{"id": "2602.23172", "title": "Latent Gaussian Splatting for 4D Panoptic Occupancy Tracking", "abstract": "Capturing 4D spatiotemporal scene structure is crucial for the safe and reliable operation of robots in dynamic environments. However, existing approaches typically address only part of the problem: they either provide coarse geometric tracking via bounding boxes or detailed 3D occupancy estimates that lack explicit temporal association and instance-level reasoning. In this work, we present Latent Gaussian Splatting (LaGS) for 4D Panoptic Occupancy Tracking (4D-POT). We revisit the underlying representation and model 3D features as a sparse set of feature-bearing Gaussians. These act as dynamic, volume-oriented keypoints that enable spatially continuous, distance-weighted aggregation of multi-view features before being splatted into a voxel grid for decoding. This point-centric formulation enables flexible, data-dependent receptive fields and long-range spatial interactions that are difficult to capture with local and dense voxel-based operators. A hierarchical Gaussian representation further enables multi-scale reasoning by combining global context from coarse super-points with fine-grained detail from higher-resolution streams. Extensive experiments on Occ3D nuScenes and Waymo demonstrate state-of-the-art performance for 4D-POT. We provide code and models at https://lags.cs.uni-freiburg.de/.", "date": "2026-02-26", "categories": []}
{"id": "2605.20448", "title": "Do Vision-Language Models Understand 3D Scenes or Just Catalogue Objects?", "abstract": "Vision-language models reliably name objects in a scene, but do they represent the 3D layout those objects inhabit? We introduce a 3,034-sample human-curated benchmark targeting three components of spatial understanding: depth-ordered occlusion (probed via three independent counterfactual operationalisations), optical-geometry inference over visible reflections, and volumetric rearrangement planning. Six frontier and open-weight VLMs, scored by trained annotators on 18,204 responses with no LLM-as-judge, reveal a sharp dissociation: models that plan rearrangements over visible layouts at 53--97% accuracy and rarely violate collision constraints fall to 6--45% on occlusion and below 7% on reflections. An embodied-reasoning model reproduces the same profile. White-box analysis on Qwen3-VL-8B-Thinking localises the failure to the visual-token merger: spatial information recoverable throughout the vision encoder becomes inaccessible after token compression and only stabilises again when clean post-merger activations are patched into the language decoder.", "date": "2026-05-19", "categories": []}
{"id": "2606.19383", "title": "3D Scene Graphs: Open Challenges and Future Directions", "abstract": "3D Scene Graphs (3DSGs) have emerged as a powerful representation for spatial AI by combining geometric grounding with semantic and relational abstractions of the environment. Their expressiveness has made them relevant to a broad range of problems in robotics and computer vision, including manipulation, navigation, task planning, scene understanding, and many others. However, the field remains fragmented: different communities adopt distinct formulations, construction pipelines, and evaluation protocols, making it difficult to compare methods, identify common assumptions, and assess remaining challenges for robust real-world deployment. This survey provides a unified and critical review of 3DSGs, with particular emphasis on open challenges and future directions. We first formalize 3DSGs under a common definition and analyze the principal modeling choices that characterize existing formulations, including node and edge attributes, hierarchical structure, dynamic scene representations, and affordance-aware extensions. We then review how 3DSGs are built from raw sensory observations, discussing the most common terminologies, conventions, and techniques. Finally, we examine downstream applications and evaluation strategies, from intrinsic graph quality to task-level performance. To support the community, we also provide a dedicated website that organizes and extends the surveyed content, accessible at https://3dscenegraphs.com/.", "date": "2026-06-15", "categories": []}
{"id": "2606.19451", "title": "3D-DLP: Self-Supervised 3D Object-Centric Scene Representation Learning", "abstract": "We introduce 3D-DLP, a self-supervised object-centric representation learning model that decomposes scene-level RGB-D or voxel observations into a set of 3D latent particles. Building on the Deep Latent Particles (DLP) framework, each particle encodes disentangled attributes, including 3D keypoint position, bounding box dimensions, and appearance features, and represents a distinct entity in the scene. The model learns interpretable per-particle segmentation maps through an end-to-end self-supervised reconstruction objective. We demonstrate on both simulated and real-world datasets that the learned latent space is interpretable and controllable: by manipulating particle positions and decoding, we can generate novel scene configurations. Furthermore, we show that leveraging these compact 3D latent particles for downstream robotic manipulation improves performance over baselines that either lack explicit 3D information or rely on memory-intensive dense 3D inputs without object-centric structure. Code and videos are available at https://eubooks3003.github.io/3d-dlp.", "date": "2026-06-17", "categories": []}
{"id": "2411.07742", "title": "Efficient 3D Perception on Multi-Sweep Point Cloud with Gumbel Spatial Pruning", "abstract": "This paper studies point cloud perception within outdoor environments. Existing methods face limitations in recognizing objects located at a distance or occluded, due to the sparse nature of outdoor point clouds. In this work, we observe a significant mitigation of this problem by accumulating multiple temporally consecutive point cloud sweeps, resulting in a remarkable improvement in perception accuracy. However, the computation cost also increases, hindering previous approaches from utilizing a large number of point cloud sweeps. To tackle this challenge, we find that a considerable portion of points in the accumulated point cloud is redundant, and discarding these points has minimal impact on perception accuracy. We introduce a simple yet effective Gumbel Spatial Pruning (GSP) layer that dynamically prunes points based on a learned end-to-end sampling. The GSP layer is decoupled from other network components and thus can be seamlessly integrated into existing point cloud network architectures. Without incurring additional computational overhead, we increase the number of point cloud sweeps from 10, a common practice, to as many as 40. Consequently, there is a significant enhancement in perception performance. For instance, in nuScenes 3D object detection and BEV map segmentation tasks, our pruning strategy improves several 3D perception baseline methods.", "date": "2024-11-12", "categories": []}
{"id": "2606.15328", "title": "SGFormer++: Semantic Graph Transformer for Incremental 3D Scene Graph Generation", "abstract": "In this paper, we propose SGFormer++, a novel Semantic Graph Transformer for 3D scene graph generation (SGG), which aims to parse point cloud scenes into semantic structural graphs, where nodes denote detected object instances and edges encode their pairwise relationships, with the core challenge lying in modeling complex global scene structure. While existing graph convolutional network (GCN)-based methods suffer from over-smoothing and limited receptive fields, SGFormer++ leverages Transformer layers as its backbone to enable global message passing. Specifically, we introduce two key components tailored for 3D SGG: (1) a Graph Embedding Layer++ that efficiently integrates edge-aware global context with linear computational complexity, and (2) a Semantic Injection Layer++ that enriches visual features with linguistic priors from large language models (LLMs) and vision-language models (VLMs), boosting semantic representation without introducing extra trainable parameters. To further address the practical challenge of incremental SGG (I-SGG), where new relationship categories arrive sequentially, we equip SGFormer++ with a novel Spatial-guided Feature Adapter, which calibrates predicate features using subject-object spatial geometry to counter scale variation, and a Cascaded Binary Prediction Head that mitigates catastrophic forgetting via task-incremental classifier expansion and logit distillation. Extensive experiments on the 3DSSG benchmark demonstrate that SGFormer++ achieves state-of-the-art performance in both standard and incremental settings: it yields a significant 4.49% absolute improvement in Predicate A@1 under the incremental setting. Code and data are available at: https://github.com/Andy20178/SGFormer.", "date": "2026-06-13", "categories": []}
{"id": "2603.04976", "title": "3D-RFT: Reinforcement Fine-Tuning for Video-based 3D Scene Understanding", "abstract": "Reinforcement Learning with Verifiable Rewards ( RLVR ) has emerged as a transformative paradigm for enhancing the reasoning capabilities of Large Language Models ( LLMs), yet its potential in 3D scene understanding remains under-explored. Existing approaches largely rely on Supervised Fine-Tuning ( SFT), where the token-level cross-entropy loss acts as an indirect proxy for optimization, leading to a misalignment between training objectives and task performances. To bridge this gap, we present Reinforcement Fine-Tuning for Video-based 3D Scene Understanding (3D-RFT ), the first framework to extend RLVR to video-based 3D perception and reasoning. 3D-RFT shifts the paradigm by directly optimizing the model towards evaluation metrics. 3D-RFT first activates 3D-aware Multi-modal Large Language Models ( MLLM s) via SFT, followed by reinforcement fine-tuning using Group Relative Policy Optimization ( GRPO) with strictly verifiable reward functions. We design task-specific reward functions directly from metrics like 3D IoU and F1-Score to provide more effective signals to guide model training. Extensive experiments demonstrate that 3D-RFT-4B achieves state-of-the-art performance on various video-based 3D scene understanding tasks. Notably, 3D-RFT-4B significantly outperforms larger models (e.g., VG LLM-8B) on 3D video detection, 3D visual grounding, and spatial reasoning benchmarks. We further reveal good properties of 3D-RFT such as robust efficacy, and valuable insights into training strategies and data impact. We hope 3D-RFT can serve as a robust and promising paradigm for future development of 3D scene understanding.", "date": "2026-03-05", "categories": []}
{"id": "2606.14307", "title": "Pano3D: Unified 3D Reconstruction and Panoptic Segmentation", "abstract": "Recent advances in 3D feedforward reconstruction neural networks have achieved remarkable success in dense reconstruction from images without any camera parameters. Yet, equipping these models with robust semantic understanding remains an open problem. Here we introduce an approach that performs 3D reconstruction and 3D panoptic segmentation in a unified framework. We build on existing 3D reconstruction models and augment them with a set-based mask decoder. The approach is jointly trained with a geometric and semantic loss, which are shown to be mutually beneficial. More precisely, the features are initialized from the geometric information and then finetuned to capture jointly geometry and semantics. We demonstrate the generality of our approach by successfully applying our framework both to online and all-to-all attention reconstruction backbones. Our method achieves state-of-the-art performance in 3D panoptic segmentation across ScanNet, ScanNet200, and ScanNet++ datasets. Ablation studies show that such joint training of a unified model equips 3D feedforward reconstruction neural networks with panoptic segmentation and yields mutually beneficial improvements.", "date": "2026-06-12", "categories": []}
{"id": "2606.08655", "title": "PhysGraph: A Physics-aware 3D Scene Graph for Perception and Reasoning", "abstract": "To perform a wide range of daily tasks, robots need to construct a 3D representation that is semantically rich, physically grounded, and structured enough to support task planning and affordance prediction. However, existing approaches primarily focus on semantic retrieval, often overlooking physical and kinematic factors. Methods that attempt to model physical properties typically rely on narrow training sets or single-object modeling, limiting scalability and generalization across diverse object types. To address these challenges, we present PhysGraph, a framework that unifies symbolic reasoning with structured 3D geometry to model kinematic and physical properties in cluttered scenes. Given RGB-D observations, PhysGraph reconstructs object-centric 3D geometry and associates object instances across views. It then decomposes objects into functional parts and infers materials and articulations through visual reasoning. Evaluated on both synthetic and real-world datasets, PhysGraph achieves state-of-the-art results in semantic segmentation, multi-object mass estimation, and articulation prediction. With its simple yet effective design, PhysGraph produces physically consistent and semantically structured scene graphs, serving as a structured 3D representation for downstream tasks such as constraint-aware 3D affordance prediction and real-to-sim transfer, both of which are demonstrated in our experiments.", "date": "2026-06-07", "categories": []}
{"id": "2511.14019", "title": "RISE: Single Static Radar-based Indoor Scene Understanding", "abstract": "Robust and privacy-preserving indoor scene understanding remains a fundamental open problem. While optical sensors such as RGB and LiDAR offer high spatial fidelity, they suffer from severe occlusions and introduce privacy risks in indoor environments. In contrast, millimeter-wave (mmWave) radar preserves privacy and penetrates obstacles, but its inherently low spatial resolution makes reliable geometric reasoning difficult. We introduce RISE, the first benchmark and system for single-static-radar indoor scene understanding, jointly targeting layout reconstruction and object detection. RISE is built upon the key insight that multipath reflections-traditionally treated as noise-encode rich geometric cues. To exploit this, we propose a Bi-Angular Multipath Enhancement that explicitly models Angle-of-Arrival and Angle-of-Departure to recover secondary (ghost) reflections and reveal invisible structures. On top of these enhanced observations, a simulation-to-reality Hierarchical Diffusion framework transforms fragmented radar responses into complete layout reconstruction and object detection. Our benchmark contains 50,000 frames collected across 100 real indoor trajectories, forming the first large-scale dataset dedicated to single, static, radar-based indoor scene understanding. Extensive experiments show that RISE reduces the Chamfer Distance by 60% (down to 16 cm) compared to the state of the art in mmWave layout reconstruction, and delivers the first mmWave-based object detection, achieving 58% IoU. These results establish RISE as a new foundation for geometry-aware and privacy-preserving indoor scene understanding using a single static radar. Our website and code are available at https://rise-cvpr.github.io.", "date": "2025-11-18", "categories": []}
{"id": "2603.07704", "title": "PARSE: Part-Aware Relational Spatial Modeling", "abstract": "Inter-object relations underpin spatial intelligence, yet existing representations -- linguistic prepositions or object-level scene graphs -- are too coarse to specify which regions actually support, contain, or contact one another, leading to ambiguous and physically inconsistent layouts. To address these ambiguities, a part-level formulation is needed; therefore, we introduce PARSE, a framework that explicitly models how object parts interact to determine feasible and spatially grounded scene configurations. PARSE centers on the Part-centric Assembly Graph (PAG), which encodes geometric relations between specific object parts, and a Part-Aware Spatial Configuration Solver that converts these relations into geometric constraints to assemble collision-free, physically valid scenes. Using PARSE, we build PARSE-10K, a dataset of 10,000 3D indoor scenes constructed from real-image layout priors and a curated part-annotated shape database, each with dense contact structures and a part-level contact graph. With this structured, spatially grounded supervision, fine-tuning Qwen3-VL on PARSE-10K yields stronger object-level layout reasoning and more accurate part-level relation understanding; furthermore, leveraging PAGs as structural priors in 3D generation models leads to scenes with substantially improved physical realism and structural complexity. Together, these results show that PARSE significantly advances geometry-grounded spatial reasoning and supports the generation of physically consistent 3D scenes.", "date": "2026-03-08", "categories": []}
{"id": "2606.06721", "title": "SCOUT: Semantic scene COverage via Uncertainty-guided Traversal", "abstract": "Robots that operate over extended periods should not merely visit space; they should progressively understand it. Yet most 3D scene graph pipelines treat perception as a post-processing stage over a fixed dataset, decoupling scene representation from the decisions that determine what is observed in the first place. We present SCOUT, an online semantic exploration framework that closes this loop by coupling active traversal with probabilistic scene graph construction. Given a prior 2D occupancy map and posed RGB-D observations, SCOUT incrementally builds an uncertainty-aware 3D scene graph whose nodes maintain fused geometry and posterior beliefs over open-vocabulary object labels, while edges encode structural relations such as on, inside, belong, and next to. These beliefs are fed back to an uncertainty-guided traversal planner, which selects viewpoints by balancing expected semantic certainty gain, geometric coverage gain, and travel cost. In this way, the robot revisits ambiguous objects when additional evidence matters and expands into unseen free space when the scene remains incomplete. The resulting system treats semantic scene completeness as an operational objective rather than a passive by-product of semantic mapping, moving toward autonomous agents that can patrol, update, and reason about evolving indoor environments with minimal human intervention.", "date": "2026-06-04", "categories": []}
{"id": "2606.05975", "title": "T-FunS3D: Task-Driven Hierarchical Open-Vocabulary 3D Functionality Segmentation", "abstract": "Open-vocabulary 3D functionality segmentation enables robots to localize functional object components in 3D scenes. It is a challenging task that requires spatial understanding and task interpretation. Current open-vocabulary 3D segmentation methods primarily focus on object-level recognition, while scene-wide part segmentation methods attempt to segment the entire scene exhaustively, making them highly resource-intensive and time consuming. Balancing segmentation performance in terms of granularity, accuracy, and speed remains a challenge. As one step towards alleviating this, we introduce T-FunS3D, a task-driven hierarchical open-vocabulary 3D functionality segmentation method that provides actionable perception for robotic applications. Our method takes as input the 3D point cloud and posed RGB-D images of an indoor scene. We construct an open-vocabulary scene graph by extracting instances and their visual embeddings in the environment. Given a task description, T-FunS3D identifies the most relevant instances in the scene graph and locates their functional components leveraging a vision-language model. Experiments on the SceneFun3D dataset demonstrate that T-FunS3D is comparable to state-of-the-art in open-vocabulary 3D functionality segmentation, while achieving faster runtime and reduced memory usage.", "date": "2026-06-04", "categories": []}
{"id": "2606.06485", "title": "PAR3D: A Unified 3D-MLLM with Part-Aware Representation for Scene Understanding", "abstract": "Recent advances in 3D multimodal large language models (3D-MLLMs) have enabled unified solutions for 3D scene understanding tasks, including visual question answering, captioning, and referring segmentation. However, existing 3D-MLLMs remain largely object-centric, limiting their ability to model fine-grained part structures that are essential for embodied interaction with 3D environments. In this work, we present PAR3D, a unified part-aware 3D-MLLM framework that enables models to understand, reason about, and ground both objects and their parts in 3D scenes. To enable training and evaluation of part-aware 3D scene understanding, we introduce ScenePart, a synthetic 3D scene dataset with part-level annotations and language instructions. We further develop Part-Aware 3D Representation Learning to enrich 3D visual representations with fine-grained part-level semantics, and propose Hierarchical Segmentation Query Generation to ground part targets via hierarchical object-part queries. Extensive experiments show that our method substantially improves part-level question answering and referring segmentation, while also achieving strong performance across object-level vision-language tasks.", "date": "2026-06-04", "categories": []}
{"id": "2606.00747", "title": "SkyShield: Occupancy as a Safety Interface for Low-Altitude UAV Autonomy", "abstract": "For low-altitude Unmanned Aerial Vehicle (UAV) autonomy, 3D spatial understanding is not merely a perception objective, but the safety interface between human instructions and physical flight. In human-scale urban airspace below 20 meters, thin geometry, occlusions, vegetation, and urban clutter define whether an aerial agent can safely enter the space ahead. However, existing UAV datasets mainly provide 2D annotations or 3D boxes, while driving-oriented occupancy benchmarks assume stable ground-level sensor rigs. Both miss the defining regime of low-altitude flight: a front-facing monocular camera observing occupied and free space from a moving aerial body with frame-wise changing 6-DoF pose and camera extrinsics. To bridge this gap, we introduce SkyShield, to the best of our knowledge the first front-view monocular semantic occupancy benchmark for urban UAV flight below 20 meters. Built on CARLA, SkyShield contains 36K front-view UAV samples across diverse urban scenes and weather conditions, pairing each image with frame-wise 6-DoF UAV pose, frame-wise dynamic camera geometry, UAV states, and front-frustum semantic occupancy labels. We further propose KAR-mIoU, a UAV-centric and dynamics-aware metric that re-weights voxel-level evaluation by kinematic reachability and time-to-collision, revealing safety-critical risks hidden by conventional mIoU. To tackle this challenging new setting, we provide SkyOcc, a geometry-first monocular baseline that integrates frame-wise UAV attitude into projection, fuses temporal occupancy features, and applies safety-prior optimization to preserve sparse collision-critical structures. Together, SkyShield, KAR-mIoU, and SkyOcc establish occupancy as a safety interface for low-altitude aerial autonomy. Code and dataset will be released publicly.", "date": "2026-05-30", "categories": []}
{"id": "2512.18954", "title": "VOIC: Visible-Occluded Integrated Guidance for 3D Semantic Scene Completion", "abstract": "Camera-based 3D Semantic Scene Completion (SSC) is a critical task for autonomous driving and robotic scene understanding. It aims to infer a complete 3D volumetric representation of both semantics and geometry from a single image. Existing methods typically focus on end-to-end 2D-to-3D feature lifting and voxel completion. However, they often overlook the interference between high-confidence visible-region perception and low-confidence occluded-region reasoning caused by single-image input, which can lead to feature dilution and error propagation.\n  To address these challenges, we introduce an offline Visible Region Label Extraction (VRLE) strategy that explicitly separates and extracts voxel-level supervision for visible regions from dense 3D ground truth. This strategy purifies the supervisory space for two complementary sub-tasks: visible-region perception and occluded-region reasoning. Building on this idea, we propose the Visible-Occluded Interactive Completion Network (VOIC), a novel dual-decoder framework that explicitly decouples SSC into visible-region semantic perception and occluded-region scene completion. VOIC first constructs a base 3D voxel representation by fusing image features with depth-derived occupancy. The visible decoder focuses on generating high-fidelity geometric and semantic priors, while the occlusion decoder leverages these priors together with cross-modal interaction to perform coherent global scene reasoning.\n  Extensive experiments on the SemanticKITTI and SSCBench-KITTI360 benchmarks demonstrate that VOIC outperforms existing monocular SSC methods in both geometric completion and semantic segmentation accuracy, achieving state-of-the-art performance.", "date": "2025-12-22", "categories": []}
{"id": "2606.03314", "title": "TASE: Truncation-Aware Semantic Embeddings for 3D Scene Understanding and Editing", "abstract": "High-fidelity semantic 3D scene representations are crucial for numerous applications, including robotics, autonomous driving, and simulation. Beyond this, the ability to edit such representations enables developers to adapt these applications more easily to specific target scenarios. Current approaches provide limited support for controllable editing. We introduce TASE, a method that projects pretrained 2D semantic features into a truncation-aware embedding space to enable flexible 3D scene editing. Our method explicitly optimizes a feature space in which progressively reducing feature channels yields increasingly abstract semantic representations, while retaining more channels preserves fine-grained detail. Additionally, we improve multi-view consistency of the features using a scale- and translation-equivariance loss. The resulting truncation-aware embedding space enables text-driven edits to 3D scenes, providing explicit control over how strongly edits adhere to the original scene content and allowing more substantial modifications than prior methods. Moreover, we propose a finetuning stage for the editing diffusion model to mitigate artifacts caused by geometric changes. Experimental results demonstrate competitive performance in 3D scene editing, substantially outperforming prior methods on edits involving large geometric modifications.", "date": "2026-06-02", "categories": []}
{"id": "2606.03460", "title": "From 3D Perception to Safety Reasoning: A Graph-Based Framework for Real-Time Underground Mine Monitoring", "abstract": "Underground coal mining requires personnel and heavy equipment to operate within shared, confined, and poorly illuminated spaces where hazards such as equipment proximity violations, structural instabilities, and occluded blind spots are difficult to anticipate. Conventional monitoring systems, including fixed cameras and rule-based proximity alerts, can detect predefined events but lack the 3D scene understanding and contextual memory needed to identify complex or evolving hazards. This paper presents a continuous monitoring framework that converts colourised 3D point clouds into structured and traceable safety reasoning outputs. The framework combines 3D semantic perception, uncertainty-based anomaly detection, rule-based hazard checks, on-device LLM reasoning, and GraphRAG -based memory analysis to identify immediate hazards and interpret longer-term safety patterns. Scene and temporal graphs serve as the explicit knowledge structure, linking perception outputs across reasoning stages. To overcome the scarcity of labeled underground data, real roadway scans, controlled object placement, and high-fidelity longwall simulation were combined to generate diverse hazard scenarios, while self-supervised pretraining improved segmentation from limited annotations. The perception model achieved 92.7% accuracy at 30 FPS with low memory usage. Across 115 hazard scenarios, rule-based checks achieved 57% coverage, increasing to 76% with contextual LLM reasoning and 93% with memory-based reasoning using historical records. Qualitative results show uncertainty-derived anomaly signals support the interpretation of out-of-distribution hazards beyond predefined classes. Overall, graph-based knowledge representation combined with 3D perception and layered safety reasoning provides a practical foundation for intelligent decision support in underground mine monitoring.", "date": "2026-06-02", "categories": []}
{"id": "2606.03915", "title": "PatchScene: Patch-based Voxel Diffusion for Large-Scale Scene Completion", "abstract": "We propose PatchScene, a novel diffusion-based framework for large-scale LiDAR scene completion. Unlike existing methods that rely on global latent representations or dense voxel grids, PatchScene adopts a patch-based voxel diffusion paradigm that explicitly generates fine-grained geometry within localized 3D regions. To ensure coherent reconstruction at both spatial and temporal scales, we introduce a confidence-guided spatio-temporal fusion mechanism that integrates overlapping patches and adjacent frames in a unified generative process. Furthermore, we design an Annular-Flow diffusion strategy that leverages the radial density pattern of LiDAR scans to progressively propagate high-fidelity information from near-range to far-range regions, enabling spatially unbounded scene completion. Extensive experiments on the SemanticKITTI benchmark demonstrate that PatchScene achieves state-of-the-art performance across all standard metrics, surpassing previous approaches in both geometric accuracy and temporal consistency. Remarkably, the model trained on 20 m LiDAR ranges generalizes effectively to 50 m scenes without retraining, highlighting its strong scalability and generalization capability for real-world autonomous driving applications.", "date": "2026-06-02", "categories": []}
{"id": "2606.03994", "title": "SimuScene: Simulation-Ready Compositional 3D Scene Reconstruction from a Single Image", "abstract": "Reconstructing interactive, simulation-ready 3D scenes from a single image is a critical bottleneck for robotic manipulation. While recent single-image lifters recover plausible per-object shapes, composing them yields scenes that collapse under physical simulation due to interpenetrating, hovering, or sinking objects. Existing physics-aware methods address this strictly as a post-hoc layout correction, leaving the underlying geometric errors unresolved. To address this, we introduce SimuScene, a compositional 3D reconstruction pipeline that puts physics in the loop of shape and layout estimation. Rather than using physics merely for layout cleanup, we utilize the physics engine as a diagnostic measurement tool during the generative process itself. By diagnostically simulating reconstructed objects under gravity, we convert penetration and support failures into quantitative correction signals that drive gravity-axis stretching and amodal shape resampling. This physics-informed feedback loop mitigates accumulated reconstruction errors and produces a stable, simulation-ready compositional 3D scene. Extensive experiments demonstrate state-of-the-art performance on physical stability and geometric alignment benchmarks. We further highlight SimuScene's utility by deploying reconstructed environments in humanoid control and robot-arm manipulation tasks.", "date": "2026-06-02", "categories": []}
{"id": "2310.15676", "title": "Recent Advances in Multi-modal 3D Intelligence: A Comprehensive Survey and Evaluation", "abstract": "Multi-modal 3D Intelligence has gained considerable attention due to its wide applications in autonomous driving and world simulation, etc. Compared to conventional single-modal 3D understanding, introducing an additional modality not only elevates the richness and precision of scene interpretation but also provides a foundation for higher-level physical world interaction. This becomes especially crucial in varied and challenging environments where solely relying on 3D data might be inadequate. While there has been a surge in the development of multi-modal 3D methods over the past six years, especially those integrating multi-camera images (3D+2D) and textual descriptions (3D+language), a comprehensive and in-depth review is notably absent. In this paper, we present a systematic survey of recent progress to bridge this gap. We begin by briefly summarizing the unique challenges among various 3D multi-modal tasks. After that, we present a novel taxonomy that delivers a thorough categorization of existing methods according to modalities and tasks, exploring their respective strengths and limitations. Furthermore, comparative results of recent approaches on several benchmark datasets, together with insightful analysis, are offered. Finally, we discuss the unresolved issues and provide several potential avenues for future research.", "date": "2023-10-24", "categories": []}
{"id": "2512.14364", "title": "Unified Semantic Transformer for 3D Scene Understanding", "abstract": "Holistic 3D scene understanding involves capturing and parsing unstructured 3D environments. Due to the inherent complexity of the real world, existing models have predominantly been developed and limited to be task-specific. We introduce UNITE, a Unified Semantic Transformer for 3D scene understanding, a novel feed-forward neural network that unifies a diverse set of 3D dense semantic indoor tasks within a single model. Our model operates on unseen scenes trained in a fully end-to-end manner and only takes a couple seconds to infer the full 3D semantic geometry. Our approach is capable of directly predicting multiple dense semantic attributes, including 3D scene segmentation, instance embeddings, open-vocabulary features, and articulations, solely from RGB images. The method is trained using a combination of 2D distillation, heavily relying on self-supervision and leverages novel multi-view losses designed to ensure 3D view consistency. We demonstrate that UNITE achieves state-of-the-art performance on several different dense indoor semantic tasks and even outperforms task-specific models, in many cases, surpassing methods that operate on ground truth 3D geometry. See the project website at unite-page.github.io", "date": "2025-12-16", "categories": []}
{"id": "2602.08058", "title": "Picasso: Holistic Scene Reconstruction with Physics-Constrained Sampling", "abstract": "In the presence of occlusions and measurement noise, geometrically accurate scene reconstructions -- which fit the sensor data -- can still be physically incorrect. For instance, when estimating the poses and shapes of objects in the scene and importing the resulting estimates into a simulator, small errors might translate to implausible configurations including object interpenetration or unstable equilibrium. This makes it difficult to predict the dynamic behavior of the scene using a digital twin, an important step in simulation-based planning and control of contact-rich behaviors. In this paper, we posit that object pose and shape estimation requires reasoning holistically over the scene (instead of reasoning about each object in isolation), accounting for object interactions and physical plausibility. Towards this goal, our first contribution is Picasso, a physics-constrained reconstruction pipeline that builds multi-object scene reconstructions by considering geometry, non-penetration, and physics. Picasso relies on a fast rejection sampling method that reasons over multi-object interactions, leveraging an inferred object contact graph to guide samples. Second, we propose the Picasso dataset, a collection of 10 contact-rich real-world scenes with ground truth annotations, as well as a metric to quantify physical plausibility, which we open-source as part of our benchmark. Finally, we provide an extensive evaluation of Picasso on our newly introduced dataset and on the YCB-V dataset, and show it largely outperforms the state of the art while providing reconstructions that are both physically plausible and more aligned with human intuition.", "date": "2026-02-08", "categories": []}
{"id": "2605.20823", "title": "RelWitness: Open-Vocabulary 3D Scene Graph Generation with Visual-Geometric Relation Witnesses", "abstract": "Open-vocabulary 3D scene graph generation seeks to describe object instances and their relations with flexible natural-language predicates. The central difficulty is not only vocabulary expansion, but supervision reliability: relation annotations in 3D scene graph datasets are selective, and many valid object-pair relations are unannotated. We propose RelWitness, a framework for open-vocabulary 3D scene graph generation from posed RGB-D sequences under incomplete relation supervision. The key concept is a relation witness: a concrete visual-geometric cue that makes a relation observable in the captured scene. Support relations require contact and vertical ordering; containment requires enclosure; proximity requires metric closeness; orientation requires facing direction; and stable relations should persist across views where both objects are visible. RelWitness constructs relation witness records from RGB views, depth maps, reconstructed 3D geometry, role-sensitive text, object-prior null views, and multi-view consistency. A visual-geometric witness verifier assigns unannotated relation candidates to verified missing positives, reliable negatives, or uncertain unlabeled cases. A witness-guided positive-unlabeled objective then learns from incomplete annotations without turning every missing label into a negative. We further introduce witness-consistent decoding and an RGB-D missing-relation audit protocol. Simulated manuscript-planning experiments on 3DSSG/3RScan and ScanNet-derived open-vocabulary splits show the intended behavior: improved unseen-relation recognition, higher witness precision, lower hallucination, and reduced redundant relation phrases. All numerical results are planning values and must be replaced by reproduced measurements before submission", "date": "2026-05-20", "categories": []}
{"id": "2606.00447", "title": "GeoSAM-3D: Geodesic Prompt Propagation for Open-Vocabulary 3D Scene Segmentation from Monocular Video", "abstract": "Open-vocabulary 3D scene segmentation usually assumes RGB-D video, calibrated multi-view imagery, or a reconstructed mesh. GeoSAM-3D studies a lighter setting: a user uploads a short monocular video, clicks or names an object in one frame, and receives a propagated 3D mask over a Gaussian scene. The implementation combines frozen image and video foundation models with a monocular 3D Gaussian Splatting reconstruction and a differentiable graph-geodesic propagation kernel over Gaussian centroids. The central design choice is to propagate prompts by heat-kernel distance on the reconstructed scene graph, rather than by Euclidean nearest neighbors in 3D. This preserves continuity around curved surfaces and reduces leakage across nearby but disconnected objects. This paper describes the repository state, the mathematical kernel implemented in geosam3d.propagate, the feature head trained from Segment Anything masks, and the validation already present in the codebase. The evaluation protocol separates implementation validation, graph propagation quality, leakage control, and interactive latency.", "date": "2026-05-30", "categories": []}
