Add Batch 2d36f5e3-0045-4304-bae6-48f69c617a46

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

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

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

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.
+
+<table><tr><td rowspan="2">Model</td><td rowspan="2">Downsample Factor</td><td rowspan="2">#Channels</td><td colspan="4">WebVid Test Set [2]</td><td colspan="4">Inter4K Test Set [22]</td><td colspan="4">Large Motion Test Set</td></tr><tr><td>PSNR (↑)</td><td>SSIM (↑)</td><td>LPIPS (↓)</td><td>FVD (↓)</td><td>PSNR (↑)</td><td>SSIM (↑)</td><td>LPIPS (↓)</td><td>FVD (↓)</td><td>PSNR (↑)</td><td>SSIM (↑)</td><td>LPIPS (↓)</td><td>FVD (↓)</td></tr><tr><td>Open-Sora-Plan (OD VAE [9])</td><td>4x8x8</td><td>4</td><td>29.1646</td><td>0.8334</td><td>0.0789</td><td>40.0616</td><td>28.6690</td><td>0.8381</td><td>0.0906</td><td>38.3960</td><td>27.5697</td><td>0.8045</td><td>0.1065</td><td>219.4829</td></tr><tr><td>IV-VAE [28]</td><td>4x8x8</td><td>4</td><td>28.7000</td><td>0.8253</td><td>0.0832</td><td>38.4457</td><td>28.7599</td><td>0.8398</td><td>0.0906</td><td>28.6293</td><td>27.2838</td><td>0.8014</td><td>0.1053</td><td>197.0937</td></tr><tr><td>Open-Sora (OPS VAE [36])</td><td>4x8x8</td><td>4</td><td>29.3753</td><td>0.8284</td><td>0.1240</td><td>90.8863</td><td>29.2721</td><td>0.8431</td><td>0.1316</td><td>55.1470</td><td>27.7586</td><td>0.8032</td><td>0.1540</td><td>355.5645</td></tr><tr><td>CV-VAE [35]</td><td>4x8x8</td><td>4</td><td>28.6795</td><td>0.8154</td><td>0.1072</td><td>70.7160</td><td>27.7437</td><td>0.8124</td><td>0.1284</td><td>50.9006</td><td>26.9456</td><td>0.7849</td><td>0.1411</td><td>293.7747</td></tr><tr><td>VidTok [24]</td><td>4x8x8</td><td>4</td><td>30.2517</td><td>0.8679</td><td>0.0645</td><td>33.9966</td><td>29.2400</td><td>0.8520</td><td>0.0865</td><td>33.6913</td><td>27.2922</td><td>0.8073</td><td>0.1044</td><td>193.3726</td></tr><tr><td>Video VAE w/o Joint Training (Ours)</td><td>4x8x8</td><td>4</td><td>30.2091</td><td>0.8656</td><td>0.0566</td><td>24.2982</td><td>28.9048</td><td>0.8543</td><td>0.0688</td><td>18.1236</td><td>27.3917</td><td>0.8078</td><td>0.0867</td><td>141.1166</td></tr><tr><td>Video VAE (Ours)</td><td>4x8x8</td><td>4</td><td>30.3140</td><td>0.8676</td><td>0.0538</td><td>22.4391</td><td>28.9227</td><td>0.8565</td><td>0.0665</td><td>17.6865</td><td>27.6236</td><td>0.8136</td><td>0.0841</td><td>138.6341</td></tr><tr><td>Cross-Modal VAE (Ours)</td><td>4x8x8</td><td>4</td><td>30.3486</td><td>0.8679</td><td>0.0530</td><td>18.1058</td><td>29.0295</td><td>0.8585</td><td>0.0649</td><td>15.0345</td><td>27.5009</td><td>0.8092</td><td>0.0820</td><td>122.4046</td></tr><tr><td>Cosmos-Tokenizer [20]</td><td>4x8x8</td><td>16</td><td>31.2545</td><td>0.8861</td><td>0.1030</td><td>24.7131</td><td>31.2002</td><td>0.8957</td><td>0.1071</td><td>19.3781</td><td>30.1619</td><td>0.8675</td><td>0.1194</td><td>109.3825</td></tr><tr><td>CogVideoX-VAE [31]</td><td>4x8x8</td><td>16</td><td>32.8940</td><td>0.9208</td><td>0.0504</td><td>14.0559</td><td>32.5122</td><td>0.9229</td><td>0.0532</td><td>10.8545</td><td>31.0906</td><td>0.8978</td><td>0.0685</td><td>68.6781</td></tr><tr><td>EasyAnimate-VAE [30]</td><td>4x8x8</td><td>16</td><td>32.1233</td><td>0.9085</td><td>0.0405</td><td>11.9533</td><td>31.5066</td><td>0.9048</td><td>0.0572</td><td>9.7797</td><td>30.5213</td><td>0.8846</td><td>0.0598</td><td>66.0293</td></tr><tr><td>CV-VAE [35]</td><td>4x8x8</td><td>16</td><td>32.2766</td><td>0.9080</td><td>0.0546</td><td>16.8478</td><td>31.6129</td><td>0.9060</td><td>0.0642</td><td>15.1846</td><td>30.7136</td><td>0.8868</td><td>0.0726</td><td>91.4076</td></tr><tr><td>Open-Sora-Plan (WF-VAE-L [18])</td><td>4x8x8</td><td>16</td><td>33.1031</td><td>0.9214</td><td>0.0400</td><td>9.3192</td><td>32.5211</td><td>0.9153</td><td>0.0505</td><td>7.9032</td><td>31.1565</td><td>0.8958</td><td>0.0602</td><td>62.4018</td></tr><tr><td>IV-VAE [28]</td><td>4x8x8</td><td>16</td><td>33.1887</td><td>0.9227</td><td>0.0381</td><td>10.2930</td><td>33.2088</td><td>0.9270</td><td>0.0408</td><td>8.4002</td><td>31.9277</td><td>0.9086</td><td>0.0493</td><td>59.5411</td></tr><tr><td>HunyuanVAE [27]</td><td>4x8x8</td><td>16</td><td>33.4381</td><td>0.9271</td><td>0.0323</td><td>8.7744</td><td>33.2339</td><td>0.9272</td><td>0.0372</td><td>9.2010</td><td>32.1441</td><td>0.9118</td><td>0.0445</td><td>51.4671</td></tr><tr><td>VidTok [24]</td><td>4x8x8</td><td>16</td><td>34.3469</td><td>0.9364</td><td>0.0337</td><td>9.6649</td><td>33.4085</td><td>0.9305</td><td>0.0427</td><td>8.2432</td><td>31.9332</td><td>0.9080</td><td>0.0534</td><td>52.9496</td></tr><tr><td>Video VAE w/o Joint Training (Ours)</td><td>4x8x8</td><td>16</td><td>33.8844</td><td>0.9334</td><td>0.0344</td><td>8.5503</td><td>32.9416</td><td>0.9297</td><td>0.0409</td><td>8.6246</td><td>31.8471</td><td>0.9073</td><td>0.0499</td><td>46.6997</td></tr><tr><td>Video VAE w/ causal conv (Ours)</td><td>4x8x8</td><td>16</td><td>33.3031</td><td>0.9233</td><td>0.0382</td><td>9.7911</td><td>33.0645</td><td>0.9266</td><td>0.0415</td><td>9.5623</td><td>32.0313</td><td>0.9087</td><td>0.0494</td><td>50.4350</td></tr><tr><td>Video VAE (Ours)</td><td>4x8x8</td><td>16</td><td>34.1558</td><td>0.9362</td><td>0.0271</td><td>7.7195</td><td>33.3184</td><td>0.9328</td><td>0.0316</td><td>6.3432</td><td>32.1503</td><td>0.9122</td><td>0.0409</td><td>41.8797</td></tr><tr><td>Cross-Modal VAE (Ours)</td><td>4x8x8</td><td>16</td><td>34.4430</td><td>0.9403</td><td>0.0271</td><td>6.3560</td><td>33.7760</td><td>0.9371</td><td>0.0315</td><td>6.8278</td><td>32.2908</td><td>0.9174</td><td>0.0408</td><td>43.4941</td></tr><tr><td>Video VAE (Ours)</td><td>8x8x8</td><td>4</td><td>27.8445</td><td>0.8080</td><td>0.0760</td><td>35.2760</td><td>26.8921</td><td>0.8033</td><td>0.0953</td><td>34.4657</td><td>24.9830</td><td>0.7438</td><td>0.1214</td><td>217.8798</td></tr><tr><td>Cross-Modal VAE (Ours)</td><td>8x8x8</td><td>4</td><td>28.5882</td><td>0.8260</td><td>0.0712</td><td>34.4697</td><td>27.3075</td><td>0.8188</td><td>0.0883</td><td>28.9388</td><td>25.7957</td><td>0.7649</td><td>0.1112</td><td>202.7046</td></tr></table>
+
+Table 1. Quantitative comparison with state-of-the-art methods.  
+
+<table><tr><td>Model</td><td>#ch</td><td>Temporal Smoothness ↑</td><td>Overall Quality ↑</td></tr><tr><td>Open-Sora-Plan</td><td>4</td><td>3.84</td><td>3.64</td></tr><tr><td>Open-Sora</td><td>4</td><td>3.16</td><td>2.74</td></tr><tr><td>CV-VAE</td><td>4</td><td>2.84</td><td>2.46</td></tr><tr><td>Our VAE</td><td>4</td><td>4.6</td><td>4.54</td></tr><tr><td>Cosmos-Tokenizer</td><td>16</td><td>3.18</td><td>2.54</td></tr><tr><td>CogVideoX-VAE</td><td>16</td><td>3.86</td><td>3.48</td></tr><tr><td>EasyAnimate-VAE</td><td>16</td><td>4.06</td><td>4.02</td></tr><tr><td>CV-VAE</td><td>16</td><td>3.36</td><td>3.00</td></tr><tr><td>Our VAE</td><td>16</td><td>4.5</td><td>4.4</td></tr></table>
+
+Table 2. User study based on MOS scores for comparisons with baselines.  
+
+<table><tr><td rowspan="2">Model</td><td rowspan="2">Factor</td><td colspan="2">UCF101</td><td colspan="2">WebVid</td></tr><tr><td>FVD↓</td><td>KVD↓</td><td>FVD↓</td><td>KVD↓</td></tr><tr><td>Vc2</td><td>1x8x8</td><td>498.70</td><td>23.38</td><td>570.48</td><td>48.45</td></tr><tr><td>Vc2 + ours</td><td>4x8x8</td><td>479.69</td><td>20.18</td><td>529.75</td><td>37.39</td></tr></table>
+
+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
+
+<table><tr><td>Model</td><td>PSNR (↑)</td><td>SSIM (↑)</td><td>LPIPS (↓)</td></tr><tr><td>Simultaneous</td><td>24.0593</td><td>0.7315</td><td>0.1293</td></tr><tr><td>Sequential</td><td>23.3681</td><td>0.6917</td><td>0.1481</td></tr><tr><td>Ours</td><td>24.6722</td><td>0.7234</td><td>0.1162</td></tr></table>
+
+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

