e0d9-47ca-b6cb-4080460048e1_content_list.json b/ICCV/2025/What we need is explicit controllability_ Training 3D gaze estimator using only facial images/130f2ff4-e0d9-47ca-b6cb-4080460048e1_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..2d764eb39e6ae38e8b7a1db9d8acde8f4c765610 --- /dev/null +++ b/ICCV/2025/What we need is explicit controllability_ Training 3D gaze estimator using only facial images/130f2ff4-e0d9-47ca-b6cb-4080460048e1_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7108100175ee8f7f8dfcf7583eff71104739eec6ba866a2b141d10502076da7d +size 83290 diff --git a/ICCV/2025/What we need is explicit controllability_ Training 3D gaze estimator using only facial images/130f2ff4-e0d9-47ca-b6cb-4080460048e1_model.json b/ICCV/2025/What we need is explicit controllability_ Training 3D gaze estimator using only facial images/130f2ff4-e0d9-47ca-b6cb-4080460048e1_model.json new file mode 100644 index 0000000000000000000000000000000000000000..035bebc58cafb14e1e0d611e43bb07021637a280 --- /dev/null +++ b/ICCV/2025/What we need is explicit controllability_ Training 3D gaze estimator using only facial images/130f2ff4-e0d9-47ca-b6cb-4080460048e1_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:258d041a6c49b33cd0676c2b281bd46befe512825baf272ca4f1125a7b1cd698 +size 105132 diff --git a/ICCV/2025/What we need is explicit controllability_ Training 3D gaze estimator using only facial images/130f2ff4-e0d9-47ca-b6cb-4080460048e1_origin.pdf b/ICCV/2025/What we need is explicit controllability_ Training 3D gaze estimator using only facial images/130f2ff4-e0d9-47ca-b6cb-4080460048e1_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..3ab27a480a40316cef2cf61fb26e6e3b4b13ed18 --- /dev/null +++ b/ICCV/2025/What we need is explicit controllability_ Training 3D gaze estimator using only facial images/130f2ff4-e0d9-47ca-b6cb-4080460048e1_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6b5d2cb17c4cdee244e4153d91123defd3c0c3b03562439def4310a7405613a2 +size 4620361 diff --git a/ICCV/2025/What we need is explicit controllability_ Training 3D gaze estimator using only facial images/full.md b/ICCV/2025/What we need is explicit controllability_ Training 3D gaze estimator using only facial images/full.md new file mode 100644 index 0000000000000000000000000000000000000000..b047701325252d6dc26198eefcbeea3eb9ec4dc7 --- /dev/null +++ b/ICCV/2025/What we need is explicit controllability_ Training 3D gaze estimator using only facial images/full.md @@ -0,0 +1,328 @@ +# What we need is explicit controllability: Training 3D gaze estimator using only facial images + +Tingwei $\mathsf{Li}^{\dagger 1,2}$ Jun $\mathsf{Bao}^{\dagger 3}$ Zhenzhong Kuang\*1,2 Buyu Liu\*3 + +1Key Laboratory of Complex Systems Modeling and Simulation, School of Computer Science, Hangzhou Dianzi University, Hangzhou, China. + +$^{2}$ Zhejiang Key Laboratory of Space Information Sensing and Transmission, Hangzhou Dianzi University, Hangzhou, China. + +$^{3}$ Harbin Institute of Technology (Shenzhen), Shenzhen, China. + +# Abstract + +This work focuses on unsupervised 3D gaze estimation. Specifically, we adopt a learning-by-synthesis approach that trains a gaze prediction model using simulated data. Unlike existing methods that lack explicit and accurate control over facial images—particularly the eye regions—we propose a geometrically meaningful 3D representation that enables diverse, precise, and explicit control over illumination, eye regions, and gaze targets using only facial images. Given a sequence of facial images, our method constructs a mesh representation where each mesh is associated with 3D Gaussians, enabling explicit lighting control. To further enhance realism, we introduce eye-focused constraints, including a rotation symmetry protocol, as well as geometry and appearance losses for the eye regions, alongside conventional learning objectives. Additionally, we incorporate a virtual screen target and rotate the eyeballs accordingly, generating more accurate pseudo gaze directions paired with realistic facial images. We validate our approach through extensive experiments on three benchmarks. The results demonstrate that our trained gaze estimators outperform all unsupervised baselines and achieve performance comparable to cross-dataset approaches. Furthermore, our method generates the most visually realistic images, as confirmed by both objective and subjective image quality metrics. Our code is available at https://github.com/ATinyBites/ControllableGaze. + +# 1. Introduction + +3D gaze estimation aims to achieve accurate gaze prediction using facial or eye images [2, 4, 34, 46, 60, 68]. Among all, unsupervised 3D gaze estimation offers an annotation-free + +solution for identifying individuals' attention, making it a more practical choice in real-world applications [22, 27, 38]. + +Existing methods for generating gaze samples typically fall into two categories: those that generate face images without any underlying 3D representation [14, 40, 69], and those that rely on 3D representations for simulation, either implicitly [48, 49, 57] or explicitly [47, 59, 61]. While some of these methods achieve satisfactory performance on certain benchmarks, they often suffer from severe limitations, including unrealistic generated images and inaccurate paired pseudo gaze ground-truths. These shortcomings hinder their ability to match the performance of cross-dataset or supervised approaches [18, 34, 41, 65-67]. + +To this end, we propose a novel unsupervised 3D gaze estimation method following the learning-by-synthesis paradigm. Specifically, we train a gaze estimator using generated realistic face images paired with accurate pseudo gaze directions, enabled by explicit controllability over lighting, eye regions, and gaze targets. Our method begins by constructing 3D mesh representations from images of an individual, where each triangle of the mesh is associated with multiple 3D Gaussians. To enhance controllability and flexibility, we introduce an explicit lighting model over these 3D Gaussians. More importantly, we adopt a comprehensive training strategy that incorporates a rotational symmetry protocol to infer occluded eyeball regions, along with a total loss function that combines conventional losses with our proposed eye-focused losses. In particular, our eye-focused loss emphasizes both the geometric and appearance properties of visible eye regions. This complements the rotational symmetry protocol, resulting in more realistic image generation. Finally, to obtain accurate and diverse pseudo gaze directions, we introduce a virtual screen that randomly places gaze targets. By rotating the eyeballs accordingly and generating images based on these adjusted eye positions, we ensure both realism and accuracy in the synthesized gaze + +data (see Fig. 1). We perform extensive experiments on three benchmarks, comparing our method to various approaches in terms of both gaze prediction and image quality. Our method significantly surpasses the state-of-the-art (SOTA) unsupervised methods across all metrics and achieves competitive performance in cross-dataset settings. + +Our contributions can be summarized as follows: 1) A SOTA method that produces accurate pseudo gaze directions with paired realistic face images. 2) Novel strategy that enforces explicit controllability over lighting, eye regions, and gaze targets. 3) Superior results compared to SOTA methods on three benchmarks on diverse metrics. + +# 2. Related work + +3D gaze estimation. The task of 3D gaze estimation [10, 32, 33, 37, 55] aims to estimate the 3D gaze direction in the camera coordinate system based on facial or eye images. Existing methods can be categorized into model-based and appearance-based approaches. Specifically, the former estimates gaze direction by fitting a 3D eyeball model under geometric constraints [4, 7, 39, 60, 62], with or without additional calibration steps [24, 29, 30, 58]. Due to its fitting nature, this line of work generates less accurate predictions but is more robust to unseen conditions and simpler. In contrast, the appearance-based methods learn a mapping function from eye or facial images to gaze directions, leading to more accurate predictions. Starting from linear regressor [11, 31], random forests [59], and k-NN [61] that deal with hand-crafted features, modern models turn to CNNs [23, 65, 66, 68] that combine representation and learning. Compared to using eye images only, including facial images [66] and additional visual content [8, 41] would enhance performance with richer context. While appearance-based methods generally deliver better accuracy, they are data-intensive and require expensive setups and specialized equipment for large-scale dataset collection, making them labor-intensive. To address these challenges, we propose a novel unsupervised method that learns a 3D gaze estimator from generated facial images. By focusing on explicit lighting, eye region, and gaze target modeling, our method outperforms the SOTA unsupervised methods across three benchmarks. + +Learning-by-synthesis in gaze estimation. Learning-by-synthesis provides an effective way to address data issues [5, 35, 44, 54, 63]. As for 3D gaze estimation, methods can be categorized into three types depending on how 3D face/eye regions are included. For instance, SynthesEyes [59] and UnityEyes [61] construct a series of variable eye region models based on head scan geometry data, synthesizing close-up eye images with various head poses and gaze directions. However, due to their complexity of modeling and high costs [6, 19], these methods allow registration and generation of a limited range of appearances. In con + +trast, 3D Morphable Model (3DMM) [7, 28, 43, 45] and 3D Gaussian Splatting (3DGS) [1, 20, 47, 56, 64] provides a cheaper way to obtain explicit 3D representations of the head. However, they focus on the entire facial region, usually neglecting the detailed control of eye areas. Therefore, directly applying these models would lead to unsatisfactory results for 3D gaze estimation. Another line of work aims to solve the generation problem without explicitly considering geometry. For example, STED[69] and FAZE[40] employ a VAE-based framework to generate facial images that comply with gaze constraints. Some methods[14] utilize the idea of generative adversarial networks (GANs) to perform gaze redirection on input eye images. Other approaches[48] use Neural Radiance Fields (NeRF)[36] to directly generate gaze samples with implicit 3D modelling. This line of work requires a large number of labeled gaze samples during training process and often relies on a pre-trained high-performance gaze estimator as supervision. Not surprisingly, the third line of work proposes to combine the afore-mentioned two. Works such as SimGAN[49] and HGSM[57] refine the synthetic images from UnityEyes through a generator, producing more realistic eye images. Meanwhile EyeNeRF[25] and ShellNeRF[26] adopt a hybrid representation combining implicit neural representations with explicit eyeball surface modeling. However, these approaches either struggle to ensure that the gaze labels of the original images remain accurate for the refined images or rely on high-quality images and detailed manual annotations for complex appearance modeling, leading to noise labels or heavy costs. Though combining 3DMM and NeRF [3, 53, 70, 71] allows more controllability over spatial position and color, its implicit nature makes precise control of operations like eyeball rotation challenging. In contrast, ours introduces representations with explicit controllability from face images alone, without requiring human annotations, and generates realistic images paired with accurate pseudo gaze directions. + +# 3. Method + +In this section, we provide the detailed explanation of how to generate realistic and diverse facial images conditioned on pseudo ground-truth gaze directions, which is the core contribution of our paper. Once these images are generated, they are used to train a gaze estimator. + +As shown in Fig. 1, given a set of facial images, our method constructs a 3D head model for each individual, incorporating controllability over lighting conditions (see Sec.3.1). This design provides an accurate yet adaptable foundation for generating facial images. To enhance realism, particularly in the eye region, we incorporate eye-focused constraints during model training. These include a rotational symmetry protocol and a comprehensive loss function that combines conventional losses with geometry and appearance losses specifically designed for the eye area (see Sec.3.2). + +![](images/6ebce276055290a0cb0b50dd68b4ada873da9b59b678eff75a8de042b337c1c6.jpg) +Figure 1. Thanks to the explicit control over lighting, eye regions, and gaze targets, our method generates realistic images with accurate pseudo gaze directions, achieving state-of-the-art results in unsupervised gaze estimation. + +After training, we implement a virtual screen and explicitly control the eyeball rotations to align with gaze targets on the screen, resulting in realistic facial images with more precise and diverse pseudo ground-truth gaze labels (see Sec. 3.3). + +# 3.1.3D Representation with Explicit Controllability + +Given a set of facial images of a static individual along with the corresponding camera parameters, our objective is to construct a 3D head representation that enables explicit controllability, allowing for enhanced diversity and finer adjustments. To achieve this, we adopt 3DGS [20] to model the head in a geometrically meaningful and accurate form, and incorporate a binding method [47] that introduces mesh deformation to regulate the movement of the 3D Gaussians. + +3DGS represents a scene using a collection of 3D Gaussians by minimizing the loss between the input images and the images rendered from these Gaussians [20]. Specifically, each 3D Gaussian is defined by $\{\pmb {\mu},\pmb {\Sigma},\sigma ,\pmb {c}\}$ where $\pmb{\mu}$ represents the center location, $\pmb{\Sigma}$ is the covariance matrix, $\sigma$ denotes opacity, and $c$ represents color. Conventionally, we have $\pmb {\mu}\in \mathbb{R}^3$ $\sigma \in \mathbb{R}$ ,and $c\in \mathbb{R}^3$ .Given any coordinate $\pmb {x}\in \mathbb{R}^3$ in world space, the Gaussian distribution is formulated as $G(\pmb {x}) = e^{-\frac{1}{2} (\pmb {x} - \pmb {\mu})^T\pmb{\Sigma}^{-1}(\pmb {x} - \pmb {\mu})}$ . Following [20], we decompose the covariance matrix into rotation and scaling components, mathematically expressed as $\pmb {\Sigma} = R\pmb {S}\pmb{S}^{T}\pmb{R}^{T}$ ensuring its positive semi-definiteness during optimization. We further denote the scaling vector as $s\in \mathbb{R}^3$ and the rotation matrix as $r\in \mathbb{R}^{3\times 3}$ . When rendering images with + +3DGS, the color $C$ of each pixel can be obtained via: + +$$ +\boldsymbol {C} = \sum_ {i = 1} \boldsymbol {c} _ {i} \alpha_ {i} \prod_ {j = 1} ^ {i - 1} (1 - \alpha_ {j}) \tag {1} +$$ + +where $c_{i}$ denotes the color represented by the $i$ -th 3D Gaussian, $\alpha_{i}$ is the blending weight obtained by multiplying the 2D projection of this 3D Gaussian with its opacity $\sigma_{i}$ . + +Mesh binding. Enabling finer control is essential for generating subtle yet diverse images. Inspired by [47], we aim to associate each 3D Gaussian with its corresponding parent triangle on the mesh. Such a protocol allows the deformation of the mesh to directly influence the movement of the Gaussians, providing greater flexibility in controlling the generated images. + +In practice, we used a head tracker as employed in [12], which takes a set of facial images as input, tracks landmarks on these images, and outputs mainly four sets of parameters: facial topology, including both point-wise connectivity and semantic information, an expression vector, rotation matrices for both the left and right eyes, and head pose. These parameters are then passed to FLAME [28] to generate meshes. Taking the eyeball meshes as an example, given the expression vector, head pose, and semantic information, we first locate the 3D points corresponding to both eyeballs. These points are then rotated according to the respective rotation matrices. Finally, the eyeball meshes are obtained by applying point-wise connectivity to reconstruct their structure. FLAME is chosen because it provides detailed meshes for + +the eyes. We would like to note that other 3DMM methods are also applicable within our framework. The next step is to bind 3D Gaussians to these deformable meshes, where each 3D Gaussian is associated with its corresponding parent triangle [47]. We follow the convention and define the local space of one triangle through $\{T, R, w\}$ . Using the following equations, we can transform any Gaussian defined in the local space of its parent triangle, or $\{\mu', r', s'\}$ , to the world space: + +$$ +\boldsymbol {r} = \boldsymbol {R r ^ {\prime}} +$$ + +$$ +\boldsymbol {\mu} = w \boldsymbol {R} \boldsymbol {\mu} ^ {\prime} + \boldsymbol {T} \tag {2} +$$ + +$$ +s = w s ^ {\prime} +$$ + +Thanks to mesh binding, deformations applied to the parent triangles directly induce corresponding motions in the associated 3D Gaussians. This representation provides a more explicit and intuitive means of control in next steps. + +Control over Lighting. To further enhance controllability over lighting for improved diversity, we introduce a lighting model. Instead of relying on spherical harmonics to determine the color $c$ of a 3D Gaussian, we propose learning a mapping function $\Omega$ that maps the 3D position encoding $f_{\mathrm{pos}}$ of the Gaussian and the lighting feature $f_{\mathrm{light}}$ of the facial image to $c$ . Mathematically, we have: + +$$ +\boldsymbol {c} = \Omega \left(\boldsymbol {f} _ {\text {p o s}}, \boldsymbol {f} _ {\text {l i g h t}}\right) \tag {3} +$$ + +where $\Omega$ is implemented as a shallow multilayer perceptron (MLP). Specifically, both $f_{\mathrm{pos}}$ and $f_{\mathrm{light}}$ are encoded in a learnable manner to capture high-frequency color information, representing positional and lighting characteristics, respectively. During the learning process, these features are progressively updated and optimized. + +Overall, our learnable lighting model facilitates 3D face reconstruction under complex lighting conditions, ensuring greater appearance diversity in the generated facial images for subsequent steps. + +# 3.2. Representation Learning with Eye-Focused Constraints + +In the previous section, we introduce our 3D representation that enables explicit controllability. In this section, we delve into its learning process, where we incorporate eye-focused constraints to enhance realism in both geometry and appearance, particularly in the eye regions. + +Rotation Symmetry of Eyeball. The eyeball is a crucial component of the eye region, particularly for gaze estimation tasks. While the previously introduced 3D representation provides a geometrically meaningful and accurate foundation for generation, its performance in modeling the eyeball remains limited. This limitation arises primarily from potential occlusions caused by the eyelids, which obscure critical information needed to accurately capture the subtle structure + +![](images/e2a6dd62e0ace18b313e80c7332124c8069cb12b7b93b7857c2ad65ca7ed99fe.jpg) +Figure 2. We observe that the 3D eyeball typically exhibits symmetric geometry. a) We define the z-axis as the direction connecting the center of the eyeball to the pupil, which also serves as the rotational symmetry axis of the eyeball. b) Ideally, if the eyeball undergoes a random rotation around its symmetry axis (the z-axis), the appearance variance before and after the rotation should be negligible. + +![](images/8fa490a50024ebc2305ba87059e39d68702229a9ac9d55ec70ad1f2655bcc511.jpg) + +![](images/b9564993f317c33560ce13089c1789a83caec30804026573f1327103d68d31e2.jpg) + +and appearance of the eyeball. This challenge becomes even more pronounced when increasing diversity during generation, as previously unseen eyeball rotations may occur. + +To address this, we introduce a Random Symmetric Rotation (RSR) process during training. This process leverages our observation that an ideal 3D eyeball should exhibit minimal appearance changes when rotated by any angle around its symmetry axis, the $z$ -axis (see Fig. 2.a). Specifically, the structure of a single eyeball is approximately rotationally symmetric around its central axis, as illustrated in Fig. 2.b. In practice, instead of rendering facial images using the original rotation matrices of the eyeballs (see Sec. 3.1), RSR randomly rotates the eyeballs and then renders the facial images accordingly. This design allows the model to explicitly learn the appearance information of previously unseen eyeball regions by assuming their visual similarity to the visible parts. Specifically, we convert the original rotation matrices of the left and right eyeballs into Euler angles denoted as two 3D vectors, or $e_l$ and $e_r$ respectively. According to our definition in Fig. 2.a, the rotational symmetry can be achieved by including random angles to the roll components of the Euler angles for both eyeballs. Mathematically, we can update these rotation matrices by: + +$$ +\boldsymbol {e} _ {l} \leftarrow \boldsymbol {e} _ {l} + (0, 0, \delta_ {l}); \boldsymbol {e} _ {r} \leftarrow \boldsymbol {e} _ {r} + (0, 0, \delta_ {r}) \tag {4} +$$ + +where $\delta_l, \delta_r \sim \mathcal{U}(0, 2\pi)$ , meaning they are uniformly sampled between 0 and $2\pi$ . + +Intuitively, RSR enables the model to learn both the appearance and geometry of previously occluded regions in the eyeballs under the symmetry assumption. This property is highly desirable, as it not only enhances realism in unseen scenarios but also imposes stronger constraints on the eyeball shape (see Sec. 4.3). + +In addition to the geometry and appearance of the previously occluded regions of the eyeballs, which is addressed by the RSR, those of the visible regions should also be carefully considered. This design ensures the learning of a high-quality 3D representation and serves as a guarantee for the accuracy of the RSR. + +Geometry Loss of Eye Area. To ensure a strong overlap between the locations of the visible eye regions and their corresponding representation, we first extract the visible regions of the eyeball, which can be obtained from the detected landmarks in Sec. 3.1. Let the input facial image and its binary mask of visible eye region be denoted as $I \in \mathbb{R}^{H \times W \times 3}$ and $M_g$ , where $H$ and $W$ are the height and width of input images, respectively. Meanwhile, we render images based on whether the 3D Gaussians are bound to the eyeballs, denoted as $M_r$ , and evaluate the differences between $M_g$ and $M_r$ . Specifically, let $\mathcal{F}$ denote the set of all 3D Gaussians, and let $S \subset \mathcal{F}$ represent the subset of 3D Gaussians whose parent triangles belong to the eyeball. We then activate the 3D Gaussians in $S$ while deactivating the rest by using: + +$$ +\boldsymbol {c} _ {i} = \left\{ \begin{array}{l l} (1, 1, 1), & \text {i f} i \in \mathcal {S}, \\ (0, 0, 0), & \text {i f} i \in \mathcal {F} \backslash \mathcal {S}. \end{array} \right. \tag {5} +$$ + +With this formulation, we focus primarily on the spatial locations of the 3D Gaussians associated with the eyeball, largely disregarding the influence of their learned colors. By rendering the facial image $M_r$ with the color function defined in Eq. 5, we define the geometry loss as: + +$$ +\mathcal {L} _ {g} = \left\| M _ {r} - M _ {g} \right\| _ {2}. \tag {6} +$$ + +Appearance Loss of Eye Area. To improve the realism of the eye area, we compute the appearance loss between the eye region of the facial image $I$ and that of the rendered image with learned 3D Gaussians, denoted as $I_{r}$ . Mathematically, this is formulated as: + +$$ +\begin{array}{l} \mathcal {L} _ {e} = \left(1 - \lambda_ {1}\right) \mathcal {L} _ {1} \left(I _ {r} ^ {e}, I ^ {e}\right) + \lambda_ {1} \mathcal {L} _ {\mathrm {D - S S I M}} \left(I _ {r} ^ {e}, I ^ {e}\right) \tag {7} \\ + \lambda_ {2} \mathcal {L} _ {\mathrm {l p i p s}} \left(I _ {r} ^ {e}, I ^ {e}\right), \\ \end{array} +$$ + +where $\lambda_{1}$ and $\lambda_{2}$ are the weights. $I_r^e = M_g\cdot I_r$ and $I^{e} = M_{g}\cdot I$ ensuring that we focus solely on the eye region identified through detected landmarks. Following [20], our loss function incorporates an $\mathcal{L}_1$ term and a D-SSIM term to minimize pixel-wise visual differences. Additionally, we integrate $\mathcal{L}_{\mathrm{lpips}}$ [16] to enhance eye texture details by reducing discrepancies in high-level features extracted from a pre-trained deep convolutional neural network [50]. + +Total Loss. Our final loss function is defined as: + +$$ +\mathcal {L} _ {\text {t o t a l}} = \mathcal {L} _ {f} + \lambda_ {4} \mathcal {L} _ {e} + \lambda_ {5} \mathcal {L} _ {g} + \lambda_ {6} \mathcal {L} _ {\text {r e g}}, \tag {8} +$$ + +where $\mathcal{L}_f$ denotes the facial appearance loss, defined as: + +$$ +\mathcal {L} _ {f} = \left(1 - \lambda_ {1}\right) \mathcal {L} _ {1} \left(I _ {\mathrm {r}}, I\right) + \lambda_ {1} \mathcal {L} _ {\mathrm {D} - \mathrm {S S I M}} \left(I _ {\mathrm {r}}, I\right). \tag {9} +$$ + +We follow [47] to define the regularization loss $\mathcal{L}_{\mathrm{reg}}$ to constrain the scale and position of 3D Gaussians using the hyperparameters $\epsilon_{\mathrm{pos}}$ and $\epsilon_{\mathrm{scal}}$ : + +$$ +\mathcal {L} _ {\text {r e g}} = \left\| \max \left(\boldsymbol {s} ^ {\prime}, \epsilon_ {\text {s c a l}}\right) \right\| _ {2} + \lambda_ {3} \| \max \left(\boldsymbol {\mu} ^ {\prime}, \epsilon_ {\text {p o s}}\right) \| _ {2}. \tag {10} +$$ + +# 3.3. Image Generation with Virtual Gaze Target + +Assuming that we learned our explicitly controllable 3D representation with the eye-focused loss, the next step is to generate facial images with corresponding gaze directions. + +A naive approach to obtain paired data, particularly the ground truth gaze direction, is to rotate the left and right eyeballs to a given angle and render images based on the head pose and lighting conditions. However, this parallel gaze assumption is only valid when the gaze target is at infinity. It becomes inaccurate for most in-lab gaze estimation datasets, where participants are instructed to focus on screens or target positioned within a few meters. This discrepancy leads to a significant performance gap. To bridge this gap, we propose to mimic the lab environment. Specifically, we introduce a virtual 2D screen in front of the learned 3D model and randomly sample gaze targets on this screen. We then define the direction from the eyeball center to the target fixation point as the rotation target for the eyeballs, ensuring that the simulated gaze aligns with real-world fixation behavior. Consequently, the gaze directions of the left eye, right eye, and face are determined by the vectors extending from their respective centers to the gaze target (see Fig. 1). Finally, we render the corresponding facial images under diverse lighting conditions. Due to space limitation, we refer the reader to the supplementary materials for additional details on the virtual gaze target and the gaze estimator. + +# 4. Experiments + +In this section, we present detailed information on the benchmarks, baselines, evaluation metrics and implementation details. Additionally, we provide both quantitative and qualitative results. For more details, results and discussions, we refer the reader to the supplementary. + +# 4.1. Datasets, Baselines and Evaluation Metrics + +Datasets. We primarily evaluate our approach on three widely used gaze estimation benchmarks: EVE [41], Columbia [51], and MPII [66]. These benchmarks are selected based on their public availability, popularity, scale, and internal diversity. Unless otherwise specified, the performance of all baselines is reported on their official test splits. EVE was collected from 54 subjects, capturing four different camera views and totaling over 12 million head images. Of these, images from 10 subjects are designated as a separate test set. For each subject, we subsampled frames from all four viewpoints at 3 fps, resulting in 3,000 frames per viewpoint. Columbia consists of 5,880 high-resolution images collected from 56 participants in a controlled indoor environment. Each subject's data includes images captured under five fixed head poses, with 21 predefined gaze directions per pose, resulting in a total of 105 images per subject. MPII includes approximately 3,000 facial images + +
MethodTest set
EVEColumbiaMPII
SupervisedGazeRefineNetGRU[41]2.49--
Gaze-Net[34]-5.92-
Spatial Weights CNN[66]--4.8
Cross-datasetETH-XGaze*[67]6.618.237.50
MPIIFaceGaze*[65, 66]10.237.75-
Gaze360[18]-9.012.1
UnsupervisedOpenFace*[4]12.9415.0711.43
FLAME tracking*[28]11.1013.1812.06
CNN with UT[52, 65]--13.91
CNN with SynthesEyes[59]--13.55
CNN with SynthesEyes+UT[52, 59]--11.12
k-NN with UnityEyes[61]--9.95
CNN with UnityEyes*[61]16.289.4911.2
SimGAN[49]--7.8
HGSM[57]--7.6
Kuang et al.[24]-9.0-
Qammaz et al.[46]-18.7-
Ours9.006.627.39
+ +Table 1. Comparisons with SOTA gaze estimation methods on the EVE, Columbia, and MPII. Best numbers are highlighted with bold for each type of methods. + +![](images/a0bfdc6372e2c0b6b1ef771a9114e5c2285c64c0aad25c9e81ee950fc2066eec.jpg) +Figure 3. The gaze distribution of the real dataset (left column) and our generated ones (right column). Ours provide more diverse and continuous gaze distributions compared to real ones. + +per subject from 15 participants, captured over a 24-hour period to encompass significant variations in illumination. + +Baselines. We evaluate three types of baselines: supervised, cross-dataset, and unsupervised methods. Specifically, we introduce three Supervised methods, where models are trained on the training split of the target dataset [34, 41, 66]. Cross-dataset methods are trained on a source dataset with human-annotated ground truth and paired images, and then directly applied to the target dataset—we include three such methods in our experiments [18, 65, 67]. Finally, and most importantly, we evaluate 11 Unsupervised methods [4, 24, 28, 46, 49, 57, 59, 65] where human-annotated gaze directions are not available in any form. Unless otherwise noted with $*$ (indicating our implementation), all reported numbers are taken from their original papers. + +Evaluation Metrics. We employ a diverse set of evaluation metrics to comprehensively assess the performance of our method, covering both gaze accuracy and image quality. For gaze accuracy, we use the 3D Angular Error [9], which measures the angular difference between predicted and ground-truth gaze directions, where lower values indicate better performance. For image quality, we follow standard practices in image generation literature [15, 17] and report Fréchet Inception Distance (FID) and Kernel Inception Distance (KID). Additionally, we conduct user studies to provide a more nuanced and reliable assessment of the generated images. Specifically, users are presented with generated images from different methods anonymously and asked to vote for the one they find most realistic. Overall, our + +evaluation includes both objective and subjective analyses. + +Implementation details. Our codebase is implemented using PyTorch [42] with Adam optimizer [21]. The learning rates for $f_{\mathrm{pos}}$ , $f_{\mathrm{light}}$ , and $\Omega$ are set to 1e-4, 2.5e-3 and 1e-4, respectively. The model is trained for 30k iterations, with density control beginning at the 10k-th iteration and applied every 2k iterations. ResNet18 [13] is employed to train a gaze estimator using our generated data by following [67]. + +# 4.2. Main Results + +Gaze Prediction Performance. First and foremost, we examine whether our method effectively enhances the gaze estimation task. To this end, we present a comprehensive comparison against all baselines across three benchmarks in terms of gaze prediction accuracy and report their angular error in Tab. 1. As expected, supervised methods consistently achieve the best performance among all baselines. Meanwhile, our approach consistently outperforms other unsupervised baselines. Specifically, we surpass the second-best unsupervised method by relative margins of $18.92\%$ , $26.44\%$ , and $2.78\%$ on EVE, Columbia, and MPII, respectively. We attribute this primarily to our method's ability to generate highly accurate gaze directions along with their corresponding facial images. An interesting observation is that, across the three benchmarks, our method even outperforms the SOTA cross-dataset baselines on two of them. Specifically, we surpass the best cross-dataset baseline by 1.13 and 0.11 degrees on Columbia and MPII, respectively. In all, our generated facial images contribute positively to real- + +
MethodFID↓KID↓Voting (%)↑
UnityEyes[61]105.240.0936.4
SimGAN[49]36.440.03322.7
GaussianAvatars[47]32.040.01724.5
Ours12.180.01146.4
+ +Table 2. Comparisons of Image Quality. To ensure a fair comparison, we convert all images to grayscale using the same function, as [49] operates exclusively in grayscale. + +![](images/f64575f2d66908bbc57c4744ec0a7576c9e7ceb8e62f773cc319e0ffeaf57281.jpg) +Figure 4. We are able to generate images with diverse gaze directions, head poses, and lighting conditions. + +world gaze estimation tasks. While our method is not yet comparable to supervised approaches, it remains a reliable choice when no human annotations are available. Moreover, our method provides a competitive alternative to models pre-trained on other human-annotated datasets. + +Image Quality Evaluation. Besides generating accurate pseudo ground-truths along with their corresponding facial images, we emphasize that our method also enhances the realism of the generated images. To validate this, we present three evaluation metrics in Tab. 2. Specifically, experiments are conducted on the MPII dataset, where each method generates $45\mathrm{k}$ images. These images are first normalized to the eye region and then compared against real eye regions to compute the FID and KID scores. Additionally, we report the average voting rate based on the generated eye images in the user study (Voting). + +Among all baselines, our method clearly stands out, outperforming others by a significant margin across all evaluation metrics. For instance, we achieve nearly twice the performance of the second-best method in terms of FID, KID, and Voting rate. Specifically, UnityEyes is constrained by a limited number of registered appearances and a substantial domain gap, leading to less realistic eye images. Although SimGAN partially mitigates this issue through image finetuning, it still suffers from artifacts such as non-circular + +![](images/77ecf34acb7c19d9b1057986c413cea5b0d0cce16ba4ca16d143c4502336c0f8.jpg) +Figure 5. Our generated eye regions appear more visually realistic (see accurately aligned landmarks in green) while maintaining precise gaze directions (angles displayed at the bottom right and visualized with a red arrow). + +![](images/4ac850eccbb7b01385b1390250e7be10f04d25cd282c5426f3c05eec3fef57f8.jpg) +Figure 6. Qualitative comparison with eye image generation methods. Ours is visually more realistic among all. + +irises. In contrast, GaussianAvatars improves realism by leveraging explicit 3D reconstruction. However, due to its lack of precise control over lighting and eye regions, the generated images appear less visually convincing than ours. Overall, our method produces the most realistic eye images compared to SOTA baselines. + +Qualitative Examples. Here we provide visual examples of our method, including both the distribution of gaze directions in both real and generated datasets, generated facial and eye images, and visual comparisons between various methods. Clearly, our approach enables the generation of diverse facial images with varying gaze directions, head poses, and lighting conditions, as supported by Fig. 3 and Fig. 4. This, in turn, allows for training more robust gaze estimators capable of handling unseen scenarios. More importantly, due to our explicit controllability and carefully designed loss functions for the eyes, the generated eye regions exhibit greater visual realism while preserving precise gaze directions in Fig. 5. Compared to other methods, our approach produces more realistic eye images, effectively reducing artifacts on the eyelids, as can be found in Fig. 6. + +# 4.3. Ablation study + +Effectiveness of Each Component. To evaluate the effectiveness of each component in our model, we conduct an ablation study on MPII and present the results in Tab. 3. Clearly, directly applying the baseline [47] results in suboptimal performance in both gaze prediction accuracy and image + +
Method3D Angular Error↓FID↓
Baseline [47]8.25°32.04
Ours (eye)7.48°19.95
Ours7.39°19.95
w/o RSR7.84°18.39
w/o Lg7.97°19.24
w/o screen8.03°20.10
w/o Le8.49°29.29
w/o Ω7.90°27.65
+ +Table 3. Ablation study results by removing the key components. Best and second-best are highlighted with bold and underline. + +![](images/e4bf30bd50c4ec0c53325ac9d2bb675168dc575d743791d24b1fa4950a5c0f42.jpg) +Figure 7. The qualitative assessment of each component. The original images are rendered by manipulating 3D Gaussians corresponding to the entire head, the periocular region, and the eyeballs (highlighted with a blue mask). Our analysis primarily focuses on the areas enclosed within the bounding boxes. + +quality. In contrast, our full model achieves the best overall prediction performance. Removing the virtual screen target, the appearance-based eye-focused loss $\mathcal{L}_e$ , or the lighting condition $\Omega$ leads to a noticeable performance drop across both metrics, confirming that these components are essential and effective. Interestingly, the impact of removing RSR or $\mathcal{L}_g$ is less straightforward. While their absence leads to a clear degradation in gaze prediction accuracy, their effect on image quality is less pronounced. This could be due to the fact that the rotation symmetry assumption does not always hold perfectly, as reflected in the FID scores. Meanwhile, the numerical gap in FID with or without $\mathcal{L}_g$ is not significant. We present a visual example in Fig. 7 to illustrate the contribution of each component, where artifacts are clearly visible when any of these components are omitted. + +Effect of Number of Image Samples. We further demonstrate the impact of the number of samples on MPII for the gaze estimation task in Fig. 8. As shown, there is a turning point (e.g. 50) for performance improvement, and the perfor + +![](images/832b4d339f12613102212230cd718c6f2625a3c41828ac1a684a3e3a3eeb3f25.jpg) +Figure 8. Impact of numbers of Figure 9. Our generated nonimages per subject. drowsy and drowsy samples. + +![](images/ba2daf57dd48577952db9dc842b603960ccb9e1a9bc0205196188b8dd8285cf2.jpg) + +mance almost plateaus after 150. Notably, ours surpasses the Cross-dataset with only 100 samples per subject, illustrating the potential of data efficiency and real-world applicability. + +Is Facial Image Necessary? A common question that may arise is whether generating full facial images is necessary or if focusing solely on the eye region would be sufficient for gaze estimation. Our answer is that facial images provide more informative cues than isolated eye regions, leading to improved performance. This is supported not only by existing literature [66] but also by our experiments in Tab. 3, where eye patches (See Ours (eye)) result in inferior performance compared to full facial images on MPII. Moreover, thanks to our explicit control, the generated images can also benefit other tasks, such as fatigue detection, where symptoms like drowsy eyes serve as critical visual cues. As illustrated in Fig. 9, by controlling mouth and eyelid closure, we can generate more vivid examples that enhance performance in such applications. + +# 5. Conclusion + +In this work, we propose a novel unsupervised approach where realistic facial images along with accurate gaze directions are generated, thereby enabling the training of 3D gaze estimators without manual supervision. To achieve this, we introduce a geometrically meaningful 3D representation that allows diverse, precise, and explicit control over lighting, eye regions, and gaze targets using only facial images. Experimental results on three benchmarks demonstrate the superiority of our approach in both gaze prediction performance and image quality measurement. + +# Acknowledgement + +This work was supported by National Natural Science Foundation of China (Grant No. 62372147, U21B2040, 62406281) and Zhejiang Provincial Natural Science Foundation of China (Grant No. LY22F020028). + +# References + +[1] Rameen Abdal, Wang Yifan, Zifan Shi, Yinghao Xu, Ryan Po, Zhengfei Kuang, Qifeng Chen, Dit-Yan Yeung, and Gordon Wetzstein. Gaussian shell maps for efficient 3d human generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9441-9451, 2024. 2 +[2] Ahmed A Abdelrahman, Thorsten Hempel, Aly Khalifa, Ayoub Al-Hamadi, and Laslo Dinges. L2cs-net: Fine-grained gaze estimation in unconstrained environments. In Proceedings of the International Conference on Frontiers of Signal Processing, pages 98-102, 2023. 1 +[3] ShahRukh Athar, Zhixin Shu, and Dimitris Samaras. Flame-in-nerf: Neural control of radiance fields for free view face animation. In Proceedings of the IEEE 17th International Conference on Automatic Face and Gesture Recognition, pages 1–8, 2023. 2 +[4] Tadas Baltrusaitis, Peter Robinson, and Louis-Philippe Morency. Openface: an open source facial behavior analysis toolkit. In Proceedings of the IEEE Winter Conference on Applications of Computer Vision, pages 1-10, 2016. 1, 2, 6 +[5] Yiwei Bao, Zhiming Wang, and Feng Lu. Gazegene: Large-scale synthetic gaze dataset with 3d eyeball annotations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2025. 2 +[6] Pascal Bérard, Derek Bradley, Maurizio Nitti, Thabo Beeler, and Markus H Gross. High-quality capture of eyes. ACM Transactions on Graphics, 33(6):223-1, 2014. 2 +[7] Volker Blanz and Thomas Vetter. A morphable model for the synthesis of 3d faces. In Seminal Graphics Papers: Pushing the Boundaries, Volume 2, pages 157-164. 2023. 2 +[8] Yihua Cheng and Feng Lu. Dvgaze: Dual-view gaze estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20632-20641, 2023. 2 +[9] Yihua Cheng, Haofei Wang, Yiwei Bao, and Feng Lu. Appearance-based gaze estimation with deep learning: A review and benchmark. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46, 2024. 6 +[10] Kenneth A Funes-Mora and Jean-Marc Odobez. Gaze estimation in the 3d space using rgb-d sensors: towards head-posed and user invariance. International Journal of Computer Vision, 118:194–216, 2016. 2 +[11] Kenneth Alberto Funes Mora, Florent Monay, and Jean-Marc Odobez. Eyediap: A database for the development and evaluation of gaze estimation algorithms from rgb and rgb-d cameras. In Proceedings of the Symposium on Eye Tracking Research and Applications, pages 255-258, 2014. 2 +[12] Philip-William Grassal, Malte Prinzler, Titus Leistner, Carsten Rother, Matthias Nießner, and Justus Thies. Neural head avatars from monocular rgb videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18653–18664, 2022. 3 +[13] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings + +of the IEEE Conference on Computer Vision and Pattern Recognition, pages 770-778, 2016. 6 +[14] Zhe He, Adrian Spurr, Xucong Zhang, and Otmar Hilliges. Photo-realistic monocular gaze redirection using generative adversarial networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6932-6941, 2019. 1, 2 +[15] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In Advances in Neural Information Processing Systems, 2020. 6 +[16] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In European Conference on Computer Vision, pages 694–711, 2016. 5 +[17] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4401-4410, 2019. 6 +[18] Petr Kellnhofer, Adria Recasens, Simon Stent, Wojciech Matusik, and Antonio Torralba. Gaze360: Physically unconstrained gaze estimation in the wild. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6912-6921, 2019. 1, 6 +[19] Glenn Kerbiriou, Quentin Avril, Fabien Danieau, and Maud Marchal. Detailed eye region capture and animation. In Computer Graphics Forum, pages 279–282, 2022. 2 +[20] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42(4), 2023. 2, 3, 5 +[21] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6 +[22] Hong Vin Koay, Joon Huang Chuah, Chee-Onn Chow, and Yang-Lang Chang. Detecting and recognizing driver distraction through various data modality using machine learning: A review, recent advances, simplified framework and open challenges (2014-2021). Engineering Applications of Artificial Intelligence, 115:105309, 2022. 1 +[23] Kyle Krafka, Aditya Khosla, Petr Kellnhofer, Harini Kannan, Suchendra Bhandarkar, Wojciech Matusik, and Antonio Torralba. Eye tracking for everyone. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2176-2184, 2016. 2 +[24] Chenyi Kuang, Jeffery O Kephart, and Qiang Ji. Towards an accurate 3d deformable eye model for gaze estimation. In International Conference on Pattern Recognition, pages 109-123, 2022. 2, 6 +[25] Gengyan Li, Abhimitra Meka, Franziska Mueller, Marcel C Buehler, Otmar Hilliges, and Thabo Beeler. Eyenerf: a hybrid representation for photorealistic synthesis, animation and relighting of human eyes. ACM Transactions on Graphics, 41 (4):1-16, 2022. 2 +[26] Gengyan Li, Kripasindhu Sarkar, Abhinitra Meka, Marcel Buehler, Franziska Mueller, Paulo Gotardo, Otmar Hilliges, and Thabo Beeler. Shellnerf: Learning a controllable high-resolution model of the eye and periocular region. In Computer Graphics Forum, page e15041, 2024. 2 + +[27] Jing Li, Zejin Chen, Yihao Zhong, Hak-Keung Lam, Junxia Han, Gaoxiang Ouyang, Xiaoli Li, and Honghai Liu. Appearance-based gaze estimation for asd diagnosis. IEEE Transactions on Cybernetics, 52(7):6504-6517, 2022. 1 +[28] Tianye Li, Timo Bolkart, Michael J Black, Hao Li, and Javier Romero. Learning a model of facial shape and expression from 4d scans. TOG, 36(6):194-1, 2017. 2, 3, 6 +[29] Jiahui Liu, Jiannan Chi, Wenxue Hu, and Zhiliang Wang. 3d model-based gaze tracking via iris features with a single camera and a single light source. IEEE Transactions on Human-Machine Systems, 51(2):75-86, 2020. 2 +[30] Conny Lu, Praneeth Chakravarthula, Kaihao Liu, Xixiang Liu, Siyuan Li, and Henry Fuchs. Neural 3d gaze: 3d pupil localization and gaze tracking based on anatomical eye model and neural refraction correction. In Proceedings of the IEEE International Symposium on Mixed and Augmented Reality, pages 375–383, 2022. 2 +[31] Feng Lu, Yusuke Sugano, Takahiro Okabe, and Yoichi Sato. Inferring human gaze from appearance via adaptive linear regression. In Proceedings of the International Conference on Computer Vision, pages 153-160, 2011. 2 +[32] Feng Lu, Takahiro Okabe, Yusuke Sugano, and Yoichi Sato. Learning gaze biases with head motion for head pose-free gaze estimation. Image and Vision Computing, 32(3):169-179, 2014. 2 +[33] Feng Lu, Yusuke Sugano, Takahiro Okabe, and Yoichi Sato. Gaze estimation from eye appearance: A head pose-free method via eye image synthesis. IEEE Transactions on Image Processing, 24(11):3680-3693, 2015. 2 +[34] Bhanuka Mahanama, Yasith Jayawardana, and Sampath Jayarathna. Gaze-net: Appearance-based gaze estimation using capsule networks. In Proceedings of the 11th Augmented Human International Conference, pages 1-4, 2020. 1, 6 +[35] Yu Meng, Jiaxin Huang, Yu Zhang, and Jiawei Han. Generating training data with language models: Towards zero-shot language understanding. In Advances in Neural Information Processing Systems, 2022. 2 +[36] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2 +[37] Kenneth Alberto Funes Mora and Jean-Marc Odobez. Person independent 3d gaze estimation from remote rgb-d cameras. In Proceedings of the IEEE International Conference on Image Processing, pages 2787-2791, 2013. 2 +[38] Carlos H Morimoto and Marcio RM Mimica. Eye gaze tracking techniques for interactive applications. Computer Vision and Image Understanding, 98(1):4-24, 2005. 1 +[39] Seonwook Park, Xuong Zhang, Andreas Bulling, and Otmar Hilliges. Learning to find eye region landmarks for remote gaze estimation in unconstrained settings. In Proceedings of the Symposium on Eye Tracking Research and Applications, pages 1-10, 2018. 2 +[40] Seonwook Park, Shalini De Mello, Pavlo Molchanov, Umar Iqbal, Otmar Hilliges, and Jan Kautz. Few-shot adaptive gaze estimation. In Proceedings of the IEEE/CVF International + +Conference on Computer Vision, pages 9368-9377, 2019. 1, 2 +[41] Seonwook Park, Emre Aksan, Xuong Zhang, and Otmar Hilliges. Towards end-to-end video-based eye-tracking. In European Conference on Computer Vision, pages 747-763, 2020. 1, 2, 5, 6 +[42] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. In Advances in Neural Information Processing Systems, 2019. 6 +[43] Pascal Paysan, Reinhard Knothe, Brian Amberg, Sami Romdhani, and Thomas Vetter. A 3d face model for pose and illumination invariant face recognition. In Proceedings of the IEEE International Conference on Advanced Video and Signal Based Surveillance, pages 296-301, 2009. 2 +[44] Xingchao Peng and Kate Saenko. Synthetic to real adaptation with generative correlation alignment networks. In Proceedings of the IEEE Winter Conference on Applications of Computer Vision, pages 1982-1991, 2018. 2 +[45] Stylianos Ploumpis, Haoyang Wang, Nick Pears, William AP Smith, and Stefanos Zafeiriou. Combining 3d morphable models: A large scale face-and-head model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10934-10943, 2019. 2 +[46] Ammar Qammaz and Antonis A Argyros. A unified approach for occlusion tolerant 3d facial pose capture and gaze estimation using mocapnets. In Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops, pages 3178-3188, 2023. 1, 6 +[47] Shenhan Qian, Tobias Kirschstein, Liam Schoneveld, Davide Davoli, Simon Giebenhain, and Matthias Nießner. Gaussianavatars: Photorealistic head avatars with rigged 3d gaussians. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20299-20309, 2024. 1, 2, 3, 4, 5, 7, 8 +[48] Alessandro Ruzzi, Xiangwei Shi, Xi Wang, Gengyan Li, Shalini De Mello, Hyung Jin Chang, Xuong Zhang, and Otmar Hilliges. Gazenerf: 3d-aware gaze redirection with neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9676-9685, 2023. 1, 2 +[49] Ashish Shrivastava, Tomas Pfister, Oncel Tuzel, Joshua Susskind, Wenda Wang, and Russell Webb. Learning from simulated and unsupervised images through adversarial training. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2107-2116, 2017. 1, 2, 6, 7 +[50] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 5 +[51] Brian A Smith, Qi Yin, Steven K Feiner, and Shree K Nayar. Gaze locking: passive eye contact detection for human-object interaction. In Proceedings of the 26th annual ACM symposium on User interface software and technology, pages 271-280, 2013. 5 +[52] Yusuke Sugano, Yasuyuki Matsushita, and Yoichi Sato. Learning-by-synthesis for appearance-based 3d gaze estima + +tion. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1821-1828, 2014. 6 +[53] Kartik Teotia, Xingang Pan, Hyeongwoo Kim, Pablo Garrido, Mohamed Elgharib, and Christian Theobalt. HQ3javatar: High-quality implicit 3d head avatar. ACM Transactions on Graphics, 43(3):1-24, 2024. 2 +[54] Josh Tobin, Rachel Fong, Alex Ray, Jonas Schneider, Wojciech Zaremba, and Pieter Abbeel. Domain randomization for transferring deep neural networks from simulation to the real world. In Proceedings of the IEEE/RSJ International Conference on Intelligent Robots and Systems, pages 23-30, 2017. 2 +[55] Evangelos Ververas, Polydefkis Gkagkos, Jiankang Deng, Michail Christos Doukas, Jia Guo, and Stefanos Zafeiriou. 3dgazenet: Generalizing 3d gaze estimation with weak-supervision from synthetic views. In European Conference on Computer Vision, pages 387-404, 2024. 2 +[56] Jie Wang, Jiu-Cheng Xie, Xianyan Li, Feng Xu, Chi-Man Pun, and Hao Gao. Gaussianhead: High-fidelity head avatars with learnable gaussian derivation. IEEE Transactions on Visualization and Computer Graphics, pages 4141-4154, 2025. 2 +[57] Kang Wang, Rui Zhao, and Qiang Ji. A hierarchical generative model for eye image synthesis and eye gaze estimation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 440-448, 2018. 1, 2, 6 +[58] Quan Wen, Derek Bradley, Thabo Beeler, Seonwook Park, Otmar Hilliges, Junhai Yong, and Feng Xu. Accurate real-time 3d gaze tracking using a lightweight eyeball calibration. In Computer Graphics Forum, pages 475-485, 2020. 2 +[59] Erroll Wood, Tadas Baltrusaitis, Xuong Zhang, Yusuke Sugano, Peter Robinson, and Andreas Bulling. Rendering of eyes for eye-shape registration and gaze estimation. In Proceedings of the IEEE International Conference on Computer Vision, pages 3756-3764, 2015. 1, 2, 6 +[60] Erroll Wood, Tadas Baltrusaitis, Louis-Philippe Morency, Peter Robinson, and Andreas Bulling. A 3d morphable eye region model for gaze estimation. In European Conference on Computer Vision, pages 297–313, 2016. 1, 2 +[61] Erroll Wood, Tadas Baltrusaitis, Louis-Philippe Morency, Peter Robinson, and Andreas Bulling. Learning an appearance-based gaze estimator from one million synthesised images. In Proceedings of the Ninth Biennial ACM Symposium on Eye Tracking Research Applications, pages 131–138, 2016. 1, 2, 6, 7 +[62] Erroll Wood, Tadas Baltrusaitis, Louis-Philippe Morency, Peter Robinson, and Andreas Bulling. Gazedirector: Fully articulated eye gaze redirection in video. In Computer Graphics Forum, pages 217–225, 2018. 2 +[63] Erroll Wood, Tadas Baltrusaitis, Charlie Hewitt, Sebastian Dziadzio, Thomas J Cashman, and Jamie Shotton. Fake it till you make it: face analysis in the wild using synthetic data alone. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3681-3691, 2021. 2 +[64] Jun Xiang, Xuan Gao, Yudong Guo, and Juyong Zhang. Flashavatar: High-fidelity head avatar with efficient gaussian embedding. In Proceedings of the IEEE/CVF Conference on + +Computer Vision and Pattern Recognition, pages 1802-1812, 2024. 2 +[65] Xucong Zhang, Yusuke Sugano, Mario Fritz, and Andreas Bulling. Appearance-based gaze estimation in the wild. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4511-4520, 2015. 1, 2, 6 +[66] Xucong Zhang, Yusuke Sugano, Mario Fritz, and Andreas Bulling. It's written all over your face: Full-face appearance-based gaze estimation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 51-60, 2017. 2, 5, 6, 8 +[67] Xucong Zhang, Seonwook Park, Thabo Beeler, Derek Bradley, Siyu Tang, and Otmar Hilliges. Eth-xgaze: A large scale dataset for gaze estimation under extreme head pose and gaze variation. In European Conference on Computer Vision, pages 365-381. Springer, 2020. 1, 6 +[68] Yuanyuan Zhang, Jing Li, and Gaoxiang Ouyang. Gaze estimation with multi-scale attention-based convolutional neural networks. In Proceedings of the International Conference on Mechatronics and Machine Vision in Practice, pages 1-6, 2023. 1, 2 +[69] Yufeng Zheng, Seonwook Park, Xuong Zhang, Shalini De Mello, and Otmar Hilliges. Self-learning transformations for improving gaze and head redirection. In Advances in Neural Information Processing Systems, 2020. 1, 2 +[70] Yufeng Zheng, Wang Yifan, Gordon Wetzstein, Michael J Black, and Otmar Hilliges. Pointavatar: Deformable point-based head avatars from videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21057-21067, 2023. 2 +[71] Wojciech Zielonka, Timo Bolkart, and Justus Thies. Instant volumetric head avatars. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4574-4584, 2023. 2 \ No newline at end of file diff --git a/ICCV/2025/What we need is explicit controllability_ Training 3D gaze estimator using only facial images/images.zip b/ICCV/2025/What we need is explicit controllability_ Training 3D gaze estimator using only facial images/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..a77bff9ea5b44e53e5134a555d7a07a5ec9ac525 --- /dev/null +++ b/ICCV/2025/What we need is explicit controllability_ Training 3D gaze estimator using only facial images/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4a32652e95443a0e6b6277bc169ea52d52f17ae3424feb7a841b159e9b44fd6a +size 533798 diff --git a/ICCV/2025/What we need is explicit controllability_ Training 3D gaze estimator using only facial images/layout.json b/ICCV/2025/What we need is explicit controllability_ Training 3D gaze estimator using only facial images/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..3131a42b2d49c77ec18dae2739d581f1c839d818 --- /dev/null +++ b/ICCV/2025/What we need is explicit controllability_ Training 3D gaze estimator using only facial images/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7cddb3e9c1f7f3ff33651ae9a33a432a07e9985bce00ece6a2e4a2e91cb5361a +size 389097 diff --git a/ICCV/2025/What's Making That Sound Right Now_ Video-centric Audio-Visual Localization/44803fdf-7cfa-4917-b402-d01e2447466c_content_list.json b/ICCV/2025/What's Making That Sound Right Now_ Video-centric Audio-Visual Localization/44803fdf-7cfa-4917-b402-d01e2447466c_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..c0cdc475ae050c88d90502ddc26860b417c61286 --- /dev/null +++ b/ICCV/2025/What's Making That Sound Right Now_ Video-centric Audio-Visual Localization/44803fdf-7cfa-4917-b402-d01e2447466c_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2049ed8e86d65c282132195209ffd12d0a19948c220d21f4c0f7249f4165f251 +size 78625 diff --git a/ICCV/2025/What's Making That Sound Right Now_ Video-centric Audio-Visual Localization/44803fdf-7cfa-4917-b402-d01e2447466c_model.json b/ICCV/2025/What's Making That Sound Right Now_ Video-centric Audio-Visual Localization/44803fdf-7cfa-4917-b402-d01e2447466c_model.json new file mode 100644 index 0000000000000000000000000000000000000000..bc8d1cae162868e7b57b0be2adb18412d8a4598a --- /dev/null +++ b/ICCV/2025/What's Making That Sound Right Now_ Video-centric Audio-Visual Localization/44803fdf-7cfa-4917-b402-d01e2447466c_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ff72a9d27f4130058e504128446090f0b4cca435ff9c949f5e869dcf680f4b29 +size 94148 diff --git a/ICCV/2025/What's Making That Sound Right Now_ Video-centric Audio-Visual Localization/44803fdf-7cfa-4917-b402-d01e2447466c_origin.pdf b/ICCV/2025/What's Making That Sound Right Now_ Video-centric Audio-Visual Localization/44803fdf-7cfa-4917-b402-d01e2447466c_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..559c444e49472824739050d7adbbea84724bc0ed --- /dev/null +++ b/ICCV/2025/What's Making That Sound Right Now_ Video-centric Audio-Visual Localization/44803fdf-7cfa-4917-b402-d01e2447466c_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:52cb1dda9c62faf2a5c40a5bec14faf5280a8be2c047d9215517b7380e6cebf8 +size 27402346 diff --git a/ICCV/2025/What's Making That Sound Right Now_ Video-centric Audio-Visual Localization/full.md b/ICCV/2025/What's Making That Sound Right Now_ Video-centric Audio-Visual Localization/full.md new file mode 100644 index 0000000000000000000000000000000000000000..dd2822f9d4fff868c42cb24e3d3afdb4c7d4a6ce --- /dev/null +++ b/ICCV/2025/What's Making That Sound Right Now_ Video-centric Audio-Visual Localization/full.md @@ -0,0 +1,342 @@ +# What's Making That Sound Right Now? Video-centric Audio-Visual Localization + +Hahyeon Choi Junhoo Lee Nojun Kwak Seoul National University + +{hahyeon.choi, mrjunoo, nojunk}@snu.ac.kr + +# Abstract + +Audio-Visual Localization (AVL) aims to identify sound-emitting sources within a visual scene. However, existing studies focus on image-level audio-visual associations, failing to capture temporal dynamics. Moreover, they assume simplified scenarios where sound sources are always visible and involve only a single object. To address these limitations, we propose AVATAR, a video-centric AVL benchmark that incorporates high-resolution temporal information. AVATAR introduces four distinct scenarios - Single-sound, Mixed-sound, Multi-entity, and Off-screen - enabling a more comprehensive evaluation of AVL models. Additionally, we present TAVLO, a novel video-centric AVL model that explicitly integrates temporal information. Experimental results show that conventional methods struggle to track temporal variations due to their reliance on global audio features and frame-level mappings. In contrast, TAVLO achieves robust and precise audio-visual alignment by leveraging high-resolution temporal modeling. Our work empirically demonstrates the importance of temporal dynamics in AVL and establishes a new standard for video-centric audio-visual localization. $^{1}$ + +# 1. Introduction + +Humans and animals perceive the world by integrating multiple sensory modalities. In particular, vision and hearing complement each other, enabling the localization and identification of sound sources across a broad spatial field. This capability is essential for filtering relevant auditory information, such as distinguishing a speaker in a crowded environment or detecting a predator's location from its roar. Inferring spatial information from multimodal cues has broad applications in areas such as robot navigation, AR/VR, and video analysis. + +Audio-Visual Localization (AVL) seeks to replicate this perceptual ability by identifying sound sources within a visual scene. A core approach in AVL research leverages the natural co-occurrence of audio and visual signals, using + +self-supervised learning to align these modalities and extract joint embeddings. This paradigm has led to significant advancements in AVL [1, 6, 23, 25, 33, 35], enabling models to effectively utilize large-scale web video datasets without requiring manual annotations. However, despite this progress, existing AVL research faces two key limitations, largely stemming from the structural constraints of current AVL benchmarks. + +The first limitation is that AVL research primarily focuses on image-level audio-visual associations rather than extending to video-based analysis. Most benchmarks [7, 29] adopt an annotation approach where annotators watch the entire video and label all sound-emitting objects in a single frame, effectively treating that frame as representative of the entire video. Consequently, existing methods [12, 24, 30, 32] operate on single-frame inputs, neglecting temporal dynamics. While this approach assesses spatial understanding within static images, it fails to capture temporal variations. Real-world AVL tasks require tracking moving sound sources and handling dynamic changes over time, making spatiotemporal modeling essential for robust performance. + +The second limitation is the oversimplified assumptions in existing AVL benchmarks [7, 29, 31, 36]. These benchmarks assume that sound-emitting objects are always visible and typically involve only a single active source. However, real-world scenarios often feature multiple simultaneous sound sources, and sound-emitting objects may be outside the visual field. These constraints limit model generalization to complex audio-visual environments. Although some studies attempt to address these issues, they primarily rely on negative pair construction using mismatched image-audio pairs [20] or synthetic mixed-audio generation [31], providing only partial solutions. + +To overcome these limitations, we propose advanced Audio-Visual localization benchmark for a spatio-TemporAl peRspective in video (AVATAR), a video-centric AVL benchmark that evaluates models on their ability to localize sound sources with high temporal resolution. As summarized in Tab. 1, AVATAR introduces four key evaluation scenarios that reflect real-world complexity: Single-sound, Mixed-sound, Multi-entity, and Off-screen. + +Table 1. Comparison of Audio-Visual Localization Datasets. A comparison of existing AVL datasets in terms of dataset scale, annotation type, and the scenarios they cover, including Single-sound, Mixed-sound, Multi-entity, and Off-screen. § Statistics for AVSBench and AVSBench-Semantic are based on the test set due to supervised use. + +
Dataset# Videos# Frames# CategoriesAvg. LengthAnnotationAnnotation typeScenario
Single-soundMixed-soundMulti-entityOff-screen
Flickr-SoundNet [29]5,0005,0005020.0sbboxImage✓XXX
VGGSS [7]5,1585,15822110.0sbboxImage✓XXX
Epic Sound Object [14]3,1729,196301.0sbboxImage✓XXX
Extended Flickr-SoundNet [20]2922925015.4sbboxImageXXX✓
Extended VGGSS [20]5,5375,53722110.0sbboxImageXXX✓
IS3 [31]3,2403,24011810.0sbbox & instance segm.ImageX✓XX
$AVSBench [36]8044,020235.0ssegm.Chunk✓✓XX
$AVSBench-Semantic [37]1,55411,520707.4ssemantic segm.Chunk✓✓XX
AVATAR (Ours)5,00024,2668010.0sbbox & instance segm.Video✓✓✓✓
+ +Furthermore, we introduce Temporal-aware Audio-Visual Localization model for fine-grained video understanding (TAVLO), a novel AVL model that explicitly incorporates temporal information. Experimental results demonstrate that existing methods struggle with tracking temporal changes, as they rely on global audio features and static frame mappings. In contrast, TAVLO is robust to temporal variations, leveraging high-resolution temporal information for precise audio-visual alignment. Our model effectively localizes sound sources even in complex environments, demonstrating significant improvements over prior approaches. + +# 2. Related Work + +# 2.1. Audio-Visual Localization + +AVL has primarily been studied through self-supervised learning, leveraging the natural co-occurrence of auditory and visual signals. Various approaches have been proposed to enhance AVL performance [12, 30, 32]. EZ-VSL [21] formulates AVL as a multiple instance learning [19] problem, introducing a cross-modal multiple-instance learning loss to improve training. SSL-TIE [18] enhances self-supervised localization by incorporating transformation invariance and equivariance. ACL-SSL [24] integrates CLIP [27] to explore multimodal learning in AVL. To address the limitations of purely self-supervised methods, DMT [10] employs semi-supervised learning to improve generalization, while SLAVC [20] explores weakly supervised strategies for precise localization with limited annotations. Beyond single-source scenarios, several studies have extended AVL to multi-source environments [13, 15, 22, 26], broadening its applicability to more complex auditory scenes. + +# 2.2. Audio-Visual Localization Benchmarks + +The most widely used datasets, Flickr-SoundNet [29] and VGGSS [7], are derived from SoundNet [3] and VG- + +GSound [5], respectively. In the egocentric domain, Epic Sound Object [14] introduces an AVL benchmark that considers viewpoint variations and occlusion challenges arising from wearer-environment interactions. Other benchmarks explore misalignment scenarios and multi-source settings. Extended datasets [20] are designed to evaluate model performance under audio-visual misalignment by generating synthetic negative samples through cross-pairing audio and video frames from different sources. IS3 [31] focuses on mixed-sound environments, combining VGGSound audio and generating synthetic images using Stable Diffusion. However, these benchmarks primarily provide image-level labels for a single frame (or up to three frames in Epic Sound Object), limiting their ability to assess temporal dynamics. To address this limitation, AVSBench [36] and AVSBench-Semantic [37] introduce the Audio-Visual Segmentation task, highlighting the inability of conventional AVL benchmarks to capture the shape and extent of sound-emitting objects. AVSBench annotates one-second clips with segmentation masks for visible sounding objects. However, this clip-based annotation approach remains restricted, as each clip is still represented by a single frame. Tab. 1 summarizes major AVL benchmarks. + +# 3. Benchmark: AVATAR + +AVATAR addresses the limitations of existing benchmarks by adopting a video-centric approach, providing a more precise evaluation framework that reflects real-world complexity. To achieve this, Section 3.1 introduces a semi-automatic annotation pipeline, which maintains high temporal resolution while reducing labeling costs and ensuring annotation quality. Subsequently, Section 3.2 defines the four key evaluation scenarios considered in AVATAR and discusses the essential factors for model assessment in each setting. + +# 3.1. Semi-Automatic Annotation Pipeline + +Annotating sound-emitting instances with gold-standard labels is labor-intensive and costly due to the need for hu- + +![](images/d448b35e87684bef0283cef4da77f512b07a0b414643e40222fb5d5535eb2133.jpg) +Figure 1. Model-Driven Labeling and Verification Process. CAV-MAE [9] and YoloV8 [28] generate candidate bounding boxes with audio class predictions, refined through Audio-guided Bounding box Filtering. Human verification filters the regions before SAM [16] performs instance segmentation, followed by a final verification for accuracy. + +man intervention. To address this, the semi-automatic annotation pipeline integrates deep learning models to minimize manual effort while maintaining high annotation quality. The pipeline consists of three stages. First, Sec. 3.1.1 describes the selection process for candidate raw videos. Next, Sec. 3.1.2 introduces an automated sampling strategy to identify key video clips and frames that require annotation, ensuring efficiency while mitigating sampling bias. This step involves two phases: (i) Clip Sampling, which extracts meaningful video segments, and (ii) Frame Sampling, which selects key frames for annotation. Finally, Sec. 3.1.3 details a model-assisted labeling process, where off-the-shelf models generate initial annotations, allowing human annotators to focus on verification and refinement. Fig. 1 provides an overview of the process. + +# 3.1.1. Obtaining Candidate Videos + +The dataset is constructed using VGGSound [5], a large-scale video dataset sourced from YouTube under the Creative Commons Attribution 4.0 International License. Using YouTube API, we retrieve the original videos based on their YouTube IDs, collecting approximately 70k raw videos. To ensure quality and consistency, we apply a filtering process with constraints on resolution $(640\times 360)$ , frame rate (20-40 fps), duration $(\geq 10s)$ , and bitrate $(\geq 100$ bps), resulting in 39k candidate videos for annotation. + +# 3.1.2. Automatic Clip and Frame Sampling + +Clip Sampling This stage extracts video segments containing meaningful audio. The Root Mean Square (RMS) energy of the audio signal is computed to distinguish between silence and audio-active regions, with a threshold of 0.01 applied at 1s intervals. Using a sliding window approach, we identify 10s segments where audio activity occurs at least five times, selecting them as candidate clips. The final annotation clips are then randomly sampled from these candidates. + +Frame Sampling Once a clip is selected, this step determines the frames to be labeled. We define audio events based on RMS energy peaks within a $\pm 0.1$ s window. Given the filtering constraints in Sec. 3.1.1, each 0.2s audio event contains at least four frames. To minimize motion blur, we select the sharpest frame by applying a Laplacian filter, a high-pass filter that enhances edge contrast. Additionally, up to five frames per clip are sampled based on audio event intensity, ensuring a balanced temporal distribution by excluding frames from overlapping events within $\pm 0.7$ s. + +# 3.1.3. Model-Driven Labeling and Verification + +Target Category Selection To effectively utilize off-the-shelf models, we first determine the target categories for annotation. We analyze existing audio-visual datasets, including VGGSound [5] (300 classes), OpenImageV7 [17] (600 classes), and AudioSet [8] (527 classes). Categories where the sound source is ambiguous (e.g., people crowd, wind noise) are excluded. Ultimately, we select 80 target categories covering a broad range of real-world domains. + +Audio-Guided Bounding-box Annotation We employ YoloV8 [28], trained on OpenImageV7 [17], for object detection, and CAV-MAE [9], trained on AudioSet [8], for audio classification. First, among the bounding boxes created by YoloV8, instances that did not correspond to the target categories are filtered out. Next, CAV-MAE's classification results guide the selection of bounding boxes associated with active sound sources, implementing an Audio-Guided Bounding Box Filtering strategy. Human annotators verify and filter these bounding boxes by reviewing the corresponding video segment ( $\pm 0.05$ s around the frame) to ensure that the labeled instance is indeed producing sound at that moment. + +Model-Assisted Instance Segmentation Verified bounding boxes are used as prompts for SAM [16] to generate instance segmentation masks automatically. Human verifica + +![](images/6148792d820e158109c64ea6a11f3962878b8dbef57d3ad6e569a4138e0fb6b0.jpg) +Figure 2. Four Scenarios and Cross-event Subset Examples. (a) Single-sound: A single source (e.g., a barking dog). (b) Mixed-sound: Overlapping sources (e.g., flute and clarinet). (c, d) Multi-entity: Multiple distinct sources, such as an interview in a crowd (c) or multiple drums (d). (e) Off-screen: The source is outside the frame (e.g., a speaker behind the camera). (f) Cross-event: The active sound source changes over time. + +![](images/e10dac9b4e85f77f5cace5378385ab821247d8e720db7bd9b864bae40b4cd456.jpg) + +tion ensures segmentation accuracy, and only in cases where the mask was not well generated, it was supplemented with human annotation under the guidance of SAM. + +# 3.2. Scenario Definitions + +A key distinction of our benchmark is the introduction of instance-level scenario definitions, enabling fine-grained evaluation of AVL models under diverse real-world conditions. We define four distinct scenarios, each addressing a specific challenge in audio-visual localization. Visual illustrations are provided in Fig. 2. + +Single-sound represents the simplest case, where only one instance within the frame emits a clear and distinct audio signal. By eliminating complex auditory interference, it evaluates a model's ability to learn and localize a one-to-one audio-visual correspondence, serving as a baseline before progressing to more complex settings. + +Mixed-sound involves multiple concurrent audio sources, requiring the model to discriminate and associate sounds with their correct visual sources. It includes three variations: (1) multiple instances of the same category, (2) instances from different categories, and (3) partially off-screen sound sources. This setting tests the model's ability to perform auditory scene separation and resolve multiple simultaneous audio-visual associations. + +Multi-entity challenges the model to identify the actual sound-emitting instance among multiple visually similar objects, as first introduced in AVATAR. For example, in a scene with multiple people, only one may be speaking, requiring the model to pinpoint the correct speaker. Unlike conventional AVL approaches that rely solely on single-frame associations, this scenario necessitates spatiotemporal reasoning to distinguish active from passive entities within the same visual category. + +Off-screen Unlike cameras with a limited field of view, microphones record omnidirectional audio, often including off-screen sound sources. This scenario evaluates a model's + +ability to avoid false positives when no visible sound-emitting object is present, serving as a robustness check to ensure accurate localization only within the visual frame. + +# 4. Approach: TAVLO + +Existing AVL studies primarily focus on audio-visual associations, overlooking temporal dynamics in video. To address this limitation, we propose a spatiotemporal AVL model that effectively integrates spatial and temporal information. The overall architecture is illustrated in Fig. 3. + +Attention mechanisms have achieved state-of-the-art performance in both NLP and Vision by enabling adaptive feature selection and capturing long-range dependencies. Motivated by this, we employ attention-based modeling to effectively integrate audio-visual cues and localize the sound-emitting object at each time step. However, the quadratic complexity of self-attention poses a significant computational challenge when applied directly to flattened video features. To overcome this, we adopt a factorized attention strategy [2, 4, 34] and design the Audio-Spatial-Temporal (AST) Attention Block, enabling efficient processing of audio-visual sequences. + +This section details our approach. Sec. 4.1 introduces modality-specific feature encoding for time-aware audiovisual localization. Sec. 4.2 describes the AST Block architecture, and Sec. 4.3 discusses the training objective. + +# 4.1. Modality-Specific Feature Encoding + +A given video $X$ consists of visual frames and audio, represented as $V \in \mathbb{R}^{T \times H_v \times W_v \times 3}$ and $A \in \mathbb{R}^{H_a \times W_a \times 1}$ , respectively. Here, $V$ corresponds to a sequence of RGB frames, while $A$ represents the spectrogram of the raw audio waveform. $T$ denotes the total number of frames in the video. $H_v, W_v$ denote the height and width of video frames, while $H_a, W_a$ represent the frequency and time dimensions of the spectrogram. + +Conventional AVL models typically adopt a two-stream neural network encoder, extracting global audio and single-frame visual features independently. However, this ap + +![](images/4d57e05de81071fb6adcc195a5c56a5dce4bcade36ab91a48d2d4a78abfa46be.jpg) +Figure 3. Overview of Our Approach. The audio and visual encoders extract features, applying positional encoding in a modality-specific manner. The AST Attention Block then applies stacked spatial and temporal attention layers, repeated $L$ times. + +proach fails to capture temporal dynamics. To remedy this, we retain the two-stream encoder structure but introduce a time-specific encoding strategy that encodes modality-specific features at each time step. + +Feature Extraction To encode visual and audio features, we employ a visual encoder $f_{v}$ and an audio encoder $f_{a}$ . For visual feature extraction, we use ResNet-18 [11] as $f_{v}$ , following prior works [20, 21, 30, 32]. Given $T$ frames as input, the encoded visual feature $\mathbf{V}$ is obtained as: + +$$ +\mathbf {V} = f _ {v} (V) \in \mathbb {R} ^ {T \times H \times W \times D _ {f}}, \tag {1} +$$ + +where $H$ and $W$ represent the spatial resolution of the downsampled feature map, and $D_{f}$ is the feature dimension. + +For audio feature extraction, we design $f_{a}$ using a rectangular 2D CNN kernel, ensuring that each audio segment aligns with its corresponding visual frame. The kernel size is defined as: + +$$ +K _ {w} = \left\lfloor \frac {W _ {a}}{T} \right\rfloor , \quad K _ {h} = H _ {a}, \tag {2} +$$ + +where $K_{w}$ adjusts the receptive field to match the $T$ -frame partitioning of the spectrogram. The audio feature $\mathbf{A}$ is then obtained as: + +$$ +\mathbf {A} = f _ {a} (A) \in \mathbb {R} ^ {T \times D _ {f}}. \tag {3} +$$ + +Positional Encoding Since self-attention is permutation-invariant, incorporating positional encoding is essential to preserve temporal and spatial structures. However, $\mathbf{A}$ contains only temporal information, whereas $\mathbf{V}$ retains both spatial and temporal dimensions. + +To ensure consistent encoding across modalities, we define spatial positional encoding $\mathrm{Pos}_{\mathrm{s}} \in \mathbb{R}^{T \times H \times W \times D_{s}}$ and temporal positional encoding $\mathrm{Pos}_{\mathrm{t}} \in \mathbb{R}^{T \times D_{t}}$ . Here, $D_{s}$ and $D_{t}$ denote the dimensions of spatial and temporal positional encodings, respectively, and $D_{s} = D_{f}$ ensures compatibility with visual features $\mathbf{V}$ . The final modality-specific fea + +ture representations are computed as: + +$$ +\tilde {\mathbf {V}} = \left[ \mathbf {V} + \operatorname {P o s} _ {s}; \operatorname {P o s} _ {t} \right] \in \mathbb {R} ^ {T \times H \times W \times D}, \tag {4} +$$ + +$$ +\tilde {\mathbf {A}} = \left[ \mathbf {A}; \operatorname {P o s} _ {\mathrm {t}} \right] \in \mathbb {R} ^ {T \times D}. \tag {5} +$$ + +where spatial positional encoding is added element-wise to visual features, while temporal encoding is concatenated across both modalities. As a result, the final dimensions of $\tilde{\mathbf{V}}$ and $\tilde{\mathbf{A}}$ are $D = D_{f} + D_{t}$ , ensuring that both modalities are encoded in a spatiotemporally consistent representation. + +# 4.2. Audio-Spatial-Temporal Attention Block + +To construct a 3D feature representation with spatial and temporal axes for the AST block's input, we first flatten the spatial dimensions $H \times W$ of $\tilde{\mathbf{V}}$ , ensuring that each times- $t$ retains $H \cdot W$ spatial positions. The processed visual feature is then concatenated with the audio feature: + +$$ +\mathbf {Z} ^ {0} = \left[ \tilde {\mathbf {A}}; \tilde {\mathbf {V}} \right] \in \mathbb {R} ^ {T \times (1 + H \cdot W) \times D}. \tag {6} +$$ + +The AST Block consists of two key components: Spatial and Temporal Attention. Spatial Attention learns fine-grained cross-modal interactions between audio and visual features at each timestamp, ensuring a precise spatial alignment between the two modalities. Following this, Temporal Attention captures long-range dependencies along the time axis, enhancing the model's ability to track temporal dynamics and maintain consistency across frames. The operations of the AST Block at layer $l$ are formulated as follows: + +$$ +\mathbf {Y} ^ {l - 1} = \text {S p a t i a l A t t e n t i o n} \left(\mathbf {Z} ^ {l - 1}\right), \tag {7} +$$ + +$$ +\mathbf {Z} ^ {l} = \text {T e m p o r a l A t t e n t i o n} \left(\mathbf {Y} ^ {l - 1}\right). \tag {8} +$$ + +Both Spatial and Temporal Attention are implemented using Multi-Head Self-Attention (MSA), with sequence lengths defined differently for each type. Spatial Attention applies self-attention across the $1 + H \cdot W$ dimension, capturing interactions between audio and visual features within a single frame. In contrast, Temporal Attention operates along the $T$ dimension, modeling dependencies over time. This structure allows each timestamp to encode both spatial and temporal relationships efficiently. As a result, these attentions can be implemented by simply adding a transpose operation at the end of the standard MSA computation. The detailed formulation of Spatial Attention is as follows: + +$$ +\mathbf {Y} _ {s} ^ {l - 1} = \operatorname {M S A} _ {\text {s p a t i a l}} \left(\operatorname {L N} \left(\mathbf {Z} ^ {l - 1}\right)\right) + \mathbf {Z} ^ {l - 1}, \tag {9} +$$ + +$$ +\mathbf {Y} _ {s ^ {\prime}} ^ {l - 1} = \operatorname {F F N} _ {\text {s p a t i a l}} \left(\operatorname {L N} \left(\mathbf {Y} _ {s} ^ {l - 1}\right)\right) + \mathbf {Y} _ {s} ^ {l - 1}, \tag {10} +$$ + +$$ +\mathbf {Y} ^ {l - 1} = \operatorname {T r a n s p o s e} \left(\mathbf {Y} _ {s ^ {\prime}} ^ {l - 1}, \dim = (1, 0)\right). \tag {11} +$$ + +# 4.3. Training Objective + +Our approach builds upon the cross-modal multiple-instance contrastive learning loss introduced in EZ-VSL [21], but incorporates two key modifications. First, we + +introduce a temporal component to explicitly model time-dependent relationships. Second, we refine the optimization goal for negative bags to enhance training stability. + +Given an AST Block with depth $L$ , the final output $\mathbf{Z}^L \in \mathbb{R}^{T \times (1 + H \cdot W) \times D}$ is decomposed into an audio representation $\hat{\mathbf{A}} \in \mathbb{R}^{T \times D}$ and a visual representation $\hat{\mathbf{V}} \in \mathbb{R}^{T \times H \times W \times D}$ . The flattened visual features are restored to their original spatial structure $H \times W$ . + +Unlike EZ-VSL, which constructs a bag of visual features corresponding to a global audio representation, our approach incorporates temporal information by redefining visual bags at the frame level. Specifically, for a given times- tamp $t$ in video $X$ , the audio segmentation representation $\hat{\mathbf{A}}^t$ is associated with a bag of visual features comprising all spatial locations within the corresponding visual representation $\hat{\mathbf{V}}^t = \{\hat{\mathbf{v}}_{xy}^t:\forall x\in [1,H],y\in [1,W]\}$ . + +To enforce alignment, $\hat{\mathbf{A}}^t$ should exhibit high similarity with at least one instance in its positive bag, while maintaining low similarity across all instances in negative bags. We define the negative bags of $\hat{\mathbf{A}}_i^t$ from video $X_{i}$ as the bag of visual features from the same timestamp $t$ in other videos $X_{j}$ within the batch $B$ , where $j\neq i$ . + +The positive and negative responses for the final optimization objective are defined as follows: + +$$ +\mathrm {p} _ {i} ^ {t} = \max _ {\hat {\mathbf {v}} \in \hat {\mathbf {V}} _ {i} ^ {t}} \langle \hat {\mathbf {A}} _ {i} ^ {t}, \hat {\mathbf {v}} \rangle , \tag {12} +$$ + +$$ +\mathrm {n} _ {i j} ^ {t} = \frac {1}{H W} \sum_ {\hat {\mathbf {v}} \in \hat {\mathbf {V}} _ {j} ^ {t}} \left\langle \hat {\mathbf {A}} _ {i} ^ {t}, \hat {\mathbf {v}} \right\rangle , \quad \forall j \neq i. \tag {13} +$$ + +Here, $\langle \cdot ,\cdot \rangle$ denotes the cosine similarity operation, which measures the similarity between two vectors. + +While EZ-VSL maximizes similarity for both positive and negative bags, we instead compute the mean similarity for negative bags. This prevents the loss function from being dominated by noisy or outlier instances, leading to a more stable optimization process that better reflects the overall distribution of negative samples. The audio-to-visual alignment loss $\mathcal{L}_{a\rightarrow v}$ is formulated as: + +$$ +\mathcal {L} _ {a \rightarrow v} = - \mathbb {E} _ {t, i} \left[ \log \frac {\exp \left(\mathrm {p} _ {i} ^ {t}\right)}{\exp \left(\mathrm {p} _ {i} ^ {t}\right) + \sum_ {j \neq i} ^ {B} \exp \left(\mathrm {n} _ {i j} ^ {t}\right)} \right]. \tag {14} +$$ + +The final training objective combines both audio-to-visual $(\mathcal{L}_{a\to v})$ and visual-to-audio $(\mathcal{L}_{v\rightarrow a})$ alignments, where $\mathcal{L}_{v\rightarrow a}$ is symmetrically defined using negative response $\mathrm{n}_{ji}^{t}$ : + +$$ +\mathcal {L} = \mathcal {L} _ {a \rightarrow v} + \mathcal {L} _ {v \rightarrow a}. \tag {15} +$$ + +During inference, direct vector similarity is used to generate an audio-visual localization map, computed as: + +$$ +\mathrm {s} _ {i, x y} ^ {t} = \langle \hat {\mathbf {A}} _ {i} ^ {t}, \hat {\mathbf {v}} _ {i, x y} ^ {t} \rangle . \tag {16} +$$ + +# 5. Experiments + +# 5.1. Experimental Setup + +Cross-event To assess the model's adaptability to dynamic real-world situations, we introduce the Cross-event subset, where the sound source changes over time. This setting evaluates whether a model can accurately track shifting audio sources. Cross-event videos are selected from AVATAR by identifying instances where a new audio-visual category emerges that was absent in earlier frames. Examples are shown in Fig. 2. + +Training Data The training dataset is based on VGGSound [5]. To prevent data leakage, we identified and removed 1,249 overlapping videos between AVATAR and VGGSound's training set. Additionally, as factors like fps can influence model training and inference settings, we selected videos with 200-300 frames to align with AVATAR's criteria. From a pool of 40k videos, a subset of 10k was randomly sampled for training. + +Baselines We evaluate three baseline models trained on VGGSound: EZ-VSL [21], SLAVC [20], and SSL-TIE [18]. These models perform inference on individual audio-image pairs, and we follow the same setting for evaluation. However, as they do not inherently model temporal dynamics, direct comparison with our method may not be entirely fair. To address this, we conduct additional experiments using varying audio window lengths (10s, 5s, 2s, 1s). Results indicate that shorter audio windows lead to a sharp performance drop or remain stable but degrade significantly in Cross-event. + +Evaluation Metrics For quantitative evaluation, we adopt Consensus Intersection over Union (CIoU) and Area Under Curve (AUC) as the primary metrics, following previous studies [18, 20, 21]. Additionally, for Off-screen scenario, we use the pixel-level True Negative (TN) percentage. Prior studies applied frame-level min-max normalization and used a heatmap threshold of 0.5 for evaluation. However, this normalization ensures that the maximum value within each frame is always 1, inevitably resulting in regions exceeding the threshold. This assumption implicitly presumes that a sound-emitting object is always present within the frame, contradicting the Off-screen scenario. To address this, we set the threshold for each model based on the top $10\%$ of pixel values across all labeled frames. + +# 5.2. Experimental Results + +# 5.2.1. Analysis of Cross-event Videos + +To assess the model's adaptability to dynamic real-world environments, we evaluate its performance in the Crossover setting, where the sound-emitting object changes over time. This conditions tests whether AVL models can effectively localize temporally varying audio sources. + +Table 2. Performance Comparison Across Scenarios. Quantitative results of various methods across different audio-visual localization scenarios: Single-sound, Mixed-sound, Multi-entity, and Off-screen. + +
Method(1) Single-sound(2) Mixed-sound(3) Multi-entity(4) Off-screen
CIoU(%)↑AUC(%)↑CIoU(%)↑AUC(%)↑CIoU(%)↑AUC(%)↑TN(%)↑\( ^\dagger \mathrm{{TN}}\left( \% \right) \uparrow \)
SLAVC(144k) [20]9.0710.606.317.886.417.9696.4695.75
EZ-VSL(10k) [21]9.6611.078.169.356.878.3296.9195.58
EZ-VSL(144k) [21]10.9212.226.978.345.807.4296.4796.45
EZ-VSL(full) [21]12.1713.387.678.916.968.4095.4395.84
SSL-TIE(144k) [18]13.1014.235.196.765.507.1290.8294.55
TAVLO(10k)13.4214.0814.1314.5212.0812.6991.1895.02
+ +${}^{ \dagger }$ TN(%) evaluated with a heatmap threshold,defined as the top $5\%$ highest pixel values,on videos that include off-screen scenarios. + +Table 3. CIoU Performance on Cross-event Scenarios. Comparison of CIoU(%) on the full AVATAR (Total) and Cross-event, with $\Delta$ indicating the performance drop. + +
MethodTotalCross-event
CIoU(%)↑CIoU(%)↑Δ
SLAVC(144k) [20]8.124.78-3.34
EZ-VSL(10k) [21]8.955.08-3.87
EZ-VSL(144k) [21]9.385.19-4.19
EZ-VSL(full) [21]10.505.26-5.24
SSL-TIE(144k) [18]10.395.03-5.36
TAVLO(10k)13.3713.04-0.33
+ +Results in Tab. 3 and Tab. 4 demonstrate a notable performance drop in baseline models on the Cross-event subset compared to the overall benchmark (Total). Specifically, existing methods exhibit declines of up to $-5.36$ CIOU and $-5.00$ AUC, indicating their inability to adapt to shifting audio sources. This limitation stems from their reliance on static mappings between global audio and single-frame visual representations, preventing effective tracking of temporally evolving sound sources. As a result, these models struggle to localize sound sources accurately when they change over time. + +In contrast, by explicitly incorporating temporal information, our model enables more effective tracking of dynamic sound sources. The results confirm this advantage, as our model exhibits only a minimal performance drop $(\Delta = -0.33$ for CIoU, $\Delta = -0.37$ for AUC) in the Cross-event setting. This demonstrates that, rather than relying on static mappings, our model leverages temporal context, maintaining stable localization performance even in dynamic environments. + +# 5.2.2. Scenario-wise Performance Evaluation + +To provide a more comprehensive assessment beyond conventional evaluation metrics, we conduct a scenario-based analysis to assess localization accuracy under diverse real-world conditions. Performance is evaluated at the frame level, considering only instances relevant to each scenario, as summarized in Tab. 2. + +In the Single-sound scenario, which represents the fun + +Table 4. AUC Performance on Cross-event Scenarios. Comparison of AUC(\%) on the full AVATAR (Total) and Cross-event, with $\Delta$ indicating the performance drop. + +
MethodTotalCross Event
AUC(%)↑AUC(%)↑Δ
SLAVC(144k) [20]9.676.66-3.07
EZ-VSL(10k) [21]10.336.73-3.60
EZ-VSL(144k) [21]10.746.86-3.88
EZ-VSL(full) [21]11.757.04-4.71
SSL-TIE(144k) [18]11.686.68-5.00
TAVLO(10k)13.9813.61-0.37