{"id": "2606.01215", "title": "Distilling Neuro-Symbolic Programs into 3D Multi-modal LLMs", "abstract": "Current 3D spatial reasoning methods face a fundamental trade-off: neuro-symbolic 3D (NS3D) concept learners achieve interpretable reasoning through compositional programs but are constrained to closed-set concept vocabularies and simple programs; end-to-end 3D multi-modal LLMs (3D MLLMs) could handle complex natural language and open-vocabulary concepts but suffer from black-box reasoning without explicit spatial verification. We introduce APEIRIA, a neuro-symbolic 3D MLLM to bridge two paradigms by distilling symbolic reasoning patterns into MLLMs with natural language chain-of-thought. Our three-stage curriculum progressively builds reasoning capabilities: a) 3D perception alignment grounds object visual-geometric features to the LLM, b) CoT-SFT teaches query decomposition and stepwise verification from symbolic program traces, and c) CoT-RL extends reasoning patterns to open-set concepts and deeply nested instructions. By transferring reasoning patterns rather than concept-specific knowledge, APEIRIA preserves key NS3D virtues: transparent reasoning and modular interchangeability of planning and perception components. Evaluations on grounding, question answering, and captioning show that APEIRIA surpasses prior NS3D methods and matches state-of-the-art 3D MLLMs on 3D spatial reasoning datasets, unifying symbolic methods' systematic reasoning with MLLMs' flexibility. Code is available at https://github.com/oceanflowlab/APEIRIA.", "date": "2026-05-31", "categories": []}
{"id": "2606.01334", "title": "HOLA: Holistic Multi-Modal Alignment for Open-Set 3D Recognition", "abstract": "Open-set 3D recognition requires models that generalize to rare or unseen categories. Recent approaches address this by distilling language-vision knowledge into 3D encoders, typically relying on heavy 2D ViTs and aligning each point cloud with a single image or caption, thus anchoring representations to partial views. We propose aligning each point cloud with multiple images and textual descriptions to capture a more holistic understanding of 3D objects. To realize this idea, it is essential to design a loss function capable of jointly aligning a 3D instance with multiple matched signals, multi-view images and multiple texts, while separating positive aggregation from negative competition. We introduce such a function, termed the decoupled multi-positive contrastive loss. Our formulation enhances the loss's hardness-aware focus on challenging negatives, avoiding the \"spotlight crowding\" that occurs when many positives share the same softmax with all the negatives. Complementing this, we present a lightweight text adapter applied only to web captions, reducing the domain gap to curated annotations and enabling effective use of large-scale unsupervised text. Our model demonstrates state-of-the-art open-vocabulary performance on long-tail benchmarks, yielding substantial zero-shot improvements while sustaining high frame rates.", "date": "2026-05-31", "categories": []}
{"id": "2606.02463", "title": "MASER: Modality-Adaptive Specialist Routing for Embodied 3D Spatial Intelligence", "abstract": "In 3D environments, Embodied Agents answer spatially relevant questions through reasoning from a mixture of modalities including natural language, RGB images, point clouds, depth maps and camera poses. Existing Vision-Language models (VLMs) are fine-tuned over a single modality. This completely ignores the question semantics which may favor a different modality than the finetuned modality. To address this, we propose MASER (Modality-Adaptive SpEcialist Routing), a lightweight framework that trains five different modality adapters of a shared VLM backbone and learns a neural routing policy that selects the best adapter based on the question during inference. We encode each question with a frozen sentence transformer and pass the embedding through a small Multi-layer Perceptron (MLP) trained on oracle adapter-accuracy labels. We evaluate our methodology over the Open3D-VQA benchmark and our evaluations show that no single modality is universally optimal -- point-cloud answers are best in 51.5% of cases. MASER routes with 51.3% oracle agreement, outperforming a Random-Forest ablation (43.5%), with only a single adapter call per question.", "date": "2026-06-01", "categories": []}
{"id": "2603.09632", "title": "X-GS: An Extensible Framework for Perceiving and Thinking via 3D Gaussian Splatting", "abstract": "3D Gaussian Splatting (3DGS) has emerged as a powerful technique for novel view synthesis, subsequently extending into numerous spatial AI applications. However, most existing 3DGS methods operate in isolation, focusing on specific domains. In this paper, we introduce X-GS, an extensible framework consisting of two major components. The X-GS-Perceiver unifies a broad range of 3DGS techniques to enable real-time online SLAM with semantic distillation. The X-GS-Thinker accommodates multimodal models, enabling them to seamlessly interface with the Perceiver to complete downstream tasks. In our implementation of X-GS, the Perceiver leverages the latest vision foundation models to improve online SLAM performance and employs three key mechanisms to accelerate semantic distillation. The Thinker can be built upon both contrastive and generative vision-language models and utilizes the Perceiver's semantic Gaussian splats to unlock capabilities such as 3D visual grounding and scene captioning. Experimental results on diverse benchmarks demonstrate the efficiency and newly unlocked multimodal capabilities of the X-GS framework.", "date": "2026-03-10", "categories": []}
{"id": "2604.20395", "title": "SpaCeFormer: Fast Proposal-Free Open-Vocabulary 3D Instance Segmentation", "abstract": "Open-vocabulary 3D instance segmentation is a core capability for robotics and AR/VR, but prior methods trade one bottleneck for another: multi-stage 2D+3D pipelines aggregate foundation-model outputs at hundreds of seconds per scene, while pseudo-labeled end-to-end approaches rely on fragmented masks and external region proposals. We present SpaCeFormer, a proposal-free space-curve transformer that runs in 0.12--0.30 seconds per scene across standard benchmarks, 2--3 orders of magnitude faster than multi-stage 2D+3D pipelines. We pair it with SpaCeFormer-3M, the largest open-vocabulary 3D instance segmentation dataset (3.0M multi-view-consistent captions over 604K instances from 7.4K scenes) built through multi-view mask clustering and multi-view VLM captioning; it reaches 21$\\times$ higher mask recall than prior single-view pipelines (54.3% vs 2.5% at IoU$>$0.5). SpaCeFormer combines spatial window attention with Morton-curve serialization for spatially coherent features, and uses a RoPE-enhanced decoder to predict instance masks directly from learned queries without external proposals. On ScanNet200 we achieve 11.1 zero-shot mAP, a 2.8$\\times$ improvement over the prior best proposal-free method; on ScanNet++ and Replica, we reach 22.9 and 24.1 mAP, surpassing all prior methods including those using multi-view 2D inputs.", "date": "2026-04-22", "categories": []}
{"id": "2604.26262", "title": "Semantic Foam: Unifying Spatial and Semantic Scene Decomposition", "abstract": "Modern scene reconstruction methods, such as 3D Gaussian Splatting, deliver photo-realistic novel view synthesis at real-time speeds, yet their adoption in interactive graphics applications has been limited. A major bottleneck is the difficulty of interacting with these representations compared to traditional, human-authored 3D assets. While previous research has attempted to impose semantic decomposition on these models, significant challenges remain regarding segmentation quality and consistency. To address this, we introduce Semantic Foam, extending the recently proposed Radiant Foam representations to semantic decomposition tasks. Our approach integrates the natural spatial volumetric decomposition of Radiant Foam's Voronoi mesh with an explicit semantic feature field parameterized at the cell level. This explicit structure enables direct spatial regularization, which prevents artifacts caused by occlusion or inconsistent supervision across views - common pitfalls for other point-based representations. Experimental results show that our method achieves comparable or superior object-level segmentation performance compared to state-of-the-art methods like Gaussian Grouping and SAGA.", "date": "2026-04-29", "categories": []}
{"id": "2605.27367", "title": "SpatialBench: Is Your Spatial Foundation Model an All-Round Player?", "abstract": "While spatial foundation models have demonstrated impressive performance on standard datasets, a critical question remains: are they truly all-round players capable of generalizing robustly across diverse downstream tasks, arbitrary viewpoints, shifting scene domains, varying input densities, and specific hardware constraints? Answering this overarching question requires a holistic assessment, yet current models are mainly evaluated on specific domains for which they were specifically designed or trained. Such evaluations are intrinsically limited by narrow paradigm coverage, limited scene domains, and arbitrary frame sampling, making it fundamentally difficult to assess their true generalization capabilities. To address this gap, we present SpatialBench, a cross-paradigm, domain-diverse benchmark for spatial foundation models with deterministic sampling. SpatialBench features unprecedented scale and rigorous deterministic design, comprising 19 datasets and 546 scenes across 5 diverse spatial domains. It comprehensively evaluates 41 models across 6 paradigms on 5 task suites under 4 different input density settings. Our extensive evaluation reveals that current models are not yet all-round players, and uncovers crucial insights for future advancement. Specifically, we demonstrate that full-context attention maximizes accuracy while bounded-memory strategies unlock long-sequence scalability. Moreover, our empirical evaluations in challenging embodied and egocentric tasks demonstrate that strict domain alignment and high data quality are far more critical to performance than simple dataset scaling. Furthermore, to address the largest data gap identified in our analysis, we go beyond evaluation by introducing a large-scale dataset, DA-Next-5M, and a strong baseline model, DA-Next, pushing the boundaries of spatial representation learning.", "date": "2026-05-26", "categories": []}
{"id": "2605.29879", "title": "DGSG-Mind: Dynamic 3D Gaussian Scene Graphs for Long-Term Scene Understanding and Grounding", "abstract": "Integrating open-vocabulary semantic information into dynamic 3D scene representations is essential for long-term embodied scene understanding. However, existing methods often suffer from fragile instance association due to incomplete cross-view cues, while their limited ability to handle object-level topological changes restricts long-term robotic task execution. Moreover, current 3D scene understanding methods either rely on simple feature matching without explicit spatial reasoning or assume offline ground-truth 3D geometry. To address these challenges, we present DGSG-Mind, a hybrid instance-aware 3D Gaussian dynamic scene graph system with an embodied reasoning agent. Our system couples a probabilistic voxel grid with explicit 3D Gaussians to enable robust cross-modal instance fusion and incremental semantic mapping. It handles dynamic changes through Gaussian-based visual relocalization and localized masked refinement guided by geometric-semantic consistency. Built on the instance Gaussian map, DGSG-Mind further constructs a hierarchical scene graph and develops the 3D Gaussian Mind, which integrates structural relations, spatial-semantic information, and visually annotated RoI Gaussian renderings for multimodal reasoning. Extensive experiments show that DGSG-Mind achieves the best zero-shot 3DVG performance among methods operating on self-reconstructed maps, while also delivering strong performance in 3D open-vocabulary semantic segmentation and scene reconstruction. We further deploy DGSG-Mind on real-world robots to demonstrate its target-oriented reasoning and dynamic update capabilities. The project page of DGSG-Mind is available at https://icr-lab.github.io/DGSG-Mind", "date": "2026-05-28", "categories": []}
{"id": "2605.31466", "title": "VolFill: Single-View Amodal 3D Scene Reconstruction with Volumetric Flow Matching", "abstract": "Reconstructing the complete geometry of a scene from a single RGB image remains challenging - especially when inferring hidden structures where visual evidence is incomplete. We introduce VolFill, a generative framework that predicts the 3D structure of the complete scene rather than relying on traditional pixel-aligned regression. Our method utilizes a hybrid 3D VAE to compress sparse truncated unsigned distance function grids into a compact latent space, paired with a latent Diffusion Transformer that denoises this representation to recover the complete scene. We condition the generation on geometry foundation models, leveraging rich spatial priors for robust reasoning. Unlike existing methods limited by per-ray constraints or unstructured point-cloud queries, VolFill provides a structured representation that supports direct surface extraction and occupancy queries at scale. Extensive experiments on the SCRREAM and NRGB-D datasets demonstrate that our approach significantly outperforms current baselines, providing a robust foundation for holistic spatial understanding.", "date": "2026-05-29", "categories": []}
{"id": "2404.07977", "title": "Gaga: Group Any Gaussians via 3D-aware Memory Bank", "abstract": "We introduce Gaga, a framework that reconstructs and segments open-world 3D scenes by leveraging inconsistent 2D masks predicted by zero-shot class-agnostic segmentation models. Contrasted to prior 3D scene segmentation approaches that rely on video object tracking or contrastive learning methods, Gaga utilizes spatial information and effectively associates object masks across diverse camera poses through a novel 3D-aware memory bank. By eliminating the assumption of continuous view changes in training images, Gaga demonstrates robustness to variations in camera poses, particularly beneficial for sparsely sampled images, ensuring precise mask label consistency. Furthermore, Gaga accommodates 2D segmentation masks from diverse sources and demonstrates robust performance with different open-world zero-shot class-agnostic segmentation models, significantly enhancing its versatility. Extensive qualitative and quantitative evaluations demonstrate that Gaga performs favorably against state-of-the-art methods, emphasizing its potential for real-world applications such as 3D scene understanding and manipulation.", "date": "2024-04-11", "categories": []}
{"id": "2605.25059", "title": "VEOcc: Voxel-Centric Online Semantic Occupancy Prediction For Embodied Scene Understanding", "abstract": "Crucial for autonomous exploration, online 3D occupancy prediction and mapping incrementally constructs dense spatial representations on the fly. However, recent Gaussian-centric methods struggle with structural boundary fidelity and rely heavily on predefined scene-size priors, fundamentally limiting their operational efficiency. In this work, we present VEOcc, a voxel-centric framework formulated as a recursive perception-and-assimilation paradigm. By eliminating the need for initial scale estimation, VEOcc enables highly streamlined, open-ended map expansion. Furthermore, to robustly aggregate noisy temporal observations within the discrete voxel space, we propose a Spatio-Temporal-Aware Online Update Strategy. It integrates Cross-Temporal Logit Aggregation (TLA) for temporal consistency, Reliability-Aware Confidence Modulation (RCM) for spatial uncertainty calibration, and Confidence-Driven Incremental State Update (CSU) for robust global state assimilation. % Extensive experiments on Occ-ScanNet and EmbodiedOcc-ScanNet demonstrate that VEOcc establishes new state-of-the-art performance in both local and embodied settings, providing an accurate and efficient solution for real-world exploration. Extensive experiments on Occ-ScanNet and EmbodiedOcc-ScanNet demonstrate that VEOcc establishes new state-of-the-art performance in both local and embodied settings. Notably, zero-shot evaluations on self-collected video sequences further confirm its robust out-of-distribution generalization capability in completely unseen real-world environments. Ultimately, our framework provides an accurate and highly efficient solution for autonomous exploration. Code and supplementary visualizations are available on our project page: https://wryzju.github.io/VEOcc/.", "date": "2026-05-24", "categories": []}
{"id": "2605.29549", "title": "Learning Representations from 3D Gaussian Splats", "abstract": "3D Gaussian Splatting (3DGS) is a recent approach for scene rendering. Although primarily designed for view synthesis, its potential for scene understanding tasks remains underexplored. In this work, we conduct a comparative evaluation of various geometric deep learning architectures for the classification of 3D scenes represented using Gaussian Splatting. We benchmark point-based and graph-based models across both traditional point cloud datasets and dedicated Gaussian Splatting datasets. Scenes are embedded into latent representations, which are evaluated through end-to-end classification, linear probing, and clustering analysis. Our study provides insight into the suitability of different geometry-aware architectures and input feature configurations for learning effective 3D Gaussian Splat representations. The results highlight consistent differences between architectural families and reveal the impact of Gaussian-specific attributes on the quality of representation.", "date": "2026-05-28", "categories": []}
{"id": "2605.30111", "title": "xModel-KD: Cross-modal Knowledge Distillation for 3D Scene Perception using LiDAR", "abstract": "Point cloud segmentation is a fundamental task in 3D scene understanding. Its progress is constrained by the high cost and time required for dense 3D annotations, making labeled samples difficult to obtain. Beyond annotation scarcity, different sensing modalities face inherent limitations. 2D images provide rich texture and appearance cues, yet they lack explicit depth and geometric structure. In contrast, 3D point clouds capture accurate spatial geometry but are sparse and contain no texture information. As a result, relying on a single modality restricts the richness of learned representations and weakens generalization. Although recent multi-modal methods that combine 3D point clouds with 2D images have demonstrated strong performance in tasks such as classification and retrieval, they typically depend on large-scale labeled datasets and have not been fully exploited for data-efficient dense prediction. To address these limitations, we propose a novel cross-modal knowledge distillation framework, xModel-KD, for 3D point cloud segmentation. Our method exploits the complementary strengths of 2D texture and 3D geometry by learning unified per-point representations through cross-modal alignment. Specifically, we design a cross-modal fusion encoder trained with a contrastive objective that enforces feature consistency between corresponding 2D and 3D representations across multiple views. By integrating powerful pre-trained backbones with a targeted fusion strategy, the proposed framework effectively transfers appearance cues from images to geometry-aware point features. Experimental results show that cross-modal fusion achieves a 2% absolute improvement in mIoU over a LiDAR-only baseline, demonstrating the benefit of leveraging complementary multi-modal information for scalable and annotation-efficient 3D scene understanding.", "date": "2026-05-28", "categories": []}
{"id": "2503.04863", "title": "Manboformer: Learning Gaussian Representations via Spatial-temporal Attention Mechanism", "abstract": "Compared with voxel-based grid prediction, in the field of 3D semantic occupation prediction for autonomous driving, GaussianFormer proposed using 3D Gaussian to describe scenes with sparse 3D semantic Gaussian based on objects is another scheme with lower memory requirements. Each 3D Gaussian function represents a flexible region of interest and its semantic features, which are iteratively refined by the attention mechanism. In the experiment, it is found that the Gaussian function required by this method is larger than the query resolution of the original dense grid network, resulting in impaired performance. Therefore, we consider optimizing GaussianFormer by using unused temporal information. We learn the Spatial-Temporal Self-attention Mechanism from the previous grid-given occupation network and improve it to GaussianFormer. The experiment was conducted with the NuScenes dataset, and the experiment is currently underway.", "date": "2025-03-06", "categories": []}
{"id": "2605.28587", "title": "Deformable Gaussian Occupancy: Decoupling Rigid and Nonrigid Motion with Factorized Distillation", "abstract": "Understanding dynamic 3D environments is essential for safe autonomous driving, particularly when reasoning about human-centric, nonrigid agents. However, existing weakly supervised occupancy prediction frameworks predominantly assume rigid-body motion and rely on simple frame-to-frame offsets, limiting their ability to capture fine-grained deformations and maintain temporal coherence. To address this issue, we propose DeGO, a deformable Gaussian occupancy framework that unifies decoupled Gaussian deformation with factorized 4D foundation-model distillation. DeGO disentangles rigid and nonrigid motion, enabling each Gaussian primitive to evolve through both deformation and offset-based updates. In parallel, a factorized 4D distillation strategy transfers cross-camera and cross-frame knowledge from the VGGT foundation model, producing foundation-aligned features that enhance temporal consistency. Experiments on the Occ3D-NuScenes benchmark demonstrate that our method achieves state-of-the-art performance under weak supervision, delivering 13.5% gains on human-centric instances and 10.9% overall improvements. These results highlight the effectiveness of deformation-aware and foundation-guided occupancy modeling for dynamic scene understanding. The code is publicly available: https://github.com/vita-epfl/DeGO", "date": "2026-05-27", "categories": []}
{"id": "2605.24321", "title": "Unified 3D Scene Understanding Through Physical World Modeling", "abstract": "Understanding 3D scenes requires flexible combinations of visual reasoning tasks, including depth estimation, novel view synthesis, and object manipulation, all of which are essential for perception and interaction. Existing approaches have typically addressed these tasks in isolation, preventing them from sharing a common representation or transferring knowledge across tasks. A conceptually simpler but practically non-trivial alternative is to unify these diverse tasks into a single model, reducing different tasks from separate training objectives to merely different prompts and allowing for joint training across all datasets. In this work, we present a physical world model for unified 3D understanding and interaction (3WM), formulated as a probabilistic graphical model in which nodes represent multimodal scene elements such as RGB, optical flow, and camera pose. Diverse tasks emerge from different inference pathways through the graph: novel view synthesis from RGB and dense flow prompts, object manipulation from RGB and sparse flow prompts, and depth estimation from RGB and camera conditioning, all zero-shot without task-specific training. 3WM outperforms specialized baselines without the need for finetuning by offering precise controllability, strong geometric consistency, and robustness in real-world scenarios, achieving state-of-the-art performance on NVS and 3D object manipulation. Beyond predefined tasks, the model supports composable inference pathways, such as moving objects aside while navigating a 3D environment, enabling complex geometric reasoning. This demonstrates that a unified model can serve as a practical alternative to fragmented task-specific systems, taking a step towards a general-purpose visual world model.", "date": "2026-05-23", "categories": []}
{"id": "2605.24728", "title": "Hylos: Operability Contracts for Model-Native Spatial Intelligence", "abstract": "Foundation models can increasingly describe, reconstruct, and generate 3D objects, assemblies, scenes, and environments, but visually plausible spatial output is not yet operable 3D. A generated object or environment becomes useful to an agent only when the system can identify its entities, frames, surfaces, constraints, provenance, admissible actions, expected effects, and validation failures. This paper introduces Hylos, a systems architecture for contract-bounded spatial intelligence. Hylos maintains scene-scale operability state over objects, assemblies, assets, surface anchors, assertions, action candidates, solver jobs, shared actuator invocations, capability gaps, and effect diffs. Durable spatial changes are routed through a SpatialTransaction: a commit boundary that resolves references, checks admissibility, protects invariants, projects effects, and returns commit, review, rollback, deferral, or capability-gap outcomes. The paper is framed as a systems/position preprint with a focused artifact study rather than a broad benchmark. The study examines causal repair: a visible misalignment appears on a dependent component, while the supported repair lies upstream in the placement structure that controls it. The successful interaction traces the symptom through scene dependencies, selects a supported upstream interaction, and applies a validated change instead of directly editing visible geometry. The broader claim is that spatial AI should be evaluated not only by visual quality, but by whether generated or edited 3D can become reliable substrate for CAD, robotics, simulation, inspection, manufacturing, and interactive world authoring.", "date": "2026-05-23", "categories": []}
{"id": "2605.25326", "title": "Perceive-then-Plan: Layout-as-Policy for Monocular 3D Scene Layout Estimation", "abstract": "Building structured 3D scene layouts from a single image requires reconciling visual observations with physical and spatial constraints, a challenge that is difficult to address with direct prediction alone. In this work, we formulate monocular 3D layout estimation as a perceive-then-plan problem with vision-language models, where a Perceiver first grounds the 3D objects and then a Planner iteratively refines the scene hypothesis through actions that improve physical plausibility while preserving consistency with the input image. We propose Layout-as-Policy (LaP), which casts the planning stage as a policy learning problem: 3D layouts are represented as structured states, and refined via discrete actions such as translation, rotation, and rescaling. Starting from an observation-aligned initialization with the geometry-enhanced Perceiver, the LaP Planner is trained to produce action sequences that progressively resolve geometric inconsistencies and enforce realistic spatial relations. To enable effective learning, we combine supervised trajectory initialization with preference-based optimization, allowing the model to learn corrective behaviors without requiring explicit reward engineering. This formulation transforms layout estimation from a one-shot prediction task into an iterative refinement process, enabling better handling of global constraints and complex object interactions. Experiments demonstrate that our approach produces layouts that are more physically coherent and better aligned with visual observations, while naturally supporting downstream tasks such as scene editing and manipulation.", "date": "2026-05-25", "categories": []}
{"id": "2512.18363", "title": "Enhancing 3D Semantic Scene Completion with a Refinement Module", "abstract": "We propose ESSC-RM, a plug-and-play Enhancing framework for Semantic Scene Completion with a Refinement Module, which can be seamlessly integrated into existing SSC models. ESSC-RM operates in two phases: a baseline SSC network first produces a coarse voxel prediction, which is subsequently refined by a 3D U-Net-based Prediction Noise-Aware Module (PNAM) and Voxel-level Local Geometry Module (VLGM) under multiscale supervision. Experiments on SemanticKITTI show that ESSC-RM consistently improves semantic prediction performance. When integrated into CGFormer and MonoScene, the mean IoU increases from 16.87% to 17.27% and from 11.08% to 11.51%, respectively. These results demonstrate that ESSC-RM serves as a general refinement framework applicable to a wide range of SSC models.", "date": "2025-12-20", "categories": []}
{"id": "2605.06088", "title": "OpenGaFF: Open-Vocabulary Gaussian Feature Field with Codebook Attention", "abstract": "Understanding open-vocabulary 3D scenes with Gaussian-based representations remains challenging due to fragmented and spatially inconsistent semantic predictions across multi-view observations. In this paper, we present OpenGaFF, a novel framework for open-vocabulary 3D scene understanding built upon 3D Gaussian Splatting. At the core of our method is a Gaussian Feature Field that models semantics as a continuous function of Gaussian geometry and appearance. By explicitly conditioning semantic predictions on geometric structure, this formulation strengthens the coupling between geometry and semantics, leading to improved spatial coherence across similar structures in 3D space. To further enforce object-level semantic consistency, we introduce a structured codebook that serves as a set of shared semantic primitives. Furthermore, a codebook-guided attention mechanism is proposed to retrieve language features via similarity matching between query embeddings and learned codebook entries, enabling robust open-vocabulary reasoning while reducing intra-object feature variance. Extensive experiments on standard 2D and 3D open-vocabulary benchmarks demonstrate that our method consistently outperforms prior approaches, achieving improved segmentation quality, stronger 3D semantic consistency and a semantically interpretable codebook that provides insight into the learned representation.", "date": "2026-05-07", "categories": []}
{"id": "2605.23287", "title": "LangFlash: Feed-forward 3D Language Gaussian Splatting from Sparse Unposed Images", "abstract": "We present LangFlash, a feed-forward framework for 3D Language Gaussian Splatting that reconstructs 3D scenes parameterized by Gaussian primitives enriched with language-aligned semantic features from sparse unposed multi-view images. Unlike optimization-based 3D methods, LangFlash directly predicts the geometry and semantics in a single forward pass, enabling low-latency 3D reconstruction and language-consistent scene understanding. To support large-scale training, we enriched the RealEstate10k dataset with coherent and dense semantic information for 3D semantic supervision. Furthermore, we propose a sparse semantic encoding scheme that combines a global semantic dictionary with locally varying per-primitive weights, preserving high-level linguistic information, while reducing representation complexity. Experimental results show that LangFlash achieves superior novel view synthesis and semantic consistency compared with previous methods. This study establishes a new paradigm for pose-free, language-grounded 3D scene reconstruction, advancing generalizable 3D vision and multimodal scene understanding. Demo is available at https://liylo.github.io/langflash.github.io/.", "date": "2026-05-22", "categories": []}
{"id": "2605.23771", "title": "PhotoFlow: Agentic 3D Virtual Photography Missions", "abstract": "Virtual photography asks an agent to enter a prepared 3D scene with no preselected camera pose or reference image, infer a suitable shot from scene information and a language intent, choose executable camera parameters, and render the final photograph. Recent progress in vision-language models makes this kind of spatial agent increasingly plausible, but the task stresses two capabilities that remain hard to evaluate together: complex 3D spatial understanding and abstract aesthetic judgment. We introduce PhotoFlow, a Director-Reviewer-Reflector agent for closed-loop camera search. The Director builds a soft photographic blueprint and proposes diverse candidate cameras; the Reviewer combines rule checks, visual critique, and pairwise incumbent selection; and the Reflector converts failures into region memory, dead-zone suppression, and high-explore relocation. We also introduce VPhotoBench, a benchmark of 47 open-license Blender scenes and 141 language-conditioned photography missions spanning subject placement, relational composition, and atmosphere/style. On held-out experiments, PhotoFlow achieves the strongest external quality-alignment composite and success rate among one-shot prediction, single-chain reflection, anchor-bank selection, and random search under a six-round rendering budget. To our knowledge, this is the first work to make language-conditioned virtual photography in arbitrary Blender scenes an executable agent task, and our results show that an LLM-centered spatial agent can already produce strong photographs in a setting designed to challenge both 3D reasoning and aesthetic choice.", "date": "2026-05-22", "categories": []}
{"id": "2605.23888", "title": "GenRecon: Bridging Generative Priors for Multi-View 3D Scene Reconstruction", "abstract": "We introduce a new approach to high-fidelity 3D scene reconstruction from multi-view RGB images that tightly couples reconstruction with a strong generative 3D prior. We cast scene reconstruction as conditional 3D generation over a set of spatially-localized, overlapping chunks that together tile the scene, scaling generation to large scene extents. Crucially, we inherit the fidelity and completeness of state-of-the-art generative shape models -- we use Trellis.2 as an example -- which we generalize to the scene level. To this end, we propose a projection-based conditioning mechanism that lifts posed multi-view image features into a coherent 3D representation aligned with the generative model, independent of view ordering and spatially anchored to the scene, yielding high-fidelity, multi-view consistent generated geometry. This enables lifting the strong object-level prior of Trellis.2 to multi-view, scene-scale generation, producing faithful, editable PBR mesh reconstructions of indoor environments. As a result, we obtain high-fidelity results that outperform cutting-edge reconstruction methods by 16%.", "date": "2026-05-22", "categories": []}
{"id": "2605.02098", "title": "From Spherical to Gaussian: A Comparative Analysis of Point Cloud Cropping Strategies in Large-Scale 3D Environments", "abstract": "Large-scale 3D point clouds can consist of hundreds of millions of points. Even after downsampling, these point clouds are too large for modern 3D neural networks. In order to develop a semantic understanding of the scene, the point clouds are divided into smaller subclouds that can be processed. Typically, this division is done using spherical crops, resulting in a loss of surrounding geometric context. To address this issue, we propose alternative methods that produce subclouds with larger crop sizes while maintaining a similar number of points. Specifically, we compare exponential, Gaussian, and linear cropping methods with the spherical method. We evaluated three 3D deep learning model architectures using multiple indoor and outdoor environment datasets. Our results demonstrate that altering the cropping strategy can enhance model performance, especially for large-scale outdoor scenes, yielding new state-of-the-art results. Code is available at https://github.com/mvg-inatech/point_cloud_cropping", "date": "2026-05-03", "categories": []}
{"id": "2605.20737", "title": "Resolving Long-Tail Ambiguity in Unsupervised 3D Point Cloud Segmentation with Language Priors", "abstract": "Existing approaches for unsupervised 3D point cloud segmentation predominantly rely on a purely visual similarity-based learning-by-clustering paradigm, which suffers from a fundamental limitation: long-tail ambiguity. In such a paradigm, features of minor classes are consistently absorbed by dominant clusters, leading to severely imbalanced predictions. To address this issue, we propose LangTail, a language-guided hierarchical learning framework that leverages the balanced world knowledge encoded in language models to mitigate long-tail ambiguity in unsupervised 3D segmentation. The key idea is to establish multi-level associations between language-derived semantic priors and visually underrepresented minor classes, thereby compensating for the biased attention of purely visual clustering toward dominant classes. Specifically, LangTail first constructs an entity-level semantic prior from language models, capturing balanced and fine-grained world knowledge across categories. These priors are injected into a hierarchical clustering framework via contrastive alignment. This guides multi-granularity semantic structure formation and prevents minor classes from being absorbed by dominant clusters, yielding more discriminative representations for underrepresented categories. Extensive experiments on ScanNet-v2, S3DIS, and nuScenes demonstrate that LangTail consistently outperforms existing methods by significant margins, \\ie, +13.5, +12.9, and +8.9 mIoU, respectively. These results demonstrate the effectiveness of language priors in improving the representation of minority classes in 3D point clouds. The code will be released at: https://github.com/Whisky0129/langtail_official.", "date": "2026-05-20", "categories": []}
{"id": "2605.20044", "title": "OP2GS: Object-Aware 3D Gaussian Splatting with Dual-Opacity Primitives", "abstract": "3D Gaussian Splatting (3DGS) provides an explicit and efficient scene representation, but its primitives lack inherent object-level identity, hindering downstream tasks such as open-vocabulary scene understanding. Existing methods typically address this by either distilling high-dimensional feature embeddings into Gaussians or by lifting 2D mask labels into 3D via heuristic refinement. However, feature-based approaches incur heavy storage and decoding overhead, while lifting-based pipelines remain vulnerable to label contamination: Gaussians necessary for appearance reconstruction often receive incorrect object labels during 2D-to-3D projection. We propose OP2GS, an object-aware Gaussian representation that augments each primitive with an explicit instance identity and a dedicated instance opacity $\\sigma^{*}$ for object-mask rendering. The original opacity $\\sigma$ remains responsible for visual reconstruction, while $\\sigma^{*}$ models whether a Gaussian should contribute to a particular object mask. This dual-opacity formulation decouples visual existence from instance occupancy: mislabeled Gaussians can remain available for image rendering while becoming transparent in the object-mask branch. To learn this representation, we introduce a random object loss that optimizes the 1D instance occupancy field using the standard transmittance-based visibility of 3DGS. Semantic descriptors are then attached at the object level through multi-view aggregation, eliminating per-Gaussian feature storage. Compared with feature-training approaches, OP2GS achieves competitive open-vocabulary performance while significantly reducing computational overhead. Compared with training-free pipelines, it leverages physically consistent occupancy learning to resolve visibility ambiguities.", "date": "2026-05-19", "categories": []}