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

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

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

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

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

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
+
+<sup>4</sup> 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].  
+
+<table><tr><td rowspan="2">Method</td><td colspan="5">Nr3D</td><td colspan="5">Sr3D</td></tr><tr><td>Overall</td><td>Easy</td><td>Hard</td><td>View Dep.</td><td>View Indep.</td><td>Overall</td><td>Easy</td><td>Hard</td><td>View Dep.</td><td>View Indep.</td></tr><tr><td>3DVG-Transformer [54]</td><td>40.8</td><td>48.5</td><td>34.8</td><td>34.8</td><td>43.7</td><td>51.4</td><td>54.2</td><td>44.9</td><td>44.6</td><td>51.7</td></tr><tr><td>LanguageRefer [36]</td><td>43.9</td><td>51.0</td><td>36.6</td><td>41.7</td><td>45.0</td><td>56.0</td><td>58.9</td><td>49.3</td><td>49.2</td><td>56.3</td></tr><tr><td>TransRefer3D [16]</td><td>42.1</td><td>48.5</td><td>36.0</td><td>36.5</td><td>44.9</td><td>57.4</td><td>60.5</td><td>50.2</td><td>49.9</td><td>57.7</td></tr><tr><td>SAT [48]</td><td>49.2</td><td>56.3</td><td>42.4</td><td>46.9</td><td>50.4</td><td>57.9</td><td>61.2</td><td>50.0</td><td>49.2</td><td>58.3</td></tr><tr><td>MVT [18]</td><td>55.1</td><td>61.3</td><td>49.1</td><td>54.3</td><td>55.4</td><td>64.5</td><td>66.9</td><td>58.8</td><td>58.4</td><td>64.7</td></tr><tr><td>ViewRefer [15]</td><td>56.0</td><td>63.0</td><td>49.7</td><td>55.1</td><td>56.8</td><td>67.0</td><td>68.9</td><td>62.1</td><td>52.2</td><td>67.7</td></tr><tr><td>MiKASA [5]</td><td>64.4</td><td>69.7</td><td>59.4</td><td>65.4</td><td>64.0</td><td>75.2</td><td>78.6</td><td>67.3</td><td>70.4</td><td>75.4</td></tr><tr><td>CoT3DRef [2]</td><td>64.4</td><td>70.0</td><td>59.2</td><td>61.9</td><td>65.7</td><td>73.2</td><td>75.2</td><td>67.9</td><td>67.6</td><td>73.5</td></tr><tr><td>ViewSRD (ours)</td><td>69.9</td><td>75.3</td><td>64.8</td><td>68.6</td><td>70.6</td><td>76.0</td><td>78.3</td><td>70.6</td><td>69.0</td><td>76.2</td></tr></table>
+
+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.  
+
+<table><tr><td>Model</td><td>Multi-Anc</td><td>Single-Anc</td><td>Overall</td></tr><tr><td>MVT [18]</td><td>52.6</td><td>56.6</td><td>55.1</td></tr><tr><td>CoT3DRef [2]</td><td>63.1</td><td>65.2</td><td>64.4</td></tr><tr><td>ViewSRD</td><td>71.5</td><td>69.5</td><td>69.9</td></tr></table>
+
+Table 3. Performance  $(\%)$  of SRD module improves MVT [18], BUTD-DETR [21] and EDA [42] on ScanRefer [6] dataset.  
+
+<table><tr><td rowspan="2">Method</td><td colspan="2">Unique (19%)</td><td colspan="2">Multiple (81%)</td><td colspan="2">Overall</td></tr><tr><td>0.25</td><td>0.5</td><td>0.25</td><td>0.5</td><td>0.25</td><td>0.5</td></tr><tr><td>MVT [18]</td><td>77.7</td><td>66.5</td><td>31.9</td><td>25.3</td><td>40.8</td><td>33.3</td></tr><tr><td>MVT+SRD</td><td>78.6</td><td>67.2</td><td>34.1</td><td>27.1</td><td>42.1 (3.2%↑)</td><td>34.3(3.0%↑)</td></tr><tr><td>BUTD-DETR [21]</td><td>82.8</td><td>64.9</td><td>44.7</td><td>33.9</td><td>50.4</td><td>38.6</td></tr><tr><td>BUTD-DETR+SRD</td><td>85.0</td><td>66.2</td><td>45.3</td><td>34.2</td><td>57.9 (14.9%↑)</td><td>45.7 (18.4%↑)</td></tr><tr><td>EDA [42]</td><td>80.4</td><td>65.3</td><td>35.6</td><td>25.1</td><td>43.6</td><td>32.3</td></tr><tr><td>EDA+SRD</td><td>81.0</td><td>67.3</td><td>36.4</td><td>28.3</td><td>44.4 (1.8%↑)</td><td>35.3 (9.3%↑)</td></tr><tr><td>ViewSRD</td><td>82.1</td><td>68.2</td><td>37.4</td><td>29.0</td><td>45.4</td><td>36.0</td></tr></table>
+
+# 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(%)  
+
+<table><tr><td>Component</td><td>Overall</td><td>Easy</td><td>Hard</td><td>View Dep.</td><td>View Indep.</td></tr><tr><td>w/o CCVTs.</td><td>62.2</td><td>68.5</td><td>56.1</td><td>60.1</td><td>63.2</td></tr><tr><td>w/o Textual M.</td><td>68.0</td><td>73.5</td><td>62.6</td><td>67.6</td><td>68.1</td></tr><tr><td>w/o Scene M.</td><td>64.6</td><td>70.5</td><td>58.9</td><td>63.8</td><td>64.9</td></tr><tr><td>w/o SRD M.</td><td>68.6</td><td>73.0</td><td>64.8</td><td>66.5</td><td>70.0</td></tr><tr><td>w/o Weight.</td><td>69.0</td><td>74.2</td><td>64.0</td><td>66.5</td><td>70.2</td></tr><tr><td>LLM-Aug.</td><td>69.1</td><td>74.5</td><td>63.7</td><td>68.0</td><td>69.5</td></tr><tr><td>ViewSRD</td><td>69.9</td><td>75.3</td><td>64.8</td><td>68.6</td><td>70.6</td></tr></table>
+
+Table 5. Ablation of view numbers on Nr3D [1].  
+
+<table><tr><td colspan="2">View Number</td><td rowspan="2">Overall</td><td rowspan="2">Easy</td><td rowspan="2">Hard</td><td rowspan="2">View Dep.</td><td rowspan="2">View Indep.</td></tr><tr><td>Train</td><td>Test</td></tr><tr><td>4</td><td>1</td><td>66.0</td><td>71.7</td><td>60.5</td><td>64.0</td><td>67.0</td></tr><tr><td>4</td><td>2</td><td>68.9</td><td>75.1</td><td>63.0</td><td>66.9</td><td>69.9</td></tr><tr><td>4</td><td>4</td><td>69.9</td><td>75.3</td><td>64.8</td><td>68.6</td><td>70.6</td></tr><tr><td>1</td><td>1</td><td>64.4</td><td>70.9</td><td>58.1</td><td>60.8</td><td>66.2</td></tr><tr><td>2</td><td>2</td><td>67.7</td><td>73.0</td><td>62.5</td><td>66.1</td><td>68.4</td></tr><tr><td>8</td><td>8</td><td>68.4</td><td>74.1</td><td>63.0</td><td>67.4</td><td>68.9</td></tr></table>
+
+# 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].  
+
+<table><tr><td>LLM decoupler</td><td>Accuracy</td></tr><tr><td>OpenChat [40]</td><td>69.6%</td></tr><tr><td>DeepSeek-R1 [28]</td><td>69.9%</td></tr><tr><td>Qwen-Plus [46]</td><td>70.5%</td></tr><tr><td>Qwen-Turbo [46]</td><td>70.7%</td></tr></table>
+
+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