+ +damental AVL task with a single sound-emitting instance, most models achieve high performance. Notably, SSLTIE (144k) marginally outperforms our model by $0.15\%$ in AUC. However, apart from this case, our model surpasses all other baselines, achieving $13.42\%$ CIoU and $14.08\%$ AUC, demonstrating superior localization precision in single-source environments. + +The Mixed-sound and Multi-entity scenarios introduce greater auditory and visual complexity, leading to significant performance degradation in baseline models. In contrast, our model achieves $14.13\%$ CIoU and $14.52\%$ AUC in the Mixed-Sound scenario, and $12.08\%$ and $12.69\%$ in the Multi-entity scenario, outperforming all baselines. Notably, despite not being explicitly designed for these conditions, our model maintains a performance similar to that in the Single-sound setting. This suggests that high-resolution temporal audio segment modeling, combined with a temporal-aware architecture leveraging motion cues, enhances its robustness. These results indicate that our approach effectively adapts to complex auditory environments and accurately differentiates multiple sound-emitting instances. + +In the Off-screen scenario, TAVLO exhibits a relatively low $\mathrm{TN}(\%)$ value. However, this can be attributed to the fact that other models frequently produce false positives in non-Off-screen frames due to the top $10\%$ pixel thresholding of the heatmap. This suggests that their reduced false positive rate in Off-screen cases may stem from lower overall performance in other scenarios. To examine this effect, we re + +![](images/afe8cc2e8ba582cae419f334727563608ecd1850f897a63af23e1267a4eb4802.jpg) +(a) Multi-entity + +![](images/d462df34c18e1f20bf2bfbfe35db74b2d86f6dff53d260c63a9778f9247260c5.jpg) +(b) Cross-event & Off-screen +Figure 4. Qualitative Comparison. (a) Multi-entity: A drum performance scene where multiple drums are visible, but only the snare drum produces sound. (b) Cross-event & Off-screen: A scenario where a woman speaks off-screen before a child begins speaking on-screen at a specific moment (4th frame). + +evaluated performance using only videos with Off-screen frames, adjusting the heatmap threshold to the top $5\%$ of pixel values. The revised results $\dagger \mathrm{TN}(\%)$ indicate that most models perform comparably under this condition. + +# 5.3. Qualitative Results + +Fig. 4 compares audio-visual localization results from the proposed model and baseline methods in two challenging videos: (a) Multi-entity and (b) Cross-event & Off-screen. + +In Fig. 4(a), a drum performance scene features multiple drums, but only the snare drum produces sound. The key challenge is distinguishing the actual sound source from visually similar but silent objects. The proposed model consistently localizes the snare drum across all five frames, whereas baselines often highlight non-sounding drums, suggesting a reliance on visual similarity rather than true audio-visual correspondence. This demonstrates the proposed model's ability to learn precise cross-modal associations. + +Fig. 4(b) depicts a scene where an off-screen woman is speaking until a child within the frame begins speaking in the fourth frame. The challenge is to avoid false positives when the off-screen speaker is active and correctly localize the on-screen speaker at the right moment. The proposed model successfully adapts to this transition, accurately identifying the child while avoiding erroneous detections during off-screen speech. In contrast, baseline models, particularly EZ-VSL(144k), EZ-VSL(full), and SSL-TIE(144k), struggle due to their reliance on global audio features, which incorporate the child's speaking voice. As a result, these models often make incorrect predictions that do not align with the actual speaking moment. This highlights a fundamental limitation of existing models in capturing fine-grained temporal changes in sound localization. + +# 6. Conclusion + +This study highlights the critical role of temporal modeling in audio-visual localization and introduces a new evaluation standard for Video-centric AVL. While prior work has largely focused on static frame-level associations, our findings demonstrate that capturing temporal dynamics is essential for accurate and robust sound source localization. To this end, we propose a novel benchmark and model, providing a systematic framework for evaluating AVL performance in dynamic environments and emphasizing the importance of precise multimodal alignment over time. + +Experimental results reveal the limitations of static image-based approaches, underscoring the necessity of video-centric methodologies in audio-visual perception. Notably, integrating temporal context significantly enhances model performance, offering valuable insights for future research in multimodal learning, fine-grained object differentiation, and localization in complex acoustic scenes. + +Despite its contributions, this study has limitations. It assumes at least partial alignment between audio and visual elements, leaving the independent localization of off-screen sounds as an open challenge. Additionally, while the benchmark introduces diverse evaluation scenarios, it does not prescribe specific methods for optimizing performance in each case. Future research should address Video-centric AVL strategies for off-screen sounds and develop scenarios-specific learning frameworks. + +# Acknowledgements + +This work was supported by NRF (2021R1A2C3006659) and IITP grants (RS-2021-II211343, RS-2022-II220320), all funded by the Korean Government. + +# References + +[1] Triantafyllos Afouras, Andrew Owens, Joon Son Chung, and Andrew Zisserman. Self-supervised learning of audio-visual objects from video. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XVIII 16, pages 208–224. Springer, 2020. +[2] Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lucic, and Cordelia Schmid. Vivit: A video vision transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6836-6846, 2021. +[3] Yusuf Aytar, Carl Vondrick, and Antonio Torralba. Soundnet: Learning sound representations from unlabeled video. Advances in neural information processing systems, 29, 2016. +[4] Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is space-time attention all you need for video understanding? In ICML, page 4, 2021. +[5] Honglie Chen, Weidi Xie, Andrea Vedaldi, and Andrew Zisserman. Vggsound: A large-scale audio-visual dataset. In ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 721-725. IEEE, 2020. +[6] Honglie Chen, Weidi Xie, Triantafyllos Afouras, Arsha Nagrani, Andrea Vedaldi, and Andrew Zisserman. Localizing visual sounds the hard way. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16867-16876, 2021. +[7] H. et al. Chen. Localizing visual sounds the hard way. In CVPR, 2021. +[8] Jort F. Gemmeke, Daniel P. W. Ellis, Dylan Freedman, Aren Jansen, Wade Lawrence, R. Channing Moore, Manoj Plakal, and Marvin Ritter. Audio set: An ontology and human-labeled dataset for audio events. In 2017 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 776-780, 2017. +[9] Yuan Gong, Andrew Rouditchenko, Alexander H. Liu, David Harwath, Leonid Karlinsky, Hilde Kuehne, and James R. Glass. Contrastive audio-visual masked autoencoder. In The Eleventh International Conference on Learning Representations, 2023. +[10] Yuxin Guo, Shijie Ma, Hu Su, Zhiqing Wang, Yuhao Zhao, Wei Zou, Siyang Sun, and Yun Zheng. Dual mean-teacher: An unbiased semi-supervised framework for audio-visual source localization. Advances in Neural Information Processing Systems, 36:48639-48661, 2023. +[11] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. +[12] Di Hu, Feiping Nie, and Xuelong Li. Deep multimodal clustering for unsupervised audiovisual learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9248-9257, 2019. +[13] Xixi Hu, Ziyang Chen, and Andrew Owens. Mix and localize: Localizing sound sources in mixtures. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10483-10492, 2022. + +[14] Chao Huang, Yapeng Tian, Anurag Kumar, and Chenliang Xu. Egocentric audio-visual object localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22910–22921, 2023. +[15] Dongjin Kim, Sung Jin Um, Sangmin Lee, and Jung Uk Kim. Learning to visually localize sound sources from mixtures without prior source knowledge. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. +[16] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. +[17] Ivan Krasin, Tom Duerig, Neil Alldrin, Vittorio Ferrari, Sami Abu-El-Haija, Alina Kuznetsova, Hassan Rom, Jasper Uijlings, Stefan Popov, Shahab Kamali, Matteo Malloci, Jordi Pont-Tuset, Andreas Veit, Serge Belongie, Victor Gomes, Abhinav Gupta, Chen Sun, Gal Chechik, David Cai, Zheyun Feng, Dhyanesh Narayanan, and Kevin Murphy. Openimages: A public dataset for large-scale multi-label and multi-class image classification. Dataset available from https://storage.googleapis.com/openimages/web/index.html, 2017. +[18] Jinxiang Liu, Chen Ju, Weidi Xie, and Ya Zhang. Exploiting transformation invariance and equivariance for self-supervised sound localisation. In Proceedings of the 30th ACM International Conference on Multimedia, pages 3742-3753, 2022. +[19] Oded Maron and Tomas Lozano-Perez. A framework for multiple-instance learning. In Neural Information Processing Systems, 1997. +[20] Shentong Mo and Pedro Morgado. A closer look at weakly-supervised audio-visual source localization. In Advances in Neural Information Processing Systems, 2022. +[21] Shentong Mo and Pedro Morgado. Localizing visual sounds the easy way. In European Conference on Computer Vision, pages 218-234. Springer, 2022. +[22] Shentong Mo and Yapeng Tian. Audio-visual grouping network for sound localization from mixtures. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10565–10574, 2023. +[23] Pedro Morgado, Yi Li, and Nuno Nvasconcelos. Learning representations from audio-visual spatial alignment. Advances in Neural Information Processing Systems, 33:4733-4744, 2020. +[24] Sooyoung Park, Arda Senocak, and Joon Son Chung. Can clip help sound source localization? In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 5711-5720, 2024. +[25] Jie Pu, Yannis Panagakis, Stavros Petridis, and Maja Pantic. Audio-visual object localization and separation using low-rank and sparsity. In 2017 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 2901-2905, 2017. +[26] Rui Qian, Di Hu, Heinrich Dinkel, Mengyue Wu, Ning Xu, and Weiyao Lin. Multiple sound sources localization from + +coarse to fine. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XX 16, pages 292-308. Springer, 2020. +[27] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PmLR, 2021. +[28] Dillon Reis, Jordan Kupec, Jacqueline Hong, and Ahmad Daoudi. Real-time flying object detection with yolov8. arXiv preprint arXiv:2305.09972, 2023. +[29] Arda Senocak, Tae-Hyun Oh, Junsik Kim, Ming-Hsuan Yang, and In So Kweon. Learning to localize sound source in visual scenes. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. +[30] Arda Senocak, Hyeonggon Ryu, Junsik Kim, Tae-Hyun Oh, Hanspeter Pfister, and Joon Son Chung. Sound source localization is all about cross-modal alignment. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7777-7787, 2023. +[31] Arda Senocak, Hyeong Sun Ryu, Junsik Kim, Tae-Hyun Oh, Hanspeter Pfister, and Joon Son Chung. Aligning sight and sound: Advanced sound source localization through audio-visual alignment. ArXiv, abs/2407.13676, 2024. +[32] Weixuan Sun, Jiayi Zhang, Jianyuan Wang, Zheyuan Liu, Yiran Zhong, Tianpeng Feng, Yandong Guo, Yanhao Zhang, and Nick Barnes. Learning audio-visual source localization via false negative aware contrastive learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6420-6429, 2023. +[33] Yapeng Tian, Jing Shi, Bochen Li, Zhiyao Duan, and Chen-liang Xu. Audio-visual event localization in unconstrained videos. In Proceedings of the European conference on computer vision (ECCV), pages 247–263, 2018. +[34] Dirk Weissenborn, Oscar Täckström, and Jakob Uszkoreit. Scaling autoregressive video models. In International Conference on Learning Representations, 2020. +[35] Hang Zhao, Chuang Gan, Andrew Rouditchenko, Carl Vondrick, Josh McDermott, and Antonio Torralba. The sound of pixels. In Proceedings of the European conference on computer vision (ECCV), pages 570-586, 2018. +[36] Jinxing Zhou, Jianyuan Wang, Jiayi Zhang, Weixuan Sun, Jing Zhang, Stan Birchfield, Dan Guo, Lingpeng Kong, Meng Wang, and Yiran Zhong. Audio-visual segmentation. In European Conference on Computer Vision, 2022. +[37] Jinxing Zhou, Xuyang Shen, Jianyuan Wang, Jiayi Zhang, Weixuan Sun, Jing Zhang, Stan Birchfield, Dan Guo, Lingpeng Kong, Meng Wang, and Yiran Zhong. Audiovisual segmentation with semantics. arXiv preprint arXiv:2301.13190, 2023. \ No newline at end of file diff --git a/ICCV/2025/What's Making That Sound Right Now_ Video-centric Audio-Visual Localization/images.zip b/ICCV/2025/What's Making That Sound Right Now_ Video-centric Audio-Visual Localization/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..161d3619b3a6c46b119ee748373014a9970e46da --- /dev/null +++ b/ICCV/2025/What's Making That Sound Right Now_ Video-centric Audio-Visual Localization/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b0a2398c36e03b91a6b89faadcdf4b8c88d1795a77c5d13275a64f1cd3d5b1d9 +size 574550 diff --git a/ICCV/2025/What's Making That Sound Right Now_ Video-centric Audio-Visual Localization/layout.json b/ICCV/2025/What's Making That Sound Right Now_ Video-centric Audio-Visual Localization/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..00fb0c9b5341f80d0a89418c3a955731c044fb12 --- /dev/null +++ b/ICCV/2025/What's Making That Sound Right Now_ Video-centric Audio-Visual Localization/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1997ac6f5fd74e907ec726185f3cc7805f42336b4d743a7b2377c4156e3fbf2d +size 365817 diff --git a/ICCV/2025/What's in a Latent_ Leveraging Diffusion Latent Space for Domain Generalization/81c74318-d891-44f5-8e24-de8ec3e2c413_content_list.json b/ICCV/2025/What's in a Latent_ Leveraging Diffusion Latent Space for Domain Generalization/81c74318-d891-44f5-8e24-de8ec3e2c413_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..2a65b109f5275c970ef8638b5df8b3aca1e6db66 --- /dev/null +++ b/ICCV/2025/What's in a Latent_ Leveraging Diffusion Latent Space for Domain Generalization/81c74318-d891-44f5-8e24-de8ec3e2c413_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:68a920600fdcddf59a7f1ec6ef9de6fcd921f8d8ac56a08e66ca593294697c29 +size 92239 diff --git a/ICCV/2025/What's in a Latent_ Leveraging Diffusion Latent Space for Domain Generalization/81c74318-d891-44f5-8e24-de8ec3e2c413_model.json b/ICCV/2025/What's in a Latent_ Leveraging Diffusion Latent Space for Domain Generalization/81c74318-d891-44f5-8e24-de8ec3e2c413_model.json new file mode 100644 index 0000000000000000000000000000000000000000..00766956e6501f5adaed73a792936f3e3b047a17 --- /dev/null +++ b/ICCV/2025/What's in a Latent_ Leveraging Diffusion Latent Space for Domain Generalization/81c74318-d891-44f5-8e24-de8ec3e2c413_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a8fa4ba3aeed3d5793f95515befaa93239f25043d32182d4da1f56108d4cc9a9 +size 118218 diff --git a/ICCV/2025/What's in a Latent_ Leveraging Diffusion Latent Space for Domain Generalization/81c74318-d891-44f5-8e24-de8ec3e2c413_origin.pdf b/ICCV/2025/What's in a Latent_ Leveraging Diffusion Latent Space for Domain Generalization/81c74318-d891-44f5-8e24-de8ec3e2c413_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..f7f8dff22209b3f53fa6d800b53eb421638bc5f0 --- /dev/null +++ b/ICCV/2025/What's in a Latent_ Leveraging Diffusion Latent Space for Domain Generalization/81c74318-d891-44f5-8e24-de8ec3e2c413_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f1ff5665645fdd6fe3ecd54361ad90a908993f3ceecef853241baa89ebcef1cf +size 12264846 diff --git a/ICCV/2025/What's in a Latent_ Leveraging Diffusion Latent Space for Domain Generalization/full.md b/ICCV/2025/What's in a Latent_ Leveraging Diffusion Latent Space for Domain Generalization/full.md new file mode 100644 index 0000000000000000000000000000000000000000..95b52ac3f337302018e5c07d362984051ce47716 --- /dev/null +++ b/ICCV/2025/What's in a Latent_ Leveraging Diffusion Latent Space for Domain Generalization/full.md @@ -0,0 +1,318 @@ +# What's in a Latent? Leveraging Diffusion Latent Space for Domain Generalization + +Xavier Thomas $^{1}$ Deepti Ghadiyaram $^{12*}$ $^{1}$ Boston University $^{2}$ Runway + +# Abstract + +Domain Generalization aims to develop models that can generalize to novel and unseen data distributions. In this work, we study how model architectures and pre-training objectives impact feature richness and propose a method to effectively leverage them for domain generalization. Specifically, given a pre-trained feature space, we first discover latent domain structures, referred to as pseudo-domains, that capture domain-specific variations in an unsupervised manner. Next, we augment existing classifiers with these complementary pseudo-domain representations making them more amenable to diverse unseen test domains. We analyze how different pre-training feature spaces differ in the domain-specific variances they capture. Our empirical studies reveal that features from diffusion models excel at separating domains in the absence of explicit domain labels and capture nuanced domain-specific information. On 5 datasets, we show that our very simple framework improves generalization to unseen domains by a maximum test accuracy improvement of over $4\%$ compared to the standard baseline Empirical Risk Minimization (ERM). Crucially, our method outperforms most algorithms that access domain labels during training. Code is available at: link. + +# 1. Introduction + +It is now a common practice to use models pre-trained on billion-scale data [20, 22, 44, 49, 51, 55, 58] as defacto backbones for diverse downstream tasks [41, 67]. In order to make these large-scale models "foundational," and offer rich feature representations, a variety of powerful pre-training strategies have been designed. Some of these objectives aim to eliminate the need for clean labeled data [8, 9, 13, 21, 78], some reap the benefits from rich text representations by aligning them with corresponding visual signals [30, 55], while others force models to build a more meaningful understanding of scenes by learning to predict large hidden regions of images [22]. Despite such tremen + +![](images/f928dde6430913384b073501ca887e439750aa430d1f277199eaf9be9cf810af.jpg) +CLIP (least domain seperation) + +![](images/9d8330f132217bc8895217d2d4fa9adea339973ff4acf922142b89936ebe66eb.jpg) +DiT (high domain seperation) + +MAE +Domains in VLCS +Figure 1. t-SNE visualization of the latent space from different pretraining objectives: CLIP [55], DiT [51], MAE [22], ResNet-50 [20] on the domain generalization benchmark VLCS [17]. VLCS is curated from 4 different datasets, thus dataset-specific biases like spatial composition and object size variations serve as different domains. Note how the diffusion features separate domains effectively, suggesting that latent domain structures can be captured without explicit supervision. Best viewed in color. +![](images/546265f12af8da389f1026c02ee63890840b2cca449d9feceaffa0955425e02e.jpg) +Caltech101 LabelMe SUN09 VOC2007 + +![](images/bfea6c0d52ad713ee0d3f5fef5327d7010f75608158502741220186a1d6016de.jpg) +ResNet-50 + +dous progress, what exactly is captured in the underlying latent landscape remains an open question. This question becomes more challenging in diffusion models mainly due to their iterative global denoising objective. + +This work aims to understand the feature landscape learnt from different pre-training models and objectives in the context of domain generalization. Robust generalization to unseen domains has been a long-standing goal in machine learning research [6, 46], particularly in scenarios where collecting domain-specific data is infeasible or expensive. In such cases, models must learn to generalize without relying on explicit domain labels even during training [38]. It has been established that most sophisticated models struggle when the test data distribution differs from that of training data [57, 63, 65] even in subtle ways, e.g., same visual scene but captured using different cameras, same patient but different brand imaging devices, same object but captured in different color schemes and so on. + +We posit that the first step to make fundamental progress + +towards designing foundational models is to examine and interpret how current state-of-the-art models structure visual information and uncover their strengths and limitations. For instance, how are object, scene, and domain-specific variations internally encoded in a latent space? Do domain-specific traits manifest in distinct regions of the latent space or are they engulfed along with low- to mid-level scene and object level information? + +We study these questions in great detail in this work. Specific to the task of domain generalization, we analyze how different pre-training objectives and architectures influence the granularity of visual information captured in their feature space. Our key insight is that certain internal states of diffusion models effectively capture abstract information such as photographic styles, camera angles, and so on. Building on this insight, we first develop an unsupervised method for discovering latent domain structures. Next, we alter a standard domain generalization classification [69] pipeline with one key difference: we augment the classifier's representations with the discovered latent domain representations. We show through extensive empirical analysis that this simple tweak to the standard pipeline assists in training a model that generalizes well to unseen domains [5]. While most prior works focus on leveraging a single feature space to design a universal model [11, 18, 42, 62, 64], we take an alternative approach and compliment existing classifier's features with domain rich features and show that this auxiliary guidance makes the overall feature space more robust to unseen domains. Our framework dubbed GUIDE: Generalization using Inferred Domains from Latent Embeddings, offers a simple and effective method to "guide" a given feature space to adapt better to unseen domains. We summarize our key contributions below: + +- We propose a method of unsupervised pseudodomain discovery from frozen pre-trained feature spaces and use them to improve any model's ability to generalize to diverse domains, making it particularly useful in scenarios where domain labels used during training are unavailable or noisy (Sec. 3.3). +- We analyze different pre-training objectives and architectures and investigate how they influence the structure of the feature latent landscape of both diffusion and conventional vision models (Sec. 4.3). +- We shine light on the ability of diffusion models to capture domain-specific information, such as photographic and artistic styles, texture variations, and demonstrate their effectiveness to domain generalization (Sec. 4.4). We obtain an average test accuracy improvement of $+2.6\%$ on 5 datasets, notably beating ERM [69] by $+4.3\%$ on the TerraIncognita dataset [4]. + +# 2. Related Work + +Diffusion features for representation learning: Diffusion models [26, 60] have significantly advanced image and video generation, prompting extensive exploration of their intermediate representations and their utility for diverse downstream tasks such as detection [12], segmentation [3, 74], classification [34], semantic correspondence [43], depth estimation [73, 81], and visual reasoning [72], showcasing their utility in both discriminative and generative domains. Recent studies [32, 43, 71] demonstrate that features extracted across layers and timesteps encode rich semantic information, ranging from coarse patterns to fine-grained details. In this work, we analyze how the latent space of diffusion models captures class and domain-specific information and leverage these representations for the task of domain generalization. + +Domain generalization: First formalized in [6], domain generalization is the challenging task of designing models capable of generalizing to unseen test domains. Various methods have been proposed to address this by learning domain-agnostic representations [29, 47], data or latent augmentation methods [27, 39, 42, 61], and meta-learning [2, 7]. Despite numerous advancements, most methods still under perform Empirical Risk Minimization (ERM) when evaluated rigorously [19], making it a very strong baseline. Teterwak et al. [64] builds a stronger baseline by incorporating improved training strategies. Matsuura and Harada [45] learn a domain-invariant feature extractor by clustering samples into latent domains using style statistics from early convolutional layers, then applying adversarial learning to reduce domain distinctions. Bui et al. [7] uses meta-learning and explicit domain labels to disentangle domain-invariant and domain-specific features, ensuring that the latter remains useful when adapting to new domains. The classifier then integrates both feature types for improved generalization. Dubey et al. [16], Thomas et al. [66] explore techniques to incorporate pseudo-domain information into classifiers to make them generalizable to unseen domains. Our work differs from these prior arts in several crucial ways: we leverage pre-trained models instead of learning a separate domain prototype network as in [16], utilize a more domain-rich feature space compared to [66], and do not rely on domain labels as in [7, 16]. + +Diffusion models for domain generalization. Prior works [23, 24, 28, 79] use text-to-image diffusion models as a data augmentation tool by generating diverse synthetic samples with variations that help models generalize better to unseen domains. However, these techniques rely on finetuning the diffusion model, expensive data augmentation steps, or access to the test data. By contrast, to the best of our knowledge, we are the first to investigate using frozen pre-trained diffusion features in an unsupervised manner for domain generalization. + +# 3. Approach + +First, we introduce the preliminaries of diffusion models (Sec. 3.1) and domain generalization (Sec. 3.2). Then, we present our two-step framework where we first learn pseudo-domain representations in an unsupervised manner and use them to adapt a classifier to unseen domains (Sec. 3.3). We stress that we do not have domain label information during both training and test phases. + +# 3.1. Preliminaries on Diffusion Models + +Diffusion models [26, 60] are probabilistic generative models designed to learn the data distribution through an iterative denoising process. In the forward diffusion process, an image $x$ is incrementally corrupted with noise $(\epsilon)$ over $T$ timesteps, resulting in a sequence of increasingly noisy images $\{x_{t}\}_{t=1}^{T}$ . In the reverse process of iterative denoising, a model $\theta$ , predicts the added noise $\epsilon_{\theta}(x_{t},t)$ at each timestep $t$ . Latent Diffusion Models [58] (LDM) extend this framework by operating on a latent representation $z$ of the image $x$ instead of directly in its high-dimensional pixel space. This latent representation is obtained by mapping the image into a lower-dimensional space using a variational autoencoder [33] with an encoder $E$ and a decoder $D$ . The diffusion process models the distribution of these lower-dimensional latent embeddings, enabling more efficient computation. The training objective is: + +$$ +L _ {\mathrm {L D M}} = \mathbb {E} _ {E (x), t, \epsilon \sim \mathcal {N} (0, 1)} \| \epsilon - \epsilon_ {\theta} (z _ {t}, t) \| _ {2} ^ {2} +$$ + +# 3.2. Domain Generalization + +Let $X$ and $Y$ be random variables denoting input and target labels respectively, and $\Phi$ a feature extractor. In supervised learning, a predictor $f$ is learnt to map feature representations of inputs $x \in X$ , i.e., $\Phi(x)$ to labels $y \in Y$ , such that $f$ generalizes to unseen test samples. We denote this as $f(\Phi(x)) \to y$ . Domain generalization is an extension of supervised learning, where training data from multiple domains is available and the goal is to learn a predictor that performs well on samples from an unseen test domain [6]. + +As in a conventional domain generalization framework, each domain $d$ is characterized by a probability distribution $P_{d}$ defined over $X$ and $Y$ . The training dataset is constructed by sampling $d^{tr}$ domains, denoted as $\{P_d^{tr}\}_{d = 1}^{d^{tr}}$ , and collecting $n_d$ labeled points from each domain, forming the dataset $\bigcup_{d = 1}^{d^{tr}}\{(x_i^d,y_i^d)\}_{i = 1}^{n_d}$ . The unseen test domain distribution is denoted as $P_{d}^{te}$ , from which $n_T$ unlabeled points $\{x_i^{d^{te}}\}_{i = 1}^{n_T}$ are sampled during evaluation. + +One popular approach for domain generalization is to learn a universal classifier on all training samples [69] that is agnostic to the underlying domains. However, this algorithm makes a strong assumption that all training samples are drawn from a single, unified distribution and minimizes + +![](images/9440f27cc059242a1fb8924b5b6c462a2817b35c5e5510a2718c247d43e17ec2.jpg) +Figure 2. Training Pipeline. The green-shaded region represents the clustering and transformation step. Green solid arrows indicate gradient flow, while red arrows represent non-gradient operations. The feature extractor $\Psi$ first clusters samples to compute the pseudo-domain centroids. The transformation function $\mathcal{T}$ then transforms these centroids to the latent space of $\Phi$ , producing transformed pseudo-domain centroids, which are concatenated with the features from $\Phi$ , and sent to the classifier. + +the average risk across them. Though simple and effective, this may not guarantee good performance, especially when the test domain lies further from the assumed unified distribution or when the training domains themselves have a very high variance [16]. To address this, motivated by findings in [7, 16, 45, 66] which leverage domain-specific representations, we complement input features with these representations. We hypothesize that augmenting input features with rich, complementary information about (pseudo) domains would make the overall feature space more robust to diverse domain variations. + +Control experiment using ground truth domain labels: To validate the above hypothesis, we conduct the following control experiment. We assume access to ground truth domain labels, cluster diffusion features explicitly into each domain, and compute cluster centroids. Next, we augment the input features $(\Phi(x))$ by concatenating them with the cluster centroids and train a classifier on them. On a popular domain generalization benchmark OfficeHome [70], we achieve a boost of $3\%$ over the strongest baseline. We acknowledge that the number of pseudo-domains we learn per dataset in GUIDE (Sec. 3.3) is different from the true domains present in each dataset. Yet, this controlled setup highlights that augmenting a feature space with domain-specific representations from seen domains yields an overall generalizable feature space for unseen domains. + +Though the standard domain generalization framework assumes access to domain labels during training, in certain applications, this information may be unavailable or incorrect. Thus, we design a robust algorithm to learn this complementary "pseudo" domain information, described next. + +# 3.3. Adaptive Domain Generalization + +Learning pseudo-domain representations: In the absence of true domain labels, we adopt an unsupervised method called Kernel Mean Embeddings (KME) [6, 46] to capture key statistical properties of a domain. KMEs offer an efficient way to summarize and represent a probability dis + +
DatasetDomain Shift TypeExample
PACS [35]Style and texture variationsMonochromatic “Sketch”, color-rich “Art Painting”
VLCS [17]Dataset-specific biasesSpatial composition and object size variations (e.g., objects in “Caltech” are centered)
Terralnognitia [4] (TI)Environmental and background shiftsLocation-specific foliage and terrain variations
OfficeHome [70] (OH)Low-level style differences“clipart” has bold edges, “real” with softer, natural edges
DomainNet [52] (DN)Style, texture, and complexity differences“Quickdraw” has simplistic and often incomplete outlines, “Sketch” features more refined strokes with shading
Synth-ArtistsSynthetic artistic stylesVariations in artist techniques, color schemes, and brushwork
Synth-PhotographySynthetic photographic stylesChanges in lighting, contrast, focus
+ +The distribution of the feature vector in the training domain $\{P_d^{tr}\}_{d = 1}^{d^{tr}}$ is a function of the number of clusters and the number of features extracted from each cluster. We use the feature vector as a input to the training process, which is done by using the K-Means++ clustering algorithm. The training data is transformed into a single-dimensional feature vector. In our case, given the probability distributions of the training domains $\{P_d^{tr}\}_{d = 1}^{d^{tr}}$ , we use the feature extractor $\Psi$ to compute feature representations for samples drawn from each $P_{d}^{tr}$ . Then, we apply K-Means++ clustering and obtain $K$ clusters as a way to capture the underlying domain structures. Given we lack information about the true number or nature of the underlying domains during training in our setup, we refer to these clusters as pseudo domains. The centroid of each cluster $\hat{\Psi}_k$ , for $k\in K$ is used as the compact representation of each pseudo-domain. Finally, we assign each training sample $x$ to its nearest cluster, such that it's pseudo-domain feature representation is $\hat{\Psi}_x = \hat{\Psi}_k$ , the centroid of the corresponding pseudo domain. We study the impact of different feature extractors $\Psi$ in Sec. 4.1 and 4.3. We show how clustering smooths out any noise or sample-specific variations and creates more stable (pseudo) domain representations in Sec. 4.4. + +Leveraging pseudo-domain representations: We take inspiration from ERM [69] and learn a single universal classifier on all training domains, with one key difference: we augment each input feature vector with its corresponding pseudo-domain representation. Specifically, we first apply a transformation function on the pseudo-domain representations to bring the latent manifold of $\Psi$ closer to $\Phi$ to mitigate feature domain drift, i.e., $\mathcal{T}:\Psi \mapsto \Phi$ . Then, we concatenate the input feature vector $\Phi(x)$ with its corresponding pseudo-domain representation $\mathcal{T}(\widehat{\Psi}_k)$ during training, to learn a "domain-adaptive" classifier (as introduced in Dubey et al. [16]). At test time, we first process the input through $\Psi$ , then assign it to the nearest cluster centroid learned during training, and finally apply $\mathcal{T}$ before passing through the classifier. We stress that in our setup, we do not assume access to domain information during training and make no assumptions about the test domains. + +# 4. Experiments + +We outline the implementation details and training setup for GUIDE in Sec 4.1, followed by a detailed analysis of the capability of different feature extractors $(\Psi)$ in capturing domain-specific information to augment class-specific features $(\Phi)$ in Sec 4.3. We empirically show how our approach leads to a more domain generalizable classifier on unseen test domains in Sec. 4.4. + +Table 1. Overview of domain shifts in each dataset, including low-level and global photographic style variations, environmental, and dataset-specific biases. Example images for each dataset in suppl. material. + +
ModelSourceFeature Dimension
ResNet-50Global Average Pooling (GAP) at layer 492048
CLIP (ViT-L/14)CLS token1024
DINOv2 (ViT-L/14)Mean over patch tokens1024
MAE (ViT-L/14)Mean over patch tokens1024
SD-2.1Mean over channels of up_/_ft1 layer [32]1280
DiT-XL/2-512Mean over tokens of block 14 [32]1152
+ +Table 2. Feature extraction details from each model. SD-2.1 features are conditioned on an empty text prompt. + +# 4.1. Implementation Details + +Datasets: We conduct our experiments on 7 datasets, summarized in Table 1. Five of these datasets (PACS, VLCS, TerraIncognita, OfficeHome, DomainNet) are part of the DomainBed [19] test bed. We present details of Synth-Artists, and Synth-Photography in Sec. 4.5. + +Training Setup: We use the default hyper parameter settings from DomainBed [19]: a batch size of 32 per domain, learning rate of $5 \times 10^{-5}$ , number of steps as 5001, no dropout in the backbone model, and a weight decay of 0 on 1 A6000 GPU. We report test accuracies using the leave-one-domain-out cross-validation methodology [19], and average the results obtained over 3 trial seeds. + +Choice of $\Phi$ : We use ResNet-50 [20], initialized with AugMix [25] pre-trained weights as in DomainBed [19]. + +Choice of $\Psi$ : We study the feature spaces from several vision encoders with varied pre-training objectives: cross-entropy loss-based ResNet [20], contrastive loss-based CLIP [55], a distillation-based loss in DINOv2 [49], and reconstruction of masked patches loss-based MAE [22]. We further study two diffusion model architectures: the convolutional UNet-based [59] Stable Diffusion 2.1 (SD-2.1) [58] and transformer-based DiT-XL-2-512 (DiT) [51]. Though the underlying pre-training objective is the same for diffusion models, we aim to study the influence of the underlying diffusion architecture on the learnt feature landscape. We provide details on the layers from which the features are extracted in Table. 2. + +Choice of $\mathcal{T}$ and cluster refinement schedule: To adapt the pseudo-domain representations as $\Phi$ evolves during training, we define $\mathcal{T}$ (Sec 3.3) as a radial basis function (RBF) kernel ridge regressor (more in suppl. material). RBF kernels are well-known for their ability to model nonlinear, distance-based relationships and have been effectively used to align second-order statistics between source and target distributions [80]. In our approach, $\mathcal{T}$ maps the centroid $\widehat{\Psi}_k$ of a given pseudo-domain $k$ to the mean of the + +features $\Phi(x)$ of the samples belonging to cluster $k$ . We employ a logarithmic schedule [66] to periodically apply $\mathcal{T}$ on $\widehat{\Psi}_k$ , starting with frequent updates and progressively reducing their frequency and thus the overall computational overhead. We note that clustering is done only once on the static $\Psi$ -feature space, but the refinement follows a logarithmic schedule. + +Number of Pseudo-Domains: For GUIDE, the number of clusters $(K)$ is the sole hyper-parameter. We follow a simple heuristic from Thomas et al. [66] to determine this: $K = \max \left(\{1,3,5\} \times n_c,200\right)$ , where $n_c$ represents the number of classes in the dataset. The upper-bound of 200 clusters helps prevent over-clustering. The number of clusters that yields the best test accuracy for each domain is used to report the scores in Table 6. + +Evaluation of domain separability: With a motive to measure expressivity [16] of the underlying pseudo-domain representations, we measure normalized mutual information (NMI) as done in prior works [45, 66]. In our setup, let $U$ and $V$ be random variables that denote pseudo domain labels and ground truth domain (or class) labels. NMI is defined as: + +$$ +\operatorname {N M I} (U, V) = \frac {2 \cdot I (U , V)}{H (U) + H (V)}, +$$ + +where $I(U,V)$ is the mutual information between $U$ and $V$ and $H(U),H(V)$ their respective entropies. NMI measures how well the discovered clusters match the ground truth domain or class labels. In our setup, a feature space that yields clusters having high domain-NMI score is an ideal candidate to complement existing class-specific features. + +# 4.2. Underlying Domains in Each Dataset + +We begin by summarizing the types of domain shifts present in the datasets we study (described in Table 1). PACS [35] image dataset captures 7 object categories and 4 domains: real-world photos, art paintings, cartoons, and sketches. Thus, the domains have stark visual distinctions driven by both global and local changes such as shapes, colors, and edges. VLCS [17] is curated from different datasets, making dataset-specific biases such as spatial composition and object size variations as different domains. OfficeHome [70] similar to PACS also has images belonging to four domains: artistic, clip-art, product catalog, and realworld images. Thus, while there is some overlap in the underlying structural characteristics of the objects across domains, the domain shifts primarily involve style differences such as variations in texture, color, and outlines. TerraIncognita [4] consists of images taken from different camera trap locations, and each camera serves as a domain. Thus, the domain shifts are driven by physical environmental aspects such as variations in foliage density, terrain patterns, and spatial patterns of vegetation. DomainNet [52] is + +
DatasetDiTSD-2.1MAECLIPDINOv2RN50
PACS0.850.820.710.540.550.59
VLCS0.580.260.200.010.050.22
TerraInc0.220.550.210.010.010.25
OfficeHome0.250.280.100.120.380.08
DomainNet0.540.510.520.320.470.46
Synth-Artists0.890.860.750.250.340.63
Synth-Photography0.350.430.310.170.230.33
+ +Table 3. Comparison of domain NMI scores across datasets. The highest domain NMI score depends both on the type of pre-training feature space and the underlying domain shifts in the dataset as noted in Sec 4.3. We note that inherent domain label noise can impact domain NMI scores. Thus, NMI is more valuable when used as a relative measure rather than an absolute indicator of domain separability. + +
DatasetDiTSD-2.1MAECLIPDINOv2RN50
PACS0.080.080.110.050.150.29
VLCS0.120.150.170.010.110.39
TerraInc0.320.350.320.010.160.30
OfficeHome0.160.220.280.100.230.59
DomainNet0.160.200.220.130.190.36
+ +Table 4. Comparison of class NMI scores across datasets. In order to choose auxiliary features for domain separation, a feature space that yields lower class NMI score along with high domain NMI is desirable, i.e. the latent space should favor grouping domains over object classes. Note that Synth-Artists and Synth-Photography datasets are omitted here as they do not have predefined class labels. + +composed of six domains such as quick-draw, infographic, real images, and so on, and exhibits a broader range of domain shifts than PACS, spanning both coarse and fine-grained variations. For example, the "quickdraw" domain consists of simple, rough sketches, while "sketch" has more detailed drawings with shading and varied strokes, showing style differences. By contrast, "real" domain captures fully detailed images, indicating shifts of varied granularities between different domains. + +# 4.3. Effect of the Choice of $\Psi$ on Domain Separation + +Next, we study how different pre-training objectives affect the separation of domain-specific signals using domain NMI $(\uparrow)$ (introduced in Sec. 4.1), which measures how well domains are separated in the latent space. We acknowledge that all models are of varied architectural complexities, trained on very different datasets, thereby making it nonviable to concretely isolate the cause of performance discrepancies in domain separation. Nevertheless, we believe our below analysis is valuable to understand the semantic information captured by different pre-training objectives. + +ResNet-50 [20] (RN50) is pre-trained on ImageNet [14] using a discriminatory cross-entropy classification loss. Consequently, the feature space evolves to aid object discrimination, making samples from the same class cluster together across domains. This is evident in the relatively low domain NMI scores (Table 3) and high class NMI across all datasets (Table 3), e.g., a class NMI of 0.29 compared to 0.08 by DiT features on the PACS dataset. + +CLIP [55] is pre-trained on internet-scale, noisy image-text pairs using a contrastive loss that aligns images with + +their textual descriptions in a joint embedding space. This objective prioritizes high-level semantic similarity, making CLIP's feature space more representative of global semantics and overall context of the image instead of object-specific details. Consequently, images of the same object may not form tight clusters if their captions emphasize different contextual attributes (e.g. "a dog on a beach" vs. "a golden retriever indoors"). Thus, CLIP, though rich in broader contextual semantics, yields low class and domain NMI scores across all datasets in (Tables 3 and 4). + +DINOv2 [49] is a self-supervised vision transformer trained by aligning representations between a student and a teacher network, across global and local crops of an image. This encourages the model to capture primarily low-level features, while also capturing global relationships to some extent [31, 49, 67]. By enforcing consistency across augmentations, DINOv2 preserves low-level features that remain invariant to these transformations. Thus DINO-v2 features are particularly effective for datasets like Office-Home (domain NMI of 0.38 in Table 3), where domain shifts are driven by low-level style differences such as bold outlines in the "clipart" domain vs softer, natural edges in the "real" domain (example images in suppl. material). By contrast, DINOv2 performs poorly on VLCS (domain NMI of 0.05 in Table 3), likely due to its over-reliance on low-level features, making it less effective at capturing high-level dataset-specific biases in VLCS, such as differences in spatial composition and object size variations. + +Masked Autoencoders [22] (MAEs) are pre-trained using a masking objective, where the model learns to reconstruct locally masked patches of an image. We conjecture that by reconstructing small, local patch details, MAE's pretraining objective may introduce a strong locality bias, and fail to capture global image context, as studied in [40, 82]. We hypothesize that this lack of global understanding limits the capability of MAEs to offer complimentary domain-specific representations. This is evident in their relatively high class-NMI scores (as seen in Table 4) and low domain-NMI scores (as seen in Table 3) across most datasets. However, MAEs achieve relatively high domain NMI scores on PACS (0.71) and DomainNet (0.52) leveraging the visual information from local details such as textures, shading, and brushstrokes. We note a similar trend with DINOv2 which also captures rich local features. This may explain why both models perform better in separating domains driven by low-level visual variations (Table 3). However, MAEs perform poorly on TerraIncognita despite its reliance on local features. Unlike PACS, we think that the domain shifts in TerraIncognita require an understanding of both local and global spatial understanding (e.g., vegetation density, terrain patterns), potentially leading to lower domain NMI. + +Conclusion: This indepth analysis indicates that comprehending different pre-training objectives is essential to max + +![](images/403f1da265a25234335b01224c310fb071cc226d5ba4fbfbf5dd49e20f583ab4.jpg) +Figure 3. t-SNE visualization of how pseudo-domains are clustered together in the latent space of DiT for PACS. Note how the sketch domain forms distinct clusters, with light and dark pencil strokes mapped to separate regions in the latent space. Best viewed in color. + +![](images/0d4ec6ae813af32ec9a364cd7e772c4ad3b36e8ab350fee18b67cf0a743e1a82.jpg) +(a) Animal portraits +(b) Oil paintings +(c) Similar color schemes +Figure 4. Pseudo-domains captured in the diffusion latent space of DiT on PACS. The clusters group images based on nuanced style-specific variances rather than class-specific variances. + +imize profit from their latents for domain separation. + +# Diffusion models for domain separation + +Next, we focus exclusively on diffusion architectures and closely study the impact of some of their architectural design choices on domain separation. As discussed in Sec. 3.1, during diffusion model pre-training, noise added to an image is iteratively removed using pixel reconstruction loss. Recent studies [50, 54] have indicated that this makes the model first capture broad structural patterns before encoding finer details. We hypothesize that this implicit hierarchical feature learning indirectly induced by the denoising objective enables diffusion models to encode global structures and fine-grained variations, assisting faithful image reconstruction. Moreover, since the generative objective is entirely agnostic to class labels, we posit that there is no incentive to group features based on class-discriminative signals. Perhaps this lack of class-driven objective allows domain-specific variations to emerge more prominently in the latent space. This is reflected in Table 3 where we observe that diffusion features achieve high domain NMI scores across most datasets compared to their non-diffusion counterparts. Figures 4, and 3 further reinforces this observation and illustrates how different clusters (pseudo-domains) in the diffusion latent space capture domain-specific class-agnostic variations. + +Within the family of diffusion feature space, we now in + +spect if transformer based DiT and U-Net based SD-2.1 behave differently for the task of domain separation. We acknowledge that both models are trained on very different datasets which makes this analysis more challenging. + +DiT [51]: Following the analysis in Kim et al. [32], we extract features from the 14th (out of 28) block of the transformer architecture of the DiT model, at timestep $t = 50$ (more in suppl. material). As noted in [32], by attending to the entire image, DiT's self-attention mechanism effectively captures global context, making it capable at distinguishing high-level semantics and stylistic differences (e.g., pencil sketches vs paintings). This proves advantageous on datasets like PACS [35] which comprises domains with varied global context (detailed in Sec. 4.2), where DiT achieves the highest domain NMI of 0.85 (Table 3). + +SD-2.1 [58]. We extract features from the second upsampling layer of the U-Net (denoted as up_ft:1) in SD-2.1 at timestep $t = 50$ (more in suppl. material). As noted in [32], these features are rich in fine-grained visual information, with convolutional-based U-Net [59] of SD-2.1 [58] capturing local spatial information [32, 68]. As a result, we observe that SD-2.1 and DiT exhibit complementary strengths. This is particularly evident on the TerraIncognita dataset, where SD-2.1 achieves the highest domain NMI of 0.55 (Table 3). This is likely due to SD-2.1's ability to capture fine-grained spatial features such as foliage density and terrain patterns, which define the domain shifts for this dataset (as described in Sec 4.2). By contrast, we find that DiT struggles with domain separation on TerraIncognita, achieving a lower domain NMI of 0.22. On the other hand, on VLCS [17] where each domain represents dataset-specific biases described in Sec. 4.2, we note that DiT (domain NMI: 0.58) outperforms SD-2.1 (domain NMI: 0.26). This highlights DiT's strength to capture global context. Interestingly, SD-2.1's bottleneck layer achieves a higher domain NMI of 0.45 compared to up_ft:1's score of 0.26. This aligns with the findings from Kim et al. [32] that U-Net's bottleneck layer captures coarser, more global features, compared to up_ft:1. + +From Table 3, we observe that OfficeHome [70] proves to be challenging for both DiT (domain NMI: 0.25) and SD-2.1 (domain NMI: 0.28). Upon inspection, we found that samples from the "real" domain visually look similar to those from both "product" and "art" in the feature spaces (suppl. material for visual examples), potentially contributing to low domain separation. On DomainNet [52], we observe moderate domain NMI scores for all pre-training objectives (except for CLIP, as discussed in Sec 4.3), with DiT achieving the highest score of 0.54, in Table 3. We attribute this to the diverse nature of domain shifts in DomainNet, which include both high- and low-level variations (described in Sec. 4.2). We believe that this variability makes it challenging for models to fully leverage their dis + +
DatasetDiTSD-2.1RN50CLIPDINOv2MAEERM
VLCS78.577.076.376.877.376.476.6
PACS87.186.984.884.784.984.683.8
OH68.468.665.764.668.365.267.2
TI48.251.349.847.448.450.247.0
Avg70.671.069.168.469.769.168.7
+ +Table 5. Domain generalization performance using different pseudodomain representations. The pseudo-domain representations obtained from the latent space of diffusion models provide the highest gains in accuracy, while those from CLIP yield minimal accuracy gains. + +tinct strengths, as no single model seems to effectively capture all domain-specific characteristics. + +Conclusion: This analysis reveals that for the same pretraining objective (diffusion denoising), the underlying architecture and the specific layer for feature extraction plays a crucial role in shaping the latent space, thereby performance on the downstream tasks. + +# 4.4. Domain Generalization Performance + +In this section, we compare GUIDE against prior domain generalization methods and examine the impact of different feature extractors $(\Psi)$ in capturing domain-specific information to enhance classification performance. + +Choice of $\Psi$ on domain generalization: Building on our findings in Sec. 4.3, we test the utility of different feature spaces for domain separation and generalization against ERM [69], a strong baseline that has been shown by Gulrajani and Lopez-Paz [19] to outperform many domain generalization algorithms. We evaluate on the DomainBed test suite, which comprises PACS [35], VLCS [17], OfficeHome [70], TerraIncognita [4], and DomainNet [52]. From Table 5, we note that diffusion features consistently outperform their non-diffusion counterparts on all datasets. Notably, DiT and SD-2.1 achieve highest accuracy while the rest show only marginal gains over ERM. CLIP seems to yield minimal gains on average on this task limiting its ability to be used "as is." GUIDE-DiT yields an average accuracy improvement of $1.9\%$ over ERM and performs best on VLCS $(+1.9\%)$ and PACS $(+3.3\%)$ . On the other hand, GUIDE-SD-2.1 outperforms on TerraIncognita, beating ERM by $+4.3\%$ . These results are inline with the analysis and domain NMI scores in Table 3, and with the theoretical findings in Thomas et al. [66], which show that higher domain NMI correlates with improved performance. + +Comparison with prior art: In Table 6, we compare GUIDE with other state-of-the-art domain generalization algorithms1 and note that GUIDE-BEST achieves the highest average performance of $66.3\%$ without using domain labels at any point. Compared to all methods, GUIDE-BEST shows the largest improvements on the PACS, TerraIncognita, and DomainNet datasets. The significant gains on DomainNet, a dataset with over + +
uses multi-layer featuresuses domain labelsAlgorithmVLCSPACSOHTIDNAvg
--ERM [69]76.683.867.247.044.163.7
✓✓MLDG [36]77.284.966.847.741.263.6
✓✓MMD [37]77.584.766.342.223.458.8
✓✓CORAL [62]78.886.268.747.641.564.5
✓✓SagNet [48]77.886.368.148.640.364.2
X✓DANN [18]78.683.665.946.738.362.6
X✓Fisher [56]77.885.567.847.441.764.0
✓XMIRO [11]79.085.470.550.444.365.9
XXMixup [75, 77]77.484.668.147.939.263.4
XXLatentDR (SA) [42]78.785.869.049.945.165.7
XXLatentDR (Pool) [42]78.086.368.449.543.965.2
X✓DA-ERM ([16])78.084.167.947.343.664.1
XXAdaClust ([66])78.987.067.748.143.664.9
XXGUIDE-DiT (ours)78.587.168.448.245.865.6
XXGUIDE-SD-2.1 (ours)77.086.968.651.345.965.9
XXGUIDE-BEST (ours)78.587.168.651.345.966.3
+ +Table 6. Comparison of GUIDE with other domain generalization algorithms on 5 datasets: utilizing the DomainBed test bed. The methods are categorized based on (1) whether they operate across multiple intermediate layers in the network and (2) whether they require explicit ground truth domain labels during training. The highest-performing method that does not rely on either is underlined. The overall best-performing method is in bold. Methods in cyan corresponds to domain-adaptive classifiers (described in Sec. 3.3). Among those methods we find GUIDE achieves the highest performance. GUIDE-BEST reports the best performance among the two diffusion latent spaces (DiT and SD-2.1) for easy reading. + +
DatasetERM / ERM++GUIDE / GUIDE++ (DiT)GUIDE / GUIDE++ (SD-2.1)
PACS83.8 / 88.087.1 / 89.286.9 / 88.6
TI47.0 / 50.748.2 / 52.751.3 / 53.6
+ +500,000 images across 325 classes and 6 domains, highlights GUIDE's ability to scale to larger datasets. Among the domain-adaptive classifier frameworks (bottom rows), GUIDE-BEST outperforms DA-ERM [16] by $+2.2\%$ and AdaClust [66] by $+1.4\%$ . Notably, the reported scores for most algorithms are obtained after extensive hyperparameter searches, whereas GUIDE achieves these gains with the default setting of DomainBed without using features from multiple layers or ground truth domain labels. Overall, results in Tables 5 and 6 validate our hypothesis that augmenting a feature space with rich domain-specific information on seen domains results in an overall generalizable feature space for unseen domains. + +Effect of enhanced training strategies: We follow the $\mathrm{ERM}++$ [64] implementation from DomainBed [19] which improves ERM by better utilization of training data, model parameter selection, and weight-space regularization techniques. From Table 7, $\mathrm{ERM}++$ improves over standard ERM by $+4.2\%$ on PACS and $+3.7\%$ on TerraIncognita. Applying the same strategies to GUIDE, we achieve even greater improvements, with $\mathrm{GUIDE}++$ outperforming ERM by $+5.4\%$ on PACS and $+6.6\%$ on TerraIncognita. These results show that GUIDE could benefit from any training optimizations proposed over ERM, such as SWAD [10] + +Effect of different $\Phi$ backbones: We use ResNet-50 as $\Phi$ for a fair comparison with prior work [19]. To test robustness across different backbones, following Angarano et al. [1], Table 8 shows that GUIDE consistently outperforms the + +Table 7. $\mathbf{ERM} + +$ [64] training strategies on GUIDE boost performance. + +
AlgorithmResNet-18 [20]ResNet-50 [20]ResNet-50-SWSL [76]ViT-B/16 [15]
ERM80.283.886.988.1
GUIDE81.887.190.089.7
+ +Table 8. Test accuracy on PACS [35] across different $\Phi$ backbones. GUIDE consistently provides gains over ERM. + +![](images/bc8f2f23a0b92695948a1886ffa3a50a0ad9e4ca40963838fad409e32ea2d901.jpg) +Figure 5. Example images from Synth-Artists and Synth-Photography, generated using Stable Diffusion XL [53]. Synth-Artists includes artistic styles such as Van Gogh and Kinkade, while the Synth-Photography captures photography effects like Tilt-Shift and Bokeh. + +baseline (ERM) across multiple backbones. + +# 4.5. Pseudo-domains for Style Discovery + +Next, we evaluate different pre-training objectives on the task of photographic and artistic style separation. Automatic style identification is valuable for curating and inspecting large-scale datasets, image retrieval, and several such applications. To study this, we first construct two datasets with controlled domain shifts using Stable Diffusion XL [53] (dataset construction details in suppl. material): (i) Synth-Photography features photographic styles such as macro, tilt-shift, bokeh, symmetry, and zoom blur. Thus, the domain shifts are primarily driven by variations in focus, sharpness, edge details, and depth contrasts. (ii) Synth-Artists, captures styles of Van Gogh, Kinkade, Warhol, Rembrandt, and Dali, making the domain shifts more high-level such as brush stroke patterns and color palettes. We show a few example images in Fig 5. On Synth-Artists, we observe that DiT achieves better domain separation, with a domain NMI score of 0.89 (Table 3). By contrast, on Synth-Photography, SD-2.1 performs better, achieving a domain NMI score of 0.43 compared to DiT's score of 0.35 (Table 3). This finding aligns with our analysis from Sec. 4.3 that DiT seem more apt for global variations and SD-2.1 for finer-grained spatially detailed variations. + +# 5. Discussion and Future Work + +In this work, we introduce GUIDE, a simple yet effective framework that improves generalization to unseen domains in the absence of domain labels during both train and test times. Limitations include the lack of a dataset specific heuristic to select optimal number of clusters, and the overhead of an additional feature extractor. Future work includes exploring ways to combine multiple models and build a generalizable latent space that works "out of the box" for diverse tasks. + +Acknowledgments. We thank Piotr Teterwak, Jonathan Granskog, and other members of our research group at BU for helpful discussions and feedback. + +# References + +[1] Simone Angarano, Mauro Martini, Francesco Salvetti, Vittorio Mazzia, and Marcello Chiaberge. Back-to-bones: Rediscovering the role of backbones in domain generalization. arXiv preprint arXiv:2209.01121, 2023. 8 +[2] Yogesh Balaji, Swami Sankaranarayanan, and Rama Chellappa. Metareg: Towards domain generalization using metaregularization. Advances in Neural Information Processing Systems, 2018. 2 +[3] Dmitry Baranchuk, Ivan Rubachev, Andrey Voynov, Valentin Khrulkov, and Artem Babenko. Label-efficient semantic segmentation with diffusion models. arXiv preprint arXiv:2112.03126, 2021. 2 +[4] Sara Beery, Grant Van Horn, and Pietro Perona. Recognition in terra incognita. In Proceedings of the European conference on computer vision (ECCV), 2018. 2, 4, 5, 7 +[5] Shai Ben-David, John Blitzer, Koby Crammer, Alex Kulesza, Fernando Pereira, and Jennifer Wortman Vaughan. A theory of learning from different domains. Machine learning, 2010. 2 +[6] Gilles Blanchard, Gyemin Lee, and Clayton Scott. Generalizing from several related classification tasks to a new unlabeled sample. Advances in Neural Information Processing Systems, 2011. 1, 2, 3 +[7] Manh-Ha Bui, Toan Tran, Anh Tran, and Dinh Phung. Exploiting domain-specific features to enhance domain generalization. Advances in Neural Information Processing Systems, 2021. 2, 3 +[8] Mathilde Caron, Piotr Bojanowski, Armand Joulin, and Matthijs Douze. Deep clustering for unsupervised learning of visual features. In Proceedings of the European conference on computer vision (ECCV), 2018. 1 +[9] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. Advances in Neural Information Processing Systems, 2020. 1 +[10] Junbum Cha, Sanghyuk Chun, Kyungjae Lee, Han-Cheol Cho, Seunghyun Park, Yunsung Lee, and Sungrae Park. Swad: Domain generalization by seeking flat minima. Advances in Neural Information Processing Systems, 2021. 8 +[11] Junbum Cha, Kyungjae Lee, Sungrae Park, and Sanghyuk Chun. Domain generalization by mutual-information regularization with pre-trained models. In Proceedings of the European conference on computer vision (ECCV), 2022. 2, 8 +[12] Shoufa Chen, Peize Sun, Yibing Song, and Ping Luo. Diffusiondet: Diffusion model for object detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 19830-19843, 2023. 2 +[13] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning + +of visual representations. In International conference on machine learning, 2020. 1 +[14] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2009. 5 +[15] A. Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020.8 +[16] Abhimanyu Dubey, Vignesh Ramanathan, Alex Pentland, and Dhruv Mahajan. Adaptive methods for real-world domain generalization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2021. 2, 3, 4, 5, 7, 8 +[17] Chen Fang, Ye Xu, and Daniel N Rockmore. Unbiased metric learning: On the utilization of multiple datasets and web images for softening bias. In Proceedings of the IEEE/CVF international conference on computer vision, 2013. 1, 4, 5, 7 +[18] Yaroslav Ganin, Evgeniya Ustinova, Hana Ajakan, Pascal Germain, Hugo Larochelle, François Laviolette, Mario March, and Victor Lempitsky. Domain-adversarial training of neural networks. Journal of machine learning research, 2016. 2, 8 +[19] Ishaan Gulrajani and David Lopez-Paz. In search of lost domain generalization. In International Conference on Learning Representations, 2021. 2, 4, 7, 8 +[20] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2016. 1, 4, 5, 8 +[21] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2020. 1 +[22] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022. 1, 4, 6 +[23] Sobhan Hemati, Mahdi Beitollahi, Amir Hossein Estiri, Bassel Al Omari, Soufiane Lamghari, Yasser H Khalil, Xi Chen, and Guojun Zhang. Beyond loss functions: Exploring data-centric approaches with diffusion model for domain generalization. Transactions on Machine Learning Research, 2022. 2 +[24] Sobhan Hemati, Mahdi Beitollahi, Amir Hossein Estiri, Bassel Al Omari, Xi Chen, and Guojun Zhang. Cross domain generative augmentation: Domain generalization with latent diffusion models. arXiv preprint arXiv:2312.05387, 2023. 2 +[25] Dan Hendrycks, Norman Mu, Ekin D Cubuk, Barret Zoph, Justin Gilmer, and Balaji Lakshminarayanan. Augmix: A simple data processing method to improve robustness and uncertainty. arXiv preprint arXiv:1912.02781, 2019. 4 +[26] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 2020. 2, 3 +[27] Minui Hong, Jinwoo Choi, and Gunhee Kim. Stylemix: Separating content and style for enhanced data augmentation. In + +Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2021. 2 +[28] Yuyang Huang, Yabo Chen, Yuchen Liu, Xiaopeng Zhang, Wenrui Dai, Hongkai Xiong, and Qi Tian. Domainfusion: Generalizing to unseen domains with latent diffusion models. In Proceedings of the European conference on computer vision (ECCV), 2025. 2 +[29] Zeyi Huang, Haohan Wang, Eric P Xing, and Dong Huang. Self-challenging improves cross-domain generalization. In Proceedings of the European conference on computer vision (ECCV), 2020. 2 +[30] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International conference on machine learning, 2021. 1 +[31] Dongsheng Jiang, Yuchen Liu, Songlin Liu, Jin'e Zhao, Hao Zhang, Zhen Gao, Xiaopeng Zhang, Jin Li, and Hongkai Xiong. From clip to dino: Visual encoders shout in multi-modal large language models. arXiv preprint arXiv:2310.08825, 2023. 6 +[32] Dahye Kim, Xavier Thomas, and Deehti Ghadiyaram. Revelio: Interpreting and leveraging semantic information in diffusion models. arXiv preprint arXiv:2411.16725, 2024. 2, 4, 7 +[33] Diederik P Kingma, Max Welling, et al. Auto-encoding variational bayes, 2013. 3 +[34] Alexander C Li, Mihir Prabhudesai, Shivam Duggal, Ellis Brown, and Deepak Pathak. Your diffusion model is secretly a zero-shot classifier. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2206-2217, 2023. 2 +[35] Da Li, Yongxin Yang, Yi-Zhe Song, and Timothy M Hospedales. Deeper, broader and artier domain generalization. In Proceedings of the IEEE international conference on computer vision, 2017. 4, 5, 7, 8 +[36] Da Li, Yongxin Yang, Yi-Zhe Song, and Timothy Hospedales. Learning to generalize: Meta-learning for domain generalization. In Proceedings of the AAAI conference on artificial intelligence, 2018. 8 +[37] Haoliang Li, Sinno Jialin Pan, Shiqi Wang, and Alex C Kot. Domain generalization with adversarial feature learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2018. 8 +[38] Jingjing Li, Zhiqi Yu, Zhekai Du, Lei Zhu, and Heng Tao Shen. A comprehensive survey on source-free domain adaptation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 1 +[39] Pan Li, Da Li, Wei Li, Shaogang Gong, Yanwei Fu, and Timothy M Hospedales. A simple feature augmentation for domain generalization. In Proceedings of the IEEE international conference on computer vision, 2021. 2 +[40] Feng Liang, Yangguang Li, and Diana Marculescu. Supmae: Supervised masked autoencoders are efficient vision learners. arXiv preprint arXiv:2205.14540, 2022. 6 +[41] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36:34892-34916, 2023. 1 + +[42] Ran Liu, Sahil Khose, Jingyun Xiao, Lakshmi Sathidevi, Keerthan Ramnath, Zsolt Kira, and Eva L Dyer. Latentdr: Improving model generalization through sample-aware latent degradation and restoration. In Proceedings of the IEEE Winter Conference on Applications of Computer Vision, 2024. 2, 7, 8 +[43] Grace Luo, Lisa Dunlap, Dong Huk Park, Aleksander Holynski, and Trevor Darrell. Diffusion hyperfeatures: Searching through time and space for semantic correspondence. Advances in Neural Information Processing Systems, 2024. 2 +[44] Dhruv Mahajan, Ross Girshick, Vignesh Ramanathan, Kaiming He, Manohar Paluri, Yixuan Li, Ashwin Bharambe, and Laurens Van Der Maaten. Exploring the limits of weakly supervised pretraining. In Proceedings of the European conference on computer vision (ECCV), 2018. 1 +[45] Toshihiko Matsuura and Tatsuya Harada. Domain generalization using a mixture of multiple latent domains. In Proceedings of the AAAI conference on artificial intelligence, 2020. 2, 3, 5 +[46] Krikamol Muandet, Kenji Fukumizu, Bharath Striperumbudur, Bernhard Schölkopf, et al. Kernel mean embedding of distributions: A review and beyond. Foundations and Trends® in Machine Learning, 2017. 1, 3 +[47] Hyeonseob Nam, HyunJae Lee, Jongchan Park, Wonjun Yoon, and Donggeun Yoo. Reducing domain gap by reducing style bias. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2021. 2 +[48] Hyeonseob Nam, HyunJae Lee, Jongchan Park, Wonjun Yoon, and Donggeun Yoo. Reducing domain gap by reducing style bias. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2021. 8 +[49] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 1, 4, 6 +[50] Yong-Hyun Park, Mingi Kwon, Jaewoong Choi, Junghyo Jo, and Youngjung Uh. Understanding the latent space of diffusion models through the lens of riemannian geometry. Advances in Neural Information Processing Systems, 2023. 6 +[51] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4195-4205, 2023. 1, 4, 7 +[52] Xingchao Peng, Qinxun Bai, Xide Xia, Zijun Huang, Kate Saenko, and Bo Wang. Moment matching for multi-source domain adaptation. In Proceedings of the IEEE international conference on computer vision, 2019. 4, 5, 7 +[53] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 8 +[54] Yurui Qian, Qi Cai, Yingwei Pan, Yehao Li, Ting Yao, Qibin Sun, and Tao Mei. Boosting diffusion models with moving average sampling in frequency domain. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2024. 6 + +[55] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PmLR, 2021. 1, 4, 5 +[56] Alexandre Rame, Corentin Dancette, and Matthieu Cord. Fishr: Invariant gradient variances for out-of-distribution generalization. In International conference on machine learning, 2022. 8 +[57] Benjamin Recht, Rebecca Roelofs, Ludwig Schmidt, and Vaishaal Shankar. DoImagenet classifiers generalize toImagenet? In International conference on machine learning, 2019. 1 +[58] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022. 1, 3, 4, 7 +[59] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention, 2015. 4, 7 +[60] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning. PMLR, 2015. 2, 3 +[61] Nathan Somavarapu, Chih-Yao Ma, and Zsolt Kira. Frustratingly simple domain generalization via image stylization. arXiv preprint arXiv:2006.11207, 2020. 2 +[62] Baochen Sun and Kate Saenko. Deep coral: Correlation alignment for deep domain adaptation. In Proceedings of the European conference on computer vision (ECCV), 2016, 2, 8 +[63] Rohan Taori, Achal Dave, Vaishaal Shankar, Nicholas Carlini, Benjamin Recht, and Ludwig Schmidt. Measuring robustness to natural distribution shifts in image classification. Advances in Neural Information Processing Systems, 2020. 1 +[64] Piotr Teterwak, Kuniaki Saito, Theodoros Tsiligkaridis, Kate Saenko, and Bryan A Plummer. Erm++: An improved baseline for domain generalization. arXiv preprint arXiv.2304.01973, 2023. 2, 8 +[65] Piotr Teterwak, Kuniaki Saito, Theodoros Tsiligkaridis, Bryan A Plummer, and Kate Saenko. Is large-scale pretraining the secret to good domain generalization? arXiv preprint arXiv:2412.02856, 2024. 1 +[66] Xavier Thomas, Dhruv Mahajan, Alex Pentland, and Abhimanyu Dubey. Adaptive methods for aggregated domain generalization. arXiv preprint arXiv:2112.04766, 2021. 2, 3, 5, 7, 8 +[67] Shengbang Tong, Zhuang Liu, Yuexiang Zhai, Yi Ma, Yann LeCun, and Saining Xie. Eyes wide shut? exploring the visual shortcomings of multimodal llms. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2024. 1, 6 + +[68] N. Tumanyan, M. Geyer, S. Bagon, and T. Dekel. Plug-and-play diffusion features for text-driven image-to-image translation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2023. 7 +[69] Vladimir N Vapnik. An overview of statistical learning theory. IEEE transactions on neural networks, 1999. 2, 3, 4, 7, 8 +[70] Hemanth Venkateswara, Jose Eusebio, Shayok Chakraborty, and Sethuraman Panchanathan. Deep hashing network for unsupervised domain adaptation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2017. 3, 4, 5, 7 +[71] Andrey Voynov, Qinghao Chu, Daniel Cohen-Or, and Kfir Aberman. p+: Extended textual conditioning in text-to-image generation. arXiv preprint arXiv:2303.09522, 2023. 2 +[72] Wenxuan Wang, Quan Sun, Fan Zhang, Yepeng Tang, Jing Liu, and Xinlong Wang. Diffusion feedback helps clip see better. arXiv preprint arXiv:2407.20171, 2024. 2 +[73] Weijia Wu, Yuzhong Zhao, Hao Chen, Yuchao Gu, Rui Zhao, Yefei He, Hong Zhou, Mike Zheng Shou, and Chunhua Shen. Datasetdm: Synthesizing data with perception annotations using diffusion models. In Advances in Neural Information Processing Systems, 2023. 2 +[74] Jiarui Xu, Sifei Liu, Arash Vahdat, Wonmin Byeon, Xiaolong Wang, and Shalini De Mello. Open-vocabulary panoptic segmentation with text-to-image diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2023. 2 +[75] Minghao Xu, Jian Zhang, Bingbing Ni, Teng Li, Chengjie Wang, Qi Tian, and Wenjun Zhang. Adversarial domain adaptation with domain mixup. In Proceedings of the AAAI conference on artificial intelligence, 2020. 8 +[76] I Zeki Yalniz, Herve Jégou, Kan Chen, Manohar Paluri, and Dhruv Mahajan. Billion-scale semi-supervised learning for image classification. arXiv preprint arXiv:1905.00546, 2019.8 +[77] Shen Yan, Huan Song, Nanxiang Li, Lincan Zou, and Liu Ren. Improve unsupervised domain adaptation with mixup training. arXiv preprint arXiv:2001.00677, 2020. 8 +[78] Xueting Yan, Ishan Misra, Abhinav Gupta, Deepti Ghadiyaram, and Dhruv Mahajan. Clusterfit: Improving generalization of visual representations. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2020. 1 +[79] Runpeng Yu, Songhua Liu, Xingyi Yang, and Xinchao Wang. Distribution shift inversion for out-of-distribution prediction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2023. 2 +[80] Yun Zhang, Nianbin Wang, Shaobin Cai, and Lei Song. Unsupervised domain adaptation by mapped correlation alignment. IEEE Access, 2018. 4 +[81] Wenliang Zhao, Yongming Rao, Zuyan Liu, Benlin Liu, Jie Zhou, and Jiwen Lu. Unleashing text-to-image diffusion models for visual perception. In Proceedings of the IEEE international conference on computer vision, 2023. 2 + +[82] Xie Zhenda, Geng Zigang, Hu Jingcheng, Zhang Zheng, Hu Han, and Cao Yue. Revealing the dark secrets of masked image modeling. arXiv preprint arXiv:2205.13543, 2022. 6 \ No newline at end of file diff --git a/ICCV/2025/What's in a Latent_ Leveraging Diffusion Latent Space for Domain Generalization/images.zip b/ICCV/2025/What's in a Latent_ Leveraging Diffusion Latent Space for Domain Generalization/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..3df85710307035518cb974f81d84576e31249c56 --- /dev/null +++ b/ICCV/2025/What's in a Latent_ Leveraging Diffusion Latent Space for Domain Generalization/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:09468d51117fc8e4cc1138312634ace341a0554c85298b1a63970dbab4a33fed +size 448314 diff --git a/ICCV/2025/What's in a Latent_ Leveraging Diffusion Latent Space for Domain Generalization/layout.json b/ICCV/2025/What's in a Latent_ Leveraging Diffusion Latent Space for Domain Generalization/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..e20eb4ca48776bd09d07b8ac265f79e29c29ebb4 --- /dev/null +++ b/ICCV/2025/What's in a Latent_ Leveraging Diffusion Latent Space for Domain Generalization/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:98947df0bab8046a3d3771858aa582bc18c07758b9f78d7d9e4d548a12579a7e +size 455787 diff --git a/ICCV/2025/When Anchors Meet Cold Diffusion_ A Multi-Stage Approach to Lane Detection/03666e3c-da14-47a9-8be7-58e6b0821b95_content_list.json b/ICCV/2025/When Anchors Meet Cold Diffusion_ A Multi-Stage Approach to Lane Detection/03666e3c-da14-47a9-8be7-58e6b0821b95_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..935a77254b66c81cb54024d42322aa52fca5a209 --- /dev/null +++ b/ICCV/2025/When Anchors Meet Cold Diffusion_ A Multi-Stage Approach to Lane Detection/03666e3c-da14-47a9-8be7-58e6b0821b95_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a55e78cddae5068ae0969a46b5cfe6a2d554c3b090f491c208487902056b5a6d +size 78188 diff --git a/ICCV/2025/When Anchors Meet Cold Diffusion_ A Multi-Stage Approach to Lane Detection/03666e3c-da14-47a9-8be7-58e6b0821b95_model.json b/ICCV/2025/When Anchors Meet Cold Diffusion_ A Multi-Stage Approach to Lane Detection/03666e3c-da14-47a9-8be7-58e6b0821b95_model.json new file mode 100644 index 0000000000000000000000000000000000000000..3cbfd3e64a76ca84064cd06e59c8907a7ace44a7 --- /dev/null +++ b/ICCV/2025/When Anchors Meet Cold Diffusion_ A Multi-Stage Approach to Lane Detection/03666e3c-da14-47a9-8be7-58e6b0821b95_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:68b4ac5a06808f9db6cb132b288056d1b244b0b18cc17db027464ca9c11d5bc8 +size 97250 diff --git a/ICCV/2025/When Anchors Meet Cold Diffusion_ A Multi-Stage Approach to Lane Detection/03666e3c-da14-47a9-8be7-58e6b0821b95_origin.pdf b/ICCV/2025/When Anchors Meet Cold Diffusion_ A Multi-Stage Approach to Lane Detection/03666e3c-da14-47a9-8be7-58e6b0821b95_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..2299bbc98042012c1ae132fbdc2abffc2c0e1a96 --- /dev/null +++ b/ICCV/2025/When Anchors Meet Cold Diffusion_ A Multi-Stage Approach to Lane Detection/03666e3c-da14-47a9-8be7-58e6b0821b95_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:17ca4528ac0a42c6fb40bdc92ccff4dc16f8ed5c74be48487d16da9611151b3b +size 4305199 diff --git a/ICCV/2025/When Anchors Meet Cold Diffusion_ A Multi-Stage Approach to Lane Detection/full.md b/ICCV/2025/When Anchors Meet Cold Diffusion_ A Multi-Stage Approach to Lane Detection/full.md new file mode 100644 index 0000000000000000000000000000000000000000..255d64db8ebc3aad47f64f7a5a12f84c1caaa677 --- /dev/null +++ b/ICCV/2025/When Anchors Meet Cold Diffusion_ A Multi-Stage Approach to Lane Detection/full.md @@ -0,0 +1,290 @@ +# When Anchors Meet Cold Diffusion: A Multi-Stage Approach to Lane Detection + +Bo-Lun Huang + +Zi-Xiang Ni + +Feng-Kai Huang + +Hong-Han Shuai $^{1*}$ + +Wen-Huang Cheng + +$^{1}$ National Yang Ming Chiao Tung University + +$^{2}$ National Taiwan University + +# Abstract + +Accurate and stable lane detection is crucial for the reliability of autonomous driving systems. A core challenge lies in predicting lane positions in complex scenarios, such as curved roads or when markings are ambiguous or absent. Conventional approaches leverage deep learning techniques to extract both high-level and low-level visual features, aiming to achieve a comprehensive understanding of the driving environment. However, these methods often rely on predefined anchors within a single-pass model, limiting their adaptability. The one-shot prediction paradigm struggles with precise lane estimation in challenging scenarios, such as curved roads or adverse conditions like low visibility at night. To address these limitations, we propose a novel cold diffusion-based framework that initializes lane predictions with predefined anchors and iteratively refines them. This approach retains the flexibility and progressive refinement capabilities of diffusion models while overcoming the constraints of traditional hot diffusion techniques. To further enhance the model's coarse-to-fine refinement capabilities, we introduce a multi-resolution image processing strategy, where images are analyzed at different timesteps to capture both global and local lane structure details. Besides, we incorporate a learnable noise variance schedule, enabling the model to dynamically adjust its learning process based on multi-resolution inputs. Experimental results demonstrate that our method significantly improves detection accuracy across a variety of challenging scenarios, outperforming state-of-the-art lane detection methods. Codes and trained weights are available at https://github.com/ntudr/CDiffLane + +# 1. Introduction + +Lane detection [16, 32, 33, 43] is a pivotal component of autonomous driving pipelines, as accurately estimating lane boundaries is essential for safe and robust vehi + +![](images/3068729365ee4be0710e14750686353196909d9ca104ead61fb5fe1b5688b61b.jpg) +Figure 1. Comparison of various lane detection approaches. (a) Examples of methods such as LaneATT [40], CLRNet [53], and GSENet [39]. (b) Standard diffusion-based approach, similar to DiffusionDet [6], where lane predictions are generated from noisy initializations. (c) Our approach adopts the concept of cold diffusion, beginning the diffusion process from predefined anchors, providing a more reliable starting point for detection. + +cle navigation. Traditional methods [12, 20, 24, 27], often reliant on hand-crafted features and post-processing techniques, require extensive parameter tuning and are sensitive to complex road conditions. Recent advances in deep learning have greatly improved lane detection performance by leveraging convolutional neural networks (CNNs) to extract rich feature representations [4, 26, 28, 53]. Approaches have evolved from segmentation-based methods [26, 28] to anchor-based [4, 53], row-wise detection [31, 32], and parametric regressors [23, 41]. These deep-learning pipelines typically operate in a single forward pass, accelerating inference while improving accuracy. + +Despite these strengths, single-pass lane detection methods can struggle under challenging scenarios such as low visibility at night or heavy occlusion, where the one-shot processing often fails to capture the detailed structure of lanes [19, 32]. For instance, lane markings may be partially obscured by other vehicles or faded paint, or harsh lighting conditions, making it difficult for a single forward pass to disambiguate foreground lane pixels from background + +clutter. Although existing designs incorporate multi-level features, they frequently fuse these features all at once—a practice that overlooks the natural progression from high-level contextual cues (e.g., overall road geometry) down to fine-grained details (e.g., lane boundary edges). As a result, the model lacks a mechanism to iteratively refine uncertain or erroneous predictions. This absence of a hierarchical, coarse-to-fine strategy not only limits the ability to resolve ambiguities in the lane positions but also exacerbates the impact of noise or occlusion in early feature extraction. + +A straightforward idea to mitigate the shortcomings is to exploit diffusion models [11, 15, 17, 18, 46, 47], which iteratively refine an initial, coarse prediction by gradually "denoising" it over multiple steps (Figure 1). This iterative mechanism naturally lends itself to self-correction and improved lane delineation. However, directly applying a classic "hot" diffusion process to anchors in lane detection [6] introduces two significant challenges. First, random noise in anchor positions can spread them into unrealistic lane shapes—unlike objects in detection tasks, lanes must follow a predictable structure on the ground plane [19, 22, 32, 40, 53]. Second, simply conditioning on the same image at every timestep fails to exploit a genuine coarse-to-fine mechanism in the image space, causing each iterative step to generate overly similar estimates. + +To overcome these issues, we propose CDiffLane, a novel framework built on a cold diffusion paradigm that replaces random corruption with a deterministic degradation process tailored for lane anchors. Rather than injecting noise that can drive anchor positions astray, our method progressively transitions from predefined lane anchors toward more refined targets, preserving strong lane priors throughout the diffusion (Figure 1(c)). Besides, we address the problem of repeated conditioning by introducing multiresolution inputs, enabling our model to progress from lower-resolution, shape-focused representations to higher-resolution details. Finally, we adopt a learnable $\alpha$ schedule in place of standard variance schedules, allowing our system to dynamically control how much each refinement step adjusts the lane anchor position. These enhancements collectively form an efficient, iterative coarse-to-fine pipeline for both training and inference, yielding lane predictions with higher accuracy (particularly under tighter IoU thresholds) and fewer false positives in cross-scenario tests. + +We demonstrate the effectiveness of CDiffLane on multiple benchmarks. On the challenging CULane dataset [28], our method achieves substantial gains over state-of-the-art baselines in $F_{1}$ at 75 and shows notable stability improvements across diverse road conditions. These results underscore the robustness and adaptability of CDiffLane, confirming that a careful blend of cold diffusion, anchored priors, and a progressive conditioning strategy can push the boundaries of lane detection performance. + +Our contributions are summarized as follows: + +- We are the first, to the best of our knowledge, to leverage a generative diffusion model for lane detection, using deterministic degradations guided by empirical lane priors. +- We introduce resolution scheduling and a learnable $\alpha$ that exploits the diffusion process, enabling the model to emphasize different feature levels at each iterative step. +- Our method demonstrates remarkable stability and outperforms state-of-the-art baselines on both CULane and TuSimple [1] datasets, particularly under higher $F_{1}$ thresholds and in challenging cross-scenario evaluations. + +# 2. Related Work + +# 2.1. Lane Detection + +The methods for lane detection can be categorized into segmentation-based, parameter-based and anchor-based approaches. In segmentation-based methods [29, 51, 52], lane lines are first predicted as dense pixel masks and then fitted into continuous curves, often requiring post-processing to enforce geometric consistency. Parameter-based methods [9, 23, 41] represent lanes using a set of curve parameters, enabling fast inference but exhibiting low tolerance to parameter errors. Meanwhile, anchor-based methods [19, 22, 32, 40, 53] rely on predefined line or row anchors and regress offsets to match lane shapes. By leveraging priors on lane geometry, they can isolate relevant features more efficiently—typically through an IoU-based loss that aligns predicted anchors with ground truth. For instance, Line-CNN [19] and LaneATT [40] incorporate line anchors and attention mechanisms to aggregate global information, while CLRNet [53] adopts learnable anchors and fuses multi-scale representations via a Feature Pyramid Network. Although these anchor-based strategies are generally faster and simpler than segmentation-based methods, they often struggle in highly varied or complex scenarios (e.g., extreme weather, occlusion, or sharply curved roads), where fixed anchors fail to adapt to lane shape variations. + +# 2.2. Diffusion Models + +Diffusion models [3, 6, 11, 15, 17, 18, 47] iteratively refine coarse inputs through a forward corruption process and a corresponding learned denoising procedure, enabling them to model complex data distributions with remarkable flexibility. While originally conceived for image generation, diffusion-based techniques have expanded to downstream tasks such as detection [6, 10], segmentation [2, 44], and other structured prediction problems [11, 47]. A large body of work adopts what is termed "hot diffusion," in which data are progressively corrupted by injecting Gaussian noise at each forward step [14, 38]. In this paradigm, the model learns to iteratively remove noise, moving from a random or partially disordered state back toward a clean sample. + +However, in tasks with strongly constrained structures, hot diffusion can disrupt geometric priors by introducing random noise. Cold diffusion [3, 25, 49] addresses this by employing deterministic degradations, such as interpolation or masking, instead of stochastic noise, and then learning a restoration process. This avoids unrealistic perturbations and proves more suitable for semi-rigid configurations like molecular shapes. To our knowledge, no previous work has applied diffusion to lane detection, because hot diffusion, with its unconstrained noise injection, is ill-suited for such tasks. By starting from structured "anchors" rather than random noise, CDiffLane iteratively refines its estimates with fewer nonsensical deviations, improving both stability and efficiency in this highly constrained setting. + +# 3. Proposed Approach + +Throughout this paper, we denote scalars by italic lowercase letters (e.g., $x$ ), vectors by bold lowercase letters (e.g., $\mathbf{x}$ ), and matrices by bold uppercase letters (e.g., $\mathbf{X}$ ). + +# 3.1. Preliminaries + +Lane Representation. In 2D lane detection, as lanes typically appear in the lower part of an image, the input image is cropped before being annotated. A lane is then represented as a sequence of $N(x,y)$ coordinates, denoted by $P = \{(x_1,y_1),\ldots ,(x_N,y_N)\}$ , where the $y$ coordinates of these points are uniformly distributed along the vertical axis of the cropped image, expressed as $y_{i} = \frac{H}{N - 1} * i$ where $H$ represents the height of the cropped image. Consequently, each $x_{i}$ value corresponds to a specific $y_{i}$ , establishing a clear mapping of lane points along the vertical axis. To reduce complexity, we parametrize each lane $P$ by five parameters $(x_{p},y_{p},\theta_{p},l,\delta_{x})$ rather than predicting all $\{(x_i,y_i)\}$ coordinates directly. Here, $(x_{p},y_{p})$ specifies the lane's starting position, $\theta_p$ is the initial slope, and $l$ indicates the total lane length. The term $\delta_x$ captures incremental horizontal shifts at each sampled vertical step, which effectively modeling how the lane deviates from the naive straight line defined by the starting position, slope and length. This parameterization both simplifies learning and preserves sufficient flexibility to represent curved lane shapes. + +Diffusion Model. While early diffusion methods typically rely on adding Gaussian noise in forward process [14, 36-38], recent work reveals that this step need not be noisy. In particular, cold diffusion [3] generalizes the notion of corruption to include any suitable degradation. Formally, rather than injecting noise, one defines a deterministic operator $D$ that systematically degrades an initial sample $\mathbf{z}_0$ . At each timestep $t \in \{1,2,\dots,T\}$ , the forward process is: + +$$ +\mathbf {z} _ {t} = q \left(\mathbf {z} _ {t} \mid \mathbf {z} _ {0}\right) = D \left(\mathbf {z} _ {0}, t\right). \tag {1} +$$ + +This flexibility allows diffusion models to better align with task-specific structural constraints without the restriction to + +noise injection. For instance, the corruption can be tailored such as blurring or interpolating. On the other hand, there exists a restoration operator $R$ that inverts $D$ , satisfying: + +$$ +R \left(\mathbf {z} _ {t}, t\right) = \mathbf {z} _ {0}. \tag {2} +$$ + +In real-world scenarios, $R$ is typically implemented as a neural network with parameters $\theta$ . During training, $R_{\theta}$ takes the degraded latent sample $\mathbf{z}_t$ as input and is trained to predict the original sample $\mathbf{z}_0$ . The training objective is commonly defined as the $\ell_2$ loss between the model's prediction $\hat{\mathbf{z}}_0$ and $\mathbf{z}_0$ : + +$$ +\mathcal {L} _ {o b j} = \frac {1}{2} \| R _ {\theta} \left(\mathbf {z} _ {t}, t\right) - \mathbf {z} _ {0} \| ^ {2}. \tag {3} +$$ + +At inference stage, $\mathbf{z}_0$ is reconstructed iteratively from the degraded sample $\mathbf{z}_t$ using the model $R_{\theta}$ , i.e., $\mathbf{z}_T \to \mathbf{z}_{T-1} \to \dots \to \mathbf{z}_0$ . The update rule is as follows: + +$$ +\mathbf {z} _ {t - 1} = \mathbf {z} _ {t} - D \left(R _ {\theta} \left(\mathbf {z} _ {t}, t\right), t\right) + D \left(R _ {\theta} \left(\mathbf {z} _ {t}, t\right), t - 1\right). \tag {4} +$$ + +# 3.2. CDiffLane Architecture + +Anchor-Based Diffusion Design. To address the challenges in lane detection, we propose a novel approach utilizing a diffusion model specifically designed to refine lane estimations progressively. Figure 2 illustrates the proposed CDiffLane framework. Unlike the approach in DiffusionDet [7], which applies Gaussian noise to degrade the ground-truth bounding box, our method aligns the ground-truth lane with predefined lane anchors. A lane anchor, represented as a straight line defined by parameters $(x_{p},y_{p},\theta_{p})$ , serves as a structured reference point to guide the detection process. In the lane diffusion process, we initiate with a collection of lane anchors $\mathbf{A} = \{\mathbf{a}_0,\mathbf{a}_1,\dots ,\mathbf{a}_{N_s - 1}\}$ where $N_{s}$ represents the number of lane anchors. These anchors are initially distributed uniformly from left to right across the image, adhering to a slope that aligns with the general appearance of road lanes. Throughout the training process, these anchors dynamically update to fit the data distribution. The primary goal of this approach is to train a lane diffusion model $R_{\theta}$ that iteratively refines the model's predictions to converge on the target lane structure. The refinement process follows a series of diffusion steps $T$ , formulated as follows: + +$$ +\mathbf {A} ^ {T} \xrightarrow {R} \mathbf {A} ^ {T - 1} \xrightarrow {R} \dots \xrightarrow {R} \mathbf {A} ^ {0}. \tag {5} +$$ + +Our method integrates a continuous forward degradation process, leveraging a predefined anchor $\mathbf{a}$ and a ground-truth lane $\mathbf{a}_{gt}$ . The degradation process is defined as: + +$$ +\begin{array}{l} \mathbf {a} ^ {t} = D _ {\alpha} \left(\mathbf {a} _ {g t}, t\right) \tag {6} \\ = \alpha_ {t} \mathbf {a} _ {g t} + (1 - \alpha_ {t}) \mathbf {a}, \text {w i t h} \alpha_ {t} \in [ 0, 1 ]. \\ \end{array} +$$ + +![](images/ef99eccf52c3c6e896cc81ae141f53ffa276130acda880d5cb49721e085d3cbc.jpg) +Figure 2. Flowchart of the proposed method. In the training phase, the model processes a degraded image and a lane degraded according to timestep $t$ , based on the model's initial lane predictions, and outputs a refined version. During inference, the model begins with a predefined set of lane anchors and iteratively refines lane locations as the input image becomes progressively clearer. + +Each anchor in $\mathbf{A}$ is matched to the closest ground-truth lane in the image. The ground-truth lane first extracts its $(x, y, \theta)$ components and duplicates them to match the number of anchors before the process begins. Here, $\mathbf{a}^t$ interpolates between the ground truth and the lane anchor, with $\alpha_t$ designed to decrease monotonically over time. This progressive shift ensures that the degraded lane representation transitions toward the anchor as time progresses. By employing this forward degradation strategy, the model initiates diffusion from an organized and semantically meaningful starting point, thereby avoiding the disordered noise that typically hampers performance. + +In our iterative