{"id": "2602.22667", "title": "Monocular Open Vocabulary Occupancy Prediction for Indoor Scenes", "abstract": "Open-vocabulary 3D occupancy is vital for embodied agents, which need to understand complex indoor environments where semantic categories are abundant and evolve beyond fixed taxonomies. While recent work has explored open-vocabulary occupancy in outdoor driving scenarios, such methods transfer poorly indoors, where geometry is denser, layouts are more intricate, and semantics are far more fine-grained. To address these challenges, we adopt a geometry-only supervision paradigm that uses only binary occupancy labels (occupied vs free). Our framework builds upon 3D Language-Embedded Gaussians, which serve as a unified intermediate representation coupling fine-grained 3D geometry with a language-aligned semantic embedding. On the geometry side, we find that existing Gaussian-to-Occupancy operators fail to converge under such weak supervision, and we introduce an opacity-aware, Poisson-based approach that stabilizes volumetric aggregation. On the semantic side, direct alignment between rendered features and open-vocabulary segmentation features suffers from feature mixing; we therefore propose a Progressive Temperature Decay schedule that gradually sharpens opacities during splatting, strengthening Gaussian-language alignment. On Occ-ScanNet, our framework achieves 59.50 IoU and 21.05 mIoU in the open-vocabulary setting, surpassing all existing occupancy methods in IoU and outperforming prior open-vocabulary approaches by a large margin in mIoU. Code will be released at https://github.com/JuIvyy/LegoOcc.", "date": "2026-02-26", "categories": []}
{"id": "2604.02060", "title": "CompassAD: Intent-Driven 3D Affordance Grounding in Functionally Competing Objects", "abstract": "When told to \"cut the cake,\" a robot must choose the knife over nearby scissors, despite both objects affording the same cutting function. In real-world scenes, multiple objects may share identical affordances, yet only one is appropriate under the given task context. We call such cases confusing pairs. However, existing 3D affordance methods largely sidestep this challenge by evaluating isolated single objects, often with explicit category names provided in the query. We formalize Intent-Driven Confusable Affordance Grounding, a new 3D affordance setting that requires predicting a per-point affordance mask on the correct object within a multi-object point cloud, conditioned on implicit natural language intent. To study this problem, we construct CompassAD, the first benchmark centered on implicit intent in confusing multi-object compositions. It comprises 30 confusing object pairs spanning 16 affordance types, 6,422 compositions, and 88K+ query-answer pairs. Furthermore, we propose CompassNet, a framework that incorporates two dedicated modules tailored to this task. Instance-bounded Cross Injection (ICI) constrains language-geometry alignment within object boundaries to prevent cross-object semantic leakage. Bi-level Contrastive Refinement (BCR) enforces discrimination at both geometric-group and point levels, sharpening distinctions between target and confusable surfaces. Extensive experiments demonstrate state-of-the-art results on both seen and unseen queries, and deployment on a robotic manipulator confirms effective transfer to real-world grasping in confusing multi-object compositions.", "date": "2026-04-02", "categories": []}
{"id": "2605.16832", "title": "Coarse Semantic Injection for LLM-Conditioned Structured Indoor Prediction", "abstract": "Large language models (LLMs) have recently been used as structured decoders for indoor understanding from 3D point-token inputs. However, point cloud encoders often under-represent thin structural elements such as doors and windows after voxelization and sparse pooling, and may miss individual furniture instances in cluttered scenes. We propose an interface-preserving semantic augmentation for LLM-conditioned structured decoding. The key idea is to associate semantic evidence with the point-cloud representation, reduce it to a coarse four-group code (furniture, walls, openings, and others), and encode it as an RGBB point interface: red for furniture, green for walls, blue for openings, and black for others, where RGBB denotes four semantic color states represented in three RGB channels rather than an additional fourth channel. This semantic color code is appended to the original raw point attributes before tokenization, so geometry and semantics share the same sparse tokenization path while the downstream language model decoder and output serialization remain unchanged. We further introduce a lightweight routed semantic shift module, with an auxiliary head used only for training-time ratio/budget regularization and analysis, to strengthen semantic cues after sparse pooling. The overall pipeline can use RGB-derived semantic evidence. Under these controlled semantic-source settings, the reported metrics improve across Structured3D, the SpatialLM dataset, and ARKitScenes, especially for opening localization and per-instance furniture detection in cluttered scenes. Ablations clarify the roles of semantic source, color coding, token fusion, and shift injection, while also showing that color/entropy effects remain nontrivial.", "date": "2026-05-16", "categories": []}
{"id": "2605.17197", "title": "OPTNet: Ordering Point Transformer Network for Post-disaster 3D Semantic Segmentation", "abstract": "Post-disaster damage assessment requires rapid and accurate semantic segmentation of 3D point clouds to identify critical infrastructure such as damaged buildings and roads. Early Point Transformers (e.g., PTv1, PTv2) relied on computationally expensive neighbor searching (k-NN) and Farthest Point Sampling (FPS). To improve efficiency, recent architectures like Point Transformer V3 (PTv3) adopted static serialization methods, such as Hilbert curves or Z-order, to organize unstructured points for window-based attention. However, these fixed orderings are not optimal for capturing the complex geometry of disaster scenes. In this paper, we propose OPTNet (Ordering Point Transformer Network), which introduces a learnable Point Sorter module. OPTNet utilizes a self-supervised ordering loss to dynamically predict an optimal permutation that maximizes the locality of the attention mechanism. We evaluate our method on the 3DAeroRelief dataset, significantly outperforming state-of-the-art baselines.", "date": "2026-05-16", "categories": []}
{"id": "2605.17478", "title": "Mamba-VGGT: Persistent Long-Sequence Video Geometry Grounded Transformer via External Sliding Window Mamba Memory", "abstract": "Visual Geometry Grounded Transformers (VGGT) have set new benchmarks in high-fidelity 3D scene reconstruction. However, as the sequence length increases, these models suffer from catastrophic geometric forgetting and accumulation drift, primarily due to the quadratic complexity of global attention which necessitates truncated temporal windows. To overcome the resulting geometric drift, we present Mamba-VGGT, an enhanced VGGT framework capable of persistent long-range reasoning. Our key contribution is a Sliding Window Mamba (SWM) memory module that maintains an explicit external memory token across temporal windows. This module leverages selective state-space modeling to distill and propagate global geometric priors, effectively bypassing the memory constraints of traditional transformers. To integrate these long-term temporal cues without disrupting the highly optimized spatial features of the pre-trained VGGT, we propose a Zero-Init Spatial Memory Injector. Utilizing zero-convolutional layers, this injector adaptively fuses persistent memory into the patch token stream, ensuring structural stability and seamless feature alignment. Extensive experiments demonstrate that our approach significantly outperforms existing VGGT-based methods in maintaining spatial consistency and reducing trajectory accumulation errors. Our work provides a scalable, linear-complexity solution for geometry-grounded world modeling in extensive 3D environments.", "date": "2026-05-17", "categories": []}
{"id": "2605.18184", "title": "Fixed External Cameras as Common Prior Maps for Active 3D Scene Graph Generation", "abstract": "Commonly available prior information, such as BIM models, floor plans, and remote sensing images, can provide valuable geometric and semantic context for autonomous robotic systems. In this paper, we treat observations from fixed external RGB cameras as Common Prior Maps (CPMs): wide-field views of the environment that initialize a semantic and geometric scene prior before any robot motion begins. We present an RGB-only framework for active, incremental 3D scene graph (3DSG) generation that seamlessly fuses observations from both onboard robot cameras and fixed external cameras within a single hardware-agnostic pipeline. By relying solely on RGB observations processed by a feed-forward 3D reconstruction model, the system treats all cameras - onboard or external - identically, requiring no hardware modifications. A graph-based active semantic exploration framework then directly leverages the partial scene graph to guide the robot toward regions of high semantic uncertainty, progressively completing and refining the prior. Experiments demonstrate that bootstrapping the scene graph with even a single external camera increases initial object recall by up to +79%, and that the richer context of the prior significantly improves the efficiency of subsequent active exploration.", "date": "2026-05-18", "categories": []}
{"id": "2605.18197", "title": "RGB-only Active 3D Scene Graph Generation for Indoor Mobile Robots", "abstract": "Current approaches to 3D scene graph generation rely on dedicated depth sensors, such as LiDAR or RGB-D cameras, for metric 3D reconstruction. This limits deployment to specialized robotic platforms and excludes settings where only RGB cameras are available, such as fixed external infrastructure. Existing pipelines also typically operate on passively collected observation trajectories, rather than selecting viewpoints based on the partially built scene representation, and therefore fail to effectively exploit the semantic and spatial information encoded within the graph during exploration. This paper presents a fully visual framework for the active, incremental construction of 3D scene graphs from RGB input only, addressing both limitations. The proposed approach unifies perception and planning around a shared structured representation that captures object semantics, 3D geometry, relational context, and information from multiple viewpoints. Because the framework is hardware-agnostic and relies only on RGB observations, it can incorporate inputs from both onboard robot cameras and fixed external cameras within the same representation. Experiments on the Replica dataset show that the RGB-only pipeline achieves F1-score parity with baselines using ground-truth depth. Active exploration experiments on ReplicaCAD further show that semantic-driven viewpoint selection detects more than twice as many objects as a geometric frontier-based baseline under the same exploration budget. Finally, the external-camera setting demonstrates that complementary RGB views can effectively bootstrap the scene graph and improve contextual understanding at no additional exploration cost.", "date": "2026-05-18", "categories": []}
{"id": "2605.15708", "title": "3D Segmentation Using Viewpoint-Dependent Spatial Relationships", "abstract": "Recent advances in 3D datasets and multimodal models have greatly improved natural language 3D scene understanding. However, most 3D referring segmentation methods do not explicitly represent the observer viewpoint, making spatial relations such as \"left,\" \"right,\" \"front,\" and \"behind\" ambiguous and difficult to evaluate. We introduce a viewpoint-aware 3D referring segmentation dataset containing 220k benchmark samples, and scalable to tens of millions of viewpoint-conditioned samples through dense viewpoint sampling. In this dataset, target objects can only be identified through observer-centric spatial relations, making viewpoint-conditioned grounding necessary. We construct the benchmark by leveraging camera poses to automatically annotate observer-centric relations (left/right, front/behind) together with viewpoint-independent relations (above/under). Using this benchmark, we evaluate several existing 3D large multimodal models in a zero-shot setting and find that current models struggle with viewpoint-dependent spatial instructions. We further study how explicit viewpoint information can be incorporated into 3D large multimodal models. We introduce a viewpoint representation that encodes camera poses and conditions the model on the observation viewpoint, improving segmentation accuracy on viewpoint-dependent relations and increasing mIoU from 0.30 to 0.47 compared to a model without viewpoint conditioning. The dataset, code, and trained models will be made publicly available upon acceptance.", "date": "2026-05-15", "categories": []}
{"id": "2605.15753", "title": "Hierarchical and Holistic Open-Vocabulary Functional 3D Scene Graphs for Indoor Spaces", "abstract": "Functional 3D scene graphs offer a versatile and flexible representation for 3D scene understanding and robotic manipulation, defined by object nodes, interactive elements, and functional relationship edges. However, their potential remains underexplored due to the limited coverage of existing benchmarks and the overly straightforward design of previous pipelines, which primarily focus on large-scale furniture but lack of hierarchical structures. Therefore, in this work, we extend the benchmark coverage by introducing dense tabletop objects and explicit multi-level functional relationships. This expansion introduces critical challenges involving small-scale, dense, and similar instances, with lack of visual anchoring in relational reasoning, instance confusion during cross-frame fusion, and attribution uncertainty under dynamic viewpoints. To address these issues, we propose an open-vocabulary pipeline based on 2D visual grounding and 3D graph optimization. Specifically, we anchor fine-grained functional edges from 2D visual evidence, and associate nodes across frames in 3D using multiple cues. Furthermore, edge association is formulated as temporal graph optimization, integrating evidence accumulation, entropy regularization, and temporal smoothing to robustly determine the functional connections of each node. Finally, global hierarchy shaping is performed to recover the hierarchical graph structure. Extensive experiments demonstrate that the proposed method can reliably infer functional 3D scene graphs in challenging real-world scenes, thereby further unlocking their potential for practical applications.", "date": "2026-05-15", "categories": []}
{"id": "2605.16065", "title": "Robust Prior-Guided Segmentation for Editable 3D Gaussian Splatting", "abstract": "3D Gaussian Splatting (3D-GS) enables real-time 3D scene reconstruction but lacks robust segmentation for editing tasks such as object removal, extraction, and recoloring. Existing approaches that lift 2D segmentations to the 3D domain suffer from view inconsistencies and coarse masks. In this paper, we propose a novel framework that leverages the Segment Anything Model High Quality (SAM-HQ) to generate accurate 2D masks, addressing the limitations of the standard SAM in boundary fidelity and fine-structure preservation. To achieve robust 3D segmentation of any target object in a given scene, we introduce a prior-guided label reassignment method that assigns labels to 3D Gaussians by enforcing multiview consistency with learned priors. Our approach achieves state-of-the-art segmentation accuracy and enables interactive, real-time object editing while maintaining high visual fidelity. Qualitative results demonstrate superior boundary preservation and practical utility in Virtual Reality (VR) and robotics, advancing 3D scene editing.", "date": "2026-05-15", "categories": []}
{"id": "2605.14393", "title": "Analogical Trajectory Transfer", "abstract": "We study analogical trajectory transfer, where the goal is to translate motion trajectories in one 3D environment to a semantically analogous location in another. Such a capacity would enable machines to perform analogical spatial reasoning, with applications in AR/VR co-presence, content creation, and robotics. However, even semantically similar scenes can still differ substantially in object placement, scale, and layout, so naively matching semantics leads to collisions or geometric distortions. Furthermore, finding where each trajectory point should transfer to has a large search space, as the mapping must preserve semantics and functionality without tearing the trajectory apart or causing collisions. Our key insight is to decompose the problem into spatially segregated subproblems and merge their solutions to produce semantically consistent and spatially coherent transfers. Specifically, we partition scenes into object-centric clusters and estimate cross-scene mappings via hierarchical smooth map prediction, using 3D foundation model features that encode contextual information from object and open-space arrangements. We then combinatorially assemble the per-cluster maps into an initial transfer and refine the result to remove collisions and distortions, yielding a spatially coherent trajectory. Our method does not require training, attains a fast runtime around 0.6 seconds, and outperforms baselines based on LLMs, VLMs, and scene graph matching. We further showcase applications in virtual co-presence, multi-trajectory transfer, camera transfer, and human-to-robot motion transfer, which indicates the broad applicability of our work to AR/VR and robotics.", "date": "2026-05-14", "categories": []}
{"id": "2605.14399", "title": "SceneForge: Structured World Supervision from 3D Interventions", "abstract": "Many multimodal learning tasks require supervision that remains consistent across edits, viewpoints, and scene-level interventions. However, such supervision is difficult to obtain from observation-level datasets, which do not expose the underlying scene state or how changes propagate through it. We present SceneForge, an intervention-driven framework that generates structured supervision from editable 3D world states. SceneForge represents each scene as a persistent world with semantic, geometric, and physical dependencies. By applying explicit interventions (e.g., object removal or camera variation) and propagating their effects through scene dependencies, SceneForge renders supervision that remains consistent with object structure and scene-level effects. This produces aligned outputs including counterfactual observations, multi-view observations, and effect-aware signals such as shadows and reflections, all derived from a shared world state rather than post hoc image-space processing. We instantiate SceneForge using Infinigen and Blender to construct a licensing-clean indoor supervision resource with a large number of counterfactual pairs and aligned annotations from over 2K scenes, covering both diverse single-view and registered multi-view settings. Under matched training budgets, incorporating SceneForge supervision improves both object removal and scene removal performance across multiple benchmarks in both quantitative and qualitative evaluation. These results indicate that modeling supervision as structured state transitions in editable worlds provides a practical and scalable foundation for intervention-consistent multimodal learning.", "date": "2026-05-14", "categories": []}
{"id": "2605.15195", "title": "VGGT-$\\Omega$", "abstract": "Recent feed-forward reconstruction models, such as VGGT, have proven competitive with traditional optimization-based reconstructors while also providing geometry-aware features useful for other tasks. Here, we show that the quality of these models scales predictably with model and data size. We do so by introducing VGGT-$\\Omega$, which substantially improves reconstruction accuracy, efficiency, and capabilities for both static and dynamic scenes. To enable training this model at an unprecedented scale, we introduce architectural changes that improve training efficiency, a high-quality data annotation pipeline that supports dynamic scenes, and a self-supervised learning protocol. We simplify VGGT's architecture by using a single dense prediction head with multi-task supervision and removing the expensive high-resolution convolutional layers. We also use registers to aggregate scene information into a compact representation and introduce register attention, which restricts inter-frame information exchange to these registers, in part replacing global attention. In this way, during training, VGGT-$\\Omega$ uses only about 30% of the GPU memory of its predecessor, allowing us to train with 15x more supervised data than prior work and to leverage vast amounts of unlabeled video data. VGGT-$\\Omega$ achieves strong results for reconstruction of static and dynamic scenes across multiple benchmarks, for example, improving over the previous best camera estimation accuracy on Sintel by 77%. We also show that the learned registers can improve vision-language-action models and support alignment with language, suggesting that reconstruction can be a powerful and scalable proxy task for spatial understanding. Project Page: http://vggt-omega.github.io/", "date": "2026-05-14", "categories": []}
{"id": "2605.04506", "title": "Ilov3Splat: Instance-Level Open-Vocabulary 3D Scene Understanding in Gaussian Splatting", "abstract": "We introduce Ilov3Splat, a novel framework for instance-level open-vocabulary 3D scene understanding built on 3D Gaussian Splatting (3D-GS). Most prior work depends on 2D rendering-based matching or point-level semantic association, which undermines cross-view consistency, lacks coherent instance-level reasoning, and limits precision in downstream 3D tasks. To address these limitations, our method jointly optimizes scene geometry and semantic representations by augmenting Gaussian splats with view-consistent feature fields. Specifically, we leverage multi-resolution hash embedding to efficiently encode language-aligned CLIP features, enabling dense and coherent language grounding in 3D space. We further train an instance feature field using contrastive loss over SAM masks, supporting fine-grained object distinction across views. At inference time, CLIP-encoded queries are matched against the learned features, followed by two-stage 3D clustering to retrieve relevant Gaussian groups. This enables our framework to identify arbitrary objects in 3D scenes based on natural language descriptions, without requiring category supervision or manual annotations. Experiments on standard benchmarks demonstrate that Ilov3Splat outperforms prior open-vocabulary 3D-GS methods in both object selection and instance segmentation, offering a flexible and accurate solution for language-driven 3D scene understanding. Project page: https://csiro-robotics.github.io/Ilov3Splat.", "date": "2026-05-06", "categories": []}
{"id": "2605.08293", "title": "Distill, Diffuse, and Semanticize (DDS): Annotation-Free 3D Scene Understanding Based on Multi-Granularity Distillation and Graph-Diffusion-Based Segmentation", "abstract": "3D semantic scene understanding is essential for digital twins, autonomous driving, smart agriculture, and embodied perception, yet dense point-wise annotation for point clouds remains expensive and difficult to scale. Existing annotation-free methods often face a trade-off between semantic recognition and structural efficiency: open-vocabulary and foundation-model-driven methods provide strong semantic priors, but often come with substantial computational costs, while structure-oriented methods based on superpoints, clustering, and graph reasoning are lightweight but often produce category-agnostic regions. We propose DDS, a resource-efficient structure-oriented framework for region-consistent and semanticized annotation-free 3D scene understanding. DDS preserves the lightweight superpoint-based organization paradigm while incorporating visual semantic cues from projected features and segmentation-derived masks. It first performs multi-granularity distillation to guide the 3D backbone at the point, mask-prototype, and inter-prototype levels, then applies graph diffusion over superpoints to propagate semantic information directly in 3D, producing coherent region representations without costly spectral decomposition or dense open-vocabulary 3D feature fields. Finally, DDS uses segmentation-cluster association to assign interpretable semantic names to category-agnostic 3D clusters. Experiments on real-world datasets show that DDS achieves the best performance among representative structure-oriented annotation-free baselines, improving oAcc, mAcc, and mIoU by up to 5.9%, 8.1%, and 2.4%, respectively. These results demonstrate that DDS improves region consistency and lightweight semantic recognition, providing a scalable and interpretable solution for annotation-free 3D scene understanding.", "date": "2026-05-08", "categories": []}
{"id": "2605.13741", "title": "LEXI-SG: Monocular 3D Scene Graph Mapping with Room-Guided Feed-Forward Reconstruction", "abstract": "Scene graphs are becoming a standard representation for robot navigation, providing hierarchical geometric and semantic scene understanding. However, most scene graph mapping methods rely on depth cameras or LiDAR sensors. In this work, we present LEXI-SG, the first dense monocular visual mapping system for open-vocabulary 3D scene graphs using only RGB camera input. Our approach exploits the semantic priors of open-vocabulary foundation models to partition the scene into rooms, deferring feed-forward reconstruction to when each room is fully observed -- enabling scalable dense mapping without sliding-window scale inconsistencies. We propose a room-based factor graph formulation to globally align room reconstructions while preserving local map consistency and naturally imposing the semantic scene graph hierarchy. Within each room, we further support open-vocabulary object segmentation and tracking. We validate LEXI-SG on indoor scenes from the Habitat-Matterport 3D and self-collected egocentric office sequences. We evaluate its performance against existing feed-forward SLAM methods, as well as established scene graphs baselines. We demonstrate improved trajectory estimation and dense reconstruction, as well as, competitive performance in open-vocabulary segmentation. LEXI-SG shows that accurate, scalable, open-vocabulary 3D scene graphs can be achieved from monocular RGB alone. Our project page and office sequences are available here: https://ori-drs.github.io/lexisg-web/.", "date": "2026-05-13", "categories": []}
{"id": "2605.11967", "title": "H2G: Hierarchy-Aware Hyperbolic Grouping for 3D Scenes", "abstract": "Hierarchical 3D grouping aims to recover scene groups across multiple granularities, from fine object parts to complete objects, without relying on semantic labels or a fixed vocabulary. The main challenge is to transform 2D foundation-model cues into coherent hierarchy supervision and embed that hierarchy in a 3D representation. We propose H2G, a hyperbolic affinity field for hierarchical 3D grouping. Our method derives semantically organized tree supervision by interpreting foundation-model affinities through Dasgupta's objective for similarity-based hierarchical clustering. This supervision is distilled into a single Lorentz hyperbolic feature field, whose geometry is well suited for tree-like branching structures. A hierarchy-aware objective aligns the field with fine-level assignments, coarse object structure, compact feature clusters, and LCA (Lowest Common Ancestor) ordering. This formulation represents multiple grouping levels in one feature space, enabling semantic hierarchical grouping grounded in 2D foundation-model knowledge.", "date": "2026-05-12", "categories": []}
{"id": "2603.16869", "title": "SegviGen: Repurposing 3D Generative Model for Part Segmentation", "abstract": "We introduce SegviGen, a framework that repurposes native 3D generative models for 3D part segmentation. Existing pipelines either lift strong 2D priors into 3D via distillation or multi-view mask aggregation, often suffering from cross-view inconsistency and blurred boundaries, or explore native 3D discriminative segmentation, which typically requires large-scale annotated 3D data and substantial training resources. In contrast, SegviGen leverages the structured priors encoded in pretrained 3D generative model to induce segmentation through distinctive part colorization, establishing a novel and efficient framework for part segmentation. Specifically, SegviGen encodes a 3D asset and predicts part-indicative colors on active voxels of a geometry-aligned reconstruction. It supports interactive part segmentation, full segmentation, and full segmentation with 2D guidance in a unified framework. Extensive experiments show that SegviGen improves over the prior state of the art by 40% on interactive part segmentation and by 15% on full segmentation, while using only 0.32% of the labeled training data. It demonstrates that pretrained 3D generative priors transfer effectively to 3D part segmentation, enabling strong performance with limited supervision. See our project page at https://fenghora.github.io/SegviGen-Page/.", "date": "2026-03-17", "categories": []}
{"id": "2605.08808", "title": "Curvature-Aware Captioning:Leveraging Geodesic Attention for 3D Scene Understanding", "abstract": "Accurate 3D scene description is fundamental to robotic navigation and augmented reality, yet current dense captioning methods face significant limitations in processing sparse point cloud data. %\nExisting approaches that apply Euclidean embedding spaces struggle to simultaneously preserve fine-grained local geometric details and model exponentially growing global semantic hierarchies, leading to either inaccurate localization or disjointed, shallow scene descriptions. %\nIn this work, we propose a novel \\textbf{\\textsc{Curvature-Aware Captioning}} framework, integrating novel non-Euclidean geodesic attention mechanisms, to resolve the localization-contextualization conflict. %\nSpecifically, self-attention within Oblique space enforces dimensional homogeneity while establishing long-range dependencies. Bidirectional geodesic cross-attention within Lorentz space models hierarchical semantic relationships across scene instances, enabling simultaneous precision in object localization and coherence in scene descriptions. %\nTheoretical analysis confirms that the curvature complementarity between the Oblique manifold and Lorentz hyperboloid resolves the Euclidean-hyperbolic conflict, ensuring feature stability via isotropic optimization while preserving inherent hierarchical relationships. Extensive experiments on ScanRefer and Nr3D benchmarks demonstrate state-of-the-art performance, with significant gains in both localization accuracy and descriptive richness.", "date": "2026-05-09", "categories": []}
{"id": "2605.09218", "title": "Flame3D: Zero-shot Compositional Reasoning of 3D Scenes with Agentic Language Models", "abstract": "3D scene understanding spans reasoning about free space, object grounding, hypothetical object insertions, complex geometric relationships, and integrating all of these with external tools and data sources. Existing 3D understanding methods typically rely on large-scale 3D-language training or focus on object grounding and simple spatial relationships. We argue that the broad generalization that motivates 3D-language training can be achieved at inference time, without 3D-specific training. We propose Flame3D, a training-free framework that represents scenes as editable visual-textual 3D memories and exposes them to an off-the-shelf MLLM through composable spatial tools. Flame3D also lets the agent synthesize custom spatial programs at inference time, enabling open-ended reasoning over layouts, empty space, and objects not yet present in the scene. External data and corrections can be added to the memory without retraining. In addition to showing competitive performance to finetuned 3D-LMM methods on ScanQA, we study multi-hop 3D reasoning capabilities of Flame3D by evaluating it on a curated compositional spatial-reasoning benchmark, Compose3D. We find that fixed tools fall short and that the agent's ability to synthesize spatial operations at inference time is essential. These results invite the question: should future progress in 3D scene understanding focus on richer scene memories and expressive compositional abstractions?", "date": "2026-05-09", "categories": []}
{"id": "2605.09719", "title": "Distilling 3D Spatial Reasoning into a Lightweight Vision-Language Model with CoT", "abstract": "Large-scale 3D vision-language models (VLMs) like LLaVA-3D offer strong spatial reasoning but are difficult to deploy due to high computational costs. We propose a knowledge distillation framework that transfers spatial reasoning from a 7B teacher to a 2.29B student model. Our approach achieves 8.7x lower inference latency and a 3x reduction in model size while retaining 54-72% of the teacher's performance. The framework utilizes VGGT as the vision encoder and a multi-task distillation pipeline with uncertainty-aware loss weighting. To improve reasoning without chain-of-thought (CoT) data, we introduce \"Hidden CoT\": learnable latent tokens that serve as an internal scratchpad before answer generation. This is the first use of latent scratchpad reasoning in distilled 3D VLMs. The student model jointly performs spatial description, depth estimation, and object detection. Experiments on ScanNet and 3D-FRONT show strong spatial understanding, reaching 68-72% accuracy on proximity and contact tasks. Our framework enables efficient 3D scene QA on resource-constrained platforms.", "date": "2026-05-10", "categories": []}
{"id": "2605.10484", "title": "OpenSGA: Efficient 3D Scene Graph Alignment in the Open World", "abstract": "Scene graph alignment establishes object correspondences between two 3D scene graphs constructed from partially overlapping observations. This enables efficient scene understanding and object-level relocalization when a robot revisits a place, as well as global map fusion across multiple agents. Such capabilities are essential for robots that require long-term memory for long-horizon tasks involving interactions with the environment. Existing approaches mainly focus on subscan-to-subscan (S2S) alignment and depend heavily on geometric point-cloud features, leaving frame-to-scan (F2S) alignment and open-set vision-language features underexplored. In addition, existing datasets for scene graph alignment remain small-scale with limited object diversity, constraining systematic training and evaluation. We present a unified and efficient scene graph alignment framework that predicts object correspondences by fusing vision-language, textual, and geometric features with spatial context. The framework comprises modules such as a distance-gated spatial attention encoder, a minimum-cost-flow-based allocator, and a global scene embedding generator to achieve accurate alignment even under large coordinate discrepancies. We further introduce ScanNet-SG, a large-scale dataset generated via an automated annotation pipeline with over 700k samples, covering 509 object categories from ScanNet labels and over 3k categories from GPT-4o-based tagging. Experiments show that our method achieves the best overall performance on both F2S and S2S tasks, substantially outperforming existing scene graph alignment methods. Our code and dataset are released at: https://autonomousrobots.nl/paper_websites/opensga.", "date": "2026-05-11", "categories": []}
{"id": "2605.10586", "title": "CausalGS: Learning Physical Causality of 3D Dynamic Scenes with Gaussian Representations", "abstract": "Learning a physical model from video data that can comprehend physical laws and predict the future trajectories of objects is a formidable challenge in artificial intelligence. Prior approaches either leverage various Partial Differential Equations (PDEs) as soft constraints in the form of PINN losses, or integrate physics simulators into neural networks; however, they often rely on strong priors or high-quality geometry reconstruction. In this paper, we propose CausalGS, a framework that learns the causal dynamics of complex dynamic 3D scenes solely from multi-view videos, while dispensing with the reliance on explicit priors. At its core is an inverse physics inference module that decouples the complex dynamics problem from the video into the joint inference of two factors: the initial velocity field representing the scene's kinematics, and the intrinsic material properties governing its dynamics. This inferred physical information is then utilized within a differentiable physics simulator to guide the learning process in a physics-regularized manner. Extensive experiments demonstrate that CausalGS surpasses the state-of-the-art on the highly challenging task of long-term future frame extrapolation, while also exhibiting advanced performance in novel view interpolation. Crucially, our work shows that, without any human annotation, the model is able to learn the complex interactions between multiple physical properties and understand the causal relationships driving the scene's dynamic evolution, solely from visual observations.", "date": "2026-05-11", "categories": []}
{"id": "2605.07356", "title": "UniD-Shift: Towards Unified Semantic Segmentation via Interpretable Share-Private Multimodal Decomposition", "abstract": "Semantic segmentation of large-scale 3D point clouds is crucial for applications such as autonomous driving and urban digital twins. However, the sparse sampling pattern of LiDAR and the view-dependent geometric distortion in image observations complicate cross-modal alignment and hinder stable fusion. Inspired by the fact that 2D images captured by cameras are representations of the 3D world, we recognize that the features learned from 2D and 3D segmentation share some common semantics, while other aspects remain modality-specific. This insight motivates a unified multimodal framework for joint 2D-3D semantic segmentation. We combine a SAM-based vision encoder with a SPTNet-based geometric encoder to extract complementary semantic and geometric representations. The resulting features from both modalities are explicitly decomposed into shared and private subspaces, where the shared components summarize semantic factors common to both domains, and the private components preserve properties that are unique to each modality. A lightweight attention-based fusion module aggregates the shared features into a consistent cross-modal representation, and a regularized training objective ensures both semantic alignment and subspace independence. Experiments on the SemanticKITTI and nuScenes benchmarks demonstrate consistent improvements in segmentation accuracy over representative multimodal baselines, accompanied by competitive computational efficiency. Cross-domain evaluation on nuScenes USA-Singapore shows stable performance under distribution shifts, demonstrating strong generalization. The implementation code is publicly available at: https://github.com/shuaizhang69/UniD-Shift.", "date": "2026-05-08", "categories": []}