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

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

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

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

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

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
+
+<table><tr><td>Method</td><td>Date</td><td>IoU↑ mIoU↑</td><td>road</td><td>sidewalk</td><td>parking</td><td>other-gmd.</td><td>building</td><td>car</td><td>truck</td><td>bicycle</td><td>motorcycle</td><td>other-veh.</td><td>vegetation</td><td>trunk</td><td>terrain</td><td>person</td><td>bicyclist</td><td>motorcyclist</td><td>fence</td><td>pole</td><td>traf-sign</td><td></td></tr><tr><td colspan="21">Stereo camera-based methods</td><td></td><td></td></tr><tr><td>StereoScene[14]</td><td>IJCAI2024</td><td>43.34</td><td>15.36</td><td>61.90</td><td>31.20</td><td>30.70</td><td>10.70</td><td>24.20</td><td>22.80</td><td>2.80</td><td>3.40</td><td>2.40</td><td>6.10</td><td>23.80</td><td>8.40</td><td>27.00</td><td>2.90</td><td>2.20</td><td>0.50</td><td>16.50</td><td>7.00</td><td>7.20</td></tr><tr><td colspan="21">Monocular temporal methods</td><td></td><td></td></tr><tr><td>VoxFormer-T[20]</td><td>CVPR2023</td><td>43.21</td><td>13.41</td><td>54.10</td><td>26.90</td><td>25.10</td><td>7.30</td><td>23.50</td><td>21.70</td><td>3.60</td><td>1.90</td><td>1.60</td><td>4.10</td><td>24.40</td><td>8.10</td><td>24.20</td><td>1.60</td><td>1.10</td><td>0.00</td><td>13.10</td><td>6.60</td><td>5.70</td></tr><tr><td>HASSC-T[36]</td><td>CVPR2024</td><td>42.87</td><td>14.38</td><td>55.30</td><td>29.60</td><td>25.90</td><td>11.30</td><td>23.10</td><td>23.00</td><td>2.90</td><td>1.90</td><td>1.50</td><td>4.90</td><td>24.80</td><td>9.80</td><td>26.50</td><td>1.40</td><td>3.00</td><td>0.00</td><td>14.30</td><td>7.00</td><td>7.10</td></tr><tr><td>HTCL[15]</td><td>ECCV2024</td><td>44.23</td><td>17.09</td><td>64.40</td><td>34.80</td><td>33.80</td><td>12.40</td><td>25.90</td><td>27.30</td><td>5.70</td><td>1.80</td><td>2.20</td><td>5.40</td><td>25.30</td><td>10.80</td><td>31.20</td><td>1.10</td><td>3.10</td><td>0.90</td><td>21.10</td><td>9.00</td><td>8.30</td></tr><tr><td>H2GFormer-T[37]</td><td>AAAI2024</td><td>43.52</td><td>14.60</td><td>57.90</td><td>30.40</td><td>30.00</td><td>6.90</td><td>24.00</td><td>23.70</td><td>5.20</td><td>0.60</td><td>1.20</td><td>5.00</td><td>25.20</td><td>10.70</td><td>25.80</td><td>1.10</td><td>0.10</td><td>0.00</td><td>14.60</td><td>7.50</td><td>9.30</td></tr><tr><td colspan="21">Monocular single-frame methods</td><td></td><td></td></tr><tr><td>MonoScene[3]</td><td>CVPR2023</td><td>34.16</td><td>11.08</td><td>54.70</td><td>27.10</td><td>24.80</td><td>5.70</td><td>14.40</td><td>18.80</td><td>3.30</td><td>0.50</td><td>0.70</td><td>4.40</td><td>14.90</td><td>2.40</td><td>19.50</td><td>1.00</td><td>1.40</td><td>0.40</td><td>11.10</td><td>3.30</td><td>2.10</td></tr><tr><td>VoxFormer-S[20]</td><td>CVPR2023</td><td>42.95</td><td>12.20</td><td>53.90</td><td>25.30</td><td>21.10</td><td>5.60</td><td>19.80</td><td>20.80</td><td>3.50</td><td>2.60</td><td>0.70</td><td>3.70</td><td>22.40</td><td>7.50</td><td>21.30</td><td>1.40</td><td>2.60</td><td>0.20</td><td>11.10</td><td>5.10</td><td>4.90</td></tr><tr><td>TPVFormer[11]</td><td>CVPR2023</td><td>34.25</td><td>11.26</td><td>55.10</td><td>27.20</td><td>27.40</td><td>6.50</td><td>14.80</td><td>19.20</td><td>3.70</td><td>1.00</td><td>0.50</td><td>2.30</td><td>13.90</td><td>2.60</td><td>20.40</td><td>1.10</td><td>2.40</td><td>0.30</td><td>11.00</td><td>2.90</td><td>1.50</td></tr><tr><td>SurroundOcc[38]</td><td>ICCV2023</td><td>34.72</td><td>11.86</td><td>56.90</td><td>28.30</td><td>30.20</td><td>6.80</td><td>15.20</td><td>20.60</td><td>1.40</td><td>1.60</td><td>1.20</td><td>4.40</td><td>14.90</td><td>3.40</td><td>19.30</td><td>1.40</td><td>2.00</td><td>0.10</td><td>11.30</td><td>3.90</td><td>2.40</td></tr><tr><td>OccFormer[46]</td><td>ICCV2023</td><td>34.53</td><td>12.32</td><td>55.90</td><td>30.30</td><td>31.50</td><td>6.50</td><td>15.70</td><td>21.60</td><td>1.20</td><td>1.50</td><td>1.70</td><td>3.20</td><td>16.80</td><td>3.90</td><td>21.30</td><td>2.20</td><td>1.10</td><td>0.20</td><td>11.90</td><td>3.80</td><td>3.70</td></tr><tr><td>IAMSSC[39]</td><td>T-ITS2024</td><td>43.74</td><td>12.37</td><td>54.00</td><td>25.50</td><td>24.70</td><td>6.90</td><td>19.20</td><td>21.30</td><td>3.80</td><td>1.10</td><td>0.60</td><td>3.90</td><td>22.70</td><td>5.80</td><td>19.40</td><td>1.50</td><td>2.90</td><td>0.50</td><td>11.90</td><td>5.30</td><td>4.10</td></tr><tr><td>DepthSSC[41]</td><td>arXiV2024</td><td>44.58</td><td>13.11</td><td>55.64</td><td>27.25</td><td>25.72</td><td>5.78</td><td>20.46</td><td>21.94</td><td>3.74</td><td>1.35</td><td>0.98</td><td>4.17</td><td>23.37</td><td>7.64</td><td>21.56</td><td>1.34</td><td>2.79</td><td>0.28</td><td>12.94</td><td>5.87</td><td>6.23</td></tr><tr><td>HASSC-S[36]</td><td>CVPR2024</td><td>43.40</td><td>13.34</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td>-</td><td></td></tr><tr><td>Symphonize[12]</td><td>CVPR2024</td><td>42.19</td><td>15.04</td><td>58.40</td><td>29.30</td><td>26.90</td><td>11.70</td><td>24.70</td><td>23.60</td><td>3.20</td><td>3.60</td><td>2.60</td><td>5.60</td><td>24.20</td><td>10.00</td><td>23.10</td><td>3.20</td><td>1.90</td><td>2.00</td><td>16.10</td><td>7.70</td><td>8.00</td></tr><tr><td>H2GFormer-S[37]</td><td>AAAI2024</td><td>44.20</td><td>13.72</td><td>56.40</td><td>28.60</td><td>26.50</td><td>4.90</td><td>22.80</td><td>23.40</td><td>4.80</td><td>0.80</td><td>0.90</td><td>4.10</td><td>24.60</td><td>9.10</td><td>23.80</td><td>1.20</td><td>2.50</td><td>0.10</td><td>13.30</td><td>6.40</td><td>6.30</td></tr><tr><td>MonoOcc-L[47]</td><td>ICRA2024</td><td>-</td><td>15.63</td><td>59.10</td><td>30.90</td><td>27.10</td><td>9.80</td><td>22.90</td><td>23.90</td><td>7.20</td><td>4.50</td><td>2.40</td><td>7.70</td><td>25.00</td><td>9.80</td><td>26.10</td><td>2.80</td><td>4.70</td><td>0.60</td><td>16.90</td><td>7.30</td><td>8.40</td></tr><tr><td>CGFormer[43]</td><td>NIPS2024</td><td>44.41</td><td>16.63</td><td>64.30</td><td>34.20</td><td>34.10</td><td>12.10</td><td>25.80</td><td>26.10</td><td>4.30</td><td>3.70</td><td>1.30</td><td>2.70</td><td>24.50</td><td>11.20</td><td>29.30</td><td>1.70</td><td>3.60</td><td>0.40</td><td>18.70</td><td>8.70</td><td>9.30</td></tr><tr><td>Ours</td><td>ICCV2025</td><td>46.50</td><td>17.46</td><td>64.60</td><td>34.10</td><td>32.00</td><td>12.50</td><td>26.90</td><td>26.70</td><td>7.50</td><td>2.90</td><td>3.30</td><td>6.20</td><td>27.30</td><td>12.50</td><td>28.00</td><td>2.30</td><td>5.10</td><td>1.90</td><td>19.50</td><td>9.20</td><td>9.20</td></tr></table>
+
+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
+
+<table><tr><td>Method</td><td>Date</td><td>IoU↑ mIoU↑</td><td>car</td><td>bicycle</td><td>motorcycle</td><td>truck</td><td>other-veh.</td><td>person</td><td>road</td><td>parking</td><td>sidewalk</td><td>other-grd.</td><td>building</td><td>fence</td><td>vegetation</td><td>terrain</td><td>pole</td><td>traf.-sign</td><td>other-struct.</td><td>other-object.</td><td></td></tr><tr><td colspan="21">LiDAR-based methods</td><td></td></tr><tr><td>SSCNet [34]</td><td>CVPR2017</td><td>53.58</td><td>16.95</td><td>31.95</td><td>0.00</td><td>0.17</td><td>10.29</td><td>0.00</td><td>0.07</td><td>65.70</td><td>17.33</td><td>41.24</td><td>3.22</td><td>44.41</td><td>6.77</td><td>43.72</td><td>28.87</td><td>0.78</td><td>0.75</td><td>8.69</td><td>0.67</td></tr><tr><td>LMSCNet [32]</td><td>3DV 2020</td><td>47.35</td><td>13.65</td><td>20.91</td><td>0.00</td><td>0.00</td><td>0.26</td><td>0.58</td><td>0.00</td><td>62.95</td><td>13.51</td><td>33.51</td><td>0.20</td><td>43.67</td><td>0.33</td><td>40.01</td><td>26.80</td><td>0.00</td><td>0.00</td><td>3.63</td><td>0.00</td></tr><tr><td colspan="21">Monocular camera-based methods</td><td></td></tr><tr><td>MonoScene [3]</td><td>CVPR2023</td><td>37.87</td><td>12.31</td><td>19.34</td><td>0.43</td><td>0.58</td><td>8.02</td><td>2.03</td><td>0.86</td><td>48.35</td><td>11.38</td><td>28.13</td><td>3.32</td><td>32.89</td><td>3.53</td><td>26.15</td><td>16.75</td><td>6.92</td><td>5.67</td><td>4.20</td><td>3.09</td></tr><tr><td>TPVFormer [11]</td><td>CVPR2023</td><td>40.22</td><td>13.64</td><td>21.56</td><td>1.09</td><td>1.37</td><td>8.06</td><td>2.57</td><td>2.38</td><td>52.99</td><td>11.99</td><td>31.07</td><td>3.78</td><td>34.83</td><td>4.80</td><td>30.08</td><td>17.52</td><td>7.46</td><td>5.86</td><td>5.48</td><td>2.70</td></tr><tr><td>OccFormer [46]</td><td>ICCV2023</td><td>40.27</td><td>13.81</td><td>22.58</td><td>0.66</td><td>0.26</td><td>9.89</td><td>3.82</td><td>2.77</td><td>54.30</td><td>13.44</td><td>31.53</td><td>3.55</td><td>36.42</td><td>4.80</td><td>31.00</td><td>19.51</td><td>7.77</td><td>8.51</td><td>6.95</td><td>4.60</td></tr><tr><td>VoxFormer [20]</td><td>CVPR2023</td><td>38.76</td><td>11.91</td><td>17.84</td><td>1.16</td><td>0.89</td><td>4.56</td><td>2.06</td><td>1.63</td><td>47.01</td><td>9.67</td><td>27.21</td><td>2.89</td><td>31.38</td><td>4.97</td><td>28.99</td><td>14.69</td><td>6.51</td><td>6.92</td><td>3.79</td><td>2.43</td></tr><tr><td>IAMSSC [39]</td><td>T-ITS2024</td><td>41.80</td><td>12.97</td><td>18.53</td><td>2.45</td><td>1.76</td><td>5.12</td><td>3.92</td><td>3.09</td><td>47.55</td><td>10.56</td><td>28.35</td><td>4.12</td><td>31.53</td><td>6.28</td><td>29.17</td><td>15.24</td><td>8.29</td><td>7.01</td><td>6.35</td><td>4.19</td></tr><tr><td>DepthSSC [41]</td><td>arXiV2024</td><td>40.85</td><td>14.28</td><td>21.90</td><td>2.36</td><td>4.30</td><td>11.51</td><td>4.56</td><td>2.92</td><td>50.88</td><td>12.89</td><td>30.27</td><td>2.49</td><td>37.33</td><td>5.22</td><td>29.61</td><td>21.59</td><td>5.97</td><td>7.71</td><td>5.24</td><td>3.51</td></tr><tr><td>Symphonies [12]</td><td>CVPR2024</td><td>44.12</td><td>18.58</td><td>30.02</td><td>1.85</td><td>5.90</td><td>25.07</td><td>12.06</td><td>8.20</td><td>54.94</td><td>13.83</td><td>32.76</td><td>6.93</td><td>35.11</td><td>8.58</td><td>38.33</td><td>11.52</td><td>14.01</td><td>9.57</td><td>14.44</td><td>11.28</td></tr><tr><td>CGFormer[43]</td><td>NIPS2024</td><td>48.07</td><td>20.05</td><td>29.85</td><td>3.42</td><td>3.96</td><td>17.59</td><td>6.70</td><td>6.63</td><td>63.85</td><td>17.15</td><td>40.72</td><td>5.53</td><td>42.73</td><td>8.22</td><td>38.80</td><td>24.04</td><td>16.24</td><td>17.45</td><td>10.18</td><td>6.77</td></tr><tr><td>Ours</td><td>ICCV2025</td><td>49.12</td><td>20.95</td><td>30.77</td><td>1.91</td><td>6.60</td><td>17.99</td><td>8.72</td><td>8.67</td><td>64.35</td><td>18.83</td><td>41.53</td><td>4.48</td><td>43.87</td><td>9.07</td><td>39.75</td><td>24.94</td><td>16.52</td><td>20.66</td><td>10.30</td><td>7.99</td></tr></table>
+
+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.  
+
+<table><tr><td>Method</td><td>Params(M)↓</td><td>Memory(M)↓</td><td>Times(s)↓</td></tr><tr><td>Monoscene[3]</td><td>149.5</td><td>19,041</td><td>0.49</td></tr><tr><td>CGFormer[43]</td><td>122.4</td><td>19,330</td><td>0.41</td></tr><tr><td>Ours</td><td>127.8</td><td>22,597</td><td>0.34</td></tr></table>
+
+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
+
+<table><tr><td>Method</td><td>IoU↑</td><td>mIoU↑</td><td>Params (M)</td><td>Memory (M)</td></tr><tr><td>Baseline</td><td>42.11</td><td>14.56</td><td>57.2</td><td>15260</td></tr><tr><td>VisFrontierNet w/o Feature Refinement</td><td>46.44 (+4.33)</td><td>15.11 (+0.55)</td><td>74.3</td><td>17349</td></tr><tr><td>+ OcclusionMAE w/o Denoising</td><td>46.38 (-0.06)</td><td>15.99 (+0.88)</td><td>86.7</td><td>18746</td></tr><tr><td>+ Feature Refinement</td><td>45.88 (-0.50)</td><td>16.59 (+0.60)</td><td>125.5</td><td>21246</td></tr><tr><td>+ Denoising</td><td>46.14 (+0.26)</td><td>17.06 (+0.47)</td><td>127.8</td><td>22597</td></tr></table>
+
+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.  
+
+<table><tr><td>Invisibility</td><td>Threshold θd(m)</td><td>IoU↑</td><td>mIoU↑</td></tr><tr><td>OOV</td><td>-</td><td>43.98</td><td>15.89</td></tr><tr><td>OOV + Occ.</td><td>1.5</td><td>45.87</td><td>16.65</td></tr><tr><td>OOV + Occ.</td><td>2.5</td><td>45.83</td><td>16.92</td></tr><tr><td>OOV + Occ.</td><td>3.5</td><td>46.14</td><td>17.06</td></tr><tr><td>OOV + Occ.</td><td>4.5</td><td>45.94</td><td>16.95</td></tr></table>
+
+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.  
+
+<table><tr><td>Rh(Voxel)</td><td>Rd(Voxel)</td><td>IoU↑</td><td>mIoU↑</td></tr><tr><td>0</td><td>0</td><td>46.03</td><td>16.32</td></tr><tr><td>0</td><td>2</td><td>45.92</td><td>16.79</td></tr><tr><td>0</td><td>3</td><td>45.83</td><td>16.92</td></tr><tr><td>0</td><td>4</td><td>45.73</td><td>16.66</td></tr><tr><td>1</td><td>3</td><td>46.19</td><td>16.52</td></tr><tr><td>1</td><td>4</td><td>46.12</td><td>16.69</td></tr></table>
+
+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