refinement process, our lane diffusion model $R_{\theta}$ is specifically designed to refine lane alignment and structure at each step $t$ . Given a time step $t$ and an image $\mathbf{I}$ , we first rescale $\mathbf{I}$ to different resolutions based on $t$ to obtain $\mathbf{I}_t$ . Next, we input a degraded lane $\mathbf{A}^t$ along with the rescaled image $\mathbf{I}_t$ into $R_{\theta}$ to obtain the parameterized lane $\mathbf{A}^{0.2}$ . At each stage, the model provides the refined lane prediction as follows: + +$$ +\Delta \mathbf {A} ^ {t}, \mathbf {I} ^ {t}, \delta^ {t}, \mathbf {c} ^ {t} = R _ {\theta} (\mathbf {A} ^ {t}, \mathbf {I} _ {t}), \tag {7} +$$ + +$$ +\hat {\mathbf {A}} ^ {0} = \mathbf {A} ^ {t} + \Delta \mathbf {A} ^ {t}, \tag {8} +$$ + +where $\mathbf{c}^t$ represents the confidence score associated with the lane prediction $\mathbf{A}^t$ at time $t$ . Notably, only the final step values of $\mathbf{l}^t$ , $\delta^t$ and $\mathbf{c}^t$ are retained, while intermediate steps discard them. This iterative refinement enables + +the model to gradually enhance the lane estimation's accuracy and robustness, yielding high-confidence predictions that align closely with the true lane structure. + +Resolution Scheduling and Learnable $\alpha$ . One of the key advantages of diffusion models in image generation is their ability to learn image structures in a progressive, coarse-to-fine manner. This gradual refinement enables the model to incrementally capture both low-level and high-level features, resulting in a highly detailed final output. However, in previous diffusion models for perception tasks [7, 34, 35, 45], the guidance for the diffusion process comes solely from a static conditioning input. When this conditioning input remains fixed throughout the process, the model tends to learn the same feature representations at each step. This limitation hinders the model's capacity to adaptively refine specific details in the image, as it continuously relies on unchanged information instead of dynamically adjusting to evolving input data. + +To address this limitation, we propose a resolution scheduling mechanism for our lane detection model, introducing input images of varying resolution at each diffusion step. Specifically, we define a scheduling parameter $\gamma_{t}$ , representing the factor by which the input image resolution is reduced at step $t$ . Formally, the input image will be scaled down by a factor of $\gamma_{t}$ and then scaled back, resulting in a lower quality effect. This parameter works with a resolution adjustment function, denoted $D_{\gamma}$ , with the adjusted image at step $t$ represented as $\mathbf{I}_t$ . Intuitively, $\gamma_{t}$ is designed to decrease gradually as $t$ approaches zero, allowing the model + +to begin with blurred images and refine them progressively. This process allows the model to focus on high-level features in the early stages when $t$ is large and to focus on finer, lower-level details as $t$ decreases. This coarse-to-fine strategy encourages the model to incrementally improve its predictions, achieving a detailed and accurate final result. + +Additionally, we introduce a learnable parameter $\alpha_{t}$ in Eq. 6, which determines the degradation ratio between the ground truth and a predefined anchor at each step. We interpret the difference in $\alpha$ values between consecutive steps as the variation magnitude $v_{t}$ , defined as $v_{t} = \alpha_{t} - \alpha_{t-1}$ . This variation magnitude is integrated into the objective function by weighting the loss at each timestep, allowing the model to adaptively adjust its learning process. For cases where the model struggles to generate precise predictions, a smaller variation magnitude reduces the impact of errors on the objective function, mitigating excessive corrections. In particular, when the input image resolution is too low to provide meaningful information, the model can minimize unnecessary modifications by dynamically reducing the magnitude. This adaptive mechanism enhances robustness across varying image qualities, improving the model to prioritize relevant features more effectively. Through this learnable adjustment, the model achieves a more context-aware refinement process, leading to improved accuracy and resilience in complex visual scenarios. + +# 3.3. Training and Inference + +Training Paradigm. Our training paradigm involves training a $R_{\theta}$ to predict the ground-truth $\mathbf{A}^0$ from predefined anchors. The pseudocode of the training procedure is provided in Alg. 1 in the Appendix. Specifically, given a degraded version of $\mathbf{A}^t$ derived from Eq. 6 and the corresponding $\mathbf{I}_t$ at timestep $t$ , the goal is to reconstruct the ground-truth $\mathbf{A}^0$ . A key challenge in this process is sampling drift in diffusion models—i.e., the misalignment between the training and sampling data distributions [5, 7, 17]. To mitigate this issue, following the approach in [17], we construct $\mathbf{A}^t$ based on the model's predictions rather than relying solely on the ground truth. More precisely, we first generate an initial estimate $\hat{\mathbf{A}}^0$ by feeding the anchor $\mathbf{A}^T$ and $\mathbf{I}_T$ into the detector. We then derive $\mathbf{A}^t$ by incorporating $\hat{\mathbf{A}}^0$ into Eq. 6, ensuring a more stable and consistent training process. + +Training Loss. The output of the lane detector consists of $N_{s}$ predicted lanes, each with an associated confidence score. During training, ground-truth lanes are dynamically assigned to one or more of the closest predicted lanes using the Hungarian algorithm. Following the approach outlined in [53], the overall training loss comprises a classification loss $\mathcal{L}_{cls}$ and a regression loss $\mathcal{L}_{reg}$ , which account for lane existence and localization accuracy, respectively. Specifically, $\mathcal{L}_{cls}$ is a focal loss used for classification, as most predicted lanes remain unassigned and are labeled as 0. This + +design encourages the model to focus on assigned lanes, thereby enhancing prediction reliability. The regression loss $\mathcal{L}_{reg}$ consists of two components $\mathcal{L}_{xytl}$ and $\mathcal{L}_{iou}$ . $\mathcal{L}_{xytl}$ is a smooth $\ell_{1}$ applied to lane parameters $(x_{p},y_{p},\theta_{p},l)$ . $\mathcal{L}_{iou}$ is the lane IoU loss, computed over the sequence of lane points $P$ . Lane points are first converted using the parameter set $(x_{p},y_{p},\theta_{p},l,\delta_{x})$ as described in Section 3.1. More details on lane IoU loss can be found in [53]. + +Additionally, we incorporate the variation magnitude $v_{t}$ into the training loss function. During training, the regression loss is weighted by $v_{t}$ , enabling the model to dynamically adjust the influence of regression at different timesteps. The total training loss is formulated as: + +$$ +\begin{array}{l} \mathcal {L} _ {\text {t o t a l}} = \lambda_ {\text {c l s}} \mathcal {L} _ {\text {c l s}} + v _ {t} \mathcal {L} _ {\text {r e g}} \tag {9} \\ = \lambda_ {c l s} \mathcal {L} _ {c l s} + v _ {t} \left(\lambda_ {x y t l} \mathcal {L} _ {x y t l} + \mathcal {L} _ {i o u}\right), \\ \end{array} +$$ + +where $\lambda_{cls}$ and $\lambda_{xytl}$ are weighting coefficients. By applying the variation magnitude $v_{t}$ to the lane regression loss, the model can dynamically modulate the impact of each step, improving its ability to handle the progressive nature of the lane refinement process. + +Inference Procedure. The inference process in our method (Alg. 2 in the Appendix) adopts a diffusion-based sampling workflow designed to iteratively refine initial anchor-based estimations. We begin by placing lane anchors on a blurred version of the input image, ensuring that the model concentrates on overarching lane structures before fine details. At each subsequent timestep, clearer versions of the image are introduced—gradually increasing resolution or reducing blurring—while the lane anchors and offsets are updated to converge toward the true lane positions. This progressive refinement framework allows the model to incorporate high-level contextual cues from earlier, coarser steps and then focus on finer-grained corrections in later stages. As a result, anchor positions become more precise, enabling the model to consistently align with lane boundaries despite challenges like occlusion, road curvature, or adverse lighting conditions. Ultimately, the diffusion-based sampling procedure guides the model from an initial, anchor-driven guess to an accurate representation of real lane geometry. + +# 4. Experiment + +# 4.1. Experimental Setup + +Datasets. We evaluate our method on two widely-recognized lane detection benchmarks, CULane [28] and TuSimple [1]. CULane features 88.9k training images, 9.7k validation images, and 34.7k test images (each $1640 \times 590$ pixels), spanning diverse environments such as urban streets, rural roads, and dense traffic. Its test set is categorized into nine conditions, e.g., crowded, night, and no line, to rigorously assess performance under varying real-world scenarios. In contrast, TuSimple is dedicated to highway + +settings, comprising 3.3k training images, 0.4k validation images, and 2.8k test images, all at $1280 \times 720$ resolution. + +Evaluation metrics. Following previous works [39, 53], we assess lane prediction accuracy using the F1-measure for the CULane dataset, calculated via the Intersection-over-Union (IoU) between predicted and ground-truth lanes. In our experiments, we report $F_{1}@50$ and $F_{1}@75$ , corresponding to IoU thresholds of 0.5 and 0.75, respectively. Notably, the higher threshold (0.75) serves as a more rigorous test for precise lane localization, emphasizing improvements in fine-grained accuracy. Additionally, we include the $mF_{1}$ score, which represents the average score over $F_{1}@50$ to $F_{1}@95$ . For the TuSimple dataset, a predicted lane is considered correct if more than $85\%$ of its pixels overlap with those in the ground truth. Due to space constraints, details are provided in the Appendix. + +Implementation Details. We build our restoration network $R_{\theta}$ on CLRNet [53], using a backbone (e.g., ResNet [13] or DLA [50]) plus an FPN [21] to extract multi-scale features. These features are then cropped according to the degraded lane $\mathbf{A}^t$ before generating final lane predictions and confidence scores. Due to space constraints, full implementation details including training configurations, data preprocessing, and inference procedures are provided in the Appendix. + +# 4.2. Quantitative Results + +Performance on CULane. As shown in Table 1, our CDiffLane outperforms the state-of-the-art methods on the CULane test set, achieving an $F_{1}$ score of 81.26 at $50\%$ IoU $(F_{1}@50)$ and 65.31 at $75\%$ IoU $(F_{1}@75)$ . Compared to the baseline (CLRNet with a DLA-34 backbone), we obtain a 0.79-point gain on $F_{1}@50$ and a 2.53-point improvement on $F_{1}@75$ . While our result shows a modest 0.13-point improvement over GSENet at $F_{1}@50$ , the margin increases to 1.23 points at $F_{1}@75$ , highlighting stronger localization precision. The smaller gain at $F_{1}@50$ aligns with previous oracle experiments [53], which found that most lanes were already detected correctly, though misclassification remained a challenge. As our cold diffusion framework focuses on refining geometric placement rather than classification, improvements at this lower threshold are limited. However, the stricter $F_{1}@75$ threshold demands more precise lane alignment, where iterative refinement proves effective. By progressively correcting misaligned anchors, CDiffLane significantly enhances performance at higher IoU thresholds. + +Moreover, Table 1 also manifest that our method reduces false positives (FP) and demonstrates robust performance in cross-scenario tests (e.g., "night" or "crowded" conditions). We attribute this to the iterative nature of our diffusion process, which converges to plausible lane boundaries over multiple steps, rather than relying on a single-pass estimate. These findings highlight two key advantages of em + +ploying cold diffusion with anchor priors. First, the iterative design refines uncertain anchor positions step by step, reducing FP rates. Second, by using deterministic degradations instead of random noise, we preserve structural consistency for lanes under challenging conditions (e.g., severe occlusions or tight curves). Hence, the cold diffusion approach offers notable benefits in aligning lane detection models with stricter precision requirements. + +Performance on TuSimple. Table 2 compares our approach against existing methods on the TuSimple benchmark. While the slight performance gaps suggest an overall trend toward saturation on this dataset, CDiffLane still achieves a notably high $F_{1}@50$ of 97.71. In addition, our method reaches a $1.62\%$ false positive rate (FP), improving upon previous scores by 0.39 points. These results indicate that our framework is capable of pushing the limits of accuracy in highway-specific lane detection tasks, even as top-line metrics converge. + +Despite a somewhat higher false negative rate (FN) compared with certain competitors, it is important to recognize the trade-off inherent in lane detection: methods targeting a lower FN often forecast more lanes to ensure coverage, thus inflating FP. In contrast, $CDiffLane$ maintains a more balanced strategy, prioritizing minimal spurious lane predictions (FP) while retaining competitive FN performance. This trade-off proves advantageous in practice, as high FP can degrade driving decisions by incorrectly signaling additional lanes, whereas our method focuses on achieving strong localization with fewer extraneous detections. Consequently, the overall robustness of $CDiffLane$ ensures that missed lanes are kept at acceptable levels, while false positives are substantially reduced, leading to a favorable balance for real-world autonomous driving scenarios. + +# 4.3. Qualitative Results + +Figure 3 demonstrates example outputs on the CULane test set under varied and difficult conditions, such as low visibility, shadows, and pronounced road curvature. Traditional anchor-based approaches (e.g., LaneATT, CLRNet) frequently produce false positives (highlighted in red), as they rely on one-shot anchor initialization. This single-pass strategy struggles to account for large contextual shifts or unusual lane shapes, leading to higher error rates when lanes deviate from predefined templates. In contrast, our method adopts a multi-step refinement process, iteratively refining lane predictions as it receives progressively more detailed image cues. This feedback loop not only captures richer global context but also compensates for suboptimal anchor initialization. Consequently, false positives diminish, and overall accuracy rises, demonstrating that an iterative diffusion approach can better handle diverse road layouts and environmental conditions, even in complex scenarios like nighttime driving or curved lane segments. + +
MethodBackboneVenuemF1↑F1@50↑F1@75↑Normal↑Crowd↑Dazzle↑Shadow↑Noline↑Arrow↑Curve↑Cross↓Night↑
CondLane [22]ResNet-101ICCV'2154.8379.4861.2393.4777.4470.9380.9154.1390.1675.21120174.80
GANet [42]ResNet-101CVPR'22-79.63-93.6778.6671.8278.3253.3889.8677.37135273.85
CLRNet [53]ResNet-101CVPR'2255.5580.1362.9693.8578.7872.4982.3354.5089.7975.57126275.51
CLRNet [53]DLA-34CVPR'2255.6480.4762.7893.7379.5975.3082.5154.5890.6274.13115575.37
CondLSTR [8]ResNet-101ICCV'23-80.77-94.1779.9075.4380.9955.0090.9773.83104775.11
CLRerNet [16]ResNet-101WACV'24-80.9164.3093.9180.0372.9882.9255.7390.5374.67111376.13
CLRerNet [16]DLA-34WACV'24-81.1264.0794.0280.2074.4183.7156.2790.3974.67116176.53
GSENet [39]ResNet-101AAAI'2456.5380.8464.2394.0579.9074.9482.2155.6390.78-116476.08
GSENet [39]DLA-34AAAI'2456.4581.1364.0893.9180.3076.3683.4156.2590.36-103676.26
Lane2Seq(anchor) [54]ViT-BaseCVPR'24-79.27-93.1177.4373.2579.4653.7490.0272.44117375.12
CDiffLaneDLA-34-58.0681.2665.3193.9380.2672.6484.0455.2090.8574.8689376.60
+ +![](images/cf101586e8a557bde36b81f123d4e3392ad785b0e5a228c0c450cf9f16db2f1b.jpg) +Figure 3. Visualization results of LaneATT, CLRerNet, and our method on the CULane. Yellow lines represent correctly predicted lanes (IoU above the threshold), classified as True Positives (TP). Red lines indicate mispredicted lanes (IoU below the threshold), categorized as False Positives (FP). Green lines denote False Negatives (FN), representing lanes present in the ground truth but missed by the model. + +Table 1. Comparison of leading lane detection models on the CULane test set. All scenarios use the $F_{1}$ score at 0.5 IoU, except Cross, which reports only false positives. Bold highlights the best result, and underlines mark the second-best. + +
MethodBackboneF1↑Acc↑FP↓FN↓
SCNN [30]VGG1695.9796.536.171.80
RESA [52]ResNet-3496.9396.823.632.48
LaneATT [40]ResNet-12296.0696.105.642.17
CondLane [22]ResNet-10197.2496.542.013.50
CANet [48]ResNet-3497.4496.662.322.79
CDiffLaneResNet-3497.7196.671.623.01
+ +# 4.4. Ablation study + +To investigate how each component of our framework contributes to the final performance, we conduct a series of ablation experiments on the CULane dataset. The $F_{1}$ scores at 50% and 75% IoU thresholds are reported in Table 3, offering a clear comparison of different diffusion paradigms (hot vs. cold) and various image degradation strategies. + +Hot vs. Cold Diffusion. We first evaluate the commonly used "hot diffusion", which adds Gaussian noise to lane anchors and initiates the denoising process from a Gaussian distribution. While this technique has shown effectiveness in object detection or segmentation [2, 7], it proves + +Table 2. Quantitative Result on TuSimple. + +
DiffusionImage Degrad.F1@50F1@75
HotNone79.2560.77
ColdNone80.8864.56
ColdResolution81.2665.31
ColdDCT81.2965.17
ColdGaussian Blur81.0664.78
ColdNoise80.9364.52
+ +Table 3. Ablation study on image degradation and diffusion version in CULane dataset. + +suboptimal for lane detection: it yields $F_{1}$ @50/F1@75 of 79.25/60.77, underperforming even the baseline CLRNet. We attribute this to hot diffusion's random noise initialization, which often places lane anchors in unrealistic regions. By contrast, our "cold diffusion" begins with a structured lane anchor that better reflects plausible lane locations, thereby reducing extraneous anchor positions. As a result, switching to cold diffusion alone raises performance to 80.88/64.56 ( $F_{1}$ @50/F1@75), highlighting the value of a deterministic, anchor-driven starting point. + +
Sche. αSche. γLrn. αLrn. γF1@50F1@75
✓✓81.0465.11
✓✓80.9364.95
✓✓81.2665.31
✓✓81.0965.03
+ +Table 4. Ablation study on $\alpha$ and $\gamma$ . 'Sche.' abbreviates schedule, and 'Lrn.' abbreviates learnable. The schedule $\alpha$ follows a cosine scheduling, while $\gamma$ . uses a linear schedule. + +Image Degradation Strategies. Next, we examine how various image degradation operations affect the coarse-to-fine refinement process. Beyond using no explicit image degradation ("None"), we test: + +1. Resolution: Downsampling and then upsampling each image to progressively introduce clarity, +2. DCT: Filtering high-frequency components via discrete cosine transform before reapplying them in later steps, +3. Gaussian Blur: Convolving the image to diffuse edges and reduce sharp details, +4. Noise: Injecting random noise at each step. + +Our results show that Resolution and DCT yield the largest gains among all degradations (see Table 3). The former achieves 81.26/65.31, while the latter reaches 81.29/65.17 at $F_{1}@50/F_{1}@75$ . Both techniques prioritize the reconstruction of coarse shapes and progressively refine finer details over multiple iterations. This approach aligns well with cold diffusion's anchor-based iterative refinement strategy, ensuring a structured and coherent recovery process. In contrast, Gaussian Blur demonstrates notably inferior performance despite its superficial resemblance to Resolution. We attribute this disparity to the critical role of lane boundaries in lane detection. While Resolution maintains these essential edge features, Gaussian Blur compromises them by smearing and weakening the defining contours, thereby reducing the model's capacity to accurately identify lane markings. Similarly, the introduction of Noise at each iteration disrupts the localized, anchor-centric design inherent in CLRNet, further exacerbating structural degradation and leading to suboptimal $F_{1}$ scores. Overall, our findings highlight the necessity of preserving lane geometry and contextual integrity in the diffusion process. They further validate that the careful selection of image degradation operations can significantly enhance the accuracy of cold diffusion-based lane detection. These insights underscore the importance of tailored preprocessing strategies to optimize model performance in real-world applications. + +Learnable $\alpha$ vs. Scheduled $\gamma$ . We further investigate the effect of treating the parameters $\alpha$ (lane-anchor interpolation) and $\gamma$ (image-resolution factor) as either learnable or fixed/scheduled. Our findings show that the best performance arises when $\alpha$ is learnable and $\gamma$ is maintained on + +a fixed schedule, giving 81.26/65.31 for $F_{1}@50/F_{1}@75$ . A learnable $\alpha$ dynamically adjusts the weighting of each iterative refinement step, amplifying confident predictions while minimizing the influence of uncertain ones. In contrast, making $\gamma$ learnable disrupts the model's training consistency; the image input then shifts unpredictably at each step, scattering the model's attention and degrading fine-grained alignment. Consequently, scheduling $\gamma$ in a stable, predetermined way emerges as the more effective strategy, balancing coherent refinement with the flexibility needed for high-precision lane detection. We perform an experiment to examine how $mF_{1}$ varies with the number of inference steps and the resulting speed (FPS). As shown in Fig. 4, the iterative strategy indeed improves performance but lowers the number of tasks the model can process per second. While this trade-off may be acceptable for tasks such as high-density map construction, it can be less ideal in time-critical settings. Notably, even when reducing the iterative steps to just one—yielding an FPS above 140—our CDiffLane still surpasses GSENet, highlighting its robustness under lower iteration budgets. + +# 4.5. Limitation + +While CDiffLane improves detection accuracy through iterative refinement, the diffusion process naturally adds extra inference steps. One promising approach to alleviate this overhead is progressive distillation, where a student model inherits the diffusion model's knowledge and produces reliable predictions in fewer iterations, thus preserving accuracy while lowering computational cost. As the first to integrate diffusion into lane detection, our main objective has been to raise accuracy; future efforts will focus on optimizing for the real-time scenarios. Notably, even at a single iterative step (exceeding 140 FPS), CDiffLane still surpasses GSENet, demonstrating strong performance under reduced iteration budgets (see Appendix C for details). + +# 5. Conclusion + +In this work, we propose CDiffLane, which integrates the cold diffusion framework into anchor-based lane detection. To enhance the model's coarse-to-fine prediction, we introduce resolution scheduling and a learnable $\alpha$ , which provide sequential information that enables the model to iteratively refine lane predictions with greater precision. Extensive experiments demonstrate that CDiffLane outperforms state-of-the-art baselines in accuracy, establishing its effectiveness in lane detection tasks. As the first framework to integrate cold diffusion into a visual perception task, we anticipate that this architecture can be extended to other perception-based applications, further advancing the field of computer vision. + +# Acknowledgments + +This work is partially supported by the National Science and Technology Council (NSTC), Taiwan (Grants: NSTC-112-2221-E-A49-059-MY3, NSTC-112-2221-E-A49-094-MY3, NSTC-112-2628-E-002-033-MY4), and by the Center of Data Intelligence, National Taiwan University (Grants: 114L900901, 114L900902, 114L900903), through the Higher Education Sprout Project, MOE, Taiwan. Additional support is provided by the Industry-Academia Innovation School, NYCU (Project No. 113UA10024-2), under the MOE's National Key Fields Program, and the Hon Hai Research Institute. + +# References + +[1] Tusimple benchmark. https://github.com/TuSimple/tusimple-benchmark/.2,5 +[2] Tomer Amit, Tal Shaharbany, Eliya Nachmani, and Lior Wolf. Segdiff: Image segmentation with diffusion probabilistic models. arXiv preprint arXiv:2112.00390, 2021. 2, 7 +[3] Arpit Bansal, Eitan Borgnia, Hong-Min Chu, Jie Li, Hamid Kazemi, Furong Huang, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Cold diffusion: Inverting arbitrary image transforms without noise. NeurIPS, 2024. 2, 3 +[4] Qing Chang and Yifei Tong. A hybrid global-local perception network for lane detection. In AAAI, 2024. 1 +[5] Sitan Chen, Sinho Chewi, Jerry Li, Yuanzhi Li, Adil Salim, and Anru Zhang. Sampling is as easy as learning the score: theory for diffusion models with minimal data assumptions. In ICLR, 2023. 5 +[6] Shoufa Chen, Peize Sun, Yibing Song, and Ping Luo. Diffusiondet: Diffusion model for object detection. In ICCV, 2023. 1, 2 +[7] Shoufa Chen, Peize Sun, Yibing Song, and Ping Luo. Diffusiondet: Diffusion model for object detection. In ICCV, 2023. 3, 4, 5, 7 +[8] Ziye Chen, Yu Liu, Mingming Gong, Bo Du, Guoqi Qian, and Kate Smith-Miles. Generating dynamic kernels via transformers for lane detection. In ICCV, 2023. 7 +[9] Ruochen Fan, Xuanrun Wang, Qibin Hou, Hanchao Liu, and Tai-Jiang Mu. Spinnet: Spinning convolutional network for lane boundary detection. Computational Visual Media, 2019. 2 +[10] Haoyang Fang, Boran Han, Shuai Zhang, Su Zhou, Cuixiong Hu, and Wen-Ming Ye. Data augmentation for object detection via controllable diffusion models. In WACV, 2024. 2 +[11] Shansan Gong, Mukai Li, Jiangtao Feng, Zhiyong Wu, and Lingpeng Kong. DiffuSeq: Sequence to sequence text generation with diffusion models. In ICLR, 2023. 2 +[12] Jie Guo, Zhihua Wei, and Duoqian Miao. Lane detection method based on improved ransac algorithm. In IEEE Twelfth International Symposium on Autonomous Decentralized Systems, 2015. 1 + +[13] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 6, 1 +[14] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. NeurIPS, 2020. 2, 3 +[15] Jonathan Ho, Chitwan Sahara, William Chan, David J Fleet, Mohammad Norouzi, and Tim Salimans. Cascaded diffusion models for high fidelity image generation. Journal of Machine Learning Research, 2022. 2 +[16] Hiroto Honda and Yusuke Uchida. Clernet: improving confidence of lane detection with laneiou. In WACV, 2024. 1, 7 +[17] Yuanfeng Ji, Zhe Chen, Enze Xie, Lanqing Hong, Xihui Liu, Zhaoqiang Liu, Tong Lu, Zhenguo Li, and Ping Luo. Ddp: Diffusion model for dense visual prediction. In ICCV, 2023. 2, 5 +[18] Lingkai Kong, Jiaming Cui, Haotian Sun, Yuchen Zhuang, B Aditya Prakash, and Chao Zhang. Autoregressive diffusion model for graph generation. In ICML, 2023. 2 +[19] Xiang Li, Jun Li, Xiaolin Hu, and Jian Yang. Line-cnn: End-to-end traffic line detection with line proposal unit. IEEE Transactions on Intelligent Transportation Systems, 2019. 1, 2 +[20] Zuo-Quan Li, Hui-Min Ma, and Zheng-Yu Liu. Road lane detection with gabor filters. In ISAI, 2016. 1 +[21] Tsung-Yi Lin, Piotr Dollar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In CVPR, 2017. 6, 1 +[22] Lizhe Liu, Xiaohao Chen, Siyu Zhu, and Ping Tan. Condilanenet: a top-to-down lane detection framework based on conditional convolution. In ICCV, 2021. 2, 7 +[23] Ruijin Liu, Zejian Yuan, Tie Liu, and Zhiliang Xiong. End-to-end lane shape prediction with transformers. In WACV, 2021. 1, 2 +[24] Chunyang Mu and Xing Ma. Lane detection based on object segmentation and piecewise fitting. TELKOMNIKA Indonesian Journal of Electrical Engineering, 2014. 1 +[25] Sergio Naval Marimont, Vasilis Siomos, Matthew Baugh, Christos Tzelepis, Bernhard Kainz, and Giacomo Tarroni. Ensembled cold-diffusion restorations for unsupervised anomaly detection. In MICCAI, 2024. 3 +[26] Davy Neven, Bert De Brabandere, Stamatios Georgoulis, Marc Proesmans, and Luc Van Gool. Towards end-to-end lane detection: an instance segmentation approach. In IEEE intelligent vehicles symposium (IV), 2018. 1 +[27] Jianwei Niu, Jie Lu, Mingliang Xu, Pei Lv, and Xiaoke Zhao. Robust lane detection using two-stage feature extraction with curve fitting. PR, 2016. 1 +[28] Xingang Pan, Jianping Shi, Ping Luo, Xiaogang Wang, and Xiaou Tang. Spatial as deep: Spatial cnn for traffic scene understanding. In AAAI, 2018. 1, 2, 5 +[29] Xingang Pan, Jianping Shi, Ping Luo, Xiaogang Wang, and Xiaou Tang. Spatial as deep: Spatial cnn for traffic scene understanding. In AAAI, 2018. 2 +[30] Angshuman Parashar, Minsoo Rhu, Anurag Mukkara, Antonio Puglielli, Rangharajan Venkatesan, Brucek Khailany, Joel Emer, Stephen W Keckler, and William J Dally. Scnn: + +An accelerator for compressed-sparse convolutional neural networks. ACM SIGARCH computer architecture news, 2017. 7 +[31] Jonah Philion. Fastdraw: Addressing the long tail of lane detection by adapting a sequential prediction network. In CVPR, 2019. 1 +[32] Zequn Qin, Huanyu Wang, and Xi Li. Ultra fast structure-aware deep lane detection. In ECCV, 2020. 1, 2 +[33] Zequn Qin, Pengyi Zhang, and Xi Li. Ultra fast deep lane detection with hybrid anchor driven ordinal classification. IEEE transactions on pattern analysis and machine intelligence, 2022. 1 +[34] Yasiru Ranasinghe, Nithin Gopalakrishnan Nair, Wele Gedara Chaminda Bandara, and Vishal M Patel. Crowddiff: Multi-hypothesis crowd density estimation using diffusion models. In CVPR, 2024. 4 +[35] Saurabh Saxena, Charles Herrmann, Junhwa Hur, Abhishek Kar, Mohammad Norouzi, Deqing Sun, and David J Fleet. The surprising effectiveness of diffusion models for optical flow and monocular depth estimation. CVPR, 2024. 4 +[36] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In ICML, 2015. 3 +[37] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, 2021. +[38] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. NeurIPS, 2019. 2, 3 +[39] Junhao Su, Zhenghan Chen, Chenghao He, Dongzhi Guan, Changpeng Cai, Tongxi Zhou, Jiashen Wei, Wenhua Tian, and Zhihuai Xie. Gsenet: Global semantic enhancement network for lane detection. In AAAI, 2024. 1, 6, 7 +[40] Lucas Tabelini, Rodrigo Berriel, Thiago M Paixao, Claudine Badue, Alberto F De Souza, and Thiago Oliveira-Santos. Keep your eyes on the lane: Real-time attention-guided lane detection. In CVPR, 2021. 1, 2, 7 +[41] Lucas Tabelini, Rodrigo Berriel, Thiago M Paixao, Claudine Badue, Alberto F De Souza, and Thiago Oliveira-Santos. Polylanenet: Lane estimation via deep polynomial regression. In ICPR, 2021. 1, 2 +[42] Jinsheng Wang, Yinchao Ma, Shaofei Huang, Tianrui Hui, Fei Wang, Chen Qian, and Tianzhu Zhang. A keypoint-based global association network for lane detection. In CVPR, 2022. 7 +[43] Chunyu Wei, Hailong Li, Junyi Shi, Guoyang Zhao, Huaiqu Feng, and Longzhe Quan. Row anchor selection classification method for early-stage crop row-following. Computers and Electronics in Agriculture, 2022. 1 +[44] Junde Wu, Rao Fu, Huihui Fang, Yu Zhang, Yehui Yang, Haoyi Xiong, Huiying Liu, and Yanwu Xu. Medsegdiff: Medical image segmentation with diffusion probabilistic model. In MIDL, 2024. 2 +[45] Fei Xie, Zhongdao Wang, and Chao Ma. Diffusiontrack: Point set diffusion model for visual object tracking. In CVPR, 2024. 4 +[46] Minkai Xu, Lantao Yu, Yang Song, Chence Shi, Stefano Ermon, and Jian Tang. Geodiff: A geometric diffusion model for molecular conformation generation. In ICLR, 2022. 2 + +[47] Dongchao Yang, Jianwei Yu, Helin Wang, Wen Wang, Chao Weng, Yuexian Zou, and Dong Yu. Diffsound: Discrete diffusion model for text-to-sound generation. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 2023. 2 +[48] Zhongyu Yang, Chen Shen, Wei Shao, Tengfei Xing, Runbo Hu, Pengfei Xu, Hua Chai, and Ruini Xue. Canet: Curved guide line network with adaptive decoder for lane detection. In ICASSP, 2023. 7 +[49] Hao Yen, François G Germain, Gordon Wichern, and Jonathan Le Roux. Cold diffusion for speech enhancement. In ICASSP, 2023. 3 +[50] Fisher Yu, Dequan Wang, Evan Shelhamer, and Trevor Darrell. Deep layer aggregation. In CVPR, 2018. 6, 1 +[51] Jia-Qi Zhang, Hao-Bin Duan, Jun-Long Chen, Ariel Shamir, and Miao Wang. Houghlanenet: Lane detection with deep hough transform and dynamic convolution. Computers & Graphics, 2023. 2 +[52] Tu Zheng, Hao Fang, Yi Zhang, Wenjian Tang, Zheng Yang, Haifeng Liu, and Deng Cai. Resa: Recurrent feature-shift aggregator for lane detection. In AAAI, 2021. 2, 7 +[53] Tu Zheng, Yifei Huang, Yang Liu, Wenjian Tang, Zheng Yang, Deng Cai, and Xiaofei He. Clrnet: Cross layer refinement network for lane detection. In CVPR, 2022. 1, 2, 5, 6, 7 +[54] Kunyang Zhou. Lane2seq: towards unified lane detection via sequence generation. In CVPR, 2024. 7 \ No newline at end of file diff --git a/ICCV/2025/When Anchors Meet Cold Diffusion_ A Multi-Stage Approach to Lane Detection/images.zip b/ICCV/2025/When Anchors Meet Cold Diffusion_ A Multi-Stage Approach to Lane Detection/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..1ec08a2a42ce4f115d50d437d5c67d54668a24b7 --- /dev/null +++ b/ICCV/2025/When Anchors Meet Cold Diffusion_ A Multi-Stage Approach to Lane Detection/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:034e25dcf1da47e4717de86e6d0e1e939f43409455e7c72976803bcbb56e1616 +size 462847 diff --git a/ICCV/2025/When Anchors Meet Cold Diffusion_ A Multi-Stage Approach to Lane Detection/layout.json b/ICCV/2025/When Anchors Meet Cold Diffusion_ A Multi-Stage Approach to Lane Detection/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..2b563bf59dc538935df0b6eb852f9a9a3dec5370 --- /dev/null +++ b/ICCV/2025/When Anchors Meet Cold Diffusion_ A Multi-Stage Approach to Lane Detection/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:78b2f8fe44989f9e08e528b9a2773d7be7c91865fda90e2a670659f66efd3cd9 +size 427104 diff --git a/ICCV/2025/When Confidence Fails_ Revisiting Pseudo-Label Selection in Semi-supervised Semantic Segmentation/1e3de233-91fb-4c3e-99af-1b078b01bace_content_list.json b/ICCV/2025/When Confidence Fails_ Revisiting Pseudo-Label Selection in Semi-supervised Semantic Segmentation/1e3de233-91fb-4c3e-99af-1b078b01bace_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..f605a4941d52cf82f096f953499b50b0c6e43988 --- /dev/null +++ b/ICCV/2025/When Confidence Fails_ Revisiting Pseudo-Label Selection in Semi-supervised Semantic Segmentation/1e3de233-91fb-4c3e-99af-1b078b01bace_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9034a46e32e08fd3339d54419f7d273f7e96ca6311acbd2cc998082bffc45b90 +size 85987 diff --git a/ICCV/2025/When Confidence Fails_ Revisiting Pseudo-Label Selection in Semi-supervised Semantic Segmentation/1e3de233-91fb-4c3e-99af-1b078b01bace_model.json b/ICCV/2025/When Confidence Fails_ Revisiting Pseudo-Label Selection in Semi-supervised Semantic Segmentation/1e3de233-91fb-4c3e-99af-1b078b01bace_model.json new file mode 100644 index 0000000000000000000000000000000000000000..0a404a8f26da6b9176b71dbd999a15b32f098946 --- /dev/null +++ b/ICCV/2025/When Confidence Fails_ Revisiting Pseudo-Label Selection in Semi-supervised Semantic Segmentation/1e3de233-91fb-4c3e-99af-1b078b01bace_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fd825f5ab945a136e6e0b596726fc1c7ea87d602c543333b8525c4064cf9bef2 +size 109961 diff --git a/ICCV/2025/When Confidence Fails_ Revisiting Pseudo-Label Selection in Semi-supervised Semantic Segmentation/1e3de233-91fb-4c3e-99af-1b078b01bace_origin.pdf b/ICCV/2025/When Confidence Fails_ Revisiting Pseudo-Label Selection in Semi-supervised Semantic Segmentation/1e3de233-91fb-4c3e-99af-1b078b01bace_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..ee6684c24f3ca18a8cc4649a2c29afa0bedbce98 --- /dev/null +++ b/ICCV/2025/When Confidence Fails_ Revisiting Pseudo-Label Selection in Semi-supervised Semantic Segmentation/1e3de233-91fb-4c3e-99af-1b078b01bace_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ab60d3af39851874d1eabd641fab102d66379c991394b6a00996b832f4b1a684 +size 4513157 diff --git a/ICCV/2025/When Confidence Fails_ Revisiting Pseudo-Label Selection in Semi-supervised Semantic Segmentation/full.md b/ICCV/2025/When Confidence Fails_ Revisiting Pseudo-Label Selection in Semi-supervised Semantic Segmentation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..c36a6c84081b713a58ff1f65669e62344ef9aef3 --- /dev/null +++ b/ICCV/2025/When Confidence Fails_ Revisiting Pseudo-Label Selection in Semi-supervised Semantic Segmentation/full.md @@ -0,0 +1,391 @@ +# When Confidence Fails: Revisiting Pseudo-Label Selection in Semi-supervised Semantic Segmentation + +Pan Liu, Jinshi Liu* +Central South University +{pan.liu, ljs11528}@csu.edu.cn + +# Abstract + +While significant advances exist in pseudo-label generation for semi-supervised semantic segmentation, pseudolabel selection remains understudied. Existing methods typically use fixed confidence thresholds to retain high-confidence predictions as pseudo-labels. However, these methods cannot cope with network overconfidence tendency, where correct and incorrect predictions overlap significantly in high-confidence regions, making separation challenging and amplifying model cognitive bias. Meanwhile, the direct discarding of low-confidence predictions disrupts spatial-semantic continuity, causing critical context loss. We propose Confidence Separable Learning (CSL) to address these limitations. CSL formulates pseudo-label selection as a convex optimization problem within the confidence distribution feature space, establishing sample-specific decision boundaries to distinguish reliable from unreliable predictions. Additionally, CSL introduces random masking of reliable pixels to guide the network in learning contextual relationships from low-reliability regions, thereby mitigating the adverse effects of discarding uncertain predictions. Extensive experimental results on the Pascal, Cityscapes, and COCO benchmarks show that CSL performs favorably against state-of-the-art methods. Code and model weights are available at: https://github.com/PanLiuCSU/CSL. + +# 1. Introduction + +Semantic segmentation is a dense prediction task involving per-pixel classification. This technique is widely applied in critical domains such as autonomous driving and medical imaging. However, fully supervised segmentation methods [6, 9, 37] often require extensive high-quality, pixel-level labeled datasets, which can be costly and resource-intensive. In light of these challenges, semi-supervised semantic segmentation [10, 13, 21] has emerged as a promising research + +![](images/4b47241f7928cb6777d59aeb3ae2995bbd48881fe4efdf1e38d6b2d2e67a4aaf.jpg) +Figure 1. prediction in confidence feature space. Blue/red points and upper histogram denote correct/incorrect predictions. The threshold method (dashed box) poorly separates them, while our approach (black curve) achieves clear discrimination. + +area. This approach seeks to leverage a limited number of labeled images alongside a substantial quantity of unlabeled images to improve segmentation performance. + +In the field of semi-supervised semantic segmentation, combining pseudo-labeling with consistency regularization has become a popular strategy [41]. Pseudo-labeling methods [8, 26, 36, 73] generate pseudo-labels for unlabeled samples based on the model predictions to optimize the model iteratively. Consistency regularization methods [29, 40, 41, 71] follow the smoothness assumption [3], encouraging the model to generate consistent predictions under different perturbations of the same sample. The effectiveness of these strategies depends heavily on the quality of the pseudo-labels, and existing methods typically assume a direct correlation between high confidence and high accuracy, relying on the maximum confidence scores to filter or weight the pseudo-labels [23, 49, 51]. + +As shown by the red box in Fig. 1, neural networks often suffer from overconfidence, causing predictions to cluster at extreme confidence levels regardless of their correct + +ness [17]. This issue becomes more pronounced when labeled data is scarce, as the limited annotations are insufficient to effectively calibrate the model's confidence estimates on the vast unlabeled dataset. Such overconfidence presents significant challenges for semi-supervised semantic segmentation. Firstly, it renders confidence scores poor indicators of prediction accuracy, since both correct and incorrect predictions may receive similarly high confidence scores due to miscalibration. Secondly, constrained by the low distinguishability of confidence, existing methods are forced to exclude low-confidence predictions to avoid introducing errors during training [1], but this also results in a lack of supervision in borderline or complex regions of the data, that are typically crucial for model improvement. + +We propose the Confidence Separable Learning (CSL) framework to address the challenge of insufficient discriminability of pseudo-labels. Based on entropy minimization theory, we find that reliable prediction should have both high maximum confidence and significant probability dispersion between non-maximum class probabilities. Therefore, we construct a confidence distribution feature space consisting of maximum confidence and residual deviation to guide the selection process. As shown in Fig. 1, while an overconfident network assigns similar confidence scores to correct and incorrect predictions, these two classes of predictions exhibit a high degree of separability in the confidence distribution feature space. Based on this property, CSL adaptively establishes sample-specific pseudolabel decision boundaries by formulating pseudo-label selection as a convex optimization problem that maximizes inter-class separability. This allows CSL to more adequately extract supervised signals from unlabelled data. + +While CSL enhances pseudo-label reliability, its inherent precision-coverage trade-off results in the persistent exclusion of low-reliability regions from supervision, potentially amplifying regional supervision bias. To address this, we integrate a Trusted Mask Perturbation (TMP) strategy into our CSL framework. By randomly masking subsets of reliable predictions, we force the model to reconstruct semantics from adjacent unreliable regions via contextual dependencies. This process maximizes mutual information between unreliable regions and pseudo-labeled areas, thereby mitigating regional bias induced by selective pseudo-labeling. Our contributions can be summarized as: + +- We demonstrate that pseudo-label selection based only on maximum confidence suffers from suboptimal performance, especially when the network is overconfident. +- We propose a semi-supervised semantic segmentation framework that separates reliable predictions by maximizing separability optimization strategy and constructs effective supervisory signals for unreliable predictions. +- Comprehensive experiments on three benchmark datasets confirm the effectiveness of our proposed method. + +# 2. Related work + +Semi-supervised learning (SSL) [27, 59, 64] aims to enhance model performance by leveraging labeled and unlabeled data. Recent investigations have concentrated on the techniques of self-training [4, 14, 42, 65, 75] and consistency regularization [30, 46, 49, 52, 66]. However, these methods limited by marginal distribution discrepancies, leading to significant erroneous supervision. + +To address this limitation, two main research directions have been explored. The first focuses on optimizing threshold-based pseudo-labeling strategies [5, 61, 67, 72]. SoftMatch [5] models confidence as a Gaussian distribution to determine optimal thresholds. Despite improvements, these methods rely on assumptions that may not be universally applicable, limiting their generalizability. The second seeks additional metrics beyond confidence, such as Monte Carlo dropout [45], ensemble methods [55], and auxiliary discriminators [33]. However, the additional metrics significantly increase the computational overhead. + +Semi-supervised semantic segmentation faces the challenge of generating accurate pixel-level supervisory signals for unlabeled images. Inspired by SSL, many works have adopted similar frameworks [54, 63, 70]. PS-MT [38] uses dual-teacher models to enhance consistency regularization. UniMatch [69] separates feature augmentation from image augmentation to reduce learning difficulty. Augmentation techniques like Cowmix [16] focus on providing diverse training samples. + +However, these methods generally do not directly improve the quality of pseudo-labels. To overcome these limitations, several approaches have been proposed [20, 28]. AEL [23] employs a sample weighting strategy to emphasize reliable predictions. DAW [51] uses labeled data distribution to determine optimal thresholds. LOGICDIAG [34] introduces symbolic reasoning to correct erroneous pseudolabels. + +Nevertheless, these methods are still constrained by the low separability of confidence scores, resulting in biased supervisory signals. Therefore, ESL [39] uses multihot labels to mitigate the effects of incorrect predictions. U2PL [60] leverages unreliable predictions through contrastive learning. CorrMatch [50] utilizes correlation maps to propagate pseudo-labels in low-confidence predictions. Furthermore, this pseudo-label selection challenge exists even in unsupervised adaptation. TransCal [56] introduces calibration loss to improve the calibration performance of labelless sets. MIC [22] employing random masking strategies attempt to enhance prediction reliability. Despite these efforts, these methods face difficulties in fully eliminating inherent biases. Unlike previous approaches, CSL introduces additional metrics to address the low discriminability of confidence scores and explicitly mitigates the supervision bias caused by pseudo-labels. + +![](images/ba3d9f024faa2fc2dfab1d4fd05c14b4f501fe7f50cd1cb75d8f537dcf9c48ea.jpg) +Figure 2. Overview of CSL pipeline for unlabeled images. The predictions of weakly augmented images are partitioned on a per-pixel basis into reliable and unreliable predictions through a convex optimization strategy in the feature space $\Phi$ . This process constructs loss weight mask $M_{i}$ and pseudo-label $\hat{y}_i^u$ , which are then used as supervisory signals for strongly augmented and masked images. + +# 3. Method + +# 3.1. Preliminaries + +In semi-supervised semantic segmentation, the training data consists of a labeled set $D_{l} = \{(x_{i}^{l},y_{i}^{l})\}_{i = 1}^{N_{l}}$ and an unlabeled set $D_{u} = \{x_{i}^{u}\}_{i = 1}^{N_{u}}$ , where $N_{l}\ll N_{u}$ . $x_{i}\in \mathbb{R}^{3\times H\times W}$ is the input image, with $H$ and $W$ denoting its height and width, respectively. $y_{i}\in \mathbb{R}^{K\times H\times W}$ is the one-hot encoded pixel label map with $K$ classes. For the semantic segmentation network $\mathcal{F}$ , the supervised loss is defined as: + +$$ +\mathcal {L} _ {S} = \frac {1}{B _ {L}} \sum_ {i = 1} ^ {B _ {L}} C E \left(y _ {i} ^ {l}, \mathcal {F} \left(A ^ {\omega} \left(x _ {i} ^ {l}\right)\right)\right), \tag {1} +$$ + +where $B_{L}$ is the batch size of labeled data, $CE(\cdot, \cdot)$ denotes the cross-entropy loss function, and $A^{\omega}(\cdot)$ is the weak data augmentation function. + +Similarly, given a batch size $B_U$ of unlabeled data, the unsupervised training loss follows a simple framework [69] with weak-to-strong consistency regularization: + +$$ +\mathcal {L} _ {u} ^ {a} = \frac {1}{B _ {U}} \sum_ {i = 1} ^ {B _ {U}} M _ {i} \odot C E \left(\hat {y} _ {i} ^ {u}, \mathcal {F} \left(A ^ {s} \left(A ^ {\omega} \left(x _ {i} ^ {u}\right)\right)\right)\right), \tag {2} +$$ + +where $\hat{y}_i^u$ is the label derived from $F(A^{\omega}(x_i^u))$ , $A^s (\cdot)$ is the strong data augmentation function, and $M_{i}$ is the indicator mask for pseudo-labels in $\hat{y}_i^u$ . In the baseline, following the confidence thresholding method, + +$$ +M _ {i} = \mathbb {I} \left(\max \left(F \left(A ^ {\omega} \left(x _ {i} ^ {u}\right)\right)\right) > \tau\right), \tag {3} +$$ + +where $\mathbb{I}(\cdot)$ is the indicator function, $\tau$ denotes the confidence threshold. However, this strategy inevitably suffers from cognitive bias and supervision loss. + +To address these issues, we propose CSL. As illustrated in Fig. 2, CSL introduces residual dispersion as an additional measure(Sec. 3.2) and through spectral relaxation, formulates $M_{i}$ construction as a convex optimization problem(Sec. 3.3). Subsequently, CSL introduces additional supervisory signals that compel the network to learn global context relationships based on regions lacking supervision, compensating for supervisory gaps in challenging areas (Sec. 3.4). + +# 3.2. Theoretical Analysis of Residual Dispersion + +The overconfidence tendency of deep networks challenges the effectiveness of maximum confidence as the sole criterion for pseudo-label selection. Based on the entropy minimization principle, we theoretically demonstrate that reliable predictions should simultaneously exhibit high maximum confidence and residual dispersion. + +Let $p_n \in \mathbb{R}^K$ denote the predicted class probabilities for pixel $n$ in an unlabeled image $x_i^u$ , $p_n(k') = \max_k p_n(k)$ represents the maximum confidence. According to entropy minimization, the ideal prediction should follow an unimodal distribution $q = [q_1, \dots, q_K]$ , where $q_{k'} = 1 - (K - 1)\epsilon$ and $q_{k \neq k'} = \epsilon$ with $\epsilon \to 0$ . Based on the decomposition + +form of cross-entropy, the goal can be defined as + +$$ +C E \left(p _ {n}, q\right) = - q _ {k ^ {\prime}} \log p _ {n} \left(k ^ {\prime}\right) - \epsilon \sum_ {k \neq k ^ {\prime}} \log p _ {n} (k). \tag {4} +$$ + +Decomposing non-maximum probabilities $p_n(k), \forall k \neq k'$ into mean-shift components $p_n(k) = \mu_{\mathrm{res}} + \delta_k$ , where $\mu_{\mathrm{res}} = \frac{1 - p_n(k')}{K - 1}$ and $\sum_{k \neq k'} \delta_k = 0$ . Applying second-order Taylor expansion to $\log p_n(k)$ around $\mu_{\mathrm{res}}$ , + +$$ +\log p _ {n} (k) = \log \mu_ {\text {r e s}} + \frac {\delta_ {k}}{\mu_ {\text {r e s}}} - \frac {\delta_ {k} ^ {2}}{2 \mu_ {\text {r e s}} ^ {2}} + \mathcal {O} \left(\delta_ {k} ^ {3}\right), \forall k \neq k ^ {\prime}. \tag {5} +$$ + +Substituting Eq. (5) into Eq. (4) and simplifying, + +$$ +C E \left(p _ {n}, q\right) \approx - \log p _ {n} \left(k ^ {\prime}\right) - \frac {\epsilon (K - 1) ^ {3}}{2 \left(1 - p _ {n} \left(k ^ {\prime}\right)\right) ^ {2}} \cdot v _ {n}, \tag {6} +$$ + +where the residual dispersion $v_{n}$ is defined as + +$$ +v _ {n} \triangleq - \frac {1}{K - 1} \sum_ {k \neq k ^ {\prime}} \delta_ {k} ^ {2}. \tag {7} +$$ + +Eq. (6) reveals that under entropy minimization objectives, trustworthy predictions require maximizing both maximum confidence $p_n(k')$ and residual dispersion $v_n$ . Notably, as $p_n(k') \to 1$ , the coefficient term of $v_n$ becomes significant, shows that $v_n$ becomes a key indicator to distinguish prediction reliability in overconfidence cases. + +# 3.3. Prediction Convex Optimization Separation + +Sec. 3.2 shows that prediction reliability nonlinearly depends on both $p_n(k')$ and $v_{n}$ . However, threshold-based methods fail to capture the nonlinear complementary relationship. To overcome this limitation, we propose a sample-adaptive method, which formulates pseudo-label selection as a convex optimization problem by spectral relaxation. + +We construct an attribute vector $h_n = [p_n(k'), v_n]^T$ for each pixel prediction, embedding it into a feature space that captures maximum confidence and residual dispersion. Based on this, we aim to separate predictions with high maximum confidence and residual dispersion to select reliable pseudo-labels. This selection is formalized as a trace maximization problem, + +$$ +\max _ {\mathbf {S}} T r \left(S ^ {T} \Phi^ {T} \Phi S\right), s. t. S \in \{0, 1 \} ^ {H W \times 2}, \tag {8} +$$ + +where $Tr(\cdot)$ denotes the trace of a matrix. $\Phi = [\mathbf{h}_1,\mathbf{h}_2,\dots,\mathbf{h}_{HW}]$ is the feature matrix of $x_{i}^{u}$ . $S$ is the selection matrix, subject to the constraint $\sum_{c}S_{n,c} = 1$ , with $c\in \{1,2\}$ indicates the class index in the selection. + +To solve this non-convex combinatorial optimization problem, we apply spectral relaxation [43], relaxing the discrete constraints on $S$ . According to the Ky Fan theorem [53], the optimal solution $S^{*}$ approximates: + +$$ +S _ {n, c} ^ {*} = \Delta \left(\arg \max _ {i \in \{1, 2 \}} | u _ {i} (n) |, c\right), \tag {9} +$$ + +![](images/36d3ce777896e53943f85d99a09fdc399d5f446dd8be903bb3e98121c15c4488.jpg) +Figure 3. Illustration of Trusted Mask Perturbation. The white area in $\hat{y}_i^u$ represents the unreliable region. + +![](images/38373a0697232b43864129b847c4df777195c7ec7b3b179809e74728eb60d845.jpg) + +![](images/a62e7f751f5e85c22ea682f52bddc283cfe6cb0fe7f1f434cb8bfa37230dda35.jpg) + +where $u_{1}, u_{2}$ are the eigenvectors of the matrix $\Phi^T\Phi$ , $|u_i(n)|$ denotes the absolute value of the $n$ -th element in $u_{i}$ , and $\Delta (\cdot ,\cdot)$ is the Kronecker delta function. + +Without loss of generality, assuming that $c = 2$ indicates the subset of predictions with high maximum confidence and residual dispersion, we define the projection matrix $Z = \Phi S_{n,2}^{*}$ , representing the projection of predictions onto the reliable component. Based on this, applying a Gaussian weighting function to construct smooth loss weights $\omega_{n}$ , + +$$ +\omega_ {n} = \prod_ {c} \exp \left(\frac {\left(h _ {n} (c) - \mu_ {c}\right) ^ {2}}{- \alpha \sigma_ {c} ^ {2}}\right), \tag {10} +$$ + +where $\mu_{c}$ is the mean of the high-reliability prediction subset in the projection matrix $Z$ for component $c$ , and $\sigma_{c}$ is its variance. $\alpha = 8$ is the smoothing parameter. To ensure that predictions with high maximum confidence and residual dispersion are always retained, the weights of these predictions are set to 1. Therefore, the unsupervised training indicator mapping $M_{i}$ for $x_{i}^{u}$ is defined as: + +$$ +M _ {i} (n) = \left\{ \begin{array}{l l} 1 & \text {i f} \left(h _ {n} (c) - \mu_ {c}\right) > 0, \forall c \in \{1, 2 \} \\ \omega_ {n} & \text {o t h e r w i s e} \end{array} . \right. \tag {11} +$$ + +# 3.4. Trusted Mask Perturbation + +While semi-supervised training can benefit from high-quality pseudo-labels, persistent exclusion of challenging regions risks entrenching the model in simplistic patterns. To mitigate this, we propose the Trusted Mask Perturbation (TMP), which selectively masks high-confidence predictions, thereby implicitly supervising unreliable regions using contextual information. + +Utilizing the reliable pixel indicator $M_{i}$ from Sec. 3.3, where $M_{i} = 1$ denotes reliable predictions, we define a patch-based perturbation mask $R_{i} \in \{0,1\}^{H \times W}$ generated as follows: + +$$ +R _ {i} (x, y) = \mathbb {I} \left(M _ {i} (x, y) = 1\right) \cdot \mathbb {I} \left(U \left(\lfloor \frac {x}{s} \rfloor , \lfloor \frac {y}{s} \rfloor\right) \geq \theta\right), \tag {12} +$$ + +
PASCAL originalTrain Size1/16(92)1/8(183)1/4(366)1/2(732)Full(1464)
Supervised321 × 32145.453.161.268.473.2
ST++ [68]321 × 32165.271.074.677.379.1
UniMatch [69]321 × 32175.277.278.879.981.2
CorrMatch [50]321 × 32176.478.579.480.681.8
ESL [39]513 × 51371.074.178.179.581.8
LogicDiag [34]513 × 51373.376.778.079.4-
RankMatch [40]513 × 51375.577.679.880.782.2
AugSeg [74]513 × 51371.175.578.880.381.4
SemiCVT [24]513 × 51368.671.375.078.580.3
FPL [44]513 × 51369.371.775.779.0-
AllSpark [54]513 × 51376.178.479.880.882.1
CSL321 × 32176.879.680.380.982.3
+ +Table 1. Comparison with SOTAs on the PASCAL VOC 2012 original dataset under different splits. All labelled images were selected from the original VOC training set, which consists of 1,464 images. The fractions denote the percentage of labeled data used for training. + +where $s = 32$ is the patch size, $\theta = 0.7$ is the masking ratio, and $U(\cdot ,\cdot)$ denotes the uniform distribution over the interval [0, 1]. + +The perturbed image $\tilde{x}_i^u$ is obtained by applying $R_{i}$ to the input image $x_{i}^{u}$ : + +$$ +\tilde {x} _ {i} ^ {u} = x _ {i} ^ {u} \odot R _ {i}, \tag {13} +$$ + +where $\odot$ denotes element-wise multiplication. As illustrated in Fig. 3, this strategy selectively masks reliable pixels, compelling the model to infer the missing content based on contextual cues from the remaining challenging regions. + +The unsupervised loss for the masked inputs is formulated as: + +$$ +\mathcal {L} _ {u} ^ {m} = \frac {1}{B _ {U}} \sum_ {i = 1} ^ {B _ {U}} M _ {i} \odot C E \left(\hat {y} _ {i} ^ {u}, F \left(\tilde {x} _ {i} ^ {u}\right)\right). \tag {14} +$$ + +The final unsupervised loss term combines both the standard and masked losses: + +$$ +\mathcal {L} _ {U} = \lambda_ {1} L _ {u} ^ {a} + \lambda_ {2} L _ {u} ^ {m}, \tag {15} +$$ + +where $\lambda_1, \lambda_2$ are loss weights set to [0.5, 0.5] by default. + +# 4. Experiments + +# 4.1. Implementation Details + +Datasets. We evaluate the proposed method on three benchmark datasets: PASCAL VOC 2012, Cityscapes, and MS COCO. PASCAL VOC 2012 [15] is a standard semantic segmentation dataset comprising 20 object classes and one background class. The training and validation sets contain 1,464 and 1,449 images, respectively. Following previous works [25, 76], we augment the training set by incorporating 9,118 images with coarse annotations from the Segmentation Boundary Dataset (SBD) [18], resulting in a blender + +set of 10,582 training images. Cityscapes [12] focuses on urban scene understanding and includes 19 classes. It consists of 2,975 finely annotated training images and 500 validation images, all at a high resolution of $1024 \times 2048$ . MS COCO [35] is a large-scale dataset with 81 object categories, containing 118k training and 5k validation images for segmentation tasks. + +Implementation Details. To ensure a fair comparison with prior methods, we adopt the ResNet [19] backbone with DeepLabv3+ [7] as the decoder for PASCAL VOC 2012 and Cityscapes, and Xception-65 [11] for MS COCO [35]. For all datasets, we use SGD with a batch size of 8 and weight decay of $1 \times 10^{-4}$ . For PASCAL VOC 2012, we set the initial learning rate to 0.001, crop size of $321 \times 321$ or $513 \times 513$ , training epochs of 80, and a decoder learning rate ten times that of the backbone. For Cityscapes, we use an initial learning rate of 0.001, crop size of $801 \times 801$ , and training epochs of 240, with OHEM [48]. For COCO, we set the initial learning rate to 0.004, crop size of $513 \times 513$ and train for 30 epochs. + +Evaluation Metrics. For PASCAL VOC 2012, we follow standard protocols [20, 60], reporting mIoU on the original validation images. For Cityscapes, we perform sliding window evaluation with a crop size of $801 \times 801$ and compute mIoU [23, 31]. For COCO, we evaluate on the standard 5k validation set with mIoU. Ablation studies are conducted on PASCAL VOC 2012 with a crop size of $321 \times 321$ . + +# 4.2. Comparison with State-of-the-Art Methods + +Original PASCAL VOC 2012. Tab. 1 presents the comparison results on the original PASCAL VOC 2012 dataset. Compared to the supervised baseline, CSL consistently achieves significant performance improvements. Specifically, under the $1/16$ , $1/8$ , $1/4$ , $1/2$ , and full data splits, CSL attains improvements of $+31.4\%$ , $+26.5\%$ , $+19.1\%$ , + +
PASCAL blenderTrain Size1/16 (662)1/8 (1323)1/4 (2646)
ST++ [68]321 × 32174.576.376.6
UniMatch [69]321 × 32176.577.077.2
CorrMatch [50]321 × 32177.677.878.3
CAC [29]321 × 32172.474.676.3
CSL321 × 32177.878.579.0
U2PL† [60]513 × 51377.279.079.3
GTA† [25]513 × 51377.880.480.5
CorrMatch† [50]513 × 51381.381.980.9
UniMatch† [69]513 × 51381.081.980.4
AugSeg† [74]513 × 51379.381.580.5
AllSpark† [54]513 × 51381.682.080.9
CSL†513 × 51381.682.481.1
+ +
PASCAL blenderTrain Size1/16 (662)1/8 (1323)1/4 (2646)
ST++ [68]513 × 51374.777.977.9
UniMatch [69]513 × 51378.178.479.2
CorrrMatch [50]513 × 51378.479.379.6
ESL [39]513 × 51376.478.679.0
PS-MT [38]513 × 51375.578.278.7
CFCG [32]513 × 51376.879.180.0
CCVC [62]513 × 51377.278.479.0
DLG [31]513 × 51377.879.379.1
RankMatch [40]513 × 51378.979.280.0
DGCL [57]513 × 51376.678.379.3
AllSpark [54]513 × 51378.380.080.4
DDFP [58]513 × 51378.378.979.8
CSL513 × 51378.979.980.3
+ +Table 2. Comparison with SOTAs on PASCAL VOC 2012 blender dataset. All labelled images were selected from the blender VOC training set under different splits, which consists of 10,582 images. $\dagger$ means using U2PL's splits. + +
Cityscapes1/16(186)1/8(732)1/4(744)1/2(1488)
Supervised63.170.573.176.2
ST++ [68]67.673.474.677.8
UniMatch [69]76.677.979.279.5
CorrMatch [50]77.378.579.480.4
ESL [39]75.177.279.080.5
DGCL [57]73.277.378.580.7
AugSeg [74]75.277.879.680.4
CCVC [62]74.976.477.3-
DAW [51]76.678.479.880.6
U2PL [60]70.374.476.579.1
CPS [8]69.874.374.676.8
DDFP [58]77.178.279.980.8
FPL [44]75.778.579.2-
CSL78.278.880.081.1
+ ++12.5%, and +9.1%, respectively. Moreover, when compared with existing SOTAs, our approach consistently demonstrates superior performance. Particularly, under the 1/8 and 1/4 splits, CSL outperforms UniMatch [69] by +2.4% and +1.5%, respectively. + +Blender PASCAL VOC 2012. We show the comparison between our method and SOTAs on the blender PASCAL VOC 2012 dataset in Tab. 2. At a crop size of $321 \times 321$ , CSL surpasses CorrMatch [50] by $+0.2\%$ , $+0.7\%$ , and $+0.7\%$ . We observe that the performance gains become more pronounced as the amount of available annotations decreases. For instance, at a crop size of $513 \times 513$ under the 1/16 splits, + +Table 3. Comparison with SOTAs on Cityscapes dataset. All methods are built upon DeepLabV3+ [7] and ResNet101 [19]. + +
COCO1/5121/2561/1281/641/32
Supervised22.528.333.138.942.1
UniMatch [69]31.938.944.448.249.8
LogicDiag [34]33.140.345.448.850.5
PseudoSeg [76]29.837.139.141.843.6
CSL35.142.345.849.751.2
+ +Table 4. Comparison with SOTAs on MS-COCO dataset. All methods are built upon XC-65 [11] for fair comparison. + +our method improves upon AllSpark [54] by $+0.6\%$ . + +Additionally, we conduct experiments using the same splits as U2PL [60], where the supervised dataset primarily consists of high-quality annotations supplemented with a small amount of coarse annotation data. Under the 1/8 splits, since most high-quality annotations are included and the coarse annotations are largely excluded, the model benefits from a robust initial representation. This enables our method to achieve an outstanding performance of $82.4\%$ , representing a $+0.9\%$ improvement over AugSeg [74]. This significant performance gain demonstrates our method's ability to effectively leverage accurate network representations in scenarios with scarce annotations. + +Cityscapes. Tab. 3 quantitatively compares the performance of CSL with SOTAs on the Cityscapes validation set. Leveraging the extensive unlabeled data, CSL achieves performance gains of $+15.1\%$ , $+8.3\%$ , $+6.9\%$ , and $+5.1\%$ over the supervised baseline under the 1/16, 1/8, 1/4, and 1/2 splits, respectively. CSL surpasses existing SOTAs across all partition protocols. For example, under the 1/16 partition, our method achieves a $+0.9\%$ performance improvement over CorrMatch [50]. + +![](images/fe7bf6dd50de7ae67e597fa7d0a677ee05448bf6b7e469c599a6de59714063fa.jpg) +(a) Dispersion Metrics + +![](images/d6c6f98a161ec47fbac38e3b184c89d2023cce3460d16b419988874da28e8478.jpg) +(b) Loss Smoothing Strategy + +![](images/3086157c825c045e75ff399b94c346fa0483be54cf98d22fc2284488683878f1.jpg) +(c) Trusted Mask Perturbation + +
SemiBaselineM1CSL ωnLu m1/16 (92)1/2 (732)
✓72.978.6
✓✓73.879.3
✓✓✓74.279.8
✓✓73.679.4
✓✓✓✓76.880.9
+ +Table 5. Ablation study on the effectiveness of different components, including optimization strategy $M_{i}$ , smooth loss weight $\omega_{n}$ , trusted mask perturbation $\mathcal{L}_u^m$ . + +MS COCO. Tab. 4 presents the quantitative comparisons of CSL with SOTAs on the MS COCO dataset. Across all partition ratios, CSL consistently outperforms existing approaches by significant margins. Notably, when utilizing only 1/512 of the labeled data, CSL achieves an impressive $35.1\%$ mIoU, surpassing the strongest baseline Logic-Diag [34] by $+2.0\%$ . + +# 4.3. Ablation Studies + +Effectiveness of components. We conducted ablation studies on the different components of CSL to demonstrate their effectiveness, as shown in Tab. 5. Starting with a semi-supervised baseline, we achieve mIoU scores of $72.9\%$ and $78.6\%$ for the 1/16 and 1/2 splits, respectively. By introducing the residual dispersion and optimization strategy $M_{i}$ , the performance improves to $73.8\%$ and $79.3\%$ . Building upon this, incorporating the smooth loss weight $\omega_{n}$ yields additional gains of $0.4\%$ and $0.5\%$ . By combining all components, our method surpasses the baseline by over $3.9\%$ and $2.3\%$ , achieving SOTA results. Notably, adding the trusted mask perturbation $\mathcal{L}_u^m$ alone to the semi-supervised baseline results in performance gains of $0.7\%$ and $0.8\%$ , illustrating its general applicability in addressing the lack of supervision in challenging regions of pseudo-labels. + +Evaluation of Dispersion Metrics. Fig. 4a compares resid + +![](images/e1395c48c60e243221f560da53b2442b289ebdf88bf26a3196a42411d3cb8b90.jpg) +Figure 4. Ablation studies on different designs of CSL. For (a), (b), and (c), experiments are conducted on the $1/2$ splits. +Figure 5. Ablation study of smoothing parameter $\alpha$ . + +ual dispersion with pixel-wise margin [47] and entropy. Both margin and entropy exhibit significant performance degradation. Because when the model is overconfident, the margin dominates by the maximum confidence, while entropy becomes insensitive to non-maximum confidence values. Under such circumstances, reliability assessment degenerates into merely comparing maximum confidence scores. Following the setup in Sec. 3.2, residual dispersion maintains magnitude consistency in scale and coefficients, ensuring robust pseudo-label selection. Meanwhile, replacing the optimization strategy with sample-wise mean thresholds [5] resulted in consistent performance degradation, validating the necessity of our design. + +Weight smoothing strategy. CSL constructs smoothed loss weights by the Gaussian weighting function. We compare it with alternative measures, including Euclidean distance and cosine similarity. Additionally, we examined the effect of retaining loss weights for reliable predictions versus uniformly smoothing all loss weights. As shown in Fig. 4b, combining Gaussian weighting with retention of reliable prediction loss weights yields the most significant performance improvement. Suggests that our approach ensures + +![](images/226fd5c4f263fcb988100a59edc3ab55e1160e49523140fb5581ec74f5d50e85.jpg) +(a) Sampling Rate + +![](images/124a4ae9f1c8d9b557ddf8e001860274dd12e84f2ae3403e8e7e9260227f696e.jpg) +(b) Pseudo-label accuracy + +
Patch Size s8163264
θ=0.378.778.578.476.2
θ=0.578.279.178.977.8
θ=0.778.879.279.478.4
θ=0.977.577.677.677.4
+ +Table 6. Ablation study with different mask perturbation parameters for $1/2$ splits. + +models learn effectively from credible examples, enhancing overall segmentation performance. + +Impact of trusted mask perturbation. Fig. 4c analyses the performance of block-wise masking [2] versus random patch masking across different masking targets. The results show that masking only reliable predictions yields the best performance. In contrast, masking untrustworthy predictions or both results in significant performance degradation. This decline may stem from the absence of supervised signals and image content in untrustworthy regions, leading to excessive perturbation. Furthermore, random patch masking consistently outperforms block-wise masking due to the increased sample diversity introduced. + +Impact of hyperparameters. The effects of the loss smoothing coefficient $\alpha$ , masking patch size $s$ , and masking rate $\theta$ on the PASCAL VOC 2012 dataset were investigated. Fig. 5 illustrates that an excessively high value of $\alpha$ leads to performance degradation, which results from the improper weighting of unreliable predictions. Consequently, $\alpha = 8$ has been established as the default value to optimize overall performance. Tab. 6 demonstrates that optimal performance enhancement is achieved from credible masking when $\theta = 0.7$ and $s = 32$ . In contrast, when the masking patch size $s = 64$ , the mIoU decreases significantly, which is attributed to the extreme size of the patch compared to the image size, ultimately leading to a loss of local contextual information. + +# 4.4. Qualitative analysis + +CSL improved semi-supervised training. To evaluate the effectiveness of CSL, we measured the average accuracy and the average sampling rate of pseudo-labels produced by + +![](images/3f8e4b36481b9c9ad6def63382f8e547a4ce8211ff9a4f50550dd4a993f65698.jpg) + +![](images/2bc5759214d62bd73702902630867529d2f724401385a1950467736652226ab9.jpg) + +![](images/409ec2c49a2e097f24a58ddd89fb91727ed0e70dda6a59fa55a4030b0d079b8c.jpg) +Figure 6. Training curves for our CSL and the threshold strategy. +(a) W/O CSL +Figure 7. Qualitative results on the PASCAL VOC 2012 dataset. (a) Pseudo labels without CSL; (b) Pseudo labels with CSL; (c) Ground truth. + +![](images/3329ed7f2f58d6928d8968e41d8d14f40ad92c28a1ba93919c4336a566bcfe3d.jpg) + +![](images/4fe23d55f7afba4d2977f11fc38b16bf11c82dd8d31876f8cea034459ce10e06.jpg) + +![](images/91eee5fdca0aa64ee904ea564158caefd8eef14f2f61ab14ce2dde019ecc6a73.jpg) +(b) W/ CSL + +![](images/a4498e17df050fc2a098da7c59a148d453716e0533b91ca3c40064a66f3f3240.jpg) + +![](images/c9e366970fd395419a458e9e6085ab9fd51ec26b396da0071225d33613a00bb5.jpg) + +![](images/e5a74f0c6b12748f560dfc643c2d398ad1fbae358420073e8eea5e24c118edd0.jpg) +(c) GT + +our method and threshold-based methods. As seen in Fig. 6, suffering from overconfidence, as the quantity of labeled data diminishes, threshold-based methods exhibit a proclivity for increasingly assertive pseudo-label selection, resulting in a notable surge in false supervision signals. In contrast, CSL demonstrates superior performance in mitigating cognitive biases and consistently generates high-quality pseudo-labels. + +Qualitative Results. Fig. 7 illustrates the visualization results for the PASCAL VOC 2012 dataset under 1/8 splits. Obviously, our predictive convex optimized segmentation strategy effectively broadens the regions encompassed by pseudo-labels. By utilizing plausible mask perturbations, CSL achieves more accurate and detailed segmentation in intricate areas, and the visualization results demonstrate its capacity to enhance segmentation performance by leveraging unlabelled data and addressing supervision gaps. + +# 5. Conclusions + +We demonstrate that through comprehensive measurement of maximum confidence and residual dispersion, a pronounced natural separation exists between reliable predictions and cognitive biases. Motivated by this, we explore pseudo-label selection in this space through a convex optimization strategy. On the other hand, we attempt to motivate the supervised signals of reliable predictions to improve prediction results in unreliable regions by masking the reliable pixels. Experiments demonstrate the significant effectiveness of decision boundaries in this space, and that masking strategy significantly improves prediction results in complex regions. In addition, comparative experiments demonstrate the superiority of CSL over other methods. + +Acknowledgements. This work is supported by the China Postdoctoral Program for State-Owned Enterprises (GZC20251179) and Hunan Youth Science Foundation Program (2025JJ60424). + +# References + +[1] Eric Arazo, Diego Ortega, Paul Albert, Noel E O'Connor, and Kevin McGuinness. Pseudo-labeling and confirmation bias in deep semi-supervised learning. In IJCNN, pages 1-8. IEEE, 2020. 2 +[2] Hangbo Bao, Li Dong, Songhao Piao, and Furu Wei. Beit: Bert pre-training of image transformers. In ICLR, 2021. 8 +[3] Olivier Chapelle, Bernhard Scholkopf, and Alexander Zien. Semi-supervised learning. IEEE Transactions on Neural Networks, 20(3):542-542, 2009. 1 +[4] Baixu Chen, Junguang Jiang, Ximei Wang, Pengfei Wan, Jian-min Wang, and Mingsheng Long. Debiased self-training for semi-supervised learning. NeurIPS, 35:32424-32437, 2022. 2 +[5] Hao Chen, Ran Tao, Yue Fan, Yidong Wang, Jindong Wang, Bernt Schiele, Xing Xie, Bhiksha Raj, and Marios Savvides. Softmatch: Addressing the quantity-quality tradeoff in semi-supervised learning. In ICLR, 2023. 2, 7 +[6] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. IEEE TPAMI, 40(4):834-848, 2017. 1 +[7] Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, and Hartwig Adam. Encoder-decoder with atrous separable convolution for semantic image segmentation. In ECCV, pages 801-818, 2018. 5, 6 +[8] Xiaokang Chen, Yuhui Yuan, Gang Zeng, and Jingdong Wang. Semi-supervised semantic segmentation with cross pseudo supervision. In CVPR, pages 2613-2622, 2021. 1, 6 +[9] Xinyue Chen, Miaojing Shi, Zijian Zhou, Lianghua He, and Sophia Tsoka. Enhancing generalized few-shot semantic segmentation via effective knowledge transfer. In AAAI, 2025. 1 +[10] Shengyang Cheng, Jianyong Huang, Xiaodong Wang, Lei Huang, and Zhiqiang Wei. Image-text aggregation for open-vocabulary semantic segmentation. Neurocomputing, page 129702, 2025. 1 +[11] François Chollet. Xception: Deep learning with depthwise separable convolutions. In CVPR, pages 1251-1258, 2017. 5, 6 +[12] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In CVPR, pages 3213-3223, 2016. 5 +[13] Jiacheng Deng, Jiahao Lu, and Tianzhu Zhang. Quantity-quality enhanced self-training network for weakly supervised point cloud semantic segmentation. IEEE TPAMI, 2025. 1 + +[14] Ye Du, Yujun Shen, Haochen Wang, Jingjing Fei, Wei Li, Liwei Wu, Rui Zhao, Zehua Fu, and Qingjie Liu. Learning from future: A novel self-training framework for semantic segmentation. NeurIPS, 35:4749-4761, 2022. 2 +[15] Mark Everingham, SM Ali Eslami, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes challenge: A retrospective. IJCV, 111:98-136, 2015. 5 +[16] G. French, S. Laine, T. Aila, and M. Mackiewicz. Semi-supervised semantic segmentation needs strong, varied perturbations. In BMVC, 2020. 2 +[17] Chuan Guo, Geoff Pleiss, Yu Sun, and Kilian Q Weinberger. On calibration of modern neural networks. In ICML, pages 1321-1330. PMLR, 2017. 2 +[18] Bharath Hariharan, Pablo Arbeláez, Lubomir Bourdev, Subhransu Maji, and Jitendra Malik. Semantic contours from inverse detectors. In ICCV, pages 991–998, 2011. 5 +[19] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 5, 6 +[20] Ruifei He, Jihan Yang, and Xiaojuan Qi. Re-distributing biased pseudo labels for semi-supervised semantic segmentation: A baseline investigation. In CVPR, pages 6930-6940, 2021. 2, 5 +[21] Seunghoon Hong, Hyeonwoo Noh, and Bohyung Han. Decoupled deep neural network for semi-supervised semantic segmentation. NeurIPS, 28, 2015. 1 +[22] Lukas Hoyer, Dengxin Dai, Haoran Wang, and Luc Van Gool. Mic: Masked image consistency for context-enhanced domain adaptation. In CVPR, pages 11721-11732, 2023. 2 +[23] Hanzhe Hu, Fangyun Wei, Han Hu, Qiwei Ye, Jinshi Cui, and Liwei Wang. Semi-supervised semantic segmentation via adaptive equalization learning. NeurIPS, 34:22106-22118, 2021. 1, 2, 5 +[24] Huimin Huang, Shiao Xie, Lanfen Lin, Ruofeng Tong, Yen-Wei Chen, Yuexiang Li, Hong Wang, Yawen Huang, and Yefeng Zheng. Semicvt: Semi-supervised convolutional vision transformer for semantic segmentation. In CVPR, pages 11340–11349, 2023. 5 +[25] Ying Jin, Jiaqi Wang, and Dahua Lin. Semi-supervised semantic segmentation via gentle teaching assistant. NeurIPS, 35:2803-2816, 2022. 5, 6 +[26] Rihuan Ke, Angelica I Aviles-Rivero, Saurabh Pandey, Saikumar Reddy, and Carola-Bibiane Schonlieb. A three-stage self-training framework for semi-supervised semantic segmentation. IEEE TIP, 31:1805-1815, 2022. 1 +[27] Jiwon Kim, Youngjo Min, Daehwan Kim, Gyuseong Lee, Junyoung Seo, Kwangrok Ryoo, and Seungryong Kim. Conmatch: Semi-supervised learning with confidence-guided consistency regularization. In ECCV, pages 674-690. Springer, 2022. 2 +[28] Donghyeon Kwon and Suha Kwak. Semi-supervised semantic segmentation with error localization network. In CVPR, pages 9957-9967, 2022. 2 +[29] Xin Lai, Zhuotao Tian, Li Jiang, Shu Liu, Hengshuang Zhao, Liwei Wang, and Jiaya Jia. Semi-supervised seman + +tic segmentation with directional context-aware consistency. In CVPR, pages 1205-1214, 2021. 1, 6 +[30] Samuli Laine and Timo Aila. Temporal ensembling for semi-supervised learning. arXiv preprint arXiv:1610.02242, 2016. 2 +[31] Peixia Li, Pulak Purkait, Thalaiyasingam Ajanthan, Majid Abdolshah, Ravi Garg, Hisham Husain, Chenchen Xu, Stephen Gould, Wanli Ouyang, and Anton Van Den Hengel. Semi-supervised semantic segmentation under label noise via diverse learning groups. In CVPR, pages 1229-1238, 2023. 5, 6 +[32] Shuo Li, Yue He, Weiming Zhang, Wei Zhang, Xiao Tan, Junyu Han, Errui Ding, and Jingdong Wang. Cfgc: Semi-supervised semantic segmentation via cross-fusion and contour guidance supervision. In CVPR, pages 16348-16358, 2023. 6 +[33] Siyuan Li, Weiyang Jin, Zedong Wang, Fang Wu, Zicheng Liu, Cheng Tan, and Stan Z Li. Semireward: A general reward model for semi-supervised learning. In ICLR, 2024. 2 +[34] Chen Liang, Wenguan Wang, Jiaxu Miao, and Yi Yang. Logic-induced diagnostic reasoning for semi-supervised semantic segmentation. In CVPR, pages 16197-16208, 2023. 2, 5, 6, 7 +[35] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, pages 740-755. Springer, 2014. 5 +[36] Jinshi Liu, Zhaohui Jiang, Ting Cao, Zhiwen Chen, Chaobo Zhang, and Weihua Gui. Generated pseudo-labels guided by background skeletons for overcoming under-segmentation in overlapping particle objects. TCSVT, 33(6):2906-2919, 2022. 1 +[37] Jinshi Liu, Zhaohui Jiang, Weihua Gui, Zhiwen Chen, and Chaobo Zhang. Occlusion segmentation: Restore and segment invisible areas for particle objects. IEEE Transactions on Automation Science and Engineering, 2024. 1 +[38] Yuyuan Liu, Yu Tian, Yuanhong Chen, Fengbei Liu, Vasileios Belagiannis, and Gustavo Carneiro. Perturbed and strict mean teachers for semi-supervised semantic segmentation. In CVPR, pages 4258-4267, 2022. 2, 6 +[39] Jie Ma, Chuan Wang, Yang Liu, Liang Lin, and Guanbin Li. Enhanced soft label for semi-supervised semantic segmentation. In CVPR, pages 1185-1195, 2023. 2, 5, 6 +[40] Huayu Mai, Rui Sun, Tianzhu Zhang, and Feng Wu. Rankmatch: Exploring the better consistency regularization for semi-supervised semantic segmentation. In CVPR, pages 3391-3401, 2024. 1, 5, 6 +[41] Sudhanshu Mittal, Maxim Tatarchenko, and Thomas Brox. Semi-supervised semantic segmentation with high-and low-level consistency. IEEE TPAMI, 43(4):1369-1379, 2019. 1 +[42] Takeru Miyato, Shin-ichi Maeda, Masanori Koyama, and Shin Ishii. Virtual adversarial training: a regularization method for supervised and semi-supervised learning. IEEE TPAMI, 41(8):1979–1993, 2018. 2 +[43] Carl Olsson, Anders P Eriksson, and Fredrik Kahl. Improved spectral relaxation methods for binary quadratic optimization problems. Computer Vision and Image Understanding, 112 (1):3-13, 2008. 4 + +[44] Pengchong Qiao, Zhidan Wei, Yu Wang, Zhennan Wang, Guoli Song, Fan Xu, Xiangyang Ji, Chang Liu, and Jie Chen. Fuzzy positive learning for semi-supervised semantic segmentation. In CVPR, pages 15465-15474, 2023. 5, 6 +[45] Mamshad Nayeem Rizve, Kevin Duarte, Yogesh S Rawat, and Mubarak Shah. In defense of pseudo-labeling: An uncertainty-aware pseudo-label selection framework for semi-supervised learning. arXiv preprint arXiv:2101.06329, 2021. 2 +[46] Mehdi Sajjadi, Mehran Javanmardi, and Tolga Tasdizen. Regularization with stochastic transformations and perturbations for deep semi-supervised learning. NeurIPS, 29, 2016. 2 +[47] Gyungin Shin, Weidi Xie, and Samuel Albanie. All you need are a few pixels: semantic segmentation with pixelpick. In ICCV, pages 1687-1697, 2021. 7 +[48] Abhinav Shrivastava, Abhinav Gupta, and Ross Girshick. Training region-based object detectors with online hard example mining. In CVPR, pages 761-769, 2016. 5 +[49] Kihyuk Sohn, David Berthelot, Nicholas Carlini, Zizhao Zhang, Han Zhang, Colin A Raffel, Ekin Dogus Cubuk, Alexey Kurakin, and Chun-Liang Li. Fixmatch: Simplifying semi-supervised learning with consistency and confidence. NeurIPS, 33:596-608, 2020. 1, 2 +[50] Boyuan Sun, Yuqi Yang, Le Zhang, Ming-Ming Cheng, and Qibin Hou. Corrmatch: Label propagation via correlation matching for semi-supervised semantic segmentation. In CVPR, pages 3097-3107, 2024. 2, 5, 6 +[51] Rui Sun, Huayu Mai, Tianzhu Zhang, and Feng Wu. Daw: exploring the better weighting function for semi-supervised semantic segmentation. NeurIPS, 36, 2024. 1, 2, 6 +[52] Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. NeurIPS, 30, 2017. 2 +[53] GQ Tian. Generalized kkm theorems, minimax inequalities, and their applications. Journal of Optimization Theory and Applications, 83:375-389, 1994. 4 +[54] Haonan Wang, Qixiang Zhang, Yi Li, and Xiaomeng Li. Allspark: Reborn labeled features from unlabeled in transformer for semi-supervised semantic segmentation. In CVPR, pages 3627-3636, 2024. 2, 5, 6 +[55] Kuo Wang, Yuxiang Nie, Chaowei Fang, Chengzhi Han, Xuewen Wu, Xiaohui Wang, Liang Lin, Fan Zhou, and Guanbin Li. Double-check soft teacher for semi-supervised object detection. In IJCAI, pages 1430-1436, 2022. 2 +[56] Ximei Wang, Mingsheng Long, Jianmin Wang, and Michael Jordan. Transferable calibration with lower bias and variance in domain adaptation. Advances in Neural Information Processing Systems, 33:19212-19223, 2020. 2 +[57] Xiaoyang Wang, Bingfeng Zhang, Limin Yu, and Jimin Xiao. Hunting sparsity: Density-guided contrastive learning for semi-supervised semantic segmentation. In CVPR, pages 3114-3123, 2023. 6 +[58] Xiaoyang Wang, Huihui Bai, Limin Yu, Yao Zhao, and Jimin Xiao. Towards the uncharted: Density-descending feature perturbation for semi-supervised semantic segmentation. In CVPR, pages 3303-3312, 2024. 6 + +[59] Yidong Wang, Hao Chen, Yue Fan, Wang Sun, Ran Tao, Wenxin Hou, Renjie Wang, Linyi Yang, Zhi Zhou, Lan-Zhe Guo, et al. Usb: A unified semi-supervised learning benchmark for classification. NeurIPS, 35:3938-3961, 2022. 2 +[60] Yuchao Wang, Haochen Wang, Yujun Shen, Jingjing Fei, Wei Li, Guoqiang Jin, Liwei Wu, Rui Zhao, and Xinyi Le. Semi-supervised semantic segmentation using unreliable pseudo-labels. In CVPR, pages 4248-4257, 2022. 2, 5, 6 +[61] Yidong Wang, Hao Chen, Qiang Heng, Wenxin Hou, Yue Fan, Zhen Wu, Jindong Wang, Marios Savvides, Takahiro Shinozaki, Bhiksha Raj, et al. Freematch: Self-adaptive thresholding for semi-supervised learning. In ICLR, 2023. 2 +[62] Zicheng Wang, Zhen Zhao, Xiaoxia Xing, Dong Xu, Xi-angyu Kong, and Luping Zhou. Conflict-based cross-view consistency for semi-supervised semantic segmentation. In CVPR, pages 19585–19595, 2023. 6 +[63] Zhenyan Wang, Zhenxue Chen, Chengyun Liu, Yaping Zhao, Xinming Zhu, and QM Jonathan Wu. Diversity augmentation and multi-fuzzy label for semi-supervised semantic segmentation. Neurocomputing, page 129681, 2025. 2 +[64] Qizhe Xie, Zihang Dai, Eduard Hovy, Thang Luong, and Quoc Le. Unsupervised data augmentation for consistency training. NeurIPS, 33:6256-6268, 2020. 2 +[65] Qizhe Xie, Minh-Thang Luong, Eduard Hovy, and Quoc V Le. Self-training with noisy student improves imagenet classification. In CVPR, pages 10687-10698, 2020. 2 +[66] Haiming Xu, Lingqiao Liu, Qiuchen Bian, and Zhen Yang. Semi-supervised semantic segmentation with prototype-based consistency regularization. NeurIPS, 35:26007-26020, 2022. 2 +[67] Yi Xu, Lei Shang, Jinxing Ye, Qi Qian, Yu-Feng Li, Baigui Sun, Hao Li, and Rong Jin. Dash: Semi-supervised learning with dynamic thresholding. In ICML, pages 11525-11536. PMLR, 2021. 2 +[68] Lihe Yang, Wei Zhuo, Lei Qi, Yinghuan Shi, and Yang Gao. St++: Make self-trainingwork better for semi-supervised semantic segmentation. In CVPR, pages 4258-4267, 2022. 5, 6 +[69] Lihe Yang, Lei Qi, Litong Feng, Wayne Zhang, and Yinghuan Shi. Revisiting weak-to-strong consistency in semi-supervised semantic segmentation. In CVPR, pages 7236-7246, 2023. 2, 3, 5, 6 +[70] Lihe Yang, Zhen Zhao, and Hengshuang Zhao. Unimatch v2: Pushing the limit of semi-supervised semantic segmentation. IEEE TPAMI, 2025. 2 +[71] Sangdoo Yun, Dongyoon Han, Sanghyuk Chun, Seong Joon Oh, Youngjoon Yoo, and Junsuk Choe. Cutmix: Regularization strategy to train strong classifiers with localizable features. In ICCV, pages 6022-6031, 2019. 1 +[72] Bowen Zhang, Yidong Wang, Wenxin Hou, Hao Wu, Jindong Wang, Manabu Okumura, and Takahiro Shinozaki. Flexmatch: Boosting semi-supervised learning with curriculum pseudo labeling. NeurIPS, 34:18408-18419, 2021. 2 +[73] Xiangyu Zhao, Zengxin Qi, Sheng Wang, Qian Wang, Xuehai Wu, Ying Mao, and Lichi Zhang. Rcps: Rectified contrastive pseudo supervision for semi-supervised medical image segmentation. IEEE Journal of Biomedical and Health Informatics, 2023. 1 + +[74] Zhen Zhao, Lihe Yang, Sifan Long, Jimin Pi, Luping Zhou, and Jingdong Wang. Augmentation matters: A simple-yet-effective approach to semi-supervised semantic segmentation. In CVPR, pages 11350-11359, 2023. 5, 6 +[75] Barret Zoph, Golnaz Ghiasi, Tsung-Yi Lin, Yin Cui, Hanxiao Liu, Ekin Dogus Cubuk, and Quoc Le. Rethinking pretraining and self-training. NeurIPS, 33:3833-3845, 2020. 2 +[76] Yuliang Zou, Zizhao Zhang, Han Zhang, Chun-Liang Li, Xiao Bian, Jia-Bin Huang, and Tomas Pfister. Pseudoseg: Designing pseudo labels for semantic segmentation. In ICLR, 2020. 