{"id": "2511.16567", "title": "POMA-3D: The Point Map Way to 3D Scene Understanding", "abstract": "In this paper, we introduce POMA-3D, the first self-supervised 3D representation model learned from point maps. Point maps encode explicit 3D coordinates on a structured 2D grid, preserving global 3D geometry while remaining compatible with the input format of 2D foundation models. To transfer rich 2D priors into POMA-3D, a view-to-scene alignment strategy is designed. Moreover, as point maps are view-dependent with respect to a canonical space, we introduce POMA-JEPA, a joint embedding-predictive architecture that enforces geometrically consistent point map features across multiple views. Additionally, we introduce ScenePoint, a point map dataset constructed from 6.5K room-level RGB-D scenes and 1M 2D image scenes to facilitate large-scale POMA-3D pretraining. Experiments show that POMA-3D serves as a strong backbone for both specialist and generalist 3D understanding. It benefits diverse tasks, including 3D question answering, embodied navigation, scene retrieval, and embodied localization, all achieved using only geometric inputs (i.e., 3D coordinates). Overall, our POMA-3D explores a point map way to 3D scene understanding, addressing the scarcity of pretrained priors and limited data in 3D representation learning. Project Page: https://matchlab-imperial.github.io/poma3d/", "date": "2025-11-20", "categories": []}
{"id": "2605.04475", "title": "Information Coordination as a Bridge: A Neuro-Symbolic Architecture for Reliable Autonomous Driving Scene Understanding", "abstract": "Reliable autonomous driving requires scene understanding that is semantically consistent across heterogeneous sensors and verifiable at the reasoning stage. However, many recent LLM-driven driving systems attach the language model as a post-processor and force it to reason over redundant or conflicting perception outputs, which can amplify hallucinated entities and unsafe conclusions. This paper proposes InfoCoordiBridge, a BEV-centric neuro-symbolic architecture that inserts an explicit coordination bridge between perception and language reasoning. InfoCoordiBridge comprises (i) a unified multi-agent perception layer that outputs typed structured facts together with modality-focused synopses, (ii) an ICA module that aligns and fuses multi-source outputs into a single SceneSummary, and (iii) an SSRE module that performs SceneSummary-grounded reasoning with verification. Experiments on nuScenes and Waymo show that ICA preserves competitive 3D detection accuracy while substantially improving fusion consistency, reducing redundancy to below 1% and achieving about 98% attribute agreement. On NuScenes-QA and a template-aligned Waymo-QA benchmark, SSRE improves factual grounding and reduces hallucinated entity mentions compared with representative VLM and agentic baselines. Overall, by coordinating multi-sensor outputs into a single conflict-aware SceneSummary before prompting, InfoCoordiBridge prevents redundant and cross-modally inconsistent perception evidence from propagating into high-level reasoning.", "date": "2026-05-06", "categories": []}
{"id": "2512.17817", "title": "Chorus: Multi-Teacher Pretraining for Holistic 3D Gaussian Scene Encoding", "abstract": "While 3DGS has emerged as a high-fidelity scene representation, encoding rich, general-purpose features directly from its primitives remains under-explored. We address this gap by introducing Chorus, a multi-teacher pretraining framework that learns a holistic feed-forward 3D Gaussian Splatting (3DGS) scene encoder by distilling complementary signals from 2D foundation models. Chorus employs a shared 3D encoder and teacher-specific projectors to learn from language-aligned, generalist, and object-aware teachers, encouraging a shared embedding space that captures signals from high-level semantics to fine-grained structure. We evaluate Chorus on a wide range of tasks: open-vocabulary semantic and instance segmentation, linear and decoder probing, data-efficient supervision, as well as LLM-based Q&A. Besides 3DGS, we also test Chorus on several benchmarks that only support point clouds by pretraining a variant using only Gaussian centers, colors, and estimated normals. Surprisingly, this encoder shows strong transfer and outperforms the point-cloud baseline while using 39.9 times fewer training scenes. Finally, we propose a render-and-distill adaptation that facilitates out-of-domain finetuning.", "date": "2025-12-19", "categories": []}
{"id": "2503.04496", "title": "Learning to Place Objects with Programs and Iterative Self Training", "abstract": "In this work we study indoor scene object placement. Given a 3D indoor scene and an object, the task is to predict placement locations within the scene. Empirical observations of data-driven approaches to the problem show their tendency to miss placement modes. We introduce a system which helps to address this flaw. We design a Domain Specific Language (DSL) that specifies object relational constraints. Upon execution, programs from our language predict possible placements from a partial scene and object. We design a generative model which writes these programs automatically. Available 3D scene datasets do not contain programs to train on, and naively extracted programs only predict the original placement location of scene objects. Training on these programs results in subpar performance so we introduce a new program bootstrapping algorithm that improves our system's performance compared to the naive approach. To quantify our qualitative observations, we introduce a new evaluation procedure which captures how well a system models per-object location distributions. We ask human annotators to label all the possible places an object can go in a scene and compare this set against locations produced by the system in question. Our system produces per-object location distributions more consistent with human annotators than those produced by existing data-driven approaches and a zero-shot approach using an LLM. While other systems degrade in performance when training data is sparse, our system does not degrade to the same degree.", "date": "2025-03-06", "categories": []}
{"id": "2603.27437", "title": "SpatialStack: Layered Geometry-Language Fusion for 3D VLM Spatial Reasoning", "abstract": "Large vision-language models (VLMs) still struggle with reliable 3D spatial reasoning, a core capability for embodied and physical AI systems. This limitation arises from their inability to capture fine-grained 3D geometry and spatial relationships. While recent efforts have introduced multi-view geometry transformers into VLMs, they typically fuse only the deep-layer features from vision and geometry encoders, discarding rich hierarchical signals and creating a fundamental bottleneck for spatial understanding. To overcome this, we propose SpatialStack, a general hierarchical fusion framework that progressively aligns vision, geometry, and language representations across the model hierarchy. Moving beyond conventional late-stage vision-geometry fusion, SpatialStack stacks and synchronizes multi-level geometric features with the language backbone, enabling the model to capture both local geometric precision and global contextual semantics. Building upon this framework, we develop VLM-SpatialStack, a model that achieves state-of-the-art performance on multiple 3D spatial reasoning benchmarks. Extensive experiments and ablations demonstrate that our multi-level fusion strategy consistently enhances 3D understanding and generalizes robustly across diverse spatial reasoning tasks, establishing SpatialStack as an effective and extensible design paradigm for vision-language-geometry integration in next-generation multimodal physical AI systems.", "date": "2026-03-28", "categories": []}
{"id": "2605.01365", "title": "VoxAfford: Multi-Scale Voxel-Token Fusion for Open-Vocabulary 3D Affordance Detection", "abstract": "Open-vocabulary 3D affordance detection requires localizing interaction regions on point clouds given novel affordance descriptions. Recent methods extend multimodal large language models (MLLMs) with special output tokens that are decoded into segmentation masks. However, these tokens are produced through autoregressive generation, which models sequential dependencies rather than spatial neighborhood relations, leaving them semantically rich but spatially impoverished for 3D localization. We propose Voxel-enhanced Affordance detection (VoxAfford), which bypasses this bottleneck by injecting multi-scale geometric features from a frozen pre-trained 3D VQVAE encoder into the output tokens after generation. Each output token uses its affordance semantics as a query to retrieve relevant geometric patterns from its paired voxel scale via cross-attention, with a learned compatibility gate controlling the injection strength. The enhanced tokens are then aggregated into a spatially-aware affordance prompt through semantic-conditioned attention and propagated alongside per-point features to generate the final mask. Experiments on open-vocabulary affordance detection tasks show that VoxAfford achieves state-of-the-art performance with approximately an 8% improvement in mIoU, and real robot experiments confirm zero-shot transfer to novel objects.", "date": "2026-05-02", "categories": []}
{"id": "2605.01498", "title": "Towards Visual Query Localization in the 3D World", "abstract": "Visual query localization (VQL) aims to predict the spatio-temporal response of the most recent occurrence in a sequence given a query. Currently, most research focuses on visual query localization in 2D videos, while its counterpart in 3D space has received little attention. In this paper, we make the first attempt to address visual query localization in the 3D world by introducing a novel benchmark, dubbed 3DVQL. Specifically, 3DVQL contains 2,002 sequences with around 170,000 frames and 6.4K response track segments from 38 object categories. Each sequence in 3DVQL is provided with multiple modalities, including point clouds, RGB images, and depth images, to support flexible research. To ensure high-quality annotations, each sequence is manually annotated with multiple rounds of verification and refinement. To the best of our knowledge, 3DVQL is the first benchmark for 3D multimodal visual query localization. To facilitate comparison in subsequent research, we implement a series of representative 3D multimodal VQL baselines using point clouds and RGB images. The experimental results show that existing methods exhibit significant performance variations across different fusion modules. To encourage future research, we propose a lift-and-attention fusion algorithm named LaF, which significantly outperforms existing baseline models. Our benchmark and model will be publicly released at https://github.com/wuhengliangliang/3DVQL.", "date": "2026-05-02", "categories": []}
{"id": "2604.27106", "title": "Reconstruction by Generation: 3D Multi-Object Scene Reconstruction from Sparse Observations", "abstract": "Accurately reconstructing complex full multi-object scenes from sparse observations remains a core challenge in computer vision and a key step toward scalable and reliable simulation for robotics. In this work, we introduce RecGen, a generative framework for probabilistic joint estimation of object and part shapes, as well as their pose under occlusion and partial visibility from one or multiple RGB-D images. By leveraging compositional synthetic scene generation and strong 3D shape priors, RecGen generalizes across diverse object types and real-world environments. RecGen achieves state-of-the-art performance on complex, heavily occluded datasets, robustly handling severe occlusions, symmetric objects, object parts, and intricate geometry and texture. Despite using nearly 80% fewer training meshes than the previous state of the art SAM3D, RecGen outperforms it by 30.1% in geometric shape quality, 9.1% in texture reconstruction, and 33.9% in pose estimation.", "date": "2026-04-29", "categories": []}
{"id": "2604.28115", "title": "FreeOcc: Training-Free Embodied Open-Vocabulary Occupancy Prediction", "abstract": "Existing learning-based occupancy prediction methods rely on large-scale 3D annotations and generalize poorly across environments. We present FreeOcc, a training-free framework for open-vocabulary occupancy prediction from monocular or RGB-D sequences. Unlike prior approaches that require voxel-level supervision and ground-truth camera poses, FreeOcc operates without 3D annotations, pose ground truth, or any learning stage. FreeOcc incrementally builds a globally consistent occupancy map via a four-layer pipeline: a SLAM backbone estimates poses and sparse geometry; a geometrically consistent Gaussian update constructs dense 3D Gaussian maps; open-vocabulary semantics from off-the-shelf vision-language models are associated with Gaussian primitives; and a probabilistic Gaussian-to-occupancy projection produces dense voxel occupancy. Despite being entirely training-free and pose-agnostic, FreeOcc achieves over $2\\times$ improvements in IoU and mIoU on EmbodiedOcc-ScanNet compared to prior self-supervised methods. We further introduce ReplicaOcc, a benchmark for indoor open-vocabulary occupancy prediction, and show that FreeOcc transfers zero-shot to novel environments, substantially outperforming both supervised and self-supervised baselines. Project page: https://the-masses.github.io/freeocc-web/.", "date": "2026-04-30", "categories": []}
{"id": "2604.26067", "title": "RADIO-ViPE: Online Tightly Coupled Multi-Modal Fusion for Open-Vocabulary Semantic SLAM in Dynamic Environments", "abstract": "We present RADIO-ViPE (Reduce All Domains Into One -- Video Pose Engine), an online semantic SLAM system that enables geometry-aware open-vocabulary grounding, associating arbitrary natural language queries with localized 3D regions and objects in dynamic environments. Unlike existing approaches that require calibrated, posed RGB-D input, RADIO-ViPE operates directly on raw monocular RGB video streams, requiring no prior camera intrinsics, depth sensors, or pose initialization. The system tightly couples multi-modal embeddings -- spanning vision and language -- derived from agglomerative foundation models (e.g., RADIO) with geometric scene information. This coupling takes place in initialization, optimization and factor graph connections to improve the consistency of the map from multiple modalities. The optimization is wrapped within adaptive robust kernels, designed to handle both actively moving objects and agent-displaced scene elements (e.g., furniture rearranged during ego-centric session). Experiments demonstrate that RADIO-ViPE achieves state-of-the-art results on the dynamic TUM-RGBD benchmark while maintaining competitive performance against offline open-vocabulary methods that rely on calibrated data and static scene assumptions. RADIO-ViPE bridges a critical gap in real-world deployment, enabling robust open-vocabulary semantic grounding for autonomous robotics and unconstrained in-the-wild video streams. Project page: https://be2rlab.github.io/radio_vipe", "date": "2026-04-28", "categories": []}
{"id": "2604.26261", "title": "Multiple Consistent 2D-3D Mappings for Robust Zero-Shot 3D Visual Grounding", "abstract": "Zero-shot 3D Visual Grounding (3DVG) is a critical capability for open-world embodied AI. However, existing methods are fundamentally bottlenecked by the poor quality of open-vocabulary 3D proposals, suffering from inaccurate categories and imprecise geometries, as well as the spatial redundancy of exhaustive multi-view reasoning. To address these challenges, we propose MCM-VG, a novel framework that achieves robust zero-shot 3DVG by explicitly establishing Multiple Consistent 2D-3D Mappings. Instead of passively relying on noisy 3D segments, MCM-VG enforces 2D-3D consistency across three fundamental dimensions to achieve precise target localization and reliable reasoning. First, a Semantic Alignment module corrects category mismatches via LLM-driven query parsing and coarse-to-fine 2D-3D matching. Second, an Instance Rectification module leverages VLM-guided 2D segmentations to reconstruct missing targets, back-projecting these reliable visual priors to establish accurate 3D geometries. Finally, to eliminate spatial redundancy, a Viewpoint Distillation module clusters 3D camera directions to extract optimal frames. By pairing these optimal RGB frames with Bird's Eye View maps into concise visual prompt sets, we formulate the final target disambiguation as a multiple-choice reasoning task for Vision-Language Models.\n  Extensive evaluations on ScanRefer and Nr3D benchmarks demonstrate that MCM-VG sets a new state-of-the-art for zero-shot 3D visual grounding. Remarkably, it achieves 62.0\\% and 53.6\\% in Acc@0.25 and Acc@0.5 on ScanRefer, outperforming previous baselines by substantial margins of 6.4\\% and 4.0\\%.", "date": "2026-04-29", "categories": []}
{"id": "2506.05199", "title": "DEGround: An Effective Baseline for Ego-centric 3D Visual Grounding with a Homogeneous Framework", "abstract": "A core task in embodied intelligence is ego-centric 3D visual grounding. Existing methods typically adopt two-stage, heterogeneous pipelines that pair a detector with a separate grounding model. Incompatible decoders and box heads hinder the transfer of object-level priors, and the split training causes redundant re-optimization. To overcome these limitations, we present DEGround, a straight, elegant, and effective framework that centers on object-level sharing over detection and grounding. It employs a set of queries that serves as the common object representation for both detection and grounding, which is decoded by a shared transformer and bounding box head. Building on this homogeneous framework, we further introduce two task-specific plug-in modules to enhance fine-grained instruction grounding. The Regional Activation Grounding module improves spatial-textual alignment by highlighting instruction-relevant regions, while the Query-wise Modulation module applies sentence-conditioned affine modulation to generate instruction-aware queries at initialization. Extensive experiments demonstrate that DEGround achieves the best performance on multiple benchmarks. Remarkably, it significantly outperforms previous methods by 7.52% at overall precision on the EmbodiedScan dataset.", "date": "2025-06-05", "categories": []}
{"id": "2512.04021", "title": "C3G: Learning Compact 3D Representations with 2K Gaussians", "abstract": "Reconstructing and understanding 3D scenes from unposed sparse views in a feed-forward manner remains as a challenging task in 3D computer vision. Recent approaches use per-pixel 3D Gaussian Splatting for reconstruction, followed by a 2D-to-3D feature lifting stage for scene understanding. However, they generate excessive redundant Gaussians, causing high memory overhead and sub-optimal multi-view feature aggregation, leading to degraded novel view synthesis and scene understanding performance. We propose C3G, a novel feed-forward framework that estimates compact 3D Gaussians only at essential spatial locations, minimizing redundancy while enabling effective feature lifting. We introduce learnable tokens that aggregate multi-view features through self-attention to guide Gaussian generation, ensuring each Gaussian integrates relevant visual features across views. We then exploit the learned attention patterns for Gaussian decoding to efficiently lift features. Extensive experiments on pose-free novel view synthesis, 3D open-vocabulary segmentation, and view-invariant feature aggregation demonstrate our approach's effectiveness. Results show that a compact yet geometrically meaningful representation is sufficient for high-quality scene reconstruction and understanding, achieving superior memory efficiency and feature fidelity compared to existing methods.", "date": "2025-12-03", "categories": []}
{"id": "2602.05330", "title": "MTPano: Multi-Task Panoramic Scene Understanding via Label-Free Integration of Dense Prediction Priors", "abstract": "Comprehensive panoramic scene understanding is critical for immersive applications, yet it remains challenging due to the scarcity of high-resolution, multi-task annotations. While perspective foundation models have achieved success through data scaling, directly adapting them to the panoramic domain often fails due to severe geometric distortions and coordinate system discrepancies. Furthermore, the underlying relations between diverse dense prediction tasks in spherical spaces are underexplored. To address these challenges, we propose MTPano, a robust multi-task panoramic foundation model established by a label-free training pipeline. First, to circumvent data scarcity, we leverage powerful perspective dense priors. We project panoramic images into perspective patches to generate accurate, domain-gap-free pseudo-labels using off-the-shelf foundation models, which are then re-projected to serve as patch-wise supervision. Second, to tackle the interference between task types, we categorize tasks into rotation-invariant (e.g., depth, segmentation) and rotation-variant (e.g., surface normals) groups. We introduce the Panoramic Dual BridgeNet, which disentangles these feature streams via geometry-aware modulation layers that inject absolute position and ray direction priors. To handle the distortion from equirectangular projections (ERP), we incorporate ERP token mixers followed by a dual-branch BridgeNet for interactions with gradient truncation, facilitating beneficial cross-task information sharing while blocking conflicting gradients from incompatible task attributes. Additionally, we introduce auxiliary tasks to fertilize the cross-task learning process. Extensive experiments demonstrate that MTPano achieves state-of-the-art performance on multiple benchmarks and delivers competitive results against task-specific panoramic specialist foundation models.", "date": "2026-02-05", "categories": []}
{"id": "2603.27507", "title": "Chat-Scene++: Exploiting Context-Rich Object Identification for 3D LLM", "abstract": "Recent advancements in multi-modal large language models (MLLMs) have shown strong potential for 3D scene understanding. However, existing methods struggle with fine-grained object grounding and contextual reasoning, limiting their ability to interpret and interact with complex 3D environments. In this paper, we present Chat-Scene++, an MLLM framework that represents 3D scenes as context-rich object sequences. By structuring scenes as sequences of objects with contextual semantics, Chat-Scene++ enables object-centric representation and interaction. It decomposes a 3D scene into object representations paired with identifier tokens, allowing LLMs to follow instructions across diverse 3D vision-language tasks. To capture inter-object relationships and global semantics, Chat-Scene++ extracts context-rich object features using large-scale pre-trained 3D scene-level and 2D image-level encoders, unlike the isolated per-object features in Chat-Scene. Its flexible object-centric design also supports grounded chain-of-thought (G-CoT) reasoning, enabling the model to distinguish objects at both category and spatial levels during multi-step inference. Without the need for additional task-specific heads or fine-tuning, Chat-Scene++ achieves state-of-the-art performance on five major 3D vision-language benchmarks: ScanRefer, Multi3DRefer, Scan2Cap, ScanQA, and SQA3D. These results highlight its effectiveness in scene comprehension, object grounding, and spatial reasoning. Additionally, without reconstructing 3D worlds through computationally expensive processes, we demonstrate its applicability to real-world scenarios using only 2D inputs.", "date": "2026-03-29", "categories": []}
{"id": "2604.05621", "title": "FunRec: Reconstructing Functional 3D Scenes from Egocentric Interaction Videos", "abstract": "We present FunRec, a method for reconstructing functional 3D digital twins of indoor scenes directly from egocentric RGB-D interaction videos. Unlike existing methods on articulated reconstruction, which rely on controlled setups, multi-state captures, or CAD priors, FunRec operates directly on in-the-wild human interaction sequences to recover interactable 3D scenes. It automatically discovers articulated parts, estimates their kinematic parameters, tracks their 3D motion, and reconstructs static and moving geometry in canonical space, yielding simulation-compatible meshes. Across new real and simulated benchmarks, FunRec surpasses prior work by a large margin, achieving up to +50 mIoU improvement in part segmentation, 5-10 times lower articulation and pose errors, and significantly higher reconstruction accuracy. We further demonstrate applications on URDF/USD export for simulation, hand-guided affordance mapping and robot-scene interaction.", "date": "2026-04-07", "categories": []}
{"id": "2604.01907", "title": "Lifting Unlabeled Internet-level Data for 3D Scene Understanding", "abstract": "Annotated 3D scene data is scarce and expensive to acquire, while abundant unlabeled videos are readily available on the internet. In this paper, we demonstrate that carefully designed data engines can leverage web-curated, unlabeled videos to automatically generate training data, to facilitate end-to-end models in 3D scene understanding alongside human-annotated datasets. We identify and analyze bottlenecks in automated data generation, revealing critical factors that determine the efficiency and effectiveness of learning from unlabeled data. To validate our approach across different perception granularities, we evaluate on three tasks spanning low-level perception, i.e., 3D object detection and instance segmentation, to high-evel reasoning, i.e., 3D spatial Visual Question Answering (VQA) and Vision-Lanugage Navigation (VLN). Models trained on our generated data demonstrate strong zero-shot performance and show further improvement after finetuning. This demonstrates the viability of leveraging readily available web data as a path toward more capable scene understanding systems.", "date": "2026-04-02", "categories": []}
{"id": "2604.15770", "title": "PLAF: Pixel-wise Language-Aligned Feature Extraction for Efficient 3D Scene Understanding", "abstract": "Accurate open-vocabulary 3D scene understanding requires semantic representations that are both language-aligned and spatially precise at the pixel level, while remaining scalable when lifted to 3D space. However, existing representations struggle to jointly satisfy these requirements, and densely propagating pixel-wise semantics to 3D often results in substantial redundancy, leading to inefficient storage and querying in large-scale scenes. To address these challenges, we present \\emph{PLAF}, a Pixel-wise Language-Aligned Feature extraction framework that enables dense and accurate semantic alignment in 2D without sacrificing open-vocabulary expressiveness. Building upon this representation, we further design an efficient semantic storage and querying scheme that significantly reduces redundancy across both 2D and 3D domains. Experimental results show that \\emph{PLAF} provides a strong semantic foundation for accurate and efficient open-vocabulary 3D scene understanding. The codes are publicly available at https://github.com/RockWenJJ/PLAF.", "date": "2026-04-17", "categories": []}
{"id": "2604.21160", "title": "Reinforcing 3D Understanding in Point-VLMs via Geometric Reward Credit Assignment", "abstract": "Point-Vision-Language Models promise to empower embodied agents with executable spatial reasoning, yet they frequently succumb to geometric hallucination where predicted 3D structures contradict the observed 2D reality. We identify a key cause of this failure not as a representation bottleneck but as a structural misalignment in reinforcement learning, where sparse geometric tokens are drowned out by noisy and broadcasted sequence-level rewards. To resolve this causal dilution, we propose Geometric Reward Credit Assignment, a framework that disentangles holistic supervision into field-specific signals and routes them exclusively to their responsible token spans. This mechanism transforms vague feedback into precise gradient updates and effectively turns generic policy optimization into targeted structural alignment. Furthermore, we internalize physical constraints via a Reprojection-Consistency term which serves as a cross-modal verifier to penalize physically impossible geometries. Validated on a calibrated benchmark derived from ShapeNetCore, our approach bridges the reliability gap by boosting 3D KPA from 0.64 to 0.93, increasing 3D bounding box intersection over union to 0.686, and raising reprojection consistency scores to 0.852. Crucially, these gains are achieved while maintaining robust 2D localization performance, marking a meaningful step from plausible textual outputs toward physically verifiable spatial predictions.", "date": "2026-04-23", "categories": []}
{"id": "2604.20093", "title": "FurnSet: Exploiting Repeats for 3D Scene Reconstruction", "abstract": "Single-view 3D scene reconstruction involves inferring both object geometry and spatial layout. Existing methods typically reconstruct objects independently or rely on implicit scene context, failing to exploit the repeated instances commonly present in realworld scenes. We propose FurnSet, a framework that explicitly identifies and leverages repeated object instances to improve reconstruction. Our method introduces per-object CLS tokens and a set-aware self-attention mechanism that groups identical instances and aggregates complementary observations across them, enabling joint reconstruction. We further combine scene-level and object-level conditioning to guide object reconstruction, followed by layout optimization using object point clouds with 3D and 2D projection losses for scene alignment. Experiments on 3D-Future and 3D-Front demonstrate improved scene reconstruction quality, highlighting the effectiveness of exploiting repetition for robust 3D scene reconstruction.", "date": "2026-04-22", "categories": []}
{"id": "2512.15577", "title": "MoonSeg3R: Monocular Online Zero-Shot Segment Anything in 3D with Reconstructive Foundation Priors", "abstract": "In this paper, we focus on online zero-shot monocular 3D instance segmentation, a novel practical setting where existing approaches fail to perform because they rely on posed RGB-D sequences. To overcome this limitation, we leverage CUT3R, a recent Reconstructive Foundation Model (RFM), to provide reliable geometric priors from a single RGB stream. We propose MoonSeg3R, which introduces three key components: (1) a self-supervised query refinement module with spatial-semantic distillation that transforms segmentation masks from 2D visual foundation models (VFMs) into discriminative 3D queries; (2) a 3D query index memory that provides temporal consistency by retrieving contextual queries; and (3) a state-distribution token from CUT3R that acts as a mask identity descriptor to strengthen cross-frame fusion. Experiments on ScanNet200 and SceneNN show that MoonSeg3R is the first method to enable online monocular 3D segmentation and achieves performance competitive with state-of-the-art RGB-D-based systems. Our code is available at https://github.com/VICO-UoE/MoonSeg3R.", "date": "2025-12-17", "categories": []}
{"id": "2602.06400", "title": "TFusionOcc: T-Primitive Based Object-Centric Multi-Sensor Fusion Framework for 3D Occupancy Prediction", "abstract": "The prediction of 3D semantic occupancy enables autonomous vehicles (AVs) to perceive the fine-grained geometric and semantic scene structure for safe navigation and decision-making. Existing methods mainly rely on either voxel-based representations, which incur redundant computation over empty regions, or on object-centric Gaussian primitives, which are limited in modeling complex, non-convex, and asymmetric structures. In this paper, we present TFusionOcc, a T-primitive-based object-centric multi-sensor fusion framework for 3D semantic occupancy prediction. Specifically, we introduce a family of Students t-distribution-based T-primitives, including the plain T-primitive, T-Superquadric, and deformable T-Superquadric with inverse warping, where the deformable T-Superquadric serves as the key geometry-enhancing primitive. We further develop a unified probabilistic formulation based on the Students t-distribution and the T-mixture model (TMM) to jointly model occupancy and semantics, and design a tightly coupled multi-stage fusion architecture to effectively integrate camera and LiDAR cues. Extensive experiments on nuScenes show state-of-the-art performance, while additional evaluations on nuScenes-C demonstrate strong robustness under most corruption scenarios. The code will be available at: https://github.com/DanielMing123/TFusionOcc", "date": "2026-02-06", "categories": []}
{"id": "2604.18831", "title": "Feasibility of Indoor Frame-Wise Lidar Semantic Segmentation via Distillation from Visual Foundation Model", "abstract": "Frame-wise semantic segmentation of indoor lidar scans is a fundamental step toward higher-level 3D scene understanding and mapping applications. However, acquiring frame-wise ground truth for training deep learning models is costly and time-consuming. This challenge is largely addressed, for imagery, by Visual Foundation Models (VFMs) which segment image frames. The same VFMs may be used to train a lidar scan frame segmentation model via a 2D-to-3D distillation pipeline. The success of such distillation has been shown for autonomous driving scenes, but not yet for indoor scenes. Here, we study the feasibility of repeating this success for indoor scenes, in a frame-wise distillation manner by coupling each lidar scan with a VFM-processed camera image. The evaluation is done using indoor SLAM datasets, where pseudo-labels are used for downstream evaluation. Also, a small manually annotated lidar dataset is provided for validation, as there are no other lidar frame-wise indoor datasets with semantics. Results show that the distilled model achieves up to 56% mIoU under pseudo-label evaluation and around 36% mIoU with real-label, demonstrating the feasibility of cross-modal distillation for indoor lidar semantic segmentation without manual annotations.", "date": "2026-04-20", "categories": []}
{"id": "2604.19411", "title": "GOLD-BEV: GrOund and aeriaL Data for Dense Semantic BEV Mapping of Dynamic Scenes", "abstract": "Understanding road scenes in a geometrically consistent, scene-centric representation is crucial for planning and mapping. We present GOLD-BEV, a framework that learns dense bird's-eye-view (BEV) semantic environment maps-including dynamic agents-from ego-centric sensors, using time-synchronized aerial imagery as supervision only during training. BEV-aligned aerial crops provide an intuitive target space, enabling dense semantic annotation with minimal manual effort and avoiding the ambiguity of ego-only BEV labeling. Crucially, strict aerial-ground synchronization allows overhead observations to supervise moving traffic participants and mitigates the temporal inconsistencies inherent to non-synchronized overhead sources. To obtain scalable dense targets, we generate BEV pseudo-labels using domain-adapted aerial teachers, and jointly train BEV segmentation with optional pseudo-aerial BEV reconstruction for interpretability. Finally, we extend beyond aerial coverage by learning to synthesize pseudo-aerial BEV images from ego sensors, which support lightweight human annotation and uncertainty-aware pseudo-labeling on unlabeled drives.", "date": "2026-04-21", "categories": []}
{"id": "2604.19609", "title": "Volume Transformer: Revisiting Vanilla Transformers for 3D Scene Understanding", "abstract": "Transformers have become a common foundation across deep learning, yet 3D scene understanding still relies on specialized backbones with strong domain priors. This keeps the field isolated from the broader Transformer ecosystem, limiting the transfer of new advances as well as the benefits of increasingly optimized software and hardware stacks. To bridge this gap, we adapt the vanilla Transformer encoder to 3D scenes with minimal modifications. Given an input 3D scene, we partition it into volumetric patch tokens, process them with full global self-attention, and inject positional information via a 3D extension of rotary positional embeddings. We call the resulting model the Volume Transformer (Volt) and apply it to 3D semantic segmentation. Naively training Volt on standard 3D benchmarks leads to shortcut learning, highlighting the limited scale of current 3D supervision. To overcome this, we introduce a data-efficient training recipe based on strong 3D augmentations, regularization, and distillation from a convolutional teacher, making Volt competitive with state-of-the-art methods. We then scale supervision through joint training on multiple datasets and show that Volt benefits more from increased scale than domain-specific 3D backbones, achieving state-of-the-art results across indoor and outdoor datasets. Finally, when used as a drop-in backbone in a standard 3D instance segmentation pipeline, Volt again sets a new state of the art, highlighting its potential as a simple, scalable, general-purpose backbone for 3D scene understanding.", "date": "2026-04-21", "categories": []}
{"id": "2603.08096", "title": "TrianguLang: Geometry-Aware Semantic Consensus for Pose-Free 3D Localization", "abstract": "Localizing objects and parts from natural language in 3D space is essential for robotics, AR, and embodied AI, yet existing methods face a trade-off between the accuracy and geometric consistency of per-scene optimization and the efficiency of feed-forward inference. We present TrianguLang, a feed-forward framework for 3D localization that requires no camera calibration at inference. Unlike prior methods that treat views independently, we introduce Geometry-Aware Semantic Attention (GASA), which utilizes predicted geometry to gate cross-view feature correspondence, suppressing semantically-plausible but geometrically-inconsistent matches without requiring ground-truth poses. Validated on five benchmarks including ScanNet++ and uCO3D, TrianguLang achieves state-of-the-art feed-forward text-guided segmentation and localization, reducing user effort from $O(N)$ clicks to a single text query. The model processes each frame at 1008x1008 resolution in $\\sim$57ms ($\\sim$18 FPS) without optimization, enabling practical deployment for interactive robotics and AR applications. Code and checkpoints are available at https://cwru-aism.github.io/triangulang/.", "date": "2026-03-09", "categories": []}
{"id": "2603.28178", "title": "ToLL: Topological Layout Learning with Asymmetric Cross-View Structural Distillation for 3D Scene Graph Generation Pretraining", "abstract": "3D Scene Graph (3DSG) generation plays a pivotal role in spatial understanding and affordance perception. To mitigate generalization issues from data scarcity, joint-embedding and generative proxy tasks are proposed to pre-train 3DSG representations on predicate label-free datasets. Currently, generative pre-training usually bypasses the semantic corruption caused by the geometric augmentations in joint-embedding, but cannot avoid a negative problem ``Geometric Shortcut.\" In this problem, exposing dense object spatial and scale priors will induce models to trivially reconstruct scenes by interpolating object positions, rather than learning the underlying topological constraints provided by edges. To address this issue, we propose a Topological Layout Learning (ToLL) for 3DSG generation pretraining framework. In detail, we design an Anchor-Conditioned Topological Geometry Reasoning. It adopts a recurrent GNN to recover the global layout of zero-centered subgraphs (the non-visible spatial features) by one anchor with sparse spatial prior. Considering the absence of spatial layout information within the objects, it creates an information bottleneck, compelling our model to recover the full scene layout by leveraging predicate representation learning. Moreover, we construct a Structural Multi-view Augmentation to avoid semantic corruption, enhancing 3DSG representations via self-distillation. The extensive experiments on special dataset demonstrate that our ToLL could often improve 3DSG pertaining quality, outperforming state-of-the-art baselines.", "date": "2026-03-30", "categories": []}