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

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

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

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

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

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 $^{1,2,4*}$
+
+<sup>1</sup>College of Computer Science and Software Engineering, Shenzhen University
+
+$^{2}$ Guangdong Laboratory of Artificial Intelligence and Digital Economy (SZ)
+
+<sup>3</sup>Shenzhen 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.  
+
+<table><tr><td>VisNumBench</td><td>Angle</td><td>Length</td><td>Scale</td><td>Quantity</td><td>Depth</td><td>Area</td><td>Volume</td><td>Total</td></tr><tr><td>VisNumBench-Synthetic</td><td>170</td><td>181</td><td>140</td><td>196</td><td>135</td><td>189</td><td>-</td><td>1011</td></tr><tr><td>VisNumBench-Real</td><td>149</td><td>162</td><td>143</td><td>147</td><td>154</td><td>-</td><td>147</td><td>902</td></tr><tr><td>Answer Format</td><td>4/5 options</td><td>3/4 options</td><td>4 options</td><td>3/4 options</td><td>4 options</td><td>4/5 options</td><td>3/4 options</td><td>3/4/5 options</td></tr></table>
+
+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.  
+
+<table><tr><td></td><td>Angle</td><td>Length</td><td>Scale</td><td>Quantity</td><td>Depth</td><td>Area</td><td>Average</td></tr><tr><td>Random</td><td>24.44</td><td>25.41</td><td>25.00</td><td>25.00</td><td>25.00</td><td>23.68</td><td>24.76</td></tr><tr><td colspan="8">Open-source MLLMs</td></tr><tr><td>Phi-3.5-vision</td><td>19.41</td><td>40.88</td><td>41.43</td><td>26.53</td><td>26.67</td><td>39.15</td><td>32.34</td></tr><tr><td>LLaVA-v1.5-7B</td><td>31.18</td><td>30.39</td><td>34.29</td><td>33.16</td><td>26.67</td><td>21.16</td><td>29.38</td></tr><tr><td>LLaVA-v1.5-13B</td><td>35.88</td><td>30.94</td><td>32.14</td><td>36.73</td><td>33.33</td><td>24.34</td><td>32.15</td></tr><tr><td>LLaVA-v1.6-34B</td><td>40.00</td><td>45.30</td><td>40.00</td><td>46.94</td><td>64.44</td><td>33.33</td><td>44.31</td></tr><tr><td>LLaVA-Onevision-7B</td><td>25.88</td><td>51.38</td><td>42.86</td><td>38.78</td><td>34.81</td><td>44.44</td><td>39.96</td></tr><tr><td>LLaVA-Onevision-72B</td><td>24.71</td><td>61.33</td><td>62.14</td><td>50.00</td><td>48.15</td><td>58.73</td><td>50.84</td></tr><tr><td>InternVL2.5-8B</td><td>26.47</td><td>41.99</td><td>49.29</td><td>34.69</td><td>41.48</td><td>46.03</td><td>39.66</td></tr><tr><td>InternVL2.5-38B</td><td>39.41</td><td>59.67</td><td>59.29</td><td>54.08</td><td>60.74</td><td>61.38</td><td>55.59</td></tr><tr><td>InternVL2.5-78B</td><td>35.29</td><td>59.67</td><td>68.57</td><td>42.86</td><td>61.48</td><td>72.49</td><td>56.18</td></tr><tr><td>Janus-Pro-7B</td><td>31.76</td><td>43.65</td><td>45.71</td><td>35.71</td><td>33.33</td><td>36.51</td><td>37.69</td></tr><tr><td>Qwen2.5-VL-3B</td><td>30.00</td><td>49.17</td><td>50.71</td><td>32.14</td><td>42.22</td><td>51.85</td><td>42.43</td></tr><tr><td>Qwen2.5-VL-7B</td><td>23.53</td><td>53.59</td><td>55.00</td><td>39.29</td><td>48.89</td><td>58.20</td><td>46.19</td></tr><tr><td>Qwen2.5-VL-72B</td><td>37.06</td><td>59.67</td><td>65.00</td><td>57.65</td><td>61.48</td><td>70.37</td><td>58.46</td></tr><tr><td colspan="8">API-based models</td></tr><tr><td>GPT-4o</td><td>35.29</td><td>43.09</td><td>54.29</td><td>37.24</td><td>54.07</td><td>43.39</td><td>43.72</td></tr><tr><td>Gemini 1.5 Flash</td><td>26.47</td><td>47.41</td><td>44.40</td><td>26.02</td><td>23.70</td><td>41.27</td><td>33.33</td></tr><tr><td>Gemini 2.0 Flash</td><td>31.18</td><td>57.46</td><td>81.43</td><td>55.10</td><td>51.11</td><td>70.90</td><td>57.57</td></tr><tr><td>Gemini 1.5 Pro</td><td>34.12</td><td>39.23</td><td>47.14</td><td>40.82</td><td>58.52</td><td>48.15</td><td>44.02</td></tr><tr><td>Human</td><td>90.00</td><td>96.00</td><td>100.00</td><td>96.00</td><td>98.00</td><td>92.00</td><td>95.33</td></tr></table>
+
+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.  
+
+<table><tr><td></td><td>Angle</td><td>Length</td><td>Scale</td><td>Quantity</td><td>Depth</td><td>Volume</td><td>Average</td></tr><tr><td>Random</td><td>25.00</td><td>25.00</td><td>25.00</td><td>27.83</td><td>25.00</td><td>25.40</td><td>25.54</td></tr><tr><td colspan="8">Open-source MLLMs</td></tr><tr><td>Phi-3.5-vision</td><td>30.20</td><td>37.65</td><td>27.97</td><td>48.30</td><td>48.70</td><td>29.93</td><td>37.25</td></tr><tr><td>LLaVA-v1.5-7B</td><td>22.82</td><td>32.72</td><td>25.87</td><td>36.73</td><td>25.32</td><td>27.21</td><td>28.49</td></tr><tr><td>LLaVA-v1.5-13B</td><td>28.86</td><td>43.21</td><td>29.37</td><td>46.94</td><td>49.35</td><td>41.50</td><td>40.02</td></tr><tr><td>LLaVA-v1.6-34B</td><td>28.86</td><td>54.94</td><td>23.08</td><td>68.03</td><td>63.64</td><td>63.27</td><td>50.55</td></tr><tr><td>LLaVA-Onevision-7B</td><td>18.12</td><td>44.44</td><td>20.28</td><td>64.63</td><td>44.81</td><td>50.34</td><td>40.58</td></tr><tr><td>LLaVA-Onevision-72B</td><td>17.45</td><td>57.41</td><td>44.76</td><td>74.83</td><td>48.70</td><td>61.22</td><td>50.78</td></tr><tr><td>InternVL2.5-8B</td><td>28.86</td><td>34.57</td><td>15.38</td><td>64.63</td><td>49.35</td><td>47.62</td><td>40.13</td></tr><tr><td>InternVL2.5-38B</td><td>30.20</td><td>51.85</td><td>26.57</td><td>83.67</td><td>61.04</td><td>58.50</td><td>52.11</td></tr><tr><td>InternVL2.5-78B</td><td>36.91</td><td>58.64</td><td>48.95</td><td>79.59</td><td>52.60</td><td>62.59</td><td>56.54</td></tr><tr><td>Janus-Pro-7B</td><td>22.82</td><td>32.10</td><td>35.66</td><td>48.98</td><td>35.71</td><td>30.61</td><td>34.26</td></tr><tr><td>Qwen2.5-VL-3B</td><td>30.20</td><td>44.44</td><td>35.66</td><td>51.70</td><td>43.51</td><td>49.66</td><td>42.57</td></tr><tr><td>Qwen2.5-VL-7B</td><td>24.16</td><td>38.89</td><td>32.17</td><td>59.18</td><td>48.70</td><td>42.86</td><td>41.02</td></tr><tr><td>Qwen2.5-VL-72B</td><td>34.23</td><td>50.62</td><td>43.36</td><td>80.27</td><td>52.60</td><td>59.18</td><td>53.33</td></tr><tr><td colspan="8">API-based models</td></tr><tr><td>GPT-4o</td><td>27.52</td><td>30.25</td><td>37.06</td><td>60.54</td><td>35.71</td><td>47.62</td><td>39.58</td></tr><tr><td>Gemini 1.5 Flash</td><td>14.77</td><td>35.80</td><td>26.57</td><td>57.14</td><td>24.68</td><td>43.54</td><td>33.70</td></tr><tr><td>Gemini 2.0 Flash</td><td>38.93</td><td>48.77</td><td>74.14</td><td>81.63</td><td>46.10</td><td>51.70</td><td>56.54</td></tr><tr><td>Gemini 1.5 Pro</td><td>30.20</td><td>45.68</td><td>27.97</td><td>68.03</td><td>64.29</td><td>55.10</td><td>48.67</td></tr><tr><td>Human</td><td>96.00</td><td>100.00</td><td>100.00</td><td>98.00</td><td>96.00</td><td>94.00</td><td>97.33</td></tr></table>
+
+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.  
+
+<table><tr><td></td><td>Average (Synthetic)</td><td>Average (Real)</td></tr><tr><td>Qwen2-VL-2B</td><td>31.85</td><td>24.94</td></tr><tr><td>Qwen2.5-VL-3B</td><td>42.24(↑+10.39)</td><td>42.57(↑+17.63)</td></tr><tr><td>Qwen2-VL-7B</td><td>41.25</td><td>41.91</td></tr><tr><td>Qwen2.5-VL-7B</td><td>46.19(↑+4.94)</td><td>41.02(↓-0.89)</td></tr><tr><td>Qwen2-VL-72B</td><td>54.20</td><td>46.56</td></tr><tr><td>Qwen2.5-VL-72B</td><td>58.46(↑+4.26)</td><td>53.33(↑+6.77)</td></tr><tr><td>InternVL2-8B</td><td>39.56</td><td>39.58</td></tr><tr><td>InternVL2.5-8B</td><td>39.66(↑+0.10)</td><td>40.13(↑+0.55)</td></tr><tr><td>InternVL2-40B</td><td>45.50</td><td>45.12</td></tr><tr><td>InternVL2.5-38B</td><td>55.59(↑+10.09)</td><td>52.11(↑+6.99)</td></tr></table>
+
+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