5, 6 \ No newline at end of file diff --git a/ICCV/2025/When Confidence Fails_ Revisiting Pseudo-Label Selection in Semi-supervised Semantic Segmentation/images.zip b/ICCV/2025/When Confidence Fails_ Revisiting Pseudo-Label Selection in Semi-supervised Semantic Segmentation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..d559483328a8eff9577697fd12707100a7256e16 --- /dev/null +++ b/ICCV/2025/When Confidence Fails_ Revisiting Pseudo-Label Selection in Semi-supervised Semantic Segmentation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:08b84b8e9ae7e2c6e0e454136faef5940746d7a2ae6f291e6ab40249ff4f8a5d +size 765674 diff --git a/ICCV/2025/When Confidence Fails_ Revisiting Pseudo-Label Selection in Semi-supervised Semantic Segmentation/layout.json b/ICCV/2025/When Confidence Fails_ Revisiting Pseudo-Label Selection in Semi-supervised Semantic Segmentation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..ae092a5cda5d657e033178cb596844c10d2cdc24 --- /dev/null +++ b/ICCV/2025/When Confidence Fails_ Revisiting Pseudo-Label Selection in Semi-supervised Semantic Segmentation/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d0ced602cb5f7b0880aa41ab67c213cbb5fd2ba6020df0fdcbe4f9056375f895 +size 498213 diff --git a/ICCV/2025/When Large Vision-Language Model Meets Large Remote Sensing Imagery_ Coarse-to-Fine Text-Guided Token Pruning/cc17937b-94d3-4352-8b5a-b4c323567743_content_list.json b/ICCV/2025/When Large Vision-Language Model Meets Large Remote Sensing Imagery_ Coarse-to-Fine Text-Guided Token Pruning/cc17937b-94d3-4352-8b5a-b4c323567743_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..1bae9227901caf8e586ba6fb3cb0e16c554f515a --- /dev/null +++ b/ICCV/2025/When Large Vision-Language Model Meets Large Remote Sensing Imagery_ Coarse-to-Fine Text-Guided Token Pruning/cc17937b-94d3-4352-8b5a-b4c323567743_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7d22d0b22e3b65a04a79f7cf9d41ab8c8de83d157c21df4067dff8c7a8e69afe +size 95789 diff --git a/ICCV/2025/When Large Vision-Language Model Meets Large Remote Sensing Imagery_ Coarse-to-Fine Text-Guided Token Pruning/cc17937b-94d3-4352-8b5a-b4c323567743_model.json b/ICCV/2025/When Large Vision-Language Model Meets Large Remote Sensing Imagery_ Coarse-to-Fine Text-Guided Token Pruning/cc17937b-94d3-4352-8b5a-b4c323567743_model.json new file mode 100644 index 0000000000000000000000000000000000000000..e3aaf1247ee99c74b5f8257957bddb11a73f661b --- /dev/null +++ b/ICCV/2025/When Large Vision-Language Model Meets Large Remote Sensing Imagery_ Coarse-to-Fine Text-Guided Token Pruning/cc17937b-94d3-4352-8b5a-b4c323567743_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:901842de9a52d6501336284121a13addc46ff7ab7c1a429594fb3c6695cf4a80 +size 126666 diff --git a/ICCV/2025/When Large Vision-Language Model Meets Large Remote Sensing Imagery_ Coarse-to-Fine Text-Guided Token Pruning/cc17937b-94d3-4352-8b5a-b4c323567743_origin.pdf b/ICCV/2025/When Large Vision-Language Model Meets Large Remote Sensing Imagery_ Coarse-to-Fine Text-Guided Token Pruning/cc17937b-94d3-4352-8b5a-b4c323567743_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..082cebe24e980aad1aa2461e51c3b517fbb9a602 --- /dev/null +++ b/ICCV/2025/When Large Vision-Language Model Meets Large Remote Sensing Imagery_ Coarse-to-Fine Text-Guided Token Pruning/cc17937b-94d3-4352-8b5a-b4c323567743_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1273140763c758a8e35572bf10ec113dd87e2c18c149d0ce88b81ba219668a7c +size 5946779 diff --git a/ICCV/2025/When Large Vision-Language Model Meets Large Remote Sensing Imagery_ Coarse-to-Fine Text-Guided Token Pruning/full.md b/ICCV/2025/When Large Vision-Language Model Meets Large Remote Sensing Imagery_ Coarse-to-Fine Text-Guided Token Pruning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..7cb11dfa8ccd9c01806f5dc46d9daced094caf49 --- /dev/null +++ b/ICCV/2025/When Large Vision-Language Model Meets Large Remote Sensing Imagery_ Coarse-to-Fine Text-Guided Token Pruning/full.md @@ -0,0 +1,416 @@ +# When Large Vision-Language Model Meets Large Remote Sensing Imagery: Coarse-to-Fine Text-Guided Token Pruning + +Junwei Luo $^{1*}$ , Yingying Zhang $^{2}$ , Xue Yang $^{3}$ , Kang Wu $^{1*}$ , Qi Zhu $^{2}$ , Lei Liang $^{2}$ , Jingdong Chen $^{2}$ , Yansheng Li $^{1\dagger}$ + +$^{1}$ Wuhan University $^{2}$ Ant Group $^{3}$ SAIS, Shanghai Jiao Tong University $\{\text {luojunwei }, \text {kangwu }, \text {yansheng.li} \} @ \text {whu.edu.cn }$ yangxue-2019-sjtu@sjtu.edu.cn + +https://github.com/VisionXLab/LRS-VQA + +# Abstract + +Efficient vision-language understanding of large Remote Sensing Images (RSIs) is meaningful but challenging. Current Large Vision-Language Models (LVLMs) typically employ limited pre-defined grids to process images, leading to information loss when handling gigapixel RSIs. Conversely, using unlimited grids significantly increases computational costs. To preserve image details while reducing computational complexity, we propose a text-guided token pruning method with Dynamic Image Pyramid (DIP) integration. Our method introduces: (i) a Region Focus Module (RFM) that leverages text-aware region localization capability to identify critical vision tokens, and (ii) a coarse-to-fine image tile selection and vision token pruning strategy based on DIP, which is guided by RFM outputs and avoids directly processing the entire large imagery. Additionally, existing benchmarks for evaluating LVLMs' perception ability on large RSI suffer from limited question diversity and constrained image sizes. We construct a new benchmark named LRS-VQA, which contains 7,333 QA pairs across 8 categories, with image length up to 27,328 pixels. Our method outperforms existing high-resolution strategies on four datasets using the same data. Moreover, compared to existing token reduction methods, our approach demonstrates higher efficiency under high-resolution settings. + +# 1. Introduction + +Benefiting from the rapid advancement of Large Language Models (LLMs) [2, 11, 14, 17, 75], Large Visual-Language Models (LVLMs) have demonstrated strong capabilities in perceiving and understanding visual information in text-based multimodal interactions [9, 29, 41, 65]. Currently, LVLMs have been extensively studied and applied across various fields like remote sensing (RS) intelligent interpre- + +![](images/1fb32b68eec8044dd58455e8dccdbc8e5eccc7ce6c4088b1c9305f6b4c9be3de.jpg) +Large RSI + +![](images/e861fd4ec9b73f6facff056d5d3b293f32ae37c7f4006fa145099b33a92e05ad.jpg) +Text Instruction + +![](images/33075a74dc2266efdcd6c605266c5938323e52c5f22c449615adb424ada1376c.jpg) +Text-Related Key Region + +![](images/91d0672d44ab42a974f32f20ab02732b47823d6aa594dfdbf898ec0ed30a9702.jpg) +(a) +Unlimited Grids: w/o Vis. Compression: + +![](images/7ca8c210441c6378034f93314e38cf688c452e80dd04de71c694ad49955b5784.jpg) +(b) +Blurred Detail Vision Part Speed LLM Part Speed +w/ Vis. Compression: +Blurred Detail Vision Part Speed LLM Part Speed + +![](images/e38c12a7819081646d610ffd537b8352b7b74e183cfe204f785f0676302c5fe6.jpg) +(c) +Dynamic Image Tiles: +Blurred Detail Vision Part Speed LLM Part Speed +Figure 1. High-resolution strategy comparison for modular LVLMs. (a) and (b) show that existing grid-based cropping methods face challenges when processing large RSIs. (c) The proposed dynamic pyramid-based token pruning strategy can dynamically select image tiles of key regions related to the input text, balancing image detail and computational cost. + +![](images/9a2097be83c4470c59f53c4f1b090b025cdbc8281c5f72116f68af99830ad7bf.jpg) + +![](images/ea62178a6e0da8b989c34b7923f02bfd059705bebd2ac86935ed51ac14b8107f.jpg) + +tation [12, 32, 53, 73, 87, 94]. + +Advances in satellite imaging technology allow for the acquisition of large remote sensing images (RSIs) that cover extensive ground areas and contain detailed land use information [18, 38]. Therefore, language-based intelligent analysis of large RSIs is valuable for applications such as complex scene understanding [37, 50], urban planning [35, 64], infrastructure development [70], and monitoring [93]. + +Some RS LVLMs handle high-resolution input through position embedding interpolation [27, 86]. Additionally, various dynamic high-resolution strategies designed for general LVLMs [33, 41, 80] can be adapted to RS LVLMs, such as EarthDial [61] (based on InternVL1.5 [10]) + +and GeoPixel[57] (using InternLM-XComposer2.5 [84]). However, these approaches for high-resolution image processing are limited when handling large RSIs exceeding $10,000 \times 10,000$ pixels. As depicted in Fig. 1, these methods typically employ limited pre-defined grid partitions, leading to detail loss when large RSIs are excessively downsampled. Conversely, unlimited grids suffer from prohibitive time and memory costs (e.g., over 5 million vision tokens for a large RSI in LLaVA-1.5 [41]). This necessitates balancing resolution and computational efficiency when processing large RSI. + +To overcome the above limitations of existing strategies in handling large RSIs, we propose a text-guided token pruning strategy that consists of two key components: a Region Focus Module (RFM) and a Dynamic Image Pyramid (DIP). The RFM identifies text-relevant key vision tokens by leveraging the capability distilled from the LLM component. Based on the RFM's output, we select critical image tiles and conduct token pruning in a coarse-to-fine manner within the DIP. This enables the LVLM to focus and zoom in key areas, avoiding to process all visual tokens, thereby facilitating flexible and efficient inference. + +Moreover, benchmarks for LVLMs in understanding large RSIs remain insufficiently developed. Although MME-Realworld [91] includes large RSIs with high-quality manually annotated questions, its RS subset suffers from limited question diversity and image sizes. To more comprehensively characterize the challenges of large RSI perception, we construct a new benchmark called LRS-VQA (Large Remote Sensing image Visual Question Answering). LRS-VQA features larger size images with lengths up to 27,328 pixels and 8 distinct question types. Our key contributions are as follows: + +- We introduce a Region Focus Module (RFM) that distills the ability from LLM part of LVLM to efficiently identify text-related key vision tokens. +- Based on RFM and DIP, we propose an efficient coarse-to-fine inference strategy for large RSIs, which achieves a balance between resolution and efficiency through key tile selection and vision token pruning. +- We construct a new benchmark named LRS-VQA, featuring larger image sizes and more diverse question types than existing benchmarks, to reflect the challenges of large RSIs perception. +- The proposed method is architecture-agnostic and shows performance improvements and efficiency gains. + +# 2. Related Work + +# 2.1. LVLMs for High-Resolution Comprehension + +Numerous LVLMs designed for high-resolution image or video perception have recently emerged. They can be primarily categorized into the following paradigms: + +Methods that carefully design image cropping and padding strategy. Methods like [10, 19, 29, 39, 41, 45, 80, 90] traverse pre-defined partition grids and obtain the most suitable grid, to divide the high-resolution image into image tiles. These methods encounter limitations when processing large RSIs as in Fig. 1. + +Methods based on multiple visual branches. These approach [34, 49, 63] employ additional high-resolution visual encoders like SAM [25] encoder or ConvNext [46] to process images with higher input resolution. + +Visual chain-of-thoughts methods. These approaches utilize a search-and-focus strategy [59, 60, 66, 69] or external RAG pipeline [92] to identify text-related key image regions. However, such pipelines require multiple LVLM inferences, making them more complex and cumbersome. + +# 2.2. Vision Token Reduction in LVLMs + +Visual token pruning for transformers has been a classic research topic, it aims to accelerate computation by dynamically reducing less important tokens. It has been extensively studied in both natural language processing [24, 79] and computer vision domains [16, 26, 51, 56, 68]. + +Recently, a variety of token pruning methods have been proposed for LVLMs, which can be categorized into three main types: token pruning conducted in the vision encoder [5, 33, 58, 67, 77, 78], token pruning performed in the LLM component [8, 20, 21, 72, 81, 88], and collaborative approaches that integrate both [6, 30, 31, 36, 74, 82, 85]. Moreover, adaptive token segmentation methods [1, 7, 28] also offer a promising approach for token reduction. + +Although vision-part pruning methods efficiently reduce vision tokens, when processing large images, it's highly time-consuming to traverse pre-defined grids and handle numerous image tiles with the vision encoder. Moreover, the absence of language guidance makes it hard to identify foreground regions in complex RSIs. The LLM-part pruning methods require feeding all vision tokens into the LLM. Therefore, long visual sequences from large images may exceed LLM length limits and incur high computational costs for attention score ranking. + +Among the collaborative approaches, the methods most closely related to ours are LLaVA-Mini [85] and FlexAttention [31]. The former employs modality pre-fusion to compress visual information; however, its LLM-like pre-fusion module still struggles with handling extremely long vision sequences. The latter integrates text-guided high-resolution features within the LLM component. Nevertheless, the allowable image size for its high-resolution view remains limited, and it can only index high-resolution features once. + +# 3. Preliminaries + +In this section, we present the foundational concepts of text-image attention within the LLM component of + +![](images/a91d23dc22c522ad0a76e041ee50bbfc80ccf28499ca5112bd84581546a40023.jpg) +Figure 2. The pipeline of the proposed method. The entire process iterates in a coarse-to-fine manner, dynamically retrieving high-resolution features from the next DIP level (leftward orange arrow) or performing token pruning at the current level (rightward orange arrow) based on the output of the RFM module at each iteration. During training, the RFM distillation text-related attention from the LLM; during inference, RFM generates the attention scores for the input vision tokens. GSD means ground sample distance. + +LVLMs. Mainstream modular LVLMs include three components [48]: a pre-trained vision encoder, a projector, and a pre-trained decoder-only LLM. Conventional high-resolution strategies [10, 19] process the original image $I_{img}$ into a thumbnail view $I_{\mathrm{thumb}}$ and a set of image tiles $I_{\mathrm{tiles}}$ . Then the vision encoder and the projector transforms $I_{\mathrm{thumb}}$ and $I_{\mathrm{tiles}}$ into vision token embeddings $T_{vis} = [T_{vis}^{lr}, T_{vis}^{hr}]$ , which contain low-resolution tokens $T_{vis}^{lr}$ from the $I_{\mathrm{thumb}}$ and higher-resolution $T_{vis}^{hr}$ from the $I_{\mathrm{tiles}}$ . $T_{vis}$ are concatenated with text token embeddings $T_{txt}$ from the tokenized text instruction, forming a multimodal token sequence $T = [T_{vis}, T_{txt}]$ . The LVLM processes $T$ via transformer-based decoder layers with causal self-attention, where the attention score $A$ can be represented as: + +$$ +A = \operatorname {A t t e n t i o n} (Q, K) = \operatorname {S o f t m a x} \left(\frac {Q K ^ {\top}}{\sqrt {d}}\right), \tag {1} +$$ + +where $A \in \mathbb{R}^{n \times n}$ is the self-attention map, $Q, K \in \mathbb{R}^{n \times d}$ are the query and key matrices derived from the input embeddings through linear transformations, $n$ is the input sequence length, and $d$ is the feature dimension. As $A$ determines the weight each token assigns to the former tokens in the sequence, by analyzing the attention from text tokens to vision tokens, we can identify text-related important vision + +tokens. Some studies [23, 81] observe that in the deep layers of LLM part in LVLMs, attention is predominantly allocated to text-related vision tokens compared to other vision tokens. This insight inspired us to design a coarse-to-fine mechanism to focus on key image regions. + +# 4. Method and Benchmark + +# 4.1. Method + +To achieve efficient understanding of ultra-high-resolution images, our core idea is coarse-to-fine focusing: we first gain a low-resolution overview, then iteratively zoom into text-related regions for details, as shown in Fig. 2. Implementing this necessitates a multi-resolution representation. Therefore, we first introduce the Dynamic Image Pyramid (DIP) in Sec. 4.1.1. Then we need to identify where to focus next. To solve this, we propose the Region Focus Module (RFM) to provide attention distribution for input vision tokens (Sec. 4.1.2). Finally, these components are integrated into an iterative refinement process (Sec. 4.1.3). + +# 4.1.1. The Construction of Dynamic Image Pyramid + +As established, our coarse-to-fine strategy requires a multiresolution representation of the image $I_{img}$ . We realize this through the DIP, a structure designed to provide image tiles + +![](images/726e46dccd0d10af9f291aad01ea5e6cb5d27cb606afb0581151583b1991224f.jpg) +Figure 3. The proposed RFM and attention distillation strategy. The left part indicates our core idea: distill accurate text-related key region localization ability from the LLM part of the LVLM. The right part shows the distillation details. We only select specific layer pairs for distillation to avoid hidden state discontinuities. "sys token" represents the tokens from the system prompt. + +![](images/4db6103da06e75842f93edef89e931f095ebc5ba14712524cd4ba8d2801e0cdd.jpg) + +at various ground sample distances (GSDs). + +First, we iteratively downsample $I_{img}$ by a factor of 2, generating a series of images $\{I_{init}^{1}, I_{init}^{2}, \ldots, I_{init}^{P}\}$ until the shorter side reaches a pre-defined minimum length. Let a basic image tile size be $B \times B$ (e.g., $336 \times 336$ for CLIP-L14), for the $p$ -th scaled image $I_{init}^{p}$ , we calculate the number of image tiles along the height and width as follows: + +$$ +N _ {h} ^ {p} = \lceil H _ {p} / B \rceil , \quad N _ {w} ^ {p} = \lceil W _ {p} / B \rceil , \tag {2} +$$ + +where $H_{p},W_{p}$ are the height and width of $I_{init}^{p}$ , respectively. For proper tiling, we compute a scale factor $r_p$ as: + +$$ +r _ {p} = \min \left(\frac {N _ {h} ^ {p} \times B}{H _ {p}}, \frac {N _ {w} ^ {p} \times B}{W _ {p}}\right). \tag {3} +$$ + +The image $I_{init}^{p}$ is then resized to $I^{p}$ using $r_p$ , followed by padding to preserve the original aspect ratio. After resizing, $I^{p}$ is partitioned into non-overlap tiles $I_{\mathrm{tiles}}^{p}$ . Tiles from all scaled images are combined in reverse order, and integrated with the thumbnail view $I_{\mathrm{thumb}}$ to form the final pyramid: $I_{DIP} = \{I_{\mathrm{thumb}}, I_{\mathrm{tiles}}^{1}, I_{\mathrm{tiles}}^{2}, \ldots, I_{\mathrm{tiles}}^{P}\}$ , which consists of $P + 1$ layers with progressively increasing resolutions. During training, we select the thumbnail $I_{\mathrm{thumb}}$ and $I_{\mathrm{tiles}}^{1}$ as visual inputs for computational efficiency. During inference, all pyramid layers become accessible to enable multi-scale inference, as in the left part of Fig. 2. + +# 4.1.2. Attention Distillation for Key Region Focus + +With the DIP established, the central challenge becomes generating a reliable guidance signal to determine where to focus next. While a sliding-window similarity map from + +VLMs like CLIP is a straightforward baseline, we posit that the attention maps from the LVLM's LLM part provide a superior localization signal. This is inspired by prior work [23, 72, 81] where LLM attention tends to converge on key visual tokens. To support this, we quantitatively validate that LLM attention significantly outperforms the CLIP/RemoteCLIP baselines by offering more accurate, text-related localization signals in Tab. 7. + +To harness this high-quality guidance efficiently, we introduce the Region Focus Module (RFM), a lightweight module trained via attention distillation to mimic the text-guided localization capability of the LLM teacher, as shown in Fig. 3. After distillation, the trained RFM can guide the selection of key image tiles and prune irrelevant vision tokens before the main LLM computation. + +Specifically, assuming the LLM has $M$ layers $l_{1}, l_{2}, \ldots, l_{M}$ , the proposed RFM adopts the same architecture with $R$ layers $r_{1}, r_{2}, \ldots, r_{R}$ , where $R < M$ . Each RFM layer $r_{i}$ is initialized from a selected LLM layer $l_{m_{i}}$ , where $\{m_{1}, m_{2}, \ldots, m_{R}\}$ forms a sparse subset of LLM layer indices. For layer-wise distillation, we select $K$ pairs $(K < R)$ from RFM and LLM as student-teacher pairs, represented as $\{(r_k, l_{m_k}) \mid k = 1, 2, \ldots, K\}$ . This strategy avoids rigidly aligning each $r_{i}$ with $l_{m_{i}}$ , which would neglect the non-teacher layers in the LLM, resulting in inconsistency between adjacent RFM layers. + +As depicted in the right part of Fig. 3, the concatenated multimodal tokens $T = [T_{vis}, T_{txt}]$ are fed into both the RFM and the LLM during training. For the $k$ -th student-teacher layer pair $(r_k, l_{mk})$ , we extract the attention scores between the last text token and all vision tokens from the + +multi-head self attention (MHSA). For the $h$ -th head, the self-attention is represented as $A_{\mathrm{stu}}^{k,h}, A_{\mathrm{tea}}^{k,h} \in \mathbb{R}^{j \times n_v}$ , where $j$ is the number of dialogue turns for multi-turn instruction and $n_v$ is the length of vision tokens. We apply the Kullback-Leibler (KL) divergence loss for distillation: + +$$ +\mathcal {L} _ {\mathrm {k l}} ^ {h} = \sum_ {k = 1} ^ {K} \left[ D _ {\mathrm {K L}} \left(A _ {\mathrm {t e a}} ^ {k, h} \| A _ {\mathrm {s t u}} ^ {k, h}\right) + \lambda_ {\mathrm {h r}} D _ {\mathrm {K L}} \left(A _ {\mathrm {t e a}} ^ {k, h, \mathrm {h r}} \| A _ {\mathrm {s t u}} ^ {k, h, \mathrm {h r}}\right) \right], \tag {4} +$$ + +where $A_{\mathrm{tea}}^{k,h,\mathrm{hr}}, A_{\mathrm{stu}}^{k,h,\mathrm{hr}}$ denote attention corresponding to $T_{vis}^{hr}$ , and $\lambda_{\mathrm{hr}}$ is the weighting factor. Additionally, we enforce stronger alignment constraints to the attention of $T_{vis}^{hr}$ by using an additional Mean Squared Error (MSE) loss: + +$$ +\mathcal {L} _ {\mathrm {m s e}} ^ {h} = \sum_ {k = 1} ^ {K} \mathrm {M S E} \left(A _ {\mathrm {s t u}} ^ {k, h, \mathrm {h r}}, A _ {\mathrm {t e a}} ^ {k, h, \mathrm {h r}}\right). \tag {5} +$$ + +The total distillation loss is computed as: + +$$ +\mathcal {L} _ {\text {d i s t i l l}} = \frac {1}{K H} \sum_ {h = 1} ^ {H} \left(\lambda_ {\mathrm {m s e}} \mathcal {L} _ {\mathrm {m s e}} ^ {h} + \lambda_ {\mathrm {k l}} \mathcal {L} _ {\mathrm {k l}} ^ {h}\right), \tag {6} +$$ + +where $K$ is the number of selected student-teacher layer pairs, $H$ is the number of attention heads, and $\lambda_{\mathrm{mse}}, \lambda_{\mathrm{kl}}$ are hyperparameters. For the flash-attention [13], we employ a specialized value matrix to extract attention maps like [88] to ensure compatibility. + +# 4.1.3. Text-Guided Token Pruning with Pyramid + +Having introduced the DIP and the RFM, this section details how they are integrated into a cohesive, iterative process. This process achieves our coarse-to-fine strategy by selectively retrieving high-resolution tiles and pruning irrelevant tokens, as illustrated in Fig. 2. + +The specific inference procedure, detailed in Alg. 1, is as follows. For simplicity, the tokens from the system prompt are omitted. We initialize the vision input using $I_{\mathrm{thmb}}$ and $I_{\mathrm{tiles}}^p$ to generate the initial $T_{vis} = [T_{vis}^{lr}, T_{vis}^{p,hr}]$ , where $p = 1$ , then the concatenated tokens $T$ are fed into the RFM. From the last layer of the RFM, we compute the average attention scores from all heads in the MHSA, and extract the attention scores $A_{\mathrm{stu}}^{K,hr}$ of $T_{\mathrm{vis}}^{p,hr}$ . By selecting the top- $\lambda$ proportion from $A_{\mathrm{stu}}^{K,hr}$ , we identify the indices of key tokens and map their coordinates to image tile-level coordinates, to select key image tiles $I_{\mathrm{key}}^{p+1}$ from $I_{\mathrm{tiles}}^{p+1}$ . Based on the number of $I_{\mathrm{key}}^{p+1}$ , we determine whether to prune the $T_{\mathrm{vis}}^{p,hr}$ directly or replace it with vision tokens $T_{\mathrm{vis}}^{p+1,hr}$ from higher-resolution tiles. If $T_{\mathrm{vis}}^{p+1,hr}$ is obtained, the tile selection process is repeated recursively until the last layer of DIP is reached. This approach enables LVLM to focus only on processing a few high-resolution image tiles in a coarse-to-fine manner, thereby reducing computational complexity while preserving critical text-related image details. + +# 4.2. The Construction of LRS-VQA Benchmark + +Considering the limitations of existing benchmarks for evaluating LVLMs' perception of RSIs in terms of question di + +Algorithm 1 Coarse-to-Fine Token Pruning with DIP +Require: $\{I_{\mathrm{tiles}}^1,I_{\mathrm{tiles}}^2,\dots,I_{\mathrm{tiles}}^{\mathrm{P}}\}$ in $I_{\mathrm{DIP}},K$ -layer RFM, initial multimodal tokens $T$ , vision encoder and projector $\mathcal{V}(\cdot)$ , token saving ratio $\alpha$ , tile number threshold $N_{\mathrm{max}}$ +Ensure: Retained tokens $T_{\mathrm{vis}}^{\mathrm{retain}}$ +1: Definition: Top- $\alpha (X)$ selects elements in $X$ with values in the top $\alpha$ proportion. +2: for $p = 1$ to $P$ do +3: Compute attention: $A_{\mathrm{stu}}^{K,hr}\gets \mathrm{RFM}(T)[T_{\mathrm{vis}}^{p,hr}]$ +4:Normalize: +5:Identify key positions: $\mathcal{P}_{\mathrm{key}}^{K,hr}\gets \mathrm{Top - }\alpha (A_{\mathrm{stu}}^{K,hr})$ +6:if $p < P$ then +7:Map coordinates and select key tiles: $I_{\mathrm{key}}^{p + 1}\gets \{t\in$ $I_{\mathrm{tiles}}^{p + 1}\mid$ mapped from $\mathcal{P}_{\mathrm{key}}^{K,hr}\}$ +8:if $|I_{\mathrm{key}}^{p + 1}| > N_{\mathrm{max}}$ then +9:Prune tokens: $T_{\mathrm{vis}}^{\mathrm{retain}}\gets \{t\in T_{\mathrm{vis}}^{p,hr}\mid \mathcal{P}_{\mathrm{key}}^{K,hr}\}$ +10: break +11: else +12:Encode selected tiles: $T_{\mathrm{vis}}^{p + 1,hr}\gets \mathcal{V}(I_{\mathrm{key}}^{p + 1})$ +13:Update $T\gets [T_{\mathrm{vis}}^{\mathrm{lr}},T_{\mathrm{vis}}^{p + 1,hr},T_{\mathrm{txt}}]$ +14:end if +15:else +16:Prune tokens at final layer: $T_{\mathrm{vis}}^{\mathrm{retain}}\gets \{t\in T_{\mathrm{vis}}^{p,hr}\mid$ $\mathcal{P}_{\mathrm{key}}^{K,hr}\}$ +17:end if +18:end for +19:return $T_{\mathrm{vis}}^{\mathrm{retain}}$ + +versity and image size, we propose a new benchmark for more comprehensive evaluation. + +# 4.2.1. Annotation Pipeline + +The annotation pipeline of LRS-VQA is illustrated in Fig. 4. We collect 3 publicly remote sensing datasets: FAIR1M-1.0 [62], GLH-Bridge [35], and STAR [37], and identify all unique targets based on the object detection labels to generate unique references from them (e.g., "the top-most ship" and "the largest dock"). Subsequently, we crop the region with the unique target from the original image, and feed it, along with a prompt containing the unique reference, into GPT-4V [54] to generate question-answer pairs about the target's color, shape, status, etc. Details of the overall construction process are provided in the Appendix A. + +# 4.2.2. Visual-Centric Quality Check and Refinement + +To ensure the quality of the generated question-answer pairs, we employ powerful Qwen2-VL [65] for quality assessment. Qwen2-VL supports manual adjustment of the maximum pixel resolution, allowing us to evaluate whether increased resolution improves accuracy while keeping the LLM unchanged. Using this approach, we filter out ques + +
DatasetImage SizeImage NumQues. Cate.Ques. NumQS Format
MME-RealWorld-RS [91]689-11500129833738Single-choice
LRS-VQA (Ours)1024-27328165787333Open-end
+ +Table 1. Comparison of existing benchmarks for evaluating LVLMs' perception capabilities in large RSIs. + +![](images/31ffc15e945f713aaaf9d4c6bdcda7eb146a5adb5cd953ed3debf32ac1e7af0b.jpg) +Figure 4. The construction pipeline of the proposed LRS-VQA dataset. The visual prompt (red box) is inspired by SoM [76]. + +tion types where accuracy can not improve with resolution. Through iterative filtering, manual review, and prompt refinement, we obtain the LRS-VQA dataset, which could effectively assess the large RSI perception capability of LVLMs, as shown in Fig. 5. LRS-VQA contains eight question-answer types: count, color, category, shape, status, reasoning, rural/urban classification, and target background. It features greater diversity in question types and larger image sizes than MME-Realworld-RS as in Tab. 1, highlighting the complexities of large RSI perception. + +# 5. Experiments + +# 5.1. Experimental Details + +Data. Our pre-training (PT) phase is the same as LLaVA-1.5 [42] (using 558K data). For the SFT phase, we collect 484K samples, including 300K sampled from LLaVA-1.5-665K, 146K sampled from RSVQA-HR [47], and 38K template-based samples from three RS datasets [15, 35, 37] using their labels. Within this 38k data, some images share the same source with LRS-VQA, but there is no overlap be + +![](images/a0c5174359dafbe94de165458c859bf79b4e0511650477026b30126d329e7173.jpg) +Figure 5. The accuracy trends of Qwen2-VL across varying input maximum pixels. This demonstrates that accuracy on both the manually annotated MME-RealWorld-RS and our proposed LRS-VQA exhibit a positive correlation with resolution improvement, proving the effectiveness of LRS-VQA in evaluating LVLM's high-resolution RSI perception capabilities. + +tween them. Details can be found in the Appendix B.1. + +Benchmarks and evaluation metrics: i) MME-RealWorld-RS: the RS part of MME-Realworld [91], containing 1,298 RSIs with expert-annotated questions in three types: color, count, and position. We follow the official evaluation script but modify the prompt by removing "The best answer is:" to address Vicuna-1.5-based models' tendency to respond with only option A. ii) LRS-VQA: it consists of 3 parts: LRS-FAIR, LRS-Bridge, and LRS-STAR, containing 2,272, 1,062, and 3,999 QA pairs, respectively. For the short open-ended format, we adopt a structured evaluation metric following [22, 71], using WordNet [52], with a semantic similarity threshold of 0.8. + +Experimental settings. We evaluate existing LVLMs using their official weights or APIs. For fair comparison, we also re-implement some methods on our 484k SFT dataset. Our approach is built upon the LLaVA-1.5 codebase, where we integrate the anyres strategy and Qwen2 language model. All experiments are conducted on 4 A100 80GB GPUs. For our method, the minimum length for DIP is 1,008 pixels. The $N_{\mathrm{max}}$ is set to 40 and 80 for Vicuna1.5 and Qwen2, respectively. The vision token saving ratio $\alpha$ is 0.25, $\lambda_{\mathrm{hr}}, \lambda_{\mathrm{mse}}, \lambda_{\mathrm{kl}}$ are set to 2.0, 1.0, 1.0. The RFMLL layer pairs are [1, 5, 11, 14] with distillation applied to the first and last pairs. Additional experimental details are provided in Appendix B.2. + +
LeaderboardDataVis. EncoderLLMMax SizeMME-RW-RSLRS-FAIRLRS-BridgeLRS-STARAvg.Acc
Qwen2-VL [65]-QwenViTQwen2-7B3,333×3,33349.7323.8038.1227.8734.88
LLaVA-OV [29]4.8MSigLipQwen2-7B2,304×2,30453.5320.6135.1126.0833.83
IXC-2.5 [84]-CLIP-L14InternLM2-7B4,096×4,09636.1225.2538.4127.3031.77
LLaVA-UHD-v2 [89]1.42MCLIP-L14+JBUVicuna-1.5-7B672×1,00840.7722.8232.5726.0830.56
LLaVA-FlexAttn [31]1.22MCLIP-L14Vicuna-1.5-7B1,008×1,00837.7519.5729.9922.7627.52
SEAL [69]0.95MCLIP-L14Vicuna-1.5-7B-30.5521.2934.7521.2926.97
MGM-HD [34]3.0MMixedVicuna-1.5-7B1,536×1,53631.5117.9035.9220.1326.36
SliME [90]2.0MCLIP-L14Vicuna-1.5-7B672×1,00828.5417.1132.0922.9925.18
LLaVA-1.5 [41]1.22MCLIP-L14Vicuna-1.5-7B336×33626.3818.7630.7022.6324.62
RSUniVLM [44]1.2MSigLIPQwen2-0.5B336×33625.3921.0232.6124.7225.93
Geochat [27]1.53MCLIP-L14Vicuna-1.5-7B504×50428.6220.1824.5413.7521.77
GPT-4o [3]----28.9222.1531.8427.4027.58
GPT-4o-mini [3]----6.6918.6731.9925.8520.80
Claude-3.5-Sonnet [4]----25.7412.9526.6913.2919.67
+ +
Comparison
LLaVA-1.5*1.04MCLIP-L14Vicuna-1.5-7B336×33634.4018.2432.2824.1727.30
LLaVA-1.5-p4*CLIP-L14672×67238.2019.1835.4326.5029.83
Geochat*CLIP-L14504×50433.6821.5135.9725.8629.25
SLiME*CLIP-L14672×1,00834.5621.9833.2025.1028.71
LLaVA-UHD-v2*CLIP-L14+JBU672×1,00838.5520.7736.5725.7930.42
Ours (LLaVA-1.5)CLIP-L14Dynamic39.0422.9736.8927.4831.59
LLaVA-Next-p4*1.04MCLIP-L14Qwen2-7B672×67239.1221.0636.2725.8030.56
LLaVA-Next-p9*1,008×1,00840.3521.1437.2526.1031.21
LLaVA-Next-p25*1,680×1,68039.6520.9936.3826.1830.80
Ours (LLaVA-Next)Dynamic41.8921.8538.2426.6732.16
+ +Table 2. Leaderboard and performance comparison on LRS-VQA and MME-RealWorld-RS. "MME-RW-RS" indicates the MMERealWorld-RS. "Data" refers to the total instruction data used during PT and SFT. $pX$ indicates the max number is $X$ for the pre-defined grids. "method*" indicates we reproduce the SFT stage of existing methods using the 484k instruction data. + +
SettingColorCountPositionAccFPS
anyres-p2541.5631.0546.1439.650.188
w/ PruMerge++ [58]43.2728.3832.7834.860.152
w/ VisionZip [77]42.7124.5534.5533.980.183
w/ FastV [8]39.5230.1043.9937.930.192
w/ PDrop [72]41.0030.5447.7739.850.184
w/ prune (CLIP)37.8027.6544.9036.860.171
w/ prune (RemoteCLIP)39.4029.8347.4538.770.148
w/ prune (Ours)43.9830.4249.1641.280.165
DIP-3layer44.0830.9449.2641.310.267
w/ prune (Ours)
+ +Table 3. Comparison with different token reduction methods based on LLaVA-Next-Qwen2. The "anyres-p25" setting means the maximum number of pre-defined grids is 25 (1680×1680 pixels). FPS is computed based on the inference time for continuous 800 images from MME-Realworld-RS. + +# 5.2. Leaderboard and Comparison + +Leaderboard on LRS-VQA. We evaluate a total of 11 open-source MLLMs, including methods for high-resolution [31, 42, 84, 90], chain-of-thought reasoning [69], multiple visual encoders [34], and RS MLLMs [27, 44]. We also evaluate 3 closed-source MLLMs. Results in Tab. 2 indicate that LRS-VQA is more challenging than MME-RealWorld + +
SettingVis. Tokens (Total)Vis. Tokens (to LLM)TFLOPs (B)
anyres-p1448352021312243.37
w/ PruMerge++ [58]83520532843.75
w/ VisionZip [77]83520928083.56
w/ FastV [8]8352021312109.21
w/ PDrop [72]8352021312101.62
w/ prune (Ours)83520576082.56
DIP-4layer55296237636.61
w/ prune (Ours)
+ +Table 4. Comparison of computational efficiency with different token reduction methods based on LLaVA-Next-Qwen2. We assume inferring a $4000 \times 4000$ pixel image (anyres-p144) and report the theoretical TFLOPs of vision tokens after the vision projector. + +RS on average, reflecting the complexities of large RSIs. + +Comparison with high-resolution methods. As shown in Tab. 2, we select 5 different high resolution strategies, performing SFT on the same dataset for comparison. Our method performs better across all 4 datasets. + +Comparison with token reduction methods. We reproduce 4 plug-and-play methods and establish 2 key baselines for comparison on the MME-RealWorld-RS. These baselines utilize sliding-window similarity maps generated + +
MethodDrop RatioColorCountPosAcc
anyres-p250%41.5631.0546.1439.65
25%42.4731.0847.4140.40
w/prune (Ours)50%42.3931.2448.4540.77
75%43.9830.4249.1641.28
90%41.5129.7746.7039.41
+ +Table 5. Ablation study on different prune ratios with LLaVA-Next-Qwen2, when pruning vision tokens based on RFM results without DIP. This table demonstrates that pruning irrelevant tokens in high-resolution RSIs can improve performance. + +
Max SizeColorCountPosAccFPS
1,008×1,00842.8729.8547.8940.290.271
2,016×2,01644.0830.9449.2641.310.267
4,032×4,03244.8629.8547.1840.720.221
8,064×8,06443.9830.4245.1139.910.204
Dynamic44.7031.0049.7241.890.238
+ +by CLIP-L14 [55] and RemoteCLIP-L14 [40], respectively. More details are in the Appendix B.2. As shown in Tab. 3, our method achieves the best performance under the fixed anyres-p25 resolution. It also secures the highest accuracy and FPS in the 3-layer DIP setting (up to $2,016 \times 2,016$ pixels). For Tab. 4, we report the average number of tiles selected by our dynamic method for images of similar sizes. These results suggest that existing methods often incur high computational costs when processing numerous text-irrelevant image tiles in large images. + +# 5.3. Ablation Results + +Different ratios for RFM-based pruning. As shown in Tab. 5, a key finding is that in large RSIs, the complex background and the small foreground regions enable high pruning rates to deliver performance benefits. + +Fixed DIP layers for inference. As shown in Tab. 6, we fix the number of DIP layers reached during inference. If an image's DIP is shallower than the specified number, the last layer is used. Results show that forcing traversal to higher resolutions can degrade performance since not all questions require image details. Therefore, our strategy dynamically selects the termination DIP layer based on the input text, balancing accuracy and efficiency. + +Attention Localization validation. We select LRSVQA (using boxes of cropped regions) and two referring-based datasets, DIOR-RSVG [83] and RRSIS-D [43] for localization accuracy evaluation, using a recall metric: a prediction is considered successful if the vision tokens retained + +Table 6. Ablation study on fixing the number of DIP layers reached in inference with LLaVA-Next-Qwen2. Rows 1-4 correspond to fixing the number of DIP layers being 2-5, respectively. + +
Pruning GuidanceLRS-VQADIOR-RSVGRRSIS-D
Teacher LLM Attn.56.1474.8173.88
CLIP Sim.20.1429.1027.53
RemoteCLIP Sim.30.4143.1241.81
RFM Attn. (Ours)47.9964.7661.17
+ +Table 7. Localization recall $(\%)$ of different pruning guidance. + +![](images/e4004ace97f8a28e6166900cb316cf0470b65e22d4780f82d503f980a95ac5e0.jpg) +Large-size RS1 4005×2919 pixels + +![](images/aa072b6c281aff405d495424696ce767fdac3e7a50a9bbcf510adfafd6125355.jpg) +Question: What is the color of the largest runway in the image? Answer: Grey. + +![](images/6dd8473e1a0a6f2d2112e199c7d988b4e45b86140aededdef54b6c126a613941.jpg) +Question: Where is the largest parking lot in the image? Answer: Bottom-left. + +![](images/067a7990b8b58886d7248725a8a2849b21c30f046c3901d2c81153da8fc603b0.jpg) +Large-size RS1 5448×4076 pixels +Figure 6. Text-related visual attention localization in the last layer if trained RFM under LLaVA-Next-Qwen2. + +![](images/d181562a97f00161f331dbe33729b1c1f8e6a804ce6fc2737f7ec69496768702.jpg) +Question: What is the shape of the parking lot in the image? Answer: Oval. + +![](images/bdb9fbb393d15823ca5ad96fb2c5df5171395959edbd705339f3955c0f16a049.jpg) +Question: How many branches does the roundabout have? Answer: 5. + +after pruning cover more than $50\%$ of the ground-truth region. The results in Tab. 7 demonstrate that at a $25\%$ token retention ratio, both the teacher LLM's attention and our student RFM achieve higher localization accuracy than the CLIP-L14 and RemoteCLIP-L14 baselines (also based on sliding-window similarity maps). This confirms that LLM attention provides a reliable localization signal and our distillation is effective. + +Visualization results. In Fig. 6, we visualize the attention maps generated by the trained RFM for vision tokens from initial input image tiles. The results demonstrate that text references to targets effectively guide the RFM's attention output, supporting our coarse-to-fine inference. + +# 6. Conclusion + +This paper presents a text-guided token pruning method tailored for efficiently processing large remote sensing images (RSIs). To address the limitations of existing high-resolution approaches when handling large RSIs, our method integrates the Region Focus Module (RFM) and Dynamic Image Pyramid (DIP) to focus on text-relevant key vision tokens while reducing computational overhead. Furthermore, we introduce a new benchmark, LRS-VQA, with 7,333 QA pairs covering 8 question types, and features larger image sizes and a diverse range of question types. + +Future work. Vision-language understanding of large RSIs remains a challenging task. Investigating approaches such as long-context transfer from LLMs to LVLMs, or chain-of-thought reasoning, is also meaningful. + +# Acknowledgment + +The authors would like to thank the anonymous reviewers for their valuable comments. This work was supported by the National Natural Science Foundation of China under Grants 42371321, 424B2006, and the Natural Science Foundation of Shanghai under Grant No. 25ZR1402268. + +# References + +[1] Marius Aasan, Odd Kolbjørnsen, Anne Schistad Solberg, and Adín Ramirez Rivera. A spitting image: Modular superpixel tokenization in vision transformers. In European Conference on Computer Vision, pages 124-142. Springer, 2024. 2 +[2] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 1 +[3] Open AI. Hello gpt-4o. 2024. 7 +[4] AI Anthropic. Claude 3.5 sonnet model card addendum. page 3, 2024. 7 +[5] Kazi Hasan Ibn Arif, JinYi Yoon, Dimitrios S. Nikolopoulos, Hans Vandierendonck, Deepu John, and Bo Ji. Hired: Attention-guided token dropping for efficient inference of high-resolution vision-language models, 2024. 2 +[6] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023. 2 +[7] Delong Chen, Samuel Cahyawijaya, Jianfeng Liu, Baoyuan Wang, and Pascale Fung. Subobject-level image tokenization. CoRR, abs/2402.14327, 2024. 2 +[8] Liang Chen, Haozhe Zhao, Tianyu Liu, Shuai Bai, Junyang Lin, Chang Zhou, and Baobao Chang. An image is worth 1/2 tokens after layer 2: Plug-and-play inference acceleration for large vision-language models. In European Conference on Computer Vision, pages 19-35. Springer, 2024. 2, 7 +[9] Zhe Chen, Weiyun Wang, Yue Cao, Yangzhou Liu, Zhangwei Gao, Erfei Cui, Jinguo Zhu, Shenglong Ye, Hao Tian, Zhaoyang Liu, et al. Expanding performance boundaries of open-source multimodal models with model, data, and test-time scaling. arXiv preprint arXiv:2412.05271, 2024. 1 +[10] Zhe Chen, Weiyun Wang, Hao Tian, Shenglong Ye, Zhang-wei Gao, Erfei Cui, Wenwen Tong, Kongzhi Hu, Jiapeng Luo, Zheng Ma, et al. How far are we to gpt-4v? closing the gap to commercial multimodal models with open-source suites. Science China Information Sciences, 67(12):220101, 2024. 1, 2, 3 +[11] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhang-hao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yong-hao Zhuang, Joseph E. Gonzalez, Ion Stoica, and Eric P. Xing. Vicuna: An open-source chatbot impressing gpt-4 with $90\%$ * chatgpt quality, 2023. 1 +[12] Muhammad Sohail Danish, Muhammad Akhtar Munir, Syed Roshaan Ali Shah, Kartik Kuckreja, Fahad Shahnbaz Khan, + +Paolo Fraccaro, Alexandre Lacoste, and Salman Khan. Geobench-vlm: Benchmarking vision-language models for geospatial tasks. arXiv preprint arXiv:2411.19325, 2024. 1 +[13] Tri Dao, Dan Fu, Stefano Ermon, Atri Rudra, and Christopher Ré. Flashattention: Fast and memory-efficient exact attention with io-awareness. Advances in neural information processing systems, 35:16344-16359, 2022. 5 +[14] DeepSeek-AI, Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, et al. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning, 2025. 1 +[15] Jian Ding, Nan Xue, Gui-Song Xia, Xiang Bai, Wen Yang, Michael Ying Yang, Serge Belongie, Jiebo Luo, Mihai Datcu, Marcello Pelillo, et al. Object detection in aerial images: A large-scale benchmark and challenges. IEEE transactions on pattern analysis and machine intelligence, 44(11): 7778-7796, 2021. 6 +[16] Peiyan Dong, Mengshu Sun, Alec Lu, Yanyue Xie, Kenneth Liu, Zhenglun Kong, Xin Meng, Zhengang Li, Xue Lin, Zhenman Fang, et al. Heatvit: Hardware-efficient adaptive token pruning for vision transformers. In 2023 IEEE International Symposium on High-Performance Computer Architecture (HPCA), pages 442–455. IEEE, 2023. 2 +[17] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 1 +[18] Xin Guo, Jiangwei Lao, Bo Dang, Yingying Zhang, Lei Yu, Lixiang Ru, Liheng Zhong, Ziyuan Huang, Kang Wu, Dingxiang Hu, et al. Skysense: A multi-modal remote sensing foundation model towards universal interpretation for earth observation imagery. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27672-27683, 2024. 1 +[19] Zonghao Guo, Ruyi Xu, Yuan Yao, Junbo Cui, Zanlin Ni, Chunjiang Ge, Tat-Seng Chua, Zhiyuan Liu, and Gao Huang. Llava-uhd: an Imm perceiving any aspect ratio and high-resolution images. In European Conference on Computer Vision, pages 390–406. Springer, 2025. 2, 3 +[20] Yuhang Han, Xuyang Liu, Pengxiang Ding, Donglin Wang, Honggang Chen, Qingsen Yan, and Siteng Huang. Rethinking token reduction in mllms: Towards a unified paradigm for training-free acceleration. arXiv preprint arXiv:2411.17686, 2024. 2 +[21] Runhui Huang, Xinpeng Ding, Chunwei Wang, Jianhua Han, Yulong Liu, Hengshuang Zhao, Hang Xu, Lu Hou, Wei Zhang, and Xiaodan Liang. Hires-llava: Restoring fragmentation input in high-resolution large vision-language models. arXiv preprint arXiv:2407.08706, 2024. 2 +[22] Drew A Hudson and Christopher D Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition, pages 6700-6709, 2019. 6 +[23] Omri Kaduri, Shai Bagon, and Tali Dekel. What's in the image? a deep-dive into the vision of vision language models. arXiv preprint arXiv:2411.17491, 2024. 3, 4 + +[24] Sehoon Kim, Sheng Shen, David Thorsley, Amir Gholami, Woosuk Kwon, Joseph Hassoun, and Kurt Keutzer. Learned token pruning for transformers. In Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 784-794, 2022. 2 +[25] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4015-4026, 2023. 2 +[26] Zhenglun Kong, Peiyan Dong, Xiaolong Ma, Xin Meng, Wei Niu, Mengshu Sun, Xuan Shen, Geng Yuan, Bin Ren, Hao Tang, et al. Spvit: Enabling faster vision transformers via latency-aware soft token pruning. In European conference on computer vision, pages 620-640. Springer, 2022. 2 +[27] Kartik Kuckreja, Muhammad Sohail Danish, Muzammal Naseer, Abhijit Das, Salman Khan, and Fahad Shahbaz Khan. Geochat: Grounded large vision-language model for remote sensing. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2024. 1, 7 +[28] Jaihyun Lew, Soohyuk Jang, Jaehoon Lee, Seungryong Yoo, Eunjji Kim, Saehyung Lee, Jisoo Mok, Siwon Kim, and Sungroh Yoon. Superpixel tokenization for vision transformers: Preserving semantic integrity in visual tokens. arXiv preprint arXiv:2412.04680, 2024. 2 +[29] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Peiyuan Zhang, Yanwei Li, Ziwei Liu, et al. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 1, 2, 7 +[30] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730–19742. PMLR, 2023. 2 +[31] Junyan Li, Delin Chen, Tianle Cai, Peihao Chen, Yining Hong, Zhenfang Chen, Yikang Shen, and Chuang Gan. Flex-attention for efficient high-resolution vision-language models. In European Conference on Computer Vision, pages 286–302. Springer, 2024. 2, 7 +[32] Qingyun Li, Yushi Chen, Xinya Shu, Dong Chen, Xin He, Yu Yi, and Xue Yang. A simple aerial detection baseline of multimodal language models. arXiv preprint arXiv:2501.09720, 2025. 1 +[33] Wentong Li, Yuqian Yuan, Jian Liu, Dongqi Tang, Song Wang, Jie Qin, Jianke Zhu, and Lei Zhang. Tokenpacker: Efficient visual projector for multimodal llm. arXiv preprint arXiv:2407.02392, 2024. 1, 2 +[34] Yanwei Li, Yuechen Zhang, Chengyao Wang, Zhisheng Zhong, Yixin Chen, Ruihang Chu, Shaoteng Liu, and Jiaya Jia. Mini-gemini: Mining the potential of multi-modality vision language models. arXiv:2403.18814, 2023. 2, 7 +[35] Yansheng Li, Junwei Luo, Yongjun Zhang, Yihua Tan, JinGang Yu, and Song Bai. Learning to holistically detect bridges from large-size vhr remote sensing imagery. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 1, 5, 6 +[36] Yanwei Li, Chengyao Wang, and Jiaya Jia. Llama-vid: An image is worth 2 tokens in large language models. 2024. 2 + +[37] Yansheng Li, Linlin Wang, Tingzhu Wang, Xue Yang, Junwei Luo, Qi Wang, Youming Deng, Wenbin Wang, Xian Sun, Haifeng Li, Bo Dang, Yongjun Zhang, Yi Yu, and Yan Junchi. Star: A first-ever dataset and a large-scale benchmark for scene graph generation in large-size satellite imagery. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 1, 5, 6 +[38] Yansheng Li, Yuning Wu, Gong Cheng, Chao Tao, Bo Dang, Yu Wang, Jiahao Zhang, Chuge Zhang, Yiting Liu, Xu Tang, et al. Meet: A million-scale dataset for fine-grained geospatial scene classification with zoom-free remote sensing imagery. IEEE/CAA Journal of Automatica Sinica, 12(5):1004-1023, 2025. 1 +[39] Zhang Li, Biao Yang, Qiang Liu, Zhiyin Ma, Shuo Zhang, Jingxu Yang, Yabo Sun, Yuliang Liu, and Xiang Bai. Monkey: Image resolution and text label are important things for large multi-modal models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26763-26773, 2024. 2 +[40] Fan Liu, Delong Chen, Zhangqingyun Guan, Xiaocong Zhou, Jiale Zhu, Qiaolin Ye, Liyong Fu, and Jun Zhou. Remoteclip: A vision language foundation model for remote sensing. IEEE Transactions on Geoscience and Remote Sensing, 2024. 8 +[41] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. arXiv preprint arXiv:2310.03744, 2023. 1, 2, 7 +[42] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in Neural Information Processing Systems, 36, 2024. 6, 7 +[43] Sihan Liu, Yiwei Ma, Xiaqing Zhang, Haowei Wang, Jiayi Ji, Xiaoshuai Sun, and Rongrong Ji. Rotated multi-scale interaction network for referring remote sensing image segmentation. In CVPR, 2024. 8 +[44] Xu Liu and Zhouhui Lian. *Rsunivlm: A unified vision language model for remote sensing via granularity-oriented mixture of experts.* arXiv preprint arXiv:2412.05679, 2024. 7 +[45] Yuliang Liu, Biao Yang, Qiang Liu, Zhang Li, Zhiyin Ma, Shuo Zhang, and Xiang Bai. Textmonkey: AnOCR-free large multimodal model for understanding document. arXiv preprint arXiv:2403.04473, 2024. 2 +[46] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2 +[47] Sylvain Lobry, Diego Marcos, Jesse Murray, and Devis Tuia. Rsvqa: Visual question answering for remote sensing data. IEEE Transactions on Geoscience and Remote Sensing, 58 (12):8555-8566, 2020. 6 +[48] Gen Luo, Xue Yang, Wenhan Dou, Zhaokai Wang, Jiawen Liu, Jifeng Dai, Yu Qiao, and Xizhou Zhu. Mono-internvl: Pushing the boundaries of monolithic multimodal large language models with endogenous visual pre-training. arXiv preprint arXiv:2410.08202, 2024. 3 +[49] Gen Luo, Yiyi Zhou, Yuxin Zhang, Xiawu Zheng, Xiaoshuai Sun, and Rongrong Ji. Feast your eyes: Mixture-of + +resolution adaptation for multimodal large language models. arXiv preprint arXiv:2403.03003, 2024. 2 +[50] Junwei Luo, Zhen Pang, Yongjun Zhang, Tingzhu Wang, Linlin Wang, Bo Dang, Jiangwei Lao, Jian Wang, Jingdong Chen, Yihua Tan, et al. Skysensegpt: A fine-grained instruction tuning dataset and model for remote sensing vision-language understanding. arXiv preprint arXiv:2406.10100, 2024.1 +[51] Lingchen Meng, Hengduo Li, Bor-Chun Chen, Shiyi Lan, Zuxuan Wu, Yu-Gang Jiang, and Ser-Nam Lim. Adavit: Adaptive vision transformers for efficient image recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12309-12318, 2022. 2 +[52] George A Miller. Wordnet: a lexical database for english. Communications of the ACM, 38(11):39-41, 1995. 6 +[53] Dilxat Muhtar, Zhenshi Li, Feng Gu, Xueliang Zhang, and Pengfeng Xiao. Lhrs-bot: Empowering remote sensing with vgi-enhanced large multimodal language model. In European Conference on Computer Vision, pages 440-457. Springer, 2024. 1 +[54] OpenAI. Gpt-4v(ision) system card. 2023. 5 +[55] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 8 +[56] Yongming Rao, Wenliang Zhao, Benlin Liu, Jiwen Lu, Jie Zhou, and Cho-Jui Hsieh. Dynamicvit: Efficient vision transformers with dynamic token sparsification. Advances in neural information processing systems, 34:13937-13949, 2021. 2 +[57] Akashah Shabbir, Mohammed Zumri, Mohammed Bennamoun, Fahad S Khan, and Salman Khan. Gepixel: Pixel grounding large multimodal model in remote sensing. arXiv preprint arXiv:2501.13925, 2025. 2 +[58] Yuzhang Shang, Mu Cai, Bingxin Xu, Yong Jae Lee, and Yan Yan. Llava-prumerge: Adaptive token reduction for efficient large multimodal models. arXiv preprint arXiv:2403.15388, 2024. 2, 7 +[59] Hao Shao, Shengju Qian, Han Xiao, Guanglu Song, Zhuofan Zong, Letian Wang, Yu Liu, and Hongsheng Li. Visual cot: Advancing multi-modal language models with a comprehensive dataset and benchmark for chain-of-thought reasoning. Advances in Neural Information Processing Systems, 37:8612-8642, 2025. 2 +[60] Haozhan Shen, Kangjia Zhao, Tiancheng Zhao, Ruochen Xu, Zilun Zhang, Mingwei Zhu, and Jianwei Yin. Zoomeye: Enhancing multimodal llms with human-like zooming capabilities through tree-based image exploration. arXiv preprint arXiv:2411.16044, 2024. 2 +[61] Sagar Soni, Akshay Dudhane, Hiyam Debary, Mustansar Fiaz, Muhammad Akhtar Munir, Muhammad Sohail Danish, Paolo Fraccaro, Campbell D Watson, Lerente J Klein, Fahad Shahbaz Khan, et al. Earthdial: Turning multi-sensory earth observations to interactive dialogues. arXiv preprint arXiv:2412.15190, 2024. 1 + +[62] Xian Sun, Peijin Wang, Zhiyuan Yan, Feng Xu, Ruiping Wang, Wenhui Diao, Jin Chen, Jihao Li, Yingchao Feng, Tao Xu, et al. Fair1m: A benchmark dataset for fine-grained object recognition in high-resolution remote sensing imagery. ISPRS Journal of Photogrammetry and Remote Sensing, 184:116-130, 2022. 5, 1 +[63] Peter Tong, Ellis Brown, Penghao Wu, Sanghyun Woo, Adithya Jairam Vedagiri IYER, Sai Charitha Akula, Shusheng Yang, Jihan Yang, Manoj Middepogu, Ziteng Wang, et al. Cambrian-1: A fully open, vision-centric exploration of multimodal llms. Advances in Neural Information Processing Systems, 37:87310-87356, 2025. 2 +[64] Junjue Wang, Zhuo Zheng, Zihang Chen, Ailong Ma, and Yanfei Zhong. Earthvqa: Towards queryable earth via relational reasoning-based remote sensing visual question answering. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 5481-5489, 2024. 1 +[65] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 1, 5, 7 +[66] Wenbin Wang, Liang Ding, Minyan Zeng, Xiabin Zhou, Li Shen, Yong Luo, and Dacheng Tao. Divide, conquer and combine: A training-free framework for high-resolution image perception in multimodal large language models. arXiv preprint, 2024. 2 +[67] Xidong Wang, Dingjie Song, Shunian Chen, Chen Zhang, and Benyou Wang. Longlava: Scaling multi-modal llms to 1000 images efficiently via a hybrid architecture. arXiv preprint arXiv:2409.02889, 2024. 2 +[68] Siyuan Wei, Tianzhu Ye, Shen Zhang, Yao Tang, and Jiajun Liang. Joint token pruning and squeezing towards more aggressive compression of vision transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2092-2101, 2023. 2 +[69] Penghao Wu and Saining Xie. V*: Guided visual search as a core mechanism in multimodal llms. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13084-13094, 2024. 2, 7 +[70] Aoran Xiao, Weihao Xuan, Junjue Wang, Jiaxing Huang, Dacheng Tao, Shijian Lu, and Naoto Yokoya. Foundation models for remote sensing and earth observation: A survey. arXiv preprint arXiv:2410.16602, 2024. 1 +[71] Junbin Xiao, Xindi Shang, Angela Yao, and Tat-Seng Chua. Next-qa: Next phase of question-answering to explaining temporal actions. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9777–9786, 2021. 6 +[72] Long Xing, Qidong Huang, Xiaoyi Dong, Jiajie Lu, Pan Zhang, Yuhang Zang, Yuhang Cao, Conghui He, Jiaqi Wang, Feng Wu, et al. Pyramiddrop: Accelerating your large vision-language models via pyramid visual redundancy reduction. arXiv preprint arXiv:2410.17247, 2024. 2, 4, 7 +[73] Xizhe Xue, Guoting Wei, Hao Chen, Haokui Zhang, Feng Lin, Chunhua Shen, and Xiao Xiang Zhu. Reo-vlm: Transforming vlm to meet regression challenges in earth observation. arXiv preprint arXiv:2412.16583, 2024. 1 + +[74] Dawei Yan, Pengcheng Li, Yang Li, Hao Chen, Qingguo Chen, Weihua Luo, Wei Dong, Qingsen Yan, Haokui Zhang, and Chunhua Shen. Tg-llava: Text guided llava via learnable latent embeddings. arXiv preprint arXiv:2409.09564, 2024. 2 +[75] An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, Huan Lin, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Yang, Jiaxi Yang, Jingren Zhou, Junyang Lin, Kai Dang, Keming Lu, Keqin Bao, Kexin Yang, Le Yu, Mei Li, Mingfeng Xue, Pei Zhang, Qin Zhu, Rui Men, Runji Lin, Tianhao Li, Tingyu Xia, Xingzhang Ren, Xuancheng Ren, Yang Fan, Yang Su, Yichang Zhang, Yu Wan, Yuqiong Liu, Zeyu Cui, Zhenru Zhang, and Zihan Qiu. Qwen2.5 technical report. arXiv preprint arXiv:2412.15115, 2024. 1 +[76] Jianwei Yang, Hao Zhang, Feng Li, Xueyan Zou, Chunyuan Li, and Jianfeng Gao. Set-of-mark prompting unleashes extraordinary visual grounding in gpt-4v, 2023. 6 +[77] Senqiao Yang, Yukang Chen, Zhuotao Tian, Chengyao Wang, Jingyao Li, Bei Yu, and Jiaya Jia. Visionzip: Longer is better but not necessary in vision language models. arXiv preprint arXiv:2412.04467, 2024. 2, 7 +[78] Linli Yao, Lei Li, Shuhuai Ren, Lean Wang, Yuanxin Liu, Xu Sun, and Lu Hou. Deco: Decoupling token compression from semantic abstraction in multimodal large language models. arXiv preprint arXiv:2405.20985, 2024. 2 +[79] Deming Ye, Yankai Lin, Yufei Huang, and Maosong Sun. Tr-bert: Dynamic token reduction for accelerating bert inference. arXiv preprint arXiv:2105.11618, 2021. 2 +[80] Jiabo Ye, Anwen Hu, Haiyang Xu, Qinghao Ye, Ming Yan, Guohai Xu, Chenliang Li, Junfeng Tian, Qi Qian, Ji Zhang, et al. Ureader: UniversalOCR-free visually-situated language understanding with multimodal large language model. In The 2023 Conference on Empirical Methods in Natural Language Processing, 2023. 1, 2 +[81] Weihao Ye, Qiong Wu, Wenhao Lin, and Yiyi Zhou. Fit and prune: Fast and training-free visual token pruning for multi-modal large language models. arXiv preprint arXiv:2409.10197, 2024. 2, 3, 4 +[82] Xubing Ye, Yukang Gan, Xiaoke Huang, Yixiao Ge, Ying Shan, and Yansong Tang. Voco-llama: Towards vision compression with large language models. arXiv preprint arXiv:2406.12275, 2024. 2 +[83] Yang Zhan, Zhitong Xiong, and Yuan Yuan. Rsvg: Exploring data and models for visual grounding on remote sensing data. IEEE TGRS, 2023. 8 +[84] Pan Zhang, Xiaoyi Dong, Yuhang Zang, Yuhang Cao, Rui Qian, Lin Chen, Qipeng Guo, Haodong Duan, Bin Wang, Linke Ouyang, Songyang Zhang, Wenwei Zhang, Yining Li, Yang Gao, Peng Sun, Xinyue Zhang, Wei Li, Jingwen Li, Wenhai Wang, Hang Yan, Conghui He, Xingcheng Zhang, Kai Chen, Jifeng Dai, Yu Qiao, Dahua Lin, and Jiaqi Wang. Internlm-xcomposer-2.5: A versatile large vision language model supporting long-contextual input and output. arXiv preprint arXiv:2407.03320, 2024. 2, 7 +[85] Shaolei Zhang, Qingkai Fang, Zhe Yang, and Yang Feng. Llava-mini: Efficient image and video large mul + +timodal models with one vision token. arXiv preprint arXiv:2501.03895, 2025. 2 +[86] Wei Zhang, Miaoxin Cai, Tong Zhang, Yin Zhuang, Jun Li, and Xuerui Mao. Earthmarker: A visual prompting multimodal large language model for remote sensing. IEEE Transactions on Geoscience and Remote Sensing, 2024. 1 +[87] Wei Zhang, Miaoxin Cai, Tong Zhang, Yin Zhuang, and Xuerui Mao. Earthgpt: A universal multi-modal large language model for multi-sensor image comprehension in remote sensing domain. arXiv preprint arXiv:2401.16822, 2024. 1 +[88] Yuan Zhang, Chun-Kai Fan, Junpeng Ma, Wenzhao Zheng, Tao Huang, Kuan Cheng, Denis Gudovskiy, Tomoyuki Okuno, Yohei Nakata, Kurt Keutzer, et al. Sparsevm: Visual token sparsification for efficient vision-language model inference. arXiv preprint arXiv:2410.04417, 2024. 2, 5 +[89] Yipeng Zhang, Yifan Liu, Zonghao Guo, Yidan Zhang, Xuesong Yang, Chi Chen, Jun Song, Bo Zheng, Yuan Yao, Zhiyuan Liu, et al. Llava-uhd v2: an mllm integrating high-resolution feature pyramid via hierarchical window transformer. arXiv preprint arXiv:2412.13871, 2024. 7 +[90] Yi-Fan Zhang, Qingsong Wen, Chaoyou Fu, Xue Wang, Zhang Zhang, Liang Wang, and Rong Jin. Beyond llava-hd: Diving into high-resolution large multimodal models. arXiv preprint arXiv:2406.08487, 2024. 2, 7 +[91] Yi-Fan Zhang, Huanyu Zhang, Haochen Tian, Chaoyou Fu, Shuangqing Zhang, Junfei Wu, Feng Li, Kun Wang, Qing-song Wen, Zhang Zhang, et al. Mme-realworld: Could your multimodal llm challenge high-resolution real-world scenarios that are difficult for humans? arXiv preprint arXiv:2408.13257, 2024. 2, 6 +[92] Zilun Zhang, Haozhan Shen, Tiancheng Zhao, Yuhao Wang, Bin Chen, Yuxiang Cai, Yongheng Shang, and Jianwei Yin. Enhancing ultra high resolution remote sensing imagery analysis with imagerag. arXiv preprint arXiv:2411.07688, 2024. 2 +[93] Yue Zhou, Litong Feng, Yiping Ke, Xue Jiang, Junchi Yan, Xue Yang, and Wayne Zhang. Towards vision-language geo-foundation models: A survey. arXiv preprint arXiv:2406.09385, 2024. 1 +[94] Qi Zhu, Jiangwei Lao, Deyi Ji, Junwei Luo, Kang Wu, Yingying Zhang, Lixiang Ru, Jian Wang, Jingdong Chen, Ming Yang, et al. Skysense-o: Towards open-world remote sensing interpretation with vision-centric visual-language modeling. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 14733-14744, 2025. 1 \ No newline at end of file diff --git a/ICCV/2025/When Large Vision-Language Model Meets Large Remote Sensing Imagery_ Coarse-to-Fine Text-Guided Token Pruning/images.zip b/ICCV/2025/When Large Vision-Language Model Meets Large Remote Sensing Imagery_ Coarse-to-Fine Text-Guided Token Pruning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..d0bc55554a1d3466aec71723fb6ff780850ab4f9 --- /dev/null +++ b/ICCV/2025/When Large Vision-Language Model Meets Large Remote Sensing Imagery_ Coarse-to-Fine Text-Guided Token Pruning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:248edf9ef422e6b5ed8f513dfc05c3421793728ed9ec890f0fcc20bda945d467 +size 895436 diff --git a/ICCV/2025/When Large Vision-Language Model Meets Large Remote Sensing Imagery_ Coarse-to-Fine Text-Guided Token Pruning/layout.json b/ICCV/2025/When Large Vision-Language Model Meets Large Remote Sensing Imagery_ Coarse-to-Fine Text-Guided Token Pruning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..53347a82ba197af18b2a8b6750331402db4f7fa9 --- /dev/null +++ b/ICCV/2025/When Large Vision-Language Model Meets Large Remote Sensing Imagery_ Coarse-to-Fine Text-Guided Token Pruning/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:50e7e55893e48e25dba30b2231a8000e8f26529bc81d6e62e71ee22794ca00bc +size 541114 diff --git a/ICCV/2025/When Lighting Deceives_ Exposing Vision-Language Models' Illumination Vulnerability Through Illumination Transformation Attack/7ce1661a-7863-4397-8464-08916ed25d3b_content_list.json b/ICCV/2025/When Lighting Deceives_ Exposing Vision-Language Models' Illumination Vulnerability Through Illumination Transformation Attack/7ce1661a-7863-4397-8464-08916ed25d3b_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..31c8dc3bc759c1666e896876e630e59523b112de --- /dev/null +++ b/ICCV/2025/When Lighting Deceives_ Exposing Vision-Language Models' Illumination Vulnerability Through Illumination Transformation Attack/7ce1661a-7863-4397-8464-08916ed25d3b_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6b9a915d0afbde4d94f3eb3c483c5b6e837cb563a39441220062e764328b588b +size 84491 diff --git a/ICCV/2025/When Lighting Deceives_ Exposing Vision-Language Models' Illumination Vulnerability Through Illumination Transformation Attack/7ce1661a-7863-4397-8464-08916ed25d3b_model.json b/ICCV/2025/When Lighting Deceives_ Exposing Vision-Language Models' Illumination Vulnerability Through Illumination Transformation Attack/7ce1661a-7863-4397-8464-08916ed25d3b_model.json new file mode 100644 index 0000000000000000000000000000000000000000..4d2582f9bb5f7d3aaccc378a9887b3bea6e39b11 --- /dev/null +++ b/ICCV/2025/When Lighting Deceives_ Exposing Vision-Language Models' Illumination Vulnerability Through Illumination Transformation Attack/7ce1661a-7863-4397-8464-08916ed25d3b_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:887e0842770f9e034bee6d59ad35311cc7f66cc48573bed7a75c53fc79bfbabd +size 106350 diff --git a/ICCV/2025/When Lighting Deceives_ Exposing Vision-Language Models' Illumination Vulnerability Through Illumination Transformation Attack/7ce1661a-7863-4397-8464-08916ed25d3b_origin.pdf b/ICCV/2025/When Lighting Deceives_ Exposing Vision-Language Models' Illumination Vulnerability Through Illumination Transformation Attack/7ce1661a-7863-4397-8464-08916ed25d3b_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..4c065e2bd9b24deaa8d243c1cb1bc1d8753abd40 --- /dev/null +++ b/ICCV/2025/When Lighting Deceives_ Exposing Vision-Language Models' Illumination Vulnerability Through Illumination Transformation Attack/7ce1661a-7863-4397-8464-08916ed25d3b_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c95fcd1fd2445079f0096afc2231afa91b75bf1c87d9e32e3dedb2962e3c7a6a +size 3744000 diff --git a/ICCV/2025/When Lighting Deceives_ Exposing Vision-Language Models' Illumination Vulnerability Through Illumination Transformation Attack/full.md b/ICCV/2025/When Lighting Deceives_ Exposing Vision-Language Models' Illumination Vulnerability Through Illumination Transformation Attack/full.md new file mode 100644 index 0000000000000000000000000000000000000000..ad55314b751736dad9ca99f872f1e063fba5fe1b --- /dev/null +++ b/ICCV/2025/When Lighting Deceives_ Exposing Vision-Language Models' Illumination Vulnerability Through Illumination Transformation Attack/full.md @@ -0,0 +1,378 @@ +# When Lighting Deceives: Exposing Vision-Language Models' Illumination Vulnerability Through Illumination Transformation Attack + +Hanqing Liu†, Shouwei Ruan†, Yao Huang†, Shiji Zhao, Xingxing Wei* + +Institute of Artificial Intelligence, + +State Key Laboratory of Virtual Reality Technology and Systems, Beihang University + +{hqliu, shouweiruan, y_huang, zhaoshijil23, xxwei}@buaa.edu.cn + +Natural Images with Normal Illumination + +![](images/abfefa09956178663e91c0769c25efcc01b221fcf8508e12a0341702d91f9c1b.jpg) +stop sign $99.9\%$ + +![](images/bd32010fc25593eae85f7ba98d9e103d2ba1fe891526deb75f992f86ed09922b.jpg) +bus $89.1\%$ + +![](images/99d95db6b78a608246f6529ac900a98b00608ef6450202a48ad48cc7f6060e56.jpg) +traffic light $99.6\%$ + +![](images/d3b9ca586a22b7305edbe2a72be5ba007b2d0dcb9a725037d599c91a28206f9f.jpg) +microwave $92.7\%$ + +![](images/884450f63bfe504167ef31f32b9a1fffe7664b5778682372e610354c1c1188e4.jpg) +dog $87.5\%$ + +Natural Images with OOD Illumination + +![](images/5372349b759721894856fa42587ec0fd42a9b9025c3aa1fdb9c75d193349c5d0.jpg) +stop sign $23.9\%$ + +![](images/66c29e3460c846327300685021ea549640da7cb0eed81d56d5438beb133dfa59.jpg) +truck $85.0\%$ + +![](images/7784fdc8a16345c07342d499ca75f121ace519c54c109f7566be3e21fe7008a7.jpg) +train $16.2\%$ + +![](images/2ed3bc922516b34ea74743a80b7d84c53fb2270dfc6e9f583a5bd3955df8ec6a.jpg) +refrigerator $27.6\%$ + +![](images/55bcde461047689fbaf447e6c14398e4362af87c6548c7432351d459ed407e8f.jpg) +cat $60.7\%$ + +Illumination-aware Adversarial Example Generated by ITA (ours) + +![](images/5510642eb42ace3b0520bad3eab9482b2bc81af66d07a7ff521b38115f00a435.jpg) +traffic light + +![](images/efde2fe4b5253f7a4339ece058fa5cb81dc08f71cf754ab926cb2255e702492b.jpg) + +![](images/611d3b3cf7f6d4ff79d60cca64f0a4becef56b34fef939135e3c69fdbffea6be.jpg) +X bird + +![](images/b8107a972d26abed34033ada1f63c450991b4a173716b09858a048bdffd10298.jpg) +car +Figure 1. Revealing VLM Vulnerabilities to Illumination Variations. Comparison of the predictions made by OpenCLIP ViT-B/16 on natural images with normal and OOD illumination versus illumination-aware adversarial examples generated by our Illumination Transformation Attack (ITA), which illustrate VLMs are vulnerable to illumination changes, underscoring the need for systematic evaluation. + +![](images/3fbd757c771815ed749e598542039da1c50368e7b9b801d7e5692ba95aaa733d.jpg) + +![](images/39d06cf6a8f79e6a7a83b65ae2dddd7fc82f2b95432f2a6d2c1ee101bdc66eee.jpg) +X +raf +ic light + +![](images/400e5e39f08e7a246343b2fab777654d48727bf522376c6a59c5f41ccc2ce542.jpg) +minivan + +![](images/d47768da21028242b01076ae5baa15dda1aec6b9bbc75473c58cb3229d24ae5f.jpg) + +![](images/0aee2391809be52a4ea071d1c4243c4b4343444755a96c1df848defc2b6277df.jpg) +X stingray + +# Abstract + +Vision-Language Models (VLMs) have achieved remarkable success in various tasks, yet their robustness to real-world illumination variations remains largely unexplored. To bridge this gap, we propose Illumination Transformation Attack (ITA), the first framework to systematically assess VLMs' robustness against illumination changes. However, there still exist two key challenges: (1) how to model global illumination with fine-grained control to achieve diverse lighting conditions and (2) how to ensure adversarial effectiveness while maintaining naturalness. To address the first challenge, we innovatively decompose global illumination into multiple parameterized point light sources based on the illumination rendering equation. This design enables us to model more diverse lighting variations that previous + +methods could not capture. Then, by integrating these parameterized lighting variations with physics-based lighting reconstruction techniques, we could precisely render such light interactions in the original scenes, finally meeting the goal of fine-grained lighting control. For the second challenge, by controlling illumination through the lighting reconstruction model's latent space rather than direct pixel manipulation, we inherently preserve physical lighting priors. Furthermore, to prevent potential reconstruction artifacts, we design additional perceptual constraints for maintaining visual consistency with original images and diversity constraints for avoiding light source convergence. Extensive experiments demonstrate that our ITA could significantly reduce the performance of advanced VLMs, e.g., LLaVA-1.6, while possessing competitive naturalness, exposing VLMS' critical illumination vulnerabilities. + +# 1. Introduction + +Recently, Vision-Language Models (VLMs) have achieved remarkable progress in various tasks, such as image captioning [3, 21] and visual question answering [6, 25, 26]. Due to their superior performance, these models are increasingly being deployed in real-world applications like autonomous driving [42, 47] and security surveillance [1, 35]. However, in such real-world applications, they may inevitably face challenging environmental variations [2, 10, 37, 38]. Among them, illumination changes are the most ubiquitous and unavoidable one. As shown in Fig. 1, while VLMs perform well under normal illumination, they often misclassify images with Out-Of-Distribution (OOD) illumination. These illumination shifts can dramatically alter scene perception without being immediately noticeable, posing significant security risks. Thus, it is urgent to prioritize and systematically investigate VLMs' illumination robustness across varying environmental conditions. + +Existing research has made some efforts in exploring illumination impacts, particularly for specific tasks like traffic sign recognition [10, 11, 14, 41, 52] and face recognition [22, 33]. They simulate illumination variations by generating localized light disturbances like spots, reflections, or laser interference. Nevertheless, they are primarily designed for CNN-based models and show significant limitations when applied to VLMs. The methods either employ overly simplistic illumination modifications (static lighting changes or shadow effects) or generate artifacts easily identifiable by VLMs. Moreover, current illumination attacks are lack of diversity, failing to capture the rich spectrum of real-world lighting conditions and variations. + +Based on the above discussions, this paper aims to generate more diverse and natural illumination-aware adversarial examples for comprehensively evaluating VLMs' illumination robustness. However, there still exist two key challenges that must be addressed: (1) How to model global illumination with fine-grained control to achieve diverse lighting conditions that reflect real-world scenarios? (2) How to ensure adversarial effectiveness while maintaining physical realism and natural appearance? + +To address these challenges, we propose a novel framework named Illumination Transformation Attack (ITA). The ITA consists of two primary components: First, to achieve fine-grained illumination modeling, we innovatively decompose global illumination into multiple parameterized point light sources following the illumination rendering equation [18]. This approach allows us to express the illumination as a linear sum of contributions from individual light sources rather than a single light spot, enabling us to model diverse lighting variations that previous methods could not capture. Then, by leveraging physics-based lighting reconstruction techniques IC-Light [48], we can accurately reconstruct illumination examples corresponding to + +specific illumination variations, ultimately achieving the intention of fine-grained lighting control. + +Second, our approach prioritizes maintaining natural appearance in the generated images. By controlling illumination through IC-Light's latent space rather than direct pixel manipulation, we inherently preserve physical lighting priors and avoid unnatural lighting that violates physical constraints. Moreover, to prevent uncontrollable effects from the diffusion model's randomness, such as overexposure and underexposure, we implement perceptual constraints that maintain visual consistency with original images, alongside diversity constraints that prevent multiple light sources from converging. This paradigm ensures the generated illumination variations remain physically plausible while effectively revealing vulnerabilities in VLMs. + +In addition, IC-Light's non-differentiable nature resulting from diffusion models and denoising operations may prevent gradient-based optimization for adversarial attacks. Thus, to overcome this limitation, we employ the Covariance Matrix Adaptation Evolution Strategy (CMA-ES) [34], a gradient-free evolutionary optimization algorithm, to iteratively optimize the illumination variation. Finally, through extensive experiments, we demonstrate that illumination-aware adversarial examples generated by ITA significantly degrade the performance of advanced VLMs under optimized illumination conditions, revealing previously unidentified vulnerabilities. The contributions of this paper can be summarized as follows: + +- We propose a novel framework called ITA for generating illumination-aware adversarial examples, marking the first work to explore the illumination robustness of VLMs. It transcends previous methods' limitations, enabling more comprehensive illumination variations. +- We innovatively formulate the generation of complex illumination scenarios as a parameterized optimization problem, enabling precise control over lighting conditions while generating illumination-aware adversarial examples that are both natural and deceptive. +- Experimental results demonstrate that our illumination-aware adversarial examples successfully expose illumination sensitivity across various VLMs, including CLIP and its variants, as well as more complex generative VLMs such as LLaVA-1.6, BLIP-2, and Instruct-BLIP. + +# 2. Related Work + +# 2.1. Vision-Language Models + +Owing to large-scale pretraining, Vision-Language Models (VLMs) [20, 25, 26, 31] have emerged as powerful tools for processing multimodal information, demonstrating exceptional capabilities in tasks requiring joint understanding of visual and textual inputs, e.g., image captioning, visual question answering (VQA), and complex reasoning tasks + +that bridge visual perception with language comprehension. Despite their impressive achievements [27, 36, 50], concerns remain regarding their robustness, particularly in challenging edge cases and safety-critical applications. Recent studies [29, 39, 43] reveal that VLMs still struggle with fine-grained visual attributes, such as spatial orientation, numerical understanding, and color perception, which may stem from inherent limitations in their visual representations. Additionally, VLMs are vulnerable to adversarial perturbations in the $L_{p}$ -norm space [8, 32]. + +While previous works have explored risks caused by challenging visual attributes or adversarial perturbation, there remains a critical gap in evaluating VLMs' robustness under natural visual attribute like illumination variations, which is a fundamental aspect of real-world perception. To bridge it, our ITA will systematically examine how changes in illumination conditions impact VLM performance. + +# 2.2. Illumination Adversarial Attacks + +Existing research on illumination-based adversarial attacks has explored several approaches to evaluate visual model's robustness. Early methods focused on direct light manipulation, including projected light patterns [11, 16], laser beam interference [10], naturalistic adversarial patches [15, 28, 30], and shadow manipulations [52]. These optical attacks, however, face significant limitations in bright environments where environmental illumination overwhelms the manipulated light signals. More recent approaches have attempted to address these limitations. Wang et al. [44] developed reflection-based adversarial illumination strategies, while Hsiao et al. [14] introduced natural light attack methods that maintain effectiveness under stronger ambient lighting conditions. Despite these advances, current illumination attack research remains constrained in several ways: (1) evaluations have primarily targeted specific applications like traffic sign recognition; (2) most studies focus on conventional CNN-based vision models rather than modern Vision-Language Models (VLMs), and (3) existing methods typically optimize for single light sources, failing to capture the real-world illumination variations. + +In contrast, our proposed ITA framework extends adversarial illumination attacks to broader and more practical scenarios. Through simultaneously optimizing multiple light sources, our approach significantly broadens the spectrum of illumination variations, enabling a comprehensive evaluation of VLMs' robustness to illumination conditions. Moreover, our approach ensures high perceptual realism, making the adversarial examples more natural and indistinguishable from real-world illumination conditions. + +# 3. Methodology + +In this section, we introduce the proposed ITA framework, as presented in Fig. 2. The ITA framework will use the + +IC-Light technique to generate illumination transformation samples with given illumination variations. Thus, we begin by providing background on IC-Light, followed by a discussion of illumination modeling and optimization strategy. + +# 3.1. Preliminary + +Given an input image $X$ , the conditional inputs consist of an optional text input $T$ and an optional light source input $X_{l}$ . The vanilla IC-Light model employs a variational auto encoder (VAE) $\mathcal{E}_e$ and a CLIP text encoder $E_{t}$ to map the inputs into a latent space, which can be formulated as: + +$$ +z _ {i} = \mathcal {E} _ {e} (X, X _ {l}), \quad z _ {t} = E _ {t} (T). \tag {1} +$$ + +The encoded representations are then concatenated and passed as cross-attention inputs to the UNet. The UNet undergoes a diffusion process and generates a transformed output, which is decoded back into pixel space using the decoder $\mathcal{E}_d$ , producing illumination variations in the image, which can be simply expressed as: + +$$ +X ^ {\prime} = \mathcal {E} _ {d} (\operatorname {D i f f} \left(z _ {i}, z _ {t}\right)), \tag {2} +$$ + +where $X'$ denotes the generated illumination transformation samples, and $\mathrm{Diff}(\cdot)$ represents the diffusion process. However, while this method is effective for reconstructing illumination transformations, it lacks fine-grained control over localized illumination conditions and is non-differentiable due to the inclusion of diffusion models. To address this, we will model the illumination distribution in Sec. 3.2 and convert it into an optimization problem in Sec. 3.3. + +# 3.2. Illumination Modeling + +To achieve precise control over illumination conditions, we innovatively model the scene lighting using multiple point light sources. Each light source is characterized by four key parameters: spatial position $(x,y)$ , intensity $I$ , and radius $r$ of influence. The complete lighting configuration $\Lambda$ represents our parameterization of the illumination environment: + +$$ +\boldsymbol {\Lambda} = \left\{\left(x _ {i}, y _ {i}, I _ {i}, r _ {i}\right) \right\} _ {i = 1} ^ {N}, \tag {3} +$$ + +where $N$ is the number of light sources. Then, to simulate realistic light behavior, we apply a Gaussian function to model the spatial distribution of each light source: + +$$ +F _ {i} (z) = I _ {i} \cdot \exp \left(- \frac {\| z - \left(x _ {i} , y _ {i}\right) \| ^ {2}}{2 r _ {i} ^ {2}}\right), \tag {4} +$$ + +where $z$ represents any point in the scene, with $F_{i}(z)$ quantifying the illumination contribution from the $i$ -th light source at that point. The parameters $(x_{i},y_{i})$ , $I_{i}$ , and $r_i$ from $\Lambda$ directly control the center position, brightness, and spread of each light source respectively. + +![](images/b49a72c17691cc00746fc779764c787dc61400cecc09e70ab7e346a938e66cde.jpg) +Figure 2. Overview of the Proposed ITA Framework. Three previously proposed illumination attack methods rely on localized lighting disturbances (top). In contrast, our method applies illumination transformations across the entire scene, generating illumination-aware adversarial examples, ensuring both naturalness and adversariality (bottom). + +The total illumination at any point is calculated as the sum of all individual light contributions: + +$$ +F (z) = \sum_ {i = 1} ^ {N} F _ {i} (z). \tag {5} +$$ + +This formulation enables us to generate diverse and realistic lighting effects by adjusting the parameters within $\Lambda$ . + +Above all, the benefit of our illumination modeling is twofold. First, unlike traditional approaches that struggle with optimizing complex nonlinear light distributions, our method preserves essential physical light properties while making the parameters easier to optimize. Second, we address limitations of previous illumination attack methods that typically use binary masks (1 for illuminated regions, 0 for non-illuminated areas) to simulate lighting effects locally, which often lead to visual discontinuities. Our Gaussian-based model holistically reconstructs illumination across the entire scene rather than applying isolated modifications, eliminating the artifacts and inconsistencies common in localized lighting modifications. + +# 3.3. Optimization Objective of ITA Framework + +Adversarial Objective. The adversarial objective of our illumination transformation attack (ITA) is to learn an optimal adversarial illumination configuration $\Lambda^{*}$ that generates illumination conditions capable of significantly degrading the performance of VLMs, which can be formulated as: + +$$ +\boldsymbol {\Lambda} ^ {*} = \arg \max _ {\boldsymbol {\Lambda}} \mathcal {L} _ {A d v} (\mathcal {I} (X, \boldsymbol {\Lambda}), Y), \tag {6} +$$ + +where $\mathcal{I}(X,\Lambda)$ denotes the relighting process applying the illumination configuration $\Lambda$ to the original image $X$ , $Y$ + +represents the truth label and $\mathcal{L}_{Adv}(\cdot)$ is designed to quantify the effectiveness of the adversarial process. + +In this work, we conduct all experiments with different versions of CLIP as the target model, leveraging its image encoder and text encoder to guide the adversarial optimization. Specifically, our adversarial attack objective is inspired by traditional adversarial attacks on classification models, where the goal is to reduce the predicted probability of the correct class. Given that CLIP can be viewed as an implicit classifier operating in an embedding space, we formulate our objective in a similar manner. To sum up, we minimize the probability assigned by the model to the ground-truth semantic label, effectively pushing the generated adversarial image $X'$ towards a misclassification. + +Given an adversarial image $X'$ and a semantic label set $Y = \{y_i\}_{i=1}^N$ containing the ground-truth label $y_t$ , we first compute a semantic similarity vector as follows: + +$$ +V = \left\{\cos \left[ E _ {v} \left(X ^ {\prime}\right), E _ {t} \left(Y _ {i}\right) \right] \right\} _ {i = 1} ^ {N}, \tag {7} +$$ + +where $\cos(\cdot)$ denotes cosine similarity, $E_v(\cdot)$ and $E_t(\cdot)$ represent the image encoder and text encoder, respectively. The vector $V$ captures the alignment between $X'$ and each semantic label. We then compute the probability distribution over semantic labels using the softmax function. Finally, the adversarial loss $\mathcal{L}_{Adv}$ is defined as the negative log-likelihood of the ground-truth semantic label: + +$$ +\mathcal {L} _ {A d v} \left(X ^ {\prime}, Y\right) = - \log \left(\frac {\exp \left(V _ {t}\right)}{\sum_ {j = 1} ^ {N} \exp \left(V _ {j}\right)}\right). \tag {8} +$$ + +This objective directly reduces the confidence of CLIP in assigning the correct label, making it more likely to misclassify $X'$ . At the same time, we need to ensure that the + +generated images do not suffer from extreme illumination variations that lead to overexposure or underexposure. To address this, we propose a perceptual constraint. Additionally, when optimizing with multiple light sources, we introduce a distance regularization loss as a diversity constraint to prevent the converging of light sources that are too close to each other. The complete optimization further incorporates these constraints to maintain visual realism and encourage diverse light placements. + +Regularization for Perception and Diversity. To ensure the perceptual naturalness of illumination adversarial examples, we impose constraints on both perception and diversity. For the perceptual constraint, we use the LPIPS (Learned Perceptual Image Patch Similarity) loss [49], which compares deep features from a pre-trained network to ensure that the adversarial image remains visually similar to the original. LPIPS aligns more closely with human perception than traditional metrics like L2 or SSIM, preserving structural details and texture consistency. This minimizes distortions, improving the stealthiness of the attack and making it harder for VLMs to distinguish adversarial examples from natural images, which can be formulated as: + +$$ +\mathcal {L} _ {\text {P e c p}} \left(X, X ^ {\prime}\right) = - \sum_ {l} \| f _ {l} (X) - f _ {l} \left(X ^ {\prime}\right) \| _ {2} ^ {2}, \tag {9} +$$ + +where $f_{l}(\cdot)$ denotes the feature of the image at layer $l$ of the pre-trained network. + +To further enhance the effectiveness and diversity of the adversarial attack, we introduce a distance regularization loss to address the converging of multiple light sources. Without proper constraints, the optimization process may collapse into a degenerate solution, where all light sources converge to a single point, reducing the attack to a simple single-light perturbation. To ensure sufficient spatial separation between light sources and maintain the adversarial effectiveness, we define the distance regularization loss as: + +$$ +\mathcal {L} _ {D i s} = \sum_ {i = 1} ^ {N} \sum_ {j = i + 1} ^ {N} \max (0, \delta - \| \boldsymbol {\Lambda} _ {i} - \boldsymbol {\Lambda} _ {j} \|), \tag {10} +$$ + +where $\delta$ is a predefined minimum distance threshold. This penalty term encourages spatial separation between light sources, promoting a more diverse and effective adversarial lighting configuration. Finally, the total loss function combines the adversarial loss, the perceptual loss, and the distance loss with weighting factors $\alpha$ and $\beta$ as follow: + +$$ +\mathcal {L} _ {\text {T o t a l}} = \mathcal {L} _ {\text {A d v}} + \alpha \cdot \mathcal {L} _ {\text {P e c p}} + \beta \cdot \mathcal {L} _ {\text {D i s}}. \tag {11} +$$ + +# 3.4. Optimization Algorithm + +In this section, we present our illumination optimization algorithm, which aims to obtain the best illumination configuration $\pmb{\Lambda}^{*}$ . Rather than optimizing for a single configuration, we explore a distribution of configurations, which + +Algorithm 1: Optimization Algorithm +Data: Original images $X$ , Semantic label list $Y$ , Number of light sources $N$ , Initial distribution parameters $\mu_0,\bar{C}_0,\Sigma_0$ Weight coefficients $\alpha ,\beta$ Result: Adversarial illumination configuration $\Lambda^{*}$ Adversarial illumination samples $X^{\prime}$ / Image-Text Pairs Preparation \*/ +1 $\{X_0,Y_0\} \leftarrow \{X,Y\}$ . /Initialization of distribution parameters \*/ +2 $\mu_0\gets 0,C_0(\Sigma_0)^2\gets \mathbf{I};$ +3 while t < tmax do /Sampling Generate adversarial illumination parameters \*/ +4 for i = 1 → K do q+1 ∼ N(μt, Ct(Σt)2); +6 $\Lambda_{i}^{t + 1}\gets \mathbf{A}\cdot \tanh (\mathbf{q}_{i}^{t + 1}) + \mathbf{B};$ (X')+1 I(X,At+1); / Calculate loss value \*/ +8 L+1 = LAdv((X')+1,Y) + α·Lpecp + β·LDis; if should stop() then break +10 end +11 end +12 Evaluation: Compute the objective function values for each sample \*/ +13 Sorting L+1 in ascending order; / Update: Distribution parameters update \*/ +14 $\mu_{t + 1}\gets \sum_{i = 1}^{K}w_{i}\mathbf{q}_{i}^{t + 1};$ C+1 (1-cc)Ct+c∑W(i(q+1-μt)(q+1-μt)T; +16 $\Sigma_{t + 1}\gets \Sigma_t\cdot \exp \left(c_\sigma \left(\frac{\|\mathbf{p}\|}{\mathbb{E}[\|\mathcal{N}(0,I)\|]} -1\right)\right);$ +17 end +18 $\mu^{*}\gets \mu_{t\mathrm{max}},C^{*}\gets C_{t\mathrm{max}},\Sigma^{*}\gets \Sigma_{t\mathrm{max}};$ +19 q' $\sim \mathcal{N}(\boldsymbol {\mu}^*,\boldsymbol {C}^* (\Sigma^*)^2)$ +20 A' $\leftarrow$ A · tanh(q*) + B, X' $\leftarrow$ I(X, A\*) + +offers several key advantages: 1) It enables a more diverse exploration of possible results, helping to identify regions of the configuration space where optimal illumination conditions may exist. 2) It also facilitates faster convergence to a region containing the best illumination configuration, reducing the risk of getting trapped in local optima and enhancing robustness by avoiding overfitting to a solution, thus ensuring better generalization across different VLMs. + +Specifically, our method optimizes adversarial illumination distributions using the Covariance Matrix Adaptation Evolution Strategy (CMA-ES) [12, 13], a gradient-free optimization algorithm well-suited for this task. Since it does not rely on gradient information, CMA-ES is ideal for optimizing adversarial illumination distributions. To further enhance the optimization process, we incorporate two techniques: Learning Rate Adaptation (LRA) and Early Stopping Policy. LRA dynamically adjusts the learning rate during the search to improve convergence speed, while Early Stopping Policy halts the optimization when further improvements are unlikely, thus preventing unnecessary iterations and saving computational resources. + +The adversarial illumination configuration is parameterized by a Gaussian variable $\mathbf{q}$ , which follows a normal dis + +Table 1. Performance Degradation (↓) of VLMs on Zero-shot Classification Under Illumination-aware Adversarial Examples Generated from COCO Dataset. We compare the accuracy and natural scores evaluated by GPT-4o of previous methods with our proposed approach. The performance for ACC is considered better with lower values, while higher values are preferred for GPT-4o. Number in bold indicates the best performance and the second-best result is underlined. + +
MethodOpenCLIP ViT-B/16Meta-CLIP ViT-L/14EVA-CLIP ViT-G/14OpenAI CLIP ViT-L/14
ACC(%)GPT-4oACC(%)GPT-4oACC(%)GPT-4oACC(%)GPT-4o
Clean973.65983.65983.65933.65
Natural Light Attack [14]94(↓3)2.32(↓1.33)97(↓1)2.14(↓1.51)97(↓1)2.16(↓1.49)93(-)2.07(↓1.58)
Shadow Attack [52]84(↓13)2.11(↓1.54)82(↓16)1.91(↓1.74)95(↓3)2.03(↓1.62)79(↓14)1.78(↓1.87)
ITA(random)71(↓26)2.56(↓1.09)81(↓17)2.24(↓1.41)93(↓5)2.48(↓1.17)67(↓26)2.33(↓1.32)
ITA(best)46(↓51)2.11(↓1.54)64(↓34)2.15(↓1.50)84(↓14)2.17(↓1.48)51(↓42)2.15(↓1.50)
+ +![](images/2b16831475b967f6e1d4274c06317761597df3d27c9b3a84d3853fc0bcf3b594.jpg) +Figure 3. Visualization Results. From up to down, the images represent the results of Clean, Natural Light Attack, Shadow Attack, and our method, along with their corresponding Top-1 labels, evaluated on OpenCLIP ViT-B/16. The red color indicates misclassified labels.(Left) Visualization of some illumination changes causing LVLMs to give incorrect answers.(Right) + +tribution with mean $\mu$ , step-size $\Sigma$ , and covariance matrix $C$ . The optimization aims to improve the adversarial illumination conditions by adjusting the parameters $\mu, C$ , and $\Sigma$ . The optimization process proceeds through three main steps: sampling, evaluation, and update. For a detailed description of the overall process and mathematical formulations, which is referred to Algorithm 1. Additional information can be found in Appendix A. + +# 4. Experiments + +In this section, we demonstrate the effectiveness of our proposed method ITA for crafting adversarial illuminating transformation examples against VLMs. + +# 4.1. Experimental Settings + +Datasets. We evaluate ITA on three representative vision-language tasks (zero-shot image classification, image cap- + +tioning, Visual Question Answering) using only the COCO dataset [5]. This approach ensures more accurate evaluations, as the same dataset provides real labels for all three tasks, eliminating the potential influence of dataset distribution. Additionally, we carefully select 30 categories, totaling 300 images with clear instance from COCO. This consistency across tasks allows for a more reliable and integrated assessment of ITA's performance. We use COCO-80 categories as our semantic label list, and the selected categories are provided in Appendix. B. + +Threat Models. For the image classification task, we select the foundation models, a diverse set of CLIP variants, including OpenCLIP ViT-B/16 [7, 17], Meta-CLIP ViT-L/14 [46], EVA-CLIP ViT-G/14 [40] and OpenAI CLIP ViT-L/14 [36]. This selection is motivated by the need to evaluate the robustness of our proposed methods across different versions and parameter scales of CLIP models. For + +Table 2. Performance Degradation (↓) of VLMs on Image Captioning Under Illumination-aware Adversarial Examples Generated from COCO Dataset. We compare the consistency (%) evaluated by GPT-4 of previous methods with our proposed approach. The performance is considered better with lower values. Number in bold indicates the best performance and the second-best result is underlined. + +
Image EncoderModels#ParamsCleanNatural Light Attack [14]Shadow Attack [52]ITA(random)ITA(best)
OpenAI CLIP ViT-L/14LLaVA-1.57B78.6077.00(↓1.63)74.60(↓4.00)69.04(↓9.56)63.73(↓14.87)
LLaVA-1.67B72.1071.70(↓0.39)71.17(↓0.93)69.19(↓2.91)61.60(↓10.51)
OpenFlamingo3B70.2069.53(↓0.67)67.80(↓2.40)66.37(↓3.85)53.93(↓16.27)
Blip-2 (FlanT5XL ViT-L)3.4B75.1070.77(↓4.33)68.57(↓6.53)67.40(↓7.73)60.93(↓14.17)
EVA-CLIP ViT-G/14Blip-2 (FlanT5XL)4.1B74.9671.27(↓3.69)68.80(↓6.17)69.50(↓5.47)62.01(↓11.88)
InstructBLIP (FlanT5XL)4.1B76.5072.07(↓4.43)69.77(↓6.73)69.03(↓7.97)63.20(↓13.30)
+ +Table 3. Performance Degradation (↓) of VLMs on VQA Task Under Illumination-aware Adversarial Examples Generated from COCO Dataset. We compare the correctness (%) evaluated by GPT-4 of previous methods with our proposed approach. The performance is considered better with lower values. Number in bold indicates the best performance and the second-best result is underlined. + +
Image EncoderModels#ParamsCleanNatural Light Attack [14]Shadow Attack [52]ITA(random)ITA(best)
OpenAI CLIP ViT-L/14LLaVA-1.57B68.0068.00(-)67.00 (↓1.00)59.00 (↓9.00)48.00 (↓20.00)
LLaVA-1.67B64.0063.00 (↓1.00)64.00 (-)50.00 (↓14.00)43.00 (↓21.00)
OpenFlamingo3B45.0039.00 (↓6.00)44.00 (↓1.00)40.00 (↓5.00)19.00 (↓26.00)
Blip-2 (FlanT5XL ViT-L)3.4B63.0058.00 (↓5.00)50.00 (↓13.00)48.00 (↓15.00)38.00 (↓25.00)
EVA-CLIP ViT-G/14Blip-2 (FlanT5XL)4.1B54.0053.00 (↓1.00)54.00(-)51.00 (↓3.00)33.00 (↓21.00)
InstructBLIP (FlanT5XL)4.1B68.0064.00 (↓4.00)62.00 (↓6.00)57.00 (↓11.00)44.00 (↓24.00)
+ +image captioning and VQA tasks, we choose current mainstream VLM models, including LLaVA-1.5 [26], LLaVA-1.6 [24], OpenFlamingo [4], Blip-2 ViT-L [21], Blip-2 FlanT5-XL [21] and InstructBLIP [23], covering a range of architectures and capabilities. The proposed method is compared to the following baselines: Shadows Attack[52] and NATual Light Attack[14]. We follow the same setup as reported in the original work, with the only difference being the substitution of the target model for the attack. + +Evaluation Metrics. For the image classification task, we evaluate top-1 accuracy and report the relative drop compared to clean samples. We employ a 4-point scale to assess the naturalness of the images using GPT-4o, focusing primarily on three fields: Visual Naturalness, Physical Consistency, and Adversarial Likelihood. Additionally, for image captioning and VQA tasks, we adopt the LLM-as-a-judge evaluation framework, following recent studies [26, 50, 51]. Specifically, we design two distinct templates to assess the consistence scores of generated captions and the correctness of responses to given questions. The details of these evaluation settings are provided in Appendix C. + +Implementation Details. In our experiments, we set the number of samples per iteration to 20 and run a total of 200 iterations. The number of light sources is 3 and the initial light distribution parameters are randomly assigned and the spatial boundaries $(x,y)$ range from 0 to the corresponding image dimensions, intensity $I$ is set between 0.5 and 1.0, and radius $r$ varies between 10 and 50. The distance threshold in distance penalty loss is set to 50. The weighting factors $\alpha$ and $\beta$ are 0.1 and 0.01 respectively. Additionally, we establish two baseline settings: the original natural images and randomly generated illumination transformation samples without optimization. All experiments are conducted on 8 NVIDIA RTX 4090 24GB GPUs, and + +the detailed computational cost is provided in Appendix. D. + +# 4.2. Comparisons with Other Methods + +Evaluation on Zero-Shot Classification Tasks. Tab. 1 presents the comparison of our method with previous illumination attack methods on zero-shot classification tasks. Our method consistently outperforms others across all CLIP versions, achieving a notable $54\%$ attack success rate on OpenCLIP ViT-B/16 while maintaining a high naturalness score. This highlights the robustness of our approach, which misleads the model effectively while preserving image quality. Additionally, the accuracy degradation across different models varies significantly. Smaller models, like OpenCLIP ViT-B/16, suffer the most (a $51\%$ drop), while larger models like EVA-CLIP ViT-G/14 and Meta-CLIP ViT-L/14 models show a more moderate drop in accuracy (around $14\%$ and $34\%$ , respectively). This suggests that models with more parameters might be slightly more resilient, but they still face considerable performance degradation under illumination transformation. We also observe a trade-off between image naturalness and classification accuracy. As the accuracy drops, the naturalness score generally decreases. Overall, our method balances adversariality with naturalness, outperforming previous techniques. As shown in Fig. 3, visualizations demonstrate that our adversarial examples appear more natural compared to others, with no alteration in the image's semantic content. Additional examples can be found in Appendix. E. + +Evaluation on Image Captioning & VQA Tasks. In addition to classification tasks, we also evaluate the effectiveness of our method on Image Captioning and VQA tasks using state-of-the-art large-scale Vision-Language Models (LVLMs). To ensure a fair and meaningful comparison, we perform transfer attacks by utilizing the samples gener + +Table 4. Further Evaluation of Naturalness Scores and Consistency Scores. Number in bold indicates the best performance and the second-best result is underlined. + +
MethodVis.(↑)Phy.(↑)Adv.(↑)Sum(↑)DISTS(↓)
Clean3.423.583.953.653.91e-08
Natural Light Attack [13]2.282.071.862.070.104
Shadow Attack [46]2.101.321.921.780.058
ITA(random)2.442.561.992.330.021
ITA(best)2.362.211.882.150.051