{"id": "2604.08626", "title": "WildDet3D: Scaling Promptable 3D Detection in the Wild", "abstract": "Understanding objects in 3D from a single image is a cornerstone of spatial intelligence. A key step toward this goal is monocular 3D object detection--recovering the extent, location, and orientation of objects from an input RGB image. To be practical in the open world, such a detector must generalize beyond closed-set categories, support diverse prompt modalities, and leverage geometric cues when available. Progress is hampered by two bottlenecks: existing methods are designed for a single prompt type and lack a mechanism to incorporate additional geometric cues, and current 3D datasets cover only narrow categories in controlled environments, limiting open-world transfer. In this work we address both gaps. First, we introduce WildDet3D, a unified geometry-aware architecture that natively accepts text, point, and box prompts and can incorporate auxiliary depth signals at inference time. Second, we present WildDet3D-Data, the largest open 3D detection dataset to date, constructed by generating candidate 3D boxes from existing 2D annotations and retaining only human-verified ones, yielding over 1M images across 13.5K categories in diverse real-world scenes. WildDet3D establishes a new state-of-the-art across multiple benchmarks and settings. In the open-world setting, it achieves 22.6/24.8 AP3D on our newly introduced WildDet3D-Bench with text and box prompts. On Omni3D, it reaches 34.2/36.4 AP3D with text and box prompts, respectively. In zero-shot evaluation, it achieves 40.3/48.9 ODS on Argoverse 2 and ScanNet. Notably, incorporating depth cues at inference time yields substantial additional gains (+20.7 AP on average across settings).", "date": "2026-04-09", "categories": []}
{"id": "2604.17454", "title": "HSG: Hyperbolic Scene Graph", "abstract": "Scene graph representations enable structured visual understanding by modeling objects and their relationships, and have been widely used for multiview and 3D scene reasoning. Existing methods such as MSG learn scene graph embeddings in Euclidean space using contrastive learning and attention based association. However, Euclidean geometry does not explicitly capture hierarchical entailment relationships between places and objects, limiting the structural consistency of learned representations. To address this, we propose Hyperbolic Scene Graph (HSG), which learns scene graph embeddings in hyperbolic space where hierarchical relationships are naturally encoded through geometric distance. Our results show that HSG improves hierarchical structure quality while maintaining strong retrieval performance. The largest gains are observed in graph level metrics: HSG achieves a PP IoU of 33.17 and the highest Graph IoU of 33.51, outperforming the best AoMSG variant (25.37) by 8.14, highlighting the effectiveness of hyperbolic representation learning for scene graph modeling. Code: https://github.com/AIGeeksGroup/HSG.", "date": "2026-04-19", "categories": []}
{"id": "2602.16385", "title": "Adaptive Multi-Scale Channel-Spatial Attention Aggregation Framework for 3D Indoor Semantic Scene Completion Toward Assisting Visually Impaired", "abstract": "Independent indoor mobility remains a critical challenge for individuals with visual impairments, largely due to the limited capability of existing assistive systems in detecting fine-grained hazardous objects such as chairs, tables, and small obstacles. These perceptual blind zones substantially increase the risk of collision in unfamiliar environments. To bridge the gap between monocular 3D vision research and practical assistive deployment, this paper proposes an Adaptive Multi-scale Attention Aggregation (AMAA) framework for monocular 3D semantic scene completion using only a wearable RGB camera. The proposed framework addresses two major limitations in 2D-to-3D feature lifting: noise diffusion during back-projection and structural instability in multi-scale fusion. A parallel channel--spatial attention mechanism is introduced to recalibrate lifted features along semantic and geometric dimensions, while a hierarchical adaptive gating strategy regulates cross-scale information flow to preserve fine-grained structural details. Experiments on the NYUv2 benchmark demonstrate that AMAA achieves an overall mIoU of 27.88%. Crucially, it yields significant relative improvements of 16.9% for small objects and 10.4% for tables over the MonoScene baseline. Furthermore, a wearable prototype based on an NVIDIA Jetson Orin NX and a ZED~2i camera validates stable real-time performance in indoor environments, demonstrating the feasibility of deploying monocular 3D scene completion for assistive navigation.", "date": "2026-02-18", "categories": []}
{"id": "2604.14025", "title": "Feed-Forward 3D Scene Modeling: A Problem-Driven Perspective", "abstract": "Reconstructing 3D representations from 2D inputs is a fundamental task in computer vision and graphics, serving as a cornerstone for understanding and interacting with the physical world. While traditional methods achieve high fidelity, they are limited by slow per-scene optimization or category-specific training, which hinders their practical deployment and scalability. Hence, generalizable feed-forward 3D reconstruction has witnessed rapid development in recent years. By learning a model that maps images directly to 3D representations in a single forward pass, these methods enable efficient reconstruction and robust cross-scene generalization. Our survey is motivated by a critical observation: despite the diverse geometric output representations, ranging from implicit fields to explicit primitives, existing feed-forward approaches share similar high-level architectural patterns, such as image feature extraction backbones, multi-view information fusion mechanisms, and geometry-aware design principles. Consequently, we abstract away from these representation differences and instead focus on model design, proposing a novel taxonomy centered on model design strategies that are agnostic to the output format. Our proposed taxonomy organizes the research directions into five key problems that drive recent research development: feature enhancement, geometry awareness, model efficiency, augmentation strategies and temporal-aware models. To support this taxonomy with empirical grounding and standardized evaluation, we further comprehensively review related benchmarks and datasets, and extensively discuss and categorize real-world applications based on feed-forward 3D models. Finally, we outline future directions to address open challenges such as scalability, evaluation standards, and world modeling.", "date": "2026-04-15", "categories": []}
{"id": "2604.14144", "title": "SpatialEvo: Self-Evolving Spatial Intelligence via Deterministic Geometric Environments", "abstract": "Spatial reasoning over three-dimensional scenes is a core capability for embodied intelligence, yet continuous model improvement remains bottlenecked by the cost of geometric annotation. The self-evolving paradigm offers a promising path, but its reliance on model consensus to construct pseudo-labels causes training to reinforce rather than correct the model's own geometric errors. We identify a property unique to 3D spatial reasoning that circumvents this limitation: ground truth is a deterministic consequence of the underlying geometry, computable exactly from point clouds and camera poses without any model involvement. Building on this insight, we present SpatialEvo, a self-evolving framework for 3D spatial reasoning, centered on the Deterministic Geometric Environment (DGE). The DGE formalizes 16 spatial reasoning task categories under explicit geometric validation rules and converts unannotated 3D scenes into zero-noise interactive oracles, replacing model consensus with objective physical feedback. A single shared-parameter policy co-evolves across questioner and solver roles under DGE constraints: the questioner generates physically valid spatial questions grounded in scene observations, while the solver derives precise answers against DGE-verified ground truth. A task-adaptive scheduler endogenously concentrates training on the model's weakest categories, producing a dynamic curriculum without manual design. Experiments across nine benchmarks demonstrate that SpatialEvo achieves the highest average score at both 3B and 7B scales, with consistent gains on spatial reasoning benchmarks and no degradation on general visual understanding.", "date": "2026-04-15", "categories": []}
{"id": "2511.22039", "title": "SparseWorld-TC: Trajectory-Conditioned Sparse Occupancy World Model", "abstract": "This paper introduces a novel architecture for trajectory-conditioned forecasting of future 3D scene occupancy. In contrast to methods that rely on variational autoencoders (VAEs) to generate discrete occupancy tokens, which inherently limit representational capacity, our approach predicts multi-frame future occupancy in an end-to-end manner directly from raw image features. Inspired by the success of attention-based transformer architectures in foundational vision and language models such as GPT and VGGT, we employ a sparse occupancy representation that bypasses the intermediate bird's eye view (BEV) projection and its explicit geometric priors. This design allows the transformer to capture spatiotemporal dependencies more effectively. By avoiding both the finite-capacity constraint of discrete tokenization and the structural limitations of BEV representations, our method achieves state-of-the-art performance on the nuScenes benchmark for 1-3 second occupancy forecasting, outperforming existing approaches by a significant margin. Furthermore, it demonstrates robust scene dynamics understanding, consistently delivering high accuracy under arbitrary future trajectory conditioning.", "date": "2025-11-27", "categories": []}
{"id": "2511.11232", "title": "DoReMi: Bridging 3D Domains via Topology-Aware Domain-Representation Mixture of Experts", "abstract": "Constructing a unified 3D scene understanding model has long been hindered by the significant topological discrepancies across different sensor modalities. While applying the Mixture-of-Experts (MoE) architecture is an effective approach to achieving universal understanding, we observe that existing 3D MoE networks often suffer from semantics-driven routing bias. This makes it challenging to address cross-domain data characterized by \"semantic consistency yet topological heterogeneity.\" To overcome this challenge, we propose DoReMi (Topology-Aware Domain-Representation Mixture of Experts). Specifically, we introduce a self-supervised pre-training branch based on multi attributes, such as topological and texture variations, to anchor cross-domain structural priors. Building upon this, we design a domain-aware expert branch comprising two core mechanisms: Domain Spatial-Guided Routing (DSR), which achieves an acute perception of local topological variations by extracting spatial contexts, and Entropy-controlled Dynamic Allocation (EDA), which dynamically adjusts the number of activated experts by quantifying routing uncertainty to ensure training stability. Through the synergy of these dual branches, DoReMi achieves a deep integration of universal feature extraction and highly adaptive expert allocation. Extensive experiments across various tasks, encompassing both indoor and outdoor scenes, validate the superiority of DoReMi. It achieves 80.1% mIoU on the ScanNet validation set and 77.2% mIoU on S3DIS, comprehensively outperforming existing state-of-the-art methods. The code will be released soon.", "date": "2025-11-14", "categories": []}
{"id": "2604.10573", "title": "Learning 3D Representations for Spatial Intelligence from Unposed Multi-View Images", "abstract": "Robust 3D representation learning forms the perceptual foundation of spatial intelligence, enabling downstream tasks in scene understanding and embodied AI. However, learning such representations directly from unposed multi-view images remains challenging. Recent self-supervised methods attempt to unify geometry, appearance, and semantics in a feed-forward manner, but they often suffer from weak geometry induction, limited appearance detail, and inconsistencies between geometry and semantics. We introduce UniSplat, a feed-forward framework designed to address these limitations through three complementary components. First, we propose a dual-masking strategy that strengthens geometry induction in the encoder. By masking both encoder and decoder tokens, and targeting decoder masks toward geometry-rich regions, the model is forced to infer structural information from incomplete visual cues, yielding geometry-aware representations even under unposed inputs. Second, we develop a coarse-to-fine Gaussian splatting strategy that reduces appearance-semantics inconsistencies by progressively refining the radiance field. Finally, to enforce geometric-semantic consistency, we introduce a pose-conditioned recalibration mechanism that interrelates the outputs of multiple heads by re-projecting predicted 3D point and semantic maps into the image plane using estimated camera parameters, and aligning them with corresponding RGB and semantic predictions to ensure cross-task consistency, thereby resolving geometry-semantic mismatches. Together, these components yield unified 3D representations that are robust to unposed, sparse-view inputs and generalize across diverse tasks, laying a perceptual foundation for spatial intelligence.", "date": "2026-04-12", "categories": []}
{"id": "2604.10789", "title": "ReplicateAnyScene: Zero-Shot Video-to-3D Composition via Textual-Visual-Spatial Alignment", "abstract": "Humans exhibit an innate capacity to rapidly perceive and segment objects from video observations, and even mentally assemble them into structured 3D scenes. Replicating such capability, termed compositional 3D reconstruction, is pivotal for the advancement of Spatial Intelligence and Embodied AI. However, existing methods struggle to achieve practical deployment due to the insufficient integration of cross-modal information, leaving them dependent on manual object prompting, reliant on auxiliary visual inputs, and restricted to overly simplistic scenes by training biases. To address these limitations, we propose ReplicateAnyScene, a framework capable of fully automated and zero-shot transformation of casually captured videos into compositional 3D scenes. Specifically, our pipeline incorporates a five-stage cascade to extract and structurally align generic priors from vision foundation models across textual, visual, and spatial dimensions, grounding them into structured 3D representations and ensuring semantic coherence and physical plausibility of the constructed scenes. To facilitate a more comprehensive evaluation of this task, we further introduce the C3DR benchmark to assess reconstruction quality from diverse aspects. Extensive experiments demonstrate the superiority of our method over existing baselines in generating high-quality compositional 3D scenes.", "date": "2026-04-12", "categories": []}
{"id": "2604.11401", "title": "GS4City: Hierarchical Semantic Gaussian Splatting via City-Model Priors", "abstract": "Recent semantic 3D Gaussian Splatting (3DGS) methods primarily rely on 2D foundation models, often yielding ambiguous boundaries and limited support for structured urban semantics. While city models such as CityGML encode hierarchically organized semantics together with building geometry, these labels cannot be directly mapped to Gaussian primitives. We present GS4City, a hierarchical semantic Gaussian Splatting method that incorporates city-model priors for urban scene understanding. GS4City derives reliable image-aligned masks from Level of Detail (LoD) 3 CityGML models via two-pass raycasting, explicitly using parent-child relations to validate and recover fine-grained facade elements. It then fuses these geometry-grounded masks with foundation-model predictions to establish scene-consistent instance correspondences, and learns a compact identity encoding for each Gaussian under joint 2D identity supervision and 3D spatial regularization. Experiments on the TUM2TWIN and Gold Coast datasets show that GS4City effectively incorporates structured building semantics into Gaussian scene representations, outperforming existing 2D-driven semantic 3DGS baselines, including LangSplat and Gaga, by up to 15.8 IoU points in coarse building segmentation and 14.2 mIoU points in fine-grained semantic segmentation. By bridging structured city models and photorealistic Gaussian scene representations, GS4City enables semantically queryable and structure-aware urban reconstruction. Code is available at https://github.com/Jinyzzz/GS4City.", "date": "2026-04-13", "categories": []}
{"id": "2510.11340", "title": "REACT3D: Recovering Articulations for Interactive Physical 3D Scenes", "abstract": "Interactive 3D scenes are increasingly vital for embodied intelligence, yet existing datasets remain limited due to the labor-intensive process of annotating part segmentation, kinematic types, and motion trajectories. We present REACT3D, a scalable zero-shot framework that converts static 3D scenes into simulation-ready interactive replicas with consistent geometry, enabling direct use in diverse downstream tasks. Our contributions include: (i) openable-object detection and segmentation to extract candidate movable parts from static scenes, (ii) articulation estimation that infers joint types and motion parameters, (iii) hidden-geometry completion followed by interactive object assembly, and (iv) interactive scene integration in widely supported formats to ensure compatibility with standard simulation platforms. We achieve state-of-the-art performance on detection/segmentation and articulation metrics across diverse indoor scenes, demonstrating the effectiveness of our framework and providing a practical foundation for scalable interactive scene generation, thereby lowering the barrier to large-scale research on articulated scene understanding. Our project page is https://react3d.github.io/", "date": "2025-10-13", "categories": []}
{"id": "2512.03370", "title": "ShelfGaussian: Shelf-Supervised Open-Vocabulary Gaussian-based 3D Scene Understanding", "abstract": "We introduce ShelfGaussian, an open-vocabulary multi-modal Gaussian-based 3D scene understanding framework supervised by off-the-shelf vision foundation models (VFMs). Gaussian-based methods have demonstrated superior performance and computational efficiency across a wide range of scene understanding tasks. However, existing methods either model objects as closed-set semantic Gaussians supervised by annotated 3D labels, neglecting their rendering ability, or learn open-set Gaussian representations via purely 2D self-supervision, leading to degraded geometry and limited to camera-only settings. To fully exploit the potential of Gaussians, we propose a Multi-Modal Gaussian Transformer that enables Gaussians to query features from diverse sensor modalities, and a Shelf-Supervised Learning Paradigm that efficiently optimizes Gaussians with VFM features jointly at 2D image and 3D scene levels. We evaluate ShelfGaussian on various perception and planning tasks. Experiments on Occ3D-nuScenes demonstrate its state-of-the-art zero-shot semantic occupancy prediction performance. ShelfGaussian is further evaluated on an unmanned ground vehicle (UGV) to assess its in the-wild performance across diverse urban scenarios. Project website: https://lunarlab-gatech.github.io/ShelfGaussian/.", "date": "2025-12-03", "categories": []}
{"id": "2604.09045", "title": "Scene-Agnostic Object-Centric Representation Learning for 3D Gaussian Splatting", "abstract": "Recent works on 3D scene understanding leverage 2D masks from visual foundation models (VFMs) to supervise radiance fields, enabling instance-level 3D segmentation. However, the supervision signals from foundation models are not fundamentally object-centric and often require additional mask pre/post-processing or specialized training and loss design to resolve mask identity conflicts across views. The learned identity of the 3D scene is scene-dependent, limiting generalizability across scenes. Therefore, we propose a dataset-level, object-centric supervision scheme to learn object representations in 3D Gaussian Splatting (3DGS). Building on a pre-trained slot attention-based Global Object Centric Learning (GOCL) module, we learn a scene-agnostic object codebook that provides consistent, identity-anchored representations across views and scenes. By coupling the codebook with the module's unsupervised object masks, we can directly supervise the identity features of 3D Gaussians without additional mask pre-/post-processing or explicit multi-view alignment. The learned scene-agnostic codebook enables object supervision and identification without per-scene fine-tuning or retraining. Our method thus introduces unsupervised object-centric learning (OCL) into 3DGS, yielding more structured representations and better generalization for downstream tasks such as robotic interaction, scene understanding, and cross-scene generalization.", "date": "2026-04-10", "categories": []}
{"id": "2604.09167", "title": "MAG-3D: Multi-Agent Grounded Reasoning for 3D Understanding", "abstract": "Vision-language models (VLMs) have achieved strong performance in multimodal understanding and reasoning, yet grounded reasoning in 3D scenes remains underexplored. Effective 3D reasoning hinges on accurate grounding: to answer open-ended queries, a model must first identify query-relevant objects and regions in a complex scene, and then reason about their spatial and geometric relationships. Recent approaches have demonstrated strong potential for grounded 3D reasoning. However, they often rely on in-domain tuning or hand-crafted reasoning pipelines, which limit their flexibility and zero-shot generalization to novel environments. In this work, we present MAG-3D, a training-free multi-agent framework for grounded 3D reasoning with off-the-shelf VLMs. Instead of relying on task-specific training or fixed reasoning procedures, MAG-3D dynamically coordinates expert agents to address the key challenges of 3D reasoning. Specifically, we propose a planning agent that decomposes the task and orchestrates the overall reasoning process, a grounding agent that performs free-form 3D grounding and relevant frame retrieval from extensive 3D scene observations, and a coding agent that conducts flexible geometric reasoning and explicit verification through executable programs. This multi-agent collaborative design enables flexible training-free 3D grounded reasoning across diverse scenes and achieves state-of-the-art performance on challenging benchmarks.", "date": "2026-04-10", "categories": []}
{"id": "2604.07010", "title": "Synthetic Dataset Generation for Partially Observed Indoor Objects", "abstract": "Learning-based methods for 3D scene reconstruction and object completion require large datasets containing partial scans paired with complete ground-truth geometry. However, acquiring such datasets using real-world scanning systems is costly and time-consuming, particularly when accurate ground truth for occluded regions is required.\n  In this work, we present a virtual scanning framework implemented in Unity for generating realistic synthetic 3D scan datasets. The proposed system simulates the behaviour of real-world scanners using configurable parameters such as scan resolution, measurement range, and distance-dependent noise. Instead of directly sampling mesh surfaces, the framework performs ray-based scanning from virtual viewpoints, enabling realistic modelling of sensor visibility and occlusion effects. In addition, panoramic images captured at the scanner location are used to assign colours to the resulting point clouds.\n  To support scalable dataset creation, the scanner is integrated with a procedural indoor scene generation pipeline that automatically produces diverse room layouts and furniture arrangements. Using this system, we introduce the \\textit{V-Scan} dataset, which contains synthetic indoor scans together with object-level partial point clouds, voxel-based occlusion grids, and complete ground-truth geometry. The resulting dataset provides valuable supervision for training and evaluating learning-based methods for scene reconstruction and object completion.", "date": "2026-04-08", "categories": []}
{"id": "2604.05638", "title": "PanopticQuery: Unified Query-Time Reasoning for 4D Scenes", "abstract": "Understanding dynamic 4D environments through natural language queries requires not only accurate scene reconstruction but also robust semantic grounding across space, time, and viewpoints. While recent methods using neural representations have advanced 4D reconstruction, they remain limited in contextual reasoning, especially for complex semantics such as interactions, temporal actions, and spatial relations. A key challenge lies in transforming noisy, view-dependent predictions into globally consistent 4D interpretations. We introduce PanopticQuery, a framework for unified query-time reasoning in 4D scenes. Our approach builds on 4D Gaussian Splatting for high-fidelity dynamic reconstruction and introduces a multi-view semantic consensus mechanism that grounds natural language queries by aggregating 2D semantic predictions across multiple views and time frames. This process filters inconsistent outputs, enforces geometric consistency, and lifts 2D semantics into structured 4D groundings via neural field optimization. To support evaluation, we present Panoptic-L4D, a new benchmark for language-based querying in dynamic scenes. Experiments demonstrate that PanopticQuery sets a new state of the art on complex language queries, effectively handling attributes, actions, spatial relationships, and multi-object interactions. A video demonstration is available in the supplementary materials.", "date": "2026-04-07", "categories": []}
{"id": "2604.05780", "title": "Sparsity-Aware Voxel Attention and Foreground Modulation for 3D Semantic Scene Completion", "abstract": "Monocular Semantic Scene Completion (SSC) aims to reconstruct complete 3D semantic scenes from a single RGB image, offering a cost-effective solution for autonomous driving and robotics. However, the inherently imbalanced nature of voxel distributions, where over 93% of voxels are empty and foreground classes are rare, poses significant challenges. Existing methods often suffer from redundant emphasis on uninformative voxels and poor generalization to long-tailed categories. To address these issues, we propose VoxSAMNet (Voxel Sparsity-Aware Modulation Network), a unified framework that explicitly models voxel sparsity and semantic imbalance. Our approach introduces: (1) a Dummy Shortcut for Feature Refinement (DSFR) module that bypasses empty voxels via a shared dummy node while refining occupied ones with deformable attention; and (2) a Foreground Modulation Strategy combining Foreground Dropout (FD) and Text-Guided Image Filter (TGIF) to alleviate overfitting and enhance class-relevant features. Extensive experiments on the public benchmarks SemanticKITTI and SSCBench-KITTI-360 demonstrate that VoxSAMNet achieves state-of-the-art performance, surpassing prior monocular and stereo baselines with mIoU scores of 18.2% and 20.2%, respectively. Our results highlight the importance of sparsity-aware and semantics-guided design for efficient and accurate 3D scene completion, offering a promising direction for future research.", "date": "2026-04-07", "categories": []}
{"id": "2511.23230", "title": "Action-guided generation of 3D functionality segmentation data", "abstract": "3D functionality segmentation aims to identify the interactive element in a 3D scene required to perform an action described in free-form language (e.g., the handle to ``Open the second drawer of the cabinet near the bed''). Progress has been constrained by the scarcity of annotated real-world data, as collecting and labeling fine-grained 3D masks is prohibitively expensive. To address this limitation, we introduce SynthFun3D, the first method for generating 3D functionality segmentation data directly from action descriptions. Given an action description, SynthFun3D constructs a plausible 3D scene by retrieving objects with part-level annotations from a large-scale asset repository and arranging them under spatial and semantic constraints. SynthFun3D renders multi-view images and automatically identifies the target functional element, producing precise ground-truth masks without manual annotation. We demonstrate the effectiveness of the generated data by training a VLM-based 3D functionality segmentation model. Augmenting real-world data with our synthetic data consistently improves performance, with gains of +2.2 mAP, +6.3 mAR, and +5.7 mIoU over real-only training. This shows that action-guided synthetic data generation provides a scalable and effective complement to manual annotation for 3D functionality understanding. Project page: tev-fbk.github.io/synthfun3d.", "date": "2025-11-28", "categories": []}
{"id": "2604.03309", "title": "TreeGaussian: Tree-Guided Cascaded Contrastive Learning for Hierarchical Consistent 3D Gaussian Scene Segmentation and Understanding", "abstract": "3D Gaussian Splatting (3DGS) has emerged as a real-time, differentiable representation for neural scene understanding. However, existing 3DGS-based methods struggle to represent hierarchical 3D semantic structures and capture whole-part relationships in complex scenes. Moreover, dense pairwise comparisons and inconsistent hierarchical labels from 2D priors hinder feature learning, resulting in suboptimal segmentation. To address these limitations, we introduce TreeGaussian, a tree-guided cascaded contrastive learning framework that explicitly models hierarchical semantic relationships and reduces redundancy in contrastive supervision. By constructing a multi-level object tree, TreeGaussian enables structured learning across object-part hierarchies. In addition, we propose a two-stage cascaded contrastive learning strategy that progressively refines feature representations from global to local, mitigating saturation and stabilizing training. A Consistent Segmentation Detection (CSD) mechanism and a graph-based denoising module are further introduced to align segmentation modes across views while suppressing unstable Gaussian points, enhancing segmentation consistency and quality. Extensive experiments, including open-vocabulary 3D object selection, 3D point cloud understanding, and ablation studies, demonstrate the effectiveness and robustness of our approach.", "date": "2026-03-31", "categories": []}
{"id": "2604.03696", "title": "FunFact: Building Probabilistic Functional 3D Scene Graphs via Factor-Graph Reasoning", "abstract": "Recent work in 3D scene understanding is moving beyond purely spatial analysis toward functional scene understanding. However, existing methods often consider functional relationships between object pairs in isolation, failing to capture the scene-wide interdependence that humans use to resolve ambiguity. We introduce FunFact, a framework for constructing probabilistic open-vocabulary functional 3D scene graphs from posed RGB-D images. FunFact first builds an object- and part-centric 3D map and uses foundation models to propose semantically plausible functional relations. These candidates are converted into factor graph variables and constrained by both LLM-derived common-sense priors and geometric priors. This formulation enables joint probabilistic inference over all functional edges and their marginals, yielding substantially better calibrated confidence scores. To benchmark this setting, we introduce FunThor, a synthetic dataset based on AI2-THOR with part-level geometry and rule-based functional annotations. Experiments on SceneFun3D, FunGraph3D, and FunThor show that FunFact improves node and relation discovery recall and significantly reduces calibration error for ambiguous relations, highlighting the benefits of holistic probabilistic modeling for functional scene understanding. See our project page at https://funfact-scenegraph.github.io/", "date": "2026-04-04", "categories": []}
{"id": "2604.02546", "title": "Contrastive Language-Colored Pointmap Pretraining for Unified 3D Scene Understanding", "abstract": "Pretraining 3D encoders by aligning with Contrastive Language Image Pretraining (CLIP) has emerged as a promising direction to learn generalizable representations for 3D scene understanding. In this paper, we propose UniScene3D, a transformer-based encoder that learns unified scene representations from multi-view colored pointmaps, jointly modeling image appearance and geometry. For robust colored pointmap representation learning, we introduce novel cross-view geometric alignment and grounded view alignment to enforce cross-view geometry and semantic consistency. Extensive low-shot and task-specific fine-tuning evaluations on viewpoint grounding, scene retrieval, scene type classification, and 3D VQA demonstrate our state-of-the-art performance. These results highlight the effectiveness of our approach for unified 3D scene understanding. https://yebulabula.github.io/UniScene3D/", "date": "2026-04-02", "categories": []}
{"id": "2604.00528", "title": "Think, Act, Build: An Agentic Framework with Vision Language Models for Zero-Shot 3D Visual Grounding", "abstract": "3D Visual Grounding (3D-VG) aims to localize objects in 3D scenes via natural language descriptions. While recent advancements leveraging Vision-Language Models (VLMs) have explored zero-shot possibilities, they typically suffer from a static workflow relying on preprocessed 3D point clouds, essentially degrading grounding into proposal matching. To bypass this reliance, our core motivation is to decouple the task: leveraging 2D VLMs to resolve complex spatial semantics, while relying on deterministic multi-view geometry to instantiate the 3D structure. Driven by this insight, we propose \"Think, Act, Build (TAB)\", a dynamic agentic framework that reformulates 3D-VG tasks as a generative 2D-to-3D reconstruction paradigm operating directly on raw RGB-D streams. Specifically, guided by a specialized 3D-VG skill, our VLM agent dynamically invokes visual tools to track and reconstruct the target across 2D frames. Crucially, to overcome the multi-view coverage deficit caused by strict VLM semantic tracking, we introduce the Semantic-Anchored Geometric Expansion, a mechanism that first anchors the target in a reference video clip and then leverages multi-view geometry to propagate its spatial location across unobserved frames. This enables the agent to \"Build\" the target's 3D representation by aggregating these multi-view features via camera parameters, directly mapping 2D visual cues to 3D coordinates. Furthermore, to ensure rigorous assessment, we identify flaws such as reference ambiguity and category errors in existing benchmarks and manually refine the incorrect queries. Extensive experiments on ScanRefer and Nr3D demonstrate that our framework, relying entirely on open-source models, significantly outperforms previous zero-shot methods and even surpasses fully supervised baselines.", "date": "2026-04-01", "categories": []}
{"id": "2604.01361", "title": "IGLOSS: Image Generation for Lidar Open-vocabulary Semantic Segmentation", "abstract": "This paper presents a new method for the zero-shot open-vocabulary semantic segmentation (OVSS) of 3D automotive lidar data. To circumvent the recognized image-text modality gap that is intrinsic to approaches based on Vision Language Models (VLMs) such as CLIP, our method relies instead on image generation from text, to create prototype images. Given a 3D network distilled from a 2D Vision Foundation Model (VFM), we then label a point cloud by matching 3D point features with 2D image features of these prototypes. Our method is state-of-the-art for OVSS on nuScenes and SemanticKITTI. Code, pre-trained models, and generated images are available at https://github.com/valeoai/IGLOSS.", "date": "2026-04-01", "categories": []}
{"id": "2604.01388", "title": "LESV: Language Embedded Sparse Voxel Fusion for Open-Vocabulary 3D Scene Understanding", "abstract": "Recent advancements in open-vocabulary 3D scene understanding heavily rely on 3D Gaussian Splatting (3DGS) to register vision-language features into 3D space. However, we identify two critical limitations in these approaches: the spatial ambiguity arising from unstructured, overlapping Gaussians which necessitates probabilistic feature registration, and the multi-level semantic ambiguity caused by pooling features over object-level masks, which dilutes fine-grained details. To address these challenges, we present a novel framework that leverages Sparse Voxel Rasterization (SVRaster) as a structured, disjoint geometry representation. By regularizing SVRaster with monocular depth and normal priors, we establish a stable geometric foundation. This enables a deterministic, confidence-aware feature registration process and suppresses the semantic bleeding artifact common in 3DGS. Furthermore, we resolve multi-level ambiguity by exploiting the emerging dense alignment properties of foundation model AM-RADIO, avoiding the computational overhead of hierarchical training methods. Our approach achieves state-of-the-art performance on Open Vocabulary 3D Object Retrieval and Point Cloud Understanding benchmarks, particularly excelling on fine-grained queries where registration methods typically fail.", "date": "2026-04-01", "categories": []}
{"id": "2604.01678", "title": "Director: Instance-aware Gaussian Splatting for Dynamic Scene Modeling and Understanding", "abstract": "Volumetric video seeks to model dynamic scenes as temporally coherent 4D representations. While recent Gaussian-based approaches achieve impressive rendering fidelity, they primarily emphasize appearance but are largely agnostic to instance-level structure, limiting stable tracking and semantic reasoning in highly dynamic scenarios. In this paper, we present Director, a unified spatio-temporal Gaussian representation that jointly models human performance, high-fidelity rendering, and instance-level semantics. Our key insight is that embedding instance-consistent semantics naturally complements 4D modeling, enabling more accurate scene decomposition while supporting robust dynamic scene understanding. To this end, we leverage temporally aligned instance masks and sentence embeddings derived from Multimodal Large Language Models to supervise the learnable semantic features of each Gaussian via two MLP decoders, enabling language-aligned 4D representations and enforcing identity consistency over time. To enhance temporal stability, we bridge 2D optical flow with 4D Gaussians and finetune their motions, yielding reliable initialization and reducing drift. For the training, we further introduce a geometry-aware SDF constraints, along with regularization terms that enforces surface continuity, enhancing temporal coherence in dynamic foreground modeling. Experiments demonstrate that Director achieves temporally coherent 4D reconstructions while simultaneously enabling instance segmentation and open-vocabulary querying.", "date": "2026-04-02", "categories": []}