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

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

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

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

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

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.  
+
+<table><tr><td>Method</td><td>LLM</td><td>Vision Encoder</td><td>MME</td><td>MMBench</td><td>POPE</td><td>Dataset Num.</td></tr><tr><td>LLaVA [B] [34]</td><td>Vicuna-7B [13]</td><td>CLIP-ViT-L-14-224 [44]</td><td>1051.2</td><td>34.4</td><td>76.5</td><td>558K</td></tr><tr><td>SEAL [D] [61]</td><td>Vicuna-7B</td><td>CLIP-ViT-L-14-224</td><td>1128.9</td><td>33.1</td><td>82.4</td><td>558K + 387K [D]</td></tr><tr><td>LLaVA + P2G [T] [8]</td><td>Vicuna-7B</td><td>CLIP-ViT-L-14-224</td><td>1223.0</td><td>—</td><td>—</td><td>558K + 427K [D]</td></tr><tr><td>LLaVA + VIsRL</td><td>Vicuna-7B</td><td>CLIP-ViT-L-14-224</td><td>1183.8</td><td>37.5</td><td>78.2</td><td>558K + 30K [D]+180K</td></tr><tr><td>LLaVA + VIsRL-Iter1</td><td>Vicuna-7B</td><td>CLIP-ViT-L-14-224</td><td>1238.3</td><td>38.6</td><td>80.4</td><td>180K</td></tr><tr><td>LLaVA-1.5 [B] [33]</td><td>Vicuna-7B</td><td>CLIP-ViT-L-14-336</td><td>1510.7</td><td>64.3</td><td>85.8</td><td>558K</td></tr><tr><td>VisCoT [D] [48]</td><td>Vicuna-7B</td><td>CLIP-ViT-L-14-336</td><td>1453.6</td><td>67.9</td><td>86.0</td><td>558K + 376K [D]</td></tr><tr><td>LLaVA-1.5 + VIsRL</td><td>Vicuna-7B</td><td>CLIP-ViT-L-14-336</td><td>1526.3</td><td>70.1</td><td>87.5</td><td>558K + 30K [D]+180K</td></tr><tr><td>LLaVA-1.5 + VIsRL-Iter1</td><td>Vicuna-7B</td><td>CLIP-ViT-L-14-336</td><td>1560.0</td><td>71.7</td><td>88.8</td><td>180K</td></tr><tr><td>LLaVA-NeXT [B] [35]</td><td>Vicuna-7B-1.5 [72]</td><td>CLIP-ViT-L-14-336</td><td>1611.1</td><td>72.3</td><td>—</td><td>558K</td></tr><tr><td>VisionLLM v2 [D] [60]</td><td>Vicuna-7B-1.5</td><td>CLIP-ViT-L-14-336</td><td>1512.5</td><td>77.1</td><td>87.5</td><td>892K</td></tr><tr><td>Insight-V-LLaVA [T] [15]</td><td>Vicuna-7B-1.5</td><td>CLIP-ViT-L-14-336</td><td>1583.9</td><td>81.7</td><td>—</td><td>558K + 215K [D]</td></tr><tr><td>LLaVA-NeXT + VIsRL</td><td>Vicuna-7B-1.5</td><td>CLIP-ViT-L-14-336</td><td>1619.2</td><td>78.8</td><td>88.4</td><td>558K + 30K [D]+180K</td></tr><tr><td>LLaVA-NeXT + VIsRL-Iter1</td><td>Vicuna-7B-1.5</td><td>CLIP-ViT-L-14-336</td><td>1637.0</td><td>80.0</td><td>89.3</td><td>180K</td></tr></table>
+
+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.  
+
+<table><tr><td>Training Phase</td><td>Base</td><td>SFT</td><td>SFT+RL1</td><td>SFT+RL1+RL2</td></tr><tr><td>LLaVA-1.5-7B [33]</td><td>0.444</td><td>0.538</td><td>0.568</td><td>0.605</td></tr><tr><td>LLaVA-NeXT-7B [35]</td><td>0.459</td><td>0.556</td><td>0.593</td><td>0.628</td></tr><tr><td>Llama-3.2-V-11B [39]</td><td>0.635</td><td>0.724</td><td>0.749</td><td>0.786</td></tr><tr><td>MiniCPM-o-2.6-8B [66]</td><td>0.428</td><td>0.519</td><td>0.544</td><td>0.591</td></tr><tr><td>PaliGemma2-10B [50]</td><td>0.377</td><td>0.479</td><td>0.507</td><td>0.562</td></tr><tr><td>Yi-VL-6B [67]</td><td>0.389</td><td>0.492</td><td>0.523</td><td>0.573</td></tr><tr><td>Qwen2.5-VL-7B [3]</td><td>0.640</td><td>0.730</td><td>0.763</td><td>0.796</td></tr></table>
+
+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.  
+
+<table><tr><td>SFT Data Num.</td><td>10k</td><td>20k</td><td>30k</td><td>50k</td><td>100k</td></tr><tr><td>Ratio</td><td>28.14%</td><td>95.48%</td><td>99.87%</td><td>99.87%</td><td>99.87%</td></tr></table>
+
+![](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.  
+
+<table><tr><td rowspan="2">Method</td><td colspan="5">Doc/Text</td><td>Chart</td><td>General VQA</td><td colspan="3">Relation Reasoning</td><td colspan="2">Fine-grained</td></tr><tr><td>DocVQA</td><td>TextCaps</td><td>TextVQA</td><td>DUDE</td><td>SROIE</td><td>InfogVQA</td><td>Flickr30k</td><td>GQA</td><td>Open images</td><td>VSR</td><td>CUB</td><td>Avg</td></tr><tr><td>VisCoT [438k] [48]</td><td>0.136</td><td>0.413</td><td>0.468</td><td>0.050</td><td>0.157</td><td>0.072</td><td>0.496</td><td>0.420</td><td>0.576</td><td>0.696</td><td>0.670</td><td>0.378</td></tr><tr><td>VisRL-SFT [30k]</td><td>0.133</td><td>0.406</td><td>0.460</td><td>0.048</td><td>0.146</td><td>0.069</td><td>0.495</td><td>0.421</td><td>0.575</td><td>0.691</td><td>0.673</td><td>0.374</td></tr><tr><td>VisRL-SFT+RL1</td><td>0.190</td><td>0.433</td><td>0.471</td><td>0.081</td><td>0.216</td><td>0.133</td><td>0.518</td><td>0.450</td><td>0.581</td><td>0.733</td><td>0.701</td><td>0.410</td></tr><tr><td>VisRL-SFT+RL1+RL2</td><td>0.231</td><td>0.464</td><td>0.504</td><td>0.124</td><td>0.252</td><td>0.172</td><td>0.519</td><td>0.487</td><td>0.590</td><td>0.745</td><td>0.715</td><td>0.437</td></tr></table>
+
+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.  
+
+<table><tr><td rowspan="2">Method</td><td rowspan="2">Doc/VQA</td><td colspan="2">Doc/Text</td><td colspan="3">Chart</td><td colspan="2">General VQA</td><td colspan="2">Relation Reasoning</td><td colspan="2">Fine-grained</td></tr><tr><td>TextCaps</td><td>TextVQA</td><td>DUDE</td><td>SROIE</td><td>InfogVQA</td><td>Flickr30k</td><td>GQA</td><td>Open images</td><td>VSR</td><td>CUB</td><td>Avg</td></tr><tr><td>Ori</td><td>0.836</td><td>0.760</td><td>0.847</td><td>0.606</td><td>0.789</td><td>0.685</td><td>0.601</td><td>0.467</td><td>0.289</td><td>0.581</td><td>0.583</td><td>0.640</td></tr><tr><td>SFT</td><td>0.778</td><td>0.751</td><td>0.856</td><td>0.566</td><td>0.713</td><td>0.631</td><td>0.627</td><td>0.524</td><td>0.345</td><td>0.626</td><td>0.736</td><td>0.650</td></tr><tr><td>DPO [45]</td><td>0.791</td><td>0.793</td><td>0.837</td><td>0.599</td><td>0.776</td><td>0.678</td><td>0.670</td><td>0.470</td><td>0.298</td><td>0.594</td><td>0.707</td><td>0.656</td></tr><tr><td>KTO [16]</td><td>0.797</td><td>0.790</td><td>0.833</td><td>0.597</td><td>0.780</td><td>0.694</td><td>0.671</td><td>0.469</td><td>0.299</td><td>0.599</td><td>0.707</td><td>0.658</td></tr><tr><td>PPO [46]</td><td>0.794</td><td>0.791</td><td>0.837</td><td>0.597</td><td>0.787</td><td>0.700</td><td>0.672</td><td>0.472</td><td>0.302</td><td>0.606</td><td>0.713</td><td>0.661</td></tr><tr><td>VisRL-RL2-Only</td><td>0.816</td><td>0.769</td><td>0.897</td><td>0.602</td><td>0.800</td><td>0.681</td><td>0.637</td><td>0.652</td><td>0.786</td><td>0.827</td><td>0.882</td><td>0.759</td></tr><tr><td>VisRL-Full</td><td>0.874</td><td>0.819</td><td>0.897</td><td>0.640</td><td>0.829</td><td>0.753</td><td>0.675</td><td>0.700</td><td>0.814</td><td>0.864</td><td>0.892</td><td>0.796</td></tr></table>
+
+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.  
+
+<table><tr><td></td><td>WP-LP</td><td>WP-LN</td><td>WN-LP</td><td>WN-LN</td><td>Data Num.</td></tr><tr><td>w GPT-4o-2024-11-20</td><td>0.00%</td><td>65.31%</td><td>1.32%</td><td>33.37%</td><td>47k</td></tr><tr><td>w SFTed Model</td><td>0.00%</td><td>54.68%</td><td>0.00%</td><td>45.32%</td><td>3k</td></tr><tr><td>w/o Bounding Box Critics</td><td>5.42%</td><td>31.02%</td><td>10.04%</td><td>53.51%</td><td>86k</td></tr><tr><td>w/o Diversity Controller</td><td>4.53%</td><td>52.02%</td><td>4.68%</td><td>38.77%</td><td>19k</td></tr><tr><td>V1sRL-Full</td><td>0.43%</td><td>64.64%</td><td>1.64%</td><td>33.29%</td><td>30k</td></tr><tr><td>V1sRL-Full-Iter1</td><td>0.45%</td><td>67.82%</td><td>0.82%</td><td>30.91%</td><td>33k</td></tr><tr><td>V1sRL-Full-Iter2</td><td>0.47%</td><td>70.12%</td><td>0.00%</td><td>29.41%</td><td>35k</td></tr></table>
+
+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