+ +![](images/0882bc3570cd5bb48f639dd7e7c83628cd65a7fe7f93737cf58086fb02bb33c8.jpg) +Figure 4. Ablation Study Results on Optimization Hyperparameters. Attack success rate of illumination-aware adversarial examples with (A) different numbers of light sources at different iteration steps on OpenCLIP and (B) different population number across different CLIP versions in 200 iteration steps. + +ated by image encoders from different VLMs on zero-shot classification tasks. A clear trend is that smaller models with fewer parameters experience larger performance drops. For example, in Image Captioning (Tab. 2), OpenFlamingo and Blip-2 show significant degradation, particularly under the ITA, with performance drops of $16\%$ and $14\%$ , respectively. Larger models like LLaVA-1.5 and LLaVA-1.6 exhibit more moderate drops, around $10\%$ , suggesting that larger models are slightly more robust to illumination transformation. Similarly, in the VQA task (Tab. 3), smaller models like OpenFlamingo and Blip-2 show more substantial drops (up to $26\%$ and $25\%$ , respectively) under ITA. In contrast, larger models like InstructBLIP experience a smaller drop of $24\%$ , indicating some resilience but still significant vulnerability. In summary, our method is the most effective across both tasks, achieving the largest performance reductions, demonstrating the transferability of our method in attacking LVLMs across different tasks. + +Evaluation of Naturalness and Consistency. Reference-free image quality evaluation through advanced VLM has been widely adopted in many studies [19, 45]. Notably, GPT-4o achieves a correlation of more than 0.4 with human ratings regarding the naturalness of synthetic images, which is very close to inter-human correlation (0.45). To better interpret the final score, we provide detailed subscores (Tab. 4). To evaluate consistency, we further introduce DISTS [9], a unified structural and textural similarity metric, which shows that the samples generated by ITA exhibit better consistency with the original images. + +# 4.3. Ablation Study + +Optimization Hyperparameters. As shown in Fig. 4, we observe that the attack success rate continuously increases with the number of optimization iterations and converges + +![](images/8f7ded5114a214e95d247ee5f293999cc6b6243c9e3b489704940aae14b7c009.jpg) +Figure 5. Ablation Study Results on Weight Factors. (a) Visual examples of illumination changes as $\alpha$ varies, along with the corresponding predicted labels in OpenCLIP. (b) Visualization of the final optimized illumination as $\beta$ varies across the same sample with three light sources. + +after 200 steps. Based on this observation, we set the default value of the number of iterations to 200 throughout the entire experiment to balance effectiveness and computational efficiency. Additionally, we analyze the number of light sources optimized jointly. The results show that the highest attack success rate is achieved when $n = 2$ . Furthermore, we conduct experiments on the number of samples in each optimization iteration, and the results demonstrate that convergence is reached when the population size is 20. + +Weight Factors. We analyze the effectiveness of the $\alpha$ and $\beta$ parameters by visualizing the results for different values in Fig. 5. The results show that when $\alpha = 0.1$ and $\beta = 0.01$ , we achieve both high naturalness and attack effectiveness. + +# 5. Conclusion + +In this paper, we propose ITA, a novel framework designed to generate realistic illumination-aware adversarial examples. By innovatively modeling global illumination using parameterized point light sources and integrating physics-based lighting reconstruction, ITA provides fine-grained control over lighting conditions, enabling diverse and realistic illumination variations. Additionally, ITA incorporates perceptual and diversity constraints, ensuring the generated illumination-aware adversarial examples are both natural and effective. Extensive experiments demonstrate that ITA significantly degrades the performance of advanced VLMs, revealing previously unrecognized vulnerabilities to illumination shifts and emphasizing the urgent need to enhance the illumination-aware robustness of VLMs. + +# Acknowledgement + +This work was supported by the Fundamental Research Funds for the Central Universities. + +# References + +[1] Sani Abba, Ali Mohammed Bizi, Jeong-A Lee, Souley Bakouri, and Maria Liz Crespo. Real-time object detection, tracking, and monitoring framework for security surveillance systems. Heliyon, 10(15), 2024. 2 +[2] Haider Al-Tahan, Quentin Garrido, Randall Balestriero, Diane Bouchacourt, Caner Hazirbas, and Mark Ibrahim. Unibench: Visual reasoning requires rethinking vision-language beyond scaling. arXiv preprint arXiv:2408.04810, 2024. 2 +[3] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in neural information processing systems, 35:23716-23736, 2022. 2 +[4] Anas Awadalla, Irena Gao, Josh Gardner, Jack Hessel, Yusuf Hanafy, Wanrong Zhu, Kalyani Marathe, Yonatan Bitton, Samir Gadre, Shiori Sagawa, et al. Openflamingo: An opensource framework for training large autoregressive vision-language models. arXiv preprint arXiv:2308.01390, 2023. 7 +[5] Xinlei Chen, Hao Fang, Tsung-Yi Lin, Ramakrishna Vedantam, Saurabh Gupta, Piotr Dollar, and C Lawrence Zitnick. Microsoft coco captions: Data collection and evaluation server. arXiv preprint arXiv:1504.00325, 2015. 6 +[6] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24185-24198, 2024. 2 +[7] Mehdi Cherti, Romain Beaumont, Ross Wightman, Mitchell Wortsman, Gabriel Ilharco, Cade Gordon, Christoph Schuhmann, Ludwig Schmidt, and Jenia Jitsev. Reproducible scaling laws for contrastive language-image learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2818-2829, 2023. 6 +[8] Xuanming Cui, Alejandro Aparcedo, Young Kyun Jang, and Ser-Nam Lim. On the robustness of large multimodal models against image adversarial attacks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24625-24634, 2024. 3 +[9] Keyan Ding, Kede Ma, Shiqi Wang, and Eero P Simoncelli. Image quality assessment: Unifying structure and texture similarity. IEEE transactions on pattern analysis and machine intelligence, 44(5):2567-2581, 2020. 8 +[10] Ranjie Duan, Xiaofeng Mao, A Kai Qin, Yuefeng Chen, Shaokai Ye, Yuan He, and Yun Yang. Adversarial laser beam: Effective physical-world attack to dnns in a blink. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16062-16071, 2021. 2, 3 +[11] Abhiram Gnanasambandam, Alex M Sherman, and Stanley H Chan. Optical adversarial attack. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 92-101, 2021. 2, 3 + +[12] Daniel Golovin, Benjamin Solnik, Subhodeep Moitra, Greg Kochanski, John Karro, and David Sculley. Google vizier: A service for black-box optimization. In Proceedings of the 23rd ACM SIGKDD international conference on knowledge discovery and data mining, pages 1487-1495, 2017. 5 +[13] Nikolaus Hansen. The cma evolution strategy: A tutorial. arXiv preprint arXiv:1604.00772, 2016. 5 +[14] Teng-Fang Hsiao, Bo-Lun Huang, Zi-Xiang Ni, Yan-Ting Lin, Hong-Han Shuai, Yung-Hui Li, and Wen-Huang Cheng. Natural light can also be dangerous: Traffic sign misinterpretation under adversarial natural light attacks. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3915-3924, 2024. 2, 3, 6, 7 +[15] Yu-Chih-Tuan Hu, Bo-Han Kung, Daniel Stanley Tan, Jun-Cheng Chen, Kai-Lung Hua, and Wen-Huang Cheng. Naturalistic physical adversarial patch for object detectors. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7848-7857, 2021. 3 +[16] Bingyao Huang and Haibin Ling. Spaa: Stealthy projector-based adversarial attacks on deep image classifiers. In 2022 IEEE Conference on Virtual Reality and 3D User Interfaces (VR), pages 534-542. IEEE, 2022. 3 +[17] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. openclip, 2021. 6 +[18] James T Kajiya. The rendering equation. In Proceedings of the 13th annual conference on Computer graphics and interactive techniques, pages 143-150, 1986. 2 +[19] Max Ku, Dongfu Jiang, Cong Wei, Xiang Yue, and Wenhu Chen. Viescore: Towards explainable metrics for conditional image synthesis evaluation. arXiv preprint arXiv:2312.14867, 2023. 8 +[20] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International conference on machine learning, pages 12888-12900. PMLR, 2022. 2 +[21] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730-19742. PMLR, 2023. 2, 7 +[22] Yanjie Li, Yiquan Li, Xuelong Dai, Songtao Guo, and Bin Xiao. Physical-world optical adversarial attacks on 3d face recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 24699-24708, 2023. 2 +[23] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36:34892-34916, 2023. 7 +[24] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26296-26306, 2024. 7 + +[25] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava next: Improved reasoning,OCR, and world knowledge, 2024. 2 +[26] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 2, 7 +[27] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, et al. Mmbench: Is your multi-modal model an all-around player? In European Conference on Computer Vision, pages 216-233. Springer, 2024. 3 +[28] Jiahuan Long, Tingsong Jiang, Wen Yao, Shuai Jia, Weijia Zhang, Weien Zhou, Chao Ma, and Xiaoqian Chen. Papmot: Exploring adversarial patch attack against multiple object tracking. In European Conference on Computer Vision, pages 128-144. Springer, 2024. 3 +[29] Jiahuan Long, Zhengqin Xu, Tingsong Jiang, Wen Yao, Shuai Jia, Chao Ma, and Xiaogian Chen. Robust sam: On the adversarial robustness of vision foundation models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 5775-5783, 2025. 3 +[30] Jiahuan Long, Wen Yao, Tingsong Jiang, and Chao Ma. Cdupatch: Color-driven universal adversarial patch attack for dual-modal visible-infrared detectors. arXiv preprint arXiv:2504.10888, 2025. 3 +[31] Jiasen Lu, Dhruv Batra, Devi Parikh, and Stefan Lee. Vilbert: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks. Advances in neural information processing systems, 32, 2019. 2 +[32] Chengzhi Mao, Scott Geng, Junfeng Yang, Xin Wang, and Carl Vondrick. Understanding zero-shot adversarial robustness for large-scale models. arXiv preprint arXiv:2212.07016, 2022. 3 +[33] Dinh-Luan Nguyen, Sunpreet S Arora, Yuhang Wu, and Hao Yang. Adversarial light projection attacks on face recognition systems: A feasibility study. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 814-815, 2020. 2 +[34] Masahiro Nomura and Masashi Shibata. cmaes: A simple yet practical python library for cma-es. arXiv preprint arXiv:2402.01373, 2024. 2 +[35] Jingxiang Qu, Ryan Wen Liu, Yuan Gao, Yu Guo, Fenghua Zhu, and Fei-Yue Wang. Double domain guided real-time low-light image enhancement for ultra-high-definition transportation surveillance. IEEE Transactions on Intelligent Transportation Systems, 25(8):9550-9562, 2024. 2 +[36] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 3, 6 +[37] Shouwei Ruan, Yinpeng Dong, Hang Su, Ning Chen, and Xingxing Wei. Towards viewpoint-invariant visual recognition via adversarial training. In ICCV, pages 1-10, 2023. 2 +[38] Shouwei Ruan, Yinpeng Dong, Hang Su, Jianteng Peng, Ning Chen, and Xingxing Wei. Improving viewpoint robust- + +ness for visual recognition via adversarial training. arXiv preprint arXiv:2307.11528, 2023. 2 +[39] Shouwei Ruan, Yinpeng Dong, Hanqing Liu, Yao Huang, Hang Su, and Xingxing Wei. Omniview-tuning: Boosting viewpoint invariance of vision-language pre-training models. In European Conference on Computer Vision, pages 309-327. Springer, 2024. 3 +[40] Quan Sun, Yuxin Fang, Ledell Wu, Xinlong Wang, and Yue Cao. Eva-clip: Improved training techniques for clip at scale. arXiv preprint arXiv:2303.15389, 2023. 6 +[41] Yitong Sun, Yao Huang, and Xingxing Wei. Embodied laser attack: Leveraging scene priors to achieve agent-based robust non-contact attacks. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 5902-5910, 2024. 2 +[42] Xiaoyu Tian, Junru Gu, Bailin Li, Yicheng Liu, Chenxu Hu, Yang Wang, Kun Zhan, Peng Jia, Xianpeng Lang, and Hang Zhao. Drivevlm: The convergence of autonomous driving and large vision-language models. arXiv preprint arXiv:2402.12289, 2024. 2 +[43] Shengbang Tong, Zhuang Liu, Yuexiang Zhai, Yi Ma, Yann LeCun, and Saining Xie. Eyes wide shut? exploring the visual shortcomings of multimodal llms, 2024. 3 +[44] Donghua Wang, Wen Yao, Tingsong Jiang, Chao Li, and Xiaogian Chen. Rfla: A stealthy reflected light adversarial attack in the physical world. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4455-4465, 2023. 3 +[45] Jifang Wang, Xue Yang, Longyue Wang, Zhenran Xu, Yiyu Wang, Yaowei Wang, Weihua Luo, Kaifu Zhang, Baotian Hu, and Min Zhang. A unified agentic framework for evaluating conditional image generation. arXiv preprint arXiv:2504.07046, 2025.8 +[46] Hu Xu, Saining Xie, Xiaqing Ellen Tan, Po-Yao Huang, Russell Howes, Vasu Sharma, Shang-Wen Li, Gargi Ghosh, Luke Zettlemoyer, and Christoph Feichtenhofer. Demystifying clip data. arXiv preprint arXiv:2309.16671, 2023. 6 +[47] Zhenhua Xu, Yujia Zhang, Enze Xie, Zhen Zhao, Yong Guo, Kwan-Yee K Wong, Zhenguo Li, and Hengshuang Zhao. Drivegpt4: Interpretable end-to-end autonomous driving via large language model. IEEE Robotics and Automation Letters, 2024. 2 +[48] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Ic-light github page, 2024. 2 +[49] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 5 +[50] Yichi Zhang, Yao Huang, Yitong Sun, Chang Liu, Zhe Zhao, Zhengwei Fang, Yifan Wang, Huanran Chen, Xiao Yang, Xingxing Wei, et al. Multitrust: A comprehensive benchmark towards trustworthy multimodal large language models. Advances in Neural Information Processing Systems, 37:49279-49383, 2024. 3, 7 +[51] Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan + +Li, Dacheng Li, Eric Xing, et al. Judging llm-as-a-judge with mt-bench and chatbot arena. Advances in Neural Information Processing Systems, 36:46595-46623, 2023. 7 +[52] Yiqi Zhong, Xianming Liu, Deming Zhai, Junjun Jiang, and Xiangyang Ji. Shadows can be dangerous: Stealthy and effective physical-world adversarial attack by natural phenomenon. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15345-15354, 2022. 2, 3, 6, 7 \ No newline at end of file diff --git a/ICCV/2025/When Lighting Deceives_ Exposing Vision-Language Models' Illumination Vulnerability Through Illumination Transformation Attack/images.zip b/ICCV/2025/When Lighting Deceives_ Exposing Vision-Language Models' Illumination Vulnerability Through Illumination Transformation Attack/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..75c12996a21be9669f1d4fe1a2ecc90459118aaf --- /dev/null +++ b/ICCV/2025/When Lighting Deceives_ Exposing Vision-Language Models' Illumination Vulnerability Through Illumination Transformation Attack/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f2d2a4e257cdb8fdd4851c3bbaba347dc8d88974fe8df57ad73c7399de0a7d05 +size 672797 diff --git a/ICCV/2025/When Lighting Deceives_ Exposing Vision-Language Models' Illumination Vulnerability Through Illumination Transformation Attack/layout.json b/ICCV/2025/When Lighting Deceives_ Exposing Vision-Language Models' Illumination Vulnerability Through Illumination Transformation Attack/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..2982629c2b2198d9450acf3137075603d022ece5 --- /dev/null +++ b/ICCV/2025/When Lighting Deceives_ Exposing Vision-Language Models' Illumination Vulnerability Through Illumination Transformation Attack/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9d1925861fff3132ab98ca7f224116de7ec1e0faa5d6d6c2385597d0324405fc +size 442439 diff --git a/ICCV/2025/When Pixel Difference Patterns Meet ViT_ PiDiViT for Few-Shot Object Detection/78d510bf-7798-4115-b8d7-929a950a8416_content_list.json b/ICCV/2025/When Pixel Difference Patterns Meet ViT_ PiDiViT for Few-Shot Object Detection/78d510bf-7798-4115-b8d7-929a950a8416_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..627acd9a0ab23b0c8813ffb4756199c68dc2d4eb --- /dev/null +++ b/ICCV/2025/When Pixel Difference Patterns Meet ViT_ PiDiViT for Few-Shot Object Detection/78d510bf-7798-4115-b8d7-929a950a8416_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cbcf68346a5541caec93bcaa5c70801dffdd7a3dbbfa2c622d94bbe4cdc6f856 +size 78757 diff --git a/ICCV/2025/When Pixel Difference Patterns Meet ViT_ PiDiViT for Few-Shot Object Detection/78d510bf-7798-4115-b8d7-929a950a8416_model.json b/ICCV/2025/When Pixel Difference Patterns Meet ViT_ PiDiViT for Few-Shot Object Detection/78d510bf-7798-4115-b8d7-929a950a8416_model.json new file mode 100644 index 0000000000000000000000000000000000000000..7631aef6b7e93a18a0f5272dfc929d2e8081e98f --- /dev/null +++ b/ICCV/2025/When Pixel Difference Patterns Meet ViT_ PiDiViT for Few-Shot Object Detection/78d510bf-7798-4115-b8d7-929a950a8416_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:576d959c70d332099c23a67bb105cd5aa0db1e48580aa7c24d138499fb9921f8 +size 97170 diff --git a/ICCV/2025/When Pixel Difference Patterns Meet ViT_ PiDiViT for Few-Shot Object Detection/78d510bf-7798-4115-b8d7-929a950a8416_origin.pdf b/ICCV/2025/When Pixel Difference Patterns Meet ViT_ PiDiViT for Few-Shot Object Detection/78d510bf-7798-4115-b8d7-929a950a8416_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..fb1d0b003cb5f84c8a070d114c42247ee3319ea1 --- /dev/null +++ b/ICCV/2025/When Pixel Difference Patterns Meet ViT_ PiDiViT for Few-Shot Object Detection/78d510bf-7798-4115-b8d7-929a950a8416_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:50cdf1cb799fe517d73d91fb3ae613fcc751bfe7beaacb8364b71d83779ad0af +size 3815383 diff --git a/ICCV/2025/When Pixel Difference Patterns Meet ViT_ PiDiViT for Few-Shot Object Detection/full.md b/ICCV/2025/When Pixel Difference Patterns Meet ViT_ PiDiViT for Few-Shot Object Detection/full.md new file mode 100644 index 0000000000000000000000000000000000000000..ac125cac12550d8ae6ce6fc370176631f0bb2d34 --- /dev/null +++ b/ICCV/2025/When Pixel Difference Patterns Meet ViT_ PiDiViT for Few-Shot Object Detection/full.md @@ -0,0 +1,318 @@ +# When Pixel Difference Patterns Meet ViT: PiDiViT for Few-Shot Object Detection + +Hongliang Zhou Yongxiang Liu* Canyu Mo Weijie Li* Bowen Peng Li Liu* The College of Electronic Science and Technology, National University of Defense Technology, China + +hongliangz2022@163.com + +lyx_bible@sina.com + +mocanyu@nudt.edu.cn + +liweijie@nudt.edu.cn + +pbowl6@nudt.edu.cn + +liuli_nudt@nudt.edu.cn + +# Abstract + +Few-shot object detection aims to detect novel classes with limited samples. Recent methods have leveraged rich semantic representations of pretrained vision transformer (ViT) to overcome limitations of model fine-tuning, thereby improving performance on novel classes. However, existing pretrained ViT schemes only perform transformer encoding in feature dimension, ignoring exploration of pixel-wise differences in low-level features and multiscale variations. The current challenges lie in: (i) extracted features suffer from blurred boundary features and smooth transition from center to boundary, leading to insufficient distinction between objects and backgrounds, and (ii) how to balance extraction of local details and global contour features under multiscale scenarios. So Pixel Difference Vision Transformer (PiDiViT) is proposed. Innovations include: (i) difference convolution fusion module (DCFM), which enhances feature differences from object centers to boundaries and effectively preserves global information by fusing pixel-wise central difference features with original features through an attention mechanism, and (ii) multiscale feature fusion module (MFFM), which adaptively fuses features extracted by five different scale convolutional kernels using a scale attention mechanism to generate attention weights, achieving an optimal balance between local detail and global semantic information extraction. PiDiViT achieves SOTA on the COCO benchmark: surpassing few-shot detection SOTA by 2.7 nAP50 (10-shot) and 4.0 nAP50 (30-shot) for novel classes, exceeding one-shot detection SOTA by 4.4 nAP50 and open-vocabulary detection SOTA by 3.7 nAP50. The code is available at https://github.com/Seaz9/PiDiViT. + +# 1. Introduction + +Few-shot object detection aims to recognize novel classes with limited samples [1, 6]. Existing methods primarily fo + +![](images/1b61e2a510f9a6b1935cdacb1c4bd88887e0f46b68a469e0bb86d0c0f4c3e991.jpg) +Figure 1. Overall performance. Comparison between PiDiViT and SOTA on COCO benchmarks (few-shot, one-shot and open-vocabulary object detection) shown in the radar chart. + +cus on fine-tuning models with scarce data for novel class detection [50]. However, the limitation of fine-tuning lies in insufficient adaptation caused by data limitations, leading to a significant performance gap between the novel and base classes [18]. Recent studies reveal that ViT [26, 33] can effectively overcome constraints of fine-tuning through their pretrained rich semantic representations. For example, DE-ViT [49] employs a projected feature subspace of ViT features to replace prototype classifier training, avoiding base class overfitting and model fine-tuning. In our exploration of ViT architectures, we made a significant discovery: existing pretrained ViT schemes perform transformer encoding in feature dimension, ignoring exploration of pixelwise differences in low-level features and multiscale variations. Current challenges lie in: (i) extracted features suffer from blurred boundary features and smooth transitions from object centers to boundaries, leading to insufficient object-background distinction, and (ii) balancing extraction of local details and global contour features under multiscale scenar + +ios. Designing explicit prior modules to address challenges can further optimize ViT features and break through learning limitations under a few-shot conditions. + +To address issues, we propose Pixel Difference Vision Transformer (PiDiViT). PiDiViT consists of two core designs: (i) difference convolution fusion module (DCFM), which enhances feature differences from object centers to boundaries and preserves global information by fusing pixel-wise central difference features with original features through an attention mechanism, and (ii) multiscale feature fusion module (MFFM), which adaptively fuses features extracted by five scale convolutional kernels using a scale attention mechanism to generate attention weights, achieving an optimal balance between local detail and global semantic information extraction in various scenarios. + +We evaluate PiDiViT on COCO [22] for few-shot, one-shot, and open-vocabulary object detection tasks, achieving SOTA performance. Performance in three different task settings demonstrates the broad adaptability of PiDiViT in few-shot and open-vocabulary detection. For few-shot object detection, it improves [7] by 2.7 and $4.0\mathrm{nAP50}$ (novel classes) for 10-shot and 30-shot settings, respectively. In one-shot object detection, it improves [49] by $4.4\mathrm{nAP50}$ on average, and in open-vocabulary object detection, it outperforms [2] by $3.7\mathrm{nAP50}$ . These results highlight PiDiViT as a precise and efficient paradigm for few-shot detection tasks. + +The contributions are summarized as follows: + +- We propose PiDiViT, which empowers pretrained ViT to excel in few-shot detection by designing explicit prior modules for pixel-wise differences and multiscale variations in low-level features of pretrained ViT. +- We design DCFM to enhance feature differences of smooth features from center to boundary and effectively preserve global information, and we develop MFFM to strengthen the balance between multiscale local and global features under the scheduling of scale attention mechanism, adapting to various scenarios. +- PiDiViT achieves SOTA performance in COCO for few-shot, one-shot, and open-vocabulary object detection, setting new benchmarks and offering a valuable reference for future detecting few-shot objects. + +# 2. Related Work + +Few-Shot Object Detection (FSOD) aims to detect novel classes with limited samples [18]. Existing methods can be categorized into two groups: fine-tuning methods [6, 9, 20, 39, 40, 43] and meta-learning methods [5, 15, 44]. Fine-tuning methods adapt pretrained features to novel classes through model fine-tuning, but suffer from complex training processes and performance gaps between novel and base classes due to insufficient fine-tuning with limited novel data [50]. Subsequently, metalearning methods with online adaptation capabilities have gradually become main- + +stream, overcoming the complexity of model fine-tuning. These methods have evolved from early prototype-based feature matching techniques [15, 44] to more advanced dense support-query feature interaction mechanisms [10-12]. Recently, DE-ViT [49] proposed a solution that replaces training prototype classifiers with projection of ViT features into subspaces, effectively avoiding base class overfitting and model fine-tuning. However, current ViT-based methods struggle to address performance gaps caused by oversmooth features from center to boundary and multiscale size feature discrepancies. Our PiDiViT effectively bridges these gaps through DCFM and MFFM modules, achieving differential enhancement of smooth center-to-boundary features while preserving global features and adaptively capturing multiscale local details and global contours. + +One-shot object detection (OSOD) extends FSOD to the more challenging scenario of detecting novel categories with a single support sample, without requiring fine-tuning [45]. Early approaches focused on improving the interaction between support and target features [14, 21, 27], but were limited by the scarcity of support samples and the computational overhead associated with inference per category [4, 50]. Recent advances, such as DE-ViT [49], have improved efficiency by leveraging the characteristics of a dot product to match the features. For feature matching, the proposed PiDiViT employs cosine similarity instead of the dot product to project features into feature subspaces. The proposed DCFM module achieves differential enhancement of smooth center-to-boundary features while preserving global information, while the MFFM adaptively balances local and global features across multiple scales. This combination of explicit prior modules yields remarkable results even under the extreme constraints of OSOD tasks. + +Open-Vocabulary Object Detection (OvD) aims to detect objects of unknown categories using category names or textual descriptions [46, 48]. Recent advances in vision language models, particularly CLIP [30], have significantly propelled the field forward by enabling the training of transferable visual models through text. Cooperative foundational models [2] have shown exceptional results by leveraging complementary strengths of CLIP and SAM to transform closed-set detectors into open-set detectors. Recently, pretrained ViT architectures have shown remarkable potential in open-vocabulary scenarios. For example, DEViT [49] achieved impressive performance in both FSOD and OVD tasks by replacing the prototype classifier training with the projection of ViT features into subspaces. Despite the advancements made by pretrained ViTs, current ViT-based approaches only perform transformations and encoding in the feature dimension, overlooking the exploration of pixel-level feature disparities and multiscale variability. The proposed PiDiViT addresses this limitation through explicit prior modules such as DCFM and MFFM, which specifi + +![](images/916c5f75763d1bf6b664ab64443fc79a68ed9c1aeff445b25e383531af510f4d.jpg) +Figure 2. Framework of PiDiViT. It contains the overall PiDiViT architecture (a), DCFM module (b), and MFFM module (c). + +![](images/6b51477a302968e55e1820e3582e72140e22c4426d155e034dc5203c37b44f66.jpg) + +cally enhance the differential features from the center to the boundary and capture multiscale feature variability. + +# 3. Method + +# 3.1. Method Overview + +Model Background. The proposed Pixel Difference Vision Transformer (PiDiViT) builds upon DE-ViT [49], which integrates pretrained Dinov2 ViT [29] into the Meta RCNN [44] framework. DE-ViT [49] refines and propagates initial proposal regions and input features through the Region Proposal Network (RPN) [32] to achieve object localization. Compared to DE-ViT [49], the innovation of our PiDiViT model lies in the DCFM and MFFM modules. + +As shown in Fig. 2 (a), PiDiViT (Fig. 2 (a)) follows a bottom-up processing pipeline: the model takes initial region proposals, category prototypes and ViT features as inputs. First, initial region proposals are generated by a class-agnostic Region Proposal Network (RPN) [32], and category prototypes and ViT features are projected into input features + +via a Cosine Similarity Projection Subspace (CSPS*, brown cuboid in Fig. 2(a)). Second, the initial proposal regions are scaled into masked regions (yellow bounding box areas in Fig. 2 (a)), and then the masked regions and input features are processed by a three-layer difference multiscale fusion layer (dashed box in Fig. 2(a)). These layers include the proposed difference convolution fusion module (DCFM) (Fig. 2 (b), green) and the multiscale feature fusion module (MFFM) (Fig. 2 (c), blue). DCFM enhances smooth feature differences from center to boundary, while MFFM extracts global and local information at different scales. The input region and the middle output region are combined through concatenated residual summation across these layers to obtain the final input region, which is then fed into the last layer for processing to generate the final output region. The input features are progressively refined through three layers to yield the output features. The final output includes + +![](images/5de52895885263bca097ac3843655d4c3bcc574f15537c58c98c93b3490787b0.jpg) +Figure 3. The process of three difference convolutions (CPDC, APDC, and RPDC) [35]. The figure elaborates on the convolution process and pixel pair selection process for each of these difference convolutions. + +predicted bounding boxes converted from regions by the region-to-box layer [49], output regions, class scores via spatial attention mapping using width-height average features and output features. The details of the innovative modules (DCFM, MFFM) are elaborated in Sec. 3.2-Sec. 3.3. + +# 3.2. Difference Convolution Fusion Module DCFM + +Existing few-shot object detection methods such as DE-ViT [49] adopt a pretrained ViT scheme that performs transformational encoding only in the feature dimension, overlooking the exploration of low-level feature discrepancy. The features extracted by DE-ViT [49] exhibit conditions such as blurred boundary features and excessively smooth transitions from the center to the boundary, resulting in a less effective discrimination between target and background features. In particular, feature discrepancy is a critical means to distinguish target features under the scarcity of labeled samples in few-shot scenarios. It has recently been shown that pixel difference convolution (PDC) [35-37] can be highly discriminative while being extremely efficient to compute, yet flexibly integrated into convolutional neural networks. However, they have rarely been integrated with ViTs. Building on these advances, we propose a novel difference convolution fusion module (DCFM) designed to explicitly integrate higher-order local differential features with standard convolutional features. + +As shown in Fig. 2 (b) (green), DCFM adaptively fuses Central PDC (CPDC) [35] and standard convolution via an attention mechanism: the CPDC enhances the smooth feature discrepancy from the center to the edge by highlighting spatial positional value changes, sharpening target edges in features to better differentiate targets from backgrounds. This is then fused with the global information extracted by standard convolution, enabling the model to accurately recognize boundary pixel difference variations while retaining ability to extract global object representation information. + +In [35, 36], the authors designed three types of PDC + +(Central PDC (CPDC), Angular PDC (APDC) and Radial PDC (RPDC)) for probing higher-order local differential features inspired by the famous Loal Binary Pattern methods [23, 28, 34]: (i) CPDC follows CD-LBP to capture center-neighbor differences, (ii) APDC models circumferential variations based on AD-LBP, and (iii) RPDC extracts radial gradient information inspired by RD-LBP. As shown in Fig. 3, these PDC variants [35] establish a multiperspective structural representation through different sampling and differential computation strategies. The key difference between PDC and standard convolution lies in that PDC incorporates pixel differential calculations into the convolution operation. The computational processes of standard convolution and PDC are shown in Eq. (1) and Eq. (2). + +$$ +y _ {1} = f (\mathbf {X}, \boldsymbol {\theta}) = \sum_ {j = 1} ^ {3 \times 3} w _ {j} x _ {j}, \quad \text {(s t a n d a r d c o n v l u t i o n)} \tag {1} +$$ + +$$ +y _ {2} = f (\Delta \mathbf {X}, \boldsymbol {\theta}) = \sum_ {(x _ {j}, x _ {j} ^ {\prime}) \in \mathcal {P}} w _ {j} \left(x _ {j} - x _ {j} ^ {\prime}\right), \quad \text {(P D C)} \tag {2} +$$ + +Here, $w_{j}$ is the weight in the $3 \times 3$ convolution kernel, where $x_{j}$ and $x_{j} - x_{j}^{\prime}$ denote the input pixel values for standard convolution and PDC [35]. $(x_{j}, x_{j}^{\prime}) \in \mathcal{P}$ is a set composed of difference pixel pairs with $j \leq 3 \times 3$ . + +Building on previous PDC work [35-37], we propose the difference convolution fusion module (DCFM) tailored for few-shot object detection. Through ablation studies (Sec. 4.5), we investigate the synergistic effects of combining three PDC variants [35] with standard convolution. Our results validate that the fusion of CPDC and standard convolution yields optimal performance gains for few-shot scenarios, APDC and RPDC are deemed unsuitable for this task due to their differential mechanisms. Based on these findings, we design DCFM as a complementary fusion of CPDC and standard convolutions. The CPDC branch models local pixel differences to extract fine-grained structural features (boundary gradient information), while the stan + +dard convolutional branch captures global high-level semantic features encompassing color, contour, and texture information. An attention mechanism dynamically weights these complementary features based on their discriminative power, enabling an adaptive balance between local structural and high-level semantic information. + +In the DCFM computation process, we concatenate a feature map $\mathbf{X}$ and feature of the region (if available) along the channel dimension. The concatenated feature map is then processed by an initial convolution layer, batch normalization, and a ReLU activation function, as shown in Eq. (3). + +$$ +\mathbf {X} _ {\operatorname {c o n v} _ {1}} = \operatorname {R e L U} \left(\operatorname {B a t c h N o r m} \left(\operatorname {C o n v} _ {1} (\mathbf {X})\right)\right) \tag {3} +$$ + +Here, $\mathrm{Conv}_1(\mathbf{X})$ is the convolution operation applied to input $\mathbf{X}$ , with a kernel size of 3, padding is $3/2$ , and a stride of 1. + +To capture global features and boundary pixel differential features, we use the standard convolution block (sc block) and the center difference convolution block (cd block) for processing. Each block independently processes the feature map $\mathbf{X}_{\mathrm{conv}_1}$ , as shown in Eq. (4). + +$$ +\mathbf {d} _ {\mathrm {s c}} = \operatorname {C o n v} _ {\mathrm {s c}} \left(\mathbf {X} _ {\text {c o n v} _ {1}}\right) \tag {4} +$$ + +$$ +\mathbf {d} _ {\mathrm {c d}} = \mathrm {P D C} _ {\mathrm {c d}} \left(\mathbf {X} _ {\text {c o n v} _ {1}}\right) +$$ + +Here, $\mathrm{Conv}_{\mathrm{sc}}$ is the sc block and $\mathrm{PDC}_{\mathrm{cd}}$ is the cd block, each block operation aiming to capture global variations or local boundary pixel differential variations of different features, producing outputs $\mathbf{d}_{\mathrm{sc}}$ and $\mathbf{d}_{\mathrm{cd}}$ . + +After obtaining the outputs of two types of convolution blocks, we concatenate them along the channel dimension and compute the attention weights through $f_{\mathrm{att}}$ . The weights are then normalized using Softmax, as shown in Eq. (5). + +$$ +\begin{array}{l} \mathbf {d} _ {\text {s t a c k}} = \operatorname {C o n c a t} \left(\mathbf {d} _ {\mathrm {s c}}, \mathbf {d} _ {\mathrm {c d}}\right) \\ f _ {\text {a t t}} = \operatorname {C o n v} _ {2} \left(\operatorname {R e L U} \left(\operatorname {B a t c h N o r m} \left(\operatorname {C o n v} _ {1}\right)\right) \right. \tag {5} \\ \end{array} +$$ + +$$ +\left[ \alpha_ {i _ {i = 1} ^ {2}} ^ {2} \right] = \operatorname {S o f t m a x} \left(f _ {\text {a t t}} \left(\mathbf {d} _ {\text {s t a c k}}\right)\right) +$$ + +Here, $[\alpha_{i_i = 1}^2 ]$ are the attention weights generated for $\mathbf{d}_{\mathrm{sc}}$ and $\mathbf{d}_{\mathrm{cd}}$ , respectively, and $f_{\mathrm{att}}$ is the attention layer that generates the weights. Softmax is applied along the channel dimension to normalize the attention weights of the outputs from each type of convolution block. + +Next, we apply the computed attention weights to the outputs of the sc block and the cd block to obtain the weighted features. The weighted results are then summed to obtain the fused features, as shown in Eq. (6) and Eq. (7). + +$$ +\mathbf {d} _ {\mathrm {s c}} ^ {w} = \mathbf {d} _ {\mathrm {s c}} \cdot \alpha_ {1} \tag {6} +$$ + +$$ +\mathbf {d} _ {\mathrm {c d}} ^ {w} = \mathbf {d} _ {\mathrm {c d}} \cdot \alpha_ {2} +$$ + +$$ +\mathbf {f} _ {\text {f u s e d}} = \mathbf {d} _ {\mathrm {s c}} ^ {w} + \mathbf {d} _ {\mathrm {c d}} ^ {w} \tag {7} +$$ + +Here, $\mathbf{d}_{\mathrm{sc}}$ represents the output features of sc block, and $\mathbf{d}_{\mathrm{cd}}$ represents output features of the cd block; $\alpha_{1}$ and $\alpha_{2}$ are the + +attention weights for the sc block and cd block, respectively; $\mathbf{d}_{\mathrm{sc}}^w$ and $\mathbf{d}_{\mathrm{cd}}^w$ represent the weighted features of the sc block and cd block. Finally, $\mathbf{f}_{\mathrm{fused}}$ represents the fused features. + +# 3.3. Multiscale Feature Fusion Module MFFM + +Existing few-shot methods, such as DE-ViT [49] refine proposal regions by stacking single-scale propagation layers, achieving promising results in few-shot detection. However, their fixed single-scale structure lacks an adaptive mechanism to balance and fuse local-global features extracted at different scales, limiting robustness in detecting objects of varying sizes and impairing performance balance and enhancement between base and novel classes detection. To address this, we propose the multiscale feature fusion module (MFFM), as shown in Fig. 2 (c) (blue). This module fuses features extracted by five-scale convolutional kernels using a scale attention mechanism to generate attention weights for fusion. By selectively extracting and fusing features at targeted scales, it ensures adaptive capture of local details and global contours for objects in diverse scenarios. + +MFFM comprises two key components: multiscale feature extraction and adaptive fusion. For extraction, it employs a parallel multi-branch architecture with $1 \times 1$ , $3 \times 3$ , $5 \times 5$ , $7 \times 7$ , and $3 \times 3$ dilated convolutions. Small-scale convolutions ( $1 \times 1$ , $3 \times 3$ ) capture fine local details, while larger-scale and dilated convolutions extract broader contextual information, enabling robust multiscale feature modeling even with limited data. For fusion, an attention-based mechanism dynamically learns the importance weights for each branch of the scale. The attention module assesses the discriminability of the features and generates fusion weights, adaptively fusing features based on their relevance. This enhances scale-invariant detection and robustness in data-scarce or complex scenarios. In the following, we detail the implementation of MFFM in the model. + +We use multiscale convolutions for feature extraction. The input feature $\mathbf{h} \in \mathbb{R}^{S \times K \times K}$ is processed through branches with five different convolution kernel sizes to extract multiscale features, as shown in Eq. (8). + +$$ +\begin{array}{l} \mathbf {h} _ {1} = \operatorname {R e L U} \left(\operatorname {B N} \left(\mathbf {W} _ {1 \times 1} * \mathbf {h} + \mathbf {b} _ {1 \times 1}\right)\right) \\ \mathbf {h} _ {2} = \operatorname {R e L U} \left(\operatorname {B N} \left(\mathbf {W} _ {3 \times 3} * \mathbf {h} + \mathbf {b} _ {3 \times 3}\right)\right) \\ \mathbf {h} _ {3} = \operatorname {R e L U} \left(\mathrm {B N} \left(\mathbf {W} _ {5 \times 5} * \mathbf {h} + \mathbf {b} _ {5 \times 5}\right)\right) \tag {8} \\ \mathbf {h} _ {4} = \operatorname {R e L U} \left(\operatorname {B N} \left(\mathbf {W} _ {7 \times 7} * \mathbf {h} + \mathbf {b} _ {7 \times 7}\right)\right) \\ \mathbf {h} _ {5} = \operatorname {R e L U} \left(\operatorname {B N} \left(\mathbf {W} _ {3 \times 3} ^ {d} * \mathbf {h} + \mathbf {b} _ {3 \times 3} ^ {d}\right)\right) \\ \end{array} +$$ + +Here, $*$ denotes convolution operation, BN is batch normalization, and $\mathbf{b}$ denotes the bias term associated with the corresponding convolution kernel $\mathbf{W}$ . Through branch operations with different receptive fields, the model can simultaneously capture local details and global contours. + +Table 1. Few-shot object detection benchmark results on COCO 2014. PiDiViT outperforms existing methods in new category detection tasks. Bold text represents SOTA result and gray text represents base class result. Backbone name is shown as abbreviation. + +
MethodYearBackboneRequires Finetune10-shot30-shot
bAPnAPnAP50nAP75bAPnAPnAP50nAP75
FSRW [15]19YOLO-ft-fullX-5.612.34.6-9.119.07.6
Meta R-CNN [44]19RN50X-8.719.16.6-12.425.310.8
TFA [40]20RN101✓33.910.019.29.235.113.424.713.2
Multi-Relation Det [5]20RN50X-16.631.316.1----
FSCE [39]21RN101✓-11.9-10.5-16.4-16.2
Retentive R-CNN [6]21RN101✓39.210.5--39.313.8--
HeteroGraph [10]21RN101✓-11.623.99.8-16.531.915.5
FsDetView [43]22RN50✓-12.527.39.8-14.730.612.2
Meta Faster R-CNN [11]22RN101✓-12.725.710.8-16.631.815.8
LVC [16]22ViT-S/8✓28.719.034.119.034.826.845.827.5
Cross-Transformer [12]22PVTv2-B2-Li✓-17.130.217.0-21.435.522.1
NIFF [9]23RN101✓39.018.8--39.020.9--
DiGeo [25]23RN101✓39.210.318.79.939.414.226.214.8
DE-ViT [49]23ViT-L/14X29.434.052.937.029.534.053.037.2
CD-VITO [7]25ViT-L/14X-35.354.937.2-35.954.538.0
PiDiViT (Ours)25ViT-L/14X32.337.057.640.931.437.358.541.7
+ +We concatenate features of different scales to obtain multiscale features $\mathbf{F} \in \mathbb{R}^{C \times K \times K}$ , and then generate the fusion weights $[\alpha_{i_{i=1}}^5]$ corresponding to the different scale convolutions using the scale attention mechanism, as shown in Eq. (9) and Eq. (10). + +$$ +\mathbf {F} = \operatorname {C o n c a t} \left(\mathbf {h} _ {1}, \mathbf {h} _ {2}, \mathbf {h} _ {3}, \mathbf {h} _ {4}, \mathbf {h} _ {5}\right) \tag {9} +$$ + +$$ +\left. \left[ \alpha_ {i _ {i = 1}} ^ {5} \right] = \sigma \left(\operatorname {C o n v} _ {2} \left(\operatorname {R e L U} \left(\operatorname {C o n v} _ {1} (\operatorname {A v g P o o l} (\mathbf {F}))\right)\right)\right) \right. \tag {10} +$$ + +Here, AvgPool is the global average pooling operation, $\mathrm{Conv}_1$ and $\mathrm{Conv}_2$ are the convolution operations, $\sigma$ is the activation function of the sigmoid, and $[\alpha_{i_i = 1}^5 ]$ is the weight vector generated by the scale attention mechanism. + +Next, multiscale features are fused by weighted summation, where features of different scales are weighted according to respective attention weights, as shown in Eq. (11). + +$$ +\mathbf {h} _ {\text {f u s i o n}} = \sum_ {i = 1} ^ {5} \alpha_ {i} \mathbf {h} _ {i} \tag {11} +$$ + +Here, $\mathbf{h}_{\mathrm{fusion}}$ is the multiscale fused feature, $\alpha_{i}$ is the fusion weight generated by the scale attention mechanism, and $\mathbf{h}_i$ is the scale feature under different convolution kernel. + +# 4. Experiment + +We extensively evaluate PiDiViT for FSOD, OSOD, and OVD tasks. Our evaluation includes comparisons with SOTA methods, analysis under different sample sizes, and qualitative comparisons with baseline models. Through ablation studies, we examine each module's contribution. The results demonstrate that PiDiViT achieves a new SOTA performance on the COCO dataset for all tasks, including 10-shot/30-shot FSOD, OSOD, and OVD. + +Evaluation Setup. We evaluate on COCO dataset [22] following standard category splits [40, 45, 52]. The data set is divided into base classes for training and novel classes for testing. We measure novel class performance using nAP, nAP50, and nAP75, and base class performance using bAP and bAP50. For OSOD, following [27], we split 80 COCO categories into 4 parts (Split-1/2/3/4), using 3 parts as base classes and the remaining part as novel. + +Implementation Details. PiDiViT builds on a two-stage detection framework such as Mask R-CNN [13], using a pretrained RPN from [52] for proposal generation. The RPN is trained only on base classes. We use frozen DINOv2 [29] ViT-L as the backbone. Following Meta R-CNN [44], we extract base class prototypes during training and add novel class prototypes afterward without changing other parameters. During the evaluation, we test base and novel classes, extracting background prototypes from COCO Stuff [3]. In inference, we select top- $T$ categories ( $T = 10$ for comparison experiments, $T = 1$ for ablations) based on feature distances of prototype and proposal. Training losses include focal loss and regression loss for category scores $c_t$ and boxes $b_t$ , BCE loss, and Dice loss [38] for regions $r_t$ . For network stacking, we stack three difference multiscale fusion layers. + +# 4.1. Comparisons with SOTA FSOD Methods + +As shown in Tab. 1, PiDiViT outperforms existing methods in COCO few-shot detection, achieving gains of $+2.7 / +4.0$ nAP50 (novel class) over the SOTA method CD-ViT [7] and $+4.7 / +5.5$ over baseline DE-ViT [49] in 10/30-shot settings. Compared with CD-ViT and DE-ViT based on pretrained ViT, we design prior modules for pixel-level feature disparities and multiscale variability of pretrained ViT's underlying features. DCFM achieves differential enhancement of smooth center-to-boundary features while preserv + +ing global information, and MFFM enables capture of local details and global contours on multiple scales. + +# 4.2. Comparisons with SOTA OSOD Methods + +As shown in Tab. 2, PiDiViT achieves an average improvement of $+4.4$ in nAP50 over the SOTA method DE-ViT [49] in the four splits of the COCO one-shot object detection task. In the extreme one-shot scenario, compared to DEViT, we design explicit prior modules (DCFM and MFFM) tailored to the pixel-level feature disparities and multiscale variability of pretrained ViT's underlying features. This enables our model to achieve significant performance gains for novel classes even under extreme one-shot constraints. + +Table 2. One-shot object detection benchmark results on COCO 2017. PiDiViT outperforms existing methods. Bold text is SOTA result. Backbone name is shown as abbreviation. + +
MethodYearBackbonenAP50
Split-1Split-2Split-3Split-4Avg
SiamMask [27]18RN5015.317.617.417.016.8
CoAE [14]19RN5023.423.620.520.422.0
AIT [4]21RN5026.026.422.322.624.3
SaFT [50]22RN10127.827.621.023.024.9
BHRL [45]22RN5026.129.022.724.525.6
DE-ViT [49]23ViT-L/1427.433.227.126.128.4
PiDiViT (Ours)25ViT-L/1432.038.630.430.232.8
+ +# 4.3. Comparisons with SOTA OVD Methods + +As shown in Tab. 3, PiDiViT significantly enhances the performance of COCO OVD, achieving improvements of $+3.7$ nAP50 over the SOTA coop-foundation-models by Bharadwaj et al. [2] and $+4.0$ over baseline DE-ViT [49]. Compared with the complementary fusion design of CLIP and SAM in coop-foundation-models [2], our explicitly designed prior modules tailored to the pixel-level feature disparities and multiscale variability of pre-trained ViT's underlying features achieve superior performance. + +# 4.4. Visualization Qualitative Comparison + +As shown in Fig. 4, we present visual comparisons between PiDiViT and DE-ViT [49] on COCO FSOD, OVD and OSOD tasks. PiDiViT demonstrates superior performance in three dimensions: First, it effectively alleviates duplicate detections observed in DE-ViT, such as cats and elephants in FSOD, sofas and vases in OVD, and potted plants in OSOD. Second, it improves classification accuracy, as exemplified by the correct identification of puppies in the OSOD task. Third, it enhances detection completeness by identifying objects missed by DE-ViT, including people in photo frames in OVD and furniture items in OSOD. Qualitative comparisons validate that the designed explicit prior modules (DCFM and MFFM) for pixel-level feature disparities and multiscale variability in pretrained ViT's underlying features achieve significant performance gains. + +Table 3. Open-vocabulary object detection benchmark results on COCO 2017. Bold text is SOTA result, gray text is all class and base class result. Backbone name is shown as abbreviation. + +
MethodYearBackboneExtra TrainnAP50bAP50AP50
OVR-CNN [48]21RN50✓22.846.039.9
ViLD [8]21EffNet-B7X27.659.551.3
RegionCLIP [52]22RN50x4✓39.361.655.7
OV-DETR [47]22ViT-B/32X29.461.052.7
BARON [41]23RN50X34.060.453.5
Detic [53]22RN50✓27.847.145.0
Rasheed et al. [31]22RN50✓36.654.049.4
Ro-ViT [17]23ViT-L/14X33.0-47.7
CoDet [24]24Swin-B✓30.652.346.6
F-VLM [19]22RN50x64X28.0-39.6
CORA [42]23RN50x4X41.744.543.8
BARON [41]23RN50✓42.754.951.7
CORA+ [42]23RN50x4✓43.160.956.2
DE-ViT [49]23ViT-L/14X50.054.052.9
Bharadwaj et al. [2]23RN101+SwinTX50.349.849.9
PiDiViT (Ours)25ViT-L/14X54.057.456.5
+ +![](images/050664ecd32cd881edfa230f62117bd9e4ed0153095dede40f5781568dcbc38d.jpg) +Figure 4. Visual performance comparison. Comparison of PiDi-ViT and baseline method DE-ViT [49] on COCO benchmarks for few-shot, open-vocabulary, and one-shot object detection tasks. Red dashed boxes indicate three cases of poor detection. + +# 4.5. Ablation Studies + +We perform ablation studies on COCO few-shot detection to validate core modules (DCFM, MFFM, and CSPS). As shown in Tab. 4, PiDiViT achieves an improvement of $+4.7$ nAP50 over DE-ViT [49] in a few-shot object detection setting $(\mathrm{T} = 1)$ . Progressive module integration experiments show that each module contributes to performance gains, with their combination yielding optimal results, validating both individual module design and synergistic effects. + +We evaluate three variants of DCFM (versions with attention-based fusion of APDC, RPDC, and CPDC [35] + +Table 4. Ablation study of all modules. Ablation results of innovative modules on COCO few-shot object detection benchmark. + +
MethodMFFMDCFMCSPSbAPnAPnAP50nAP75
DE-ViT [49]XXX28.631.149.434.8
Ours✓XX30.232.751.636.2
OursX✓X30.132.451.435.5
OursXX✓29.631.650.134.6
Ours✓✓✓31.234.854.138.3
+ +with standard convolution) in ablation studies, as shown in Tab. 5. The CPDC fused with standard convolution via attention effectively captures pixel-level boundary differential information and global context, thereby enhancing few-shot object feature representation and achieving optimal performance. In contrast, APDC (extracting angular differences) and RPDC (extracting radial differences) exhibit limited performance when fused with standard convolution, indicating their weaker adaptability to few-shot learning tasks. Therefore, in PiDiViT, we adopt the DCFM version combining CPDC with standard convolution. + +Table 5. Ablation study of DCFM. Ablation results of three difference layer combinations on COCO few-shot detection set. + +
MethodStrategiesbAPnAPnAP50nAP75
DE-ViT [49]X28.631.149.434.8
Ours+APDC28.830.949.034.3
Ours+RPDC28.931.049.434.3
Ours+CPDC29.632.451.435.8
+ +We investigate the impact of the strength of the CPDC block in the fusion of CPDC [35] and standard convolution in DCFM. As shown in Fig. 5, we evaluate seven strength gradients, where 0 reduces to the standard convolution and 1 applies the full CPDC. The results show that as the strength increases, the performance initially improves and then declines, with the optimal performance achieved at a strength of 0.875, which represents the best balance between differential boundary information and the global context. This optimal strength value is maintained across all subsequent PiDiViT experiments to ensure consistent performance gains. + +We ablate MFFM's convolution kernel combinations and fusion methods, as shown in Tab. 6. Beginning with a static weighted fusion of $1 \times 1$ , $3 \times 3$ , and $5 \times 5$ kernels, we enhance design by adding $7 \times 7$ kernels and $3 \times 3$ dilated convolutions (d2, d4). The results indicate that conv1,3,5,7,3d2 configuration achieves optimal performance by best balancing the extraction of multiscale features encompassing both global contours and local details. Additionally, attention-based fusion surpasses static fusion by dynamically weighting scale contributions. Thus, we adopt conv1,3,5,7,3d2 with attention fusion as MFFM's final configuration in PiDiViT. + +![](images/860fd40d4564ac6dff2e441ff10bdf5318b771d71f280e4c0e2118ff2fd2439d.jpg) +Figure 5. Ablation study of CPDC strength under CPDC [35] and standard convolution fusion. The ablation curve illustrates impact of CPDC strength settings in DCFM on the performance of COCO few-shot detection tasks. Best CPDC strength is 0.875. + +Table 6. Ablation study of MFFM. Ablation results of different multiscale combination schemes on COCO few-shot benchmark. + +
MethodKernel Comb.FusionbAPnAPnAP50nAP75
DE-ViT [49]conv3X28.631.149.434.8
Oursconv1,3,5Static29.332.350.735.6
Oursconv1,3,5,7,3d2Attention30.232.751.636.2
Oursconv1,3,5,7,3d4Attention29.532.651.236.6
+ +# 5. Conclusion + +We propose PiDiViT, a novel ViT architecture specifically designed for few-shot object detection. In PiDiViT, we design explicit prior modules (DCFM and MFFM) for pixel-level feature disparities and multiscale variability of pretrained ViT's underlying features. DCFM achieves differential enhancement of smooth center-to-boundary features while preserving global information, and MFFM enables the capture of local details and global contours across multiple scales. Experiments demonstrate that PiDiViT achieves SOTA results on COCO FSOD, OSOD, and OVD tasks. + +Despite these advancements, PiDiViT still has some limitations. The inherent bias in the pretrained ViT may influence the projections of class-specific feature, particularly when the class distribution is imbalanced. To address the issue of ViT bias, we plan to design a fairness-constrained algorithm for different classes in the future. This algorithm will dynamically adjust the class weights during training to mitigate the impact of pretrained ViT bias and further optimize the performance balance between the novel and base classes in scenarios with imbalanced class distributions. + +# 6. Acknowledgement + +This work was supported by the National Natural Science Foundation of China under Grant 62376283. + +# References + +[1] Simone Antonelli, Danilo Avola, Luigi Cinque, Donato Crisostomi, Gian Luca Foresti, Fabio Galasso, Marco Raoul Marini, Alessio Mecca, and Daniele Pannone. Few-shot object detection: A survey. CSUR, 54(11s):1-37, 2022. 1 +[2] Rohit Bharadwaj, Muzammal Naseer, Salman Khan, and Fahad Shahbaz Khan. Enhancing novel object detection via cooperative foundational models. arXiv preprint arXiv:2311.12068, 2023. 2, 7 +[3] Holger Caesar, Jasper Uijlings, and Vittorio Ferrari. Cocostuff: Thing and stuff classes in context. In CVPR, pages 1209-1218, 2018. 6 +[4] Ding-Jie Chen, He-Yen Hsieh, and Tyng-Luh Liu. Adaptive image transformer for one-shot object detection. In CVPR, pages 12247-12256, 2021. 2, 7 +[5] Qi Fan, Wei Zhuo, Chi-Keung Tang, and Yu-Wing Tai. Few-shot object detection with attention-rpn and multi-relation detector. In CVPR, pages 4013–4022, 2020. 2, 6 +[6] Zhibo Fan, Yuchen Ma, Zeming Li, and Jian Sun. Generalized few-shot object detection without forgetting. In CVPR, pages 4527-4536, 2021. 1, 2, 6 +[7] Yuqian Fu, Yu Wang, Yixuan Pan, Lian Huai, Xingyu Qiu, Zeyu Shangguan, Tong Liu, Yanwei Fu, Luc Van Gool, and Xingqun Jiang. Cross-domain few-shot object detection via enhanced open-set object detector. In ECCV, pages 247-264, 2025. 2, 6 +[8] Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, and Yin Cui. Open-vocabulary object detection via vision and language knowledge distillation. arXiv preprint arXiv:2104.13921, 2021. 7 +[9] Karim Guirguis, Johannes Meier, George Eskandar, Matthias Kayser, Bin Yang, and Jürgen Beyerer. Niff: Alleviating forgetting in generalized few-shot object detection via neural instance feature forging. In CVPR, pages 24193–24202, 2023. 2, 6 +[10] Guangxing Han, Yicheng He, Shiyuan Huang, Jiawei Ma, and Shih-Fu Chang. Query adaptive few-shot object detection with heterogeneous graph convolutional networks. In ICCV, pages 3263-3272, 2021. 2, 6 +[11] Guangxing Han, Shiyuan Huang, Jiawei Ma, Yicheng He, and Shih-Fu Chang. Meta faster r-cnn: Towards accurate few-shot object detection with attentive feature alignment. In AAAI, pages 780-789, 2022. 6 +[12] Guangxing Han, Jiawei Ma, Shiyuan Huang, Long Chen, and Shih-Fu Chang. Few-shot object detection with fully cross-transformer. In CVPR, pages 5321-5330, 2022. 2, 6 +[13] Kaiming He, Georgia Gkioxari, Piotr Dollár, and Ross Girshick. Mask r-cnn. In ICCV, pages 2961-2969, 2017. 6 +[14] Ting-I Hsieh, Yi-Chen Lo, Hwann-Tzong Chen, and Tyng-Luh Liu. One-shot object detection with co-attention and co-excitation. In NeurIPS, 2019. 2, 7 +[15] Bingyi Kang, Zhuang Liu, Xin Wang, Fisher Yu, Jiashi Feng, and Trevor Darrell. Few-shot object detection via feature reweighting. In ICCV, pages 8420-8429, 2019. 2, 6 +[16] Prannay Kaul, Weidi Xie, and Andrew Zisserman. Label, verify, correct: A simple few shot object detection method. In CVPR, pages 14237-14247, 2022. 6 + +[17] Dahun Kim, Anelia Angelova, and Weicheng Kuo. Region-aware pretraining for open-vocabulary object detection with vision transformers. In CVPR, pages 11144-11154, 2023. 7 +[18] Mona Köhler, Markus Eisenbach, and Horst-Michael Gross. Few-shot object detection: A comprehensive survey. IEEE TNNLS, 2023. 1, 2 +[19] Weicheng Kuo, Yin Cui, Xiuye Gu, AJ Piergiovanni, and Anelia Angelova. Open-vocabulary object detection upon frozen vision and language models. In ICLR, 2023. 7 +[20] Weijie Li, Wei Yang, Yuenan Hou, Li Liu, Yongxiang Liu, and Xiang Li. Saratr-x: Towards building a foundation model for sar target recognition. IEEE TIP, 2025. 2 +[21] Xiang Li, Lin Zhang, Yau Pun Chen, Yu-Wing Tai, and Chi-Keung Tang. One-shot object detection without fine-tuning. arXiv preprint arXiv:2005.03819, 2020. 2 +[22] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, pages 740-755, 2014. 2, 6 +[23] Li Liu, Lingjun Zhao, Yunli Long, Gangyao Kuang, and Paul Fieguth. Extended local binary patterns for texture classification. IMAVIS, 30(2):86-99, 2012. 4 +[24] Chuofan Ma, Yi Jiang, Xin Wen, Zehuan Yuan, and Xiaojuan Qi. Codet: Co-occurrence guided region-word alignment for open-vocabulary object detection. In NeurIPS, 2024. 7 +[25] Jiawei Ma, Yulei Niu, Jincheng Xu, Shiyuan Huang, Guangxing Han, and Shih-Fu Chang. Digeo: Discriminative geometry-aware learning for generalized few-shot object detection. In CVPR, pages 3208-3218, 2023. 6 +[26] Luke Melas-Kyriazi, Christian Rupprecht, Iro Laina, and Andrea Vedaldi. Deep spectral methods: A surprisingly strong baseline for unsupervised semantic segmentation and localization. In CVPR, pages 8364-8375, 2022. 1 +[27] Claudio Michaelis, Ivan Ustyuzhaninov, Matthias Bethge, and Alexander S Ecker. One-shot instance segmentation. arXiv preprint arXiv:1811.11507, 2018. 2, 6, 7 +[28] Timo Ojala, Matti Pietikainen, and Topi Maenpaa. Multiresolution gray-scale and rotation invariant texture classification with local binary patterns. IEEE TPAMI, 24(7):971-987, 2002. 4 +[29] Maxime Oquab, Timothée Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 3, 6 +[30] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763, 2021. 2 +[31] Hanoona Rasheed, Muhammad Maaz, Muhammad Uzair Khattak, Salman Khan, and Fahad Shahbaz Khan. Bridging the gap between object and image-level representations for open-vocabulary detection. In NeurIPS, pages 33781-33794, 2022. 7 +[32] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region + +proposal networks. Advances in neural information processing systems, 28, 2015. 3 +[33] Oriane Simeoni, Chloé Sekkat, Gilles Puy, Antonín Vopecký, Éloi Zablocki, and Patrick Pérez. Unsupervised object localization: Observing the background to discover objects. In CVPR, pages 3176-3186, 2023. 1 +[34] Zhuo Su, Matti Pietikainen, and Li Liu. Bird: Learning binary and illumination robust descriptor for face recognition. In BMVC, 2019. 4 +[35] Zhuo Su, Wenzhe Liu, Zitong Yu, Dewen Hu, Qing Liao, Qi Tian, Matti Pietikainen, and Li Liu. Pixel difference networks for efficient edge detection. In ICCV, pages 5117-5127, 2021. 4, 7, 8 +[36] Zhuo Su, Jiehua Zhang, Longguang Wang, Hua Zhang, Zhen Liu, Matti Pietikainen, and Li Liu. Lightweight pixel difference networks for efficient visual representation learning. IEEE TPAMI, 2023. 4 +[37] Zhuo Su, Li Liu, Matthias Müller, Jiehua Zhang, Diana Wofk, Ming-Ming Cheng, and Matti Pietikainen. Rapid salient object detection with difference convolutional neural networks. IEEE TPAMI, pages 1-16, 2025. 4 +[38] Carole H Sudre, Wenqi Li, Tom Vercauteren, Sebastien Ourselin, and M Jorge Cardoso. Generalised dice overlap as a deep learning loss function for highly unbalanced segmentations. In DLMIA, pages 240-248, 2017. 6 +[39] Bo Sun, Banghuai Li, Shengcai Cai, Ye Yuan, and Chi Zhang. Fsce: Few-shot object detection via contrastive proposal encoding. In CVPR, pages 7352-7362, 2021. 2, 6 +[40] Xin Wang, Thomas E. Huang, Trevor Darrell, Joseph E. Gonzalez, and Fisher Yu. Frustratingly simple few-shot object detection. In ICML, pages 9919-9928, 2020. 2, 6 +[41] Size Wu, Wenwei Zhang, Sheng Jin, Wentao Liu, and Chen Change Loy. Aligning bag of regions for open-vocabulary object detection. In CVPR, pages 15254-15264, 2023. 7 +[42] Xiaoshi Wu, Feng Zhu, Rui Zhao, and Hongsheng Li. Cora: Adapting clip for open-vocabulary detection with region prompting and anchor pre-matching. In CVPR, pages 7031-7040, 2023. 7 +[43] Yang Xiao, Vincent Lepetit, and Renaud Marlet. Few-shot object detection and viewpoint estimation for objects in the wild. IEEE TPAMI, 45(3):3090-3106, 2022. 2, 6 +[44] Xiaopeng Yan, Ziliang Chen, Anni Xu, Xiaoxi Wang, Xiaodan Liang, and Liang Lin. Meta r-cnn: Towards general solver for instance-level low-shot learning. In ICCV, pages 9577-9586, 2019. 2, 3, 6 +[45] Hanqing Yang, Sijia Cai, Hualian Sheng, Bing Deng, Jianqiang Huang, Xian-Sheng Hua, Yong Tang, and Yu Zhang. Balanced and hierarchical relation learning for one-shot object detection. In CVPR, pages 7591-7600, 2022. 2, 6, 7 +[46] Qihang Yu, Ju He, Xueqing Deng, Xiaohui Shen, and Liang-Chieh Chen. Convolutions die hard: Open-vocabulary segmentation with single frozen convolutional clip. In NeurIPS, pages 32215-32234, 2023. 2 +[47] Yuhang Zang, Wei Li, Kaiyang Zhou, Chen Huang, and Chen Change Loy. Open-vocabulary detr with conditional matching. In ECCV, pages 106-122, 2022. 7 + +[48] Alireza Zareian, Kevin Dela Rosa, Derek Hao Hu, and Shih-Fu Chang. Open-vocabulary object detection using captions. In CVPR, pages 14393-14402, 2021. 2, 7 +[49] Xinyu Zhang, Yuting Wang, and Abdeslam Boularias. Detect every thing with few examples. arXiv preprint arXiv:2309.12969, 2023. 1, 2, 3, 4, 5, 6, 7, 8 +[50] Yizhou Zhao, Xun Guo, and Yan Lu. Semantic-aligned fusion transformer for one-shot object detection. In CVPR, pages 7601-7611, 2022. 1, 2, 7 +[51] Yan Zhao, Lingjun Zhao, Ding Ding, Dewen Hu, Gangyao Kuang, and Li Liu. Few-shot class-incremental sar target recognition via cosine prototype learning. IEEE TGRS, 61: 1-18, 2023. 3 +[52] Yiwu Zhong, Jianwei Yang, Pengchuan Zhang, Chunyuan Li, Noel Codella, Liunian Harold Li, Luowei Zhou, Xiyang Dai, Lu Yuan, Yin Li, et al. Regionclip: Region-based language-image pretraining. In CVPR, pages 16793-16803, 2022. 6, 7 +[53] Xingyi Zhou, Rohit Girdhar, Armand Joulin, Philipp Krahenbuhl, and Ishan Misra. Detecting twenty-thousand classes using image-level supervision. In ECCV, pages 350-368, 2022. 7 \ No newline at end of file diff --git a/ICCV/2025/When Pixel Difference Patterns Meet ViT_ PiDiViT for Few-Shot Object Detection/images.zip b/ICCV/2025/When Pixel Difference Patterns Meet ViT_ PiDiViT for Few-Shot Object Detection/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..cbf9cf61e4d3d0d467c48ed4ce95aebf964ff868 --- /dev/null +++ b/ICCV/2025/When Pixel Difference Patterns Meet ViT_ PiDiViT for Few-Shot Object Detection/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:899f9d1894b33e8860a67c5649d16c25f777bd7b91ad9b705984bfda4de940ea +size 750133 diff --git a/ICCV/2025/When Pixel Difference Patterns Meet ViT_ PiDiViT for Few-Shot Object Detection/layout.json b/ICCV/2025/When Pixel Difference Patterns Meet ViT_ PiDiViT for Few-Shot Object Detection/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..d60e18bdf7f19874d7820ad6c59b2b7ffb93864b --- /dev/null +++ b/ICCV/2025/When Pixel Difference Patterns Meet ViT_ PiDiViT for Few-Shot Object Detection/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5240bd02aca8247580d192de51f795cea831be1d40b563691c34a16bca71c9d4 +size 357171 diff --git a/ICCV/2025/When Schrodinger Bridge Meets Real-World Image Dehazing with Unpaired Training/c7189c5a-d018-4b96-8859-6c17a385c5d1_content_list.json b/ICCV/2025/When Schrodinger Bridge Meets Real-World Image Dehazing with Unpaired Training/c7189c5a-d018-4b96-8859-6c17a385c5d1_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..db130ebc172848fe91ac6553f4636e3d2b9198f2 --- /dev/null +++ b/ICCV/2025/When Schrodinger Bridge Meets Real-World Image Dehazing with Unpaired Training/c7189c5a-d018-4b96-8859-6c17a385c5d1_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:89a5757413a334bfd4f05cf4cf08846db202dca450e2400372fec76387053429 +size 86362 diff --git a/ICCV/2025/When Schrodinger Bridge Meets Real-World Image Dehazing with Unpaired Training/c7189c5a-d018-4b96-8859-6c17a385c5d1_model.json b/ICCV/2025/When Schrodinger Bridge Meets Real-World Image Dehazing with Unpaired Training/c7189c5a-d018-4b96-8859-6c17a385c5d1_model.json new file mode 100644 index 0000000000000000000000000000000000000000..e89e8ea0645bd19bd78c570d0e7437605dd62e27 --- /dev/null +++ b/ICCV/2025/When Schrodinger Bridge Meets Real-World Image Dehazing with Unpaired Training/c7189c5a-d018-4b96-8859-6c17a385c5d1_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:74dcc457906c036f5999843fc94ee0c11f8718a579dff13a34bf03d4bd245654 +size 107094 diff --git a/ICCV/2025/When Schrodinger Bridge Meets Real-World Image Dehazing with Unpaired Training/c7189c5a-d018-4b96-8859-6c17a385c5d1_origin.pdf b/ICCV/2025/When Schrodinger Bridge Meets Real-World Image Dehazing with Unpaired Training/c7189c5a-d018-4b96-8859-6c17a385c5d1_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..14dc8271749c7af964385a7e892b5fd826e50dc0 --- /dev/null +++ b/ICCV/2025/When Schrodinger Bridge Meets Real-World Image Dehazing with Unpaired Training/c7189c5a-d018-4b96-8859-6c17a385c5d1_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e40bb40f502daa36d458ff253df19cbacab2e4d0c251afe0b262e79cc74f1852 +size 9979222 diff --git a/ICCV/2025/When Schrodinger Bridge Meets Real-World Image Dehazing with Unpaired Training/full.md b/ICCV/2025/When Schrodinger Bridge Meets Real-World Image Dehazing with Unpaired Training/full.md new file mode 100644 index 0000000000000000000000000000000000000000..064b995f2ffcad9c04842b0d189c22bade552c5d --- /dev/null +++ b/ICCV/2025/When Schrodinger Bridge Meets Real-World Image Dehazing with Unpaired Training/full.md @@ -0,0 +1,394 @@ +# When Schrödinger Bridge Meets Real-World Image Dehazing with Unpaired Training + +Yunwei Lan $^{1,2}$ , Zhigao Cui $^{1*}$ , Xin Luo $^{2}$ , Chang Liu $^{2}$ , Nian Wang $^{1}$ , Menglin Zhang $^{2}$ , Yanzhao Su $^{1}$ , Dong Liu $^{2*}$ + +$^{1}$ Rocket Force University of Engineering, Xi'an 710025, China + +$^{2}$ MOE Key Laboratory of Brain-Inspired Intelligent Perception and Cognition, + +University of Science and Technology of China, Hefei 230093, China + +# Abstract + +Recent advancements in unpaired dehazing, particularly those using GANs, show promising performance in processing real-world hazy images. However, these methods tend to face limitations due to the generator's limited transport mapping capability, which hinders the full exploitation of their effectiveness in unpaired training paradigms. To address these challenges, we propose DehazeSB, a novel unpaired dehazing framework based on the Schrödinger Bridge. By leveraging optimal transport (OT) theory, DehazeSB directly bridges the distributions between hazy and clear images. This enables optimal transport mappings from hazy to clear images in fewer steps, thereby generating high-quality results. To ensure the consistency of structural information and details in the restored images, we introduce detail-preserving regularization, which enforces pixel-level alignment between hazy inputs and dehazed outputs. Furthermore, we propose a novel prompt learning to leverage pre-trained CLIP models in distinguishing hazy images and clear ones, by learning a haze-aware vision-language alignment. Extensive experiments on multiple real-world datasets demonstrate our method's superiority. + +# 1. Introduction + +Images captured under hazy conditions usually display reduced brightness and diminished contrast, significantly degrading visual quality and hindering further applications in visual tasks such as object detection [11]. Consequently, image dehazing, which aims to restore clear images from hazy ones, has received considerable attention over the past + +![](images/44d073c809c276f2dd1e8bf7e46df90ccfd23014c09c260b68857e1f550f0c7a.jpg) + +![](images/51a512d15156b3d80cd07ad7482711706ab9cf3a88c732b6a0ed26371ca380e6.jpg) + +![](images/9980f37482ae28ae0cc5b008c0087e420117b184d48c68a3fc9578834df57755.jpg) +Figure 1. Existing GAN-based unpaired dehazing methods struggle to handle images with complex structures due to the limited mapping capability of the generator. While diffusion model-based methods can achieve more intricate transport mappings, they necessitate transforming images into Gaussian noise and then regenerating them through a large number of sampling steps. Our Schrödinger Bridge enables the direct learning of optimal transport mapping from hazy to clear image distributions in fewer time steps, achieving high efficiency and high-quality dehazing. + +decade. The degradation process of hazy images typically adheres to the Atmospheric Scattering Model (ASM) [31]. + +$$ +I (x) = J (x) t (x) + A (1 - t (x)), \tag {1} +$$ + +where $I(x)$ represents the hazy image, and $J(x)$ is the corresponding clear one. $A$ and $t(x)$ are the atmospheric light and transmission map. + +Early studies [3, 16, 51] primarily rely on physical priors to estimate the parameters necessary for the ASM, subsequently obtaining clear images. Although these methods can effectively dehaze images, they are susceptible to over-dehazing since hand-crafted priors are often scenario-specific. Recently, the advent of deep learning has spurred + +a shift towards utilizing deep learning for dehazing [4, 7, 13, 25, 33, 48]. These methods typically utilize synthetic paired data for training since real-world hazy and clear image pairs are normally tough to collect. Although improved performance is observed, such paired training paradigm often fails to generalize to real-world scenarios, due to the ill-presenting performance of trained networks that lack real-world information from hazy images. To tackle the bottleneck of paired training, recent studies [24, 40, 43, 47] explore unpaired training for dehazing without the requirements of data pairs. This paradigm enables existing model architectures to be directly trained on real-world data, making it more suitable for real-world image dehazing. + +As shown in Figure 1, existing unpaired dehazing methods [9, 40, 43, 47] predominantly leverage Generative Adversarial Networks (GANs) [14] to generate dehazed images. Typically, these methods first produce preliminary dehazed results using a generator, often based on an Encoder-Decoder. They then align the distribution of these results with that of unpaired clear images using a discriminator. However, these methods often struggle to handle images with complex and detailed structures due to limitations in the generator's ability to model transport mappings. Recent advancements in diffusion models [18, 35, 39] offer an effective alternative. Diffusion models can simulate more intricate transport mappings through an iterative process of adding and removing noise, generating diverse and high-quality dehazed results. Although this capability is difficult for GANs to match, diffusion models must first transform the image into Gaussian noise and then generate the image through a large number of sampling steps. This process imposes limitations on their potential for unpaired dehazing and leads to inferior computational efficiency. + +Another key challenge in unpaired dehazing is the way to ensure the structural consistency between hazy and dehazed images, so as to suppress misaligned information caused by unpaired clear images. While CycleGAN [50] serves as a classical solution and has been widely used in previous unpaired dehazing methods [9, 22, 43, 44], it requires additional training of generators and discriminators, leading to additional computational cost. Additionally, we observe that a pre-trained CLIP model [34] can effectively distinguish between hazy and clear images. This insight motivates us to explore improving dehazing by maximizing the distance between hazy textual prompts and dehazed images. However, we find that directly providing textual descriptions is not the optimal solution. This may be due to the fact that the concentration and types of haze in real-world images are extremely complex, and simple textual descriptions fail to capture these characteristics. + +To address the aforementioned challenges, we propose a novel framework for real-world image dehazing, namely DehazeSB, consisting of the following contributions: + +- We apply Schrödinger Bridge (SB) in our unpaired dehazing framework. Leveraging SB based on optimal transport (OT) theory, we establish direct bridges between hazy and clear image distributions. This enables us to find optimal transport mappings from hazy to clear images, thereby improving dehazing performance. Moreover, SB can generate high-quality images in just a few or even one step, significantly reducing computational overhead. +- To enforce a structure-consistent constraint while minimizing training overhead, we introduce detail-preserving regularization. This ensures content and texture consistency during the dehazing process. +- We propose a novel prompt learning strategy to tailor the textual representations for hazy images. This learned prompt plays a more effective role in representing hazy images instead of direct textual descriptions. +- Extensive quantitative and qualitative experiments show that our method surpasses existing state-of-the-art dehazing methods, including paired or unpaired training. + +# 2. Related Work + +# 2.1. Image Dehazing + +Early prior-based methods focus on obtaining the ASM's parameters, and then restoring clear images by reserving the ASM. For example, DCP [16] proposes a dark channel prior to estimate the atmospheric light and transmission map. With the development of deep learning, some efforts concentrate on designing networks to estimate the parameters of ASM or directly learn the mapping from hazy to clear images. DehazeNet [4] is the pioneer in estimating transmission maps in an end-to-end manner. C2PNet [48] customizes a curriculum learning to improve dehazing performance. Despite the existence of various proposed networks, their dehazing performance on real-world images remains limited since they are trained on synthetic datasets. + +To address the bottleneck of paired training, recent studies have focused on unpaired training strategies to improve the model's generalization capabilities. For example, RefineDNet [47] integrates the dark channel prior (DCP) into a two-stage dehazing network for unpaired learning. D4+ [44] decomposes the ASM and develops a self-augmented dehazing framework built on CycleGAN [50]. UCL-Dehaze [40] proposes an unpaired contrastive learning paradigm for image dehazing. These methods all employ GANs to generate dehazed images. Despite achieving better performance on real-world hazy images, they still fail to produce satisfactory dehazing results due to the limited transmission mapping capability of the generator. + +# 2.2. Schrödinger Bridge based Image Translation + +The Schrödinger Bridge [37] problem involves finding the stochastic process that transitions from an initial distribu + +![](images/0b3abc31a94f10ef350f673998112aeb160e09d83a04599b6a8aa28eb85b8b88.jpg) +Figure 2. Overview of our method. We first learn a haze-aware prompt using prompt learning. This learned prompt is then used to guide the training of the Schrödinger Bridge (The part within the black dashed box on the right). For SB training, we discretize the unit interval $[0,1]$ into $N$ intervals with uniform spacing, where $t_j < t_{j+1}$ , $t_0 = 0$ , and $t_N = 1$ . $x(t_0) = x_0$ denotes the hazy image sampled from $\pi_0$ . $x_1$ denotes the unpaired clear image sampled from $\pi_1$ . $x_1(x(t_j))$ denotes the generated dehazed image from intermediate states $x(t_j)$ using generator $G$ . The generation process involves generating $x(t_{j+1})$ from $x(t_j)$ (details are provided within the green dashed box). During the training of the SB, we use the hazy image $x_0$ to generate an intermediate state $x(t_i)$ through an iterative generation process with $j = 0,1,2,\ldots,i-1$ . Next, we feed $x(t_i)$ into the generator $G: q_\theta(x_1 \mid x(t_i))$ to produce a dehazed image $x_1(x(t_i))$ . Finally, we calculate the $L_{adv}$ , $L_{SB}$ and $L_p$ using the unpaired clear image $x_1$ , the generated dehazed image $x_1(x(t_i))$ , the hazy image $x_0$ , and the learned prompt. Furthermore, we impose a detail-preserving regularization between $x_0$ and $x_1(x(t_i))$ to enforce structural and detail consistency, ensuring that the dehazed image aligns pixel-by-pixel with the hazy image. In the inference process, we input hazy images directly into generator $G$ to produce dehazed images with different numbers of function evaluations (NFEs). + +![](images/a437654f066c081b2d46423aedbcbbc7536201172ca2206d81190c27a7f2b8f7.jpg) + +tion to a target distribution while adhering to a reference measure. The Schrödinger Bridge provides the flexibility to select arbitrary distributions as both the initial and terminal distributions, making it highly versatile. For instance, $\mathrm{I}^2\mathrm{SB}$ [27] proposes the Image-to-Image Schrödinger Bridge, learning nonlinear diffusion processes between two given distributions. UNSB [20] is the first to achieve high-resolution unpaired image translation using adversarial learning. CUNSB-RFIE [8] introduces a context-aware unpaired neural Schrödinger Bridge for retinal fundus image enhancement. However, despite these advancements, applying the Schrödinger Bridge to image dehazing remains challenging, since the degradation mechanisms of images captured under hazy conditions are extremely complex. + +# 3. Preliminary + +# 3.1. Schrödinger Bridge + +The Schrödinger Bridge problem involves finding an optimal stochastic process $\{x(t):t\in [0,1]\}$ that connects two probability distributions $\pi_0$ and $\pi_{1}$ on $\mathbb{R}^d$ , while respecting a reference measure $\mathbb{W}$ , typically chosen to be the Wiener measure. Formally, let $\Omega$ represent the path space and $\mathcal{P}(\Omega)$ denote the probability measure on $\Omega$ . The Schrödinger + +Bridge problem can be mathematically formulated as: + +$$ +\mathbb {Q} ^ {S B} = \underset {\mathbb {Q} \in \mathcal {P} (\Omega)} {\arg \min } D _ {K L} (\mathbb {Q} \| \mathbb {W}) \quad \text {s . t .} \quad \mathbb {Q} _ {0} = \pi_ {0}, \mathbb {Q} _ {1} = \pi_ {1}, \tag {2} +$$ + +where $\mathbb{Q}^{SB}$ denotes the Schrodinger Bridge solution, and $\mathbb{Q}_t$ corresponds to the marginal distribution of $\mathbb{Q}$ at time $t$ . $D_{KL}(\mathbb{Q} \parallel \mathbb{W})$ represents the Kullback-Leibler divergence. This formulation establishes a probabilistic coupling that enables bidirectional sampling between the distributions. Specifically, given samples $x_0 \sim \pi_0$ , the bridge $\mathbb{Q}^{SB}$ facilitates the generation of corresponding samples $x_1 \sim \pi_1$ through the interpolating process. The reverse direction is equally attainable through time reversal. + +# 3.2. Static Schrödinger Bridge + +The Schrödinger Bridge enables the direct learning of optimal transport between arbitrary distributions. However, the inherent temporal dependencies in the full process make the direct optimization of $\mathbb{Q}^{SB}$ challenging. The Static Schrödinger Bridge (SSB) offers a solution by simplifying the Schrödinger Bridge problem into a more tractable formulation. This reduction establishes an equivalence between the Schrödinger Bridge $\mathbb{Q}^{SB}$ and its static formulation $\mathbb{Q}_{01}^{SB}$ , which characterizes the joint distribution of the initial distribution $\pi_0$ and the terminal distribution $\pi_1$ . + +Let $\Pi (\pi_0,\pi_1)$ denote the space of joint distributions with margins $\pi_0$ and $\pi_{1}$ . The SSB problem can be formulated as an entropy-regularized optimal transport problem: + +$$ +\mathbb {Q} _ {0 1} ^ {\mathrm {S B}} = \underset {\gamma \in \Pi \left(\pi_ {0}, \pi_ {1}\right)} {\arg \min } \mathbb {E} _ {\left(x _ {0}, x _ {1}\right) \sim \gamma} \left[ \| x _ {0} - x _ {1} \| ^ {2} \right] - 2 \tau H (\gamma), \tag {3} +$$ + +where $H(\gamma)$ denotes the entropy, and $\tau$ controls the amount of randomness in the trajectory. Under this formulation, the marginal distribution $x(t):t\in [0,1]$ can be obtained through conditional Gaussian distributions: + +$$ +p (x (t) \mid x _ {0}, x _ {1}) \sim \mathcal {N} (x (t); t x _ {1} + (1 - t) x _ {0}, t (1 - t) \tau I), \tag {4} +$$ + +where $I$ is the identity matrix. A crucial self-similarity property emerges in $\mathbb{Q}_{01}^{\mathrm{SB}}$ , as demonstrated in UNSB [20]. For any sub-interval $[t_{\alpha}, t_{\beta}] \subseteq [0, 1]$ , the restricted process $Q_{t_{\alpha}t_{\beta}}^{SB}$ remains a Schrödinger Bridge, which can be formulated as conditional Gaussian distributions: + +$$ +p (x (t) \mid x (t _ {\alpha}), x (t _ {\beta})) \sim \mathcal {N} (x (t); \mu (t), \sigma (t)), +$$ + +$$ +\mu (t) = s (t) x \left(t _ {\beta}\right) + (1 - s (t)) x \left(t _ {\alpha}\right), \tag {5} +$$ + +$$ +\sigma (t) = s (t) (1 - s (t)) \tau (t _ {\beta} - t _ {\alpha}) I, +$$ + +where $s(t) \coloneqq (t - t_{\alpha}) / (t_{\beta} - t_{\alpha})$ . This enables recursive interpolation between intermediate states. + +# 4. Method + +As shown in Figure 2, We propose a novel unpaired dehazing framework, DehazeSB, based on the Schrödinger Bridge. It comprises three components: prompt learning, Schrödinger Bridge, and detail-preserving regularization. + +# 4.1. Prompt Learning + +Inspired by CLIP's [26, 30, 34] demonstrated capability for cross-modal alignment, we observe that its discriminative power extends to distinguishing hazy and clear images when provided with textual descriptors (as validated in Figure 3). This insight motivates us to leverage prompts to facilitate dehazing. However, we have empirically observed that generic prompts such as "hazy image" yield suboptimal results. This limitation arises from the wide variation in haze concentration and distribution found in real-world hazy images, which simplistic prompts fail to adequately capture. + +To address this problem, we propose a novel prompt learning that exclusively optimizes haze-aware prompts. Specifically, we randomly initialize a learnable textual prompt $T_{hazy}$ . Subsequently, we leverage the pre-trained CLIP's image encoder $\Phi_{image}$ and text encoder $\Phi_{text}$ to separately extract features from hazy images $I_{hazy}$ , clear images $I_{clear}$ , and initialized haze-aware prompt $T_{hazy}$ . We train this prompt using a binary cross-entropy (BCE) loss: + +$$ +L _ {B C E} = - (y \log \hat {y} + (1 - y) \log (1 - \hat {y})), \tag {6} +$$ + +![](images/143277f07b479ae9e685429ec0e6985d4394aa53cec6fd86cbb24fb58791b35b.jpg) +Hazy image. 0.93 + +![](images/b4501b220c3aa5b9bc4bab2ccb05d8087ceec5148b6d74914c94fec602f9dea6.jpg) + +![](images/b4319940ff649530927d30ffe8a57d09472b4452c97b4e6519b24dbd96d111fa.jpg) +Rain image. 0.63 + +![](images/7e6ab014326d89982dde5d2a1bf1981fb54eaec103bc490020a4d03dfd3f2334.jpg) +Low-light image. 0.81 + +![](images/5ec4694ec7626b48547b7083b1ce6282510101031bbfce95ad3992515d57d24c.jpg) +Clear image. 0.42 +Our learned prompt. 0.97 +Figure 3. CLIP scores of different prompts. We observe that appropriate prompts are effective in distinguishing between hazy and clear images (for instance, the prompt "hazy image" receives a higher score on hazy images). However, due to the complex degradation characteristics of real-world hazy images, simplistic prompts like "hazy image" fail to accurately capture these nuances. In contrast, our learned haze-aware prompt demonstrates superior effectiveness, achieving the highest scores. + +![](images/b61cf8d732695cf168192e34a1c6fdb76fa2d846613048baec58f3df895a7eab.jpg) +Hazy image. 0.88 + +![](images/cc3065c089ebc9a9a2c4b07858cfabc940e627ac286611ab5af08b47d1e96b6a.jpg) + +![](images/17b1d16cf58f8f0dcf57aedd1b5a8ea496730cfad20cac1971717241676029ff.jpg) +Rain image. 0.81 + +![](images/9026860f6651f6c8d723c1ce0934dbcccc5e6bf1b64fc32019c214a022623645.jpg) +Low-light image. 0.37 + +![](images/727e856144d303215c536db40a28f1d4cec25d5d548d72d76ed9742b32a82fad.jpg) +Clear image. 0.07 +Our learned prompt. 0.95 + +$$ +\hat {y} = \frac {e ^ {\cos \left(\left(\Phi_ {\text {i m a g e}} (I) , \Phi_ {\text {t e x t}} \left(T _ {\text {h a z y}}\right)\right)\right)}}{\sum_ {i \in \{h a z y , c l e a r \}} e ^ {\cos \left(\left(\Phi_ {\text {i m a g e}} \left(I _ {i}\right) , \Phi_ {\text {t e x t}} \left(T _ {\text {h a z y}}\right)\right)\right)}}, \tag {7} +$$ + +where $I \in \{I_{hazy}, I_{clear}\}$ , and $y$ is the label of $I$ . 0 is for negative sample $I_{clear}$ and 1 is for positive sample $I_{hazy}$ . The learned haze-aware prompt subsequently guides the training of the Schrödinger Bridge. Specifically, given a generated dehazed image $I_{dehaze}$ (denoted by $x_1(x(t_i))$ in Figure 2) and its hazy counterpart $I_{hazy}$ (denoted by $x_0$ ), we guide the dehazing by pushing the learned haze-aware prompt away from dehazed images. This loss function can also be expressed as Eq. 6, where $\hat{y}$ is formulated as: + +$$ +\hat {y} = \frac {e ^ {\cos \left(\left(\Phi_ {\text {i m a g e}} (I) , \Phi_ {\text {t e x t}} \left(T _ {\text {h a z y}}\right)\right)\right)}}{\sum_ {i \in \{h a z y , d e h a z e \}} e ^ {\cos \left(\left(\Phi_ {\text {i m a g e}} \left(I _ {i}\right) , \Phi_ {\text {t e x t}} \left(T _ {\text {h a z y}}\right)\right)\right)}}, \tag {8} +$$ + +where $I \in \{I_{hazy}, I_{dehaze}\}$ . Specifically, 0 is for negative sample $I_{dehaze}$ and 1 is for positive sample $I_{hazy}$ . + +# 4.2. Schrödinger Bridge + +For SB training, we discretize the unit interval $[0,1]$ into $N$ intervals with uniform spacing ( $N = 5$ in our framework). According to the Markov chain decomposition, we can simulate the SB between hazy and clear images as: + +$$ +p \left(\left\{x \left(t _ {n}\right) \right\}\right) = p \left(x \left(t _ {N}\right) \mid x \left(t _ {N - 1}\right)\right) \dots p \left(x \left(t _ {1}\right) \mid x \left(t _ {0}\right)\right) p \left(x \left(t _ {0}\right)\right), \tag {9} +$$ + +where $t_i < t_{i+1}$ , $t_0 = 0$ , and $t_N = 1$ . $\{x(t_n)\}$ represents a partition $\{x(t_i)\}_{i=0}^N$ of the unit interval [0, 1]. Given the initial hazy image distribution $p(x(t_0))$ , learning the transition probabilities $p(x(t_{i+1}) | x(t_i))$ enables recursive + +sampling to approximate the target distribution $p(x(t_{N}))$ , where $p(x(t_0))$ and $p(x(t_N))$ are also referred to as $p(x_0)$ and $p(x_{1})$ . Therefore, we need to design a network with parameters $\theta_{i}$ as the generator $G: q_{\theta_i}(x(t_{i + 1}) \mid x(t_i))$ to modulate $p(x(t_{i + 1}) \mid x(t_i))$ , allowing us to sample from the initial distribution and obtain the target distribution. + +However, it is impractical to devise a separate generator $G$ for each time step. Following the UNSB [20], we devise an elaborated generator $G: q_{\theta}(x(t_N) \mid x(t_i))$ conditioned on time step $t_i$ , which can also be denoted as $G: q_{\theta}(x_1 \mid x(t_i))$ . This generator directly predicts the target $x_1$ from any intermediate state $x(t_i)$ . In doing so, the optimization objective can be expressed as: + +$$ +\begin{array}{l} L _ {S B} := \mathbb {E} _ {q _ {\theta} (x (t _ {i}), x _ {1})} \left[ \| x (t _ {i}) - x _ {1} \| ^ {2} \right] \\ - 2 \tau (1 - t _ {i}) H \left(q _ {\theta} \left(x \left(t _ {i}\right), x _ {1}\right)\right) \tag {10} \\ \end{array} +$$ + +$$ +\begin{array}{l} \text {s . t .} L _ {\mathrm {a d v}} := D _ {\mathrm {K L}} \left(q _ {\theta} \left(x _ {1}\right) \| p \left(x _ {1}\right)\right) = 0. \end{array} +$$ + +Using Lagrange multipliers, we can transform the optimization objective into a loss function: + +$$ +L := L _ {a d v} + \lambda_ {S B} L _ {S B}. \tag {11} +$$ + +The training pipeline is illustrated in Figure 2. Starting from an initial sample $x(t_0) = x_0 \sim \pi_0$ , we progressively generate intermediate states $\{x(t_j)\}_{j=1}^i$ through a Markov chain process. After obtaining $x(t_i)$ , we use the generator $G: q_{\theta}(x_1 \mid x(t_i))$ to produce the dehazed image $x_1(x(t_i))$ . We then calculate the $L_{SB}$ and $L_{adv}$ in Eq.10 using the pairs $(x(t_i), x_1(x(t_i)))$ and $(x_1(x(t_i)), x_1)$ . Note that we randomly choose different time steps $t_i$ to optimize during the training, thus our model can generate target images with an arbitrary number of function evaluations (NFEs). For the generation process of $x(t_{j+1})$ at each step $j = 0, 1, 2, \dots, i-1$ , we first predict the target domain sample $x_1(x(t_j))$ using the generator $G: q_{\theta}(x_1 \mid x(t_j))$ . Subsequently we generate $x(t_{j+1})$ through the Eq. 5, which can be expressed as: + +$$ +\begin{array}{l} p (x (t _ {j + 1}) \mid x (t _ {j}), x _ {1} (x (t _ {j}))) \sim \\ \mathcal {N} (x (t _ {j + 1}); \mu (t _ {j + 1}), \sigma (t _ {j + 1})), \\ \mu \left(t _ {j + 1}\right) = s \left(t _ {j + 1}\right) x _ {1} \left(x \left(t _ {j}\right)\right) + \left(1 - s \left(t _ {j + 1}\right)\right) x \left(t _ {j}\right), \tag {12} \\ \sigma \left(t _ {j + 1}\right) = s \left(t _ {j + 1}\right) \left(1 - s \left(t _ {j + 1}\right)\right) \tau \left(1 - t _ {j}\right) I. \\ \end{array} +$$ + +To align the distribution of generated images $q_{\theta}(x_1)$ with the target distribution $p(x_1)$ , we replace the KL divergence in Eq.10 with a dual-discriminator adversarial strategy. Specifically, We first employ a Markovian discriminator which distinguishes samples on the patch level since it is effective at capturing details and textures in local characteristics. To complement patch-level analysis, we introduce a CLIP-based discriminator [21] that assesses overall structural coherence and scene consistency. This mitigates limitations of purely local evaluation, which can + +overlook implausible global arrangements. By combining these discriminators, our method simultaneously enhances local realism and global plausibility. Finally, We compute $H(q_{\theta}(x(t_i), x_1))$ in a manner similar to the approach used in UNSB [20], as detailed in our supplementary materials. + +# 4.3. Detail-Preserving Regularization + +Another critical challenge in unpaired image dehazing is detail preservation. The generated images must not only display high visual quality but also maintain fidelity in details and textures compared to the original images. Although CycleGAN [50] is a classic solution for this problem, it necessitates the training of additional generators and discriminators, leading to high computational demands. To overcome this limitation, we introduce a detail-preserving regularization that includes PatchNCE regularization, physical prior regularization, and high-frequency detail regularization. + +# 4.3.1. PatchNCE Regularization + +Following the CUT [32], we employ PatchNCE regularization to enhance the generator's ability to preserve content consistency during image dehazing. Specifically, PatchNCE regularization leverages contrastive learning to enforce similarity between corresponding patches in the input and output. This helps the generator learn to maintain important structural details and textures, thereby improving the overall quality of the dehazed images. For more detailed information, please refer to our supplementary materials. + +# 4.3.2. Physical Prior Regularization + +Previous studies [5, 43, 47] have demonstrated that using physical priors can enhance real-world image dehazing. Moreover, they enable us to preserve as much detail as possible. Specifically, we process the hazy image $I$ (also referred to as $x_0$ ) using Dark Channel Prior (DCP) [16], obtaining the atmospheric light $A$ , coarse transmission map $t$ , and dehazed image $J_{dcp}$ . We then feed $t$ into a UNet [36] to obtain a refined transmission map $t_{ref}$ . Finally, we reconstruct the hazy image $I_{phy}$ according to the Eq. 1 and formulate the physical prior regularization as follows: + +$$ +L _ {p h y} = L _ {r e c} (I, I _ {p h y}), \tag {13} +$$ + +where $I_{phy} = Jt_{ref} + A(1 - t_{ref})$ refers to the reconstructed hazy image derived using the ASM, and $J$ denotes the generated dehazed image $x_{1}(x(t_{i}))$ . $L_{rec}$ is the combined distance of $L_{1}$ and LPIPS. [46]. + +# 4.3.3. High-Frequency Detail Regularization + +To enhance structural fidelity in the restored images while preserving high-frequency details, we introduce a high-frequency detail regularization that includes Discrete Fourier Transform (DFT) loss, Structural Similarity Index Measure (SSIM) loss, and Sobel Gradient loss. We provide more detailed information in our supplementary materials. + +![](images/462db8411ee1ec702cffa20b9e01ae9a522dfcc332452737d1f32892ceb429ce.jpg) +Figure 4. Visual comparison of samples from RTTS and Haze2020. Our method can effectively remove haze and generate high-quality images with natural color and realistic contrast. More than 40 visual results are presented in our supplementary materials. + +
RTTSHaze2020
TypeMethodFID↓NIQE↓MUSIQ↑MANIQA ↑FID↓NIQE↓MUSIQ↑MANIQA↑
Prior-basedDCP [16]73.0174.27152.9350.11991.9483.82754.3750.121
PairedMBFormer [33]66.0234.87453.5760.13982.5964.12855.5500.141
PairedC2PNet [48]66.1175.03753.9600.14283.9594.20654.5650.138
PairedKANet [13]64.9634.33954.5130.11084.8883.74256.4110.127
PairedDEANet [6]66.3554.80553.6280.13284.1973.97654.7570.135
PairedDiff-Plugin [28]65.7875.33450.7400.09580.9654.40752.7520.111
PairedOneRestore [15]65.9345.32151.3800.11484.2764.36552.0660.116
PairedSGDN [10]78.0185.71644.8580.08189.0615.91446.3260.094
UnpairedCUT [32]98.7713.93859.7980.14675.6254.26158.6680.135