{"id": "2604.01882", "title": "A3R: Agentic Affordance Reasoning via Cross-Dimensional Evidence in 3D Gaussian Scenes", "abstract": "Affordance reasoning in 3D Gaussian scenes aims to identify the region that supports the action specified by a given text instruction in complex environments. Existing methods typically cast this problem as one-shot prediction from static scene observations, assuming sufficient evidence is already available for reasoning. However, in complex 3D scenes, many failure cases arise not from weak prediction capacity, but from incomplete task-relevant evidence under fixed observations. To address this limitation, we reformulate fine-grained affordance reasoning as a sequential evidence acquisition process, where ambiguity is progressively reduced through complementary 3D geometric and 2D semantic evidence. Building on this formulation, we propose A3R, an agentic affordance reasoning framework that enables an MLLM-based policy to iteratively select evidence acquisition actions and update the affordance belief through cross-dimensional evidence acquisition. To optimize such sequential decision making, we further introduce a GRPO-based policy learning strategy that improves evidence acquisition efficiency and reasoning accuracy. Extensive experiments on scene-level benchmarks show that A3R consistently surpasses static one-shot baselines, demonstrating the advantage of agentic cross-dimensional evidence acquisition for fine-grained affordance reasoning in complex 3D Gaussian scenes.", "date": "2026-04-02", "categories": []}
{"id": "2509.16483", "title": "Octree Diffusion for Semantic Scene Generation and Completion", "abstract": "The completion, extension, and generation of 3D semantic scenes are an interrelated set of capabilities that are useful for robotic navigation and exploration. Existing approaches seek to decouple these problems and solve them one-off. Additionally, these approaches are often domain-specific, requiring separate models for different data distributions, e.g.\\ indoor vs.\\ outdoor scenes.\n  To unify these techniques and provide cross-domain compatibility, we develop a single framework that can perform scene completion, extension, and generation in both indoor and outdoor scenes, which we term Octree Latent Semantic Diffusion. Our approach operates directly on an efficient dual octree graph latent representation: a hierarchical, sparse, and memory-efficient occupancy structure. This technique disentangles synthesis into two stages: (i) structure diffusion, which predicts binary split signals to construct a coarse occupancy octree, and (ii) latent semantic diffusion, which generates semantic embeddings decoded by a graph VAE into voxel-level semantic labels. To perform semantic scene completion or extension, our model leverages inference-time latent inpainting, or outpainting respectively. These inference-time methods use partial LiDAR scans or maps to condition generation, without the need for retraining or finetuning.\n  We demonstrate high-quality structure, coherent semantics, and robust completion from single LiDAR scans, as well as zero-shot generalization to out-of-distribution LiDAR data. These results indicate that completion-through-generation in a dual octree graph latent space is a practical and scalable alternative to regression-based pipelines for real-world robotic perception tasks.", "date": "2025-09-20", "categories": []}
{"id": "2604.00372", "title": "Dynamic Graph Neural Network with Adaptive Features Selection for RGB-D Based Indoor Scene Recognition", "abstract": "Multi-modality of color and depth, i.e., RGB-D, is of great importance in recent research of indoor scene recognition. In this kind of data representation, depth map is able to describe the 3D structure of scenes and geometric relations among objects. Previous works showed that local features of both modalities are vital for promotion of recognition accuracy. However, the problem of adaptive selection and effective exploitation on these key local features remains open in this field. In this paper, a dynamic graph model is proposed with adaptive node selection mechanism to solve the above problem. In this model, a dynamic graph is built up to model the relations among objects and scene, and a method of adaptive node selection is proposed to take key local features from both modalities of RGB and depth for graph modeling. After that, these nodes are grouped by three different levels, representing near or far relations among objects. Moreover, the graph model is updated dynamically according to attention weights. Finally, the updated and optimized features of RGB and depth modalities are fused together for indoor scene recognition. Experiments are performed on public datasets including SUN RGB-D and NYU Depth v2. Extensive results demonstrate that our method has superior performance when comparing to state-of-the-arts methods, and show that the proposed method is able to exploit crucial local features from both modalities of RGB and depth.", "date": "2026-04-01", "categories": []}
{"id": "2512.05597", "title": "Fast SceneScript: Fast and Accurate Language-Based 3D Scene Understanding via Multi-Token Prediction", "abstract": "Recent perception-generalist approaches based on language models have achieved state-of-the-art results across diverse tasks, including 3D scene layout estimation and 3D object detection, via unified architecture and interface. However, these approaches rely on autoregressive next-token prediction, which is inherently slow. In this work, we introduce Fast SceneScript, a novel structured language model for accurate and efficient 3D scene understanding. Our method employs multi-token prediction (MTP) to reduce the number of autoregressive iterations and significantly accelerate inference. While MTP improves speed, unreliable token predictions can significantly reduce accuracy. To filter out unreliable tokens, we adapt self-speculative decoding (SSD) for structured language models and introduce confidence-guided decoding (CGD) with an improved scoring mechanism for token reliability. Furthermore, we design a parameter-efficient mechanism that reduces the parameter overhead of MTP. Extensive experiments on synthetic and real-world benchmarks demonstrate that Fast SceneScript can generate up to 9 tokens per decoder inference step without compromising accuracy, while adding only $\\sim7.5\\%$ additional parameters.", "date": "2025-12-05", "categories": []}
{"id": "2603.25165", "title": "Towards Foundation Models for 3D Scene Understanding: Instance-Aware Self-Supervised Learning for Point Clouds", "abstract": "Recent advances in self-supervised learning (SSL) for point clouds have substantially improved 3D scene understanding without human annotations. Existing approaches emphasize semantic awareness by enforcing feature consistency across augmented views or by masked scene modeling. However, the resulting representations transfer poorly to instance localization, and often require full finetuning for strong performance. Instance awareness is a fundamental component of 3D perception, thus bridging this gap is crucial for progressing toward true 3D foundation models that support all downstream tasks on 3D data. In this work, we introduce PointINS, an instance-oriented self-supervised framework that enriches point cloud representations through geometry-aware learning. PointINS employs an orthogonal offset branch to jointly learn high-level semantic understanding and geometric reasoning, yielding instance awareness. We identify two consistent properties essential for robust instance localization and formulate them as complementary regularization strategies, Offset Distribution Regularization (ODR), which aligns predicted offsets with empirically observed geometric priors, and Spatial Clustering Regularization (SCR), which enforces local coherence by regularizing offsets with pseudo-instance masks. Through extensive experiments across five datasets, PointINS achieves on average +3.5% mAP improvement for indoor instance segmentation and +4.1% PQ gain for outdoor panoptic segmentation, paving the way for scalable 3D foundation models.", "date": "2026-03-26", "categories": []}
{"id": "2603.29798", "title": "SceneTeract: Agentic Functional Affordances and VLM Grounding in 3D Scenes", "abstract": "Embodied AI depends on interactive 3D environments that support meaningful activities for diverse users, yet assessing their functional affordances remains a core challenge. We introduce SceneTeract, a framework that verifies 3D scene functionality under agent-specific constraints. Our core contribution is a grounded verification engine that couples high-level semantic reasoning with low-level geometric checks. SceneTeract decomposes complex activities into sequences of atomic actions and validates each step against accessibility requirements (e.g., reachability, clearance, and navigability) conditioned on an embodied agent profile, using explicit physical and geometric simulations. We deploy SceneTeract to perform an in-depth evaluation of (i) synthetic indoor environments, uncovering frequent functional failures that prevent basic interactions, and (ii) the ability of frontier Vision-Language Models (VLMs) to reason about and predict functional affordances, revealing systematic mismatches between semantic confidence and physical feasibility even for the strongest current models. Finally, we leverage SceneTeract as a reward engine for VLM post-training, enabling scalable distillation of geometric constraints into reasoning models. We release the SceneTeract verification suite and data to bridge perception and physical reality in embodied 3D scene understanding.", "date": "2026-03-31", "categories": []}
{"id": "2512.20770", "title": "OccuFly: A 3D Vision Benchmark for Semantic Scene Completion from the Aerial Perspective", "abstract": "Semantic Scene Completion (SSC) is essential for 3D perception in mobile robotics, as it enables holistic scene understanding by jointly estimating dense volumetric occupancy and per-voxel semantics. Although SSC has been widely studied in terrestrial domains such as autonomous driving, aerial settings like autonomous flying remain largely unexplored, thereby limiting progress on downstream applications. Furthermore, LiDAR sensors are the primary modality for SSC data generation, which poses challenges for most uncrewed aerial vehicles (UAVs) due to flight regulations, mass and energy constraints, and the sparsity of LiDAR point clouds from elevated viewpoints. To address these limitations, we propose a LiDAR-free, camera-based data generation framework. By leveraging classical 3D reconstruction, our framework automates semantic label transfer by lifting <10% of annotated images into the reconstructed point cloud, substantially minimizing manual 3D annotation effort. Based on this framework, we introduce OccuFly, the first real-world, camera-based aerial SSC benchmark, captured across multiple altitudes and all seasons. OccuFly provides over 20,000 samples of images, semantic voxel grids, and metric depth maps across 21 semantic classes in urban, industrial, and rural environments, and follows established data organization for seamless integration. We benchmark both SSC and metric monocular depth estimation on OccuFly, revealing fundamental limitations of current vision foundation models in aerial settings and establishing new challenges for robust 3D scene understanding in the aerial domain. Visit https://github.com/markus-42/occufly.", "date": "2025-12-23", "categories": []}
{"id": "2602.23153", "title": "Efficient Encoder-Free Fourier-based 3D Large Multimodal Model", "abstract": "Large Multimodal Models (LMMs) that process 3D data typically rely on heavy, pre-trained visual encoders to extract geometric features. While recent 2D LMMs have begun to eliminate such encoders for efficiency and scalability, extending this paradigm to 3D remains challenging due to the unordered and large-scale nature of point clouds. This leaves a critical unanswered question: How can we design an LMM that tokenizes unordered 3D data effectively and efficiently without a cumbersome encoder? We propose Fase3D, the first efficient encoder-free Fourier-based 3D scene LMM. Fase3D tackles the challenges of scalability and permutation invariance with a novel tokenizer that combines point cloud serialization and the Fast Fourier Transform (FFT) to approximate self-attention. This design enables an effective and computationally minimal architecture, built upon three key innovations: First, we represent large scenes compactly via structured superpoints. Second, our space-filling curve serialization followed by an FFT enables efficient global context modeling and graph-based token merging. Lastly, our Fourier-augmented LoRA adapters inject global frequency-aware interactions into the LLMs at a negligible cost. Fase3D achieves performance comparable to encoder-based 3D LMMs while being significantly more efficient in computation and parameters. Project website: https://tev-fbk.github.io/Fase3D.", "date": "2026-02-26", "categories": []}
{"id": "2603.27238", "title": "An Instance-Centric Panoptic Occupancy Prediction Benchmark for Autonomous Driving", "abstract": "Panoptic occupancy prediction aims to jointly infer voxel-wise semantics and instance identities within a unified 3D scene representation. Nevertheless, progress in this field remains constrained by the absence of high-quality 3D mesh resources, instance-level annotations, and physically consistent occupancy datasets. Existing benchmarks typically provide incomplete and low-resolution geometry without instance-level annotations, limiting the development of models capable of achieving precise geometric reconstruction, reliable occlusion reasoning, and holistic 3D understanding. To address these challenges, this paper presents an instance-centric benchmark for the 3D panoptic occupancy prediction task. Specifically, we introduce ADMesh, the first unified 3D mesh library tailored for autonomous driving, which integrates over 15K high-quality 3D models with diverse textures and rich semantic annotations. Building upon ADMesh, we further construct CarlaOcc, a large-scale, physically consistent panoptic occupancy dataset generated using the CARLA simulator. This dataset contains over 100K frames with fine-grained, instance-level occupancy ground truth at voxel resolutions as fine as 0.05 m. Furthermore, standardized evaluation metrics are introduced to quantify the quality of existing occupancy datasets. Finally, a systematic benchmark of representative models is established on the proposed dataset, which provides a unified platform for fair comparison and reproducible research in the field of 3D panoptic perception. Code and dataset are available at https://mias.group/CarlaOcc.", "date": "2026-03-28", "categories": []}
{"id": "2603.27970", "title": "AffordMatcher: Affordance Learning in 3D Scenes from Visual Signifiers", "abstract": "Affordance learning is a complex challenge in many applications, where existing approaches primarily focus on the geometric structures, visual knowledge, and affordance labels of objects to determine interactable regions. However, extending this learning capability to a scene is significantly more complicated, as incorporating object- and scene-level semantics is not straightforward. In this work, we introduce AffordBridge, a large-scale dataset with 291,637 functional interaction annotations across 685 high-resolution indoor scenes in the form of point clouds. Our affordance annotations are complemented by RGB images that are linked to the same instances within the scenes. Building upon our dataset, we propose AffordMatcher, an affordance learning method that establishes coherent semantic correspondences between image-based and point cloud-based instances for keypoint matching, enabling a more precise identification of affordance regions based on cues, so-called visual signifiers. Experimental results on our dataset demonstrate the effectiveness of our approach compared to other methods.", "date": "2026-03-30", "categories": []}
{"id": "2603.28548", "title": "Seen2Scene: Completing Realistic 3D Scenes with Visibility-Guided Flow", "abstract": "We present Seen2Scene, the first flow matching-based approach that trains directly on incomplete, real-world 3D scans for scene completion and generation. Unlike prior methods that rely on complete and hence synthetic 3D data, our approach introduces visibility-guided flow matching, which explicitly masks out unknown regions in real scans, enabling effective learning from real-world, partial observations. We represent 3D scenes using truncated signed distance field (TSDF) volumes encoded in sparse grids and employ a sparse transformer to efficiently model complex scene structures while masking unknown regions. We employ 3D layout boxes as an input conditioning signal, and our approach is flexibly adapted to various other inputs such as text or partial scans. By learning directly from real-world, incomplete 3D scans, Seen2Scene enables realistic 3D scene completion for complex, cluttered real environments. Experiments demonstrate that our model produces coherent, complete, and realistic 3D scenes, outperforming baselines in completion accuracy and generation quality.", "date": "2026-03-30", "categories": []}
{"id": "2603.28732", "title": "Pandora: Articulated 3D Scene Graphs from Egocentric Vision", "abstract": "Robotic mapping systems typically approach building metric-semantic scene representations from the robot's own sensors and cameras. However, these \"first person\" maps inherit the robot's own limitations due to its embodiment or skillset, which may leave many aspects of the environment unexplored. For example, the robot might not be able to open drawers or access wall cabinets. In this sense, the map representation is not as complete, and requires a more capable robot to fill in the gaps. We narrow these blind spots in current methods by leveraging egocentric data captured as a human naturally explores a scene wearing Project Aria glasses, giving a way to directly transfer knowledge about articulation from the human to any deployable robot. We demonstrate that, by using simple heuristics, we can leverage egocentric data to recover models of articulate object parts, with quality comparable to those of state-of-the-art methods based on other input modalities. We also show how to integrate these models into 3D scene graph representations, leading to a better understanding of object dynamics and object-container relationships. We finally demonstrate that these articulated 3D scene graphs enhance a robot's ability to perform mobile manipulation tasks, showcasing an application where a Boston Dynamics Spot is tasked with retrieving concealed target items, given only the 3D scene graph as input.", "date": "2026-03-30", "categories": []}
{"id": "2503.07507", "title": "PE3R: Perception-Efficient 3D Reconstruction", "abstract": "Recent advances in 2D-to-3D perception have enabled the recovery of 3D scene semantics from unposed images. However, prevailing methods often suffer from limited generalization, reliance on per-scene optimization, and semantic inconsistencies across viewpoints. To address these limitations, we introduce PE3R, a tuning-free framework for efficient and generalizable 3D semantic reconstruction. By integrating multi-view geometry with 2D semantic priors in a feed-forward pipeline, PE3R achieves zero-shot generalization across diverse scenes and object categories without any scene-specific fine-tuning. Extensive evaluations on open-vocabulary segmentation and multi-view depth estimation show that PE3R not only achieves up to 9$\\times$ faster inference but also sets new state-of-the-art accuracy in both semantic and geometric metrics. Our approach paves the way for scalable, language-driven 3D scene understanding. Code is available at github.com/hujiecpp/PE3R.", "date": "2025-03-10", "categories": []}
{"id": "2512.05272", "title": "Inferring Compositional 4D Scenes without Ever Seeing One", "abstract": "Scenes in the real world are often composed of several static and dynamic objects. Capturing their 4-dimensional structures, composition and spatio-temporal configuration in-the-wild, though extremely interesting, is equally hard. Therefore, existing works often focus on one object at a time, while relying on some category-specific parametric shape model for dynamic objects. This can lead to inconsistent scene configurations, in addition to being limited to the modeled object categories. We propose COM4D (Compositional 4D), a method that consistently and jointly predicts the structure and spatio-temporal configuration of 4D/3D objects using only static multi-object or dynamic single object supervision. We achieve this by a carefully designed training of spatial and temporal attentions on 2D video input. The training is disentangled into learning from object compositions on the one hand, and single object dynamics throughout the video on the other, thus completely avoiding reliance on 4D compositional training data. At inference time, our proposed attention mixing mechanism combines these independently learned attentions, without requiring any 4D composition examples. By alternating between spatial and temporal reasoning, COM4D reconstructs complete and persistent 4D scenes with multiple interacting objects directly from monocular videos. Furthermore, COM4D provides state-of-the-art results in existing separate problems of 4D object and composed 3D reconstruction despite being purely data-driven.", "date": "2025-12-04", "categories": []}
{"id": "2603.24713", "title": "Lookalike3D: Seeing Double in 3D", "abstract": "3D object understanding and generation methods produce impressive results, yet they often overlook a pervasive source of information in real-world scenes: repeated objects. We introduce the task of lookalike object detection in indoor scenes, which leverages repeated and complementary cues from identical and near-identical object pairs. Given an input scene, the task is to classify pairs of objects as identical, similar or different using multiview images as input. To address this, we present Lookalike3D, a multiview image transformer that effectively distinguishes such object pairs by harnessing strong semantic priors from large image foundation models. To support this task, we collected the 3DTwins dataset, containing 76k manually annotated identical, similar and different pairs of objects based on ScanNet++, and show an improvement of 104% IoU over baselines. We demonstrate how our method improves downstream tasks such as enabling joint 3D object reconstruction and part co-segmentation, turning repeated and lookalike objects into a powerful cue for consistent, high-quality 3D perception. Our code, dataset and models will be made publicly available.", "date": "2026-03-25", "categories": []}
{"id": "2603.25494", "title": "AdaSFormer: Adaptive Serialized Transformers for Monocular Semantic Scene Completion from Indoor Environments", "abstract": "Indoor monocular semantic scene completion (MSSC) is notably more challenging than its outdoor counterpart due to complex spatial layouts and severe occlusions. While transformers are well suited for modeling global dependencies, their high memory cost and difficulty in reconstructing fine-grained details have limited their use in indoor MSSC. To address these limitations, we introduce AdaSFormer, a serialized transformer framework tailored for indoor MSSC. Our model features three key designs: (1) an Adaptive Serialized Transformer with learnable shifts that dynamically adjust receptive fields; (2) a Center-Relative Positional Encoding that captures spatial information richness; and (3) a Convolution-Modulated Layer Normalization that bridges heterogeneous representations between convolutional and transformer features. Extensive experiments on NYUv2 and Occ-ScanNet demonstrate that AdaSFormer achieves state-of-the-art performance. The code is publicly available at: https://github.com/alanWXZ/AdaSFormer.", "date": "2026-03-26", "categories": []}
{"id": "2603.24146", "title": "LightSplat: Fast and Memory-Efficient Open-Vocabulary 3D Scene Understanding in Five Seconds", "abstract": "Open-vocabulary 3D scene understanding enables users to segment novel objects in complex 3D environments through natural language. However, existing approaches remain slow, memory-intensive, and overly complex due to iterative optimization and dense per-Gaussian feature assignments. To address this, we propose LightSplat, a fast and memory-efficient training-free framework that injects compact 2-byte semantic indices into 3D representations from multi-view images. By assigning semantic indices only to salient regions and managing them with a lightweight index-feature mapping, LightSplat eliminates costly feature optimization and storage overhead. We further ensure semantic consistency and efficient inference via single-step clustering that links geometrically and semantically related masks in 3D. We evaluate our method on LERF-OVS, ScanNet, and DL3DV-OVS across complex indoor-outdoor scenes. As a result, LightSplat achieves state-of-the-art performance with up to 50-400x speedup and 64x lower memory, enabling scalable language-driven 3D understanding. For more details, visit our project page https://vision3d-lab.github.io/lightsplat/.", "date": "2026-03-25", "categories": []}
{"id": "2508.03643", "title": "Uni3R: Unified 3D Reconstruction and Semantic Understanding via Generalizable Gaussian Splatting from Unposed Multi-View Images", "abstract": "Reconstructing and semantically interpreting 3D scenes from sparse 2D views remains a fundamental challenge in computer vision. Conventional methods often decouple semantic understanding from reconstruction or necessitate costly per-scene optimization, thereby restricting their scalability and generalizability. In this paper, we introduce Uni3R, a novel feed-forward framework that jointly reconstructs a unified 3D scene representation enriched with open-vocabulary semantics, directly from unposed multi-view images. Our approach leverages a Cross-View Transformer to robustly integrate information across arbitrary multi-view inputs, which then regresses a set of 3D Gaussian primitives endowed with semantic feature fields. This unified representation facilitates high-fidelity novel view synthesis, open-vocabulary 3D semantic segmentation, and depth prediction, all within a single, feed-forward pass. Extensive experiments demonstrate that Uni3R establishes a new state-of-the-art across multiple benchmarks, including 25.07 PSNR on RE10K and 55.84 mIoU on ScanNet. Our work signifies a novel paradigm towards generalizable, unified 3D scene reconstruction and understanding. The code is available at https://github.com/HorizonRobotics/Uni3R.", "date": "2025-08-05", "categories": []}
{"id": "2512.02487", "title": "Masking Matters: Unlocking the Spatial Reasoning Capabilities of LLMs for 3D Scene-Language Understanding", "abstract": "Recent advances in 3D scene-language understanding have leveraged Large Language Models (LLMs) for 3D reasoning by transferring their general reasoning ability to 3D multi-modal contexts. However, existing methods typically adopt standard decoders from language modeling, which rely on a causal attention mask. This design introduces two fundamental conflicts in 3D scene understanding: sequential bias among order-agnostic 3D objects and restricted object-instruction attention, hindering task-specific reasoning. To overcome these limitations, we propose 3D Spatial Language Instruction Mask (3D-SLIM), an effective masking strategy that replaces the causal mask with an adaptive attention mask tailored to the spatial structure of 3D scenes. Our 3D-SLIM introduces two key components: a Geometry-adaptive Mask that constrains attention based on spatial density rather than token order, and an Instruction-aware Mask that enables object tokens to directly access instruction context. This design allows the model to process objects based on their spatial relationships while being guided by the user's task. 3D-SLIM is simple, requires no architectural modifications, and adds no extra parameters, yet it yields substantial performance improvements across diverse 3D scene-language tasks. Extensive experiments across multiple benchmarks and LLM baselines validate its effectiveness and underscore the critical role of decoder design in 3D multi-modal reasoning.", "date": "2025-12-02", "categories": []}
{"id": "2603.22852", "title": "Gau-Occ: Geometry-Completed Gaussians for Multi-Modal 3D Occupancy Prediction", "abstract": "3D semantic occupancy prediction is crucial for autonomous driving. While multi-modal fusion improves accuracy over vision-only methods, it typically relies on computationally expensive dense voxel or BEV tensors. We present Gau-Occ, a multi-modal framework that bypasses dense volumetric processing by modeling the scene as a compact collection of semantic 3D Gaussians. To ensure geometric completeness, we propose a LiDAR Completion Diffuser (LCD) that recovers missing structures from sparse LiDAR to initialize robust Gaussian anchors. Furthermore, we introduce Gaussian Anchor Fusion (GAF), which efficiently integrates multi-view image semantics via geometry-aligned 2D sampling and cross-modal alignment. By refining these compact Gaussian descriptors, Gau-Occ captures both spatial consistency and semantic discriminability. Extensive experiments across challenging benchmarks demonstrate that Gau-Occ achieves state-of-the-art performance with significant computational efficiency.", "date": "2026-03-24", "categories": []}
{"id": "2603.23478", "title": "UniFunc3D: Unified Active Spatial-Temporal Grounding for 3D Functionality Segmentation", "abstract": "Functionality segmentation in 3D scenes requires an agent to ground implicit natural-language instructions into precise masks of fine-grained interactive elements. Existing methods rely on fragmented pipelines that suffer from visual blindness during initial task parsing. We observe that these methods are limited by single-scale, passive and heuristic frame selection. We present UniFunc3D, a unified and training-free framework that treats the multimodal large language model as an active observer. By consolidating semantic, temporal, and spatial reasoning into a single forward pass, UniFunc3D performs joint reasoning to ground task decomposition in direct visual evidence. Our approach introduces active spatial-temporal grounding with a coarse-to-fine strategy. This allows the model to select correct video frames adaptively and focus on high-detail interactive parts while preserving the global context necessary for disambiguation. On SceneFun3D, UniFunc3D achieves state-of-the-art performance, surpassing both training-free and training-based methods by a large margin with a relative 59.9\\% mIoU improvement, without any task-specific training. Code will be released on our project page: https://jiaying.link/unifunc3d.", "date": "2026-03-24", "categories": []}
{"id": "2404.12352", "title": "Point-In-Context: Understanding Point Cloud via In-Context Learning", "abstract": "The rise of large-scale models has catalyzed in-context learning as a powerful approach for multitasking, particularly in natural language and image processing. However, its application to 3D point cloud tasks has been largely unexplored. In this paper, we introduce Point-In-Context (PIC), a pioneering framework for 3D point cloud understanding that leverages in-context learning with a standard transformer architecture. PIC uniquely enables the execution of multiple tasks after a single, unified training phase, eliminating the need for fine-tuning. To extend masked point modeling to 3D in-context learning, we introduce a Joint Sampling module, a simple yet effective technique that emphasizes the mapping relationship between input and target. PIC treats both inputs and targets as coordinate-based, addressing the segmentation challenge by associating label points with pre-defined XYZ coordinates for each category. However, relying on such fixed label-coordinate assignments limits the model's ability to generalize to unseen domains. To address this limitation, we further propose two innovative training strategies: In-Context Labeling and In-Context Enhancing. These strategies are integrated into PIC++, which enhances dynamic in-context labeling and model training. Besides its multitask capability, PIC++ demonstrates generalization across part segmentation datasets by employing dynamic in-context labels and regular in-context pairs. Remarkably, PIC++, trained once without fine-tuning, can generalize effectively to unseen datasets and perform novel part segmentation through customized prompts. Overall, PIC is a general framework that seamlessly integrates additional tasks or datasets through a unified data format via in-context learning. Extensive experiments substantiate PIC's versatility and adaptability in handling diverse tasks and segmenting multiple datasets simultaneously.", "date": "2024-04-18", "categories": []}
{"id": "2510.01049", "title": "KeySG: Hierarchical Keyframe-Based 3D Scene Graphs", "abstract": "In recent years, 3D scene graphs have emerged as a powerful world representation, offering both geometric accuracy and semantic richness. Combining 3D scene graphs with large language models enables robots to reason, plan, and navigate in complex human-centered environments. However, current approaches for constructing 3D scene graphs are semantically limited to a predefined set of relationships, and their serialization in large environments can easily exceed an LLM's context window. We introduce KeySG, a framework that represents 3D scenes as a hierarchical graph consisting of floors, rooms, objects, and functional elements, where nodes are augmented with multi-modal information extracted from keyframes selected to optimize geometric and visual coverage. The keyframes allow us to efficiently leverage VLMs to extract scene information, alleviating the need to explicitly model relationship edges between objects, enabling more general, task-agnostic reasoning and planning. Our approach can process complex and ambiguous queries while mitigating the scalability issues associated with large scene graphs by utilizing a hierarchical multi-modal retrieval-augmented generation (RAG) pipeline to extract relevant context from the graph. Evaluated across three distinct benchmarks, 3D object semantic segmentation, functional element segmentation, and complex query retrieval, KeySG outperforms prior approaches on most metrics, demonstrating its superior semantic richness and efficiency.", "date": "2025-10-01", "categories": []}
{"id": "2603.20239", "title": "Rheos: Modelling Continuous Motion Dynamics in Hierarchical 3D Scene Graphs", "abstract": "3D Scene Graphs (3DSGs) provide hierarchical, multi-resolution abstractions that encode the geometric and semantic structure of an environment, yet their treatment of dynamics remains limited to tracking individual agents. Maps of Dynamics (MoDs) complement this by modeling aggregate motion patterns, but rely on uniform grid discretizations that lack semantic grounding and scale poorly. We present Rheos, a framework that explicitly embeds continuous directional motion models into an additional dynamics layer of a hierarchical 3DSG that enhances the navigational properties of the graph. Each dynamics node maintains a semi-wrapped Gaussian mixture model that captures multimodal directional flow as a principled probability distribution with explicit uncertainty, replacing the discrete histograms used in prior work. To enable online operation, Rheos employs reservoir sampling for bounded-memory observation buffers, parallel per-cell model updates and a principled Bayesian Information Criterion (BIC) sweep that selects the optimal number of mixture components, reducing per-update initialization cost from quadratic to linear in the number of samples. Evaluated across four spatial resolutions in a simulated pedestrian environment, Rheos consistently outperforms the discrete baseline under continuous as well as unfavorable discrete metrics. We release our implementation as open source.", "date": "2026-03-09", "categories": []}
{"id": "2603.21166", "title": "Training-Free Instance-Aware 3D Scene Reconstruction and Diffusion-Based View Synthesis from Sparse Images", "abstract": "We introduce a novel, training-free system for reconstructing, understanding, and rendering 3D indoor scenes from a sparse set of unposed RGB images. Unlike traditional radiance field approaches that require dense views and per-scene optimization, our pipeline achieves high-fidelity results without any training or pose preprocessing. The system integrates three key innovations: (1) A robust point cloud reconstruction module that filters unreliable geometry using a warping-based anomaly removal strategy; (2) A warping-guided 2D-to-3D instance lifting mechanism that propagates 2D segmentation masks into a consistent, instance-aware 3D representation; and (3) A novel rendering approach that projects the point cloud into new views and refines the renderings with a 3D-aware diffusion model. Our method leverages the generative power of diffusion to compensate for missing geometry and enhances realism, especially under sparse input conditions. We further demonstrate that object-level scene editing such as instance removal can be naturally supported in our pipeline by modifying only the point cloud, enabling the synthesis of consistent, edited views without retraining. Our results establish a new direction for efficient, editable 3D content generation without relying on scene-specific optimization. Project page: https://jiatongxia.github.io/TID3R/", "date": "2026-03-22", "categories": []}
{"id": "2603.21487", "title": "GaussianSSC: Triplane-Guided Directional Gaussian Fields for 3D Semantic Completion", "abstract": "We present \\emph{GaussianSSC}, a two-stage, grid-native and triplane-guided approach to semantic scene completion (SSC) that injects the benefits of Gaussians without replacing the voxel grid or maintaining a separate Gaussian set. We introduce \\emph{Gaussian Anchoring}, a sub-pixel, Gaussian-weighted image aggregation over fused FPN features that tightens voxel--image alignment and improves monocular occupancy estimation. We further convert point-like voxel features into a learned per-voxel Gaussian field and refine triplane features via a triplane-aligned \\emph{Gaussian--Triplane Refinement} module that combines \\emph{local gathering} (target-centric) and \\emph{global aggregation} (source-centric). This directional, anisotropic support captures surface tangency, scale, and occlusion-aware asymmetry while preserving the efficiency of triplane representations. On SemanticKITTI~\\cite{behley2019semantickitti}, GaussianSSC improves Stage~1 occupancy by +1.0\\% Recall, +2.0\\% Precision, and +1.8\\% IoU over state-of-the-art baselines, and improves Stage~2 semantic prediction by +1.8\\% IoU and +0.8\\% mIoU.", "date": "2026-03-23", "categories": []}
{"id": "2511.03992", "title": "Camera-Aware Cross-View Alignment for Referring 3D Gaussian Splatting Segmentation", "abstract": "Referring 3D Gaussian Splatting Segmentation (R3DGS) aims to ground free-form language queries in 3D Gaussian fields. However, existing methods rely on single-view pseudo supervision, leading to viewpoint drift and inconsistent predictions across views. We propose CaRF (Camera-aware Referring Field), a camera-aware cross-view alignment framework for view-consistent referring in 3D Gaussian splatting. CaRF introduces Camera-conditioned Alignment Modulation (CAM) to inject camera geometry into Gaussian-text interactions, and Gaussian-level Cross-view Logit Alignment (GCLA) to explicitly align referring responses of the same Gaussians across calibrated views during training. By turning cross-view discrepancy into an optimizable objective, CaRF enables geometry-aware and view-consistent reasoning directly in the Gaussian space. Extensive experiments on three benchmarks demonstrate that CaRF achieves state-of-the-art performance, improving mIoU by 16.8%, 4.3%, and 2.0% on Ref-LERF, LERF-OVS, and 3D-OVS, respectively. Our code is available at https://github.com/eR3R3/CaRF.", "date": "2025-11-06", "categories": []}