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

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

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

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

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

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
+
+<table><tr><td rowspan="2">Method</td><td rowspan="2">Backbone</td><td rowspan="2">Training Dataset</td><td colspan="2">Novel+base</td><td colspan="2">Novel</td></tr><tr><td>R@50/100</td><td>mR@50/100</td><td>R@50/100</td><td>mR@50/100</td></tr><tr><td>VS3 [43]</td><td>GLIP</td><td>VG</td><td>11.0 / 12.8</td><td>5.1 / 5.7</td><td>- / -</td><td>0.0 / 0.0</td></tr><tr><td>SGTR [20]</td><td>BLIP</td><td>VG</td><td>12.6 / 18.2</td><td>3.5 / 5.4</td><td>- / -</td><td>0.0 / 0.0</td></tr><tr><td>PGSG [21]</td><td>BLIP</td><td>VG</td><td>15.8 / 19.1</td><td>6.2 / 8.3</td><td>- / -</td><td>3.7 / 5.2</td></tr><tr><td>Ours</td><td>GLIP</td><td>VG</td><td>13.7 / 19.7</td><td>7.5 / 10.6</td><td>10.1 / 14.3</td><td>5.2 / 7.5</td></tr><tr><td>VS3* [43]</td><td>GLIP</td><td>VG</td><td>15.6 / 17.3</td><td>- / -</td><td>0.0 / 0.0</td><td>- / -</td></tr><tr><td>OVSGTR* [5]</td><td>GroundingDINO</td><td>VG</td><td>17.7 / 20.0</td><td>6.4 / 7.5</td><td>0.3 / 0.4</td><td>- / -</td></tr><tr><td>Ours*</td><td>GLIP</td><td>VG</td><td>14.3 / 20.4</td><td>8.7 / 12.7</td><td>9.0 / 12.6</td><td>- / -</td></tr><tr><td>OVSGTR* [5]</td><td>GroundingDINO</td><td>VG, COCO, Flickr30k, SBU</td><td>20.5 / 23.9</td><td>3.9 / 4.6</td><td>13.5 / 16.2</td><td>- / -</td></tr><tr><td>OVSGTR+Ours*</td><td>GroundingDINO</td><td>VG, COCO, Flickr30k, SBU</td><td>21.1 / 25.0</td><td>4.2 / 4.9</td><td>14.2 / 17.7</td><td>- / -</td></tr></table>
+
+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.  
+
+<table><tr><td rowspan="2">Method</td><td rowspan="2">Backbone</td><td rowspan="2">Training Dataset</td><td colspan="2">Novel+base</td><td colspan="2">Novel</td></tr><tr><td>R@50/100</td><td>mR@50/100</td><td>R@50/100</td><td>mR@50/100</td></tr><tr><td>SVRP [8]</td><td>Faster-RCNN</td><td>VG</td><td>33.5 / 35.9</td><td>8.3 / 10.8</td><td>- / -</td><td>- / -</td></tr><tr><td>CaCao [40]</td><td>CLIP</td><td>VG,CC3M,COCO caption</td><td>- / -</td><td>10.3 / 12.6</td><td>- / -</td><td>- / -</td></tr><tr><td>PGSG [21]</td><td>BLIP</td><td>VG</td><td>26.9 / 33.9</td><td>10.8 / 13.9</td><td>- / -</td><td>5.2 / 7.7</td></tr><tr><td>Ours</td><td>GLIP</td><td>VG</td><td>27.8 / 37.5</td><td>11.1 / 14.5</td><td>11.9 / 18.3</td><td>5.9 / 8.2</td></tr></table>
+
+Table 2. Experimental results of open-vocabulary scene graph generation of the PredCls task on VG.  
+
+<table><tr><td rowspan="2">Method</td><td colspan="2">Novel+Base</td><td colspan="2">Novel</td></tr><tr><td>R@50/100</td><td>mR@50/100</td><td>R@50/100</td><td>mR@50/100</td></tr><tr><td>SGTR [20]</td><td>36.1 / 38.4</td><td>11.0 / 16.7</td><td>- / -</td><td>0.0 / 0.0</td></tr><tr><td>PGSG [21]</td><td>41.3 / 43.3</td><td>20.8 / 23.0</td><td>- / -</td><td>3.8 / 8.9</td></tr><tr><td>Ours</td><td>46.3 / 59.4</td><td>19.8 / 24.6</td><td>14.6 / 18.7</td><td>4.5 / 10.1</td></tr></table>
+
+$\%$  [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.  
+
+<table><tr><td rowspan="2">Method</td><td colspan="2">Novel+Base</td><td colspan="2">Novel</td></tr><tr><td>R@50</td><td>R@100</td><td>R@50</td><td>R@100</td></tr><tr><td>IMP [38]</td><td>0.77</td><td>0.94</td><td>0.00</td><td>0.00</td></tr><tr><td>VCTREE [31]</td><td>1.04</td><td>1.17</td><td>0.00</td><td>0.00</td></tr><tr><td>TDE [32]</td><td>1.00</td><td>1.10</td><td>0.00</td><td>0.00</td></tr><tr><td>VS3-SwinT [43]</td><td>5.88</td><td>7.20</td><td>0.00</td><td>0.00</td></tr><tr><td>OVSGTR-T [5]</td><td>13.53</td><td>16.36</td><td>9.20</td><td>11.19</td></tr><tr><td>OVSGTR-T+Ours</td><td>14.87</td><td>18.27</td><td>12.27</td><td>15.26</td></tr></table>
+
+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
+
+<table><tr><td rowspan="2">Method</td><td colspan="4">Novel+Base</td></tr><tr><td>R@50</td><td>R@100</td><td>mR@50</td><td>mR@100</td></tr><tr><td>Frozen</td><td>10.66</td><td>15.36</td><td>4.75</td><td>6.60</td></tr><tr><td>w/o Context</td><td>11.33</td><td>16.67</td><td>5.89</td><td>6.68</td></tr><tr><td>w Context</td><td>12.94</td><td>18.48</td><td>7.17</td><td>9.60</td></tr></table>
+
+![](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.  
+
+<table><tr><td rowspan="2">Method</td><td colspan="2">Novel+base</td><td colspan="2">Novel</td></tr><tr><td>R@100</td><td>mR@100</td><td>R@100</td><td>mR@100</td></tr><tr><td>w/o Debias</td><td>16.24</td><td>6.62</td><td>10.23</td><td>4.36</td></tr><tr><td>LA [28]</td><td>12.58</td><td>6.56</td><td>10.47</td><td>3.21</td></tr><tr><td>PGLA [23]</td><td>13.42</td><td>9.12</td><td>10.76</td><td>5.63</td></tr><tr><td>Dynamic Fitting</td><td>18.48</td><td>9.60</td><td>13.45</td><td>6.06</td></tr></table>
+
+Table 6. Ablation study on different bias estimation strategies.  
+
+<table><tr><td rowspan="2">Method</td><td colspan="2">Novel+Base</td><td colspan="2">Novel</td></tr><tr><td>R@100</td><td>mR@100</td><td>R@100</td><td>mR@100</td></tr><tr><td>w/o Pseudo label</td><td>18.48</td><td>9.60</td><td>13.45</td><td>6.06</td></tr><tr><td>Only fg</td><td>19.12</td><td>10.38</td><td>13.78</td><td>6.58</td></tr><tr><td>Bg &amp; fg</td><td>19.71</td><td>10.60</td><td>14.32</td><td>7.51</td></tr></table>
+
+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