UnpairedUNSB [8]56.6683.83455.0480.13272.8124.13755.3840.132
UnpairedYOLY [24]69.9964.55351.8740.12985.6483.93754.1640.138
UnpairedRefineDNet [47]65.0764.01256.1170.10888.2293.97254.7790.122
UnpairedD4 [43]69.4004.63758.4450.13987.5363.97153.4580.119
UnpairedD4+ [44]67.6874.27453.5990.13187.1013.86553.6120.124
UnpairedOurs53.1203.76060.1140.15369.7963.74359.2560.150
+ +Table 1. Quantitative results on RTTS and Haze2020. The best results are denoted in bold, and the second-best results are underlined. + +The final loss function can be formulated as: + +$$ +L = L _ {a d v} + \lambda_ {S B} L _ {S B} + \lambda_ {p} L _ {p} + \tag {14} +$$ + +$$ +\lambda_ {N C E} L _ {N C E} + \lambda_ {p h y} L _ {p h y} + \lambda_ {h f d} L _ {h f d}, +$$ + +where $L_{NCE}, L_{phy}$ , and $L_{hfd}$ denote the PatchNCE, physical prior, and high-frequency detail regularization loss. Each term is weighted by its corresponding coefficients $\lambda$ : 1, 1, 1, 0.5, and 0.5 respectively. + +# 5. Experiments and Discussion + +Our method enables direct training on unpaired real-world hazy and clear images. Specifically, we collect over 7,000 real hazy images from the RESIDE [23] for our training set + +of hazy images. Additionally, we gather over 10,000 clear images from the SOTS (a subset of RESIDE) and ADE20K [49] to serve as our training set of clear images. To comprehensively evaluate our method, we conduct extensive quantitative and qualitative comparisons using multiple real-world datasets: RTTS, URHI, Haze2020, OHAZE, IHAZE, and Fattal's dataset. Among them, RTTS and URHI are subsets of RESIDE, each containing over 4,000 hazy images. Haze2020, introduced by DA [38], includes more than 1,000 real-world images. OHAZE [2] and IHAZE [1] contain 45 and 55 pairs of hazy images, respectively, generated using a haze machine. Fattal's dataset [12] comprises over 30 real-world hazy images. More training details and parameters are in the supplementary materials. + +![](images/737d3ef6ac574bd7fcb5cddda602303b7d00cfc3d3232eb606a84c239b10523e.jpg) +Figure 5. Visual comparison of samples from OHAZE and IHAZE. + +
MethodPSNR↑SSIM↑LPIPS↓VSI↑
DCP17.0050.8190.2420.945
MBFormer15.2430.6720.3390.911
C2PNet18.0140.7080.3170.929
KANet17.7130.8140.2650.939
DEANet14.6870.6910.4400.880
Diff-Plugin16.4700.5260.3640.920
OneRestore17.4580.7400.2820.924
SGDN18.1630.6820.3810.953
CUT15.9270.7130.3410.936
UNSB15.0180.6480.3920.940
YOLY14.9740.4060.4250.914
RefineDNet18.6930.7550.2600.954
D416.7670.6900.2900.939
D4+17.3360.6770.3010.947
Ours18.8290.8380.2230.961
+ +Table 2. Quantitative results on OHAZE. + +We compare our method against several state-of-the-art methods. Among them, DCP is a prior-based dehazing method. CUT and UNSB are originally proposed for unpaired image translation. For images with available ground truth (GT), we evaluate them using full-reference metrics such as PSNR, SSIM [41], LPIPS, and VSI [45]. For images without GT, we use no-reference metrics, including FID [17], NIQE [29], MUSIQ [19], and MANIQA [42]. + +# 5.1. Results on RTTS and Haze2020 + +We conduct qualitative comparisons on RTTS and Haze2020, and zoom in on local details within partial images. As shown in Figure 4, DCP effectively processes images but tends to over-enhance results due to the limitations of hand-crafted priors. C2PNet struggles with images because it is trained on synthetic datasets, which have a domain gap with real-world data. KANet shows promis + +ing performance by employing an advanced data synthesis strategy. However, KANet still falls short of fully removing haze due to the inherent limitations of paired training. Unpaired dehazing methods like YOLY, RefineDNet, and D4+ can dehaze images to some extent. However, these methods often produce sub-optimal results due to the limited transport mapping capabilities of GAN generators, resulting in artifacts and insufficient detail. In contrast, our proposed method leverages the Schrödinger Bridge to directly learn optimal transport mappings in an unpaired training framework. This approach generates more realistic and natural images with rich details, textures, and high contrast. + +Table 1 presents a quantitative comparison of Haze2020 and RTTS. Our method achieves the best quantitative results on RTTS and excels in several metrics on Haze2020. For NIQE, our approach ranks second, slightly behind KANet. This discrepancy likely results from the introduction of physical prior regularization in our method. We introduce physical prior regularization to ensure our method effectively removes haze while preserving the original image's fine details and textures. Although this regularization slightly compromised NIQE scores, we believe the overall improvement in image quality justifies this trade-off. + +# 5.2. Results on OHAZE and IHAZE + +Figure 5 and Table 2 present results on OHAZE and IHAZE. Notably, we do not retrain our model on these datasets but directly evaluate them to demonstrate the generalization capability of our method. The experimental results indicate that our method outperforms other approaches, achieving superior metrics while also producing visually pleasing dehazed images. The results for URHI and Fattal's dataset are provided in the supplementary materials. + +# 5.3. Ablation Study + +We conduct a series of ablation studies on Haze2020 and OHAZE to validate the effectiveness of each component, + +![](images/30bf65e2db007745ddf6ccc209da3592f850d936eb5ab90bc1cee82def98e5df.jpg) +(a) + +![](images/0a737c473226ef6933c9f08a5f60f056010876afe6eb0fe9098cc4a30301e477.jpg) +(b) + +![](images/f7982a1d0bfefda1b2762dc29b0dbcf7a3ecd9212034bcda421f5e4369656098.jpg) +(c) +Figure 6. Ablation study of each component. (a) Hazy image. (b) Backbone. (c) Backbone + PL. (d) Backbone + DPR. (e) Ours. + +![](images/a9aa75877705cffb6f56068aa11cac21387f8bb38200b6d47f95ac58d55667d1.jpg) +(d) + +![](images/06d1b043f77712d7e7b44174d97a681069c6a866b6c556a00c25ec731fa10cf4.jpg) +(e) + +
Backbone✓✓✓✓
PL✓✓
DPR✓✓
FID ↓80.26073.96875.19269.796
NIQE ↓3.7123.6903.9813.743
MUSIQ ↑54.41459.10956.55059.256
MANIQA ↑0.1050.1440.1330.150
PSNR ↑17.41017.44518.68718.829
SSIM ↑0.7770.7960.8390.838
LPIPS ↓0.2820.2980.2530.223
VSI ↑0.9390.9370.9520.961
+ +including the Schrödinger Bridge (SB), Prompt Learning (PL), and Detail-Preserving Regularization (DPR). Note that our method employs unpaired training, which does not utilize ground truth (GT). Therefore, PatchNCE regularization is essential to prevent gradient explosion. In the following ablation study, our backbone includes PatchNCE regularization by default. We construct the following variants: Backbone: Removing both PL and DPR. Backbone + PL: Removing DPR. Backbone + DPR: Removing PL. Ours: Our method. As shown in Table 3 and Figure 6, our proposed method achieves the best performance in both visual appeal and most metrics, striking an effective balance between perceptual quality and distortion. This validates that each component plays a critical role in improving dehazing performance. + +# 5.3.1. Ablation of Schrödinger Bridge + +We perform an ablation study to evaluate how discriminators affect the results within our framework. As illustrated in Figure 7, by employing both a Markovian Discriminator (MD) and a CLIP-based Discriminator (CD), we can simultaneously preserve the local details and ensure the global quality of the dehazed images, resulting in high-quality results. Moreover, the MD plays a crucial role in our method due to our use of PatchNCE regularization, which minimizes the distance between a patch in the dehazed image and its corresponding patch in the hazy image at the same location. Without the MD to evaluate images at the patch level, the method will fail to effectively remove haze. + +Additionally, we investigate the impact of the number of intervals used in our method, with detailed information + +![](images/f96345d77fa5a244a57118318b2d63b9a810f03569304a3d1a52b99cad433da5.jpg) +(a) +Figure 7. Ablation study of discriminators. (a) Hazy image. (b) Remove MD. (c) Remove CD. (d) Ours. + +![](images/5242840ea36e24cb35f4e57db4f85d74ccba4752b83354de7a0df3c11e6c6ada.jpg) +(b) + +![](images/6a29b583b6c165736fe64121b86e910aa012775fa8ddfa391ace6749f70237a3.jpg) +(c) + +![](images/263af9b1c061c4019faf0b20da74d689a24a1144590391692aedcb5606bcac72.jpg) +(d) + +Table 3. Ablation study of each component. + +
Backbone✓✓✓✓
PPR✓✓
HFDR✓✓
FID ↓73.96874.37273.72769.796
NIQE ↓3.6903.7533.6553.743
MUSIQ ↑59.10955.77058.52659.256
MANIQA ↑0.1440.1430.1420.150
PSNR ↑17.44518.08517.53418.829
SSIM ↑0.7960.8120.8050.838
LPIPS ↓0.2980.2900.2670.223
VSI ↑0.9370.9400.9400.961
+ +Table 4. Ablation study of detail-preservation regularization. + +provided in the supplementary materials. + +# 5.3.2. Ablation of detail-preserving regularization + +We validate the effectiveness of detail-preservation regularization, which includes physical prior regularization (PPR) and high-frequency detail regularization (HFDR). Table 4 demonstrates that both types of regularization contribute to improved performance. We also analyze the impact of varying weights of them in our supplementary materials. + +# 6. Conclusion and Discussion + +In this paper, we apply the Schrödinger Bridge in real-world image dehazing with unpaired training, named DehazeSB. By learning the optimal transport mappings from hazy to clear image distributions and enforcing detail-preserving regularization, we achieve high-quality dehazing. Furthermore, inspired by the observation that pre-trained CLIP models can accurately distinguish between hazy and clear images, we introduce a novel prompt learning to enhance dehazing. Extensive experiments demonstrate its effectiveness in producing high-quality dehazed results. + +While our Schrödinger Bridge-based approach shows significant superiority in unpaired image dehazing, it has some limitations. First, it depends on a learned haze-aware prompt, making it sensitive to the quality and diversity of the training data. If the training set lacks variability in haze conditions, the model's generalization may suffer. Second, although our detail-preserving improves fidelity in textures and structures, it struggles with highly complex or fine-grained details, especially in images with dense haze. + +# References + +[1] Cosmin Ancuti, Codruta O Ancuti, Radu Timofte, and Christophe De Vleeschouwer. I-haze: A dehazing benchmark with real hazy and haze-free indoor images. In Advanced Concepts for Intelligent Vision Systems: 19th International Conference, ACIVS 2018, Poitiers, France, September 24-27, 2018, Proceedings 19, pages 620-631. Springer, 2018. 6 +[2] Codruta O Ancuti, Cosmin Ancuti, Radu Timofte, and Christophe De Vleeschouwer. O-haze: a dehazing benchmark with real hazy and haze-free outdoor images. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 754-762, 2018. 6 +[3] Dana Berman, Shai Avidan, et al. Non-local image dehazing. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1674-1682, 2016. 1 +[4] Bolun Cai, Xiangmin Xu, Kui Jia, Chunmei Qing, and Dacheng Tao. Dehazenet: An end-to-end system for single image haze removal. IEEE transactions on image processing, 25(11):5187-5198, 2016. 2 +[5] Zeyuan Chen, Yangchao Wang, Yang Yang, and Dong Liu. Psd: Principled synthetic-to-real dehazing guided by physical priors. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7180-7189, 2021. 5 +[6] Zixuan Chen, Zewei He, and Zhe-Ming Lu. Dea-net: Single image dehazing based on detail-enhanced convolution and content-guided attention. IEEE Transactions on Image Processing, 2024. 6 +[7] Hang Dong, Jinshan Pan, Lei Xiang, Zhe Hu, Xinyi Zhang, Fei Wang, and Ming-Hsuan Yang. Multi-scale boosted dehazing network with dense feature fusion. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2157-2167, 2020. 2 +[8] Xuanzhao Dong, Vamsi Krishna Vasa, Wenhui Zhu, Peijie Qiu, Xiwen Chen, Yi Su, Yujuan Xiong, Zhangsihao Yang, Yanxi Chen, and Yalin Wang. Cunsb-rfie: Context-aware unpaired neural schrodinger bridge in retinal fundus image enhancement. In Proceedings of the Winter Conference on Applications of Computer Vision (WACV), pages 4502-4511, 2025. 3, 6 +[9] Deniz Engin, Anil Genç, and Hazim Kemal Ekenel. Cycle-dehaze: Enhanced cyclegan for single image dehazing. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 825-833, 2018. 2 +[10] Wenxuan Fang, Junkai Fan, Yu Zheng, Jiangwei Weng, Ying Tai, and Jun Li. Guided real image dehazing using ycbcr color space. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2906-2914, 2025. 6 +[11] Ali Farhadi and Joseph Redmon. Yolov3: An incremental improvement. In Computer vision and pattern recognition, pages 1-6. Springer Berlin/Heidelberg, Germany, 2018. 1 +[12] Raanan Fattal. Dehazing using color-lines. ACM transactions on graphics (TOG), 34(1):1-14, 2014. 6 +[13] Yuxin Feng, Long Ma, Xiaozhe Meng, Fan Zhou, Risheng Liu, and Zhuo Su. Advancing real-world image dehazing: + +perspective, modules, and training. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 2, 6 +[14] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139-144, 2020. 2 +[15] Yu Guo, Yuan Gao, Yuxu Lu, Huilin Zhu, Ryan Wen Liu, and Shengfeng He. Onerestore: A universal restoration framework for composite degradation. In European Conference on Computer Vision, pages 255-272. Springer, 2024. 6 +[16] Kaiming He, Jian Sun, and Xiaou Tang. Single image haze removal using dark channel prior. IEEE transactions on pattern analysis and machine intelligence, 33(12):2341-2353, 2010. 1, 2, 5, 6 +[17] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 7 +[18] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2 +[19] Junjie Ke, Qifei Wang, Yilin Wang, Peyman Milanfar, and Feng Yang. Musiq: Multi-scale image quality transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5148-5157, 2021. 7 +[20] Beomsu Kim, Gihyun Kwon, Kwanyoung Kim, and Jong Chul Ye. Unpaired image-to-image translation via neural schrödinger bridge. In The Twelfth International Conference on Learning Representations, 2024. 3, 4, 5 +[21] Nupur Kumari, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Ensembling off-the-shelf models for gan training. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10651-10662, 2022. 5 +[22] Yunwei Lan, Zhigao Cui, Chang Liu, Jialun Peng, Nian Wang, Xin Luo, and Dong Liu. Exploiting diffusion prior for real-world image dehazing with unpaired training. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 4455-4463, 2025. 2 +[23] Boyi Li, Wenqi Ren, Dengpan Fu, Dacheng Tao, Dan Feng, Wenjun Zeng, and Zhangyang Wang. Benchmarking single-image dehazing and beyond. IEEE Transactions on Image Processing, 28(1):492-505, 2018. 6 +[24] Boyun Li, Yuanbiao Gou, Shuhang Gu, Jerry Zitao Liu, Joey Tianyi Zhou, and Xi Peng. You only look yourself: Unsupervised and untrained single image dehazing neural network. International Journal of Computer Vision, 129:1754-1767, 2021. 2, 6 +[25] Zhuoyuan Li, Junqi Liao, Chuanbo Tang, Haotian Zhang, Yuqi Li, Yifan Bian, Xihua Sheng, Xinmin Feng, Yao Li, Changsheng Gao, et al. Ustc-td: A test dataset and benchmark for image and video coding in 2020s. arXiv preprint arXiv:2409.08481, 2024. 2 +[26] Zhexin Liang, Chongyi Li, Shangchen Zhou, Ruicheng Feng, and Chen Change Loy. Iterative prompt learning for unsupervised backlit image enhancement. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8094-8103, 2023. 4 + +[27] Guan-Horng Liu, Arash Vahdat, De-An Huang, Evangelos A. Theodorou, Weili Nie, and Anima Anandkumar. I²sb: Image-to-image schrödinger bridge. In Proceedings of the 40th International Conference on Machine Learning. JMLR.org, 2023. 3 +[28] Yuhao Liu, Zhanghan Ke, Fang Liu, Nanxuan Zhao, and Rynson WH Lau. Diff-plugin: Revitalizing details for diffusion-based low-level tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4197-4208, 2024. 6 +[29] Anish Mittal, Rajiv Soundararajan, and Alan C Bovik. Making a "completely blind" image quality analyzer. IEEE Signal processing letters, 20(3):209-212, 2012. 7 +[30] Igor Morawski, Kai He, Shusil Dangi, and Winston H Hsu. Unsupervised image prior via prompt learning and clip semantic guidance for low-light image enhancement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5971-5981, 2024. 4 +[31] Shree K Nayar and Srinivasa G Narasimhan. Vision in bad weather. In Proceedings of the seventh IEEE international conference on computer vision, pages 820-827. IEEE, 1999. 1 +[32] Taesung Park, Alexei A Efros, Richard Zhang, and JunYan Zhu. Contrastive learning for unpaired image-to-image translation. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part IX 16, pages 319-345. Springer, 2020. 5, 6 +[33] Yuwei Qiu, Kaihao Zhang, Chenxi Wang, Wenhan Luo, Hongdong Li, and Zhi Jin. Mb-taylorformer: Multi-branch efficient transformer expanded by taylor formula for image dehazing. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12802-12813, 2023. 2, 6 +[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning Transferable Visual Models From Natural Language Supervision. In ICML, pages 8748-8763, 2021. 2, 4 +[35] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2 +[36] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234-241. Springer, 2015. 5 +[37] Erwin Schrödinger. Sur la théorie relativiste de l'électron et l'interprétation de la mécanique quantique. In Annales de l'institut Henri Poincaré, pages 269-310, 1932. 2 +[38] Yuanjie Shao, Lerenhan Li, Wenqi Ren, Changxin Gao, and Nong Sang. Domain adaptation for image dehazing. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2808-2817, 2020. 6 + +[39] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations, 2021. 2 +[40] Yongzhen Wang, Xuefeng Yan, Fu Lee Wang, Haoran Xie, Wenhan Yang, Xiao-Ping Zhang, Jing Qin, and Mingqiang Wei. Ucl-dehaze: Towards real-world image dehazing via unsupervised contrastive learning. IEEE Transactions on Image Processing, 2024. 2 +[41] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 7 +[42] Sidi Yang, Tianhe Wu, Shuwei Shi, Shanshan Lao, Yuan Gong, Mingdeng Cao, Jiahao Wang, and Yujiu Yang. Maniaq: Multi-dimension attention network for no-reference image quality assessment. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1191-1200, 2022. 7 +[43] Yang Yang, Chaoyue Wang, Risheng Liu, Lin Zhang, Xiaojie Guo, and Dacheng Tao. Self-augmented unpaired image dehazing via density and depth decomposition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2037-2046, 2022. 2, 5, 6 +[44] Yang Yang, Chaoyue Wang, Xiaojie Guo, and Dacheng Tao. Robust unpaired image dehazing via density and depth decomposition. International Journal of Computer Vision, 132 (5):1557-1577, 2024. 2, 6 +[45] Lin Zhang, Ying Shen, and Hongyu Li. Vsi: A visual saliency-induced index for perceptual image quality assessment. IEEE Transactions on Image processing, 23(10): 4270-4281, 2014. 7 +[46] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 5 +[47] Shiyu Zhao, Lin Zhang, Ying Shen, and Yicong Zhou. Refinednet: A weakly supervised refinement framework for single image dehazing. IEEE Transactions on Image Processing, 30:3391-3404, 2021. 2, 5, 6 +[48] Yu Zheng, Jiahui Zhan, Shengfeng He, Junyu Dong, and Yong Du. Curricular contrastive regularization for physics-aware single image dehazing. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5785-5794, 2023. 2, 6 +[49] Bolei Zhou, Hang Zhao, Xavier Puig, Tete Xiao, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Semantic understanding of scenes through the ade20k dataset. International Journal of Computer Vision, 127:302-321, 2019. 6 +[50] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Proceedings of the IEEE international conference on computer vision, pages 2223-2232, 2017. 2, 5 +[51] Qingsong Zhu, Jiaming Mai, and Ling Shao. A fast single image haze removal algorithm using color attenuation prior. IEEE transactions on image processing, 24(11):3522-3533, 2015. 1 \ No newline at end of file diff --git a/ICCV/2025/When Schrodinger Bridge Meets Real-World Image Dehazing with Unpaired Training/images.zip b/ICCV/2025/When Schrodinger Bridge Meets Real-World Image Dehazing with Unpaired Training/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..540eba730922b82b4b79c043810b02b0947e13fc --- /dev/null +++ b/ICCV/2025/When Schrodinger Bridge Meets Real-World Image Dehazing with Unpaired Training/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ca50902fbe267158ffa4d996ee198bcc33391ac4b338b2e37bbe0bcba85f1485 +size 801160 diff --git a/ICCV/2025/When Schrodinger Bridge Meets Real-World Image Dehazing with Unpaired Training/layout.json b/ICCV/2025/When Schrodinger Bridge Meets Real-World Image Dehazing with Unpaired Training/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..5e3e5058cd59849cf427b41645b7dfcaa6059bcd --- /dev/null +++ b/ICCV/2025/When Schrodinger Bridge Meets Real-World Image Dehazing with Unpaired Training/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7d974275dc5e25f11fb5274d6022fa5ccb52949b064c4d2f3375c9e393a255f5 +size 499750 diff --git a/ICCV/2025/When and Where do Data Poisons Attack Textual Inversion_/3fa7061d-705e-43ad-9c51-b882bb9873d1_content_list.json b/ICCV/2025/When and Where do Data Poisons Attack Textual Inversion_/3fa7061d-705e-43ad-9c51-b882bb9873d1_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..9d3ba29725b7d136cada56208c97e626c7b25fe4 --- /dev/null +++ b/ICCV/2025/When and Where do Data Poisons Attack Textual Inversion_/3fa7061d-705e-43ad-9c51-b882bb9873d1_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:732482f9622b9c8020fc9c4f34cef586e0d1dfd8c781e61e7390393d367bdff9 +size 85508 diff --git a/ICCV/2025/When and Where do Data Poisons Attack Textual Inversion_/3fa7061d-705e-43ad-9c51-b882bb9873d1_model.json b/ICCV/2025/When and Where do Data Poisons Attack Textual Inversion_/3fa7061d-705e-43ad-9c51-b882bb9873d1_model.json new file mode 100644 index 0000000000000000000000000000000000000000..e7419415857e135db48cd2be6089ff5bed4c88ab --- /dev/null +++ b/ICCV/2025/When and Where do Data Poisons Attack Textual Inversion_/3fa7061d-705e-43ad-9c51-b882bb9873d1_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b8b801c54a61235dd8aa38fa089a0df1dfd994520e68f289ce32cc7edf141619 +size 111830 diff --git a/ICCV/2025/When and Where do Data Poisons Attack Textual Inversion_/3fa7061d-705e-43ad-9c51-b882bb9873d1_origin.pdf b/ICCV/2025/When and Where do Data Poisons Attack Textual Inversion_/3fa7061d-705e-43ad-9c51-b882bb9873d1_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..1b70870f9d65bb6e081b53796d7cd946504d64c6 --- /dev/null +++ b/ICCV/2025/When and Where do Data Poisons Attack Textual Inversion_/3fa7061d-705e-43ad-9c51-b882bb9873d1_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ac906406492a180b2fdc4c79627aa934db3dd070cd4e59b568c73be2261f7fdc +size 9672755 diff --git a/ICCV/2025/When and Where do Data Poisons Attack Textual Inversion_/full.md b/ICCV/2025/When and Where do Data Poisons Attack Textual Inversion_/full.md new file mode 100644 index 0000000000000000000000000000000000000000..e7c2bbe32de425dd63d03d7cc98a05ed5d7de7ef --- /dev/null +++ b/ICCV/2025/When and Where do Data Poisons Attack Textual Inversion_/full.md @@ -0,0 +1,365 @@ +# When and Where do Data Poisons Attack Textual Inversion? + +Jeremy Styborski $^{1*}$ Mingzhi Lyu $^{2*}$ Jiayou Lu $^{1}$ Nupur Kapur $^{1}$ Adams Wai-Kin Kong $^{1}$ \ + $^{1}$ College of Computing and Data Science, Nanyang Technological University, Singapore\ + $^{2}$ Rapid-Rich Object Search (ROSE) Lab, Nanyang Technological University, Singapore + +{styb0001,lyum0002,jiayou001,nupur003,AdamsKong}@ntu.edu.sg + +# Abstract + +Poisoning attacks pose significant challenges to the robustness of diffusion models (DMs). In this paper, we systematically analyze when and where poisoning attacks textual inversion (TI), a widely used personalization technique for DMs. We first introduce Semantic Sensitivity Maps, a novel method for visualizing the influence of poisoning on text embeddings. Second, we identify and experimentally verify that DMs exhibit non-uniform learning behavior across timesteps, focusing on lower-noise samples. Poisoning attacks inherit this bias and inject adversarial signals predominantly at lower timesteps. Lastly, we observe that adversarial signals distract learning away from relevant concept regions within training data, corrupting the TI process. Based on these insights, we propose Safe-Zone Training (SZT), a novel defense mechanism comprised of 3 key components: (1) JPEG compression to weaken high-frequency poison signals, (2) restriction to high timesteps during TI training to avoid adversarial signals at lower timesteps, and (3) loss masking to constrain learning to relevant regions. Extensive experiments across multiple poisoning methods demonstrate that SZT greatly enhances the robustness of TI against all poisoning attacks, improving generative quality beyond prior published defenses. + +Code: www=qgithub.com/JStyborski/Diff_Lab + +Data: www.github.com/JStyborski/NC10 + +# 1. Introduction + +The image quality and prompt fidelity offered by diffusion-based image generation models such as DALL-E 2 [57], Stable Diffusion [58], and Imagen [60] have popularized (or vilified) the use of AI-generated images in the art, marketing, and media industries. The subsequent proliferation of personalization and editing methods [15, 20, 25, 35, 59, 82] and tools [2, 7, 72] allows anyone, even non-experts and non-artists, to quickly retrain existing models to generate their desired images, including images of novel concepts + +![](images/dc3544503a7257cc7bf685742d3ec8699f7956ec91dd09dd69f8438acb52f1b2.jpg) +Figure 1. Applying SZT to TI mitigates poison signals and recovers shape, color, and texture of desired concepts. Novel Concepts10 training data poisoned by ADM+. Generation prompt: "A $R*$ with a wheat field in the background" + +not found in the model training data. + +Unfortunately, the surge in generative AI corresponds to misfortune for artists and copyright owners, who may see their works easily reproduced by retrained generative models. Consequently, the subfield of data poisoning for generative AI has focused on how to inject images with signals that nullify their use in training datasets for generative models. Individuals looking to protect their data are encouraged to "poison" their images with imperceptible adversarial perturbations such that generative models will fail to learn to recreate the novel concepts contained within. + +The concept of data poisoning stems from adversarial examples (AEs) [17, 36, 49, 70], where the inputs are slightly perturbed in order to maximize loss. Intuitively, these AEs learn to incorporate signals that maximally mislead AI models into making incorrect predictions. The AE method is likewise effective against generative models; multiple state-of-the-art poisoning methods for generative models apply the AE method to the diffusion model (DM) objective to inject poison signals into images [6, 39, 61, 78]. Although the textual inversion (TI) technique is widely used to adapt DMs to novel concepts, there are no systematic studies that identify when and where AEs attack the TI process. + +# 1.1. Research Gap and Contributions + +State-of-the-art poisoning methods for DMs are largely cannibalizing previous adversarial methods (e.g., Glaze [63] from Adv-VAE [71] or AdvDM [39] from PGD [36, 49]) or creating new loss functions that nonetheless follow the same AE algorithm (e.g., SDS [78] or DiffusionGuard [6]). Although these poisoning methods are useful research milestones that explore novel applications for AEs, they neglect to study how poison signals are learned and which DM vulnerabilities are being exploited. Furthermore, many poisoning works neglect to understand the intricacies of the downstream tasks, such as TI, for which the AEs are designed. + +In this paper, we examine the learning behavior of TI [15], a lightweight personalization technique for adapting DMs to novel concepts, when trained on poisoned datasets. To enable our analysis, we devise a novel visualization method, semantic sensitivity maps (SSM), to display loss sensitivity to individual text tokens throughout training. Using SSMs, we observe several heretofore unseen trends in poison learning: (1) DMs concentrate poison learning at lower-middle timesteps (approximately $t \in [0,600]$ ) and (2) adversarial signals tend to "distract" learning away from the novel concept of interest and towards irrelevant regions. + +We scrutinize the noising process and the noise-prediction training objective that are common to DMs and the TI process to demonstrate that their combined effect inherently favors low-noise images. Crucially, this learning bias towards lower-middle timesteps is inherent in most AE algorithms for DMs, which implies that the AEs themselves are mostly learned at lower-middle timesteps. This exposes a weakness where many poisoning techniques are less effective at higher timesteps. In experiments, we demonstrate that restricting TI training to higher timesteps is a simple yet effective defense against multiple poisons. + +We find that TI prioritizes learning in regions that contain concept objects rather than background objects. This behavior stems from the tendency to minimize noise-prediction loss by focusing on common features across training images. Poisoning disrupts this process by introducing dominant adversarial signals and suppressing the learnability of clean concept information, misleading the TI training process. To counteract this, we propose a loss-masking strategy during optimization, explicitly training the model to focus on the concept regions within poisoned images. This ensures that TI effectively captures the intended concept and mitigates the influence of adversarial signals. + +We also analyze JPEG compression, a common defense technique in poisoning literature. We find that JPEG compression helps purify AEs by reshaping the distributions of adversarial signal magnitudes and frequencies to benign distributions, similar to those of clean images. Although JPEG compression alone cannot enable TI to learn from poisoned images, using it as a first step to weaken + +high-frequency adversarial signals significantly improves the performance of other defense methods. + +Finally, we combine our analyses to present safe-zone training (SZT), a simple and lightweight defense method for training TI. SZT consists of three main components: (1) JPEG compression to weaken high-frequency adversarial signals, (2) training at high timesteps to avoid adversarial signals at lower-middle timesteps, and (3) loss masking to prioritize learning on the novel concept. We show that SZT outperforms multiple established defenses across six different poisons in accurately generating novel concepts and reaches the performance of TI on clean data. + +Our contributions are summarized below: + +- We devise SSMs and use them to visualize loss attribution during the TI training process. +- We analyze the noising process and noise-prediction objective to understand the mechanism by which DMs, AEs, and TI focus on lower-middle timesteps. +- We develop a loss-masking strategy to encourage DMs to learn from the concept regions in the training data. +- We combine our findings and establish SZT to methodically improve the poison robustness of TI. + +# 2. Related Works + +We assume that the reader is familiar with standard DMs [24, 32, 67, 68] and generation techniques [12, 23, 50, 66]. We mainly focus on latent diffusion models (LDMs) [58]. + +Although DMs can easily generate most concepts from their training dataset with high quality, they predictably suffer in generation quality for novel (unseen) concepts. Editing techniques [3, 20, 29, 50-52, 82] apply modifications to an input image while enforcing structural or composition constraints, permitting generation with novel concept images. However, editing techniques are still bound to the prior from the DM's training data. In order to "inject" new concepts and styles into the DM, personalization techniques [15, 18, 19, 25, 35, 53, 59, 73, 74, 83] quickly finetune some component of the DM on a small dataset of novel concept images. Many personalization techniques learn new text embeddings [15, 18, 35, 73, 74] or co-opt existing tokens [19, 59] to capture concept semantics. Without retraining the entire DM from scratch, personalization techniques are the best way to adapt DMs to new concepts. In particular, TI [15] is a lightweight personalization method that requires training only one new text embedding and is often included within other methods [18, 35, 73]. + +AEs [5, 17, 36, 49, 70] expose the brittleness of neural net predictions by learning minor perturbations to maximize loss. Although originally designed for classification models, AEs have also been applied to object detection and segmentation [76], reinforcement learning [27], watermarking [48], self-supervised learning [4, 28, 30], VAEs [34, 71], and diffusion models [6, 37, 39, 61, 63, 64, 78]. Due to their + +ability to defeat deep learning models, AEs have become a landmark technique for data protection. The concept of AEs extends to adversarial poisoning [14, 26], where entire datasets are poisoned to nullify attempts to scrape or steal data. Typical defenses against AEs include simple detection [1], augmentation [8, 9, 11, 45, 80, 81], which modifies input images to increase sample diversity and dilute poison signals, multitask architectures [69, 79], which use multiple networks to mitigate shortcut learning [16, 47], and adversarial training, which simply adds AEs to the training set. + +AEs for DMs fall into two main categories. Encoder attacks [61, 63, 64] utilize the encoder of an LDM to learn a small perturbation to an input image that minimizes the distance between the perturbed image encoding and the encoding of some unrelated target image. Diffuser attacks [6, 37, 39, 61, 78] aim to learn a small perturbation to an input image that maximizes the DM loss (often noise-prediction). Other attack methods combine encoder attacks and diffuser attacks [38, 61]. Defenses against AEs for DMs include augmentation [65] and regeneration [13, 44, 77, 84], where poisoned images are randomly noised and then denoised with a DM to remove adversarial signals. Regeneration methods are the most actively researched, but they are time-consuming and complex as they require preprocessing entire datasets with a DM in order to "purify" them. + +# 3. Preliminaries + +# 3.1. Latent Diffusion Models + +We mainly focus on LDMs [58] as they are quick to train, they generate high-quality outputs, and tools for using/modifying them are widely available. In LDMs, an input image $x$ is first encoded to a latent of lower dimension via a variational encoder, $z_0 = \mathcal{E}(x)^1$ . The latent is then iteratively noised to timestep $t$ according to noise variance schedule $\beta_t$ . An equivalent one-step noising process is + +$$ +z _ {t} = \sqrt {\bar {\alpha} _ {t}} z _ {0} + \sqrt {1 - \bar {\alpha} _ {t}} \epsilon , \tag {1} +$$ + +where $\overline{\alpha}_t = \Sigma_{s=1}^t (1 - \beta_s)$ and $\epsilon \sim \mathcal{N}(\mathbf{0}, I)$ . $\overline{\alpha}_t$ monotonically decreases from 1 to 0 as $t$ increases from 0 to $T$ , following a variance-preserving schedule [68]. A generative DM learns the denoising process, where the goal is to maximize the likelihood of generating the unnoised input $(z_0)$ . DDPM [24] showed that this objective can be practically realized with a simple noise-prediction task. The model predicts the Gaussian noise $\epsilon$ added to input $z_0$ when given noisy sample $z_t$ , and the learning objective is given as + +$$ +\underset {\theta} {\arg \min } E _ {x, t \sim \mathcal {U} (0, T), \epsilon} \left[ \left| \left| \epsilon_ {\theta} \left(z _ {t}, t\right) - \epsilon \right| \right| _ {2} ^ {2} \right], \tag {2} +$$ + +where $x$ is drawn from the training dataset, $\mathcal{U}(0,T)$ is the uniform distribution between timesteps 0 and $T$ , and $\epsilon_{\theta}$ is + +a model that predicts a noise vector $\epsilon$ given a noised latent $z_{t}$ . For text-conditional models, $\epsilon_{\theta}(z_t,t,\tau_\theta (y))$ also accepts features vector $\tau_{\theta}(y)$ , where $\tau_{\theta}$ is a language model and input prompt $y$ is provided by the user or training data. + +# 3.2. Textual Inversion + +In this paper, we consider how AEs affect the downstream personalization task of TI [15], which seeks to learn a new text token that captures the semantics of a novel concept. TI trains a new text embedding corresponding to a new token $R^{*}$ ; the new embedding is the only finetuned parameter. TI trains via a noise-prediction task as + +$$ +\underset {\theta_ {R ^ {*}}} {\arg \min } E _ {x _ {f}, t \sim \mathcal {U} (0, T), y ^ {*}, \epsilon} \left[ | | \epsilon_ {\theta} \left(z _ {t}, t, \tau_ {\theta} \left(y ^ {*}\right)\right) - \epsilon | | _ {2} ^ {2} \right], \tag {3} +$$ + +where sample $x_{f}$ is drawn from the finetuning dataset and $y^{*}$ is a prompt containing token $R^{*}$ . + +# 3.3. Adversarial Examples + +The primary method to create AEs against LDMs is by attacking the diffuser with an adversarial perturbation $\delta$ injected into the input data. The adversarial objective is to find a $\delta$ that maximizes the noise-prediction loss, as + +$$ +\begin{array}{l} \underset {\delta} {\arg \max } E _ {x _ {p}, t \sim \mathcal {U} (0, T), \epsilon} \left[ | | \epsilon_ {\theta} \left(z _ {t} ^ {\prime}, t\right) - \epsilon | | _ {2} ^ {2} \right] \tag {4} \\ \begin{array}{l l} \text {s . t .} & | | \delta | | _ {b} \leq \kappa , \end{array} \\ \end{array} +$$ + +where sample $x_{p}$ is drawn from the dataset to poison and $z_{t}^{\prime} = \sqrt{\overline{\alpha}_{t}}\mathcal{E}(x_{p} + \delta) + \sqrt{1 - \overline{\alpha}_{t}}\epsilon$ is the adversarial latent following Eq. 1. The perturbation $\delta$ is commonly constrained in the $l^{\infty}$ norm to a small value $\kappa$ to limit its visibility. + +Encoder attacks against LDMs seek to perturb an image such that the encoding of the poisoned image $x_{p} + \delta$ is similar to that of some unrelated target image $x_{g}$ . Encoder attacks optimize the objective function + +$$ +\underset {\delta} {\arg \min } E _ {x _ {p}} \left[ \left| \left| \mathcal {E} \left(x _ {p} + \delta\right) - \mathcal {E} \left(x _ {g}\right) \right| \right| _ {2} ^ {2} \right] \text {s . t .} \| \delta \| _ {b} \leq \kappa . \tag {5} +$$ + +# 4. Analysis + +# 4.1. Visualizing the Effect of Adversarial Examples + +To observe the influence of AEs against TI, it is critical to visualize the impact of each text embedding on a generated image. Let $\widehat{e} = \{e_0,\dots e_n,\dots ,e_{L - 1}\}$ , where $e_0$ represents the start token embedding vector and $e_n$ represents the embedding vector of the nth token. For a text input consisting of $l$ tokens, $\widehat{e}$ is padded to length $L$ with $\{e_{l + 1},\ldots ,e_{L - 1}\}$ with end token embeddings. Each $e_n$ is obtained via lookup in a token-embedding dictionary. As an example, given the text prompt, "a puppy wearing a hat", embeddings $\{e_1,\dots ,e_5\}$ correspond to "a", "puppy", "wearing", "a", and "hat" respectively. + +![](images/2f21851cd1cbec62e47c0fd6150510bd96ff53a8dfdbf2ac0530bb4f6aa1d63d.jpg) +Figure 2. Cross attention maps (a) and SSMs (b) for tokens “”, “puppy”, and “hat” with text prompt “a puppy wearing a hat” at noise timestep $t = 500$ . + +A popular text-attribute method for LDMs visualizes the cross-attention maps [41] in $\epsilon_{\theta}$ . However, cross-attention maps are incapable of capturing the effect of each embedding $e_n$ . This is because the self-attention layers in the text encoder $\tau_{\theta}$ entangle signals from the embedding vectors of $\hat{e}$ before they reach the cross-attention layers of $\epsilon_{\theta}$ . Fig. 2(a) visualizes text attribution using cross-attention maps. The cross-attention map corresponding to the start token $e_0$ is bright, highlighting many regions, but the maps corresponding to words "puppy" and "hat" are dim and blurry. This drawback has been noted in other works [46]. + +To accurately identify the spatial regions over which a specific token has influence, we developed a new visualization method called semantic sensitivity maps (SSMs). We define SSMs by + +$$ +\begin{array}{l} S S M (x, t, \widehat {e}, n) = E _ {\widehat {e} _ {\Delta n}} \left[ \left\| \epsilon_ {\theta} \left(z _ {t}, t, c\right) - \epsilon_ {\theta} \left(z _ {t}, t, c _ {\Delta n}\right) \right\| _ {c h} ^ {2} \right] \\ \mathrm {s . t .} c = \tau_ {\theta} (\hat {e}), c _ {\Delta n} = \tau_ {\theta} (\hat {e} _ {\Delta n}), \tag {6} \\ \end{array} +$$ + +where $x$ is an input image and $c$ is the condition tensor. $c_{\Delta n}$ is analogous to $c$ , except that it is derived from $\widehat{e}_{\Delta n}$ , a modified version of $\widehat{e}$ that replaces $e_n$ with a randomly sampled vector from the embedding dictionary. For notational simplicity, we denote that $\tau_{\theta}(\widehat{e})$ accepts token embeddings as input instead of the prompt. $||\cdot||_{ch}^2$ indicates sum of squares across channels such that the output is a one-channel image that matches the input height and width. Fig. 2(b) displays the SSMs corresponding to the embeddings for "puppy" and "hat". Compared to cross-attention attribution maps in Fig. 2(a), SSMs more accurately reflect the regions corresponding to each input embedding. Crucially, SSMs avoid attributing semantic tokens to the start token. + +Using SSMs, we can investigate when and where adversarial signals attack the TI process. Fig. 3 shows the SSMs for a clean image and its corresponding AEs for various noisy timesteps. The input images contain a novel concept not contained in the training dataset. The image prompt is "a $R^*$ laying on top of a grass covered field". Fig. 3 shows the SSMs for the trainable text embedding corresponding to $R^*$ before and after TI training. We observe + +![](images/8a416ab27f96da648802b0937e303b1b848914c81fa23fe4d0d6f3a8198d619b.jpg) +Figure 3. Comparison between SSMs of clean and AE samples at different noise timesteps (a) before TI training and (b) after 500 steps of TI training. + +that before TI training, SSMs of the clean sample and AEs highlight the novel object in the range $t \in [300, 700]$ . This indicates that the DM prior can already correlate unknown objects with unknown tokens. After TI training, the novel concept is more pronounced in SSMs of clean images, but the SSMs of AEs become extremely noisy, particularly at lower-middle timestep values. Notably, the attribution for AEs extends far outside the novel concept region. To explain this behavior, we investigate the temporal and spatial learning properties of TI in subsequent sections. + +# 4.2. Timestep Learning Bias + +Fig. 3 shows that concept learning is focused on lower-middle timesteps ( $t < 700$ ) and that poison signals have the greatest impact at lower timesteps. To understand this phenomenon, we revisit the noising and training processes. + +We begin by noting that all training objectives (i.e., model training in Eq. 2, TI in Eq. 3, or AEs in Eq. 4) utilize the expected noise-prediction loss over a uniform $t$ distribution. The only differences between the objectives are the training datasets and the trainable parameters. + +Recalling that $\overline{\alpha}_t$ from Eq. 1 monotonically decreases from 1 at $t = 0$ to 0 at $t = T$ , we can see that the noising process is interpolating between input $z_0$ and noise $\epsilon$ . At $t = 0$ , $z_{t=0} = \sqrt{\overline{\alpha}_0} z_0 + \sqrt{1 - \overline{\alpha}_0} \epsilon = z_0$ . Therefore, at $t = 0$ , the input to the noise-prediction model $\epsilon_\theta(z_0, 0)$ contains no information about the sampled noise $\epsilon$ . Since $\epsilon$ is distributed as $\mathcal{N}(\mathbf{0}, I)$ , the optimal noise prediction to minimize the expected loss at $t = 0$ is $\epsilon_\theta(z_0, 0) = \mathbf{0}$ . Predictably, the loss at $t = 0$ is a maximum because the noise-prediction model cannot accurately predict the ground-truth noise. Conversely, at $t = T$ , $z_{t=T} = \sqrt{\overline{\alpha}_T} z_0 + \sqrt{1 - \overline{\alpha}_T} \epsilon = \epsilon$ , and the + +![](images/4b5a1e849fef531d7bf2388cd9548d7cf8c9649d7e8c5426b17f437dbc130f64.jpg) +Figure 4. Loss distributions for Stable Diffusion 1.5 on 1000 clean (green) and 1000 ADM+ poisoned (red) LAION Aesthetic images at each of 21 interspersed timesteps. The dark lines represent median loss values. The dark and light bands encompass $50\%$ and $95\%$ , respectively, of the loss distributions. + +noise-prediction model simply needs to return the input to minimize loss. At $t = T$ , the loss approaches a minimum. + +To validate our argument, we calculate the noise-prediction loss of Stable Diffusion 1.5 for 1000 images sampled from the LAION Aesthetic [62] dataset at 21 equally-spaced timesteps (21000 samples total) and plot the resulting distributions in Fig. 4. As predicted, the expected loss decreases from a maximum at $t = 0$ to 0 at $t = T$ . Applying $\mathrm{ADM}+$ [39] poisoning to the images increases the loss values in the lower-middle timesteps (as expected by a loss-maximizing AE algorithm), but the losses at $t = 0$ and at high timesteps remain unchanged. + +We must also consider how the noise-prediction objective affects gradients during TI training. Fig. 5 displays the average loss gradient magnitude with respect to the trainable text embedding at each timestep after 1000 steps of TI training with Stable Diffusion 1.5. Curves are obtained by calculating 1000 losses for each concept of the NovelConcepts10 dataset across 21 equally-spaced timesteps and then averaging gradient magnitudes at each timestep across concepts. As expected, the graph shows that at high timesteps, gradients decay to 0 as the loss distributions converge to 0. At lower-middle timesteps, moderate loss allows the model to backpropagate real (or adversarial) signals to the text embedding, corresponding to high gradients. However, near $t = 0$ , high losses correspond to low gradients. Intuitively, this is because the noise-prediction model minimizes loss at $t = 0$ by predicting near- $\mathbf{0}$ noise regardless of the inputs, $z_{t}$ and $\tau_{\theta}(y^{*})$ . Since the prediction is independent of the inputs, no loss signal is backpropagated to the trainable text embedding. The loss gradient magnitudes in Fig. 5 correspond to the intensities of the sensitivity maps in Fig. 3. + +This behavior is particularly applicable to diffusion-based AEs and TI. All diffusion-based AEs seek to maximize the expected noise-prediction loss (Eq. 2), which includes an expectation over a uniform timestep range + +![](images/2ae116e5350a4c6e9b18a6480ba1e9fdc147493c5e47e5f30210a2920bee0649.jpg) +Figure 5. Trainable text embedding loss gradients after 1000 steps of TI training on clean and ADM+ poisoned NovelConcepts10 images. Results at each timestep are averaged across all concepts. The dark lines represent median gradient values. The dark bands encompass $50\%$ of the gradient distributions. + +$\mathcal{U}(0,T)$ . As the loss gradient is near $\mathbf{0}$ at high timesteps, most of the adversarial signals will be learned at lower-middle timesteps. Likewise, TI applies an expectation over a uniform timestep range (Eq. 3) thus focusing learning at lower-middle timesteps. Assuming that the adversarial signals contained within the training images are biased towards lower-middle timesteps, then we can bias the TI task towards high timesteps to avoid adversarial signals and learn clean concept features. + +In Appendix B, we give a detailed derivation of the noise-prediction error along with additional plots for the loss and noise-prediction distributions. In Appendix C, we show the resulting effect on loss gradients throughout TI training and across multiple poisons. + +# 4.3. Spatial Learning Bias + +As observed in Sec. 4.1, AEs disrupt TI by "distracting" the optimization process and expanding the spatial distribution of the trainable token attribution beyond the novel concept region. To understand this phenomenon, we revisit the adversarial objectives in Eq. 4 and Eq. 5. + +We begin by noting that both adversarial objectives involve an $l^2$ norm which includes a sum over all spatial coordinates (height, width, channels) of the error tensor. This summation does not apply preferential weight to any spatial region. Furthermore, the $l^2$ vector norm is convex and non-decreasing, implying that it increases as the magnitudes of individual elements increase. This encourages all spatial elements to contribute to optimizing the $l^2$ objective. This incentive combines with the adversarial perturbation constraint $||\delta||_b$ , which is commonly implemented with the $l^\infty$ norm. We note that $||\delta||_\infty$ is non-increasing for all infinitesimal changes in perturbation elements $\delta_i$ where $|\delta_i| < ||\delta||_\infty$ ; this permits increases in most perturbation elements without affecting the constraint. In summary, the $l^2$ norm objectives in Eq. 4 and Eq. 5 encourage all elements + +![](images/d3e619302393f75980d5417ce3d9e70dcb33624ad9619da19ce1d33e4b3cce63.jpg) +Figure 6. Loss difference inside ("LossIn") and outside ("LossOut") novel concept regions during TI training on clean, $\mathrm{ADM}+$ , $\mathrm{SDS}+$ , and EA images of seven concepts from NovelConcepts10. + +of the error terms to contribute and the $l^{\infty}$ constraint on $\delta$ encourages all elements of $\delta$ to contribute maximally. Practically, this implies that adversarial perturbations maximally infect every region of a poisoned image. + +To empirically verify this phenomenon, we study the losses attributed to novel concept regions for clean images and AEs during TI training. We apply TI to clean, ADM+, SDS+, and EA versions of the NovelConcepts10 dataset. We track the loss contributions from inside and outside the novel concept regions throughout training. We first average losses across their contributing pixels and then average the results across all concepts. The results in Fig. 6 indicate that TI for clean images can effectively minimize the loss inside the novel concept regions and ignores signals outside these regions. This implies that TI effectively learns new concepts and ignores background information. However, when trained on AEs, TI losses inside and outside of the concept regions are reduced by the same proportion. This indicates that AEs mislead TI to learn equally from all regions and ignore the semantic significance of the novel concept. This aligns with the observation from Fig. 3 that all regions of an adversarial input are attributed to the trainable token. + +# 4.4. JPEG as Poison Defense + +The adversarial signals injected into images by poisoning techniques tend towards higher frequencies in pixel space. Consequently, JPEG compression, which relies on discrete cosine transforms to identify and remove high-frequency signals, has been noted as an effective poison defense [45]. We investigate the mechanism by which JPEG eliminates adversarial signals and identify two crucial behaviors: (1) JPEG compression converts bimodal poison noise into unimodal Gaussian-like noise in pixel space and (2) JPEG centralizes the frequency spectrum in latent space, forcing poisoned latent frequencies towards those of clean latents. + +We empirically verify (1) in Fig. 7 by examining the dis + +![](images/497dafe12ca8aff050c106ef7861593356da4424a5818855ffa20079e63961c3.jpg) +Figure 7. Histograms of pixel-space perturbations (relative to clean images) for ADM+ poisoned images without (red) and with (green) JPEG compression. + +distribution of poison perturbations (relative to clean images) with and without applying JPEG compression. We use a JPEG quality of 25 and aggregate results across all 50 images in the NovelConcepts10 dataset. The bimodal distribution of poison perturbations aligns with loss objective and constraint discussion from Sec. 4.3. After JPEG compression, the poison perturbations follow a unimodal Gaussian-like distribution. We note in Fig. 12 of Appendix D that this behavior holds for all poisons studied. Intuitively, the carefully-optimized poison signals have been replaced by a distribution of random errors due to JPEG compression. + +Due to space limitations, we move empirical verification of (2) and subsequent analysis to Appendix D. We note that the effect of (2) is that JPEG-compressed poisoned images force poison latents to conform to the power spectra of natural image latents, as would be expected by LDMs. + +# 4.5. Safe-Zone Training + +In the above analyses, we have demonstrated when and where AEs poison TI. Combining takeaways from our analyses, we propose Safe-Zone Training (SZT) to mitigate the influence of adversarial signals in poisoned images. Applying lessons from Sec. 4.4, we first apply JPEG compression to obfuscate the adversarial signals in the pixel domain and centralize the power spectra in the latent domain. Applying the conclusions from Sec. 4.2, we restrict training to high timesteps in order to avoid adversarial signals at lower-middle timesteps. Finally, SZT incorporates the behaviors noted in Sec. 4.3 and utilizes a binary loss mask to allow only signals from the novel concept region to backpropagate to the trainable text embedding. In summary, SZT is a simple algorithm that involves applying JPEG compression, adjusting the timestep sampling range, and inserting a mask during loss calculation. Crucially, SZT does not require additional networks to regenerate or purify poison data. The SZT objective can be formulated as + +$$ +\underset {\theta_ {R ^ {*}}} {\arg \min } E _ {x _ {j} ^ {\prime}, t, y ^ {*}, \epsilon} \left[ | | \left(\epsilon_ {\theta} \left(z _ {t} ^ {\prime}, t, \tau_ {\theta} \left(y ^ {*}\right)\right) - \epsilon\right) \odot M _ {z} \| _ {2} ^ {2} \right] \tag {7} +$$ + +$$ +\mathrm {s . t .} x _ {j} ^ {\prime} = J P E G (x _ {p} ^ {\prime}), t \sim \mathcal {U} (t _ {t h}, T), M _ {z} = R e s i z e (M _ {x}), +$$ + +where $x_{j}^{\prime}$ is the JPEG-compressed version of poisoned image $x_{p}^{\prime} = x_{p} + \delta$ , $z_{t}^{\prime} = \sqrt{\overline{\alpha}_{t}}\mathcal{E}(x_{j}^{\prime}) + \sqrt{1 - \overline{\alpha}_{t}}\epsilon$ , timestep $t$ is uniformly sampled from the range above threshold $t_{th}$ , $M_{z}$ is the latent-space mask resized from the pixel-space mask of the novel concept region $M_{x}$ , and $\odot$ denotes element-wise multiplication. + +# 5. Experiments + +# 5.1. Experiment Settings + +# 5.1.1. Datasets + +We test the ability of TI to learn new concepts from poisoned images. To ensure that the LDM is learning new concepts not contained within the prior, we collected NovelConcepts10, a dataset of five images for each of 10 distinct objects (50 total images). An extended dataset description and example images are given in Appendix E. We also evaluate on CustomConcept101 [35], which was curated to study personalization tasks. For each image in a dataset, we create a binary mask of the novel concept region using the Segment Anything Model [33]. We utilize six data poisons: ADM+ [39], ADM-, SDS+ [78], SDS- [78], EA [61], and DA [61]. We place poisoning details in Appendix F. + +# 5.1.2. Model and Training + +We conduct all experiments with Stable Diffusion v1.5 [58]. We utilize a modified Diffusers library from HuggingFace [72] to perform TI. For each concept of a dataset, we run TI [15] with a default learning rate of $5e^{-4}$ (no decay, no warmup) for 5000 training steps and a batch size of 1. For CustomConcept101, we utilize 2500 training steps without noticeable loss of performance. For each learned concept, we generate five images for each of the 24 object prompts from the DreamBooth pipeline [59] (120 images total). + +# 5.1.3. Evaluation Metrics + +For each learned concept, we measure the DINOv2 encoding similarity [54] and FID score [22] between training images and the images generated after TI. For poisoned datasets, we calculate metrics between generated images and clean versions of the training images. We also measure prompt fidelity with CLIP Score [21, 56] between the generation prompt and the generated images. We report metrics first averaged across all generated images per concept and then across all concepts in a dataset. + +# 5.2. Results and Analysis + +# 5.2.1. Personalization Settings Ablations + +We begin by abulating training hyperparameters for TI across clean, $\mathrm{ADM + }$ , ADM-, and EA versions of NovelConcepts10 for learning rate in $(5e^{-5}, 5e^{-4}, 5e^{-3})$ , learning rate schedule in (constant, linear decay, cosine decay), and training steps in (1000, 5000, 10000). Results and further discussion are given in Appendix G. In summary, we + +find that maintaining default settings (constant $5e^{-4}$ learning rate with 5000 training steps) is sufficient to learn novel concepts without overfitting to the training images. + +# 5.2.2. Timestep Range Ablation + +Based on findings about timestep learning bias in Sec. 4.2, we ablate methods to restrict timesteps to high values during TI training. We consider multiple sampling methods to favor high timesteps: thresholding, power distributions, and tanh distributions. Fig. 15 in Appendix H displays the probability curves for the power and tanh distributions. We ablate $t$ -sampling methods for clean, ADM+, ADM-, and EA versions of NovelConcepts10. Full results are given in Appendix H. We find that simple high-thresholding works best, improving DINOv2 similarity for ADM+ data from 0.10 ( $t \sim \mathcal{U}(0,1000)$ ) to 0.21 for $t \geq 600$ and 0.28 for $t \geq 700$ . Using $t \geq 600$ also maintains or improves scores for clean and EA datasets. In our subsequent proposed defenses, we mainly utilize $t \geq 600$ , denoted "T600". + +# 5.2.3. Masking Ablation + +To verify the effectiveness of loss masking (LM), we compare it to various masking strategies: input image masking (IM), latent $(z_{t})$ masking (ZM), and masking both input image and loss (LIM). We ablate the masking method on all poisoned versions of NovelConcepts10. Results in Appendix I demonstrate that LM outperforms all other masking strategies across all poisoning methods, except EA. We hypothesize that applying masks directly to the input image or latent $z_{t}$ may destroy useful information before the features are processed by the U-Net. Intuitively, masked inputs lack contextual information about the relationship between the novel concept and the background, distorting the U-Net predictions. By applying the mask to the final loss, we preserve background information and permit a more coherent understanding of the scene. + +In Appendix J, we experiment with dilating the mask size to incorporate additional context in loss backpropagation. We find that relatively small amounts of mask dilation (i.e., 16 pixels for $512 \times 512$ masks) can improve LM for TI. + +# 5.2.4. Safe-Zone Training Performance + +Our implementation of SZT is highly customizable, empowering the user to control JPEG compression quality, timestep restriction method/shape, and loss mask size/shape. In our ablations, we find that combining JPEG compression (quality = 25), timestep restriction $t \geq 500$ and LM with 16 pixels of mask dilation performs best. We defer further discussion on SZT configurations to Appendix J and encourage users to explore additional settings. + +We demonstrate the robustness of SZT across poisoned versions of the CustomConcept101 dataset. Table 1 compares the DINOv2 similarity performance of SZT and its novel components with established defenses, including Regen [84], PDMPure [77], AdvClean [65], and JPEG com + +Table 1. CustomConcept101 DINOv2 Similarity for various poison defenses. + +
Defense → +Poison ↓NominalRegen +[84]PDMPure +[77]AdvClean +[65]JPEGT600LMJPEG ++T600JPEG ++LMSZT
Clean0.460.460.250.450.450.500.470.470.460.48
ADM+[39]0.100.170.230.150.360.250.300.410.440.45
ADM-0.330.250.230.350.410.270.430.450.450.46
SDS+[78]0.100.160.250.160.340.230.320.400.450.44
SDS-[78]0.310.220.230.350.400.230.420.450.450.47
EA[61]0.100.090.240.190.330.120.290.440.440.46
DA[61]0.290.250.270.350.390.220.420.440.460.47
Psn Avg0.210.190.240.260.370.220.360.430.450.46
+ +![](images/4d468ab364f72246eb3177905c2668ac96a57820745c28a19056ebb79acd03ee.jpg) +Figure 8. Generated images after TI on NovelConcepts10 BlueElephant poisoned by ADM+. Generation prompt: “a $R*$ on the beach” + +pression; we apply Regen default settings with Stable Diffusion 1.5, utilize published repositories for PDMPure and AdvClean, and set JPEG quality to 25. + +JPEG is a moderately-strong defense, outperforming established methods like Regen and PDMPure on multiple poisons. T600, despite requiring only one parameter change and no extra computation overhead, modestly improves performance against adversarial poisons like ADM+ and SDS+ and outperforms some established defenses on these poisons. LM is a strong standalone defense, improving the generative quality for every poison. Adding JPEG compression as a preprocessing step for T600 or LM further improves performance across all poisons. In particular, JPEG+T600 is an incredibly efficient defense, delivering competitive image quality across all poisons without additional networks or masking. SZT is consistently the strongest defense, achieving state-of-the-art image quality across almost all poisons. Notably, SZT performance on all poison data (average of 0.46) matches the quality achieved by nominal TI on clean data, effectively nullifying all poisons. We also report FID and CLIP Score for CustomConcept101 in Appendix K. We perform the same study on NovelConcepts10 in Appendix L and find that the trends are largely similar. + +We include studies on additional LDM architectures and parameterizations in Appendix O and additional personalization methods in Appendix P. Across all models and methods, SZT and its ablations are the strongest defenses. + +For a qualitative comparison, we display images generated after TI training on the BlueElephant concept poisoned by ADM+ in Fig. 8. Existing defenses (Regen, PDM-Pure, and AdvClean) can occasionally improve generated + +image quality, resulting in generated images that contain a resemblance of the desired concept. Images generated after JPEG compression display some faithfulness to the desired concept. T600 also presents as a moderately effective defense, often removing poison noise but occasionally failing to capture the desired concept. LM alone tends to learn the correct concept shape but not the correct texture or color, as LM does not attempt to mitigate adversarial signals within the masked region. JPEG+LM improves on this weakness. JPEG+T600 strikes a good balance between accurately learning the desired concept and displaying prompt context. SZT is able to accurately learn the desired concept shape and texture. We note that defenses utilizing the LM operation (i.e., LM, JPEG+LM, and SZT) tend to hyperfocus on the concept, generating images that are dominated by the concept. This behavior is reduced for SZT, which uses dilated masks to incorporate some background information from training images. We include further qualitative results for additional poisons and concepts from Custom-Concept101 and NovelConcepts10 in Appendix M. + +# 6. Conclusion + +The SZT defense method is derived from systematic analyses of the underlying mechanisms of DMs as well as observations on the learning tendencies of TI. Despite the simplicity of SZT, it is an effective defense method that performs beyond other established defenses and nullifies many poisons. We hope that our work exposes the vulnerabilities of existing poisons and spurs further research on robust poisons for DMs. + +# Acknowledgments + +This research is supported by the National Research Foundation, Singapore and Infocomm Media Development Authority under its Trust Tech Funding Initiative, and ROSE @ NTU. Any opinions, findings and conclusions or recommendations expressed in this material are those of the author(s) and do not reflect the views of National Research Foundation, Singapore and Infocomm Media Development Authority, or ROSE @ NTU. + +# References + +[1] Ahmed Aldahdooh, Wassim Hamidouche, Sid Ahmed Fezza, and Olivier Déforges. Adversarial example detection for dnn models: a review and experimental comparison. Artificial Intelligence Review, 55(6):4403-4462, 2022. 3 +[2] AUTOMATIC1111. Stable Diffusion Web UI, 2022. 1 +[3] Tim Brooks, Aleksander Holynski, and Alexei A. Efros. Instructpix2pix: Learning to follow image editing instructions, 2023. 2 +[4] Nicholas Carlini and Andreas Terzis. Poisoning and backdooring contrastive learning. In International Conference on Learning Representations, 2022. 2 +[5] Nicholas Carlini and David Wagner. Towards evaluating the robustness of neural networks, 2017. 2 +[6] June Suk Choi, Kyungmin Lee, Jongheon Jeong, Saining Xie, Jinwoo Shin, and Kimin Lee. Diffusionguard: A robust defense against malicious diffusion-based image editing. In The Thirteenth International Conference on Learning Representations, 2025. 1, 2, 3 +[7] CivitAI, 2025. 1 +[8] Jeremy M Cohen, Elan Rosenfeld, and J. Zico Kolter. Certified adversarial robustness via randomized smoothing, 2019. 3 +[9] Ekin D. Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V. Le. Randaugment: Practical automated data augmentation with a reduced search space, 2019. 3 +[10] J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei. ImageNet: A Large-Scale Hierarchical Image Database. In CVPR09, 2009. 14 +[11] Terrance DeVries and Graham W. Taylor. Improved regularization of convolutional neural networks with cutout, 2017. 3 +[12] Prafulla Dhariwal and Alex Nichol. Diffusion models beat gans on image synthesis, 2021. 2 +[13] Hadi M. Dolatabadi, Sarah Erfani, and Christopher Leckie. The devil's advocate: Shattering the illusion of unexploitable data using diffusion models, 2024. 3 +[14] Liam Fowl, Micah Goldblum, Ping yeh Chiang, Jonas Geiping, Wojtek Czaja, and Tom Goldstein. Adversarial examples make strong poisons, 2021. 3 +[15] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H. Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion, 2022. 1, 2, 3, 7 + +[16] Robert Geirhos, Patricia Rubisch, Claudio Michaelis, Matthias Bethge, Felix A. Wichmann, and Wieland Brendel. Imagenet-trained CNNs are biased towards texture; increasing shape bias improves accuracy and robustness. In International Conference on Learning Representations, 2019. 3 +[17] Ian J. Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples, 2015. 1, 2 +[18] Yuchao Gu, Xintao Wang, Jay Zhangjie Wu, Yujun Shi, Yunpeng Chen, Zihan Fan, Wuyou Xiao, Rui Zhao, Shuning Chang, Weijia Wu, Yixiao Ge, Ying Shan, and Mike Zheng Shou. Mix-of-show: Decentralized low-rank adaptation for multi-concept customization of diffusion models, 2023. 2 +[19] Ligong Han, Yinxiao Li, Han Zhang, Peyman Milanfar, Dimitris Metaxas, and Feng Yang. Svdiff: Compact parameter space for diffusion fine-tuning, 2023. 2 +[20] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control, 2022. 1, 2 +[21] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning, 2022. 7 +[22] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium, 2018. 7 +[23] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance, 2022. 2 +[24] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models, 2020. 2, 3 +[25] Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models, 2021. 1, 2, 26 +[26] Hanxun Huang, Xingjun Ma, Sarah Monazam Erfani, James Bailey, and Yisen Wang. Unlearnable examples: Making personal data unexploitable, 2021. 3 +[27] Sandy Huang, Nicolas Papernot, Ian Goodfellow, Yan Duan, and Pieter Abbeel. Adversarial attacks on neural network policies, 2017. 2 +[28] Ziyu Jiang, Tianlong Chen, Ting Chen, and Zhangyang Wang. Robust pre-training by adversarial contrastive learning, 2020. 2 +[29] Xuan Ju, Ailing Zeng, Yuxuan Bian, Shaoteng Liu, and Qiang Xu. Pnp inversion: Boosting diffusion-based editing with 3 lines of code. In The Twelfth International Conference on Learning Representations, 2024. 2 +[30] Minseon Kim, Jihoon Tack, and Sung Ju Hwang. Adversarial self-supervised contrastive learning, 2020. 2 +[31] Diederik P Kingma and Max Welling. Auto-encoding variational bayes, 2022. 12 +[32] Diederik P. Kingma, Tim Salimans, Ben Poole, and Jonathan Ho. Variational diffusion models, 2023. 2 +[33] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dólar, and Ross Girshick. Segment anything. arXiv:2304.02643, 2023. 7 + +[34] Jernej Kos, Ian Fischer, and Dawn Song. Adversarial examples for generative models, 2017. 2 +[35] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion, 2023. 1, 2, 7, 26 +[36] Alexey Kurakin, Ian Goodfellow, and Samy Bengio. Adversarial examples in the physical world, 2017. 1, 2 +[37] Thanh Van Le, Hao Phung, Thuan Hoang Nguyen, Quan Dao, Ngoc Tran, and Anh Tran. Anti-dreambooth: Protecting users from personalized text-to-image synthesis, 2023. 2, 3, 26 +[38] Chumeng Liang and Xiaoyu Wu. Mist: Towards improved adversarial examples for diffusion models, 2023. 3, 14, 15 +[39] Chumeng Liang, Xiaoyu Wu, Yang Hua, Jiaru Zhang, Yiming Xue, Tao Song, Zhengui Xue, Ruhui Ma, and Haibing Guan. Adversarial example does good: Preventing painting imitation from diffusion models via adversarial examples, 2023. 1, 2, 3, 5, 7, 8, 14, 15, 21, 22, 23, 27, 28, 29, 30, 31, 32, 33 +[40] Shanchuan Lin, Bingchen Liu, Jiashi Li, and Xiao Yang. Common diffusion noise schedules and sample steps are flawed, 2024. 26 +[41] Bingyan Liu, Chengyu Wang, Tingfeng Cao, Kui Jia, and Jun Huang. Towards understanding cross and self-attention in stable diffusion for text-guided image editing. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7817-7826, 2024. 4 +[42] Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow, 2022. 26 +[43] Xingchao Liu, Xiwen Zhang, Jianzhu Ma, Jian Peng, and Qiang Liu. Instaflow: One step is enough for high-quality diffusion-based text-to-image generation, 2024. 26 +[44] Yepeng Liu, Yiren Song, Hai Ci, Yu Zhang, Haofan Wang, Mike Zheng Shou, and Yuheng Bu. Image watermarks are removable using controllable regeneration from clean noise, 2024. 3 +[45] Zhuoran Liu, Zhengyu Zhao, and Martha Larson. Image shortcut squeezing: Countering perturbative availability poissons with compression, 2023. 3, 6 +[46] Shilin Lu, Yanzhu Liu, and Adams Wai-Kin Kong. Tf-Icon: Diffusion-based training-free cross-domain image composition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2294–2305, 2023. 4 +[47] Kaifeng Lyu, Zhiyuan Li, Runzhe Wang, and Sanjeev Arora. Gradient descent on two-layer nets: Margin maximization and simplicity bias, 2021. 3 +[48] Mingzhi Lyu, Yi Huang, and Adams Wai-Kin Kong. Adversarial attack for robust watermark protection against inpainting-based and blind watermark removers. In Proceedings of the 31st ACM International Conference on Multimedia, page 8396-8405, New York, NY, USA, 2023. Association for Computing Machinery. 2 +[49] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks, 2019. 1, 2 + +[50] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jia-jun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equations, 2022. 2 +[51] Daiki Miyake, Akihiro Iohara, Yu Saito, and Toshiyuki Tanaka. Negative-prompt inversion: Fast image inversion for editing with text-guided diffusion models, 2024. +[52] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models, 2022. 2 +[53] Saman Motamed, Danda Pani Paudel, and Luc Van Gool. Lego: Learning to disentangle and invert personalized concepts beyond object appearance in text-to-image diffusion models, 2024. 2 +[54] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Mahmoud Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael Rabbat, Vasu Sharma, Gabriel Synnaeve, Hu Xu, Hervé Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. Dinov2: Learning robust visual features without supervision, 2024. 7 +[55] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis, 2023. 26 +[56] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision, 2021. 7 +[57] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents, 2022. 1 +[58] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2022. 1, 2, 3, 7, 12 +[59] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation, 2023. 1, 2, 7, 26 +[60] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S. Sara Mahdavi, Rapha Gontijo Lopes, Tim Salimans, Jonathan Ho, David J Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding, 2022. 1 +[61] Hadi Salman, Alaa Khaddaj, Guillaume Leclerc, Andrew Ilyas, and Aleksander Madry. Raising the cost of malicious ai-powered image editing, 2023. 1, 2, 3, 7, 8, 14, 15, 21, 22, 23, 27, 28, 29, 30, 31, 32, 33 +[62] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, Patrick Schramowski, Srivatsa Kundurthy, Katherine Crowson, Ludwig Schmidt, Robert Kaczmarczyk, and Jenia + +Jitsev. Laion-5b: An open large-scale dataset for training next generation image-text models, 2022. 5, 14 +[63] Shawn Shan, Jenna Cryan, Emily Wenger, Haitao Zheng, Rana Hanocka, and Ben Y. Zhao. Glaze: Protecting artists from style mimicry by text-to-image models, 2023. 2, 3 +[64] Shawn Shan, Wenxin Ding, Josephine Passananti, Stanley Wu, Haitao Zheng, and Ben Y. Zhao. Nightshade: Prompt-specific poisoning attacks on text-to-image generative models, 2024. 2, 3 +[65] Shidoto. Shidoto/adversecleaner: Remove adversarial noise from images, 2025. 3, 7, 8, 22, 23, 27, 28, 29, 30, 31, 32, 33 +[66] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models, 2022. 2 +[67] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution, 2020. 2 +[68] Yang Song, Jascha Sohl-Dickstein, Diederik P. Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations, 2021. 2, 3 +[69] Jeremy Styborski, Mingzhi Lyu, Yi Huang, and Adams Kong. Exploiting supervised poison vulnerability to strengthen self-supervised defense, 2024. 3 +[70] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. Intriguing properties of neural networks, 2014. 1, 2 +[71] Pedro Tabacof, Julia Tavares, and Eduardo Valle. Adversarial images for variational autoencoders, 2016. 2 +[72] Patrick von Platen, Suraj Patil, Anton Lozhkov, Pedro Cuenca, Nathan Lambert, Kashif Rasul, Mishig Davaadorj, Dhruv Nair, Sayak Paul, William Berman, Yiyi Xu, Steven Liu, and Thomas Wolf. Diffusers: State-of-the-art diffusion models. https://github.com/huggingface/diffusers, 2022.1, 7 +[73] Andrey Voynov, Qinghao Chu, Daniel Cohen-Or, and Kfir Aberman. P+: Extended textual conditioning in text-to-image generation, 2023. 2 +[74] Yuxiang Wei, Yabo Zhang, Zhilong Ji, Jinfeng Bai, Lei Zhang, and Wangmeng Zuo. Elite: Encoding visual concepts into textual embeddings for customized text-to-image generation, 2023. 2 +[75] Wikipedia. Multivariate normal distribution, 2025. 12 +[76] Cihang Xie, Jianyu Wang, Zhishuai Zhang, Yuyin Zhou, Lingxi Xie, and Alan Yuille. Adversarial examples for semantic segmentation and object detection, 2017. 2 +[77] Haotian Xue and Yongxin Chen. Pixel is a barrier: Diffusion models are more adversarially robust than we think, 2024. 3, 7, 8, 22, 23, 27, 28, 29, 30, 31, 32, 33 +[78] Haotian Xue, Chumeng Liang, Xiaoyu Wu, and Yongxin Chen. Toward effective protection against diffusion-based mimicry through score distillation. In The Twelfth International Conference on Learning Representations, 2024. 1, 2, 3, 7, 8, 14, 15, 21, 22, 23, 26, 27, 28, 29, 30, 31, 32, 33 +[79] Yihao Xue, Siddharth Joshi, Eric Gan, Pin-Yu Chen, and Baharan Mirzasoleiman. Which features are learnt by contrastive learning? on the role of simplicity bias in class collapse and feature suppression, 2023. 3 + +[80] Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. Cutmix: Regularization strategy to train strong classifiers with localizable features, 2019. 3 +[81] Hongyi Zhang, Moustapha Cisse, Yann N. Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization, 2018. 3 +[82] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models, 2023. 1, 2 +[83] Xulu Zhang, Xiao-Yong Wei, Jinlin Wu, Tianyi Zhang, Zhaoxiang Zhang, Zhen Lei, and Qing Li. Compositional inversion for stable diffusion models, 2024. 2 +[84] Xuandong Zhao, Kexun Zhang, Zihao Su, Saastha Vasan, Ilya Grishchenko, Christopher Kruegel, Giovanni Vigna, Yu-Xiang Wang, and Lei Li. Invisible image watermarks are provably removable using generative ai, 2024. 3, 7, 8, 22, 23, 27, 28, 29, 30, 31, 32, 33 \ No newline at end of file diff --git a/ICCV/2025/When and Where do Data Poisons Attack Textual Inversion_/images.zip b/ICCV/2025/When and Where do Data Poisons Attack Textual Inversion_/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..17e362f336d291d22528ea344d8f0f1412849225 --- /dev/null +++ b/ICCV/2025/When and Where do Data Poisons Attack Textual Inversion_/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e0fca862b4c63763a8ab56d447a90ac84ec2d38444352941e28d315f25c39f49 +size 402347 diff --git a/ICCV/2025/When and Where do Data Poisons Attack Textual Inversion_/layout.json b/ICCV/2025/When and Where do Data Poisons Attack Textual Inversion_/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..05cbd9f09fa900fcd9c2506ca00bc559382ccb68 --- /dev/null +++ b/ICCV/2025/When and Where do Data Poisons Attack Textual Inversion_/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a76044b9281a075a26f639f237fd13e6daf3fa95fd1beb42c4329c8de4b4c6b8 +size 483973 diff --git a/ICCV/2025/Where am I_ Cross-View Geo-localization with Natural Language Descriptions/f1023190-1ee9-442e-9a72-f9dec8a76f13_content_list.json b/ICCV/2025/Where am I_ Cross-View Geo-localization with Natural Language Descriptions/f1023190-1ee9-442e-9a72-f9dec8a76f13_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..9bfc264204d9bf56e9d5e6aa0778f338b3ff0bcd --- /dev/null +++ b/ICCV/2025/Where am I_ Cross-View Geo-localization with Natural Language Descriptions/f1023190-1ee9-442e-9a72-f9dec8a76f13_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dcbcf8886eca90dffaf74d37975e6b7f296497d5b74a423fb12c313d04b060c0 +size 85876 diff --git a/ICCV/2025/Where am I_ Cross-View Geo-localization with Natural Language Descriptions/f1023190-1ee9-442e-9a72-f9dec8a76f13_model.json b/ICCV/2025/Where am I_ Cross-View Geo-localization with Natural Language Descriptions/f1023190-1ee9-442e-9a72-f9dec8a76f13_model.json new file mode 100644 index 0000000000000000000000000000000000000000..d82d073a23e841ea746433549130defb8e7c10cc --- /dev/null +++ b/ICCV/2025/Where am I_ Cross-View Geo-localization with Natural Language Descriptions/f1023190-1ee9-442e-9a72-f9dec8a76f13_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b4ca40f290bfda58a273fb447e2beadda98adf3ec27e0aec39f50c52fafa5fe5 +size 106593 diff --git a/ICCV/2025/Where am I_ Cross-View Geo-localization with Natural Language Descriptions/f1023190-1ee9-442e-9a72-f9dec8a76f13_origin.pdf b/ICCV/2025/Where am I_ Cross-View Geo-localization with Natural Language Descriptions/f1023190-1ee9-442e-9a72-f9dec8a76f13_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..7259872ce55f94c59019cf346016d020efacd15e --- /dev/null +++ b/ICCV/2025/Where am I_ Cross-View Geo-localization with Natural Language Descriptions/f1023190-1ee9-442e-9a72-f9dec8a76f13_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:92a918cae35f950362bdba56c52ff1e4cae4870d040e648080f48df7239955a1 +size 3514346 diff --git a/ICCV/2025/Where am I_ Cross-View Geo-localization with Natural Language Descriptions/full.md b/ICCV/2025/Where am I_ Cross-View Geo-localization with Natural Language Descriptions/full.md new file mode 100644 index 0000000000000000000000000000000000000000..6ec5d1342853cbe157ee2938a4db7234d0287e49 --- /dev/null +++ b/ICCV/2025/Where am I_ Cross-View Geo-localization with Natural Language Descriptions/full.md @@ -0,0 +1,340 @@ +# Where am I? Cross-View Geo-localization with Natural Language Descriptions + +Junyan Ye $^{1,2*}$ , Honglin Lin $^{2*}$ , Leyan Ou $^{1}$ + +Dairong Chen $^{4,1}$ , Zihao Wang $^{1}$ , Qi Zhu $^{1}$ , Conghui He $^{2,3}$ , Weijia Li $^{1\dagger}$ + +1Sun Yat-Sen University, 2Shanghai AI Laboratory, 3Sensetime Research, 4Wuhan University + +![](images/2903a68feb9d0d838c16eac0ede80a18aac052270207aef86657dc708cfb3286.jpg) +Figure 1. Towards Text-guided Geo-localization. In scenarios where GPS signals are interfered with, users must describe their surroundings using natural language, providing various location cues to determine their position $(Up)$ . To address this, we introduce a text-based cross-view geo-localization task, which retrieves satellite imagery or OSM data only based on text queries for position localization (Down). + +# Abstract + +Cross-view geo-localization identifies the locations of street-view images by matching them with geo-tagged satellite images or OSM. However, most existing studies focus on image-to-image retrieval, with fewer addressing text-guided retrieval, a task vital for applications like pedestrian navigation and emergency response. In this work, we introduce a novel task for cross-view geo-localization with natural language descriptions, which aims to retrieve corresponding satellite images or OSM database based on scene text descriptions. To support this task, we construct the CVG-Text dataset by collecting cross-view data from multiple cities and employing a scene text generation approach that leverages the annotation capabilities of Large Multimodal Models to produce high-quality scene text descriptions with localization details. Additionally, we propose a novel text-based retrieval localization method, CrossText2Loc, which demonstrates excellent long-text retrieval capabilities. In terms of explainability, it not only provides similarity scores but also offers retrieval reasons. More can be found at https://github.com/yejy53/CVG-Text. + +# 1. Introduction + +Accurate positioning of ground-level images is crucial for various applications, including pedestrian navigation [35], mobile robot localization [32], and noisy GPS signals correction in crowded urban areas [16, 37]. Traditional localization methods typically rely on 3D point cloud positioning [17, 26] or cross-view retrieval using GPS-tagged satellite images [10, 44, 45]. However, most research focuses on matching image-to-3D data or image-to-image data. Recent studies have begun to explore a novel localization approach based on natural language text, which holds significant value for many practical applications [35, 39, 40]. As shown in Figure 1, taxi drivers rely on verbal instructions from passengers to determine their location [21], or pedestrians describe their position during emergency calls [7]. + +Recent natural language localization methods, such as Text2Pose[21] and Text2Loc [40], have been limited to using natural language to identify individual locations within point clouds. Constructing 3D maps using LiDAR or photogrammetry is expensive on a global scale [1, 13, 27], and the storage costs for 3D maps are also high, often requiring costly cloud infrastructure, which hinders localization on mobile devices. Notably, the cross-view retrieval geo + +localization paradigm [41, 45, 56] that utilizes OSM1 map data or satellite imagery, although oriented towards coarse-grained localization, can still meet the needs of most tasks and has clear advantages over 3D data in terms of coverage and storage costs. Therefore, this paper introduces a novel cross-view geo-localization task, i.e., exploring the use of natural language descriptions to retrieve corresponding OSM or satellite images. + +To tackle this challenging task, a dataset is needed that (i) contains foundational cross-view data with street-view, OSM, and satellite images, and (ii) includes text data capable of simulating human users in describing street-view scenes, while providing high-quality scene localization cues. With the development of large multimodal models (LMMs), annotating text using LMMs seems to be an effective solution [8, 14, 46]. However, LMMs may suffer from vague descriptions or hallucination phenomena [15, 31]. To address these issues, we propose the Cross-View Geo-localization dataset, CVG-Text. We first collected street-view data from over 30,000 locations in three cities, New York, Brisbane, and Tokyo. Then, based on the geographical coordinates, we obtained corresponding paired data of OSM and satellite images. Subsequently, we developed a progressive text description framework that leverages LMM, GPT-4o [28] as the core for generation, combining Optical Character Recognition (OCR) and Open-World Segmentation [29, 52] techniques to generate high-quality scene description text from street-view images while reducing vague descriptions. + +Although the textual data constructed above can provide user-like street scece descriptions, it still has still have a significant domain gap compared to satellite images or OSM data. Moreover, in order to fully capture the scene's detailed information, the generated text descriptions are generally long, often exceeding the text encoding limits of image-text retrieval methods. To address this issue, we propose a novel Cross-view Text-based Localization method, CrossText2Loc. This method includes a length-extended text encoding module, Extended Embedding, which fully leverages the long and complex text descriptions in the dataset. Through contrastive learning strategies, it effectively learns cross-domain matching information. It also features an Explainable Retrieval Module (ERM), which provides natural language explanations alongside the retrieval results. This overcomes the limitations of traditional cross-view retrieval methods that only provide similarity scores, lacking interpretability and making it difficult to make confident decisions. We evaluate the performance of mainstream text-image retrieval methods and our method on this novel task, and the experimental results demonstrate that our CrossText2Loc has significant advantages in recall metrics and interpretability. Our main contributions are as follows: + +We introduce and formalize the Cross-View Geo + +localization task based on natural language descriptions, utilizing scene text descriptions to retrieve corresponding OSM or satellite images for geographical localization. + +- We propose CVG-Text, a dataset with well-aligned streetviews, satellite images, OSM, and text descriptions across three cities and over 30,000 coordinates. Additionally a progressive scene text generation framework based on LMM is presented, which reduces vague descriptions and generates high-quality scene text. +- We introduce CrossText2Loc, a novel text-based localization method that excels in handling long texts and interpretability. It achieves an improvement of over $10\%$ in Top-1 recall compared to existing methods, while offering retrieval reason and confidence beyond similarity scores. + +# 2. Related Work + +# 2.1. Cross-view Geo-localization + +Cross-view geo-localization identifies the geographic locations of street-view images by matching them with geographic reference satellite databases or OSM databases for coarse localization [23, 34, 41, 45]. For example, works like Sample4G [10] employ a contrastive learning framework to match features of street-view images with satellite image features, achieving high-accuracy satellite data retrieval. However, current cross-view retrieval and localization tasks primarily focus on image-to-image, with limited consideration for text, which presents certain shortcomings in practical applications. Additionally, existing cross-view retrieval methods mainly provide similarity score, with little research dedicated to confident and interpretable retrieval localization. Our proposed text retrieval localization task is based on the retrieval and localization of natural language, addressing the gap in existing research regarding text-guided scene localization applications. This approach enables more transparent and interpretable retrieval localization through the use of natural language. + +# 2.2. Visual Language Navigation and Localization + +Visual Language Navigation (VLN) requires agents to navigate specific environments based on natural language instructions [2, 6]. Previous tasks have primarily focused on decision-making based on images and natural language. Recent works have begun to shift from visual language navigation to direct visual language localization tasks. For instance, Loc4Plan[35] and AnyLoc[18] emphasize the necessity of visual spatial localization prior to navigation. Text2Pose [21] and Text2Loc [40] explore the use of natural language to identify individual positions in outdoor point cloud maps. However, point cloud retrieval tasks tend to incur high storage and computational costs when handling large-scale areas. + +# 2.3. Data Synthesis via LMMs + +The rapid development of multimodal large models (LMMs) has demonstrated their outstanding ability to generate high-quality natural language descriptions [3, 28, 42, 47]. Many studies have leveraged LMMs for automated data annotation [14], such as LatteCLIP [4], which synthesizes text using LMMs for unsupervised CLIP fine-tuning. However, text retrieval localization tasks impose higher demands on the annotation capabilities of LMMs, requiring them to accurately identify key localization details in street-view images, such as store signs and other critical information, while minimizing interference from hallucination phenomena [31]. In our fine-grained text synthesis, we incorporate street-view image, OCR, and open-world segmentation to enhance GPT's capture capability and reduce hallucination. + +# 2.4. Multi-modality Alignment + +In this work, our task involves retrieving satellite images or OSM images based on scene text descriptions synthesized by LMMs, which can be seen as a subtask of text-to-image retrieval [9, 11, 43, 48, 54, 55]. The CLIP [30] model introduced a contrastive learning approach between image-text pairs, establishing a new paradigm for recent text retrieval tasks. However, such models often have fixed maximum sequence length limitations, typically defaulting to 77 tokens, making it challenging to handle complex long-text scene descriptions. This can lead to the loss of fine-grained textual information, negatively impacting model performance. Inspired by works like Long-CLIP [50], our retrieval method employs stretching to extend the model's acceptable text length, enabling the retrieval of long text descriptions for environmental contexts. + +# 3. CVG-Text Dataset + +# 3.1. Overview + +We introduce CVG-Text, a multimodal cross-view retrieval localization dataset designed to evaluate text-based scene localization tasks. CVG-Text covers three cities: New York, Brisbane, and Tokyo, encompassing over 30,000 scene data points. The data from New York and Tokyo is more oriented toward urban environments, while the Brisbane data leans towards suburban scenes. Each individual point includes corresponding street-view images, OSM data, satellite images, and associated scene text descriptions. The dataset is randomly split into training and test sets with a ratio of 5:1. More details can be found in the supplementary materials. + +Statistical Overview of Text Data. Figure 2 presents the feature statistics of the scene text. The t-SNE dimensionality reduction visualization indicates a relatively dispersed distribution of the text data, reflecting high diversity. Texts from the same city exhibit clustering, while texts from different cities are clearly distinguishable, highlighting regional varia + +![](images/0f8ed9645c9613c9e0beed23edf7a4a9513f6bc72cc771f5e640be33fcb6ea6d.jpg) + +![](images/f1b116ef51557c02917acc58d21121dd51fe4dd1e30a09178873f7bca46e532e.jpg) + +![](images/e19951aca0ec3981dfd9f4b6963755410a34b31f8f6a8f76375cb0616321ea2f.jpg) +(c) Token Length +Figure 2. Textual Feature Statistics Overview. (a) t-SNE visualization of text data from different cities; (b) text similarity matrix; (c) token length distribution histogram; (d) comparison of text statistics across different datasets. + +![](images/916f18bb9d7ca194e1d0c367c68426a95cdafa298898c695f8a9bc10e1f848bf.jpg) +(d) Text statistics + +tions in text features. These differences are likely tied to the unique styles and cultural characteristics of each city. The text similarity matrix reveals low similarity, demonstrating the independence of the texts, which effectively represent each distinct scene and reduce the risk of confusion between different texts. The average text length generated by the multimodal large model GPT-4o exceeds 126 tokens, with two prominent peaks at 100 and 145 tokens, corresponding to single-view and panoramic images, respectively. Compared to datasets such as Text2Pose and GeoText-1652, our data shows superior performance in vocabulary richness, token length, and entropy, reflecting a higher quality of text. + +Comparisons with Existing Datasets. Table 1 compares CVG-Text with existing datasets. In contrast to common cross-view retrieval datasets, such as CVUSA [38] and VIGOR [56], CVG-Text includes aligned text modality information, enabling the evaluation of text-based scene localization tasks and interpretability analysis for cross-view retrieval. Furthermore, CVG-Text demonstrates superior data completeness, encompassing panoramic street-views, single-perspective street-views, aerial images, and OSM data. + +Compared to GeoText-1652 [9] dataset, which is primarily used for drone navigation, the text descriptions are directly derived from drone images and are used for drone image retrieval. Our task, however, focuses on addressing the needs of pedestrians, tourists, and other users, with text originating from street-view images and used for cross-domain retrieval of satellite or OSM images. There is a significant difference in both task objectives and the source of text. Furthermore, the coverage of drone images is more limited compared to the OSM and satellite images, making it difficult to achieve large-scale geo-localization. + +Table 1. Comparison of the proposed CVG-Text with existing cross-view datasets. # Ref. and # G-Query represent the number of reference images and ground query images, respectively. + +
DatasetTextPano1Single2OSM# G-Query# Ref.