{"id": "2504.00992", "title": "SuperDec: 3D Scene Decomposition with Superquadric Primitives", "abstract": "We present SuperDec, an approach for creating compact 3D scene representations via decomposition into superquadric primitives. While most recent works leverage geometric primitives to obtain photorealistic 3D scene representations, we propose to leverage them to obtain a compact yet expressive representation. We propose to solve the problem locally on individual objects and leverage the capabilities of instance segmentation methods to scale our solution to full 3D scenes. In doing that, we design a new architecture which efficiently decompose point clouds of arbitrary objects in a compact set of superquadrics. We train our architecture on ShapeNet and we prove its generalization capabilities on object instances extracted from the ScanNet++ dataset as well as on full Replica scenes. Finally, we show how a compact representation based on superquadrics can be useful for a diverse range of downstream applications, including robotic tasks and controllable visual content generation and editing.", "date": "2025-04-01", "categories": []}
{"id": "2510.04714", "title": "Object-Centric Representation Learning for Enhanced 3D Semantic Scene Graph Prediction", "abstract": "3D Semantic Scene Graph Prediction aims to detect objects and their semantic relationships in 3D scenes, and has emerged as a crucial technology for robotics and AR/VR applications. While previous research has addressed dataset limitations and explored various approaches including Open-Vocabulary settings, they frequently fail to optimize the representational capacity of object and relationship features, showing excessive reliance on Graph Neural Networks despite insufficient discriminative capability. In this work, we demonstrate through extensive analysis that the quality of object features plays a critical role in determining overall scene graph accuracy. To address this challenge, we design a highly discriminative object feature encoder and employ a contrastive pretraining strategy that decouples object representation learning from the scene graph prediction. This design not only enhances object classification accuracy but also yields direct improvements in relationship prediction. Notably, when plugging in our pretrained encoder into existing frameworks, we observe substantial performance improvements across all evaluation metrics. Additionally, whereas existing approaches have not fully exploited the integration of relationship information, we effectively combine both geometric and semantic features to achieve superior relationship prediction. Comprehensive experiments on the 3DSSG dataset demonstrate that our approach significantly outperforms previous state-of-the-art methods. Our code is publicly available at https://github.com/VisualScienceLab-KHU/OCRL-3DSSG-Codes.", "date": "2025-10-06", "categories": []}
{"id": "2603.18510", "title": "OnlinePG: Online Open-Vocabulary Panoptic Mapping with 3D Gaussian Splatting", "abstract": "Open-vocabulary scene understanding with online panoptic mapping is essential for embodied applications to perceive and interact with environments. However, existing methods are predominantly offline or lack instance-level understanding, limiting their applicability to real-world robotic tasks. In this paper, we propose OnlinePG, a novel and effective system that integrates geometric reconstruction and open-vocabulary perception using 3D Gaussian Splatting in an online setting. Technically, to achieve online panoptic mapping, we employ an efficient local-to-global paradigm with a sliding window. To build local consistency map, we construct a 3D segment clustering graph that jointly leverages geometric and semantic cues, fusing inconsistent segments within sliding window into complete instances. Subsequently, to update the global map, we construct explicit grids with spatial attributes for the local 3D Gaussian map and fuse them into the global map via robust bidirectional bipartite 3D Gaussian instance matching. Finally, we utilize the fused VLM features inside the 3D spatial attribute grids to achieve open-vocabulary scene understanding. Extensive experiments on widely used datasets demonstrate that our method achieves better performance among online approaches, while maintaining real-time efficiency.", "date": "2026-03-19", "categories": []}
{"id": "2512.11903", "title": "Aion: Towards Hierarchical 4D Scene Graphs with Temporal Flow Dynamics", "abstract": "Autonomous navigation in dynamic environments requires spatial representations that capture both semantic structure and temporal evolution. 3D Scene Graphs (3DSGs) provide hierarchical multi-resolution abstractions that encode geometry and semantics, but existing extensions toward dynamics largely focus on individual objects or agents. In parallel, Maps of Dynamics (MoDs) model typical motion patterns and temporal regularities, yet are usually tied to grid-based discretizations that lack semantic awareness and do not scale well to large environments. In this paper we introduce Aion, a framework that embeds temporal flow dynamics directly within a hierarchical 3DSG, effectively incorporating the temporal dimension. Aion employs a graph-based sparse MoD representation to capture motion flows over arbitrary time intervals and attaches them to navigational nodes in the scene graph, yielding more interpretable and scalable predictions that improve planning and interaction in complex dynamic environments. We provide the code at https://github.com/IacopomC/aion", "date": "2025-12-10", "categories": []}
{"id": "2603.17583", "title": "Edit-As-Act: Goal-Regressive Planning for Open-Vocabulary 3D Indoor Scene Editing", "abstract": "Editing a 3D indoor scene from natural language is conceptually straightforward but technically challenging. Existing open-vocabulary systems often regenerate large portions of a scene or rely on image-space edits that disrupt spatial structure, resulting in unintended global changes or physically inconsistent layouts. These limitations stem from treating editing primarily as a generative task. We take a different view. A user instruction defines a desired world state, and editing should be the minimal sequence of actions that makes this state true while preserving everything else. This perspective motivates Edit-As-Act, a framework that performs open-vocabulary scene editing as goal-regressive planning in 3D space. Given a source scene and free-form instruction, Edit-As-Act predicts symbolic goal predicates and plans in EditLang, a PDDL-inspired action language that we design with explicit preconditions and effects encoding support, contact, collision, and other geometric relations. A language-driven planner proposes actions, and a validator enforces goal-directedness, monotonicity, and physical feasibility, producing interpretable and physically coherent transformations. By separating reasoning from low-level generation, Edit-As-Act achieves instruction fidelity, semantic consistency, and physical plausibility - three criteria that existing paradigms cannot satisfy together. On E2A-Bench, our benchmark of 63 editing tasks across 9 indoor environments, Edit-As-Act significantly outperforms prior approaches across all edit types and scene categories.", "date": "2026-03-18", "categories": []}
{"id": "2603.17605", "title": "ReLaGS: Relational Language Gaussian Splatting", "abstract": "Achieving unified 3D perception and reasoning across tasks such as segmentation, retrieval, and relation understanding remains challenging, as existing methods are either object-centric or rely on costly training for inter-object reasoning. We present a novel framework that constructs a hierarchical language-distilled Gaussian scene and its 3D semantic scene graph without scene-specific training. A Gaussian pruning mechanism refines scene geometry, while a robust multi-view language alignment strategy aggregates noisy 2D features into accurate 3D object embeddings. On top of this hierarchy, we build an open-vocabulary 3D scene graph with Vision Language derived annotations and Graph Neural Network-based relational reasoning. Our approach enables efficient and scalable open-vocabulary 3D reasoning by jointly modeling hierarchical semantics and inter/intra-object relationships, validated across tasks including open-vocabulary segmentation, scene graph generation, and relation-guided retrieval. Project page: https://dfki-av.github.io/ReLaGS/", "date": "2026-03-18", "categories": []}
{"id": "2603.17647", "title": "Part-Aware Open-Vocabulary 3D Affordance Grounding via Prototypical Semantic and Geometric Alignment", "abstract": "Grounding natural language questions to functionally relevant regions in 3D objects -- termed language-driven 3D affordance grounding -- is essential for embodied intelligence and human-AI interaction. Existing methods, while progressing from label-based to language-driven approaches, still face challenges in open-vocabulary generalization, fine-grained geometric alignment, and part-level semantic consistency. To address these issues, we propose a novel two-stage cross-modal framework that enhances both semantic and geometric representations for open-vocabulary 3D affordance grounding. In the first stage, large language models generate part-aware instructions to recover missing semantics, enabling the model to link semantically similar affordances. In the second stage, we introduce two key components: Affordance Prototype Aggregation (APA), which captures cross-object geometric consistency for each affordance, and Intra-Object Relational Modeling (IORM), which refines geometric differentiation within objects to support precise semantic alignment. We validate the effectiveness of our method through extensive experiments on a newly introduced benchmark, as well as two existing benchmarks, demonstrating superior performance in comparison with existing methods.", "date": "2026-03-18", "categories": []}
{"id": "2411.16253", "title": "Open-Vocabulary Octree-Graph for 3D Scene Understanding", "abstract": "Open-vocabulary 3D scene understanding is indispensable for embodied agents. Recent works leverage pretrained vision-language models (VLMs) for object segmentation and project them to point clouds to build 3D maps. Despite progress, a point cloud is a set of unordered coordinates that requires substantial storage space and does not directly convey occupancy information or spatial relation, making existing methods inefficient for downstream tasks, e.g., path planning and text-based object retrieval. To address these issues, we propose \\textbf{Octree-Graph}, a novel scene representation for open-vocabulary 3D scene understanding. Specifically, a Chronological Group-wise Segment Merging (CGSM) strategy and an Instance Feature Aggregation (IFA) algorithm are first designed to get 3D instances and corresponding semantic features. Subsequently, an adaptive-octree structure is developed that stores semantics and depicts the occupancy of an object adjustably according to its shape. Finally, the Octree-Graph is constructed where each adaptive-octree acts as a graph node, and edges describe the spatial relations among nodes. Extensive experiments on various tasks are conducted on several widely-used datasets, demonstrating the versatility and effectiveness of our method. Code is available \\href{https://github.com/yifeisu/OV-Octree-Graph}{here}.", "date": "2024-11-25", "categories": []}
{"id": "2603.16868", "title": "MessyKitchens: Contact-rich object-level 3D scene reconstruction", "abstract": "Monocular 3D scene reconstruction has recently seen significant progress. Powered by the modern neural architectures and large-scale data, recent methods achieve high performance in depth estimation from a single image. Meanwhile, reconstructing and decomposing common scenes into individual 3D objects remains a hard challenge due to the large variety of objects, frequent occlusions and complex object relations. Notably, beyond shape and pose estimation of individual objects, applications in robotics and animation require physically-plausible scene reconstruction where objects obey physical principles of non-penetration and realistic contacts. In this work we advance object-level scene reconstruction along two directions. First, we introduceMessyKitchens, a new dataset with real-world scenes featuring cluttered environments and providing high-fidelity object-level ground truth in terms of 3D object shapes, poses and accurate object contacts. Second, we build on the recent SAM 3D approach for single-object reconstruction and extend it with Multi-Object Decoder (MOD) for joint object-level scene reconstruction. To validate our contributions, we demonstrate MessyKitchens to significantly improve previous datasets in registration accuracy and inter-object penetration. We also compare our multi-object reconstruction approach on three datasets and demonstrate consistent and significant improvements of MOD over the state of the art. Our new benchmark, code and pre-trained models will become publicly available on our project website: https://messykitchens.github.io/.", "date": "2026-03-17", "categories": []}
{"id": "2511.03571", "title": "OneOcc: Semantic Occupancy Prediction for Legged Robots with a Single Panoramic Camera", "abstract": "Robust 3D semantic occupancy is crucial for legged/humanoid robots, yet most semantic scene completion (SSC) systems target wheeled platforms with forward-facing sensors. We present OneOcc, a vision-only panoramic SSC framework designed for gait-introduced body jitter and 360{\\deg} continuity. OneOcc combines: (i) Dual-Projection fusion (DP-ER) to exploit the annular panorama and its equirectangular unfolding, preserving 360{\\deg} continuity and grid alignment; (ii) Bi-Grid Voxelization (BGV) to reason in Cartesian and cylindrical-polar spaces, reducing discretization bias and sharpening free/occupied boundaries; (iii) a lightweight decoder with Hierarchical AMoE-3D for dynamic multi-scale fusion and better long-range/occlusion reasoning; and (iv) plug-and-play Gait Displacement Compensation (GDC) learning feature-level motion correction without extra sensors. We also release two panoramic occupancy benchmarks: QuadOcc (real quadruped, first-person 360{\\deg}) and Human360Occ (H3O) (CARLA human-ego 360{\\deg} with RGB, Depth, semantic occupancy; standardized within-/cross-city splits). OneOcc sets a new state of the art on QuadOcc, outperforming strong vision baselines and remaining competitive with classical LiDAR baselines; on H3O it gains +3.83 mIoU (within-city) and +8.08 (cross-city). Modules are lightweight, enabling deployable full-surround perception for legged/humanoid robots. Datasets and code will be publicly available at https://github.com/MasterHow/OneOcc.", "date": "2025-11-05", "categories": []}
{"id": "2601.10168", "title": "RAG-3DSG: Enhancing 3D Scene Graphs with Re-Shot Guided Retrieval-Augmented Generation", "abstract": "Open-vocabulary 3D Scene Graph (3DSG) can enhance various downstream tasks in robotics by leveraging structured semantic representations, yet current 3DSG construction methods suffer from semantic inconsistencies caused by noisy cross-image aggregation under occlusions and constrained viewpoints. To mitigate the impact of such inconsistency, we propose RAG-3DSG, which introduces re-shot guided uncertainty estimation. By measuring the semantic consistency between original limited viewpoints and re-shot optimal viewpoints, this method quantifies the underlying semantic ambiguity of each graph object. Based on this quantification, we devise an Object-level Retrieval-Augmented Generation (RAG) that leverages low-uncertainty objects as semantic anchors to retrieve more reliable contextual knowledge, enabling a Vision-Language Model to rectify the predictions of uncertain objects and optimize the final 3DSG. Extensive evaluations across three challenging benchmarks and real-world robot trials demonstrate that RAG-3DSG achieves superior recall and precision, effectively mitigating semantic noise to provide highly reliable scene representations for robotics tasks.", "date": "2026-01-15", "categories": []}
{"id": "2603.13910", "title": "Scene Generation at Absolute Scale: Utilizing Semantic and Geometric Guidance From Text for Accurate and Interpretable 3D Indoor Scene Generation", "abstract": "We present GuidedSceneGen, a text-to-3D generation framework that produces metrically accurate, globally consistent, and semantically interpretable indoor scenes. Unlike prior text-driven methods that often suffer from geometric drift or scale ambiguity, our approach maintains an absolute world coordinate frame throughout the entire generation process. Starting from a textual scene description, we predict a global 3D layout encoding both semantic and geometric structure, which serves as a guiding proxy for downstream stages. A semantics- and depth-conditioned panoramic diffusion model then synthesizes 360{\\deg} imagery aligned with the global layout, substantially improving spatial coherence. To explore unobserved regions, we employ a video diffusion model guided by optimized camera trajectories that balances coverage and collision avoidance, achieving up to 10x faster sampling compared to exhaustive path exploration. The generated views are fused using 3D Gaussian Splatting, yielding a consistent and fully navigable 3D scene in absolute scale. GuidedSceneGen enables accurate transfer of object poses and semantic labels from layout to reconstruction, and supports progressive scene expansion without re-alignment. Quantitative results and a user study demonstrate greater 3D consistency and layout plausibility compared to recent panoramic text-to-3D baselines.", "date": "2026-03-14", "categories": []}
{"id": "2603.14076", "title": "SGR-OCC: Evolving Monocular Priors for Embodied 3D Occupancy Prediction via Soft-Gating Lifting and Semantic-Adaptive Geometric Refinement", "abstract": "3D semantic occupancy prediction is a cornerstone for embodied AI, enabling agents to perceive dense scene geometry and semantics incrementally from monocular video streams. However, current online frameworks face two critical bottlenecks: the inherent depth ambiguity of monocular estimation that causes \"feature bleeding\" at object boundaries , and the \"cold start\" instability where uninitialized temporal fusion layers distort high-quality spatial priors during early training stages. In this paper, we propose SGR-OCC (Soft-Gating and Ray-refinement Occupancy), a unified framework driven by the philosophy of \"Inheritance and Evolution\". To perfectly inherit monocular spatial expertise, we introduce a Soft-Gating Feature Lifter that explicitly models depth uncertainty via a Gaussian gate to probabilistically suppress background noise. Furthermore, a Dynamic Ray-Constrained Anchor Refinement module simplifies complex 3D displacement searches into efficient 1D depth corrections along camera rays, ensuring sub-voxel adherence to physical surfaces. To ensure stable evolution toward temporal consistency, we employ a Two-Phase Progressive Training Strategy equipped with identity-initialized fusion, effectively resolving the cold start problem and shielding spatial priors from noisy early gradients. Extensive experiments on the EmbodiedOcc-ScanNet and Occ-ScanNet benchmarks demonstrate that SGR-OCC achieves state-of-the-art performance. In local prediction tasks, SGR-OCC achieves a completion IoU of 58.55$\\%$ and a semantic mIoU of 49.89$\\%$, surpassing the previous best method, EmbodiedOcc++, by 3.65$\\%$ and 3.69$\\%$ respectively. In challenging embodied prediction tasks, our model reaches 55.72$\\%$ SC-IoU and 46.22$\\%$ mIoU. Qualitative results further confirm our model's superior capability in preserving structural integrity and boundary sharpness in complex indoor environments.", "date": "2026-03-14", "categories": []}
{"id": "2603.14232", "title": "S2GS: Streaming Semantic Gaussian Splatting for Online Scene Understanding and Reconstruction", "abstract": "Existing offline feed-forward methods for joint scene understanding and reconstruction on long image streams often repeatedly perform global computation over an ever-growing set of past observations, causing runtime and GPU memory to increase rapidly with sequence length and limiting scalability. We propose Streaming Semantic Gaussian Splatting (S2GS), a strictly causal, incremental 3D Gaussian semantic field framework: it does not leverage future frames and continuously updates scene geometry, appearance, and instance-level semantics without reprocessing historical frames, enabling scalable online joint reconstruction and understanding. S2GS adopts a geometry-semantic decoupled dual-backbone design: the geometry branch performs causal modeling to drive incremental Gaussian updates, while the semantic branch leverages a 2D foundation vision model and a query-driven decoder to predict segmentation masks and identity embeddings, further stabilized by query-level contrastive alignment and lightweight online association with an instance memory. Experiments show that S2GS matches or outperforms strong offline baselines on joint reconstruction-and-understanding benchmarks, while significantly improving long-horizon scalability: it processes 1,000+ frames with much slower growth in runtime and GPU memory, whereas offline global-processing baselines typically run out of memory at around 80 frames under the same setting.", "date": "2026-03-15", "categories": []}
{"id": "2603.14301", "title": "4D Synchronized Fields: Motion-Language Gaussian Splatting for Temporal Scene Understanding", "abstract": "Current 4D representations decouple geometry, motion, and semantics: reconstruction methods discard interpretable motion structure; language-grounded methods attach semantics after motion is learned, blind to how objects move; and motion-aware methods encode dynamics as opaque per-point residuals without object-level organization. We propose 4D Synchronized Fields, a 4D Gaussian representation that learns object-factored motion in-loop during reconstruction and synchronizes language to the resulting kinematics through a per-object conditioned field. Each Gaussian trajectory is decomposed into shared object motion plus an implicit residual, and a kinematic-conditioned ridge map predicts temporal semantic variation, yielding a single representation in which reconstruction, motion, and semantics are structurally coupled and enabling open-vocabulary temporal queries that retrieve both objects and moments. On HyperNeRF, 4D Synchronized Fields achieves 28.52 dB mean PSNR, the highest among all language-grounded and motion-aware baselines, within 1.5 dB of reconstruction-only methods. On targeted temporal-state retrieval, the kinematic-conditioned field attains 0.884 mean accuracy, 0.815 mean vIoU, and 0.733 mean tIoU, surpassing 4D LangSplat (0.620, 0.433, and 0.439 respectively) and LangSplat (0.415, 0.304, and 0.262). Ablation confirms that kinematic conditioning is the primary driver, accounting for +0.45 tIoU over a static-embedding-only baseline. 4D Synchronized Fields is the only method that jointly exposes interpretable motion primitives and temporally grounded language fields from a single trained representation. Code will be released.", "date": "2026-03-15", "categories": []}
{"id": "2511.17361", "title": "SuperQuadricOcc: Real-Time Self-Supervised Semantic Occupancy Estimation with Superquadric Volume Rendering", "abstract": "Self-supervision for semantic occupancy estimation is appealing as it removes the labour-intensive manual annotation, thus allowing one to scale to larger autonomous driving datasets. Superquadrics offer an expressive shape family very suitable for this task, yet their deployment in a self-supervised setting has been hindered by the lack of efficient rendering methods to bridge the 3D scene representation and 2D training pseudo-labels. To address this, we introduce SuperQuadricOcc, the first self-supervised occupancy model to leverage superquadrics for scene representation. To overcome the rendering limitation, we propose a real-time volume renderer that preserves the fidelity of the superquadric shape during rendering. It relies on spatial superquadric-voxel indexing, restricting each ray sample to query only nearby superquadrics, thereby greatly reducing memory usage and computational cost. Using drastically fewer primitives than previous Gaussian-based methods, SuperQuadricOcc achieves state-of-the-art performance on the Occ3D-nuScenes dataset, while running at real-time inference speeds with substantially reduced memory footprint.", "date": "2025-11-21", "categories": []}
{"id": "2603.12599", "title": "A Prediction-as-Perception Framework for 3D Object Detection", "abstract": "Humans combine prediction and perception to observe the world. When faced with rapidly moving birds or insects, we can only perceive them clearly by predicting their next position and focusing our gaze there. Inspired by this, this paper proposes the Prediction-As-Perception (PAP) framework, integrating a prediction-perception architecture into 3D object perception tasks to enhance the model's perceptual accuracy. The PAP framework consists of two main modules: prediction and perception, primarily utilizing continuous frame information as input. Firstly, the prediction module forecasts the potential future positions of ego vehicles and surrounding traffic participants based on the perception results of the current frame. These predicted positions are then passed as queries to the perception module of the subsequent frame. The perceived results are iteratively fed back into the prediction module. We evaluated the PAP structure using the end-to-end model UniAD on the nuScenes dataset. The results demonstrate that the PAP structure improves UniAD's target tracking accuracy by 10% and increases the inference speed by 15%. This indicates that such a biomimetic design significantly enhances the efficiency and accuracy of perception models while reducing computational resource consumption.", "date": "2026-03-13", "categories": []}
{"id": "2603.12663", "title": "Learning Geometric and Photometric Features from Panoramic LiDAR Scans for Outdoor Place Categorization", "abstract": "Semantic place categorization, which is one of the essential tasks for autonomous robots and vehicles, allows them to have capabilities of self-decision and navigation in unfamiliar environments. In particular, outdoor places are more difficult targets than indoor ones due to perceptual variations, such as dynamic illuminance over twenty-four hours and occlusions by cars and pedestrians. This paper presents a novel method of categorizing outdoor places using convolutional neural networks (CNNs), which take omnidirectional depth/reflectance images obtained by 3D LiDARs as the inputs. First, we construct a large-scale outdoor place dataset named Multi-modal Panoramic 3D Outdoor (MPO) comprising two types of point clouds captured by two different LiDARs. They are labeled with six outdoor place categories: coast, forest, indoor/outdoor parking, residential area, and urban area. Second, we provide CNNs for LiDAR-based outdoor place categorization and evaluate our approach with the MPO dataset. Our results on the MPO dataset outperform traditional approaches and show the effectiveness in which we use both depth and reflectance modalities. To analyze our trained deep networks we visualize the learned features.", "date": "2026-03-13", "categories": []}
{"id": "2510.06754", "title": "UniFField: A Generalizable Unified Neural Feature Field for Visual, Semantic, and Spatial Uncertainties in Any Scene", "abstract": "Comprehensive visual, geometric, and semantic understanding of a 3D scene is crucial for successful execution of robotic tasks, especially in unstructured and complex environments. Additionally, to make robust decisions, it is necessary for the robot to evaluate the reliability of perceived information. While recent advances in 3D neural feature fields have enabled robots to leverage features from pretrained foundation models for tasks such as language-guided manipulation and navigation, existing methods suffer from two critical limitations: (i) they are typically scene-specific, and (ii) they lack the ability to model uncertainty in their predictions. We present UniFField, a unified uncertainty-aware neural feature field that combines visual, semantic, and geometric features in a single generalizable representation while also predicting uncertainty in each modality. Our approach, which can be applied zero shot to any new environment, incrementally integrates RGB-D images into our voxel-based feature representation as the robot explores the scene, simultaneously updating uncertainty estimation. We evaluate our uncertainty estimations to accurately describe the model prediction errors in scene reconstruction and semantic feature prediction. Furthermore, we successfully leverage our feature predictions and their respective uncertainty for an active object search task using a mobile manipulator robot, demonstrating the capability for robust decision-making.", "date": "2025-10-08", "categories": []}
{"id": "2603.12144", "title": "O3N: Omnidirectional Open-Vocabulary Occupancy Prediction", "abstract": "Understanding and reconstructing the 3D world through omnidirectional perception is an inevitable trend in the development of autonomous agents and embodied intelligence. However, existing 3D occupancy prediction methods are constrained by limited perspective inputs and predefined training distribution, making them difficult to apply to embodied agents that require comprehensive and safe perception of scenes in open world exploration. To address this, we present O3N, the first purely visual, end-to-end Omnidirectional Open-vocabulary Occupancy predictioN framework. O3N embeds omnidirectional voxels in a polar-spiral topology via the Polar-spiral Mamba (PsM) module, enabling continuous spatial representation and long-range context modeling across 360{\\deg}. The Occupancy Cost Aggregation (OCA) module introduces a principled mechanism for unifying geometric and semantic supervision within the voxel space, ensuring consistency between the reconstructed geometry and the underlying semantic structure. Moreover, Natural Modality Alignment (NMA) establishes a gradient-free alignment pathway that harmonizes visual features, voxel embeddings, and text semantics, forming a consistent \"pixel-voxel-text\" representation triad. Extensive experiments on multiple models demonstrate that our method not only achieves state-of-the-art performance on QuadOcc and Human360Occ benchmarks but also exhibits remarkable cross-scene generalization and semantic scalability, paving the way toward universal 3D world modeling. The source code will be made publicly available at https://github.com/MengfeiD/O3N.", "date": "2026-03-12", "categories": []}
{"id": "2510.16410", "title": "REALM: An MLLM-Agent Framework for Open World 3D Reasoning Segmentation and Editing on Gaussian Splatting", "abstract": "Bridging the gap between complex human instructions and precise 3D object grounding remains a significant challenge in vision and robotics. Existing 3D segmentation methods often struggle to interpret ambiguous, reasoning-based instructions, while 2D vision-language models that excel at such reasoning lack intrinsic 3D spatial understanding. In this paper, we introduce REALM, an innovative MLLM-agent framework that enables open-world reasoning-based segmentation without requiring extensive 3D-specific post-training. We perform segmentation directly on 3D Gaussian Splatting representations, capitalizing on their ability to render photorealistic novel views that are highly suitable for MLLM comprehension. As directly feeding one or more rendered views to the MLLM can lead to high sensitivity to viewpoint selection, we propose a novel Global-to-Local Spatial Grounding strategy. Specifically, multiple global views are first fed into the MLLM agent in parallel for coarse-level localization, aggregating responses to robustly identify the target object. Then, several close-up novel views of the object are synthesized to perform fine-grained local segmentation, yielding accurate and consistent 3D masks. Extensive experiments show that REALM achieves remarkable performance in interpreting both explicit and implicit instructions across LERF, 3D-OVS, and our newly introduced REALM3D benchmarks. Furthermore, our agent framework seamlessly supports a range of 3D interaction tasks, including object removal, replacement, and style transfer, demonstrating its practical utility and versatility. Project page: https://ChangyueShi.github.io/REALM.", "date": "2025-10-18", "categories": []}
{"id": "2603.09173", "title": "Point Cloud as a Foreign Language for Multi-modal Large Language Model", "abstract": "Multi-modal large language models (MLLMs) have shown remarkable progress in integrating visual and linguistic understanding. Recent efforts have extended these capabilities to 3D understanding through encoder-based architectures that rely on pre-trained 3D encoders to extract geometric features. However, such approaches suffer from semantic misalignment between geometric and linguistic spaces, resolution sensitivity, and substantial computational overhead. In this work, we present SAGE, the first end-to-end 3D MLLM that directly processes raw point clouds without relying on a pre-trained 3D encoder. Our approach introduces a lightweight 3D tokenizer that combines geometric sampling and neighbourhood aggregation with vector quantization to convert point clouds into discrete tokens--treating 3D data as a foreign language that naturally extends the LLM's vocabulary. Furthermore, to enhance the model's reasoning capability on complex 3D tasks, we propose a preference optimization training strategy with a semantic alignment-based reward, specifically designed for open-ended 3D question answering where responses are descriptive. Extensive experiments across diverse 3D understanding benchmarks demonstrate that our end-to-end approach outperforms existing encoder-based methods while offering significant advantages in computational efficiency, generalization across LLM backbones, and robustness to input resolution variations. Code is available at: github.com/snehaputul/SAGE3D.", "date": "2026-03-10", "categories": []}
{"id": "2603.09737", "title": "$M^2$-Occ: Resilient 3D Semantic Occupancy Prediction for Autonomous Driving with Incomplete Camera Inputs", "abstract": "Semantic occupancy prediction enables dense 3D geometric and semantic understanding for autonomous driving. However, existing camera-based approaches implicitly assume complete surround-view observations, an assumption that rarely holds in real-world deployment due to occlusion, hardware malfunction, or communication failures. We study semantic occupancy prediction under incomplete multi-camera inputs and introduce $M^2$-Occ, a framework designed to preserve geometric structure and semantic coherence when views are missing. $M^2$-Occ addresses two complementary challenges. First, a Multi-view Masked Reconstruction (MMR) module leverages the spatial overlap among neighboring cameras to recover missing-view representations directly in the feature space. Second, a Feature Memory Module (FMM) introduces a learnable memory bank that stores class-level semantic prototypes. By retrieving and integrating these global priors, the FMM refines ambiguous voxel features, ensuring semantic consistency even when observational evidence is incomplete. We introduce a systematic missing-view evaluation protocol on the nuScenes-based SurroundOcc benchmark, encompassing both deterministic single-view failures and stochastic multi-view dropout scenarios. Under the safety-critical missing back-view setting, $M^2$-Occ improves the IoU by 4.93%. As the number of missing cameras increases, the robustness gap further widens; for instance, under the setting with five missing views, our method boosts the IoU by 5.01%. These gains are achieved without compromising full-view performance. The source code will be publicly released at https://github.com/qixi7up/M2-Occ.", "date": "2026-03-10", "categories": []}
{"id": "2412.17635", "title": "LangSurf: Language-Embedded Surface Gaussians for 3D Scene Understanding", "abstract": "Applying Gaussian Splatting to perception tasks for 3D scene understanding is becoming increasingly popular. Most existing works primarily focus on rendering 2D feature maps from novel viewpoints, which leads to an imprecise 3D language field with outlier languages, ultimately failing to align objects in 3D space. By utilizing masked images for feature extraction, these approaches also lack essential contextual information, leading to inaccurate feature representation. To this end, we propose a Language-Embedded Surface Field (LangSurf), which accurately aligns the 3D language fields with the surface of objects, facilitating precise 2D and 3D segmentation with text query, widely expanding the downstream tasks such as removal and editing. The core of LangSurf is a joint training strategy that flattens the language Gaussian on the object surfaces using geometry supervision and contrastive losses to assign accurate language features to the Gaussians of objects. In addition, we also introduce the Hierarchical-Context Awareness Module to extract features at the image level for contextual information then perform hierarchical mask pooling using masks segmented by SAM to obtain fine-grained language features in different hierarchies. Extensive experiments on open-vocabulary 2D and 3D semantic segmentation demonstrate that LangSurf outperforms the previous state-of-the-art method LangSplat by a large margin. As shown in Fig. 1, our method is capable of segmenting objects in 3D space, thus boosting the effectiveness of our approach in instance recognition, removal, and editing, which is also supported by comprehensive experiments. https://langsurf.github.io.", "date": "2024-12-23", "categories": []}
{"id": "2603.07570", "title": "Efficient RGB-D Scene Understanding via Multi-task Adaptive Learning and Cross-dimensional Feature Guidance", "abstract": "Scene understanding plays a critical role in enabling intelligence and autonomy in robotic systems. Traditional approaches often face challenges, including occlusions, ambiguous boundaries, and the inability to adapt attention based on task-specific requirements and sample variations. To address these limitations, this paper presents an efficient RGB-D scene understanding model that performs a range of tasks, including semantic segmentation, instance segmentation, orientation estimation, panoptic segmentation, and scene classification. The proposed model incorporates an enhanced fusion encoder, which effectively leverages redundant information from both RGB and depth inputs. For semantic segmentation, we introduce normalized focus channel layers and a context feature interaction layer, designed to mitigate issues such as shallow feature misguidance and insufficient local-global feature representation. The instance segmentation task benefits from a non-bottleneck 1D structure, which achieves superior contour representation with fewer parameters. Additionally, we propose a multi-task adaptive loss function that dynamically adjusts the learning strategy for different tasks based on scene variations. Extensive experiments on the NYUv2, SUN RGB-D, and Cityscapes datasets demonstrate that our approach outperforms existing methods in both segmentation accuracy and processing speed.", "date": "2026-03-08", "categories": []}