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

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

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

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

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

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 <box coordinates> 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
+
+<table><tr><td rowspan="2">Method</td><td rowspan="2">Shot</td><td colspan="2">Detection</td><td colspan="2">Referring</td><td colspan="2">Visual Referring</td></tr><tr><td>NAT</td><td>SYN</td><td>NAT</td><td>SYN</td><td>NAT</td><td>SYN</td></tr><tr><td>Claude-sonnet</td><td>0</td><td>48.2</td><td>86.7</td><td>51.1</td><td>90.3</td><td>53.9</td><td>87.7</td></tr><tr><td>NVLM-D-72B [13]</td><td>0</td><td>41.5</td><td>77.5</td><td>42.0</td><td>73.7</td><td>37.3</td><td>51.7</td></tr><tr><td>Molmo-7B[14]</td><td>0</td><td>32.0</td><td>67.2</td><td>32.4</td><td>38.0</td><td>33.1</td><td>28.4</td></tr><tr><td>Molmo-72B[14]</td><td>0</td><td>40.6</td><td>83.3</td><td>41.2</td><td>65.4</td><td>36.7</td><td>73.6</td></tr><tr><td>Llama3.2-Vision-11B [15]</td><td>0</td><td>32.1</td><td>48.7</td><td>29.1</td><td>52.4</td><td>29.7</td><td>52.4</td></tr><tr><td>PaliGemma-3B-448[7]</td><td>0</td><td>27.6</td><td>42.0</td><td>1.2</td><td>9.5</td><td>2.3</td><td>4.8</td></tr><tr><td rowspan="3">Phi3-4B[1]</td><td>0</td><td>32.1</td><td>41.2</td><td>32.8</td><td>55.3</td><td>32.8</td><td>47.2</td></tr><tr><td>3</td><td>34.1</td><td>33.5</td><td>32.0</td><td>27.1</td><td>32.1</td><td>38.5</td></tr><tr><td>5</td><td>31.1</td><td>17.0</td><td>32.1</td><td>18.9</td><td>32.2</td><td>46.7</td></tr><tr><td rowspan="3">Phi3.5-Vision-3.5B[1]</td><td>0</td><td>23.2</td><td>35.0</td><td>27.5</td><td>53.7</td><td>27.5</td><td>63.5</td></tr><tr><td>3</td><td>23.3</td><td>19.5</td><td>28.8</td><td>41.0</td><td>28.8</td><td>20.8</td></tr><tr><td>5</td><td>25.2</td><td>29.3</td><td>30.8</td><td>11.1</td><td>30.8</td><td>19.0</td></tr><tr><td rowspan="3">LLava 1.6-7B[31]</td><td>0</td><td>24.5</td><td>16.3</td><td>21.4</td><td>10.1</td><td>20.8</td><td>16.6</td></tr><tr><td>3</td><td>7.0</td><td>16.4</td><td>15.2</td><td>8.8</td><td>17.8</td><td>17.0</td></tr><tr><td>5</td><td>11.4</td><td>16.4</td><td>10.9</td><td>9.1</td><td>9.7</td><td>17.0</td></tr><tr><td rowspan="3">Idefics2-8B[26]</td><td>0</td><td>19.5</td><td>64.3</td><td>29.6</td><td>36.6</td><td>33.8</td><td>49.5</td></tr><tr><td>3</td><td>21.1</td><td>66.3</td><td>28.4</td><td>34.2</td><td>31.1</td><td>39.6</td></tr><tr><td>5</td><td>34.7</td><td>67.2</td><td>28.3</td><td>42.6</td><td>30.9</td><td>34.5</td></tr><tr><td rowspan="3">Idefics3-8B[25]</td><td>0</td><td>24.3</td><td>28.4</td><td>24.3</td><td>52.8</td><td>22.1</td><td>19.2</td></tr><tr><td>3</td><td>26.9</td><td>40.3</td><td>26.9</td><td>20.67</td><td>21.9</td><td>40.6</td></tr><tr><td>5</td><td>22.3</td><td>21.4</td><td>22.3</td><td>18.1</td><td>20.9</td><td>58.3</td></tr><tr><td rowspan="3">VILA-1.5-8B[29]</td><td>0</td><td>23.5</td><td>40.0</td><td>13.0</td><td>23.7</td><td>15.8</td><td>17.0</td></tr><tr><td>3</td><td>25.1</td><td>17.0</td><td>28.8</td><td>21.2</td><td>28.8</td><td>17.0</td></tr><tr><td>5</td><td>23.2</td><td>17.0</td><td>30.8</td><td>20.7</td><td>30.8</td><td>17.0</td></tr><tr><td rowspan="3">Qwen2-VL-1.5B[51]</td><td>0</td><td>19.2</td><td>26.3</td><td>22.1</td><td>20.6</td><td>20.9</td><td>20.2</td></tr><tr><td>3</td><td>25.2</td><td>23.3</td><td>21.4</td><td>21.8</td><td>20.2</td><td>16.3</td></tr><tr><td>5</td><td>25.3</td><td>23.8</td><td>21.7</td><td>16.5</td><td>20.9</td><td>17.7</td></tr><tr><td rowspan="3">Qwen2-VL-7B[51]</td><td>0</td><td>32.5</td><td>55.7</td><td>32.5</td><td>34.2</td><td>35.2</td><td>57.4</td></tr><tr><td>3</td><td>35.6</td><td>53.8</td><td>36.0</td><td>17.0</td><td>34.1</td><td>64.2</td></tr><tr><td>5</td><td>37.2</td><td>54.9</td><td>37.2</td><td>17.7</td><td>29.3</td><td>72.0</td></tr><tr><td rowspan="3">Qwen2-VL-72B[51]</td><td>0</td><td>41.6</td><td>88.8</td><td>44.6</td><td>93.6</td><td>41.7</td><td>74.7</td></tr><tr><td>3</td><td>43.9</td><td>89.3</td><td>43.6</td><td>93.1</td><td>43.2</td><td>85.9</td></tr><tr><td>5</td><td>43.9</td><td>89.9</td><td>44.9</td><td>92.6</td><td>42.3</td><td>87.9</td></tr><tr><td rowspan="3">InternVL-4B[12]</td><td>0</td><td>26.6</td><td>41.5</td><td>29.8</td><td>63.4</td><td>30.7</td><td>52.2</td></tr><tr><td>3</td><td>27.7</td><td>17.0</td><td>27.4</td><td>25.3</td><td>29.5</td><td>41.7</td></tr><tr><td>5</td><td>33.4</td><td>17.0</td><td>28.1</td><td>39.1</td><td>30.4</td><td>52.5</td></tr><tr><td rowspan="3">InternVL-2-8B[12]</td><td>0</td><td>20.0</td><td>58.7</td><td>23.0</td><td>71.9</td><td>24.8</td><td>23.0</td></tr><tr><td>3</td><td>30.5</td><td>52.3</td><td>24.2</td><td>51.7</td><td>31.7</td><td>64.4</td></tr><tr><td>5</td><td>27.8</td><td>43.9</td><td>25.0</td><td>53.7</td><td>31.4</td><td>50.5</td></tr><tr><td rowspan="3">GPT-4o</td><td>0</td><td>47.6</td><td>89.2</td><td>47.3</td><td>88.4</td><td>42.6</td><td>73.5</td></tr><tr><td>3</td><td>38.9</td><td>88.4</td><td>37.5</td><td>87.7</td><td>35.7</td><td>86.7</td></tr><tr><td>5</td><td>41.9</td><td>86.0</td><td>39.8</td><td>89.1</td><td>38.4</td><td>87.4</td></tr></table>
+
+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
+
+<table><tr><td rowspan="2">Models</td><td colspan="2">SYN</td><td colspan="2">NAT</td></tr><tr><td>Top1</td><td>Top3</td><td>Top1</td><td>Top3</td></tr><tr><td>Random</td><td>24.6</td><td>53.2</td><td>14.6</td><td>36.3</td></tr><tr><td>Siglip-so400m-patch14-384</td><td>55.3</td><td>87.9</td><td>0.0</td><td>57.2</td></tr><tr><td>CLIP-ViT-Base-Patch32</td><td>42.1</td><td>71.7</td><td>12.0</td><td>59.8</td></tr><tr><td>CLIP-ViT-Base-Patch16</td><td>41.6</td><td>78.4</td><td>0.9</td><td>40.8</td></tr><tr><td>CLIP-ViT-Large-Patch14</td><td>41.2</td><td>78.6</td><td>0.0</td><td>31.6</td></tr><tr><td>CLIP-ViT-Large-Patch14-336</td><td>47.5</td><td>79.2</td><td>0.0</td><td>36.3</td></tr></table>
+
+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
+
+<table><tr><td rowspan="2">Model</td><td rowspan="2">Category</td><td rowspan="2">Level</td><td colspan="3">Accuracy</td></tr><tr><td>Detection</td><td>Referring</td><td>Visual Referring</td></tr><tr><td rowspan="3">Qwen2-VL-72B</td><td rowspan="3">Orientation</td><td>Easy</td><td>98.6</td><td>97.1</td><td>88.3</td></tr><tr><td>Medium</td><td>95.8</td><td>96.8</td><td>83.0</td></tr><tr><td>Hard</td><td>95.7</td><td>96.2</td><td>71.7</td></tr><tr><td rowspan="3">GPT-4o</td><td rowspan="3">Orientation</td><td>Easy</td><td>96.2</td><td>96.5</td><td>97.6</td></tr><tr><td>Medium</td><td>96.2</td><td>89.2</td><td>94.7</td></tr><tr><td>Hard</td><td>98.6</td><td>88.5</td><td>96.2</td></tr><tr><td rowspan="3">Qwen2-VL-72B</td><td rowspan="3">Size</td><td>Easy</td><td>94.2</td><td>98.8</td><td>62.3</td></tr><tr><td>Medium</td><td>50.5</td><td>74.3</td><td>31.9</td></tr><tr><td>Hard</td><td>46.0</td><td>62.0</td><td>33.3</td></tr><tr><td rowspan="3">GPT-4o</td><td rowspan="3">Size</td><td>Easy</td><td>93.3</td><td>96.9</td><td>42.5</td></tr><tr><td>Medium</td><td>59.1</td><td>59.1</td><td>39.1</td></tr><tr><td>Hard</td><td>36.8</td><td>35.3</td><td>11.3</td></tr><tr><td rowspan="3">Qwen2-VL-72B</td><td rowspan="3">Color</td><td>Easy</td><td>100.0</td><td>100.0</td><td>99.1</td></tr><tr><td>Medium</td><td>100.0</td><td>100.0</td><td>98.6</td></tr><tr><td>Hard</td><td>60.1</td><td>80.2</td><td>73.5</td></tr><tr><td rowspan="3">GPT-4o</td><td rowspan="3">Color</td><td>Easy</td><td>99.8</td><td>99.0</td><td>96.1</td></tr><tr><td>Medium</td><td>100.0</td><td>100.0</td><td>93.3</td></tr><tr><td>Hard</td><td>66.1</td><td>66.2</td><td>57.9</td></tr></table>
+
+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.  
+
+<table><tr><td># Distractors</td><td>Orientation</td><td>Color</td><td>Focus</td><td>Shape</td><td>Size</td><td>Location</td><td>Pattern</td><td>Avg F1</td></tr><tr><td colspan="9">Qwen2-VL-72B</td></tr><tr><td>&lt;7</td><td>18.4</td><td>92.8</td><td>18.6</td><td>53.4</td><td>39.7</td><td>30.3</td><td>57.9</td><td>44.5</td></tr><tr><td>7-15</td><td>9.3</td><td>93.7</td><td>13.0</td><td>55.8</td><td>37.8</td><td>31.6</td><td>41.2</td><td>40.4</td></tr><tr><td>15-25</td><td>23.5</td><td>95.8</td><td>11.1</td><td>59.1</td><td>39.5</td><td>0.0</td><td>37.5</td><td>38.1</td></tr><tr><td>&gt;25</td><td>16.7</td><td>96.1</td><td>15.0</td><td>52.2</td><td>36.5</td><td>20.0</td><td>25.4</td><td>37.4</td></tr><tr><td colspan="9">GPT-4o</td></tr><tr><td>&lt;7</td><td>37.1</td><td>88.2</td><td>38.1</td><td>49.4</td><td>35.7</td><td>23.6</td><td>57.0</td><td>47.0</td></tr><tr><td>7-15</td><td>46.2</td><td>90.1</td><td>40.0</td><td>47.5</td><td>32.6</td><td>27.5</td><td>53.3</td><td>48.2</td></tr><tr><td>15-25</td><td>43.8</td><td>91.5</td><td>45.2</td><td>47.5</td><td>31.0</td><td>12.5</td><td>45.1</td><td>45.2</td></tr><tr><td>&gt;25</td><td>40.5</td><td>89.8</td><td>48.4</td><td>48.1</td><td>30.4</td><td>4.8</td><td>29.4</td><td>41.6</td></tr></table>
+
+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