CVUSA [38]✓44k44k
CVACT [25]✓128k128k
VIGOR [56]✓90k105k
CVGlobal [45]✓✓130K130K
Uni.-1652 [53]✓14K90k
Geotext-1652 [9]✓✓14K90k
Ours✓✓✓✓30k60k
+ +1 Pano refers to the Panoramic street-view images. +2 Single refers to the Single-view street-view images. + +# 3.2. Data Collection + +We collected panoramic and single-perspective street-view images from different city areas using the Google Street View ${}^{2}$ and Google Places API ${}^{3}$ . The resolution of the panoramic street-view images is ${2048} \times {1024}$ ,with the north direction aligned to the middle column. The resolution of the single-perspective images is not fixed, but all are high-definition images. Based on the latitude and longitude coordinates of the street-view images, we collected corresponding satellite image tiles using Google Maps ${\mathrm{{API}}}^{4}$ . The size of the satellite images is ${512} \times {512}$ ,with a zoom level of 20 and a ground resolution of around ${0.12}\mathrm{\;m}$ ,aligned with the center of the street-view images. + +Additionally, we collected corresponding OSM data for each region in vector format, encompassing global geographic and map information. OSM data contains numerous points of interest (POIs) with rich label information, such as restaurant names and bus stops, which are highly useful for geo-localization tasks. We utilize raster tiles provided by the OSM official website as retrieval targets, retaining various POI identifiers. The size of the OSM raster data is $512 \times 512$ , which is close to the image size and corresponding geographical extent of the satellite images. Since OSM data is dynamically updated, we collected street-view data from recent years to minimize discrepancies between the OSM and street-view data. Existing cross-view datasets typically include street-view data from before 2021; thus, rather than adding text annotations to existing datasets, we opted to collect new data. + +# 3.3. Text Data Synthesis + +Figure 3 illustrates the main process of text synthesis based on GPT-4o. We first utilize Paddle OCR5 to capture text within street-view images, enabling GPT to focus on key semantic text information present in the images while re + +```txt +$\mathbf{2}$ https://www.google.com/streetview/ $\mathbf{3}$ https : // developers . google . com / maps / documentation/places/web-service $\mathbf{4}$ https://www.google.com/maps $\mathbf{5}$ https://github.com/PaddlePaddle/PaddleOCR +``` + +# Task Description: + +Describe your surroundings to help identify location... + +# Data Description: + +- Street-view Image; +OCR result ; +- Open World Segmentation; + +# Describe Paths: + +1 - Road characteristics.. +2 - Scene trademark to +3-Building color & location... +4 - ... + +If there is no corresponding information, skip + +![](images/f50fec6aa11c5d5ff25b18f277f63df803f1a3824e82a8e8df69df96fc456a0f.jpg) +Figure 3. Overall Process for Street-View Text Description Generation using GPT-4o. + +ducing hallucination phenomena. Additionally, we perform open-world segmentation on the street-view images to provide more semantic details and location information for the subsequent text synthesis process. With the help of the semantic segmentation results, OCR outputs related to moving objects, such as vehicles, can also be filtered out. + +Next, we designed a systematic prompting scheme to generate scene text descriptions using GPT-4o. The system prompts are divided into four parts: task requirements, input data introduction, description paths, and response demands. Specifically, we employ a progressive scene analysis strategy to guide GPT in sequentially describing road features, building signage, and the overall environment, including aspects such as color, material, and distribution. In real-world applications, users may not know the exact geographic orientation. Therefore, text descriptions often use simple directional cues like front, back, left, and right to describe the scene. For more implementation details, refer to the supplementary materials. + +In terms of quality control, we first filter out incorrectly formatted responses and conduct GPT-based review of both images and textual descriptions to ensure accuracy, consistency, content completeness, relevance, clarity, and comprehensiveness. Samples that do not meet the standards are re-synthesized. Additionally, we extracted $20\%$ of the samples (approximately 6,000) for manual expert review, involving 10 human evaluators and requiring around 100 hours of work. The pass rate for the manual checks was $77.6\%$ . Furthermore, street-view retrieval experiments using a retrained CLIP model showed that the Top-1 recall rate for text-to-street view was over $85.5\%$ , validating the high quality of the dataset's text descriptions. + +# 4. Method + +# 4.1. Overview + +Problem Formulation. In this work, we introduce a novel task for cross-view geo-localization with natural language. The objective of this task is to leverage natural language + +descriptions to retrieve its corresponding OSM or satellite images, of which the location information is usually available. The input for this task is a text description $Q$ -Text describing a street scene. In practical applications, users often have location prior, such as knowing they are in a particular district of New York City, but not the precise location. Therefore, the retrieval model utilizes this location prior to narrow the search scope $M$ , querying a smaller subset of satellite images, $R$ -sat, and OSM image, $R$ -OSM. + +Model Architecture. To address the challenge of the task, we propose the CrossText2Loc architecture (Figure 4). It aligns the image and text domains through an enhanced text embedding module and a contrastive learning loss $L_{itc}$ enabling efficient text retrieval tasks. Furthermore, during inference, we introduce an optional Explainable Retrieval Module (ERM) to provide natural language analysis, enhancing the interpretability and confidence of retrieval decisions. + +# 4.2. Image-text Contrastive Learning + +CrossText2Loc adopts a dual-stream architecture, consisting of a text encoder and a visual encoder that extract features from text queries and reference satellite or OSM images, respectively. Since the text describes scenes corresponding to street view images, there is a significant domain gap between these and the reference satellite and OSM data. As a result, the features obtained by the pre-trained encoder are spatially distant. Therefore, we align the image and text embedding spaces through contrastive learning, while jointly training both the text and image encoders. We set the batch size to $n$ and align the representations of images and text by minimizing the contrastive learning loss function $L_{itc}$ . To achieve this, we uniformly encode panoramic and single-view texts as $t$ , and satellite and OSM images as $v$ . The loss function is expressed as: + +$$ +L _ {i t c} = \sum_ {i = 1} ^ {n} \sum_ {j = 1} ^ {n} - \log \frac {\exp \left(\operatorname {s i m} \left(v _ {i} , t _ {j}\right) / \tau\right)}{\sum_ {k = 1} ^ {n} \exp \left(\operatorname {s i m} \left(v _ {i} , t _ {k}\right) / \tau\right)} \tag {1} +$$ + +Where $v_{i}$ and $t_{j}$ represent the $i$ -th image and the $j$ -th text embedding vector respectively. $\tau$ is a learnable temperature used to control the sharpness of softmax distribution. + +Scene description texts are generally complex and lengthy, and their matching with OSM or satellite images is not based on direct surface-level word matching, but requires deeper semantic understanding. However, previous text encoders, such as CLIP[30], have limited positional embeddings, often resulting in truncation and insufficient global text representation. To address this issue, we adopt the Expanded Positional Embedding (EPE) method to extend the positional embeddings of the text encoder, which are then fed into the Transformer block. During subsequent embedding in the Transformer architecture, attention mechanisms are used to capture critical localization cues. Specially, we utilize linear interpolation to expand the positional embeddings to accommodate a sequence length of $N$ (300) tokens. The expanded + +![](images/34b97388dbc70b4c1653a5276a838357abfbfb6b4350987318ad26b3cac27a8f.jpg) +Figure 4. The proposed CrossText2Loc method. Street-view texts serve as query inputs, with satellite and OSM images as references. + +positional embedding $P^{*}$ can be computed from the original positional embedding $P$ as follows: + +$$ +P ^ {*} (x) = \left(1 - \left(x - \lfloor x \rfloor\right)\right) \cdot P (\lfloor x \rfloor) + \left(x - \lfloor x \rfloor\right) \cdot P (\lceil x \rceil) \tag {2} +$$ + +Where $P^{*}(x)$ represents the expanded positional embedding at position $x$ , $P(\lfloor x\rfloor)$ and $P(\lceil x\rceil)$ are the values from the original positional embedding at the indices $\lfloor x\rfloor$ and $\lceil x\rceil$ , with $\lfloor \cdot \rfloor$ and $\lceil \cdot \rceil$ denoting the floor and ceiling functions respectively. Due to the scene texts in our case are generated by GPT for simulating user data, there are no prominent short titles at the beginning which carry significant importance. Therefore, unlike LongCLIP [50] uses Knowledge Preserving Stretching, we use a full interpolation method. + +# 4.3. Explainable Retrieval Module (ERM) + +We also propose an optional Explainable Retrieval Module, designed to enhance the interpretability and trustworthiness of geolocation retrieval predictions through natural language. Attention Heatmap Generation. Inspired by the method described in [5], we generate attention heatmaps to reveal the model's focus during retrieval. We initialize the correlation heatmap before the starting layer $s$ with an identity matrix $I$ . Then, we iteratively process each attention layer from the starting layer $s$ to the output layer $L$ . In each attention mechanism, we accumulate non-negative gradient contributions to generate attention heatmaps that highlight the regions the model focuses on. The image or text correlation heatmap $R$ can be calculated as: + +$$ +R ^ {(l)} = R ^ {(l - 1)} + \frac {1}{H} \sum_ {h = 1} ^ {H} \max \left(0, \nabla A _ {h} ^ {(l)} \odot A _ {h} ^ {(l)}\right) R ^ {(l - 1)} \tag {3} +$$ + +Where $A_h^{(l)}$ represents the attention weight of the $h$ -th attention head at layer $l$ , $\nabla A_h^{(l)}$ is the gradient of the attention weight, $\odot$ denotes the Hadamard product, $H$ is the total number of attention heads. + +LMM Interpretation. After generating the image and text heatmaps, we feed them into the LMM (GPT-4o) with a + +
MethodSatellite imageOSM data
NewYorkBrisbaneTokyoNewYorkBrisbaneTokyo
R@1L@50R@1L@50R@1L@50R@1L@50R@1L@50R@1L@50
ViLT[19]11.5815.5811.0014.5010.8315.505.839.928.6711.754.679.17
X-VLM[48]15.7416.8615.6717.6012.4614.3416.1417.2620.4621.949.5310.94
SigLIP-B/16[49]19.6721.0819.5822.0015.5817.9220.1721.5825.5827.4211.9213.67
SigLIP-SO400M[49]33.5034.8334.2536.8328.4231.5027.7529.5829.7531.5817.5019.50
EVA2-CLIP-B/16[12]25.1726.5828.4231.7522.5025.2518.5820.8327.3328.8313.9215.67
EVA2-CLIP-L/14[12]34.0835.6735.6738.0031.0034.0826.3328.6730.9232.6719.8322.50
CLIP-B/16[30]26.6728.1729.9232.5824.0027.2527.4229.4230.8332.6717.7519.92
CLIP-L/14[30]35.0837.0834.0837.2528.0830.5031.5033.5832.5034.6721.0023.17
BLIP[22]34.5837.2534.5038.1729.7533.6752.9255.9243.0046.3330.6734.50
Ours (w/o ERM)46.2548.7543.5847.4236.8339.5859.0862.0046.0848.6734.3338.33
Ours50.3353.0747.5851.8041.7543.8662.3365.3948.7551.5036.9241.22
+ +Table 2. Quantitative comparison of different methods on CVG-Text. R@1 represents the Top-1 image recall rate; L@50 represents the recall rate where localization error is less than 50 meters. [Key: Best, Second Best] + +carefully designed prompt. First, the LMM observes the heatmaps to identify key clues in the image and text that the model focuses on during retrieval, such as specific landmarks or geographical features. Then, the LMM compares and reasons over the highlighted regions in the text and image, providing an explanation for why the model retrieved the query in a particular way, i.e., the rationale behind the matching of the query and reference data. Finally, the LMM outputs the confidence level of the retrieval rationale, simulating user decision-making. LMMs are capable of providing retrieval explanations in natural language, beyond just similarity measures, which is highly beneficial for the interpretability and trustworthiness of geo-localization retrieval. + +Confidence re-ranking. The confidence scores obtained through explainable retrieval can be used to re-rank the original top-5 similarity scores from the retrieval match. We set the results with top-1 confidence scores lower than 0.5 and normalize both the similarity and confidence scores to the [0, 1] range, summing them to obtain the re-ranked Top-5 results. In the subsequent experimental section, we demonstrate that the effective re-ranking strategy helps simulate user retrieval decisions and improves recall in tasks. + +# 5. Experiment + +# 5.1. Experimental Setup + +Implementation Details. We used the CLIP-L/14@336px model [30] pre-trained by OpenAI as the backbone, with the Adam optimizer [20], a learning rate of 1e-5, and cosine learning rate decay. The batch size was set to 128, and training was conducted over 40 epochs on four NVIDIA A100 GPUs. The image resolution was set to default $336 \times 336$ , and the text context length $N$ was 300 tokens. We initialized the temperature coefficient $\tau$ from the checkpoint. + +Moreover, as mentioned in Section 4.1, based on practical application requirements and utilizing user location priors, we set the retrieval range $M$ to 100. During testing, the reference database consists of 100 samples from the nearby area, covering an area of approximately $10\mathrm{km}^2$ . $M$ is a configurable parameter, and we present additional results with different $M$ settings in the supplementary material. The image resolution for all comparison methods follows their respective default best settings, and all methods are trained on the same proposed dataset. + +Evaluation Metrics. Following previous cross-view geolocation works [10, 33], we use the image recall accuracy of the top K images as an evaluation metric to assess text retrieval localization performance. Specifically, given a query text for a certain location, if its ground-truth OSM and satellite image is within the top K retrieval results, the location is considered "successfully localized." Additionally, we also provide the localization recall rate metric [21, 40]. Similarly, the localization recall rate refers to the proportion of retrieved results where the distance to the actual location is below a specified threshold. + +# 5.2. Geo-localization Performance + +We evaluated the performance of various text-based retrieval methods under different settings of satellite images and OSM data, with the results shown in Table 2. Among the existing approaches, BILP achieves optimal performance, as it is not constrained by limitations on text embedding length.Our method, even without the ERM, achieved the best results. Compared to the baseline CLIP method, the proposed approach improved Recall@1 by $14.1\%$ and Recall@10 by $14.8\%$ , demonstrating its advantages in this task. + +Next, we evaluated the performance of methods in different cities and across different data sources (Satellite/OSM). + +![](images/b66d19f8318919bb16b5af183b60639a7682efe16e8bbb9299bd0fae5550bbfd.jpg) + +I am in front of a restaurant named 'Burger Mania', which features a modern facade with a neon sign. The establishment offers a variety of menu items, including burgers, sandwiches, pancakes, craft beer, coffee, tea, and cake, as indicated by the signage above the entrance. There is a welcoming atmosphere with greenery at the base of the window... + +The model's text attention focuses on the store sign "Burger Mania" and the food offerings. In the OSM data, the model's attention is primarily on restaurant information, with the retrieved OSM data containing a matching "Burger Mania" location. Therefore, the retrieval reason is identifying the same "Burger Mania" establishment. Confidence: 0.95 + +![](images/97f8e0d28f6376fe11e79e869d769db11720a4ea8c6a2e48a5d53dbee80de52c.jpg) + +The model's text attention focuses on the zebra crossing and the mesh like white lines on the ground. In the satellite data, the model's attention is also directed towards these markings, with the retrieved satellite image containing intersections and mesh-like white lines. Thus, the retrieval reason is recognizing the same intersection and ground mesh patterns. Confidence: 0.85 + +I am at an intersection with a clear zebra crossing sign and a clear white gridline on the ground. In front of me is a multi-storey building with a large billboard prominently displayed on it, showing an advertisement. Traffic lights are present to control vehicle flow. To my left, there are several fast food restaurants including 'Black Burger' and a convenience store. The area is bustling with activity, typical of a city center... + +![](images/4ede223e12c1a3628567207ba45c33d36f03b741e97eaf1f526de4bb6b0de815.jpg) + +![](images/e7c6629978920f8862e58dec5ff9032e316250f65f8a274e3fc5c84a18243af0.jpg) + +![](images/36139c14f1422e9a59dc48d0322e8bc2a602be4de82ad7e9a7262a7c291295be.jpg) + +![](images/614c7df9c1fdbc1670b609da4891a65295b0cd44f9e317559f243ada57e93a5c.jpg) + +![](images/d37fa3cc35087fe653d124d8b0704d42144b5aac589a213942e65392926931d0.jpg) + +![](images/f5f6c258d35329714e3d1dd78a178406850ec4c260710c61dcbbcf97189c06d5.jpg) +Figure 5. Qualitative retrieval results on CVG-Text Dataset. The left side of the figure displays the original street-view data, synthetic text data with corresponding response heatmaps, and retrieval reason provided by our ERM module. The right side shows the top three retrieval results with corresponding response heatmaps; green indicates correct matches and red denotes incorrect results. + +![](images/2d3fc5064f7654038097953fe440d0028fb4e1ffbdd9fcbaa4166652330dd80a.jpg) + +![](images/379262e588218c8d7b913c43709a3d603e6609abfc3f6b12d42a5f108a7b3168.jpg) + +![](images/b2d023b5f4a357615b78993bf115313890d2d48b18b15b0f52983e9eca50a83e.jpg) + +In New York, OSM outperformed satellite imagery due to richer POI data, such as bus stops and store names, which are hard to identify in satellite images. In contrast, the level of detail in OSM data for Brisbane is limited, and in this case, the localization performance of satellite images is comparable to that of OSM. In Tokyo, due to the poor pre-training of CLIP on Japanese, the model's response to certain Japanese words in the street view description text and Japanese POIs in OSM data is weak, leading to the least favorable performance in Tokyo. The supplementary materials also include cross-city evaluations and a collaborative retrieval method for OSM and satellite images. + +# 5.3. Explainable Retrieval + +We present the results of the Explainable Retrieval Module (ERM) in Figure 5. This method uses text and the heatmap responses of the retrieved images to highlight the key features the localization model focuses on. In the first example, the model focuses on "Burger Mania". Interestingly, even in non-top-1 results, it still emphasizes the burger icon. In the second example, the model focuses on the "zebra crossing" and "white gridline", with the first three retrieval results all highlighting the zebra crossing, and the best retrieval result matching both features. We provide a quantitative evaluation of the ERM module's retrieval rationale in the supplementary materials, using similarity matching with human-written + +explanations, CIDEr[36], ROUGE_L[24], and multidimensional human scoring. + +Moreover, the confidence scores obtained from LMMs essentially simulate the user's confidence decision-making process. When the confidence score is low, it indicates that, even though the similarity score of the result may be high, it is still not convincing. In this case, by applying the re-ranking strategy, we can obtain better retrieval results, as shown in Table 2. The practical value of the Explainable Retrieval Module (ERM) lies in assisting users to assess the rationale behind the localization decision. They can select the most reasonable result from similar candidates, whereas previous methods only relied on the highest similarity match. If the rationale provided by ERM is not convincing enough, users can choose to re-rank the remaining search results or add additional descriptions to provide more visual clues. + +Additionally, we generate feature heatmap responses based on text and reference images. By analyzing the top-ranked words in the model's attention, we identified eight focal subcategories, such as "LandMarker" for landmarks and "SignName" for store names or bus stop names. A list of these words and categories is in the supplementary materials. Attention scores for each category in the scene are shown in Figure 6. Results reveal that the cross-view retrieval model mainly focuses on landmarks and road information, with less attention to vehicles, sky, and weather, as + +![](images/9e581977cb522a40ba1c23897e69c3c2102740dc869c970f88f76157a3993767.jpg) +Figure 6. The average attention score for each object category (left) and the word cloud of selected attention words (right). + +
MethodsText qualityRetrieval result
LenTTRSimi.R@1-OSMR@1-Sat
Baseline1080.740.2225.1738.00
Ours1260.760.1759.0846.25
+ +Table 3. Comparison between Baseline (directly using GPT-generated Text) and our CVG-Text dataset in New York. + +
MethodSatelliteOSM
R@1R@5R@1R@5
SigLIP19.6753.3320.1741.83
SigLIP + EPE29.5064.0045.2570.25
CLIP35.0871.4231.5055.42
CLIP + EPE46.2581.5859.0882.75
+ +Table 4. Ablation Study of Expanded Positional Embedding (EPE) Module in New York. + +they are less relevant for localization. This highlights key clues for understanding cross-view retrieval models. + +# 5.4. Ablation Studies + +We evaluated the text synthetic effects of those directly generated by GPT versus those generated through the integrated process of CVG-Text, as shown in Table 3. CVG-Text exhibits higher text length and vocabulary complexity (TTR), with lower text similarity, indicating better text quality. The inclusion of OCR assistance helps precisely capture textual details in street view images, reducing GPT's tendency for vague descriptions and hallucinations, and effectively generates text with key localization cues, significantly improving OSM retrieval performance. Open-World Segmentation further enhances GPT's semantic and spatial understanding, improving satellite image retrieval performance. + +We employed two different learning architectures, CLIP [30] and SigLIP [49], to validate the contribution of our proposed Expand Positional Embedding (EPE) module. As + +
MethodOSMSatelliteR@1 average
R@1R@10R@1R@10
SAFA [33]19.2544.8877.4095.3048.32
Geo-Dtr [51]24.1053.3086.4598.8055.28
Sample4G [10]27.1062.9091.7099.2059.40
Text-only59.0890.0046.2591.0052.67
Sample4G [10] +Text67.3097.2098.4099.8082.80
+ +Table 5. Impact of Adding Text Branch for Cross-View Retrieval. + +shown in Table 4, the use of the EPE module significantly improves the performance of both text retrieval methods on satellite images and OSM, with R@1 recall rates increasing by $10.5\%$ and $26.3\%$ , respectively. This demonstrates the effectiveness of extending text encoding length for handling longer text content in this task. + +# 5.5. Further Discussion on Cross-view retrieval task + +Previous cross-view retrieval tasks have primarily focused on image-to-image queries. Leveraging the CVG-Text dataset, we expanded this task to multimodal image-text joint queries. Building on the state-of-the-art cross-view street-view retrieval method, Sample4G [10], we added an additional text query branch via fusing the similarity scores. We conducted satellite and OSM retrieval experiments in the New York area of the CVG-Text dataset. As shown in Table 5, compared to using only street-view images as queries, the addition of a text branch improved Top-1 recall by $6.7\%$ for the satellite retrieval task and by $40.2\%$ for the OSM retrieval task, effectively enhancing retrieval accuracy. Since the detailed scene descriptions in CVG-Text align more closely with POI data in OSM, the improvement for OSM retrieval is more pronounced than for satellite image data. + +# 6. Conclusion + +In this work, we explore the task of cross-view geolocation using natural language descriptions and introduce the CVG-Text dataset, which includes well-aligned street-views, satellite images, OSM images, and text descriptions. We also propose the CrossText2Loc text retrieval localization method, which excels in handling long-text retrieval and interpretability for this task. This work represents another advancement in the field of natural language-based localization. It also introduces new application scenarios for cross-view localization, encouraging subsequent researchers to explore and innovate further. + +# Acknowledgements + +This work was supported in part by the Natural Science Foundation of Guangdong Province, China (Grant No. 2025A1515010400), the National Natural Science Foundation of China (Grant No. 42201358) and Shanghai Artificial Intelligence Laboratory. + +# References + +[1] Sameer Agarwal, Yasutaka Furukawa, Noah Snavely, Ian Simon, Brian Curless, Steven M Seitz, and Richard Szeliski. Building rome in a day. Communications of the ACM, 54(10): 105-112, 2011. 1 +[2] Peter Anderson, Qi Wu, Damien Teney, Jake Bruce, Mark Johnson, Niko Sünderhauf, Ian Reid, Stephen Gould, and Anton Van Den Hengel. Vision-and-language navigation: Interpreting visually-grounded navigation instructions in real environments. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3674-3683, 2018. 2 +[3] Anthropic. The claude 3 model family: Opus, sonnet, haiku, 2024. Accessed: 2024-09-23. 3 +[4] Anh-Quan Cao, Maximilian Jaritz, Matthieu Guillaumin, Raoul de Charette, and Loris Bazzani. Latteclip: Unsupervised clip fine-tuning via lmm-synthetic texts. arXiv preprint arXiv:2410.08211, 2024.3 +[5] Hila Chefer, Shir Gur, and Lior Wolf. Generic attention-model explainability for interpreting bi-modal and encoder-decoder transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 397-406, 2021. 5 +[6] Howard Chen, Alane Suhr, Dipendra Misra, Noah Snavely, and Yoav Artzi. Touchdown: Natural language navigation and spatial reasoning in visual street environments. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12538-12547, 2019. 2 +[7] Jiaqi Chen, Daniel Barath, Iro Armeni, Marc Pollefeys, and Hermann Blum. "where am i?" scene retrieval with language. arXiv preprint arXiv:2404.14565, 2024. 1 +[8] Lin Chen, Jinsong Li, Xiaoyi Dong, Pan Zhang, Conghui He, Jiaqi Wang, Feng Zhao, and Dahua Lin. Sharegpt4v: Improving large multi-modal models with better captions. arXiv preprint arXiv:2311.12793, 2023. 2 +[9] Meng Chu, Zhedong Zheng, Wei Ji, and Tat-Seng Chua. Towards natural language-guided drones: Geotext-1652 benchmark with spatially relation matching. arXiv preprint arXiv:2311.12751, 2023. 3, 4 +[10] Fabian Deuser, Konrad Habel, and Norbert Oswald. Sample4geo: Hard negative sampling for cross-view geolocation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 16847-16856, 2023. 1, 2, 6, 8 +[11] Zi-Yi Dou, Yichong Xu, Zhe Gan, Jianfeng Wang, Shuohang Wang, Lijuan Wang, Chenguang Zhu, Pengchuan Zhang, Lu Yuan, Nanyun Peng, et al. An empirical study of training end-to-end vision-and-language transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18166-18176, 2022. 3 +[12] Yuxin Fang, Quan Sun, Xinggang Wang, Tiejun Huang, Xinlong Wang, and Yue Cao. Eva-02: A visual representation for neon genesis. Image and Vision Computing, 149:105171, 2024. 6 +[13] Jan-Michael Frahm, Pierre Fite-Georgel, David Gallup, Tim Johnson, Rahul Raguram, Changchang Wu, Yi-Hung Jen, + +Enrique Dunn, Brian Clipp, Svetlana Lazebnik, et al. Building rome on a cloudless day. In Computer Vision-ECCV 2010: 11th European Conference on Computer Vision, Heraklion, Crete, Greece, September 5-11, 2010, Proceedings, Part IV 11, pages 368-381. Springer, 2010. 1 +[14] Owen He, Ansh Jain, Axel Adonai Rodriguez-Leon, Arnav Taduvayi, and Matthew Louis Mauriello. From crowdsourcing to large multimodal models: Toward enhancing image data annotation with gpt-4v. 2023. 2, 3 +[15] Lei Huang, Weijiang Yu, Weitao Ma, Weihong Zhong, Zhangyin Feng, Haotian Wang, Qianglong Chen, Weihua Peng, Xiaocheng Feng, Bing Qin, et al. A survey on hallucination in large language models: Principles, taxonomy, challenges, and open questions. arXiv preprint arXiv:2311.05232, 2023. 2 +[16] Zilong Huang, Jun He, Junyan Ye, Lihan Jiang, Weijia Li, Yiping Chen, and Ting Han. Scene4u: Hierarchical layered 3d scene reconstruction from single panoramic image for your immerse exploration. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 26723-26733, 2025. 1 +[17] Arnold Irschara, Christopher Zach, Jan-Michael Frahm, and Horst Bischof. From structure-from-motion point clouds to fast location recognition. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 2599-2606. IEEE, 2009. 1 +[18] Nikhil Keetha, Avneesh Mishra, Jay Karhade, Krishna Murthy Jatavallabhula, Sebastian Scherer, Madhava Krishna, and Sourav Garg. Anyloc: Towards universal visual place recognition. IEEE Robotics and Automation Letters, 9(2):1286-1293, 2023. 2 +[19] Wonjae Kim, Bokyung Son, and Ildoo Kim. Vilt: Vision-and-language transformer without convolution or region supervision. In International conference on machine learning, pages 5583-5594. PMLR, 2021. 6 +[20] Diederik P Kingma. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6 +[21] Manuel Kolmet, Qunjie Zhou, Aljosa Osep, and Laura Leal-Taixe. Text2pos: Text-to-point-cloud cross-modal localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6687-6696, 2022. 1, 2, 6 +[22] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International conference on machine learning, pages 12888-12900. PMLR, 2022. 6 +[23] Weijia Li, Yawen Lai, Linning Xu, Yuanbo Xiangli, Jinhua Yu, Conghui He, Gui-Song Xia, and Dahua Lin. Omnicity: Omnipotent city understanding with multi-level and multiview images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17397-17407, 2023. 2 +[24] Chin-Yew Lin. Rouge: A package for automatic evaluation of summaries. In Text summarization branches out, pages 74-81, 2004. 7 +[25] Liu Liu and Hongdong Li. Lending orientation to neural networks for cross-view geo-localization. In Proceedings of + +the IEEE/CVF conference on computer vision and pattern recognition, pages 5624-5633, 2019. 4 +[26] Simon Lynen, Bernhard Zeisl, Dror Aiger, Michael Bosse, Joel Hesch, Marc Pollefeys, Roland Siegwart, and Torsten Sattler. Large-scale, real-time visual-inertial localization revisited. The International Journal of Robotics Research, 39 (9):1061-1084, 2020. 1 +[27] Pierre Moulon, Pascal Monasse, Romuald Perrot, and Renaud Marlet. Openmvg: Open multiple view geometry. In Reproducible Research in Pattern Recognition: First International Workshop, RRPR 2016, Cancún, Mexico, December 4, 2016, Revised Selected Papers 1, pages 60-74. Springer, 2017. 1 +[28] OpenAI. Hello gpt-4o. https://openai.com/index/hello-gpt-4o/, 2024. 2, 3 +[29] Lu Qi, Jason Kuen, Yi Wang, Jieuxiang Gu, Hengshuang Zhao, Philip Torr, Zhe Lin, and Jiaya Jia. Open world entity segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(7):8743-8756, 2022. 2 +[30] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 3, 5, 6, 8 +[31] Vipula Rawte, Amit Sheth, and Amitava Das. A survey of hallucination in large foundation models. arXiv preprint arXiv:2309.05922, 2023. 2, 3 +[32] Paul-Edouard Sarlin, Cesar Cadena, Roland Siegwart, and Marcin Dymczyk. From coarse to fine: Robust hierarchical localization at large scale. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12716–12725, 2019. 1 +[33] Yujiao Shi, Liu Liu, Xin Yu, and Hongdong Li. Spatial-aware feature aggregation for image based cross-view geolocation. Advances in Neural Information Processing Systems, 32, 2019. 6, 8 +[34] Yujiao Shi, Xin Yu, Dylan Campbell, and Hongdong Li. Where am i looking at? joint location and orientation estimation by cross-view matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4064-4072, 2020. 2 +[35] Huilin Tian, Jingke Meng, Wei-Shi Zheng, Yuan-Ming Li, Junkai Yan, and Yunong Zhang. Loc4plan: Locating before planning for outdoor vision and language navigation. arXiv preprint arXiv:2408.05090, 2024. 1, 2 +[36] Ramakrishna Vedantam, C. Lawrence Zitnick, and Devi Parikh. Cider: Consensus-based image description evaluation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2015. 7 +[37] Xiaolong Wang, Runsen Xu, Zhuofan Cui, Zeyu Wan, and Yu Zhang. Fine-grained cross-view geo-localization using a correlation-aware homography estimator. Advances in Neural Information Processing Systems, 36, 2024. 1 +[38] Scott Workman, Richard Souvenir, and Nathan Jacobs. Wide-area image geolocation with aerial reference imagery. In Proceedings of the IEEE International Conference on Computer Vision, pages 3961-3969, 2015. 3, 4 + +[39] Yan Xia, Yusheng Xu, Cheng Wang, and Uwe Stilla. Vpc-net: Completion of 3d vehicles from mls point clouds. ISPRS Journal of Photogrammetry and Remote Sensing, 174:166-181, 2021. 1 +[40] Yan Xia, Letian Shi, Zifeng Ding, Joao F Henriques, and Daniel Cremers. Text2loc: 3d point cloud localization from natural language. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14958-14967, 2024. 1, 2, 6 +[41] Zimin Xia, Yujiao Shi, Hongdong Li, and Julian FP Kooij. Adapting fine-grained cross-view localization to areas without fine ground truth. In European Conference on Computer Vision, pages 397-415. Springer, 2025. 2 +[42] Zhiyuan Yan, Junyan Ye, Weijia Li, Zilong Huang, Shenghai Yuan, Xiangyang He, Kaiqing Lin, Jun He, Conghui He, and Li Yuan. Gpt-imageval: A comprehensive benchmark for diagnosing gpt4o in image generation. arXiv preprint arXiv:2504.02782, 2025.3 +[43] Shuyu Yang, Yinan Zhou, Zhedong Zheng, Yaxiong Wang, Li Zhu, and Yujiao Wu. Towards unified text-based person retrieval: A large-scale multi-attribute and language search benchmark. In Proceedings of the 31st ACM International Conference on Multimedia, pages 4492–4501, 2023. 3 +[44] Junyan Ye, Qiyan Luo, Jinhua Yu, Huaping Zhong, Zhimeng Zheng, Conghui He, and Weijia Li. Sg-bev: Satellite-guided bev fusion for cross-view semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27748-27757, 2024. 1 +[45] Junyan Ye, Zhutao Lv, Weijia Li, Jinhua Yu, Haote Yang, Huaping Zhong, and Conghui He. Cross-view image geolocation with panorama-bev co-retrieval network. arXiv preprint arXiv:2408.05475, 2024. 1, 2, 4 +[46] Junyan Ye, Baichuan Zhou, Zilong Huang, Junan Zhang, Tianyi Bai, Hengrui Kang, Jun He, Honglin Lin, Zihao Wang, Tong Wu, et al. Loki: A comprehensive synthetic data detection benchmark using large multimodal models. arXiv preprint arXiv:2410.09732, 2024. 2 +[47] Junyan Ye, Jun He, Xiang Zhang, Yi Lin, Honglin Lin, Conghui He, and Weijia Li. Satellite image synthesis from street view with fine-grained spatial textual guidance: A novel framework. IEEE Geoscience and Remote Sensing Magazine, 2025. 3 +[48] Yan Zeng, Xinsong Zhang, and Hang Li. Multi-grained vision language pre-training: Aligning texts with visual concepts. In International Conference on Machine Learning, pages 25994-26009. PMLR, 2022. 3, 6 +[49] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11975-11986, 2023. 6, 8 +[50] Beichen Zhang, Pan Zhang, Xiaoyi Dong, Yuhang Zang, and Jiaqi Wang. Long-clip: Unlocking the long-text capability of clip. In European Conference on Computer Vision, pages 310–325. Springer, 2025. 3, 5 +[51] Xiaohan Zhang, Xingyu Li, Waqas Sultani, Yi Zhou, and Safwan Wshah. Cross-view geo-localization via learning disentangled geometric layout correspondence. In Proceed- + +ings of the AAAI Conference on Artificial Intelligence, pages 3480-3488, 2023. 8 +[52] Junwei Zheng, Ruiping Liu, Yufan Chen, Kunyu Peng, Chengzhi Wu, Kailun Yang, Jiaming Zhang, and Rainer Stiefelhagen. Open panoramic segmentation. In European Conference on Computer Vision, pages 164-182. Springer, 2025. 2 +[53] Zhedong Zheng, Yunchao Wei, and Yi Yang. University-1652: A multi-view multi-source benchmark for drone-based geo-localization. In Proceedings of the 28th ACM international conference on Multimedia, pages 1395–1403, 2020. 4 +[54] Zhedong Zheng, Liang Zheng, Michael Garrett, Yi Yang, Mingliang Xu, and Yi-Dong Shen. Dual-path convolutional image-text embeddings with instance loss. ACM Transactions on Multimedia Computing, Communications, and Applications (TOMM), 16(2):1-23, 2020. 3 +[55] Baichuan Zhou, Haote Yang, Dairong Chen, Junyan Ye, Tianyi Bai, Jinhua Yu, Songyang Zhang, Dahua Lin, Conghui He, and Weijia Li. Urbench: A comprehensive benchmark for evaluating large multimodal models in multi-view urban scenarios. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 10707-10715, 2025. 3 +[56] Sijie Zhu, Taojiannan Yang, and Chen Chen. Vigor: Cross-view image geo-localization beyond one-to-one retrieval. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3640-3649, 2021. 2, 3, 4 \ No newline at end of file diff --git a/ICCV/2025/Where am I_ Cross-View Geo-localization with Natural Language Descriptions/images.zip b/ICCV/2025/Where am I_ Cross-View Geo-localization with Natural Language Descriptions/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..d0fc27e16472830bdb38639f6743a1885e877978 --- /dev/null +++ b/ICCV/2025/Where am I_ Cross-View Geo-localization with Natural Language Descriptions/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:110c8082f4e632041b0d381419a3548a67219a253e09ce573b4dbb848dfcaf59 +size 613360 diff --git a/ICCV/2025/Where am I_ Cross-View Geo-localization with Natural Language Descriptions/layout.json b/ICCV/2025/Where am I_ Cross-View Geo-localization with Natural Language Descriptions/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..a94d74831c81638411b1c680e0848d921e9e6608 --- /dev/null +++ b/ICCV/2025/Where am I_ Cross-View Geo-localization with Natural Language Descriptions/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:580efd5a2f2ec497939388b188987a22a969e5f626bc31b38706b2231cdfd7f9 +size 405867 diff --git a/ICCV/2025/Where, What, Why_ Towards Explainable Driver Attention Prediction/8087ca0e-98c5-4d7a-a52f-09d0f857ac52_content_list.json b/ICCV/2025/Where, What, Why_ Towards Explainable Driver Attention Prediction/8087ca0e-98c5-4d7a-a52f-09d0f857ac52_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..5a0b6290fd2538be70a5f0a3223969c4d7b03de0 --- /dev/null +++ b/ICCV/2025/Where, What, Why_ Towards Explainable Driver Attention Prediction/8087ca0e-98c5-4d7a-a52f-09d0f857ac52_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2320b42c2c9fffdea9d50320f92fd7229173808d22e2f7ba7219fac56d2b61ae +size 94834 diff --git a/ICCV/2025/Where, What, Why_ Towards Explainable Driver Attention Prediction/8087ca0e-98c5-4d7a-a52f-09d0f857ac52_model.json b/ICCV/2025/Where, What, Why_ Towards Explainable Driver Attention Prediction/8087ca0e-98c5-4d7a-a52f-09d0f857ac52_model.json new file mode 100644 index 0000000000000000000000000000000000000000..e74395eff89271e357af13482ea3fe7a3949d41f --- /dev/null +++ b/ICCV/2025/Where, What, Why_ Towards Explainable Driver Attention Prediction/8087ca0e-98c5-4d7a-a52f-09d0f857ac52_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9733f706465df2ca034d9326f1ac6f5021f8a01d5f286813612a6d8cfa412a68 +size 127423 diff --git a/ICCV/2025/Where, What, Why_ Towards Explainable Driver Attention Prediction/8087ca0e-98c5-4d7a-a52f-09d0f857ac52_origin.pdf b/ICCV/2025/Where, What, Why_ Towards Explainable Driver Attention Prediction/8087ca0e-98c5-4d7a-a52f-09d0f857ac52_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..d2f0d4d68613eb22bc86d64c2218aba27f0554ad --- /dev/null +++ b/ICCV/2025/Where, What, Why_ Towards Explainable Driver Attention Prediction/8087ca0e-98c5-4d7a-a52f-09d0f857ac52_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9300911d3ec2c814f8d0bce47eba6afb8da7708cc7b8b39e775c89f9191e9eed +size 1903079 diff --git a/ICCV/2025/Where, What, Why_ Towards Explainable Driver Attention Prediction/full.md b/ICCV/2025/Where, What, Why_ Towards Explainable Driver Attention Prediction/full.md new file mode 100644 index 0000000000000000000000000000000000000000..0e1dc44cee2f0ad34b34523098a6f50475300143 --- /dev/null +++ b/ICCV/2025/Where, What, Why_ Towards Explainable Driver Attention Prediction/full.md @@ -0,0 +1,379 @@ +# Where, What, Why: Towards Explainable Driver Attention Prediction + +Yuchen Zhou $^{1,2*}$ , Jiayu Tang $^{1*}$ , Xiaoyan Xiao $^{1}$ , Yueyao Lin $^{1}$ , Linkai Liu $^{1}$ , Zipeng Guo $^{1}$ , Hao Fei $^{2}$ , Xiaobo Xia $^{2}$ , Chao $\mathrm{Gou}^{1\dagger}$ + +$^{1}$ Sun Yat-sen University, $^{2}$ National University of Singapore + +https://github.com/yuchen2199/Explainable-Driver-Attention-Prediction + +# Abstract + +Modeling task-driven attention in driving is a fundamental challenge for both autonomous vehicles and cognitive science. Existing methods primarily predict where drivers look by generating spatial heatmaps, but fail to capture the cognitive motivations behind attention allocation in specific contexts, which limits deeper understanding of attention mechanisms. To bridge this gap, we introduce Explainable Driver Attention Prediction, a novel task paradigm that jointly predicts spatial attention regions (where), parses attended semantics (what), and provides cognitive reasoning for attention allocation (why). To support this, we present W3DA, the first large-scale explainable driver attention dataset. It enriches existing benchmarks with detailed semantic and causal annotations across diverse driving scenarios, including normal conditions, safety-critical situations, and traffic accidents. We further propose LLada, a Large Language model-driven framework for driver attention prediction, which unifies pixel modeling, semantic parsing, and cognitive reasoning within an end-to-end architecture. Extensive experiments demonstrate the effectiveness of LLada, exhibiting robust generalization across datasets and driving conditions. This work serves as a key step toward a deeper understanding of driver attention mechanisms, with significant implications for autonomous driving, intelligent driver training, and human-computer interaction. + +# 1. Introduction + +As an old proverb goes, "The eyes are the window to the soul." Human eyes serve as a crucial gateway to cognition, reflecting how human perceive, interpret, and interact with their environment [6, 54, 59, 60, 77, 79, 81]. Modeling and understanding human visual attention provide fundamental insights into cognitive abilities, experiential knowledge, + +![](images/401705f8337b15443ac7ebf7040db7d46f6e72744c23aba27dfb2eac35e217b3.jpg) +Figure 1. Prior studies focus only on predicting where drivers allocate their attention, offering an implicit and shallow representation of attention. In contrast, our work proposes a more explicit and in-depth paradigm that answers where to attend, what to attend, and why attention is allocated there, revealing underlying factors behind attention allocation such as traffic rule knowledge, driving safety perception, and goal-driven behaviors. Our work unlocks a deeper understanding of driver attention mechanisms, providing new insights to the research community. + +and decision-making strategies, especially in complex, task-driven scenarios such as driving [8, 28, 32, 33, 53, 80, 82]. + +In recent years, driver attention prediction has emerged as a key research direction and has achieved remarkable progress [22, 27, 50, 51, 64]. Existing paradigms [10, 20, 31, 78, 84, 85] primarily focus on predicting where drivers look by generating spatial heatmaps, as shown in Fig. 1(a). These methods fundamentally perform pixel-space regression, offering only a shallow and implicit representation of attention mechanisms. However, these approaches fail to capture the underlying cognitive motivations behind attention allocation in specific driving contexts. For instance, in the scenario depicted in Fig. 1, a driver attends to the red traffic light to comply with traffic rules, monitors an + +oncoming cyclist at an intersection to ensure driving safety, and observes the road ahead to navigate toward the intended destination. Unfortunately, no existing work has effectively modeled these semantic-level (i.e., what is being attended to) and cognitive-level (i.e., why attention is allocated) aspects of driver attention. This critical limitation hinders both a deeper understanding of driver attention mechanisms and the real-world applicability of these models. + +To bridge these critical gaps, we propose a novel unified task paradigm, termed Explainable Driver Attention Prediction, as shown in Fig. 1(b). This paradigm requires models to simultaneously predict pixel-level attention regions (answering where), provide semantic-level parsing (answering what), and perform cognitive-level reasoning to explain the underlying causes of attention allocation (answering why). By integrating spatial, semantic, and cognitive knowledge, this enables a more comprehensive and explainable modeling of driver attention. + +Specifically, our work introduces several key advancements over existing driver attention prediction methods: (1) We introduce W3DA, a large-scale explainable driver attention dataset, which integrates multiple mainstream driver attention benchmarks and incorporates additional annotations for semantic parsing (what) and cognitive reasoning (why). It provides spatial, semantic, and cognitive labels for attention regions, explicitly enabling where, what, and why reasoning in driver attention modeling. (2) We propose LLada, a Large language model-driven framework for driver attention prediction, which is the first unified architecture capable of jointly predicting attention regions (where), semantic interpretations (what), and cognitive reasoning (why) in driver attention modeling. (3) Unlike existing models trained and evaluated separately on individual datasets, we leverage the W3DA dataset for cross-domain collaborative training, enabling a unified modeling approach for normal driving, safety-critical situations, and traffic accidents. + +Extensive experiments validate the effectiveness of LLada, demonstrating its superior performance over all state-of-the-art (SOTA) methods on $\mathrm{W}^{3}\mathrm{DA}$ in both attention map prediction and textual explanation generation. It consistently excels across various driving scenarios. Moreover, trained solely on $\mathrm{W}^{3}\mathrm{DA}$ , LLada outperforms most independently trained specialized models across the entire DR(eye)VE [51], BDDA [64], and DADA-2000 [19] datasets, achieving highly competitive performance and demonstrating strong generalization. Further qualitative analyses confirm its strong alignment with human driver attention. Our work establishes a solid foundation for comprehensive and explainable attention modeling, paving the way for advancements in cognitive science and real-world applications such as interpretable autonomous driving and intelligent driver training. Before delving into details, we + +clearly emphasize our contributions as follows: + +- (Paradigm) We introduce Explainable Driver Attention Prediction, a novel paradigm that extends traditional driver attention prediction by jointly reasoning over where, what, and why to achieve a more comprehensive understanding of attention in driving. +- (Dataset) We present W $^3$ DA, a large-scale explainable driver attention dataset, which enhances existing benchmarks with semantic (what) and causal (why) annotations, spanning normal, safety-critical, and accident scenarios. +- (Methodology) We propose LLada, the first large language model-driven driver attention prediction framework, enabling end-to-end training and joint prediction of attention regions (where), semantic parsing (what), and cognitive reasoning (why). +- (Experiments) Extensive experiments demonstrate that LLada consistently outperforms existing SOTA methods across multiple tasks, scenarios, metrics, and settings, demonstrating superior robustness and generalization. + +# 2. Related Work + +# 2.1. Driver Attention Prediction + +Datasets. Specialized driver attention datasets have significantly advanced the field. Early works like DR(eye)VE [51] introduced a pipeline to analyze driver gaze in normal traffic, recording attention data from eight drivers. Similarly, the LBW dataset [31] provided gaze data collected from 28 drivers in real driving conditions. Cheng et al. introduced IVGaze [11], which collected in-vehicle gaze data but lacked corresponding driving footage. The BDD-A dataset by Xia et al. [64] focused on safety-critical events like emergency braking and traffic congestion. DADA2000 [19] extended this to traffic accidents, including 2,000 videos and corresponding gaze data. While existing datasets have provided large-scale driver attention data, they primarily associate spatial heatmaps with driving scenes without capturing the cognitive reasoning behind attention allocation, which limits a deeper understanding of driver attention mechanisms. To address this, we introduce W $^3$ DA, the first large-scale dataset for explainable driver attention modeling. W $^3$ DA extends beyond heatmaps by providing fine-grained annotations that link attention heatmaps to their semantic meaning and underlying causes, offering unprecedented insights into driver attention mechanisms. + +Models. Attention prediction models have evolved from early image processing techniques like ITTI [29] and GBVS [23] to deep learning-based architectures. CNNs became fundamental, enabling richer visual feature representations. Palazzi et al. [51] integrated RGB, optical flow, and segmentation, while Xia et al. [64] employed convolutional LSTMs. Deng et al. [14] proposed a fully convolutional encoder-decoder. Multi-resolution networks [25] and + +heterogeneous networks [26] further incorporated multilevel visual content. ASIAFNet [37] emphasized semantic-based object-level attention. Beyond CNNs, Inverse Reinforcement Learning [3] and Generative Adversarial Networks [2, 3] have been explored to model driver attention. Amadori et al. [1] incorporated driving inputs and vehicle states in VR simulations. Unsupervised methods [85] explored scene dynamics without reliance on labeled attention data. Shi et al. [57] proposed FOD-Net with saliency guidance. Recent advances integrate Transformers [18] to enhance global context modeling. ACT-Net [21] pioneered CNN-Transformer fusion, followed by FBLNet [10], which introduces a feedback loop mechanism for experience accumulation, further bridging attention modeling with human-like learning. Despite these advancements, existing models still fail to explicitly capture the cognitive motivations behind driver attention. To address this limitation, we propose LLada, a unified end-to-end framework that jointly models spatial (where), semantic (what), and cognitive (why) factors, enabling a more comprehensive and explainable approach to driver attention prediction. + +# 2.2. Multimodal Large Language Models + +Recent advancements in multimodal large language models (MLLMs) [38, 39, 42, 43, 47-49, 63, 69, 71, 72, 76, 83] have driven significant progress across image-level captioning [7], object-level detection [74], and pixel-level segmentation [35, 65], with notable applications in autonomous driving [13, 58, 62, 75] and robotics [40, 45, 73]. Building on this foundation, recent studies have explored incorporating human attention as an additional modality to enhance model performance and interpretability. Voila-A [68] aligned MLLMs with the human gaze (captured via AR/VR devices), ensuring better attention correspondence between models and users. GazeXplain [9] predicted visual scanpaths and generated natural language explanations for fixations using BLIP [36]. However, its explanations remained semantic-level parsing, describing what is attended to without explicitly uncovering the underlying cognitive motivations behind attention allocations. In contrast, LLada introduces a unified framework that jointly models spatial (where), semantic (what), and cognitive (why) factors, enabling a more comprehensive and explainable approach to driver attention modeling. To the best of our knowledge, this is the first attempt to model underlying cognitive motivations of attention using MLLMs. + +# 3. W³DA Dataset + +Existing driving attention datasets primarily predict where drivers look via spatial heatmaps. However, they lack semantic and cognitive understanding, limiting their ability to comprehensively model task-driven visual attention in driving. To bridge this gap, we move beyond spatial attention + +prediction to a deeper, more explainable understanding of driver attention. A key challenge is the absence of a dataset that provides semantic parsing (what) and cognitive reasoning (why) for driver attention. + +To address this, we introduce W3DA, a semantic and reasoning-aware driving attention dataset that enhances conventional where annotations with high-quality what and why labels. As shown in Fig. 2(a), W3DA unifies multiple mainstream driving attention datasets, integrating DR(eye)VE [51] and LBW [31] for normal driving, BDDA [64] for critical situations, and DADA-2000 [19, 20] for traffic accidents. This diverse data composition allows for more robust cross-domain training, improving model generalization and real-world applicability. Furthermore, as depicted in Fig. 2(b-d), we introduce an efficient and highly reliable data annotation pipeline that assigns semantic labels to attention targets (what) and provides context-aware cognitive reasoning for attention causes (why), enabling a more comprehensive understanding of driver attention. + +# 3.1. Data Annotation Pipeline + +Creating a large-scale, high-quality dataset for explainable driver attention prediction presents several inherent challenges. First, while vast amounts of driver attention data exist, many driving scenarios—such as cruising, highway driving, or following another vehicle—exhibit stable attention patterns, with drivers primarily fixating on the preceding vehicle or the road's vanishing point. In such cases, attention shifts are minimal, making frame-by-frame annotation inefficient and resource-intensive. Secondly, manual annotation by human experts incurs substantial time and financial costs. While leveraging powerful MLLMs for automated annotation is a promising alternative, MLLMs are prone to hallucinations, often generating erroneous or contextually inappropriate annotations that compromise dataset reliability. To address these challenges, we introduce an efficient and highly reliable semi-automatic pipeline for dataset annotation. + +For the first challenge, we introduce an attention-aware key sample selection strategy that avoids redundant frame-by-frame annotation. This approach prioritizes identifying key moments indicative of significant attention shifts by evaluating perceptual and attention-related changes between consecutive samples, ensuring the annotations capture high-information moments. Our selection process is grounded in three key criteria: (1) semantic similarity of driving scenes, capturing perceptual variations in the environment; (2) spatial divergence of driver attention, capturing variations in the spatial distribution of attended regions; and (3) semantic similarity of driver attention, reflecting contextual shifts in attended regions. Specifically, criteria (1) and (3) are computed using the CLS token from the CLIP-Large image encoder [56], while (2) is evaluated + +![](images/831ea9471878e7b98b74f25918e9c9631822173a146e0f1175c0e4e4922968a5.jpg) +Figure 2. The W3DA annotation pipeline. (a) W3DA integrates multiple driver attention benchmarks, covering normal driving, safety-critical scenarios, and accidents. (b) To reduce annotation redundancy in scenarios with stable attention patterns (e.g., cruising, following another vehicle), we propose a key sample selection method, leveraging semantic scene similarity, spatial attention divergence, and semantic attention similarity to identify key moments of significant attention shifts. (c) We employ visual and contextual prompts with a structured chain-of-thought to guide an advanced MLLM in generating preliminary context-aware annotations. (d) Finally, human experts validate and refine these annotations to ensure accuracy, contextual relevance, and alignment with real driving scenarios. + +via KL divergence between attention heatmaps. To implement this, we initialize the key sample set $K$ with the first sample from the original driving video $V \in \mathbb{R}^{T \times H \times W \times 3}$ . For each subsequent sample in $V$ , we evaluate these three metrics relative to the most recent key sample in $K$ . A sample is added to $K$ if the KL divergence (criterion 2) exceeds a predefined threshold or if the semantic similarity metrics (criteria 1 and 3) fall below their respective thresholds. Otherwise, it is considered redundant and excluded from further consideration. + +For the second challenge, we propose an effective approach that integrates visual and contextual prompts with a clever chain-of-thought strategy to guide an advanced MLLM in generating preliminary annotations that are both useful and contextually relevant. Specifically, we provide the MLLM with comprehensive contextual cues, including driving behavior, weather conditions, and scene categories, encouraging it to generate annotations aligned with the current scenario. Additionally, we overlay attention maps onto images using a simple grayscale mask, directing the MLLM's focus toward the relevant regions where driver attention is concentrated. We further employ a chain-of-thought to guide the MLLM through a progressive analysis of the semantics and underlying causes of driver attention. This process prompts the MLLM first to determine the number of attention regions within the scene, then identify the content within each region (answering what), and finally explore the causal factors behind the attention allocation (answering why). In practice, we utilize the Qwen-VL-Max API as our MLLM to implement this approach. + +Despite the MLLM's strong visual reasoning capabilities, hallucinations remain a potential issue. To address this, + +we involve human experts to review and refine the model's outputs based on their domain expertise, ensuring the final annotations are accurate, reasonable, and contextually aligned with the driving scenario. + +# 3.2. Dataset Statistics + +Building on our pipeline, W³DA comprises 69,980 key samples extracted from 3,548 video scenes, covering diverse driving conditions. Specifically, it includes 22,839 samples from normal driving, 22,950 samples from critical situations, and 24,191 samples from accidents. W³DA spans a variety of weather conditions (sunny, rainy, cloudy, overcast, snowy, foggy), geographic locations (urban, rural, highway, suburban, mountain, tunnel), and time periods (morning, afternoon, night). Furthermore, W³DA provides rich textual annotations: the average length of what annotations is 5.40 words, while the average length of why annotations is 24.32 words, offering fine-grained semantic and cognitive insights into driver attention. To maintain a rigorous experimental setup, we adhere to the original training-validation-test splits of the four source datasets, ensuring no data leakage. W³DA advances driver attention modeling, offering valuable insights for both academia and industry. More details can be found in the supplementary material. + +# 4. Explainable Driver Attention Prediction + +# 4.1. Problem Definition + +The goal of Explainable Driver Attention Prediction is to predict where a driver is likely to attend, what regions are attended, and why attention is allocated. This task goes beyond conventional attention prediction by unifying + +spatial, semantic, and causal reasoning, enabling a more explainable and cognition-driven understanding of driver attention. Formally, given a driving scene represented as an image $\mathbf{I} \in \mathbb{R}^{H \times W \times C}$ with contextual information $\mathbf{C}$ (e.g., weather, road conditions, driving behavior), the goal is to estimate a pixel-wise attention map $\mathbf{A} \in \mathbb{R}^{H \times W}$ , a set of semantic descriptions of attended regions $S = \{s_i\}_{i=1}^N$ and a set of causal explanations for attention allocation $\mathcal{E} = \{e_i\}_{i=1}^M$ , where $s_i$ denotes the semantic label of the $i$ -th attended region, and $e_i$ represents the explanation behind the attention allocated to the $i$ -th region. + +# 4.2. LLada Architecture + +LLada comprises four core components: 1) a pretrained visual encoder $\mathcal{F}_{\mathrm{vis}}$ , 2) a large language model $\mathcal{F}_{\mathrm{LLM}}$ , 3) a special attention token [ATTN], and 4) a cognitive-aware attention decoder $\mathcal{F}_{\mathrm{dec}}$ , as illustrated in Fig. 3. Given a driving scene, the visual input $\mathbf{I}$ and contextual information $\mathbf{C}$ are first processed by the pretrained visual encoder $\mathcal{F}_{\mathrm{vis}}$ and the LLM $\mathcal{F}_{\mathrm{LLM}}$ , respectively. Then, the attention token [ATTN] encodes high-level cognitive cues within $\mathcal{F}_{\mathrm{LLM}}$ and enables interaction between the language and vision modalities through the attention decoder $\mathcal{F}_{\mathrm{dec}}$ . Finally, $\mathcal{F}_{\mathrm{dec}}$ generates the spatial attention map $\mathbf{A}$ , while the $\mathcal{F}_{\mathrm{LLM}}$ outputs the semantic descriptions $\mathcal{S}$ and causal explanations $\mathcal{E}$ . Next, we provide a detailed description of each component. + +Visual Encoder. We utilize a visual encoder consisting of the pretrained vision foundation model CLIP-ViT-L [56] and a linear projector $\phi$ , following the setup of LLaVA [42]. Given the visual input $\mathbf{I}$ , CLIP extracts image features, which are then projected by $\phi$ into the visual token space, producing embeddings $\mathbf{h}_{\mathrm{vis}} \in \mathbb{R}^{N \times d}$ for further processing by $\mathcal{F}_{\mathrm{LLM}}$ . + +Large Language Model. The driving scene context $\mathbf{C}$ is tokenized into context tokens by the LLM tokenizer $\mathcal{T}$ : $\mathbf{h}_{\mathrm{con}} = \mathcal{T}(\mathbf{C})$ . The visual tokens $\mathbf{h}_{\mathrm{vis}}$ and context tokens $\mathbf{h}_{\mathrm{con}}$ are then concatenated with a fixed prompt sequence $\mathbf{h}_{\mathrm{prompt}}$ and fed into the LLM, denoted as $\mathcal{F}_{\mathrm{LLM}}$ . The output token embeddings $\tilde{\mathbf{y}}_{\mathrm{txt}}$ are generated as: + +$$ +\tilde {\mathbf {y}} _ {\mathrm {t x t}} = \mathcal {F} _ {\mathrm {L L M}} ([ \mathbf {h} _ {\text {p r o m p t}}, \mathbf {h} _ {\text {v i s}}, \mathbf {h} _ {\text {c o n}} ]), \tag {1} +$$ + +where $[\cdot ]$ denotes concatenation. The final text responses are generated by applying a linear classifier to $\tilde{\mathbf{y}}_{\mathrm{xt}}$ for next-. token prediction. During attention prediction, these responses include semantic descriptions of the attended regions $S$ and causal explanations for attention allocation $\mathcal{E}$ In practice, we use Vicuna-7B [12] as the $\mathcal{F}_{\mathrm{LLM}}$ + +Attention Token. To adapt the LLM for attention prediction, we introduce a special attention token, [ATTN], extending the original LLM vocabulary. This token encodes high-level cognitive cues within the LLM and facilitates cross-modal interaction between language and vision + +via $\mathcal{F}_{\mathrm{dec}}$ . During attention prediction, the output token sequence $\tilde{\mathbf{y}}_{\mathrm{txt}}$ includes the [ATTN] token. We extract its corresponding embedding $\tilde{\mathbf{y}}_{\mathrm{txt}}[\mathrm{ATTN}]$ , which is then projected into the $\mathcal{F}_{\mathrm{dec}}$ space via an MLP projector $\psi$ : + +$$ +\mathbf {h} _ {\text {a t t n}} = \psi \left(\tilde {\mathbf {y}} _ {\mathrm {t x t}} [ \mathrm {A T T N} ]\right). \tag {2} +$$ + +This enables $\mathcal{F}_{\mathrm{dec}}$ to decode high-level cognitive attention information from $\mathbf{h}_{\mathrm{attn}}$ . + +Cognitive-aware Attention Decoder. The attention decoder $\mathcal{F}_{\mathrm{dec}}$ facilitates the interaction between the attention token embedding $\mathbf{h}_{\mathrm{attn}}$ and the visual features $\mathbf{h}_{\mathrm{vis}}$ , decoding context-aware cognitive information to generate the pixelwise attention map $\mathbf{A} \in \mathbb{R}^{H \times W}$ . Specifically, we first introduce a cross-attention mechanism to guide the decoder's awareness of high-level cognitive cues from $\mathbf{h}_{\mathrm{attn}}$ and derive cognitively-driven visual features $\mathbf{h}_{\mathrm{vis}}'$ : + +$$ +\mathbf {h} _ {\mathrm {d e c}} ^ {\prime} = \mathbf {h} _ {\mathrm {v i s}} + \operatorname {R e p a t} \left(C A \left(\mathbf {h} _ {\mathrm {a t t n}}, \mathbf {h} _ {\mathrm {v i s}}\right)\right), \tag {3} +$$ + +where $CA(q, kv)$ denotes the cross-attention operation and $\mathrm{Repeat}(\cdot)$ is a replication operation, which adds $CA(\mathbf{h}_{\mathrm{attn}}, \mathbf{h}_{\mathrm{vis}})$ to each token in $\mathbf{h}_{\mathrm{vis}}$ . Next, the $\mathbf{h}_{\mathrm{vis}}'$ are reshaped from a 2D shape of $\frac{HW}{p^2} \times C$ to a standard 3D feature map of size $\frac{H}{p} \times \frac{W}{p} \times C$ , where $C$ is the feature channel size and $p$ is the image patch size. Then, we apply a series of five $3 \times 3$ convolutional layers with batch normalization and ReLU activation to reduce the feature dimensionality to $\frac{H}{p} \times \frac{W}{p}$ . Finally, a bilinear upsampling operation is applied to generate the attention map $\mathbf{A}$ at the full image size of $H \times W$ . + +Training Objectives. The LLada model is trained end-to-end using the attention map prediction loss $\mathcal{L}_{\mathrm{map}}$ and the textual explanation generation loss $\mathcal{L}_{\mathrm{txt}}$ . The overall objective $\mathcal{L}$ is the weighted sum of these losses: + +$$ +\mathcal {L} = \lambda_ {\text {m a p}} \mathcal {L} _ {\text {m a p}} + \lambda_ {\text {t x t}} \mathcal {L} _ {\text {t x t}}, \tag {4} +$$ + +where $\lambda_{\mathrm{map}}$ and $\lambda_{\mathrm{txt}}$ are scaling factors. Specifically, the attention map loss $\mathcal{L}_{\mathrm{map}}$ consists of binary cross-entropy (BCE) and Kullback-Leibler (KL) divergence, encouraging accurate pixel-wise attention maps (answering Where?): + +$$ +\mathcal {L} _ {\text {m a p}} = \lambda_ {\mathrm {b c e}} \mathbf {B C E} (\hat {\mathbf {A}}, \mathbf {A}) + \lambda_ {\mathrm {k l}} \mathbf {K L} (\hat {\mathbf {A}}, \mathbf {A}), \tag {5} +$$ + +where $\hat{\mathbf{A}}$ is the predicted attention map and $\mathbf{A}$ is the ground truth. The textual explanation generation loss $\mathcal{L}_{\mathrm{txi}}$ is formulated as the autoregressive cross-entropy (CE) loss, guiding the LLM to generate correct semantic descriptions (answering What?) and causal explanations (answering Why?). of the attended regions: + +$$ +\mathcal {L} _ {\mathrm {t x t}} = \lambda_ {\text {w h a t}} \mathbf {C E} (\hat {\mathcal {S}}, \mathcal {S}) + \lambda_ {\text {w h y}} \mathbf {C E} (\hat {\mathcal {E}}, \mathcal {E}), \tag {6} +$$ + +where $\hat{S}$ and $\hat{\mathcal{E}}$ are the predicted semantic descriptions and causal explanations, respectively. That $S$ and $\mathcal{E}$ are the ground truth descriptions and explanations. + +![](images/f5149336db24d26e493f419e04dd695ea9db06cb22baf2fa087360399b1298ef.jpg) +Figure 3. Overall architecture of LLada. Given a driving scene represented as an image with contextual information, the image is processed by the visual encoder (CLIP encoder and projector) to extract visual tokens, while the contextual information, along with prompt sequences, is tokenized by the LLM tokenizer. The visual and text tokens are then fed into the LLM. To adapt the LLM for explainable attention prediction tasks, a special attention token [ATTN] is introduced to encode high-level cognitive cues. Finally, [ATTN] performs cross-attention with the visual tokens, generating the attention map (answering Where?) through cognitive-aware attention decoder, while the LLM produces the semantic descriptions (answering What?) and causal explanations (answering Why?) of attended regions. + +# 5. Experiment + +# 5.1. Experimental Setting + +Implementation Details. We train our model on 4 NVIDIA A100 GPUs, utilizing the DeepSpeed engine for efficient optimization. The training process employs the AdamW [46] optimizer with a learning rate of 0.0003, and WarmupDecayLR as the learning rate scheduler with 100 warmup iterations. For loss balancing, we set the scaling factors $\mathcal{L}_{\mathrm{map}}$ , $\mathcal{L}_{\mathrm{txt}}$ , $\mathcal{L}_{\mathrm{bce}}$ , $\mathcal{L}_{\mathrm{kl}}$ , $\mathcal{L}_{\mathrm{what}}$ , and $\mathcal{L}_{\mathrm{why}}$ to 2, 1, 1, 0.1, 1, and 1, respectively. Training is conducted with a batch size of 8 per device, using gradient accumulation with a step size of 5 to accommodate memory constraints. To retain the knowledge in the pre-trained LLM $\mathcal{F}_{\mathrm{LLM}}$ , we apply LoRA [24] for parameter-efficient fine-tuning while keeping the visual encoder $\mathcal{F}_{\mathrm{vis}}$ entirely frozen. The cognitive-aware attention decoder $\mathcal{F}_{\mathrm{dec}}$ is trained from scratch. + +Evaluation Metrics. We follow the evaluation setup of prior driver attention prediction studies [10, 20] and assess the spatial accuracy of generated attention maps using $\mathrm{AUC_J}$ , $\mathrm{AUC_B}$ , SIM, CC, KLdiv, and NSS. To evaluate the language quality of the generated what and why textual explanations, we adopt BLEU, METEOR, ROUGE, and CIDEr-R. This comprehensive evaluation allows us to effectively measure the model's performance in spatial, semantic, and cognitive reasoning for driver attention. + +# 5.2. Results on W3DA + +Attention Map Prediction (Where). We first evaluate the attention map prediction quality on the W3DA dataset, comparing our proposed LLada with 10 state-of-the-art methods: ITTI [29], GBVS [23], DeepGaze I [34], DeepGaze + +IIE [41], MLNet [17], CDNN [14], FBNet [16], ConvNeXt [44], ERFNet [70], and GazeXplain [9]. ITTI and GBVS are early saliency models based on handcrafted features, while the others are learning-based approaches. DeepGaze I and DeepGaze IIE use official pre-trained models due to unavailable training codes. MLNet, CDNN, FBNet, ConvNeXt, and ERFNet are specialized attention prediction models optimized solely for this task, whereas GazeXplain and LLada jointly optimize attention map prediction and textual explanation generation. Table 1 presents the performance comparison. LLada surpasses all SOTA methods in every scenario. In particular, LLada achieves substantial gains on the KLdiv, $\mathrm{AUC_J}$ , and $\mathrm{AUC_B}$ metrics. For KLdiv, it outperforms the second-best method (ERFNet) by $38.40\%$ , $22.79\%$ , and $11.65\%$ in normal driving, safety-critical situations, and traffic accidents, respectively. Compared to GazeXplain, which jointly optimizes attention map prediction (where) and textual explanation generation (what & why), LLada consistently achieves superior performance across all metrics, demonstrating its effectiveness. + +Textual Explanation Generation (What & Why). We further evaluate the quality of textual explanations on the W3DA dataset, comparing our proposed LLada with the multi-task model GazeXplain. To ensure a comprehensive evaluation, we also establish strong baseline methods to validate LLada's effectiveness. Unlike our end-to-end approach, these baselines adopt a two-stage paradigm: first, a conventional attention prediction model generates an attention map; second, an MLLM interprets the attention map to generate textual explanations for what and why. Specifically, we select DeepGaze I, DeepGaze IIE, MLnet, CDNN, + +Table 1. Comparison of attention map prediction performance. Bold and underline indicate the best and second-best results. † marks multi-task models jointly optimizing Attention Map Prediction (where) and Textual Explanation Generation (what/why). + +
MethodNormal DrivingSafety-Critical SituationTraffic Accident
KLdiv↓CC↑SIM↑AUC1↑AUCB↑NSS↑KLdiv↓CC↑SIM↑AUC1↑AUCB↑NSS↑KLdiv↓CC↑SIM↑AUC1↑AUCB↑NSS↑
ITTI [29]3.2160.0930.0800.6760.6650.6622.8070.0490.1190.6180.6130.4103.3390.0330.0730.5950.6000.266
GBVS [23]2.5720.2940.1390.8680.8572.0572.2380.2460.1760.8390.8041.5012.8260.1730.1050.8140.7651.212
DeepGaze I [34]3.1030.2070.0800.8850.5461.2542.5500.2370.1370.8740.5951.2973.0600.2060.0820.8830.6731.412
DeepGaze IIE [41]3.0710.2260.0820.8870.6101.3822.5050.2960.1410.9260.5891.6413.0260.2270.0850.9170.5841.476
MLnet [17]2.1290.5470.4600.9140.8364.4941.9530.5280.4330.9280.8745.2542.8970.3440.2880.8930.7843.008
CDNN [14]2.6140.4650.3940.8870.7903.6872.6460.4010.3500.8850.7673.8313.5680.2830.2540.8510.7142.446
FBnet [16]2.9800.4060.3430.8690.7693.1032.5850.4310.3640.8870.8033.9753.1970.3290.2460.8730.7892.727
ConvNeXt [44]2.0420.5700.4120.9160.8484.6611.7650.5670.4130.9380.8775.4383.0490.3770.2480.8910.8063.425
ERFNet [70]1.9790.5580.4250.9230.8404.3041.5930.5380.4100.9420.8685.2012.1810.3910.2530.9300.8463.035
GazeXplain† [9]2.5780.4770.3890.8570.8663.9452.7690.3830.3210.8480.7432.2993.1090.3710.2360.9020.8041.598
LLada† (Ours)1.2190.5830.4360.9520.9085.3761.2300.5790.4200.9500.9125.2711.9270.3960.2620.9340.8893.216
+ +Table 2. Comparison of textual explanation generation performance. Bold and underline show the best and second-best performances. + +
MethodNormal DrivingSafety-Critical SituationTraffic Accident
BLEU↑METEOR↑ROUGE↑CIDEr-R↑BLEU↑METEOR↑ROUGE↑CIDEr-R↑BLEU↑METEOR↑ROUGE↑CIDEr-R↑
DeepGaze I [34] + LLaVA [42]0.2810.2740.3570.2870.1300.1780.2690.0780.1790.2130.2910.167
DeepGaze IIE [41] + LLaVA [42]0.2810.2750.3610.3010.2530.2830.3460.2890.2530.2660.3440.272
MLnet [17] + LLaVA [42]0.2940.2830.3980.2790.2850.2940.3950.3120.2580.2640.3730.260
CDNN [14] + LLaVA [42]0.2990.2870.4030.2830.2850.2940.3970.3120.2630.2700.3770.265
ConvNeXt [44] + LLaVA [42]0.2910.2820.3960.2680.1170.1730.2680.0780.2250.2470.3450.209
ERFNet [70] + LLaVA [42]0.2940.2830.3960.2740.2870.2970.3980.3070.2700.2720.3750.258
GazeXplain [9]0.3070.2990.2160.4220.1850.2900.3670.5540.1670.1950.4380.656
LLada (Ours)0.4360.3600.5820.9630.4440.3750.5931.2330.3760.3180.5201.002
+ +ConvNeXt, and ERFNet as first-stage predictors. For the second stage, we fine-tune LLaVA [42] (CLIP-ViT-L [56] + Vicuna-7B [12]) via LoRA [24], using the same MLLM as LLada to process grayscale attention maps for textual explanation generation. As shown in Table 2, LLada consistently outperforms all baselines across scenarios and metrics, surpassing both GazeXplain and all two-stage baselines, which demonstrates its superior cognitive reasoning capabilities in driver attention analysis. The results from Table 1 (Attention Map Prediction) and Table 2 (Textual Explanation Generation) collectively confirm that our LLada surpasses all previous methods, showcasing its capability for a deeper understanding of driver attention mechanisms. + +# 5.3. Results on Conventional Attention Prediction + +To evaluate the effectiveness of LLada trained on the cross-domain dataset $\mathrm{W}^3\mathrm{DA}$ , we compare it with domain-specific attention prediction models fully trained on DR(eye)VE [51], BDDA [64], and DADA [19, 20]. Performance is evaluated on the complete test sets of these datasets. As shown in Table 3, LLada demonstrates competitive performance, particularly in terms of KLdiv, outperforming the second-best domain-specific models by $29.81\%$ , $20.69\%$ , and $5.49\%$ on the DR(eye)VE, BDDA, and DADA test sets, respectively. Additionally, GazeXplain, trained with cross-domain multi-task learning on $\mathrm{W}^3\mathrm{DA}$ , also achieves promising results. Notably, $\mathrm{W}^3\mathrm{DA}$ provides 39,642 keyframes for training, while domain-specific models use 28,632 (DR(eye)VE), 26,325 (BDDA), and 33,939 + +(DADA) frames. Despite the cross-domain setting, training LLada on $\mathrm{W}^{3}\mathrm{DA}$ does not significantly increase cost, highlighting its strong generalization across domains and the efficiency of key-frame selection in $\mathrm{W}^{3}\mathrm{DA}$ . + +# 5.4. Qualitative Analyses + +We quantitatively compare our LLada with the latest GazeXplain [9] on explainable driver attention prediction, as shown in Fig. 4. LLada's pixel-wise attention map and textual explanations align more closely with the Ground Truth, capturing attention mechanisms across spatial, semantic, and cognitive levels. In complex scenarios, such as busy traffic flow (Row 2) and pedestrian interactions (Row 3), LLada outperforms GazeXplain. LLada correctly focuses on the pedestrian in front of the vehicle (Row 3) and provides context-aware, task-driven explanations like "assessing the pedestrian's movement and speed to avoid a potential collision." In contrast, GazeXplain misses this key focus and offers less relevant, scene-independent explanations. See supplementary materials for more examples. + +# 5.5. Ablation Study + +To validate the effectiveness of LLada in jointly reasoning over where, what, and why, we conduct ablation studies. First, we remove the where task (attention map prediction) to assess its impact on what and why (textual explanation generation). As shown in Fig. 5, removing where causes a significant performance drop in all textual explanation metrics. This demonstrates that where plays a crucial role in + +Table 3. Comparison with domain-specific models for attention map prediction. Bold and underline denote the top two performances. +(a) DR(eye)VE test set. + +
TrainingMethodKLdiv↓CC↑
DR(eye)VEMLNet [17]2.000.44
RMDN [5]1.770.41
DR(eye)VE [51]1.400.56
HWS [64]1.720.51
SCAFNet [20]1.350.59
DVAM [15]1.380.57
TransConvNet [67]1.370.58
MTSF [30]1.720.56
WDAGazeXplain1.480.49
LLada1.040.67
+ +(b) BDDA test set. + +
TrainingMethodKLdiv↓CC↑SIM↑
BDDAHWS [64]2.070.480.35
U2NET [55]1.470.560.36
MINET [52]10.500.490.36
DRIVE [4]13.830.320.26
SCAFNet [20]1.480.560.40
DBNET [61]1.850.560.38
FBLNet [10]1.400.640.47
MTSF [30]1.610.510.36
W3DAGazeXplain2.850.380.32
LLada1.160.600.47
+ +(c) DADA test set. + +
TrainingMethodKLdiv↓CC↑SIM↑
DADAHWS [64]2.770.330.22
U2NET [55]1.860.470.30
MINET [52]9.990.390.30
DRIVE [4]4.030.370.25
SCAFNet [20]2.170.480.34
DBNET [61]2.770.400.26
PGNET [66]5.280.460.37
FBLNet [10]1.920.500.33
W3DAGazeXplain1.930.460.29
our LLada1.820.480.36
+ +![](images/a0f157988ebe6c1ba9d664270c6d6b4a8975dfbc1cca424646de7fa70453e84f.jpg) +GazeXplain + +(What?) Attention Semantics + +[Road ahead] + +(Why?) Attention Causes +__________ +[To monitor the path and any potential obstacles or changes in the road conditions that may require a pavement]. + +![](images/f86673f22f25c3dab89270e64df54e02ee8f4d3975e10a9af3bebfbcfa3ef562.jpg) +LLada (Ours) + +(What?) Attention Semantics +[Road ahead and side road] +(Why?) Attention Causes +[To be aware of any potential entry points to the highway, as +indicated by the side road and signage, and prepare for any + +![](images/61de55be83c9728e6b3a377e6d8c05ed97be62e2fd2f7544da1487cdcd461b5e.jpg) +Ground Truth + +(What?) Attention Semantics + +[Road ahead and highway entrance with signage] +(Why?) Attention Causes: +[To be aware of the highway entrance and any indications of +signage , and prepare for necessary lone changes or stops tosafely merge onto the highway . + +![](images/fed131ce8ce3ea3b8934f7a161bc1b071ff3e6cfffa24ab6277f8ae9ed2e8a4c.jpg) + +(What?) Attention Semantics +[Road ahead] [Left side of the road] [Right side of the road] +(Why?) Attention Causes +[To monitor the path and any potential obstacles or changes in the road conditions that may require a pavement.] +To be aware of vehicles, pedestrians, or other elements on the left +side that could impact driving decisions. +[To check for parked cars, moving vehicles, or pedestrians on therightside that might enter the driver ’ s lane or effect navigation ] + +![](images/162b407c61dac4153bd2b1f1fec482b5bf34bcbfb49fcdd7296509402c14ce9a.jpg) + +(What?) Attention Semantics +[Road ahead and vehicle in front] +(Why?) Attention Causes +[To maintain a safe following distance from the vehicle ahead] +while monitoring the road conditions to detect any potential obstacles or pedestrians, ensuring safe driving. + +![](images/ad81c46fdcfda7080bcf22fb587bbd7891fcea4cbc34e24ac887b2b747d6487e.jpg) + +(What?) Attention Semantics + +[Vehicle directly in front and pedestrian crossing] + +(Why? Attention Causes + +[To maintain a safe following distance from the vehicle ahead] + +and ensure the pedestrian crosses the road safely, adjusting speed as needed to avoid collisions. + +![](images/0cdebc404bd70e29c23af0ebab94cc5035505605b40bda4245700a69be9cf50d.jpg) +Figure 4. Quantitative examples. More in supplementary materials. + +(What?) Attention Semantics +[Pedestrian crossing] +(Why?) Attention Causes +To assess the presence and movement of The pedestrian who is crossing or about to cross, ensuring safe navigation through the area. + +![](images/fb8206a2864f04de2fc81a41ffe52827df8934c956198a6ce66c48264aedcfa5.jpg) +Figure 5. Impact of where on textual explanation generation. + +![](images/6f78e8e8bfc9d5c433a74cfdf424232680440299cacc38d02f22b809166338b7.jpg) +Figure 6. Impact of what and why on attention map prediction. + +what and why reasoning by providing essential spatial-level cues for more accurate and coherent explanations. + +Next, we construct a baseline without what and why reasoning. We then progressively incorporate these tasks to assess their impact on the attention map prediction, as shown in Fig. 6. Introducing the what task improves where performance, indicating that textual descriptions provide valuable semantic cues for attention localization. Adding the why task further boosts where performance, confirming that causal reasoning strengthens attention map prediction. Overall, these findings highlight the mutual reinforcement between where and what/why reasoning, particularly + +the strong facilitative effect of attention map prediction on textual explanation generation, demonstrating the effectiveness of LLada's joint optimization framework. + +# 6. Conclusion + +We introduce Explainable Driver Attention Prediction, a novel paradigm that extends beyond traditional pixel-wise attention modeling by jointly predicting where drivers look, what they attend to, and why their attention is allocated. This enables a cognitive-level understanding of driver attention, bridging perception and reasoning in autonomous driving. To support this paradigm, we present W $^3$ DA, the first large-scale explainable driver attention dataset, which significantly enriches existing benchmarks with detailed semantic and causal annotations across diverse real-world driving scenarios. Furthermore, we propose LLada, a Large Language Model-driven framework that unifies attention prediction, semantic parsing, and cognitive reasoning in an end-to-end architecture, offering human-like interpretability beyond conventional models. Extensive quantitative and qualitative analyses demonstrate LLada's strong generalization across diverse driving conditions and datasets, highlighting the impact of cognitive cues in driver attention modeling. By systematically decoding the latent cognitive factors underlying attention allocation, our work pushes the frontier of task-driven attention modeling in driving, providing new insights for cognitive science, transparent autonomous systems, and human-computer interaction. + +# Acknowledgements + +This work is supported in part by National Natural Science Foundation of China under Grant 62373387, in part by Natural Science Foundation of Guangdong Province (No. 2023A1515030264), in part by Beijing Natural Science Foundation (4252048), and in part by the Shenzhen Fundamental Research Program (Grant No. JCYJ20240813151301003). Xiaobo Xia is supported by MoE Key Laboratory of Brain-inspired Intelligent Perception and Cognition, University of Science and Technology of China (Grant No. 2421002). + +# References + +[1] Pierluigi Vito Amadori, Tobias Fischer, and Yiannis Demiris. Hammerdrive: A task-aware driving visual attention model. IEEE T-ITS, 23(6):5573-5585, 2021. 3 +[2] Javier Araluce, Luis M Bergasa, Manuel Ocaña, Rafael Barea, Elena López-Guillén, and Pedro Revenga. Aragan: A driver attention estimation model based on conditional generative adversarial network. In IV, pages 1066-1072. IEEE, 2022. 3 +[3] Sonia Baee, Erfan Pakdamanian, Inki Kim, Lu Feng, Vicente Ordonez, and Laura Barnes. Medirl: Predicting the visual attention of drivers via maximum entropy deep inverse reinforcement learning. In ICCV, pages 13178-13188, 2021. 3 +[4] Wentao Bao, Qi Yu, and Yu Kong. Drive: Deep reinforced accident anticipation with visual explanation. In ICCV, pages 7619-7628, 2021. 8 +[5] Loris Bazzani, Hugo Larochelle, and Lorenzo Torresani. Recurrent mixture density network for spatiotemporal visual attention. In ICLR, 2017. 8 +[6] Giuseppe Cartella, Marcella Cornia, Vittorio Cuculo, Alessandro D'Amelio, Dario Zanca, Giuseppe Boccignone, and Rita Cucchiara. Trends, applications, and challenges in human attention modelling. In *IJCAI*, pages 7971-7979, 2024. Survey Track. 1 +[7] Lin Chen, Xilin Wei, Jinsong Li, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Zehui Chen, Haodong Duan, Zhenyu Tang, Li Yuan, et al. Sharegpt4video: Improving video understanding and generation with better captions. NeurIPS, 37:19472-19495, 2025. 3 +[8] Shi Chen, Ming Jiang, Jinhui Yang, and Qi Zhao. Air: Attention with reasoning capability. In ECCV, pages 91-107. Springer, 2020. 1 +[9] Xianyu Chen, Ming Jiang, and Qi Zhao. Gazexplain: Learning to predict natural language explanations of visual scanpaths. In ECCV, pages 314-333. Springer, 2024. 3, 6, 7 +[10] Yilong Chen, Zhixiong Nan, and Tao Xiang. Fblnet: Feedback loop network for driver attention prediction. In ICCV, pages 13371-13380, 2023. 1, 3, 6, 8 +[11] Yihua Cheng, Yaning Zhu, Zongji Wang, Hongquan Hao, Yongwei Liu, Shiqing Cheng, Xi Wang, and Hyung Jin Chang. What do you see in vehicle? comprehensive vision solution for in-vehicle gaze estimation. In CVPR, pages 1556–1565, 2024. 2 + +[12] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhang-hao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yong-hao Zhuang, Joseph E. Gonzalez, Ion Stoica, and Eric P. Xing. Vicuna: An open-source chatbot impressing gpt-4 with $90\%$ * chatgpt quality, 2023. 5, 7 +[13] Can Cui, Yunsheng Ma, Xu Cao, Wenqian Ye, Yang Zhou, Kaizhao Liang, Jintai Chen, Juanwu Lu, Zichong Yang, Kuei-Da Liao, et al. A survey on multimodal large language models for autonomous driving. In WACV, pages 958-979, 2024. 3 +[14] Tao Deng, Hongmei Yan, Long Qin, Thuyen Ngo, and B. S. Manjunath. How do drivers allocate their potential attention? driving fixation prediction via convolutional neural networks. IEEE T-ITS, 21(5):2146-2154, 2020. 2, 6, 7 +[15] Tao Deng, Lianfang Jiang, Yi Shi, Jiang Wu, Zhangbi Wu, Shun Yan, Xianshi Zhang, and Hongmei Yan. Driving visual saliency prediction of dynamic night scenes via a spatiotemporal dual-encoder network. IEEE T-ITS, 25(3):2413-2423, 2023. 8 +[16] Guanqun Ding, Nevrez Imamoglu, Ali Caglayan, Masahiro Murakawa, and Ryosuke Nakamura. Fbnet: Feedback-recursive cnn for saliency detection. In MVA, pages 1-5. IEEE, 2021. 6, 7 +[17] Samuel F Dodge and Lina J Karam. Visual saliency prediction using a mixture of deep neural networks. IEEE T-IP, 27 (8):4080–4090, 2018. 6, 7, 8 +[18] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR. OpenReview.net, 2021. 3 +[19] Jianwu Fang, Dingxin Yan, Jiahuan Qiao, Jianru Xue, He Wang, and Sen Li. Dada-2000: Can driving accident be predicted by driver attention? analyzed by a benchmark. In ITSC, pages 4303-4309. IEEE, 2019. 2, 3, 7 +[20] Jianwu Fang, Dingxin Yan, Jiahuan Qiao, Jianru Xue, and Hongkai Yu. Dada: Driver attention prediction in driving accident scenarios. IEEE T-ITS, 23(6):4959-4971, 2022. 1, 3, 6, 7, 8 +[21] Chao Gou, Yuchen Zhou, and Dan Li. Driver attention prediction based on convolution and transformers. The Journal of Supercomputing, 78(6):8268-8284, 2022. 3 +[22] Chao Gou, Yuchen Zhou, Yao Xiao, Xiao Wang, and Hui Yu. Cascade learning for driver facial monitoring. IEEE T-IV, 8 (1):404-412, 2023. 1 +[23] Jonathan Harel, Christof Koch, and Pietro Perona. Graph-based visual saliency. NeurIPS, 19, 2006. 2, 6, 7 +[24] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. ICLR, 1(2):3, 2022. 6, 7 +[25] Zhongxu Hu, Chen Lv, Peng Hang, Chao Huang, and Yang Xing. Data-driven estimation of driver attention using calibration-free eye gaze and scene features. IEEE T-IE, 69(2):1800-1808, 2021. 2 +[26] Zhongxu Hu, Yiran Zhang, Qinghua Li, and Chen Lv. A novel heterogeneous network for modeling driver attention + +with multi-level visual content. IEEE T-ITS, 23(12):24343-24354, 2022. 3 +[27] Zhixin Huang, Yuchen Zhou, and Chao Gou. Task-oriented scanpath prediction with spatial-temporal information in driving scenarios. In PRCV, pages 177-191. Springer, 2024. 1 +[28] Zhixin Huang, Yuchen Zhou, Jie Zhu, and Chao Gou. Driver scanpath prediction based on inverse reinforcement learning. In ICASSP, pages 8306-8310, 2024. 1 +[29] Laurent Itti, Christof Koch, and Ernst Niebur. A model of saliency-based visual attention for rapid scene analysis. IEEE T-PAMI, 20(11):1254-1259, 1998. 2, 6, 7 +[30] Lisheng Jin, Bingdong Ji, Baicang Guo, Huanhuan Wang, Zhuotong Han, and Xingchen Liu. Mtsf: Multi-scale temporal-spatial fusion network for driver attention prediction. IEEE T-ITS, 2024. 8 +[31] Isaac Kasahara, Simon Stent, and Hyun Soo Park. Look both ways: Self-supervising driver gaze estimation and road scene saliency. In ECCV. Springer Nature Switzerland, 2022. 1, 2, 3 +[32] Iuliia Kotseruba and John K Tsotsos. Attention for vision-based assistive and automated driving: A review of algorithms and datasets. IEEE T-ITS, 23(11):19907-19928, 2022. 1 +[33] Iuliia Kotseruba and John K. Tsotsos. Data limitations for modeling top-down effects on drivers' attention. In IV, 2024. 1 +[34] Matthias Kümmerer, Lucas Theis, and Matthias Bethge. Deep gaze I: boosting saliency prediction with feature maps trained onImagenet. In *ICLRW*, 2015. 6, 7 +[35] Xin Lai, Zhuotao Tian, Yukang Chen, Yanwei Li, Yuhui Yuan, Shu Liu, and Jiaya Jia. Lisa: Reasoning segmentation via large language model. In CVPR, pages 9579-9589, 2024. 3 +[36] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International conference on machine learning, pages 12888-12900. PMLR, 2022. 3 +[37] Qiang Li, Chunsheng Liu, Faliang Chang, Shuang Li, Hui Liu, and Zehao Liu. Adaptive short-temporal induced aware fusion network for predicting attention regions like a driver. IEEE T-ITS, 23(10):18695-18706, 2022. 3 +[38] Yunxin Li, Baotian Hu, Xinyu Chen, Lin Ma, Yong Xu, and Min Zhang. Lmeye: An interactive perception network for large language models. IEEE T-MM, 26:10952-10964, 2024. 3 +[39] Yunxin Li, Zhenyu Liu, Zitao Li, Xuanyu Zhang, Zhenran Xu, Xinyu Chen, Haoyuan Shi, Shenyuan Jiang, Xintong Wang, Jifang Wang, et al. Perception, reason, think, and plan: A survey on large multimodal reasoning models. arXiv preprint arXiv:2505.04921, 2025. 3 +[40] Fei Lin, Yonglin Tian, Yunzhe Wang, Tengchao Zhang, Xinyuan Zhang, and Fei-Yue Wang. Airvista: Empowering uavs with 3d spatial reasoning abilities through a multimodal large language model agent. In ITSC, pages 476-481. IEEE, 2024. 3 + +[41] Akis Linardos, Matthias Kümmerer, Ori Press, and Matthias Bethge. Deepgaze iie: Calibrated prediction in and out-of-domain for state-of-the-art saliency modeling. In ICCV, pages 12919-12928, 2021. 6, 7 +[42] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. NeurIPS, 36:34892-34916, 2023. 3,5,7 +[43] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In CVPR, pages 26296-26306, 2024. 3 +[44] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In CVPR, pages 11976-11986, 2022. 6, 7 +[45] Zijun Long, George Killick, Richard McCreadie, and Gerardo Aragon-Camarasa. Robollm: Robotic vision tasks grounded on multimodal large language models. In ICRA, pages 12428-12435. IEEE, 2024. 3 +[46] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6 +[47] Run Luo, Yunshui Li, Longze Chen, Wanwei He, Ting-En Lin, Ziqiang Liu, Lei Zhang,zikai Song, Xiaobo Xia, Tongliang Liu, et al. Deem: Diffusion models serve as the eyes of large language models for image perception. arXiv preprint arXiv:2405.15232, 2024. 3 +[48] Run Luo, Haonan Zhang, Longze Chen, Ting-En Lin, Xiong Liu, Yuchuan Wu, Min Yang, Minzheng Wang, Pengpeng Zeng, Lianli Gao, et al. Mmevol: Empowering multimodal large language models with evol-instruct. arXiv preprint arXiv:2409.05840, 2024. +[49] Run Luo, Lu Wang, Wanwei He, and Xiaobo Xia. Gui-r1: A generalist r1-style vision-language action model for gui agents. arXiv preprint arXiv:2504.10458, 2025.3 +[50] Anwesan Pal, Sayan Mondal, and Henrik I Christensen. "looking at the right stuff"-guided semantic-gaze for autonomous driving. In CVPR, pages 11883-11892, 2020. 1 +[51] Andrea Palazzi, Davide Abati, simone Calderara, Francesco Solera, and Rita Cucchiara. Predicting the driver's focus of attention: The dr(eye)ve project. IEEE T-PAMI, 41(7):1720-1733, 2019. 1, 2, 3, 7, 8 +[52] Youwei Pang, Xiaqi Zhao, Lihe Zhang, and Huchuan Lu. Multi-scale interactive network for salient object detection. In CVPR, pages 9413-9422, 2020. 8 +[53] Warren Woodrich Pettine, Dhruva Venkita Raman, A David Redish, and John D Murray. Human generalization of internal representations through prototype learning with goal-directed attention. Nature Human Behaviour, 7(3):442-463, 2023. 1 +[54] Michael I Posner and Steven E Petersen. The attention system of the human brain. Annual Review of Neuroscience, 13 (1):25-42, 1990. 1 +[55] Xuebin Qin, Zichen Zhang, Chenyang Huang, Masood Dehghan, Osmar R Zaiane, and Martin Jagersand. U2-net: Going deeper with nested u-structure for salient object detection. Pattern Recognition, 106:107404, 2020. 8 +[56] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, + +Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PMLR, 2021. 3, 5, 7 +[57] Yi Shi, Shixuan Zhao, Jiang Wu, Zhangbi Wu, and Hongmei Yan. Fixated object detection based on saliency prior in traffic scenes. IEEE T-CSVT, 34(3):1413-1426, 2024. 3 +[58] Chonghao Sima, Katrin Renz, Kashyap Chitta, Li Chen, Hanxue Zhang, Chengen Xie, Jens Beibwenger, Ping Luo, Andreas Geiger, and Hongyang Li. Drivelm: Driving with graph visual question answering. In ECCV, pages 256-274. Springer, 2024. 3 +[59] Yuehao Song, Xinggang Wang, Jingfeng Yao, Wenyu Liu, Jinglin Zhang, and Xiangmin Xu. Vitgaze: gaze following with interaction features in vision transformers. Visual Intelligence, 2(1):1-15, 2024. 1 +[60] Lisa J Stephenson, S Gareth Edwards, and Andrew P Bayliss. From gaze perception to social cognition: The shared-attention system. Perspectives on Psychological Science, 16 (3):553–576, 2021. 1 +[61] Han Tian, Tao Deng, and Hongmei Yan. Driving as well as on a sunny day? predicting driver's fixation in rainy weather conditions via a dual-branch visual model. IEEE/CAA Journal of Automatica Sinica, 9(7):1335-1338, 2022. 8 +[62] Xiaoyu Tian, Junru Gu, Bailin Li, Yicheng Liu, Yang Wang, Zhiyong Zhao, Kun Zhan, Peng Jia, Xianpeng Lang, and Hang Zhao. Drivevm: The convergence of autonomous driving and large vision-language models. In CoRL, 2024. 3 +[63] Shengqiong Wu, Hao Fei, Leigang Qu, Wei Ji, and Tat-Seng Chua. Next-gpt: Any-to-any multimodal llm. In ICML, 2024. 3 +[64] Ye Xia, Danqing Zhang, Jinkyu Kim, Ken Nakayama, Karl Zipser, and David Whitney. Predicting driver attention in critical situations. In ACCV, pages 658-674. Springer, 2019. 1, 2, 3, 7, 8 +[65] Zhuofan Xia, Dongchen Han, Yizeng Han, Xuran Pan, Shiji Song, and Gao Huang. Gsva: Generalized segmentation via multimodal large language models. In CVPR, pages 3858-3869, 2024. 3 +[66] Chenxi Xie, Changqun Xia, Mingcan Ma, Zhirui Zhao, Xiaowu Chen, and Jia Li. Pyramid grafting network for one-stage high resolution saliency detection. In CVPR, pages 11717-11726, 2022. 8 +[67] Chuan Xu, Bo Jiang, and Yan Su. Transconvnet: Perform perceptually relevant driver's visual attention predictions. Computers and Electrical Engineering, 115:109104, 2024. 8 +[68] Kun Yan, Zeyu Wang, Lei Ji, Yuntao Wang, Nan Duan, and Shuai Ma. Voila-a: Aligning vision-language models with user's gaze attention. NeurIPS, 37:1890-1918, 2025. 3 +[69] An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, et al. Qwen2.5 technical report. arXiv preprint arXiv:2412.15115, 2024.3 +[70] Kaihui Yang, Junwei Han, Guangyu Guo, Chaowei Fang, Yingzi Fan, Lechao Cheng, and Dingwen Zhang. Progressive adapting and pruning: Domain-incremental learning for saliency prediction. ACM T-MCCA, 2024. 6, 7 + +[71] Shuo Yang, Zijian Yu, Zhenzhe Ying, Yuqin Dai, Guoqing Wang, Jun Lan, Jinfeng Xu, Jinze Li, and Edith C. H. Ngai. Rama: Retrieval-augmented multi-agent framework for mis-information detection in multimodal fact-checking, 2025. 3 +[72] Yang Yang, Zhiying Cui, Junjie Xu, Changhong Zhong, Wei-Shi Zheng, and Ruixuan Wang. Continual learning with bayesian model based on a fixed pre-trained feature extractor. Visual Intelligence, 1(1):5, 2023. 3 +[73] Yifu Yuan, Haiqin Cui, Yibin Chen, Zibin Dong, Fei Ni, Longxin Kou, Jinyi Liu, Pengyi Li, Yan Zheng, and Jianye Hao. From seeing to doing: Bridging reasoning and decision for robotic manipulation. arXiv preprint arXiv:2505.08548, 2025.3 +[74] Yuhang Zang, Wei Li, Jun Han, Kaiyang Zhou, and Chen Change Loy. Contextual object detection with multimodal large language models. IJCV, pages 1-19, 2024. 3 +[75] Hao Zhang, Hongyang Li, Xingyu Liao, Feng Li, Shilong Liu, Lionel M Ni, and Lei Zhang. Da-bev: Depth aware bev transformer for 3d object detection. CoRR, 2023. 3 +[76] Hao Zhang, Hongyang Li, Feng Li, Tianhe Ren, Xueyan Zou, Shilong Liu, Shijia Huang, Jianfeng Gao, Leizhang, Chunyuan Li, et al. Llava-grounding: Grounded visual chat with large multimodal models. In ECCV, pages 19-35. Springer, 2024. 3 +[77] Ruohan Zhang, Akanksha Saran, Bo Liu, Yifeng Zhu, Sihang Guo, Scott Niekum, Dana Ballard, and Mary Hayhoe. Human gaze assisted artificial intelligence: A review. In *IJ-CAI*, page 4951. NIH Public Access, 2020. 1 +[78] Chunyu Zhao, Wentao Mu, Xian Zhou, Wenbo Liu, Fei Yan, and Tao Deng. $Salm^2$ : An extremely lightweight saliency mamba model for real-time cognitive awareness of driver attention. In AAAI, pages 1647-1655, 2025. 1 +[79] Yuchen Zhou, Guang Tan, Mengtang Li, and Chao Gou. Learning from easy to hard pairs: Multi-step reasoning network for human-object interaction detection. In ACM MM, pages 4368-4377, 2023. 1 +[80] Yuchen Zhou, Guang Tan, Rui Zhong, Yaokun Li, and Chao Gou. Pit: Progressive interaction transformer for pedestrian crossing intention prediction. IEEE T-ITS, 24(12):14213-14225, 2023. 1 +[81] Yuchen Zhou, Linkai Liu, and Chao Gou. Learning from observer gaze: Zero-shot attention prediction oriented by human-object interaction recognition. In CVPR, pages 28390-28400, 2024. 1 +[82] Yuchen Zhou, Xinxin Liu, Zipeng Guo, Ming Cai, and Chao Gou. Hktsg: A hierarchical knowledge-guided traffic scene graph representation learning framework for intelligent vehicles. IEEE T-IV, 2024. 1 +[83] Yiwei Zhou, Xiaobo Xia, Zhiwei Lin, Bo Han, and Tongliang Liu. Few-shot adversarial prompt learning on vision-language models. NIPS, 37:3122-3156, 2024. 3 +[84] Yuchen Zhou, Chao Gou, Zipeng Guo, Yihua Cheng, and Hyung Jin Chang. Behavior-aware knowledge-embedded model for driver attention prediction. IEEE T-CSVT, 2025. 1 +[85] Pengfei Zhu, Mengshi Qi, Xia Li, Weijian Li, and Huadong Ma. Unsupervised self-driving attention prediction via uncertainty mining and knowledge embedding. In ICCV, pages 8524-8534, 2023. 1, 3 \ No newline at end of file diff --git a/ICCV/2025/Where, What, Why_ Towards Explainable Driver Attention Prediction/images.zip b/ICCV/2025/Where, What, Why_ Towards Explainable Driver Attention Prediction/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..e2b9a1c120f9ee37d14c11cab2a203721854107f --- /dev/null +++ b/ICCV/2025/Where, What, Why_ Towards Explainable Driver Attention Prediction/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:aeca8cf41e716ea98c0635756d2bcf12ea8571b9615678314cf79f9bacf55650 +size 464588 diff --git a/ICCV/2025/Where, What, Why_ Towards Explainable Driver Attention Prediction/layout.json b/ICCV/2025/Where, What, Why_ Towards Explainable Driver Attention Prediction/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..25bd99f037fbf47a81bddeaa4ac7db49850fb70c --- /dev/null +++ b/ICCV/2025/Where, What, Why_ Towards Explainable Driver Attention Prediction/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:902c43d60b0d12815ebcfd467bac93bc029f609d7bee393bc8ba1b7653dfac81 +size 522283 diff --git a/ICCV/2025/Who Controls the Authorization_ Invertible Networks for Copyright Protection in Text-to-Image Synthesis/3f7258c8-a390-4cf7-9b1e-42cbf47a42a9_content_list.json b/ICCV/2025/Who Controls the Authorization_ Invertible Networks for Copyright Protection in Text-to-Image Synthesis/3f7258c8-a390-4cf7-9b1e-42cbf47a42a9_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..5165a7002f55769457c573b4229ef510866de52a --- /dev/null +++ b/ICCV/2025/Who Controls the Authorization_ Invertible Networks for Copyright Protection in Text-to-Image Synthesis/3f7258c8-a390-4cf7-9b1e-42cbf47a42a9_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ecc21ccf34ee7990b3cddc27f50ef074a6e371a5afb9fb6781a32ea9825f82b2 +size 75010 diff --git a/ICCV/2025/Who Controls the Authorization_ Invertible Networks for Copyright Protection in Text-to-Image Synthesis/3f7258c8-a390-4cf7-9b1e-42cbf47a42a9_model.json b/ICCV/2025/Who Controls the Authorization_ Invertible Networks for Copyright Protection in Text-to-Image Synthesis/3f7258c8-a390-4cf7-9b1e-42cbf47a42a9_model.json new file mode 100644 index 0000000000000000000000000000000000000000..a04069bcc4570afd2bec82e6c717fe882918bfc1 --- /dev/null +++ b/ICCV/2025/Who Controls the Authorization_ Invertible Networks for Copyright Protection in Text-to-Image Synthesis/3f7258c8-a390-4cf7-9b1e-42cbf47a42a9_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:95900f7a8b429a9ba4197e6988e7150f5265b3dcb3bfbe57dd7e6d5719645bf6 +size 90548 diff --git a/ICCV/2025/Who Controls the Authorization_ Invertible Networks for Copyright Protection in Text-to-Image Synthesis/3f7258c8-a390-4cf7-9b1e-42cbf47a42a9_origin.pdf b/ICCV/2025/Who Controls the Authorization_ Invertible Networks for Copyright Protection in Text-to-Image Synthesis/3f7258c8-a390-4cf7-9b1e-42cbf47a42a9_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..195ca2251a0d5f02ba363e649b77e497b13bc87c --- /dev/null +++ b/ICCV/2025/Who Controls the Authorization_ Invertible Networks for Copyright Protection in Text-to-Image Synthesis/3f7258c8-a390-4cf7-9b1e-42cbf47a42a9_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2708b23bb76b5579d3b1b2d7396ad354cb155453941be1fc81e39e544012b45f +size 6091003 diff --git a/ICCV/2025/Who Controls the Authorization_ Invertible Networks for Copyright Protection in Text-to-Image Synthesis/full.md b/ICCV/2025/Who Controls the Authorization_ Invertible Networks for Copyright Protection in Text-to-Image Synthesis/full.md new file mode 100644 index 0000000000000000000000000000000000000000..479dd43a4a67f4e18ccbc2705ab11dce1e627898 --- /dev/null +++ b/ICCV/2025/Who Controls the Authorization_ Invertible Networks for Copyright Protection in Text-to-Image Synthesis/full.md @@ -0,0 +1,295 @@ +# Who Controls the Authorization? Invertible Networks for Copyright Protection in Text-to-Image Synthesis + +Baoyue Hu $^{1,3}$ , Yang Wei $^{2,3,\text{四}}$ , Junhao Xiao $^{1,3}$ , Wendong Huang $^{1,3}$ , Xiuli Bi $^{2,3,\text{四}}$ , and Bin Xiao $^{2,3,4}$ $^{1}$ School of Computer Science and Technology, $^{2}$ School of Artificial Intelligence, $^{3}$ Chongqing Key Laboratory of Image Cognition, Chongqing University of Posts and Telecommunications, Chongqing, China + $^{4}$ Jinan Inspur Data Technology Co., Ltd., Jinan, China + +d230201013@stu.cqupt.edu.cn, weiyang@cqupt.edu.cn, d230201040@stu.cqupt.edu.cn +d220201013@stu.cqupt.edu.cn, bixl@cqupt.edu.cn, xiaobin@cqupt.edu.cn + +# Abstract + +To defend against personalized generation, a new form of infringement that is more concealed and destructive, the existing copyright protection methods is to add adversarial perturbations in images. However, these methods focus solely on countering illegal personalization, neglecting the requirement for legitimate personalization. Moreover, none of these methods are capable of directly verifying and tracing the copyright from adversarial examples. In response to these limitations, we propose a traceable and authorizable copyright traceability method that embeds the copyright watermark into images through a series of invertible compound coupling modules. We introduce a novel information exchange mechanism for invertible neural network and design a contrastive learning-based optimization strategy tailored to address personalized infringement issues. Our method effectively mitigates the malicious use of unauthorized personalized generation models by inducing watermark-like artifacts and obscuring privacy details in generated images. Additionally, it facilitates copyright traceability and supports authorized legitimate personalization, thereby offering broader practical applicability. Experimental results demonstrate that our method can almost losslessly restore original image and extract copyright watermark, achieving FID exceeding 300 and causing visually noticeable artifacts in unauthorized personalized images. Furthermore, it exhibits consistent robustness against adversarial purification and text prompt modifications. + +# 1. Introduction + +Diffusion models (DM) have achieved remarkable success due to their exceptional generative abilities. Notably, text- + +