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

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

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

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

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

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 Hassan<sup>1</sup>, Anabia Sohail<sup>2</sup>, Muzammal Naseer<sup>2</sup>, Naoufel Werghi<sup>2</sup>
+
+<sup>1</sup>Abu Dhabi University, UAE <sup>2</sup>Khalifa 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'.
+
+<table><tr><td>Networks</td><td>Input Dimension</td><td>Output Dimension</td><td>Total Model Layers</td></tr><tr><td>N</td><td>224 × 224 × 3</td><td>1 × 2048</td><td>174 without LL</td></tr><tr><td>N0</td><td>1 × 2048</td><td>1 × 16</td><td>4 without LL</td></tr><tr><td>NEp</td><td>1 × 4112</td><td>1 × 16</td><td>4 without LL</td></tr><tr><td>NUs</td><td>1 × 4112</td><td>1 × 2048</td><td>4 without LL</td></tr><tr><td>NDs</td><td>1 × 4112</td><td>1 × 2048</td><td>4 without LL</td></tr><tr><td>NNp</td><td>1 × 4096</td><td>1 × 2048</td><td>4 without LL</td></tr></table>
+
+# 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.  
+
+<table><tr><td>Dataset</td><td>Segments</td><td>Precision</td><td>Recall</td><td>F1</td><td>IoU</td></tr><tr><td rowspan="3">ODIR</td><td>100</td><td>0.88</td><td>0.91</td><td>0.89</td><td>0.80</td></tr><tr><td>50</td><td>0.84</td><td>0.87</td><td>0.85</td><td>0.74</td></tr><tr><td>10</td><td>0.77</td><td>0.81</td><td>0.79</td><td>0.65</td></tr><tr><td rowspan="3">BIOMISA</td><td>100</td><td>0.92</td><td>0.94</td><td>0.93</td><td>0.86</td></tr><tr><td>50</td><td>0.85</td><td>0.90</td><td>0.87</td><td>0.77</td></tr><tr><td>10</td><td>0.79</td><td>0.82</td><td>0.80</td><td>0.67</td></tr><tr><td rowspan="3">IDRiD</td><td>100</td><td>0.91</td><td>0.93</td><td>0.92</td><td>0.85</td></tr><tr><td>50</td><td>0.85</td><td>0.89</td><td>0.87</td><td>0.76</td></tr><tr><td>10</td><td>0.76</td><td>0.79</td><td>0.77</td><td>0.63</td></tr></table>
+
+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.  
+
+<table><tr><td>Dataset</td><td>Model</td><td>Precision</td><td>Recall</td><td>F1</td><td>IoU</td><td>GMACs</td><td>Time (ms)</td></tr><tr><td rowspan="3">ODIR</td><td>ViT</td><td>0.86</td><td>0.87</td><td>0.86</td><td>0.76</td><td>19.60</td><td>384</td></tr><tr><td>ResNet-50</td><td>0.84</td><td>0.86</td><td>0.85</td><td>0.73</td><td>17.10</td><td>462</td></tr><tr><td>PVT</td><td>0.88</td><td>0.91</td><td>0.89</td><td>0.80</td><td>10.40</td><td>323</td></tr><tr><td rowspan="3">BIOMISA</td><td>ViT</td><td>0.88</td><td>0.90</td><td>0.89</td><td>0.80</td><td>17.40</td><td>341</td></tr><tr><td>ResNet-50</td><td>0.89</td><td>0.91</td><td>0.90</td><td>0.81</td><td>14.90</td><td>427</td></tr><tr><td>PVT</td><td>0.92</td><td>0.94</td><td>0.93</td><td>0.86</td><td>8.63</td><td>306</td></tr><tr><td rowspan="3">IDRiD</td><td>ViT</td><td>0.89</td><td>0.91</td><td>0.90</td><td>0.81</td><td>16.20</td><td>356</td></tr><tr><td>ResNet-50</td><td>0.87</td><td>0.90</td><td>0.88</td><td>0.79</td><td>12.90</td><td>392</td></tr><tr><td>PVT</td><td>0.91</td><td>0.93</td><td>0.92</td><td>0.85</td><td>8.04</td><td>284</td></tr></table>
+
+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.  
+
+<table><tr><td rowspan="2">K/σ</td><td colspan="3">ODIR</td><td colspan="3">BIOMISA</td><td colspan="3">IDRiD</td></tr><tr><td>1%</td><td>2%</td><td>3%</td><td>1%</td><td>2%</td><td>3%</td><td>1%</td><td>2%</td><td>3%</td></tr><tr><td>5</td><td>11</td><td>8</td><td>5</td><td>9</td><td>6</td><td>4</td><td>13</td><td>10</td><td>6</td></tr><tr><td>10</td><td>16</td><td>12</td><td>9</td><td>13</td><td>10</td><td>7</td><td>17</td><td>14</td><td>11</td></tr><tr><td>15</td><td>21</td><td>17</td><td>13</td><td>16</td><td>12</td><td>9</td><td>24</td><td>21</td><td>19</td></tr></table>
+
+# 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.  
+
+<table><tr><td>Dataset</td><td>Method</td><td>Precision</td><td>Recall</td><td>F1</td><td>IoU</td></tr><tr><td rowspan="2">ODIR</td><td>Zero-shot</td><td>0.88</td><td>0.91</td><td>0.89</td><td>0.80</td></tr><tr><td>Few-shot</td><td>0.93</td><td>0.95</td><td>0.94</td><td>0.88</td></tr><tr><td rowspan="2">BIOMISA</td><td>Zero-shot</td><td>0.92</td><td>0.94</td><td>0.93</td><td>0.86</td></tr><tr><td>Few-shot</td><td>0.95</td><td>0.97</td><td>0.96</td><td>0.92</td></tr><tr><td rowspan="2">IDRiD</td><td>Zero-shot</td><td>0.91</td><td>0.93</td><td>0.92</td><td>0.85</td></tr><tr><td>Few-shot</td><td>0.94</td><td>0.96</td><td>0.95</td><td>0.90</td></tr></table>
+
+# 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.  
+
+<table><tr><td>Dataset</td><td>Method</td><td>Approach</td><td>Precision</td><td>Recall</td><td>F1</td><td>IoU</td></tr><tr><td rowspan="3">ODIR</td><td>Proposed</td><td>Unsupervised</td><td>0.88</td><td>0.91</td><td>0.89</td><td>0.80</td></tr><tr><td>DMSF [10]</td><td>Supervised</td><td>0.83</td><td>0.86</td><td>0.84</td><td>0.73</td></tr><tr><td>MSAG [34]</td><td>Supervised</td><td>0.81</td><td>0.84</td><td>0.82</td><td>0.70</td></tr><tr><td rowspan="4">BIOMISA</td><td>Proposed</td><td>Unsupervised</td><td>0.92</td><td>0.94</td><td>0.93</td><td>0.86</td></tr><tr><td>MSAG [34]</td><td>Supervised</td><td>0.89</td><td>0.91</td><td>0.90</td><td>0.81</td></tr><tr><td>DMSF [10]</td><td>Supervised</td><td>0.90</td><td>0.92</td><td>0.91</td><td>0.83</td></tr><tr><td>DRUNET [2]</td><td>Supervised</td><td>0.87</td><td>0.89</td><td>0.88</td><td>0.78</td></tr><tr><td rowspan="4">IDRiD</td><td>Proposed</td><td>Unsupervised</td><td>0.91</td><td>0.93</td><td>0.92</td><td>0.85</td></tr><tr><td>MSAG [34]</td><td>Supervised</td><td>0.86</td><td>0.89</td><td>0.87</td><td>0.77</td></tr><tr><td>DMSF [10]</td><td>Supervised</td><td>0.87</td><td>0.90</td><td>0.88</td><td>0.79</td></tr><tr><td>DRUNET [2]</td><td>Supervised</td><td>0.85</td><td>0.86</td><td>0.85</td><td>0.74</td></tr><tr><td rowspan="3">FIVES</td><td>Proposed</td><td>Unsupervised</td><td>0.92</td><td>0.93</td><td>0.92</td><td>0.86</td></tr><tr><td>SegDRoWS [21]</td><td>Supervised</td><td>0.87</td><td>0.89</td><td>0.88</td><td>0.79</td></tr><tr><td>MSAG [34]</td><td>Supervised</td><td>0.82</td><td>0.85</td><td>0.83</td><td>0.72</td></tr></table>
+
+![](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

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

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

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

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