Algorithms① UP② AP③ CT
Stegastamp [31] (CVPR 20')X-✓
RoSteALS [3] (CVPR 23')X-✓
AdvDM [20] (ICML 23')✓XX
MetaCloak [22] (CVPR 24')✓XX
Wm-AdvDM [40] (CVPR 24')✓X✓
Ours✓✓✓
+ +Table 1. Comparison of the functions of representative copyright protection methods. ①, ② and ③ represent Unauthorized Personalization Resistance, Authorized Personalization Support and Copyright Traceability, respectively. + +to-image generation models [26, 28] like Stable Diffusion can produce highly realistic and diverse images according to textual descriptions. However, the widespread adoption of AI-generated tools and the inherent lack of traceability in their outputs have introduced new security challenges. The rapid advancement of text-to-image diffusion models has revolutionized the field of image generation [8, 13], surpassing the boundaries of traditional copyright protection techniques [24, 36]. In this context, users can easily generate infringing images, including artworks that mimic the style of specific artists and deepfake images containing sensitive personal information. To address this emerging form of infringement, researchers are actively exploring effective copyright protection methods. + +Existing copyright protection methods include traditional watermarking techniques and adversarial approaches specifically designed to defend against diffusion models. Watermarking is a common proactive defense mechanism for traditional copyright infringement, embedding traceable watermarks into copyrighted images to verify their origin and assert ownership. However, these conventional watermarking techniques [3, 24, 31] that support copyright track- + +ing are ineffective against infringement facilitated by diffusion models, which generate entirely new images. + +Among all infringement methods leveraging diffusion models, personalized generation [11, 18, 29] poses a particularly significant challenge due to its low entry barrier, minimal requirements, strong concealment, and resistance to traditional defensive measures. Personalized generation requires only a few target-related images to fine-tune existing text-to-image models, enabling the creation of infringing images that highly similar to the target's identity or artistic style without authorization. To counter this form of infringement, researchers have proposed adversarial-based defense strategies [19, 20, 22, 32] that introduce imperceptible perturbations into original images. These perturbations force diffusion models to learn features deviating from the original image semantics, significantly degrading their generative performance and reducing infringement risks. However, these methods still have limitations. + +On one hand, existing adversarial-based antipersonalization methods overlook the necessity of authorized usage. While these approaches effectively deter unauthorized malicious personalization, their restrictive measures may also render images permanently unusable for legitimate personalization under authorized conditions. On the other hand, these methods exhibit significant limitations in terms of traceability. They primarily rely on introducing visually anomalous artifacts in unauthorized generated images for protection but lack a verifiable mechanism for copyright attribution. While Wm-AdvDM [40] can reveal copyright attribution from unauthorized outputs, it still fails to directly verify copyright from the copyrighted image. + +To address the shortcomings of both traditional watermarking and existing anti-personalization methods, as well as to meet the demand for image copyright protection that balances security and flexibility, we propose a novel traceable and authorizable copyright protection framework. As illustrated in Figure 1, the proposed invertible framework effectively mitigates infringement risks and meets protection requirements in network environments. This framework offers a triple-layered defense mechanism: (1) it effectively mitigates infringement caused by unauthorized personalization; (2) it supports authorized personalization; and (3) it enables the extraction of copyright watermarks at any time, ensuring traceability of infringement activities. Furthermore, it allows authorized users to restore images in a local environment, circumventing the risk of information leakage during network transmission. A comparison of our method with several representative copyright protection techniques is presented in Table 1. + +Contributions. Our main contributions are as follows. + +1. We propose a copyright protection strategy using invertible neural networks to generate images that prevent unauthorized personalization, enable effective copyright + +![](images/ade2fbdcd2ef373b342a56b0b5438c2c1be5d5aacbbdfdf462e444f75d1e3ee0.jpg) +Figure 1. Infringes encountered in the network environment and corresponding protection measures for copyrighted image by our proposed invertible copyright traceability method, which defends against various image infringements while supporting authorized personalization or other high-resolution tasks. + +tracking in case of misuse, and support authorized personalization. + +2. We investigate the relationship between diffusion model and different frequency domain components, incorporating a novel mechanism into our model to enhance the interaction between the original image and the embedded copyright watermark, thereby improving robustness against unauthorized personalization. +3. To address the dual challenges of content attribute leakage and untraceable copyright in personalized infringement scenarios, we propose a contrastive learning-based copyright protection optimization strategy. +4. Experimental results demonstrate that our method forces unauthorized generated images to deviate significantly from the real image distribution, achieving an FID score exceeding 300 on the VGGFace2 dataset while exhibiting artifacts resembling copyright watermarks. Furthermore, the results of authorized personalization show no significant differences from the original. + +# 2. Related Work + +# 2.1. Image Watermarking for Copyright Protection + +The underlying logic of watermarking technology lies essentially in exploiting redundancy in images to hide copyright watermarks while evading the human eye or even detection systems. Early methods for image copyright protection, such as least significant bit and histogram shifting, converted copyright watermarks into binary bits and em + +bedded them into pixel space. In pursuit of greater robustness, researchers explored frequency-domain watermarking [35], employing mathematical transformations to transform images into frequency domain and embedding watermarks by modulating appropriate frequency coefficients. With the emergence of deep learning, neural network-based watermarking approaches [3, 31], including those utilizing CNNs and GANs, have introduced new possibilities for enhancing the performance of image watermarking. + +# 2.2. Copyright Protection against Unauthorized Personalized Generation + +To address the security issues brought by high-performance personalized generation model, recent studies have employed adversarial-based methods to protect image copyrights. These methods add adversarial perturbation in images to form adversarial examples that significantly degrade generation performance of diffusion model. Such adversarial examples serve as 'copyrighted images' that effectively prevent unauthorized personalization of original image. + +AdvDM [20] is a pioneering work that protects user images from being personalized by diffusion models. It uses the training loss of a pre-trained diffusion model as the objective function for generating adversarial examples to design the perturbations. Its successor, Mist [19], further considers texture loss to achieve more robust and visually clear images. Similarly, Anti-DreamBooth [32] adopts the strategy of maximizing the training loss of diffusion models as its optimization objective and enhances protection performance through alternating training. SimAC [33] designs its method based on the intrinsic properties of the generative model. Some studies also focus on improving the robustness of adversarial perturbations [21, 22], or the visual quality of adversarial examples [1]. Recently, Zhu et al. [40] proposed a watermarking method with copyright traceability by embedding personal watermarks into images utilizes a pre-trained generator [16]. However, according to the perspective of ArtFlow [2], the pooling operations in the encoder result in spatial information loss, and the decoder cannot achieve lossless image reconstruction. + +# 2.3. Invertible Neural Network + +Dinh et al. [9] first proposed the Invertible Neural Network (INN) for unsupervised learning. Its basic idea is to find an inverse generator $G$ that maps a high-dimensional latent variable $x$ to an output $y$ , where $y = G(x)$ is close to the distribution of a real image. The inverse model $G^{-1}$ shares the same parameters as $G$ , and the input $x$ can be obtained directly through $x = G^{-1}(y)$ . Dinh et al. [10] later incorporated convolutional layers into coupling models and introduced multi-scale layers to reduce computational cost and increase regularization ability. Glow [17] further refined INN for image-related tasks by applying in + +vertible $1 \times 1$ convolution. Since then, INN has been applied in various image-related tasks, such as deep image hiding [15, 38, 39], image super-resolution [37], etc., and also for generating adversarial examples [5, 27]. + +We are the first to employ INN for anti-infringement defense, specifically addressing illegal image personalization—a novel application of INN in copyright protection. We innovatively adapt the invertible framework to simultaneously address two challenges: preventing personalized infringement along with enabling near-lossless watermark extraction and original image restoration to meet practical demands. This dual functionality ensures copyright security while expanding user-centric applications. + +# 3. Methodology + +In previous sections, we identified critical limitations of existing anti-personalization methodologies based on practical needs. To address these limitations, we propose a novel INN-based copyright protection framework that enables copyright traceability and authorized personalization, advancing anti-personalization research. + +# 3.1. Overall Architecture + +The proposed traceable and authorizable method adopts the basic architecture and principles of the INN [10]. The entire network can be divided into a forward process for generating adversarial examples and a backward process for reconstruction. The two processes differ only in the direction of the data flow, with the same module $i$ sharing identical parameters. The network structure is illustrated in Figure 2. It consists of $N$ affine coupling modules with the same structure. Based on the characteristics of this task, we analyze the relationship between the diffusion model and frequency components. Leveraging this correlation, we design the compound coupling module, as detailed in Section 3.2. + +Our goal is to ensure that, in the forward process, information from copyright watermark that contributes to personalized generation is embedded into copyrighted image, while discarding irrelevant information, which is output as a "key". This key will be provided to authorized personalizers or copyright authenticators for authorized personalization or verification. To achieve this, we propose an optimization scheme implemented through a contrastive learning-based adversarial loss, as detailed in Section 3.3. + +The threat scenario and the corresponding authorization protocol are discussed in the appendix. + +# 3.2. Compound Coupling Module + +A widely accepted notion in image watermarking technology is that distinct frequency components exhibit varying levels of effectiveness in terms of embedding capacity and robustness. We start by examining the relationship between + +![](images/719e4617b9e7d9705b346168242e5b2491f5ea91ea8ac2e4467fb1b73efd9bf2.jpg) +Figure 2. INN-based invertible copyright protection framework against unauthorized personalization. The green and purple arrows indicate forward and backward data flows, respectively, corresponding to: (1) "Protection" processes for generating copyrighted images and (2) "Reconstruction" processes for restoring bath original image and copyright watermark image. + +![](images/81e4c3c5456fca841c68de0a0696111b4aa0d79acdea0484e77953c3bd4167e5.jpg) +Figure 3. Impact of watermark embedding in different frequency components on images generated by personalized diffusion model. In this example, the personalized diffusion model used is Dream-Booth, denoted as $\varepsilon(\cdot)$ . To better visualize the impact on frequency coefficients, only one color component is shown for each sub-band (LL, LH, HL, and HH). + +information in various frequency components and the performance of personalized diffusion models. + +In the context of discrete wavelet transform (DWT), LL, LH, HL, and HH refer to sub-bands that represent different frequency components of the original image $I_{ori}$ . We sequentially apply interference to the ring-shaped region of four DWT sub-bands, using Gaussian-distributed noise to simulate the impact of adversarial perturbations. We modify one sub-band at a time, then apply the inverse wavelet transform (IWT) to obtain the watermarked image $I_{ll}^{w}$ , $I_{lh}^{w}$ , $I_{hl}^{w}$ , and $I_{hh}^{w}$ . The personalized diffusion model, denoted as $\varepsilon(\cdot)$ , is applied to each watermarked image, resulting in the generated images $\varepsilon(I_{ll}^{w})$ , $\varepsilon(I_{lh}^{w})$ , $\varepsilon(I_{hl}^{w})$ , and $\varepsilon(I_{hh}^{w})$ . We can visually assess the abnormal artifacts in the generated images to evaluate the impact of watermark embedding in different frequency sub-bands on personalization. + +As shown in Figure 3, perturbations in the low-frequency sub-band cause more pronounced visual artifacts in the gen + +erated image $\varepsilon(I_{ll}^{w})$ compared to those in the mid- and high-frequency sub-bands. Low-frequency typically capture the structural information of images, which is more robust to common image attacks. This characteristic remains effective against personalized attacks based on diffusion models, as evidenced by our validation. + +Building on these findings, we aim to train our network to embed robust copyright watermark features into the copyrighted image, leveraging the property that low-frequency information can be better preserved during the diffusion generation process. As shown in Figure 2, the original image and the copyright watermark are transformed into the frequency domain using DWT, and the resulting frequency components are simultaneously fed into a sequence of invertible compound coupling modules. The outputs of the last coupling module are passed through IWT to generate the copyrighted image and the key image. + +The input data of compound coupling module $x_{\mathrm{ori}}^i$ and $x_{\mathrm{w}}^i$ are transformed by learning functions and coupled alternately. The computation formulas for the forward process are as follows: + +$$ +\begin{array}{l} x _ {\mathrm {o r i}} ^ {i + 1} = x _ {\mathrm {o r i}} ^ {i} + \phi \left(x _ {\mathrm {w}} ^ {i}\right), \tag {1} \\ x _ {\mathrm {w}} ^ {i + 1} = x _ {\mathrm {w}} ^ {i} \odot \psi \left(x _ {\mathrm {o r i}} ^ {i + 1}\right) \odot \exp \left(\alpha \left(\rho \left(x _ {\mathrm {o r i}} ^ {i + 1}\right)\right)\right) + \eta \left(x _ {\mathrm {o r i}} ^ {i + 1}\right), \\ \end{array} +$$ + +where $\alpha$ is a sigmoid function multiplied by a constant factor, and $\odot$ denotes the dot product operation. The learning functions $\rho(\cdot), \varphi(\cdot)$ , and $\eta(\cdot)$ are represented using the dense block in [34]. To better align with this task, we introduce a function $\psi(\cdot)$ , which employs the mechanism in [14] to enhance the exchange of information flow between the frequency components of the $x_{\mathrm{ori}}^i$ and $x_{\mathrm{w}}^i$ . + +In the backward process, the direction of information flow is reversed compared to the forward process, where the + +![](images/81cd6968d5b8c4e431803b971179d893cf814f935e2c280d248f10f8fcc200db.jpg) +Figure 4. Conceptual of the contrastive learning-based copyright protection optimization strategy. + +inputs $x_{\mathrm{c}}^{i}$ and $x_{\mathrm{key}}^{i}$ are used, and the computation formulas are as follows: + +$$ +\begin{array}{l} x _ {\text {k e y}} ^ {i} = x _ {\text {k e y}} ^ {i + 1} - \eta \left(x _ {\mathrm {c}} ^ {i + 1}\right) \odot \exp \left(- \alpha \left(\rho \left(x _ {\mathrm {c}} ^ {i + 1}\right)\right)\right) \oslash \psi \left(x _ {\mathrm {c}} ^ {i + 1}\right), \tag {2} \\ x _ {\mathrm {c}} ^ {i} = x _ {\mathrm {c}} ^ {i + 1} - \phi \left(x _ {\mathrm {k e y}} ^ {i}\right). \\ \end{array} +$$ + +# 3.3. Loss Functions + +The loss function consists of two types of loss: perceptual loss, which constrains the modification to preserve the original image's integrity, and adversarial loss, which enhances the ability of copyrighted images to resist unauthorized personalization. To ensure that the diffusion model captures copyright watermark features while deviating from original image features, we proposed a contrastive learning-based adversarial loss. + +Perceptual loss. The forward process gradually integrates the content of copyright watermark into the copyrighted image, while the copyrighted image is required to remain close to the original image. To achieve this, we comprehensively measure the differences between the images at the pixel level, feature level and frequency level. The perceptual loss $\mathcal{L}_{\mathrm{perp}}$ is defined as follows, + +$$ +\begin{array}{l} \mathcal {L} _ {\text {p e r p}} = \lambda_ {p i x} \| \left(x _ {\mathrm {c}}, x _ {\mathrm {o r i}}\right) \| _ {2} ^ {2} + \lambda_ {f e a t} \| \theta \left(x _ {\mathrm {c}}\right), \theta \left(x _ {\mathrm {o r i}}\right) \| _ {2} ^ {2} \tag {3} \\ + \lambda_ {f r e q} \| \left(x _ {\mathrm {c}}\right) _ {l l}, \left(x _ {\mathrm {o r i}}\right) _ {l l} \| _ {2} ^ {2}, \\ \end{array} +$$ + +where $\theta(\cdot)$ denotes the features of the pretrained VGG16 model, $x_{\mathrm{ori}}$ and $x_{\mathrm{c}}$ are the original and copyrighted images, respectively, and $(\cdot)_{ll}$ indicates the operation of extracting low-frequency sub-bands after wavelet decomposition. $\lambda_{pix}$ , $\lambda_{feat}$ and $\lambda_{freq}$ are scalar hyperparameters that weight the contributions of different loss terms at the pixel, feature, and frequency levels, respectively. + +Adversarial loss. We design the adversarial loss based on the contrastive learning approach. The output of network + +(i.e., the copyrighted image) is treated as an anchor, with the objective of minimizing the representation distance between the anchor and the copyright watermark, while maximizing its representation distance from the original image. Figure 4 illustrates the conceptual of this strategy. The adversarial loss $\mathcal{L}_{\mathrm{advc}}$ is defined as follows: + +$$ +\mathcal {L} _ {\mathrm {a d v c}} = \max \left(\| \mathcal {E} \left(x _ {\mathrm {c}}\right) - \mathcal {E} \left(x _ {\mathrm {w}}\right) \| _ {2} ^ {2} - \| \mathcal {E} \left(x _ {\mathrm {c}}\right) - \mathcal {E} \left(x _ {\mathrm {o r i}}\right) \| _ {2} ^ {2} + \alpha , 0\right). \tag {4} +$$ + +We follow [30, 40] to calculate the loss based on the latent representation $\mathcal{E}(x)$ , which is the output of the Variational Auto Encoder. $x_{\mathrm{w}}$ denotes the copyright watermark, and $\alpha$ is a hyperparameter that regulates the margin difference between two samples. + +Additionally, we use another adversarial loss $\mathcal{L}_{\mathrm{adv}}$ to more intuitively minimize the distance between the representation of the copyrighted image and the copyright watermark, ensuring that copyright can be traced in images generated by unauthorized personalization, as follows, + +$$ +\mathcal {L} _ {\mathrm {a d v t}} = \left\| \mathcal {E} \left(x _ {\mathrm {c}}\right) - \mathcal {E} \left(x _ {\mathrm {w}}\right) \right\| _ {2} ^ {2}. \tag {5} +$$ + +Total loss. The total loss function $\mathcal{L}_{total}$ is a weighted sum of $\mathcal{L}_{\mathrm{perp}}, \mathcal{L}_{\mathrm{advc}}$ and $\mathcal{L}_{\mathrm{advt}}$ , as follows, + +$$ +\mathcal {L} _ {\text {t o t a l}} = \mathcal {L} _ {\text {p e r p}} + \lambda_ {a d v c} \mathcal {L} _ {\text {a d v c}} + \lambda_ {a d v t} \mathcal {L} _ {\text {a d v t}}, \tag {6} +$$ + +where $\lambda_{advc}$ and $\lambda_{advt}$ are weights to balance different loss terms. + +# 4. Experiment + +# 4.1. Experiment Setting + +Datasets. Our experiments are performed on human subjects using the dataset of VGGFace2 [4], and following Anti-DreamBooth [32] fifty identities are selected randomly from it. VGGFace2 is a comprehensive dataset with over 3.3 million face images from 9,131 unique identities. + +Implementation Details and Comparison methods. The personalized model we use is Dreambooth based on the SD v2-1 base. We train both the text encoder and UNet model of DreamBooth, with batch size of 1 and learning rate of $5 \times 10^{-6}$ for 1000 training steps. We leverage instance prompts with 50 inferring steps to generate 16 images per prompt in the inference phase. The prompts [32] include one used in training and another unseen, to ensure the practicality of the experimental results. We set the weight of adversarial loss $\lambda_{advc} = 0.5$ , $\lambda_{adv} = 0.1$ and $\alpha = 0.5$ . The perceptual loss weights are $\lambda_{pix} = 50$ , $\lambda_{feat} = 0.005$ and $\lambda_{freq} = 10$ . In contrast to other methods that use adversarial examples for copyright protection, we do not set a fixed perturbation bound. Instead, we adjust the perturbation magnitude dynamically based on the loss function. + +We evaluate the proposed copyright protection method with ASPL [32], FSMG [32], MetaClock [22] and Wm-AdvDM [40] which are adopted state-of-the-art baselines. + +
Method"A photo of sks person""A dslr portrait of sks person"
FID ↑FDFR ↑ISM ↓FID ↑FDFR ↑ISM ↓
No Defense85.900.03910.4887134.220.13250.5147
DefenseASPL [32] (ICCV 23')214.410.16370.3208216.95 (+2.53)0.1825 (+0.0188)0.3115 (-0.0093)
FSMG [32] (ICCV 23')217.090.20370.2956214.45 (-2.64)0.2262 (+0.0225)0.2814 (-0.0142)
MetaClock [22] (CVPR 24')193.170.20620.3984154.89 (-38.28)0.1637 (-0.0425)0.5263 (+0.1279)
Wm-AdvDM [40] (CVPR 24')160.120.07120.3613151.68 (-8.44)0.1033 (+0.0321)0.5156 (+0.1543)
Ours300.350.18870.4611314.71 (+14.36)0.2350 (+0.0463)0.4963 (+0.0352)
+ +Table 2. Comparing the defense performance of current state-of-the-art methods on VGGFace2. + +![](images/8655315b9efa2d7b84dabfac2337cdb97e09d8d659b6a42eee82ade507ea4186.jpg) +Figure 5. The performance of current state-of-the-art methods against unauthorized personalization. + +Evaluation metrics. We evaluate our method from three perspectives. For the image generated through unauthorized and authorized personalization, we use Frechet Inception Distance (FID) [12] to measure the similarity between the generated and clean image, and we measure the face detection failure rate (FDFR) using Retina Face detector [7] to determine whether the face is present in the generated image, and use Identity Score Matching (ISM) [32] to measure the similarity of identity, the embedding extractor is ArcFace [6] recognizer. For the traceability of watermarks, we use Normalized Cross Correlation (NC) to measure the similarity between the original watermark and the extracted watermark. For the copyrighted image, we use three common visual quality evaluation indicators: PSNR, SSIM, and MSE to measure the difference from the original images. + +# 4.2. Effectiveness of Unauthorized and Authorized Personalization + +Effectiveness of Unauthorized Personalization. We have demonstrated the performance of current state-of-the-art methods against unauthorized personalization. A high FID mean that the generated image deviates from the original one, indicating that diffusion model fails to capture the original face's feature from the copyrighted image. A high FDFR and a low ISM means that the face in the generated image is difficult to be recognized by the detector and the + +original face features are not preserved in it, which indicates that the facial privacy features are not easily leaked by unauthorized personalization. The average scores of generated images are presented in Table 2. Our method effectively guides the diffusion model to shift from capturing facial features to learning unique features of the watermark, significantly increasing the FID metric to 300 and achieving state-of-the-art on VGGFace2. + +Unfortunately, our method is less effective than other algorithms in disrupting facial consistency. However, our primary focus is to ensure the traceability of generated images for copyright protection, and the relevant experimental results are presented in Section 4.3. + +The visualization results are presented in Figure 5 and additional results on other objects are provided in the appendix. The traceable method Wm-AdvDM [40] introduces grid-like artifacts in personalized images, but traces of a copyright watermark remain insignificant, especially under unseen prompt. Our method generates watermark-like artifacts that are clearly observable in unauthorized personalized images. Although facial contours remain discernible, these artifacts effectively obscure critical biometric details such as ocular features, achieving both copyright traceability and privacy protection. + +It is worth noting that our method maintains high performance even with unseen prompt, and in some cases, it + +![](images/79efee2ff11b849307e2200e29c0c8c73cfab2de4b88ec87228629c08a36f092.jpg) +Figure 6. The effect of our method supports authorized personalization. "Input" refers to the input images for personalized training, while "Generated" denotes the output images of diffusion model trained on three different sets: the original images, the unauthorized copyrighted images, and the authorized restored images protected by our method. + +outperforms seen prompt. We attribute this to our method's ability to establish a strong association between the copyright watermark and the target token "sks", ensuring that the diffusion model captures sufficient watermark features rather than facial privacy features. This capability is an advancement, as other methods exhibit varying degrees of performance degradation when the prompt changes. In contrast, our method consistently preserves its protective efficacy and copyright traceability under diverse text prompts, demonstrating consistent robustness to prompt variations. + +In addition, we perform personalization tasks on general object dataset ImageNet and style transfer tasks on WikiArt. The corresponding results are presented in the appendix. + +Effectiveness of Authorized Personalization. We verify the effect of our method in supporting authorized personalization. Figure 6 shows the final generated images of our methods under unauthorized and authorized personalization, as well as the original one without protection. The generated images obtained with authorization exhibit high similarity to those generated from the original, with FID differences less than $1\%$ . This is due to our invertible method, which allows for near-lossless restoration of the original images. In the case of unauthorized personalization, the generated image trained on copyrighted images displays visible artifacts resembling the copyright watermark. + +# 4.3. Copyright Traceability + +As mentioned earlier, our method supports copyright traceability in various infringement scenarios, including: + +(1) Tracing copyright from generated images. "Tracing copyright from generated images" refers to the presence of watermark-like artifacts in all generated images trained using copyrighted images, effectively declaring ownership of the image content, much like a visible watermark. As + +
Purification TechniquesFID ↑FDFR ↑ISM ↓
No300.350.18870.4611
JPEG300.590.22620.4496
SR306.110.23750.4031
Gaussian175.760.08130.4714
Oracle*85.900.03910.4887
+ +Table 3. The effects of applying purification techniques on copyrighted images. Oracle* denotes training on clean image. + +![](images/2f71c89d841061ed187f6502d46f759905e43106bedb4a1c91833a4869fdd289.jpg) +Figure 7. Examples of copyrighted images after purification, corresponding generated images, and extracted watermarks. + +shown in Figure 5 and 7, Our copyright protection method demonstrates certain robustness, where changing the text prompts and applying adversarial purification techniques do not impact the tracing ability. More results with different prompts can be found in the appendix, further supporting the stability of the copyright watermark across different generated images. + +(2) Tracing copyright from copyrighted images. The copyright owner can directly extract the copyright watermark from the image when the copyrighted image is stolen by attackers, which is actually a more common form of image infringement. A higher NC indicates that the extracted watermark image is more similar to the watermark embedded by the owner. Figure 7 shows that the copyright watermark extracted from the copyrighted image still retains strong copyright traceability, as evidenced by its perceptual quality. The average NC value calculated for the watermark extracted in 200 VGGFace2 images is 0.9929, which is sufficient for the owner to assert copyright ownership. + +# 4.4. Robustness of Copyrighted Images + +To evaluate the robustness of the copyrighted images generated by our method, we apply several widely used adversarial purification techniques, including JPEG compression [23], super-resolution (SR) [25], and Gaussian filtering. These purification techniques directly manipulate the + +
MethodPSNR ↑SSIM ↑MSE ↓
ASPL [32]35.01480.868620.5187
FSMG [32]34.84790.867321.3396
MetaClock [22]29.41990.7232395.2789
Wm-AdvDM [40]35.26510.902519.6919
Ours30.69510.806790.6988
Ours (Authorized)∞0.986153.1304
+ +copyrighted image to remove the protective perturbations that we added on the image. We use a quality factor of 75 for the JPEG, a scale factor of 4 for the SR, and the kernel size of gaussian blur is set to $3^{*}3$ . The FID and FDFR metrics are calculated between clean images and unauthorized generated images, demonstrating the effectiveness of our method against adversarial purification. + +As shown in Table 3, JPEG compression and superresolution have no impact on our method's ability to resist personalized infringement. Gaussian filtering smooths the image and partially disrupts the injected watermark, restoring the generation ability of the diffusion model. Nevertheless, it fails to fully recover the model's performance to the level achieved when trained on clean images. + +Figure 7 demonstrates that purified images maintain robust anti-infringement ability despite exhibiting variations when generated with the same seed. This conclusion is evidenced by two observations: (1) the consistent presence of identifiable artifacts in generated images, and (2) the high fidelity of watermarks extracted from purified images. + +# 4.5. Visual Quality of Copyrighted Image + +Visual quality determines the applicability and effectiveness of images in downstream tasks within practical network environments. We present the evaluation results in Table 4, which compares the visual quality across different methods. "Ours(Authorized)" in the table indicates performance with restored images post-authorization. Based on the values of these image quality metrics, it can be observed that the copyrighted image generated by these protection methods exhibits relatively poor quality compared to invisible watermarking methods in other application domains. Although the MSE is moderate, as it reflects pixel-level sensitivity, the superior PSNR and SSIM indicate the perceptual quality advantages of our method. + +Figure 6 provides some visualization of copyrighted images and restored images. Furthermore, the appendix presents visual comparisons of images generated by comparison methods. Despite aiming for imperceptible watermarks, they embed systematic adversarial perturbations within the images, which appear as regular patterns throughout the image content upon close inspection. We attribute this unavoidable visual quality degradation to the + +Table 4. The quantitative evaluation results of visual quality. + +
SettingFID ↑ISM ↓PSNR ↑SSIM ↑
w/o DWT317.350.472828.95800.7801
w/o COP314.100.486730.73250.8067
w/o Ladvc214.220.402036.13360.9249
Ours324.110.436730.76340.8074
+ +Table 5. Ablation Study by cutting of different modules. "COP" and "Ladvc" refers to the compound coupling module, and the proposed contrastive learning-based adversarial loss, respectively. + +powerful learning capability of diffusion model and the consequent increased difficulty in generating effective adversarial perturbations. + +This challenge serves as one of the motivations for developing our invertible algorithm, which mitigates visual degradation in adversarial examples for ant-personalization. As shown in the table, the restored images obtained under authorized have high similarity to the original images. The restored images not only support authorized personalization but also meet the requirements for high-resolution tasks. + +# 4.6. Ablation Study + +We tested each module's effectiveness on twenty randomly selected identities. As shown in Table 5, removing any module results in a certain performance degradation, manifested as reduced copyright protection effectiveness or visual quality. In particular, the proposed contrastive learning-based optimization strategy has the most significant effect in resisting unauthorized personalization, while the utilization of frequency components contributes to enhancing the visual quality of copyrighted images. Corresponding qualitative results are presented in the appendix. + +# 5. Conclusion + +In this paper, we identify three key challenges in existing anti-personalization methods. Our proposed invertible copyright protection framework takes a novel approach to addressing critical issues, including unauthorized personalization, untraceable copyright, lack of authorization, and image quality degradation. Considering the varying impact of watermark embedding across frequency components, we perform operations in the frequency domain and introduce a novel information exchange mechanism within our invertible neural network. And we design a specialized contrastive learning-based optimization strategy that simultaneously resists personalized infringement while ensuring copyright verifiability. Extensive experiments demonstrate its effectiveness. As infringement techniques become increasingly sophisticated, enhancing the stealthiness and robustness of proactive defense mechanisms remains a crucial research challenge worthy of further investigation. + +# Acknowledgments + +# References + +[1] Namhyuk Ahn, Wonhyuk Ahn, KiYoon Yoo, Daesik Kim, and Seung-Hun Nam. Imperceptible protection against style imitation from diffusion models. CoRR, 2024. 3 +[2] Jie An, Siyu Huang, Yibing Song, Dejing Dou, Wei Liu, and Jiebo Luo. Artflow: Unbiased image style transfer via reversible neural flows. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 862-871, 2021. 3 +[3] Tu Bui, Shruti Agarwal, Ning Yu, and John Collomosse. Rosteals: Robust steganography using autoencoder latent space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 933-942, 2023. 1, 3 +[4] Qiong Cao, Li Shen, Weidi Xie, Omkar M Parkhi, and Andrew Zisserman. Vggface2: A dataset for recognising faces across pose and age. In 2018 13th IEEE international conference on automatic face & gesture recognition (FG 2018), pages 67-74. IEEE, 2018. 5 +[5] Zihan Chen, Ziyue Wang, Jun-Jie Huang, Wentao Zhao, Xiao Liu, and Dejian Guan. Imperceptible adversarial attack via invertible neural networks. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 414-424, 2023. 3 +[6] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4690-4699, 2019. 6 +[7] Jiankang Deng, Jia Guo, Evangelos Ververas, Irene Kotsia, and Stefanos Zafeiriou. Retinaface: Single-shot multi-level face localisation in the wild. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5203-5212, 2020. 6 +[8] Prafulla Dhariwal and Alex Nichol. Diffusion models beat gans on image synthesis. 34:8780-8794, 2021. 1 +[9] Laurent Dinh, David Krueger, and Yoshua Bengio. Nice: Non-linear independent components estimation. Computer Science, 2014. 3 +[10] Laurent Dinh, Jascha Sohl-Dickstein, and Samy Bengio. Density estimation using real nvp. In International Conference on Learning Representations, 2017. 3 + +This work was supported in part by the National Natural Science Foundation of China (U24B20173,62406047), the Natural Science Foundation of Chongqing for Distinguished Young Scholars (CSTB2022NSCQ-JQX0001), the Science and Technology Research Program of Chongqing Education Commission (KJQN202400608), Special Grants for Postdoctoral Research Program of Chongqing (2023CQBSHTB3160), the Natural Science Foundation of Chongqing for Innovation and Development Joint Foundation (CSTB2024NSCQ-LMX0002). +[11] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022. 2 +[12] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6 +[13] Ajay Jain Ho, Jonathan and Pieter Abbeel. Denoising diffusion probabilistic models. pages 6840-6851, 2020. 1 +[14] Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7132-7141, 2018. 4 +[15] Junpeng Jing, Xin Deng, Mai Xu, Jianyi Wang, and Zhenyu Guan. Hinet: Deep image hiding by invertible network. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4733-4742, 2021. 3 +[16] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, pages 694-711. Springer, 2016. 3 +[17] Durk P Kingma and Prafulla Dhariwal. Glow: Generative flow with invertible 1x1 convolutions. Advances in neural information processing systems, 31, 2018. 3 +[18] Zhang B. Zhang R. Shechtman E. Kumari, N. and J. Y. Zhu. Multi-concept customization of text-to-image diffusion. In /Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1931-1941, 2023. 2 +[19] Chumeng Liang and Xiaoyu Wu. Mist: Towards improved adversarial examples for diffusion models. arXiv preprint arXiv:2305.12683, 2023. 2, 3 +[20] Chumeng Liang, Xiaoyu Wu, Yang Hua, Jiaru Zhang, Yiming Xue, Tao Song, Zhengui Xue, Ruhui Ma, and Haibing Guan. Adversarial example does good: Preventing painting imitation from diffusion models via adversarial examples. In International Conference on Machine Learning, ICML 2023, 23-29 July 2023, Honolulu, Hawaii, USA, pages 20763-20786. PMLR, 2023. 1, 2, 3 +[21] Hanwen Liu, Zhicheng Sun, and Yadong Mu. Countering personalized text-to-image generation with influence watermarks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12257-12267, 2024. 3 +[22] Yixin Liu, Chenrui Fan, Yutong Dai, Xun Chen, Pan Zhou, and Lichao Sun. Metacloak: Preventing unauthorized subject-driven text-to-image diffusion-based synthesis via meta-learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24219-24228, 2024. 1, 2, 3, 5, 6, 8 +[23] Zihao Liu, Qi Liu, Tao Liu, Nuo Xu, Xue Lin, Yanzhi Wang, and Wujie Wen. Feature distillation: Dnn-oriented jpeg compression against adversarial examples. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 860-868. IEEE, 2019. 7 + +[24] C. M., Lou, C. D., and Chang. Dual-wrapped digital watermarking scheme for image copyright protection. COMPUTERS AND SECURITY, 2007. 1 +[25] Aamir Mustafa, Salman H Khan, Munawar Hayat, Jianbing Shen, and Ling Shao. Image super-resolution as a defense against adversarial attacks. IEEE Transactions on Image Processing, 29:1711-1724, 2019. 7 +[26] English Z. Lacey K. Blattmann A. Dockhorn T. Müller J. Podell, D. and year=2023 Rombach, R. booktitle=The Twelfth International Conference on Learning Representations. Sdxl: Improving latent diffusion models for high-resolution image synthesis. 1 +[27] Phillip Pope, Yogesh Balaji, and Soheil Feizi. Adversarial robustness of flow-based generative models. In International Conference on Artificial Intelligence and Statistics, pages 3795-3805. PMLR, 2020. 3 +[28] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1 +[29] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22500-22510, 2023. 2 +[30] Hadi Salman, Alaa Khaddaj, Guillaume Leclerc, Andrew Ilyas, and Aleksander Madry. Raising the cost of malicious ai-powered image editing. pages 29894-29918, 2023. 5 +[31] Matthew Tancik, Ben Mildenhall, and Ren Ng. Stegastamp: Invisible hyperlinks in physical photographs. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2117-2126, 2020. 1, 3 +[32] Thanh Van Le, Hao Phung, Thuan Hoang Nguyen, Quan Dao, Ngoc N Tran, and Anh Tran. Anti-dreambooth: Protecting users from personalized text-to-image synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2116-2127, 2023. 2, 3, 5, 6, 8 +[33] Feifei Wang, Zhentao Tan, Tianyi Wei, Yue Wu, and Qidong Huang. Simac: A simple anti-customization method for protecting face privacy against text-to-image synthesis of diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12047-12056, 2024. 3 +[34] Xintao Wang, Ke Yu, Shixiang Wu, Jinjin Gu, Yihao Liu, Chao Dong, Yu Qiao, and Chen Change Loy. Esrgan: Enhanced super-resolution generative adversarial networks. In Proceedings of the European conference on computer vision (ECCV) workshops, pages 0–0, 2018. 4 +[35] X. Y Wang and H Zhao. A novel synchronization invariant audio watermarking scheme based on dwt and dct. IEEE Transactions on Signal Processing, 54:p.4835-4840, 2006. 3 +[36] R. B. Wolfgang, C. I. Podilchuk, and E. J. Delp. Perceptual watermarks for digital images and video. Proceedings of the IEEE, 87(7):1108-1126, 2002. 1 + +[37] Mingqing Xiao, Shuxin Zheng, Chang Liu, Zhouchen Lin, and Tie-Yan Liu. Invertible rescaling network and its extensions. International Journal of Computer Vision, 131(1): 134-159, 2023. 3 +[38] Youmin Xu, Chong Mou, Yujie Hu, Jingfen Xie, and Jian Zhang. Robust invertible image steganography. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7875-7884, 2022. 3 +[39] Hang Yang, Yitian Xu, Xuhua Liu, and Xiaodong Ma. Pris: Practical robust invertible network for image steganography. Engineering Applications of Artificial Intelligence, 133: 108419, 2024. 3 +[40] Peifei Zhu, Tsubasa Takahashi, and Hirokatsu Kataoka. Watermark-embedded adversarial examples for copyright protection against diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24420–24430, 2024. 1, 2, 3, 5, 6, 8 \ No newline at end of file diff --git a/ICCV/2025/Who Controls the Authorization_ Invertible Networks for Copyright Protection in Text-to-Image Synthesis/images.zip b/ICCV/2025/Who Controls the Authorization_ Invertible Networks for Copyright Protection in Text-to-Image Synthesis/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..919246065aca031c8d815a669a3df4ffdd031e2c --- /dev/null +++ b/ICCV/2025/Who Controls the Authorization_ Invertible Networks for Copyright Protection in Text-to-Image Synthesis/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9d5a12f881d85647a6cb3e639fb0d980fe465f0309b88d704447635eb5b1414c +size 744724 diff --git a/ICCV/2025/Who Controls the Authorization_ Invertible Networks for Copyright Protection in Text-to-Image Synthesis/layout.json b/ICCV/2025/Who Controls the Authorization_ Invertible Networks for Copyright Protection in Text-to-Image Synthesis/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..03b610fa112d04b9e7e570f1b54190f32eecb8c8 --- /dev/null +++ b/ICCV/2025/Who Controls the Authorization_ Invertible Networks for Copyright Protection in Text-to-Image Synthesis/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a24b5103566a6899bef72d610cd550efa75b310790e813bc7af3a967034d670e +size 337556 diff --git a/ICCV/2025/Who is a Better Talker_ Subjective and Objective Quality Assessment for AI-Generated Talking Heads/31ef8567-3c00-4c44-b71f-af1f2b0b5c24_content_list.json b/ICCV/2025/Who is a Better Talker_ Subjective and Objective Quality Assessment for AI-Generated Talking Heads/31ef8567-3c00-4c44-b71f-af1f2b0b5c24_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..1dc718f58c85b8b13c1cd7f96a324eb63118d681 --- /dev/null +++ b/ICCV/2025/Who is a Better Talker_ Subjective and Objective Quality Assessment for AI-Generated Talking Heads/31ef8567-3c00-4c44-b71f-af1f2b0b5c24_content_list.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e4b5a8d80a24580143877f99a8092b12021d4943c5bb03dfb4ac79842374e179 +size 89056 diff --git a/ICCV/2025/Who is a Better Talker_ Subjective and Objective Quality Assessment for AI-Generated Talking Heads/31ef8567-3c00-4c44-b71f-af1f2b0b5c24_model.json b/ICCV/2025/Who is a Better Talker_ Subjective and Objective Quality Assessment for AI-Generated Talking Heads/31ef8567-3c00-4c44-b71f-af1f2b0b5c24_model.json new file mode 100644 index 0000000000000000000000000000000000000000..bba03bc6ebbd503b1185cc86aa9d4ca8527d9629 --- /dev/null +++ b/ICCV/2025/Who is a Better Talker_ Subjective and Objective Quality Assessment for AI-Generated Talking Heads/31ef8567-3c00-4c44-b71f-af1f2b0b5c24_model.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9a96b89a78769fdba05c722f89a28caf3b7e46978cdc17629933078593034fb4 +size 113945 diff --git a/ICCV/2025/Who is a Better Talker_ Subjective and Objective Quality Assessment for AI-Generated Talking Heads/31ef8567-3c00-4c44-b71f-af1f2b0b5c24_origin.pdf b/ICCV/2025/Who is a Better Talker_ Subjective and Objective Quality Assessment for AI-Generated Talking Heads/31ef8567-3c00-4c44-b71f-af1f2b0b5c24_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..f1ccbaf129b62046c0b88af63e9efb8636744b49 --- /dev/null +++ b/ICCV/2025/Who is a Better Talker_ Subjective and Objective Quality Assessment for AI-Generated Talking Heads/31ef8567-3c00-4c44-b71f-af1f2b0b5c24_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:55350b9417de1b07b0581bfa23ec82848430a955952aa0e1b937b7efcd504da8 +size 1912750 diff --git a/ICCV/2025/Who is a Better Talker_ Subjective and Objective Quality Assessment for AI-Generated Talking Heads/full.md b/ICCV/2025/Who is a Better Talker_ Subjective and Objective Quality Assessment for AI-Generated Talking Heads/full.md new file mode 100644 index 0000000000000000000000000000000000000000..3bfc412edb31cae0e71c85c93e3bf51c6b97d26e --- /dev/null +++ b/ICCV/2025/Who is a Better Talker_ Subjective and Objective Quality Assessment for AI-Generated Talking Heads/full.md @@ -0,0 +1,335 @@ +# Who is a Better Talker: Subjective and Objective Quality Assessment for AI-Generated Talking Heads + +Yingjie Zhou $^{1,2}$ Jiezhang Cao $^{*3}$ Zicheng Zhang $^{1,2}$ Farong Wen $^{1,2}$ + +Yanwei Jiang $^{1,2}$ Jun Jia $^{1,2}$ Xiaohong Liu $^{*1,2}$ Xiongkuo Min $^{*,1,2}$ Guangtao Zhai $^{*,1,2}$ + +1 Shanghai Jiao Tong University 2 PengCheng Laboratory 3 Harvard Medical School + +# Abstract + +Speech-driven methods for portraits are figuratively known as "Talkers" because of their capability to synthesize speaking mouth shapes and facial movements. Especially with the rapid development of the Text-to-Image (T2I) models, AI-Generated Talking Heads (AGTHs) have gradually become an emerging digital human media. However, challenges persist regarding the quality of these talkers and AGTHs they generate, and comprehensive studies addressing these issues remain limited. To address this gap, this paper presents the largest AGTH quality assessment dataset THQA-10K to date, which selects 12 prominent T2I models and 14 advanced talkers to generate AGTHs for 14 prompts. After excluding instances where AGTH generation is unsuccessful, the THQA-10K dataset contains 10,457 AGTHs. Then, volunteers are recruited to subjectively rate the AGTHs and give the corresponding distortion categories. In our analysis for subjective experimental results, we evaluate the performance of talkers in terms of generalizability and quality, and also expose the distortions of existing AGTHs. Finally, an objective quality assessment method based on the first frame, Y-T slice and tone-lip consistency is proposed. Experimental results show that this method can achieve state-of-the-art (SOTA) performance in AGTH quality assessment. The work is released at https://github.com/zyj-2000/Talker. + +# 1. Introduction + +Digital humans represent an emerging digital media technology focused on generating realistic representations of virtual characters endowed with human-like characters [67]. Currently, the majority of high-quality digital humans are predominantly crafted and manipulated by skilled designers, necessitating extensive expertise and experience. In particular, the design process is cumbersome and time-consuming in terms of character modeling and facial an + +![](images/17ab3700f2e8bb525ef8833fcf82a2cacd03951a49373c64f53805d4838d3fc9.jpg) +Figure 1. Manual approach to digital human head design versus Text-to-Image and Talker-based approaches. + +imation. This method of manual design obviously suffers from low efficiency and high cost, hindering the popularization and promotion of digital humans. Fortunately, advancements in artificial intelligence (AI) have significantly facilitated the design of digital humans. On one hand, various types of text-to-image (T2I) models [1-3, 9-11, 25, 31, 32, 34, 35] have been able to generate various types of character images, allowing for a more diverse appearance of digital humans. On the other hand, a variety of speech-driven methods [6, 7, 23, 33, 42, 43, 46, 47, 49-51, 63, 64], which can be viewed as "Talkers," have been developed to achieve the effect of Talking Head (TH). Although these talkers have improved the efficiency of digital human design, they inevitably face a variety of quality problems, which adversely impact the user experience. Therefore, it is imperative to evaluate these Talkers to provide objective and reliable reference metrics for various generative methods, thereby fostering the ongoing development of the digital human domain and enhancing user experience with AI-Generated Talking Head (AGTH) videos. More specifically, considering the AGTH generation process, two aspects should be considered to assess the quality of the talkers. A) Generalization: A better talker should produce high-quality AGTHs across a wide range of portrait images. + +Table 1. The comparison of digital human databases and THQA-10K. + +
DatabaseModalScaleMethodsDistortionsDescription
DHH-QA [55]Mesh + UV1,540Zhang et al. [55], Zhou et al. [69]7 model distortionsScanned Real Human Heads
DDHQA [54]Mesh + UV800Zhang et al. [56], Chen et al. [5]7 model distortions and 2 motion distortionsDynamic 3D Digital Human
SJTU-H3D [53]Mesh + UV1,120Zhang et al. [53]7 model distortionsStatic 3D Digital Humans
6G-DTQA [60]Mesh + UV400Zhang et al. [60]3 model distortions and 2 stream media issuesDynamic 3D Digital Human
THQA-3D [73]Mesh + UV1,000Zhou et al. [73]5 stream media issuesScanned Real Human Heads
CDHQA [77]Video254None3 generative distortionsInteractive Digital Human
THQA [72]Video + Audio800None9 generative distortionsAI-Generated Talking Heads
AHQA [78]Video1200Zhou et al. [78]4 generative distortionsAnimated Humans
ReLU-QA [75]Image840None4 relighted methodsRelighted Human Heads
MEMO-Bench [71]Image7,145NoneSentimental ErrorEmotional Human Heads
THQA-10K (Ours)Image + Audio10,457None10 generative distortionsAI-Generated Talking Heads
+ +B) Quality: A better talker must generate superior AGTHs for identical images and speech inputs. + +Unfortunately, there has been limited research addressing these quality concerns and proposing credible solutions. To tackle this challenge, this paper first introduces the largest and most comprehensive AGTH Quality Assessment dataset named THQA-10K. This dataset encompasses a diverse selection of character materials, considering various ages and genders, and includes 14 prompts for character image generation. Each prompt is paired with five tailored speech sentences serving as driving audio. Furthermore, the dataset employs 12 leading T2I models and 14 talkers to generate AGTHs, resulting in a total of 10,457 instances within THQA-10K. Subsequently, volunteers are recruited to conduct subjective evaluations of the AGTHs, focusing on both distortion categories and visual quality scores. The findings reveal 10 distinct types of distortions among the AGTHs, alongside significant variations in quality depending on the talkers utilized. This highlights the critical need for quality assessments in this field. Leveraging the THQA-10K dataset and subjective ratings, we propose FSCD, an objective quality assessment method for AGTH based on the first frame, Y-T Slice [38] and tone-lip consistency. Experimental results validate the efficacy and superiority of this method, offering reliable objective metrics for the continued advancement of AGTH technologies. The principal contributions of this paper are as follows: + +- The THQA-10K dataset, comprising 10,457 AGTHs, has been constructed. To our knowledge, this dataset represents the largest collection created for the purpose of AGTH quality assessment, incorporating 12 prominent T2I models and 14 speech-driven methods. +- An objective quality assessment algorithm, referred to as FSCD, has been designed. This method integrates quality features from the first frame, Y-T slice, and tone-lip consistency to deliver an effective and robust objective evaluation of AGTH quality. +- The proposed THQA-10K dataset is representative and comprehensive to advance the field of digital human design. The designed FSCD method can achieve state-of-the-art (SOTA) objective quality assessment performance, which provides a reliable quality indicator for the field. + +# 2. Related Works + +# 2.1. Talker: Talking Head Driven Methods + +Head-driven animation has been a significant area of research within the field of computer animation, primarily because the human head encompasses a wealth of facial details, expressive movements, and identity information. Traditional methods typically utilize computer-aided animation software, such as Maya or Blender. These approaches require the binding of facial bones to the character model and the establishment of corresponding controllers to manipulate facial movements. Subsequently, keyframe animation techniques are employed, where appropriate keyframes are set for each controller to create various mouth shapes corresponding to different phonemes. This process is notably time-consuming and often necessitates extensive debugging. To address these challenges, the current standard solution in film and television production involves the use of facial capture sensors that track the key points of real THs [12, 21, 24, 37]. The captured data is then imported into computer-aided animation software for further refinement and design. However, high-precision facial capture sensors are prohibitively expensive. + +In recent years, advancements in AI have opened new avenues for the production of THs. The emergence of T2I models has not only simplified character image design but has also introduced speech-driven methodologies that directly apply speech to face images for AGTH generation. This approach alleviates a series of cumbersome processes and reduces equipment costs significantly. Speech-driven methods can be further classified into image-based [6, 13, 23, 43, 43, 49] and video-based [7, 33, 42, 46, 47, 50, 51, 63] techniques, depending on the input modality. Nonetheless, the absence of human design and oversight raises concerns regarding the quality of AGTHs, which is the central focus of this paper. + +# 2.2. Digital Human Quality Assessment + +With the development of digital human quality assessment of digital humans has become an emerging research component. As shown in Table 1, many relevant datasets and ob + +![](images/9d1d3f96dc3d3e8334a0a166c087ccadbec14eb4dde21d9973f47e1eec0d0b95.jpg) +(a) + +![](images/7f387c8fbfca7ad2f3a970fddc195ef7d50d24310f8f2912b950e5ae2c2b3e84.jpg) +(b) +Figure 2. Features of selected prompts and speeches. (a) Word cloud of selected prompts. (b) Word cloud of the speech text content. (c) Results of the resonance peak estimation for the selected speech. Speech samples that show resonance peak merging have been removed. + +![](images/e18486633c4c7db54d8706573f8a38949fcb9d5aaf400773303431728f779520.jpg) +(c) + +jective quality assessment methods have been established, providing a rich data base and feasible solutions for the development of digital human quality assessment. Nonetheless, attention to AGTH is still lacking, with only Zhou et al. constructing a THQA dataset [72], which also fails to provide an effective quality assessment metric. Actually, PSNR and SSIM [44] are still two commonly used quality metrics in the field of THs. However, it is clear that these two metrics are no longer suitable for the evaluation of AGTHs, due to the lack of corresponding reference videos. Although metrics such as Frechet Inception Distance (FID) [14], LSE-C [8], LES-D [8], and CPBD [30] have also been used in the domain of THs, these metrics only focus on a certain dimension of AGTH and do not provide a comprehensive and effective assessment of AGTH. Therefore, there is an urgent need for effective objective indicators to measure the quality of AGTH in the field of AGTH. In Zhou et al.'s experiments [72], they identified 9 common distortions present in AGTHs. However, the limitations of the THQA dataset are apparent. First, it consists of only 800 AGTHs, which constrains its ability to encompass AGTHs generated by a variety of contemporary models. Second, the exclusive reliance on StyleGAN [15, 16] for character image generation neglects the potential influence of current mainstream T2I models on quality. These constraints hinder the development of effective methods for objectively assessing AGTH quality. To address these deficiencies, this paper introduces a larger and more comprehensive THQA-10K dataset. Utilizing the THQA-10K dataset, we conduct extensive subjective experiments and design targeted objective quality assessment methods, thereby providing reliable reference metrics for AGTH quality evaluation. + +# 3. Database Construction + +# 3.1. Prompts and Speeches + +To ensure diversity and representativeness among persons and to account for potential variations in AGTHs due to differences in gender and age, we select 14 distinct prompts for character portrait generation, each assigned a unique prompt ID (PID). Fig. 2(a) displays the frequency and content of these prompts, all of which are set to “ $8k$ ” quality to + +maximize the resolution and detail of the generated characters. Additionally, photographic elements are incorporated into some prompts to enhance realism. Overall, these prompts offer a comprehensive description of various human head features. + +Aligned with the age and gender associated with each prompt, we select five corresponding speeches from the Common Voice speech dataset. Each set of five speeches is sourced from the same speaker, ensuring consistent phonological features across the selected speeches for each prompt. To illustrate the diversity in speech content and features, we perform speech recognition and feature extraction on the selected audio samples, with results presented in Fig. 2(b-c). Several observations can be made from Fig. 2(b-c): 1) The speeches encompass a broad vocabulary, predominantly comprising common words, suggesting a rich variety of phonemes; 2) The first formant frequency of most speeches falls between 600 and $1150\mathrm{Hz}$ , indicating diverse mouth shapes during articulation. The second formant frequency ranges between 1750 and $4000\mathrm{Hz}$ , reflecting differences in tongue positioning during pronunciation; 3) For each PID, the five selected speech samples display clustering in their audio features, while the samples from different PIDs are more widely separated, indicating consistency within each PID; 4) Audio durations range from 3.06 to 10.12 seconds, providing a balanced representation of short interactions and longer conversations. + +# 3.2. Generative Models + +To generate AGTHs from prompts, two types of generative models are required. The first, text-to-image (T2I) models, are designed to produce portraits based on the provided prompts, serving as the foundational material for subsequent facial animation. Although previous research [19, 58, 61] has highlighted variations in the quality of images produced by different T2I models, there has been limited targeted discussion addressing the quality of generated portraits. To comprehensively evaluate the generalization performance of talkers and to investigate the impact of differences in the quality of generated portraits on AGTHs, we select 12 prominent T2I models for image generation. + +Table 2. Details of T2I models employed for generation. + +
TypeLabelT2I ModelYearOutput Resolution
Closed sourceDL3Dalle3 [3]20231,024×1,024
MJ6MidjourneyV6 [25]20231,024×1,024
IDGIdeagram [1]20241,024×1,024
Open sourceSD2Stable Diffusion 2.1 [35]2022512×512
SD1Stable Diffusion 1.5 [35]2022512×512
SDXStable Diffusion XL [32]20231,024×1,024
FCSFooocus [11]20231,024×1,024
KDSKandinsky [2]20231,024×1,024
ODEOpenDalleV1.1 [31]20231,024×1,024
PTSProteus [34]20241,024×1,024
FLUFLUX.1 [10]20241,024×1,024
SD3Stable Diffusion 3 [9]20241,024×1,024
+ +![](images/371c7469d43a33e408e822549bd801331ed45d742be58e239adcb3ba8982e9f5.jpg) +Figure 3. Generated portraits of different PIDs. + +![](images/5b3480e6868c18332de9260071350d1d9e07d9df48f80a843f02c25b8e49bbc0.jpg) + +![](images/f0e34f4d06985e3b4d7b4a557bc4c94ba82741f451248b705e85b1e903e38cb5.jpg) + +![](images/4253c4a9164df75fe2c246cabf0714f33480791700e48939d1277ccffaede9cb.jpg) + +![](images/840fd38b0021969424c5203c4354e899a571e0e30985dc05f3c1787946cd926a.jpg) + +![](images/7e849082ba6f3a535574a1a3b52e173c3ae65bcccededf05158a2b6ee9a453e6.jpg) + +![](images/8e03496061ee6b5d1c723a2cfe7d88539663653d4acd6108d8c5738b7261f627.jpg) +Figure 4. Visualization of the number of successful generation. (a) Portraits that can be successfully generated. The black block indicates success while the white one indicates failure. (b) Number of AGTHs that can be successfully generated. + +Details regarding the selected T2I models are presented in Table 2, with a subset of the generated portraits illustrated in Fig. 3. + +Another critical generative model is represented by speech-driven methods, commonly referred as talkers. To ensure that the constructed dataset encompasses a wide range of existing mainstream driving methods, 14 talkers are utilized to target the generated character images for AGTH generation. Table 3 provides details of the selected talkers, including their respective output resolutions. It is important to note that variations in output resolution exist and the selected driving methods are totally implemented using source code provided by the original authors. Additionally, for the video-based driving methods, the input video consists of a repeated driving image with a duration set to one second and a frame rate of 25 frames per second. Ultimately, a total of 10,457 AGTHs are successfully generated for 14 prompts and their corresponding 70 speeches, culminating in the creation of the THQA-10K dataset. + +# 3.3. Data Statistics and Subjective Experiment + +To assess the generalization capability of various T2I models and talkers, we conduct a statistical analysis of the number of portraits and AGTHs successfully generated by these generative models, as illustrated in Fig. 4. From Fig. 4(a), it is evident that the two T2I models, SD1 and SD2, demonstrate limited effectiveness in generating portraits across a range of prompts. However, for IDG, its' suboptimal performance can be attributed to the constraints posed by prompt-sensitive vocabulary. Additionally, Fig. 4(b) highlights the + +Table 3. Details of talkers employed for generation. + +
TypeLabelMethodsYearHead MotionOutput Resolution
Image-basedMIMakeIttalk [64]2020✓256×256
AHAuido2Head [43]2021✓256×256
STSadtalker [49]2023✓512×512
DTDreamtalk [23]2023✓256×256
ETEAT [13]2023✓256×256
EMEchoMimic [6]2024✓512×512
Video-basedWLWav2Lip [33]2020✗1,024×1,024
VRVideo-Retalking [7]2022✗1,024×1,024
SHStyleHeat [47]2022✗1,024×1,024
DNDINet [51]2023✗1,024×1,024
ILIP-LAP [63]2023✗1,024×1,024
TLTalkLip [42]2023✗1,024×1,024
MTMuseTalk [50]2024✗1,024×1,024
EGEmoGen [46]2024✗1,024×1,024
+ +![](images/5eafa2481bd172587c6b8675b104cc1c8911fd4a29b24144dfeb689234f4f71f.jpg) +(a) + +![](images/5ac181fd09da4a74c6c4ae0b468ebe307c62e06b09174c106ab63a476d89254f.jpg) +(b) + +generalization performance of different talkers. Notably, the SH and TL talkers exhibit the strongest generalization abilities. Remarkably, the disparity in the total number of AGTHs generated between the highest and lowest performing talkers reaches 128, underscoring a significant variation in generalization performance across the different talkers. + +To investigate the quality of AGTHs and their distortions in detail, we recruit 13 male and 12 female participants to conduct subjective quality assessments of all AGTHs in the THQA-10K dataset. The subjective evaluation is carried out in a well-controlled laboratory in accordance with the guidelines outlined in ITU-R BT.500-13 [4]. Participants can view AGTHs on an iMac monitor with a resolution of $4,096 \times 2,304$ . Given that the AGTHs include audio components, a wired headset is utilized to ensure low-latency and high-quality audio transmission, while also minimizing potential interference between participants due to the audio output. The AGTHs are organized into 100 phases, with each phase comprising at most 120 AGTHs. To mitigate visual fatigue and discomfort associated with prolonged viewing, all participants are required to take a 15-minute break after completing each phase. Furthermore, each participant is limited to a maximum of 6 assessment phases per day. + +# 3.4. Data Processing + +At the end of the subjective experiment, we receive a total of $261,425 = 25 \times 10,457$ subjective ratings. Each rating can be described as $\{s_{ij}, D_{ij}\}$ , where $s_{ij}$ and $D_{ij}$ are the quality rating and labeled distortion of the $j$ -th AGTH by the + +![](images/a5d62194cecbfe625b9d72c031782ad277c4f38d5eb2b12cb325d9aa7a94ba91.jpg) + +![](images/f30f09af0ae7b4cde747c585fa3d11f200035bbafec7f7d0102501f296c3fcea.jpg) + +![](images/4dee46fc979457ba85f898f4e28fbf07fd09caa890d5b4b032f0c93baa6306c4.jpg) +(a) +(c) +Figure 5. Distribution of MOSs. + +![](images/c60360514910c844887e8e0b99fd38929c984b46d6f26f15664a1f4f9b24bdad.jpg) +(b) +(d) + +$i$ -th subject. In particular, $D_{ij}$ is a ten-dimensional 0-1 distortion vector, with each dimension denoting a corresponding distortion type. According to existing works [26, 53-55, 59, 62, 65, 66, 68, 70, 74-76], $s_{ij}$ is processed as z-scores according to the following equation: + +$$ +z _ {i j} = \frac {s _ {i j} - \mu_ {i}}{\sigma_ {i}}, \tag {1} +$$ + +where $\mu_{i} = \frac{1}{N_{i}}\sum_{j = 1}^{N_{i}}s_{ij}$ , $\sigma_{i} = \sqrt{\frac{1}{N_{i} - 1}\sum_{j = 1}^{N_{i}}(s_{ij} - \mu_{i})}$ and $N_{i}$ represents the total number of AGTHs evaluated by subject $i$ . Following the rejection procedure outlined in [4], ratings from unreliable subjects are excluded. The remaining z-scores $z_{ij}$ are linearly rescaled to the range [0, 5]. Finally, the mean opinion scores (MOSs) for the $j$ -th AGTH are computed by averaging the rescaled z-scores. For the distortion vector $D_{ij}$ , the summation operation is employed to count the distortion type of the $j$ -th AGTH: + +$$ +D _ {j} = \sum_ {i = 1} ^ {N _ {j}} D _ {i j}, \tag {2} +$$ + +where $N_{j}$ denotes the number of subjects who classified distortions of the $j$ -th AGTH. A threshold vector $T_{j}$ , defined as $N_{j} / 2$ across all ten dimensions, is utilized to derive the final distortions $\mathcal{D}_j$ : + +$$ +\mathcal {D} _ {j} = u \left(D _ {j} - T _ {j}\right), \tag {3} +$$ + +where $u(\cdot)$ denotes the step function. The whole process can be interpreted as for each category of distortion for each AGTH, more than half of the subjects need to believe that the distortion exists before it is officially acknowledged. + +# 3.5. Mean Opinion Score Analysis + +Based on the results of the subjective experiments, we conduct a comprehensive analysis of mean opinion scores (MOSs) and distortion of AGTHs in order to evaluate the various types of talkers. Initially, the overall MOS distribution is plotted as shown in Fig. 5(a). To further analyze the effects of various possible factors on the MOSs, + +the violin plots shown in Fig. 5(b-d) provide a more intuitive picture of the relationship between different PIDs, T2Is, talkers and MOSs. By observing Fig. 5, we can draw some valuable conclusions: 1) The majority of AGTHs received quality scores centered around 3, indicating that some talkers are capable of meeting user expectations regarding audiovisual quality. However, it is noteworthy that only a limited number of AGTHs achieved MOSs exceeding 4 points, while a significant proportion of AGTHs received low-quality scores in the range of 0 to 2 points. This suggests considerable potential for improving the audiovisual quality produced by talkers; 2) AGTHs generated from different PIDs and T2Is exhibit a similar distribution of MOSs. This observation indicates that the PIDs and T2Is utilized in constructing the THQA-10K dataset are both representative and universal, thereby facilitating generalization to other PIDs and T2Is. Additionally, it implies that variations in PIDs and T2Is are not the primary determinants affecting the quality of AGTHs; 3) There are significant disparities in the quality distribution of AGTHs produced by different talkers. Notably, two types of talkers, EM and ST, generate the highest quality AGTHs, whereas two other types, EG and SH, yield lower quality outputs. + +# 3.6. Distortion Visualization & Analysis + +An examination of the AGTHs in the THQA-10K dataset reveals that the 9 distortion types previously identified in the THQA dataset by Zhou et al. [72] remain prevalent. Additionally, new distortions have emerged, predominantly characterized by misalignment of facial keypoints. To illustrate the impact of each distortion type and quantify their occurrence, we select representative samples for visualization and cataloged the distortions identified in the subjective assessments. The results are depicted in Fig. 6. From this analysis, several key insights can be derived: 1) A total of 17,191 distortions are identified across 10,457 AGTHs, indicating that distortions are a common issue within existing AGTHs. Notably, individual AGTHs often exhibit a combination of multiple distortion types; 2) Among the various distortion types, blur (BL), noise (NI) and artifacts (AF) remain the predominant issues, suggesting that current talkers still face limitations in these areas; 3) The use of Y-T slices to represent the two distortion types, little lip motion (LLM) and muscle twitch (MT), as demonstrated by Zhou et al., offers a more pronounced indication of the visual effects associated with these distortions. Specifically, LLM manifests as parallel lines or minor fluctuations in the mouth texture within the Y-T slice, while MT is characterized by periodic repetitive textures. The newly identified misaligned keypoints (MK) distortion results in a noticeable displacement of facial features, leading to severe distortion. + +To further investigate the frequency of various distortion types across different talkers, we plot a heatmap to visu + +![](images/ff57ed19a27dfc24623c40a4801fe4ad5dd69caa6db20f34c721ba435b236d3f.jpg) +Figure 6. Visualization of distortions. The left side shows how often each distortion occurs, while the right side shows typical cases. + +![](images/02c1436c90d743b86e22c2466c9c6df02b52ac6d7840c9b558342b5331ba8e51.jpg) + +![](images/d8c297199e6bcf43b6c7cd15c14b417df9c076cc084a4eeeda327582be6b00e5.jpg) + +![](images/76105e9121e98350da77d4ce3596cc2ff97c0c9183d82544c62313cc90992fe8.jpg) +Figure 7. Statistical analysis of different distortions. + +![](images/16ac9c2383a309773b189ba28bd0bda45b66905b4b2ed611646c605b09ae4438.jpg) + +ally illustrate the relationship between talkers and distortions. The data presented in Fig. 7(a) reveal another two key findings: 1) Specific distortions tend to be concentrated among one or a few talkers. For instance, LLM predominantly occurs with the IL talker, while distortion from distorted background (DB) is mainly observed with the ET talker, and MK is frequently associated with the SH and EG talkers; 2) Among all talkers included in the comparison, ST and EM exhibit the least frequency of distortion, highlighting their superiority and explaining why these two methods achieve the highest MOSs. Finally, in conjunction with the MOS analysis method of Sec. 3.5, the effect of each type of distortion on the MOSs is plotted as Fig. 7(b). The results indicate that different distortion types have varying effects on human audiovisual perception. Notably, the three most common distortions do not significantly detract from the overall quality of AGTHs, and in some cases, they still yield high MOSs. Conversely, MK has the most pronounced negative impact on AGTH quality, primarily due to the intolerable displacement of facial features, resulting in significant distortion. + +# 4. Proposed Method + +# 4.1. First Frame & Slice Process + +Considering that each AGTH video is generated by a stationary talker, it is reasonable to assert that the video qual + +ity remains relatively stable throughout the duration of the AGTH. Consequently, the initial frame of an AGTH provides a rich set of spatial features suitable for quality assessment. In terms of temporal feature selection, this paper proposes a slice-based temporal feature extraction scheme, as illustrated in Fig. 8. Firstly, we advocate for the use of Y-T slices over X-T slices, as Y-T slices capture more comprehensive facial information due to the inherent symmetry of the face. Additionally, to ensure that the tangent line intersects the mouth, which is a critical aspect of AGTHs, we utilize the image generated by the T2I model as a reference and apply landmark detection to identify the key points $p_k$ of the mouth within the generated image. Subsequently, we compute the coordinates of the mouth's centroid $C_o$ : + +$$ +C _ {o} = \frac {1}{K} \sum_ {k = 1} ^ {K} p _ {k}, \tag {4} +$$ + +where $K$ denotes the total number of key points on the face. However, this derived center point cannot be directly employed to guide the slicing of corresponding AGTHs, as the videos produced by different talkers exhibit varying resolutions $R_{v}$ . Therefore, a scale transformation is applied to determine the location of the mouth centroid in the AGTH videos: + +$$ +C _ {v} = \frac {R _ {v}}{R _ {o}} C _ {o}, \tag {5} +$$ + +where $R_{o}$ is the resolution of the generated portrait and $C_{v}$ denotes the center of the human face in AGTHs. Following this, the entire AGTH video can be projected as a Y-T slice within the three-dimensional space defined by XY-T coordinates. Ultimately, the Y-T slice is resized to match the resolution of the first frame, facilitating subsequent processing. This entire slicing process underscores that the Y-T slice effectively captures temporal features over the video duration, in contrast to conventional frame extraction methods. + +![](images/c8843aba77a55c26e173c0a684547801835915d464a227df80ff6346643ecd7e.jpg) +Figure 8. The proposed framework for FSCD. The method mainly consists of four compartmentalized modules: the Y-T slice process, tone-lip consistency detection, backbone feature extraction, and quality regression. + +# 4.2. Tone-lip Consistency Detection + +AGTH is a media that exhibits a high sensitivity to audio-visual synchronization, particularly in relation to the synchronization between lip movements and speech. To address this sensitivity, we initially perform a cropping of the mouth region in the AGTH and extracted Mel-frequency cepstral coefficients (MFCCs) from the accompanying audio. Subsequently, we employ the classical SyncNet [8] to assess audio-lip consistency, yielding two key outputs: lip sync error confidence (LSE-C) and distance (LSE-D). In contrast to existing methodologies that integrate these two features directly into the final quality regression layer, this study proposes an innovative approach whereby the two LSE features are expanded into a tensor of dimensions corresponding to those of the first frame. This tensor is then utilized as two additional images within the first frame, enabling the backbone network to learn the relative significance of tonal lip consistency autonomously, rather than relying on predetermined weights based on prior knowledge. + +# 4.3. Backbone & Quality Regression + +The obtained First frame, Y-T Slice, LSE-C, and LSE-D can be collectively viewed as a new image called FSCD. Given the excellent performance achieved by the swim-transformer (Swin-T) [22, 52, 57, 79] in several computer vision tasks, Swin-T is used to extract quality features from FSCDs. During the training phase, the predicted audiovisual quality is compared with the actual MOS using the Mean Squared Error (MSE) as the loss function, facilitating the gradual optimization of the algorithm for enhanced performance: + +$$ +L o s s = \frac {1}{n} \sum_ {l = 1} ^ {n} \left(\hat {Q} _ {l} - Q _ {l}\right) ^ {2}, \tag {6} +$$ + +where $\hat{Q}_l$ and $Q_{l}$ represent the predicted quality and MOS of $l$ th AGTH, and $n$ indicates the size of training batch. + +# 5. Experiments + +# 5.1. Experiment Details & Criteria + +To validate the effectiveness of the proposed method, we select 15 quality assessment algorithms applicable to AGTHs for comparison. This selection includes 4 classical image quality assessment (IQA) methods, 2 methods for audio and lip consistency, and 9 video quality assessment (VQA) methods. Among these, RAPIQUE, SimpVQA, VSFA, FAST-VQA, and BVQA are deep learning-based methods, while the remaining methods rely on manually extracted features. All selected methods are tested on THQA-10K, THQA [72] and THQA-3D [73] datasets, utilizing a five-fold cross-validation scheme. The average test results from the five folds are recorded as the performance of each method. Notably, the five-fold data partitioning ensures that there is no content overlap, and all algorithms employed are derived from the source code provided by their authors. + +In terms of evaluation criteria, we adopt four commonly used metrics for assessing the performance of objective multimedia quality assessment algorithms: Spearman Rank Correlation Coefficient (SRCC), Kendall's Rank Correlation Coefficient (KRCC), Pearson Linear Correlation Coefficient (PLCC), and Root Mean Squared Error (RMSE). + +# 5.2. Performance Analysis + +The performance of the proposed FSCD method, along with other competing methods, on three datasets is presented in Table 4. An analysis of this table yields several key conclusions: 1) FSCD demonstrates optimal performance on all datasets, surpassing the next best algorithm by $+2\%$ in SRCC at least. This result strongly supports the effectiveness of the proposed FSCD method for assessing AGTH quality; 2) The methods that achieve suboptimal performance differ among three datasets. In contrast, FSCD consistently achieves optimal performance across all datasets, highlighting the robustness and generalizability of the proposed method; 3) Existing IQA methods, single audio-lip + +Table 4. Performance results on the proposed THQA-10K, THQA and THQA-3D databases. Best in RED, second in BLUE. + +
TypeModelsTHQA-10KTHQATHQA-3D
SRCC↑PLCC↑KRCC↑RMSE↓SRCC↑PLCC↑KRCC↑RMSE↓SRCC↑PLCC↑KRCC↑RMSE↓
IQABRISQUE [27]0.42710.44510.29931.02620.48560.59700.34540.82270.67490.74530.50600.5717
NIQE [28]0.00890.04360.00511.14920.05350.16430.04020.98110.22430.47410.12320.7707
CPBD [30]0.05530.06860.03711.14760.05750.08760.03760.99080.21450.31360.14320.8273
IL-NIQE [48]0.04900.06340.02861.14800.05370.21600.02760.97120.22930.48710.15370.7600
SyncLSE-C [8]0.07060.16340.04681.13490.00560.21090.00480.97230.17280.22970.13550.8499
LSE-D [8]0.05800.11230.03851.14310.13660.23360.08550.96710.00790.10540.00080.8684
VQAVIIDEO [29]0.13540.17820.09011.13190.17770.18910.13540.95950.10560.23080.07210.8387
TLVQM [17]0.43770.46790.30701.01300.02540.03550.02091.08530.18870.31120.12720.8240
VIDEVAL [40]0.38690.41470.27061.04310.03170.03580.02311.19160.22520.35440.15560.8118
V-BLIINDS [36]0.47400.49770.33340.99410.49490.64030.35330.79760.52980.64120.39070.6674
RAPIQUE [41]0.35760.38460.24901.05790.17890.19080.12771.01620.37480.46800.26600.7643
SimpVQA [39]0.77750.80390.59310.68320.68000.75920.50520.63610.63210.72580.47170.5983
VSFA [20]0.75370.77540.57260.73430.76010.81060.58300.59660.74630.78110.55960.5726
FAST-VQA [45]0.73510.75420.55190.80260.63890.74410.46770.69830.77780.79840.59640.5503
BVQA [18]0.63350.74050.45220.76340.72870.79850.55490.60940.78710.82980.60810.5983
FSCD (Ours)0.80660.83220.62280.63330.78120.84090.59510.50550.82350.85050.64630.4577
+ +Table 5. Ablation study results on databases, where $w/o$ stands for 'without'. Best in RED, second in BLUE. + +
DimensionTHQA-10KTHQATHQA-3D
SRCC↑PLCC↑KRCC↑RMSE↓SRCC↑PLCC↑KRCC↑RMSE↓SRCC↑PLCC↑KRCC↑RMSE↓
w/o F0.65100.68920.47260.84970.68300.75060.50350.57650.73680.78440.54510.5468
w/o S0.73300.76720.55090.73290.72050.79680.53250.52720.75060.77300.56550.5593
w/o C0.79270.81690.60490.67640.76600.81740.58600.50860.80190.84290.61600.4658
w/o D0.76100.78790.57440.72210.74620.81100.55790.51050.79150.83590.60370.4710
FSCD0.80660.83220.62280.63330.78120.84090.59510.50550.82350.85050.64630.4577
+ +consistency detection techniques, and VQA methods are constrained in their ability to evaluate AGTH quality. This limitation primarily arises from the inability of these methods to fully leverage the spatio-temporal features and multimodal information inherent in AGTHs. Moreover, compared to the extraction of temporal features through the Y-T slice, conventional approaches often struggle to adequately account for the relationships among frames. + +# 5.3. Ablation Experiments + +To further evaluate the effectiveness of each component within the FSCD framework, ablation experiments are conducted, with results summarized in Table 5, from which following observations can be drawn: 1) Each component of FSCD contributes positively to the overall performance. This enhancement can be attributed to the fact that the four components address different quality aspects, allowing for a synergistic effect; 2) Regarding the importance of individual components, the quality characteristics provided by the first frame rank highest, followed by the Y-T Slice, LSE-D, and LSE-C. This suggests that spatial features are paramount in influencing the quality of AGTHs, with temporal features and coherence features following in importance; 3) A comparative analysis of the performance results presented in Tables 4 and 5 reveals a noteworthy finding: even when utilizing only the Y-T Slice and the tone and lip coherence features, without the first frame, competitive performance is still achievable on three datasets. This highlights the validity and significance of the Y-T slice as em + +ployed by FSCD in the assessment of AGTH quality. + +# 6. Conclusion + +As digital human technology continues to advance, various speech-driven methods, commonly referred as "Talkers," have emerged to enhance the efficiency of digital human face design. To thoroughly investigate the generalization performance and generation quality of different talkers, this study introduces the THQA-10K dataset, comprising a total of 10,457 AI-Generated Talking Head (AGTH) videos. Specifically, 12 Text-to-Image (T2I) methods are employed to generate character portraits, while 14 advanced talkers are utilized to produce AGTHs. Through comprehensive data analysis and subjective experiments conducted on the THQA-10K dataset, we validate both the comprehensiveness and generalizability of the dataset. Additionally, we assess the generalization performance of existing talkers and identify potential quality issues and their distributions. Finally, we propose an objective quality assessment method named FSCD, leveraging the first frame, Y-T slice, and tone-lip consistency. Experimental results substantiate the effectiveness and robustness of FSCD, which is anticipated to inform the ongoing development of talkers. + +Acknowledgements. This work was supported in part by the Major Key Project of PCL (PCL2023A10-2), National Natural Science Foundation of China (623B2073, 62101326, 62225112, 62271312, 62132006) and STCSM (22DZ2229005). + +# References + +[1] Ideagram 2.0. https://ideogram.ai/, 2024.1,4 +[2] Shakhmatov Arseniy, Razzhigaev Anton, Nikolich Aleksandr, Arkhipkin Vladimir, Pavlov Igor, Kuznetsov Andrey, and Dimitrov Denis. kandinsky 2.1, 2023. 4 +[3] James Betker, Gabriel Goh, Li Jing, Tim Brooks, Jianfeng Wang, Linjie Li, Long Ouyang, Juntang Zhuang, Joyce Lee, Yufei Guo, et al. Improving image generation with better captions. Computer Science, 2(3):8, 2023. 1, 4 +[4] RECOMMENDATION ITU-R BT. Methodology for the subjective assessment of the quality of television pictures. International Telecommunication Union, 2002. 4, 5 +[5] Shi Chen, Zicheng Zhang, Yingjie Zhou, Wei Sun, and Xiongkuo Min. A no-reference quality assessment metric for dynamic 3d digital human. Displays, 80:102540, 2023. 2 +[6] Zhiyuan Chen, Jiajiong Cao, Zhiquan Chen, Yuming Li, and Chenguang Ma. Echomimic: Lifelike audio-driven portrait animations through editable landmark conditions. arXiv preprint arXiv:2407.08136, 2024. 1, 2, 4 +[7] Kun Cheng, Xiaodong Cun, Yong Zhang, Menghan Xia, Fei Yin, Mingrui Zhu, Xuan Wang, Jue Wang, and Nannan Wang. Videoretalking: Audio-based lip synchronization for talking head video editing in the wild. In ACM Special Interest Group on Computer Graphics Asia 2022, pages 1-9, 2022. 1, 2, 4 +[8] Joon Son Chung and Andrew Zisserman. Out of time: automated lip sync in the wild. In Asian Conference on Computer Vision 2016 Workshops, pages 251-263, 2017. 3, 7, 8 +[9] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In Forty-first International Conference on Machine Learning, 2024. 1, 4 +[10] FLUX.1-dev. https://huggingface.co/black-forest-labs/FLUX.1-dev, 2024.4 +[11] Fooocus. https://github.com/llllysaviel/Fooocus, 2023.1,4 +[12] Yun Fu, Renxiang Li, Thomas S Huang, and Mike Danielsen. Real-time multimodal human-ayar interaction. IEEE Transactions on Circuits and Systems for Video Technology, 18(4):467-477, 2008. 2 +[13] Yuan Gan, Zongxin Yang, Xihang Yue, Lingyun Sun, and Yi Yang. Efficient emotional adaptation for audio-driven talking-head generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22634-22645, 2023. 2, 4 +[14] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 3 +[15] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4401-4410, 2019. 3 + +[16] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8110-8119, 2020. 3 +[17] Jari Korhonen. Two-level approach for no-reference consumer video quality assessment. IEEE Transactions on Image Processing, 28(12):5923-5938, 2019. 8 +[18] Bowen Li, Weixia Zhang, Meng Tian, Guangtao Zhai, and Xianpei Wang. Blindly assess quality of in-the-wild videos via quality-aware pre-training and motion perception. IEEE Transactions on Circuits and Systems for Video Technology, 32(9):5944-5958, 2022. 8 +[19] Chunyi Li, Tengchuan Kou, Yixuan Gao, Yuqin Cao, Wei Sun, Zicheng Zhang, Yingjie Zhou, Zhichao Zhang, Weixia Zhang, Haoning Wu, et al. Aigia-20k: A large database for ai-generated image quality assessment. arXiv preprint arXiv:2404.03407, 2(3):5, 2024. 3 +[20] Dingquan Li, Tingting Jiang, and Ming Jiang. Quality assessment of in-the-wild videos. In ACM International Conference on Multimedia, pages 2351-2359, 2019. 8 +[21] Xiaohong Liu, Xiongkuo Min, Qiang Hu, Xiaoyun Zhang, Jie Guo, Guangtao Zhai, Shushi Wang, Yingjie Zhou, Lu Liu, Jingxin Li, et al. Ntire 2025 xgc quality assessment challenge: Methods and results. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 1389-1402, 2025. 2 +[22] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 7 +[23] Yifeng Ma, Shiwei Zhang, Jiayu Wang, Xiang Wang, Yingya Zhang, and Zhidong Deng. Dreamtalk: When expressive talking head generation meets diffusion probabilistic models. arXiv preprint arXiv:2312.09767, 2023. 1, 2, 4 +[24] Yong-You Ma, Hui Zhang, and Shou-Wei Jiang. Realistic modeling and animation of human body based on scanned data. Journal of Computer Science and Technology, 19(4): 529-537, 2004. 2 +[25] Midjourney. https://www.midjourney.com/home, 2023.1,4 +[26] Xiongkuo Min, Huiyu Duan, Wei Sun, Yucheng Zhu, and Guangtao Zhai. Perceptual video quality assessment: A survey. Science China Information Sciences, 67(11):211301, 2024. 5 +[27] Anish Mittal, Anush Krishna Moorthy, and Alan Conrad Bovik. No-reference image quality assessment in the spatial domain. IEEE Transactions on Image Processing, 21 (12):4695-4708, 2012. 8 +[28] Anish Mittal, Rajiv Soundararajan, and Alan C Bovik. Making a "completely blind" image quality analyzer. IEEE Signal Processing Letters, 20(3):209-212, 2012. 8 +[29] Anish Mittal, Michele A Saad, and Alan C Bovik. A completely blind video integrity oracle. IEEE Transactions on Image Processing, 25(1):289-300, 2015. 8 +[30] Niranjan D Narvekar and Lina J Karam. A no-reference image blur metric based on the cumulative probability of blur + +detection (cpbd). IEEE Transactions on Image Processing, 20(9):2678-2683, 2011. 3, 8 +[31] OpenDalleV1.1. https://huggingface.co/dataautogpt3/OpenDalleV1.1, 2023.1,4 +[32] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 1, 4 +[33] KR Prajwal, Rudrabha Mukhopadhyay, Vinay P Namboodiri, and CV Jawahar. A lip sync expert is all you need for speech to lip generation in the wild. In ACM International Conference on Multimedia, pages 484-492, 2020. 1, 2, 4 +[34] ProteusV0.2. https://huggingface.co/datautogpt3/ProteusV0.2, 2024.1,4 +[35] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 1, 4 +[36] Michele A Saad, Alan C Bovik, and Christophe Charrier. Blind prediction of natural video quality. IEEE Transactions on Image Processing, 23(3):1352-1365, 2014. 8 +[37] Oliver Schreer, Roman Englert, Peter Eisert, and Ralf Tanger. Real-time vision and speech driven avatars for multimedia applications. IEEE Transactions on Multimedia, 10 (3):352-360, 2008. 2 +[38] Yanhu Shan, Shiquan Wang, Zhang Zhang, and Kaiqi Huang. An xt slice based method for action recognition. In 2011 IEEE International Conference on Computer Vision Workshops, pages 1897-1903, 2011. 2 +[39] Wei Sun, Xiongkuo Min, Wei Lu, and Guangtao Zhai. A deep learning based no-reference quality assessment model forUGC videos. In ACM International Conference on Multimedia, 2022. 8 +[40] Zhengzhong Tu, Yilin Wang, Neil Birkbeck, Balu Adsumilli, and Alan C Bovik. Ugc-vqa: Benchmarking blind video quality assessment for user generated content. IEEE Transactions on Image Processing, 30:4449-4464, 2021. 8 +[41] Zhengzhong Tu, Xiangxu Yu, Yilin Wang, Neil Birkbeck, Balu Adsumilli, and Alan C Bovik. Rapique: Rapid and accurate video quality prediction of user generated content. IEEE Open Journal of Signal Processing, 2:425-440, 2021. 8 +[42] Jiadong Wang, Xinyuan Qian, Malu Zhang, Robby T Tan, and Haizhou Li. Seeing what you said: Talking face generation guided by a lip reading expert. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14653-14662, 2023. 1, 2, 4 +[43] Suzhen Wang, Lincheng Li, Yu Ding, Changjie Fan, and Xin Yu. Audio2head: Audio-driven one-shot talking-head generation with natural head motion. arXiv preprint arXiv:2107.09293, 2021. 1, 2, 4 +[44] Zhou Wang, A.C. Bovik, H.R. Sheikh, and E.P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4): 600-612, 2004. 3 + +[45] Haoning Wu, Chaofeng Chen, Jingwen Hou, Liang Liao, Annan Wang, Wenxiu Sun, Qiong Yan, and Weisi Lin. Fastvqa: Efficient end-to-end video quality assessment with fragment sampling. In European Conference on Computer Vision, pages 538-554. Springer, 2022. 8 +[46] Jingyuan Yang, Jiawei Feng, and Hui Huang. Emogen: Emotional image content generation with text-to-image diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6358-6368, 2024. 1, 2, 4 +[47] Fei Yin, Yong Zhang, Xiaodong Cun, Mingdeng Cao, Yanbo Fan, Xuan Wang, Qingyan Bai, Baoyuan Wu, Jue Wang, and Yujiu Yang. Styleheat: One-shot high-resolution editable talking face generation via pre-trained stylegan. In European Conference on Computer Vision, pages 85-101, 2022. 1, 2, 4 +[48] Lin Zhang, Lei Zhang, and Alan C Bovik. A feature-enriched completely blind image quality evaluator. IEEE Transactions on Image Processing, 24(8):2579-2591, 2015. 8 +[49] Wenxuan Zhang, Xiaodong Cun, Xuan Wang, Yong Zhang, Xi Shen, Yu Guo, Ying Shan, and Fei Wang. Sadtalker: Learning realistic 3d motion coefficients for stylized audio-driven single image talking face animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8652-8661, 2023. 1, 2, 4 +[50] Yue Zhang, Minhao Liu, Zhaokang Chen, Bin Wu, Yubin Zeng, Chao Zhan, Yingjie He, Junxin Huang, and Wenjiang Zhou. Musetalk: Real-time high quality lip synchronization with latent space inpainting. arXiv preprint arXiv:2410.10122, 2024. 2, 4 +[51] Zhimeng Zhang, Zhipeng Hu, Wenjin Deng, Changjie Fan, Tangjie Lv, and Yu Ding. Dinet: Deformation inpainting network for realistic face visually dubbing on high resolution video. arXiv preprint arXiv:2303.03988, 2023. 1, 2, 4 +[52] Zicheng Zhang, Wei Sun, Yingjie Zhou, Wei Lu, Yucheng Zhu, Xiongkuo Min, and Guangtao Zhai. Eep-3dqa: Efficient and effective projection-based 3d model quality assessment. In 2023 IEEE international conference on Multimedia and expo (ICME), pages 2483–2488. IEEE, 2023. 7 +[53] Zicheng Zhang, Wei Sun, Yingjie Zhou, Haoning Wu, Chunyi Li, Xiongkuo Min, Xiaohong Liu, Guangtao Zhai, and Weisi Lin. Advancing zero-shot digital human quality assessment through text-prompted evaluation. arXiv preprint arXiv:2307.02808, 2023. 2, 5 +[54] Zicheng Zhang, Yingjie Zhou, Wei Sun, Wei Lu, Xiongkuo Min, Yu Wang, and Guangtao Zhai. Ddh-qa: A dynamic digital humans quality assessment database. In IEEE International Conference on Multimedia and Expo, pages 2519–2524, 2023. 2 +[55] Zicheng Zhang, Yingjie Zhou, Wei Sun, Xiongkuo Min, Yuzhe Wu, and Guangtao Zhai. Perceptual quality assessment for digital human heads. In IEEE International Conference on Acoustics, Speech and Signal Processing, pages 1-5, 2023. 2, 5 +[56] Zicheng Zhang, Yingjie Zhou, Wei Sun, Xiongkuo Min, and Guangtao Zhai. Geometry-aware video quality assessment for dynamic digital human. In IEEE International Conference on Image Processing, pages 1365-1369, 2023. 2 + +[57] Zicheng Zhang, Yingjie Zhou, Wei Sun, Xiongkuo Min, and Guangtao Zhai. Simple baselines for projection-based full-reference and no-reference point cloud quality assessment. arXiv preprint arXiv:2310.17147, 2023. 7 +[58] Zicheng Zhang, Haoning Wu, Chunyi Li, Yingjie Zhou, Wei Sun, Xiongkuo Min, Zijian Chen, Xiaohong Liu, Weisi Lin, and Guangtao Zhai. A-bench: Are lmm masters at evaluating ai-generated images? arXiv preprint arXiv:2406.03070, 2024. 3 +[59] Zicheng Zhang, Yingjie Zhou, Chunyi Li, Baixuan Zhao, Xiaohong Liu, and Guangtao Zhai. Quality assessment in the era of large models: A survey. ACM Transactions on Multimedia Computing, Communications and Applications, 2024. 5 +[60] Zicheng Zhang, Yingjie Zhou, Long Teng, Wei Sun, Chunyi Li, Xiongkuo Min, Xiao-Ping Zhang, and Guangtao Zhai. Quality-of-experience evaluation for digital twins in 6g network environments. IEEE Transactions on Broadcasting, 2024. 2 +[61] Zicheng Zhang, Junying Wang, Yijin Guo, Farong Wen, Zijian Chen, Hanqing Wang, Wenzhe Li, Lu Sun, Yingjie Zhou, Jianbo Zhang, Bowen Yan, Ziheng Jia, Jiahao Xiao, Yuan Tian, Xiangyang Zhu, Kaiwei Zhang, Chunyi Li, Xiaohong Liu, Xiongkuo Min, Qi Jia, and Guangtao Zhai. Aibench: Towards trustworthy evaluation under the $45^{\circ}$ law. https://aiben.ch/, 2025.3 +[62] Zicheng Zhang, Yingjie Zhou, Chunyi Li, Wei Sun, Xiongkuo Min, Xiaohong Liu, and Guangtao Zhai. Mmpcqa+: Advancing multi-modal learning for point cloud quality assessment. ACM Transactions on Multimedia Computing, Communications and Applications, 21(4):1-22, 2025. 5 +[63] Weizhi Zhong, Chaowei Fang, Yinqi Cai, Pengxu Wei, Gangming Zhao, Liang Lin, and Guanbin Li. Identity-preserving talking face generation with landmark and appearance priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9729-9738, 2023. 1, 2, 4 +[64] Yang Zhou, Xintong Han, Eli Shechtman, Jose Echevarria, Evangelos Kalogerakis, and Dingzeyu Li. Makelttalk: speaker-aware talking-head animation. ACM Transactions On Graphics, 39(6):1-15, 2020. 1, 4 +[65] Yu Zhou, Yanjing Sun, Leida Li, Ke Gu, and Yuming Fang. Omnidirectional image quality assessment by distortion discrimination assisted multi-stream network. IEEE Transactions on Circuits and Systems for Video Technology, 32(4): 1767-1777, 2021. 5 +[66] Yu Zhou, Weikang Gong, Yanjing Sun, Leida Li, Jinjian Wu, and Xinbo Gao. Pyramid feature aggregation for hierarchical quality prediction of stitched panoramic images. IEEE Transactions on Multimedia, 25:4177-4186, 2022. 5 +[67] Yingjie Zhou, Yaodong Chen, Kaiyue Bi, Lian Xiong, and Hui Liu. An implementation of multimodal fusion system for intelligent digital human generation. arXiv preprint arXiv:2310.20251, 2023. 1 +[68] Yu Zhou, Weikang Gong, Yanjing Sun, Leida Li, Ke Gu, and Jinjian Wu. Quality assessment for stitched panoramic + +images via patch registration and bidimensional feature aggregation. IEEE Transactions on Multimedia, 2023. 5 +[69] Yingjie Zhou, Zicheng Zhang, Wei Sun, Xiongkuo Min, Xianghe Ma, and Guangtao Zhai. A no-reference quality assessment method for digital human head. In IEEE International Conference on Image Processing, pages 36-40, 2023. 2 +[70] Yingjie Zhou, Zicheng Zhang, Wei Sun, Xiongkuo Min, and Guangtao Zhai. Perceptual quality assessment for point clouds: A survey. ZTE Communications, 21(4):3, 2023. 5 +[71] Yingjie Zhou, Zicheng Zhang, Jiezhang Cao, Jun Jia, Yanwei Jiang, Farong Wen, Xiaohong Liu, Xiongkuo Min, and Guangtao Zhai. Memo-bench: A multiple benchmark for text-to-image and multimodal large language models on human emotion analysis. arXiv preprint arXiv:2411.11235, 2024. 2 +[72] Yingjie Zhou, Zicheng Zhang, Wei Sun, Xiaohong Liu, Xiongkuo Min, Zhihua Wang, Xiao-Ping Zhang, and Guangtao Zhai. Thqa: A perceptual quality assessment database for talking heads. In IEEE International Conference on Image Processing, pages 15-21, 2024. 2, 3, 5, 7 +[73] Yingjie Zhou, Zicheng Zhang, Wei Sun, Xiaohong Liu, Xiongkuo Min, and Guangtao Zhai. Subjective and objective quality-of-experience assessment for 3d talking heads. In ACM International Conference on Multimedia, pages 6033-6042, 2024. 2, 7 +[74] Yingjie Zhou, Zicheng Zhang, Farong Wen, Jun Jia, Yanwei Jiang, Xiaohong Liu, Xiongkuo Min, and Guangtao Zhai. 3dgcqa: A quality assessment database for 3d ai-generated contents. arXiv preprint arXiv:2409.07236, 2024.5 +[75] Yingjie Zhou, Zicheng Zhang, Farong Wen, Jun Jia, Xiongkuo Min, Jia Wang, and Guangtao Zhai. Reli-qa: A multidimensional quality assessment dataset for relighted human heads. In IEEE Visual Communications and Image Processing, 2024. 2 +[76] Yingjie Zhou, Jiezhang Cao, Zicheng Zhang, Farong Wen, Yanwei Jiang, Jun Jia, Xiaohong Liu, Xiongkuo Min, and Guangtao Zhai. Q-agent: Quality-driven chain-of-thought image restoration agent through robust multimodal large language model. arXiv preprint arXiv:2504.07148, 2025. 5 +[77] Yingjie Zhou, Jing Wan, Sitong Liu, Yinghan Xia, Zhixiang Lu, Farong Wen, Zicheng Zhang, Yu Wang, Yu Zhou, Xiaohong Liu, Xiongkuo Min, Jiezhang Cao, and Guangtao Zhai. Cdhqa: A quality assessment database for conversational digital human. In International Conference on Image and Graphics, 2025. 2 +[78] Yingjie Zhou, Zicheng Zhang, Jun Jia, Yanwei Jiang, Xiaohong Liu, Xiongkuo Min, and Guangtao Zhai. Who is a better imitator: Subjective and objective quality assessment of animated humans. IEEE Transactions on Circuits and Systems for Video Technology, 2025. 2 +[79] Yingjie Zhou, Zicheng Zhang, Wei Sun, Xiongkuo Min, and Guangtao Zhai. Ct-pcqa: A convolutional neural network and transformer combined method for point cloud quality assessment. Signal Processing: Image Communication, page 117371, 2025. 7 \ No newline at end of file diff --git a/ICCV/2025/Who is a Better Talker_ Subjective and Objective Quality Assessment for AI-Generated Talking Heads/images.zip b/ICCV/2025/Who is a Better Talker_ Subjective and Objective Quality Assessment for AI-Generated Talking Heads/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..4b9bd17cddff936981ac681a0c898562e95b1242 --- /dev/null +++ b/ICCV/2025/Who is a Better Talker_ Subjective and Objective Quality Assessment for AI-Generated Talking Heads/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:98053773adc3192ca5571b312a20889533e95c1c30eadebc36faf00bbb5d672a +size 961659 diff --git a/ICCV/2025/Who is a Better Talker_ Subjective and Objective Quality Assessment for AI-Generated Talking Heads/layout.json b/ICCV/2025/Who is a Better Talker_ Subjective and Objective Quality Assessment for AI-Generated Talking Heads/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..84679d98c3683960914a0887db6fb9cbcc3dab9b --- /dev/null +++ b/ICCV/2025/Who is a Better Talker_ Subjective and Objective Quality Assessment for AI-Generated Talking Heads/layout.json @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:63ed38e65f18dceab172fa400d2fc1a80711dba1cd56fe4dc78eeb5677ccc326 +size 411858