{"id": "2603.07660", "title": "Holi-Spatial: Evolving Video Streams into Holistic 3D Spatial Intelligence", "abstract": "The pursuit of spatial intelligence fundamentally relies on access to large-scale, fine-grained 3D data. However, existing approaches predominantly construct spatial understanding benchmarks by generating question-answer (QA) pairs from a limited number of manually annotated datasets, rather than systematically annotating new large-scale 3D scenes from raw web data. As a result, their scalability is severely constrained, and model performance is further hindered by domain gaps inherent in these narrowly curated datasets.\n  In this work, we propose Holi-Spatial, the first fully automated, large-scale, spatially-aware multimodal dataset, constructed from raw video inputs without human intervention, using the proposed data curation pipeline. Holi-Spatial supports multi-level spatial supervision, ranging from geometrically accurate 3D Gaussian Splatting (3DGS) reconstructions with rendered depth maps to object-level and relational semantic annotations, together with corresponding spatial Question-Answer (QA) pairs.\n  Following a principled and systematic pipeline, we further construct Holi-Spatial-4M, the first large-scale, high-quality 3D semantic dataset, containing 12K optimized 3DGS scenes, 1.3M 2D masks, 320K 3D bounding boxes, 320K instance captions, 1.2M 3D grounding instances, and 1.2M spatial QA pairs spanning diverse geometric, relational, and semantic reasoning tasks.\n  Holi-Spatial demonstrates exceptional performance in data curation quality, significantly outperforming existing feed-forward and per-scene optimized methods on datasets such as ScanNet, ScanNet++, and DL3DV. Furthermore, fine-tuning Vision-Language Models (VLMs) on spatial reasoning tasks using this dataset has also led to substantial improvements in model performance.", "date": "2026-03-08", "categories": []}
{"id": "2603.08254", "title": "DynamicVGGT: Learning Dynamic Point Maps for 4D Scene Reconstruction in Autonomous Driving", "abstract": "Dynamic scene reconstruction in autonomous driving remains a fundamental challenge due to significant temporal variations, moving objects, and complex scene dynamics. Existing feed-forward 3D models have demonstrated strong performance in static reconstruction but still struggle to capture dynamic motion. To address these limitations, we propose DynamicVGGT, a unified feed-forward framework that extends VGGT from static 3D perception to dynamic 4D reconstruction. Our goal is to model point motion within feed-forward 3D models in a dynamic and temporally coherent manner. To this end, we jointly predict the current and future point maps within a shared reference coordinate system, allowing the model to implicitly learn dynamic point representations through temporal correspondence. To efficiently capture temporal dependencies, we introduce a Motion-aware Temporal Attention (MTA) module that learns motion continuity. Furthermore, we design a Dynamic 3D Gaussian Splatting Head that explicitly models point motion by predicting Gaussian velocities using learnable motion tokens under scene flow supervision. It refines dynamic geometry through continuous 3D Gaussian optimization. Extensive experiments on autonomous driving datasets demonstrate that DynamicVGGT significantly outperforms existing methods in reconstruction accuracy, achieving robust feed-forward 4D dynamic scene reconstruction under complex driving scenarios.", "date": "2026-03-09", "categories": []}
{"id": "2603.08521", "title": "OccTrack360: 4D Panoptic Occupancy Tracking from Surround-View Fisheye Cameras", "abstract": "Understanding dynamic 3D environments in a spatially continuous and temporally consistent manner is fundamental for robotics and autonomous driving. While recent advances in occupancy prediction provide a unified representation of scene geometry and semantics, progress in 4D panoptic occupancy tracking remains limited by the lack of benchmarks that support surround-view fisheye sensing, long temporal sequences, and instance-level voxel tracking. To address this gap, we present OccTrack360, a new benchmark for 4D panoptic occupancy tracking from surround-view fisheye cameras. OccTrack360 provides substantially longer and more diverse sequences (174~2234 frames) than prior benchmarks, together with principled voxel visibility annotations, including an all-direction occlusion mask and an MEI-based fisheye field-of-view mask. To establish a strong fisheye-oriented baseline, we further propose Focus on Sphere Occ (FoSOcc), a framework that addresses two core challenges in fisheye occupancy tracking: distorted spherical projection and inaccurate voxel-space localization. FoSOcc includes a Center Focusing Module (CFM) to enhance instance-aware spatial localization through supervised focus guidance, and a Spherical Lift Module (SLM) that extends perspective lifting to fisheye imaging under the Unified Projection Model. Extensive experiments on Occ3D-Waymo and OccTrack360 show that our method improves occupancy tracking quality with notable gains on geometrically regular categories, and establishes a strong baseline for future research on surround-view fisheye 4D occupancy tracking. The benchmark and source code will be made publicly available at https://github.com/YouthZest-Lin/OccTrack360.", "date": "2026-03-09", "categories": []}
{"id": "2504.08603", "title": "FindAnything: Open-Vocabulary and Object-Centric Mapping for Robot Exploration in Any Environment", "abstract": "Geometrically accurate and semantically expressive map representations have proven invaluable for robot deployment and task planning in unknown environments. Nevertheless, real-time, open-vocabulary semantic understanding of large-scale unknown environments still presents open challenges, mainly due to computational requirements. In this paper we present FindAnything, an open-world mapping framework that incorporates vision-language information into dense volumetric submaps. Thanks to the use of vision-language features, FindAnything combines pure geometric and open-vocabulary semantic information for a higher level of understanding. It proposes an efficient storage of open-vocabulary information through the aggregation of features at the object level. Pixelwise vision-language features are aggregated based on eSAM segments, which are in turn integrated into object-centric volumetric submaps, providing a mapping from open-vocabulary queries to 3D geometry that is scalable also in terms of memory usage. We demonstrate that FindAnything performs on par with the state-of-the-art in terms of semantic accuracy while being substantially faster and more memory-efficient, allowing its deployment in large-scale environments and on resourceconstrained devices, such as MAVs. We show that the real-time capabilities of FindAnything make it useful for downstream tasks, such as autonomous MAV exploration in a simulated Search and Rescue scenario. Project Page: https://ethz-mrl.github.io/findanything/.", "date": "2025-04-11", "categories": []}
{"id": "2603.06166", "title": "FreeOcc: Training-free Panoptic Occupancy Prediction via Foundation Models", "abstract": "Semantic and panoptic occupancy prediction for road scene analysis provides a dense 3D representation of the ego vehicle's surroundings. Current camera-only approaches typically rely on costly dense 3D supervision or require training models on data from the target domain, limiting deployment in unseen environments. We propose FreeOcc, a training-free pipeline that leverages pretrained foundation models to recover both semantics and geometry from multi-view images. FreeOcc extracts per-view panoptic priors with a promptable foundation segmentation model and prompt-to-taxonomy rules, and reconstructs metric 3D points with a reconstruction foundation model. Depth- and confidence- aware filtering lifts reliable labels into 3D, which are fused over time and voxelized with a deterministic refinement stack. For panoptic occupancy, instances are recovered by fitting and merging robust current-view 3D box candidates, enabling instance-aware occupancy without any learned 3D model. On Occ3D-nuScenes, FreeOcc achieves 16.9 mIoU and 16.5 RayIoU train-free, on par with state-of-the-art weakly supervised methods. When employed as a pseudo-label generation pipeline for training downstream models, it achieves 21.1 RayIoU, surpassing the previous state-of-the-art weakly supervised baseline. Furthermore, FreeOcc sets new baselines for both train-free and weakly supervised panoptic occupancy prediction, achieving 3.1 RayPQ and 3.9 RayPQ, respectively. These results highlight foundation-model-driven perception as a practical route to training-free 3D scene understanding.", "date": "2026-03-06", "categories": []}
{"id": "2603.06279", "title": "Can we Trust Unreliable Voxels? Exploring 3D Semantic Occupancy Prediction under Label Noise", "abstract": "3D semantic occupancy prediction is a cornerstone of robotic perception, yet real-world voxel annotations are inherently corrupted by structural artifacts and dynamic trailing effects. This raises a critical but underexplored question: can autonomous systems safely rely on such unreliable occupancy supervision? To systematically investigate this issue, we establish OccNL, the first benchmark dedicated to 3D occupancy under occupancy-asymmetric and dynamic trailing noise. Our analysis reveals a fundamental domain gap: state-of-the-art 2D label noise learning strategies collapse catastrophically in sparse 3D voxel spaces, exposing a critical vulnerability in existing paradigms. To address this challenge, we propose DPR-Occ, a principled label noise-robust framework that constructs reliable supervision through dual-source partial label reasoning. By synergizing temporal model memory with representation-level structural affinity, DPR-Occ dynamically expands and prunes candidate label sets to preserve true semantics while suppressing noise propagation. Extensive experiments on SemanticKITTI demonstrate that DPR-Occ prevents geometric and semantic collapse under extreme corruption. Notably, even at 90% label noise, our method achieves significant performance gains (up to 2.57% mIoU and 13.91% IoU) over existing label noise learning baselines adapted to the 3D occupancy prediction task. By bridging label noise learning and 3D perception, OccNL and DPR-Occ provide a reliable foundation for safety-critical robotic perception in dynamic environments. The benchmark and source code will be made publicly available at https://github.com/mylwx/OccNL.", "date": "2026-03-06", "categories": []}
{"id": "2504.13596", "title": "Collaborative Learning of Local 3D Occupancy Prediction and Versatile Global Occupancy Mapping", "abstract": "Vision-based 3D semantic occupancy prediction is vital for autonomous driving, enabling unified modeling of static infrastructure and dynamic agents. Global occupancy maps serve as long-term memory priors, providing valuable historical context that enhances local perception. This is particularly important in challenging scenarios such as occlusion or poor illumination, where current and nearby observations may be unreliable or incomplete. Priors aggregated from previous traversals under better conditions help fill gaps and enhance the robustness of local 3D occupancy prediction. In this paper, we propose Long-term Memory Prior Occupancy (LMPOcc), a plug-and-play framework that incorporates global occupancy priors to boost local prediction and simultaneously updates global maps with new observations. To realize the information gain from global priors, we design an efficient and lightweight Current-Prior Fusion module that adaptively integrates prior and current features. Meanwhile, we introduce a model-agnostic prior format to enable continual updating of global occupancy and ensure compatibility across diverse prediction baselines. LMPOcc achieves state-of-the-art local occupancy prediction performance validated on the Occ3D-nuScenes benchmark, especially on static semantic categories. Furthermore, we verify LMPOcc's capability to build large-scale global occupancy maps through multi-vehicle crowdsourcing, and utilize occupancy-derived dense depth to support the construction of 3D open-vocabulary maps. Our method opens up a new paradigm for continuous global information updating and storage, paving the way towards more comprehensive and scalable scene understanding in large outdoor environments.", "date": "2025-04-18", "categories": []}
{"id": "2510.16688", "title": "Pursuing Minimal Sufficiency in Spatial Reasoning", "abstract": "Spatial reasoning, the ability to ground language in 3D understanding, remains a persistent challenge for Vision-Language Models (VLMs). We identify two fundamental bottlenecks: inadequate 3D understanding capabilities stemming from 2D-centric pre-training, and reasoning failures induced by redundant 3D information. To address these, we first construct a Minimal Sufficient Set (MSS) of information before answering a given question: a compact selection of 3D perception results from \\textit{expert models}. We introduce MSSR (Minimal Sufficient Spatial Reasoner), a dual-agent framework that implements this principle. A Perception Agent programmatically queries 3D scenes using a versatile perception toolbox to extract sufficient information, including a novel SOG (Situated Orientation Grounding) module that robustly extracts language-grounded directions. A Reasoning Agent then iteratively refines this information to pursue minimality, pruning redundant details and requesting missing ones in a closed loop until the MSS is curated. Extensive experiments demonstrate that our method, by explicitly pursuing both sufficiency and minimality, significantly improves accuracy and achieves state-of-the-art performance across two challenging benchmarks. Furthermore, our framework produces interpretable reasoning paths, offering a promising source of high-quality training data for future models. Source code is available at https://github.com/gyj155/mssr.", "date": "2025-10-19", "categories": []}
{"id": "2510.16714", "title": "SceneCOT: Eliciting Grounded Chain-of-Thought Reasoning in 3D Scenes", "abstract": "Existing research on 3D Large Language Models (LLMs) still struggles to achieve grounded question-answering, primarily due to the under-exploration of the mechanism of human-like scene-object grounded reasoning. This paper bridges the gap by presenting a novel framework. We first introduce a grounded Chain-of-Thought reasoning method in 3D scenes (SCENECOT), decoupling a complex reasoning task into simpler and manageable problems, and building corresponding visual clues based on multimodal expert modules. To enable such a method, we develop SCENECOT-185K, the first large-scale grounded CoT reasoning dataset, consisting of 185K high-quality instances. Extensive experiments across various complex 3D scene reasoning benchmarks demonstrate that our new framework achieves strong performance with high grounding-QA coherence. To the best of our knowledge, this is the first successful application of CoT reasoning to 3D scene understanding, enabling step-by-step human-like reasoning and showing potential for extension to broader 3D scene understanding scenarios.", "date": "2025-10-19", "categories": []}
{"id": "2603.04614", "title": "SGR3 Model: Scene Graph Retrieval-Reasoning Model in 3D", "abstract": "3D scene graphs provide a structured representation of object entities and their relationships, enabling high-level interpretation and reasoning for robots while remaining intuitively understandable to humans. Existing approaches for 3D scene graph generation typically combine scene reconstruction with graph neural networks (GNNs). However, such pipelines require multi-modal data that may not always be available, and their reliance on heuristic graph construction can constrain the prediction of relationship triplets. In this work, we introduce a Scene Graph Retrieval-Reasoning Model in 3D (SGR3 Model), a training-free framework that leverages multi-modal large language models (MLLMs) with retrieval-augmented generation (RAG) for semantic scene graph generation. SGR3 Model bypasses the need for explicit 3D reconstruction. Instead, it enhances relational reasoning by incorporating semantically aligned scene graphs retrieved via a ColPali-style cross-modal framework. To improve retrieval robustness, we further introduce a weighted patch-level similarity selection mechanism that mitigates the negative impact of blurry or semantically uninformative regions. Experiments demonstrate that SGR3 Model achieves competitive performance compared to training-free baselines and on par with GNN-based expert models. Moreover, an ablation study on the retrieval module and knowledge base scale reveals that retrieved external information is explicitly integrated into the token generation process, rather than being implicitly internalized through abstraction.", "date": "2026-03-04", "categories": []}
{"id": "2603.05473", "title": "Towards 3D Scene Understanding of Gas Plumes in LWIR Hyperspectral Images Using Neural Radiance Fields", "abstract": "Hyperspectral images (HSI) have many applications, ranging from environmental monitoring to national security, and can be used for material detection and identification. Longwave infrared (LWIR) HSI can be used for gas plume detection and analysis. Oftentimes, only a few images of a scene of interest are available and are analyzed individually. The ability to combine information from multiple images into a single, cohesive representation could enhance analysis by providing more context on the scene's geometry and spectral properties. Neural radiance fields (NeRFs) create a latent neural representation of volumetric scene properties that enable novel-view rendering and geometry reconstruction, offering a promising avenue for hyperspectral 3D scene reconstruction. We explore the possibility of using NeRFs to create 3D scene reconstructions from LWIR HSI and demonstrate that the model can be used for the basic downstream analysis task of gas plume detection. The physics-based DIRSIG software suite was used to generate a synthetic multi-view LWIR HSI dataset of a simple facility with a strong sulfur hexafluoride gas plume. Our method, built on the standard Mip-NeRF architecture, combines state-of-the-art methods for hyperspectral NeRFs and sparse-view NeRFs, along with a novel adaptive weighted MSE loss. Our final NeRF method requires around 50% fewer training images than the standard Mip-NeRF and achieves an average PSNR of 39.8 dB with as few as 30 training images. Gas plume detection applied to NeRF-rendered test images using the adaptive coherence estimator achieves an average AUC of 0.821 when compared with detection masks generated from ground-truth test images.", "date": "2026-03-05", "categories": []}
{"id": "2603.04254", "title": "EmbodiedSplat: Online Feed-Forward Semantic 3DGS for Open-Vocabulary 3D Scene Understanding", "abstract": "Understanding a 3D scene immediately with its exploration is essential for embodied tasks, where an agent must construct and comprehend the 3D scene in an online and nearly real-time manner. In this study, we propose EmbodiedSplat, an online feed-forward 3DGS for open-vocabulary scene understanding that enables simultaneous online 3D reconstruction and 3D semantic understanding from the streaming images. Unlike existing open-vocabulary 3DGS methods which are typically restricted to either offline or per-scene optimization setting, our objectives are two-fold: 1) Reconstructs the semantic-embedded 3DGS of the entire scene from over 300 streaming images in an online manner. 2) Highly generalizable to novel scenes with feed-forward design and supports nearly real-time 3D semantic reconstruction when combined with real-time 2D models. To achieve these objectives, we propose an Online Sparse Coefficients Field with a CLIP Global Codebook where it binds the 2D CLIP embeddings to each 3D Gaussian while minimizing memory consumption and preserving the full semantic generalizability of CLIP. Furthermore, we generate 3D geometric-aware CLIP features by aggregating the partial point cloud of 3DGS through 3D U-Net to compensate the 3D geometric prior to 2D-oriented language embeddings. Extensive experiments on diverse indoor datasets, including ScanNet, ScanNet++, and Replica, demonstrate both the effectiveness and efficiency of our method. Check out our project page in https://0nandon.github.io/EmbodiedSplat/.", "date": "2026-03-04", "categories": []}
{"id": "2603.02133", "title": "SimRecon: SimReady Compositional Scene Reconstruction from Real Videos", "abstract": "Compositional scene reconstruction seeks to create object-centric representations rather than holistic scenes from real-world videos, which is natively applicable for simulation and interaction. Conventional compositional reconstruction approaches primarily emphasize on visual appearance and show limited generalization ability to real-world scenarios. In this paper, we propose SimRecon, a framework that realizes a \"Perception-Generation-Simulation\" pipeline towards cluttered scene reconstruction, which first conducts scene-level semantic reconstruction from video input, then performs single-object generation, and finally assembles these assets in the simulator. However, naively combining these three stages leads to visual infidelity of generated assets and physical implausibility of the final scene, a problem particularly severe for complex scenes. Thus, we further propose two bridging modules between the three stages to address this problem. To be specific, for the transition from Perception to Generation, critical for visual fidelity, we introduce Active Viewpoint Optimization, which actively searches in 3D space to acquire optimal projected images as conditions for single-object completion. Moreover, for the transition from Generation to Simulation, essential for physical plausibility, we propose a Scene Graph Synthesizer, which guides the construction from scratch in 3D simulators, mirroring the native, constructive principle of the real world. Extensive experiments on the ScanNet dataset validate our method's superior performance over previous state-of-the-art approaches.", "date": "2026-03-02", "categories": []}
{"id": "2603.02134", "title": "OnlineX: Unified Online 3D Reconstruction and Understanding with Active-to-Stable State Evolution", "abstract": "Recent advances in generalizable 3D Gaussian Splatting (3DGS) have enabled rapid 3D scene reconstruction within seconds, eliminating the need for per-scene optimization. However, existing methods primarily follow an offline reconstruction paradigm, lacking the capacity for continuous reconstruction, which limits their applicability to online scenarios such as robotics and VR/AR. In this paper, we introduce OnlineX, a feed-forward framework that reconstructs both 3D visual appearance and language fields in an online manner using only streaming images. A key challenge in online formulation is the cumulative drift issue, which is rooted in the fundamental conflict between two opposing roles of the memory state: an active role that constantly refreshes to capture high-frequency local geometry, and a stable role that conservatively accumulates and preserves the long-term global structure. To address this, we introduce a decoupled active-to-stable state evolution paradigm. Our framework decouples the memory state into a dedicated active state and a persistent stable state, and then cohesively fuses the information from the former into the latter to achieve both fidelity and stability. Moreover, we jointly model visual appearance and language fields and incorporate an implicit Gaussian fusion module to enhance reconstruction quality. Experiments on mainstream datasets demonstrate that our method consistently outperforms prior work in novel view synthesis and semantic understanding, showcasing robust performance across input sequences of varying lengths with real-time inference speed.", "date": "2026-03-02", "categories": []}
{"id": "2603.02548", "title": "SemGS: Feed-Forward Semantic 3D Gaussian Splatting from Sparse Views for Generalizable Scene Understanding", "abstract": "Semantic understanding of 3D scenes is essential for robots to operate effectively and safely in complex environments. Existing methods for semantic scene reconstruction and semantic-aware novel view synthesis often rely on dense multi-view inputs and require scene-specific optimization, limiting their practicality and scalability in real-world applications. To address these challenges, we propose SemGS, a feed-forward framework for reconstructing generalizable semantic fields from sparse image inputs. SemGS uses a dual-branch architecture to extract color and semantic features, where the two branches share shallow CNN layers, allowing semantic reasoning to leverage textural and structural cues in color appearance. We also incorporate a camera-aware attention mechanism into the feature extractor to explicitly model geometric relationships between camera viewpoints. The extracted features are decoded into dual-Gaussians that share geometric consistency while preserving branch-specific attributes, and further rasterized to synthesize semantic maps under novel viewpoints. Additionally, we introduce a regional smoothness loss to enhance semantic coherence. Experiments show that SemGS achieves state-of-the-art performance on benchmark datasets, while providing rapid inference and strong generalization capabilities across diverse synthetic and real-world scenarios.", "date": "2026-03-03", "categories": []}
{"id": "2603.02609", "title": "VLMFusionOcc3D: VLM Assisted Multi-Modal 3D Semantic Occupancy Prediction", "abstract": "This paper introduces VLMFusionOcc3D, a robust multimodal framework for dense 3D semantic occupancy prediction in autonomous driving. Current voxel-based occupancy models often struggle with semantic ambiguity in sparse geometric grids and performance degradation under adverse weather conditions. To address these challenges, we leverage the rich linguistic priors of Vision-Language Models (VLMs) to anchor ambiguous voxel features to stable semantic concepts. Our framework initiates with a dual-branch feature extraction pipeline that projects multi-view images and LiDAR point clouds into a unified voxel space. We propose Instance-driven VLM Attention (InstVLM), which utilizes gated cross-attention and LoRA-adapted CLIP embeddings to inject high-level semantic and geographic priors directly into the 3D voxels. Furthermore, we introduce Weather-Aware Adaptive Fusion (WeathFusion), a dynamic gating mechanism that utilizes vehicle metadata and weather-conditioned prompts to re-weight sensor contributions based on real-time environmental reliability. To ensure structural consistency, a Depth-Aware Geometric Alignment (DAGA) loss is employed to align dense camera-derived geometry with sparse, spatially accurate LiDAR returns. Extensive experiments on the nuScenes and SemanticKITTI datasets demonstrate that our plug-and-play modules consistently enhance the performance of state-of-the-art voxel-based baselines. Notably, our approach achieves significant improvements in challenging weather scenarios, offering a scalable and robust solution for complex urban navigation.", "date": "2026-03-03", "categories": []}
{"id": "2603.02896", "title": "3D-DRES: Detailed 3D Referring Expression Segmentation", "abstract": "Current 3D visual grounding tasks only process sentence level detection or segmentation, which critically fails to leverage the rich compositional contextual reasonings within natural language expressions. To address this challenge, we introduce Detailed 3D Referring Expression Segmentation (3D-DRES), a new task that provides a phrase to 3D instance mapping, aiming at enhancing fine-grained 3D vision language understanding. To support 3D-DRES, we present DetailRefer, a new dataset comprising 54,432 descriptions spanning 11,054 distinct objects. Unlike previous datasets, DetailRefer implements a pioneering phrase-instance annotation paradigm where each referenced noun phrase is explicitly mapped to its corresponding 3D elements. Additionally, we introduce DetailBase, a purposefully streamlined yet effective baseline architecture that supports dual-mode segmentation at both sentence and phrase levels. Our experimental results demonstrate that models trained on DetailRefer not only excel at phrase-level segmentation but also show surprising improvements on traditional 3D-RES benchmarks.", "date": "2026-03-03", "categories": []}
{"id": "2603.00912", "title": "VGGT-Det: Mining VGGT Internal Priors for Sensor-Geometry-Free Multi-View Indoor 3D Object Detection", "abstract": "Current multi-view indoor 3D object detectors rely on sensor geometry that is costly to obtain (i.e., precisely calibrated multi-view camera poses) to fuse multi-view information into a global scene representation, limiting deployment in real-world scenes. We target a more practical setting: Sensor-Geometry-Free (SG-Free) multi-view indoor 3D object detection, where there are no sensor-provided geometric inputs (multi-view poses or depth). Recent Visual Geometry Grounded Transformer (VGGT) shows that strong 3D cues can be inferred directly from images. Building on this insight, we present VGGT-Det, the first framework tailored for SG-Free multi-view indoor 3D object detection. Rather than merely consuming VGGT predictions, our method integrates VGGT encoder into a transformer-based pipeline. To effectively leverage both the semantic and geometric priors from inside VGGT, we introduce two novel key components: (i) Attention-Guided Query Generation (AG): exploits VGGT attention maps as semantic priors to initialize object queries, improving localization by focusing on object regions while preserving global spatial structure; (ii) Query-Driven Feature Aggregation (QD): a learnable See-Query interacts with object queries to 'see' what they need, and then dynamically aggregates multi-level geometric features across VGGT layers that progressively lift 2D features into 3D. Experiments show that VGGT-Det significantly surpasses the best-performing method in the SG-Free setting by 4.4 and 8.6 mAP@0.25 on ScanNet and ARKitScenes, respectively. Ablation study shows that VGGT's internally learned semantic and geometric priors can be effectively leveraged by our AG and QD.", "date": "2026-03-01", "categories": []}
{"id": "2503.21449", "title": "Towards Generating Realistic 3D Semantic Training Data for Autonomous Driving", "abstract": "Semantic scene understanding is crucial for robotics and computer vision applications. In autonomous driving, 3D semantic segmentation plays an important role for enabling safe navigation. Despite significant advances in the field, the complexity of collecting and annotating 3D data is a bottleneck in this developments. To overcome that data annotation limitation, synthetic simulated data has been used to generate annotated data on demand. There is still, however, a domain gap between real and simulated data. More recently, diffusion models have been in the spotlight, enabling close-to-real data synthesis. Those generative models have been recently applied to the 3D data domain for generating scene-scale data with semantic annotations. Still, those methods either rely on image projection or decoupled models trained with different resolutions in a coarse-to-fine manner. Such intermediary representations impact the generated data quality due to errors added in those transformations. In this work, we propose a novel approach able to generate 3D semantic scene-scale data without relying on any projection or decoupled trained multi-resolution models, achieving more realistic semantic scene data generation compared to previous state-of-the-art methods. Besides improving 3D semantic scene-scale data synthesis, we thoroughly evaluate the use of the synthetic scene samples as labeled data to train a semantic segmentation network. In our experiments, we show that using the synthetic annotated data generated by our method as training data together with the real semantic segmentation labels, leads to an improvement in the semantic segmentation model performance. Our results show the potential of generated scene-scale point clouds to generate more training data to extend existing datasets, reducing the data annotation effort. Our code is available at https://github.com/PRBonn/3DiSS.", "date": "2025-03-27", "categories": []}
{"id": "2507.23599", "title": "DA-Occ: Direction-Aware 2D Convolution for Efficient and Geometry-Preserving 3D Occupancy Prediction in Autonomous Driving", "abstract": "Efficient and high-accuracy 3D occupancy prediction is vital for the performance of autonomous driving systems. However, existing methods struggle to balance precision and efficiency: high-accuracy approaches are often hindered by heavy computational overhead, leading to slow inference speeds, while others leverage pure bird's-eye-view (BEV) representations to gain speed at the cost of losing vertical spatial cues and compromising geometric integrity. To overcome these limitations, we build on the efficient Lift-Splat-Shoot (LSS) paradigm and propose a pure 2D framework, DA-Occ, for 3D occupancy prediction that preserves fine-grained geometry. Standard LSS-based methods lift 2D features into 3D space solely based on depth scores, making it difficult to fully capture vertical structure. To improve upon this, DA-Occ augments depth-based lifting with a complementary height-score projection that explicitly encodes vertical geometric information. We further employ direction-aware convolution to extract geometric features along both vertical and horizontal orientations, effectively balancing accuracy and computational efficiency. On the Occ3D-nuScenes, the proposed method achieves an mIoU of 39.3% and an inference speed of 27.7 FPS, effectively balancing accuracy and efficiency. In simulations on edge devices, the inference speed reaches 14.8 FPS, further demonstrating the method's applicability for real-time deployment in resource-constrained environments.", "date": "2025-07-31", "categories": []}
{"id": "2407.11564", "title": "SGIFormer: Semantic-guided and Geometric-enhanced Interleaving Transformer for 3D Instance Segmentation", "abstract": "In recent years, transformer-based models have exhibited considerable potential in point cloud instance segmentation. Despite the promising performance achieved by existing methods, they encounter challenges such as instance query initialization problems and excessive reliance on stacked layers, rendering them incompatible with large-scale 3D scenes. This paper introduces a novel method, named SGIFormer, for 3D instance segmentation, which is composed of the Semantic-guided Mix Query (SMQ) initialization and the Geometric-enhanced Interleaving Transformer (GIT) decoder. Specifically, the principle of our SMQ initialization scheme is to leverage the predicted voxel-wise semantic information to implicitly generate the scene-aware query, yielding adequate scene prior and compensating for the learnable query set. Subsequently, we feed the formed overall query into our GIT decoder to alternately refine instance query and global scene features for further capturing fine-grained information and reducing complex design intricacies simultaneously. To emphasize geometric property, we consider bias estimation as an auxiliary task and progressively integrate shifted point coordinates embedding to reinforce instance localization. SGIFormer attains state-of-the-art performance on ScanNet V2, ScanNet200 datasets, and the challenging high-fidelity ScanNet++ benchmark, striking a balance between accuracy and efficiency. The code, weights, and demo videos are publicly available at https://rayyoh.github.io/sgiformer.", "date": "2024-07-16", "categories": []}
{"id": "2602.22716", "title": "SoPE: Spherical Coordinate-Based Positional Embedding for Enhancing Spatial Perception of 3D LVLMs", "abstract": "3D Large Vision-Language Models (3D LVLMs) built upon Large Language Models (LLMs) have achieved remarkable progress across various multimodal tasks. However, their inherited position-dependent modeling mechanism, Rotary Position Embedding (RoPE), remains suboptimal for 3D multimodal understanding. The vanilla RoPE formulation fails to preserve essential three-dimensional spatial structures when encoding 3D tokens, and its relative distance computation overlooks angular dependencies, hindering the model's ability to capture directional variations in visual representations. To overcome these limitations, we introduce Spherical Coordinate-based Positional Embedding (SoPE). Our method maps point-cloud token indices into a 3D spherical coordinate space, enabling unified modeling of spatial locations and directional angles. This formulation preserves the inherent geometric structure of point-cloud data, enhances spatial awareness, and yields more consistent and expressive geometric representations for multimodal learning. In addition, we introduce a multi-scale frequency mixing strategy to fuse feature information across different frequency domains. Experimental results on multiple 3D scene benchmarks validate the effectiveness of our approach, while real-world deployment experiments further demonstrate its strong generalization capability.", "date": "2026-02-26", "categories": []}
{"id": "2602.21552", "title": "Generalizing Visual Geometry Priors to Sparse Gaussian Occupancy Prediction", "abstract": "Accurate 3D scene understanding is essential for embodied intelligence, with occupancy prediction emerging as a key task for reasoning about both objects and free space. Existing approaches largely rely on depth priors (e.g., DepthAnything) but make only limited use of 3D cues, restricting performance and generalization. Recently, visual geometry models such as VGGT have shown strong capability in providing rich 3D priors, but similar to monocular depth foundation models, they still operate at the level of visible surfaces rather than volumetric interiors, motivating us to explore how to more effectively leverage these increasingly powerful geometry priors for 3D occupancy prediction. We present GPOcc, a framework that leverages generalizable visual geometry priors (GPs) for monocular occupancy prediction. Our method extends surface points inward along camera rays to generate volumetric samples, which are represented as Gaussian primitives for probabilistic occupancy inference. To handle streaming input, we further design a training-free incremental update strategy that fuses per-frame Gaussians into a unified global representation. Experiments on Occ-ScanNet and EmbodiedOcc-ScanNet demonstrate significant gains: GPOcc improves mIoU by +9.99 in the monocular setting and +11.79 in the streaming setting over prior state of the art. Under the same depth prior, it achieves +6.73 mIoU while running 2.65$\\times$ faster. These results highlight that GPOcc leverages geometry priors more effectively and efficiently. Code will be released at https://github.com/JuIvyy/GPOcc.", "date": "2026-02-25", "categories": []}
{"id": "2511.12370", "title": "Changes in Real Time: Online Scene Change Detection with Multi-View Fusion", "abstract": "Online Scene Change Detection (SCD) is an extremely challenging problem that requires an agent to detect relevant changes on the fly while observing the scene from unconstrained viewpoints. Existing online SCD methods are significantly less accurate than offline approaches. We present the first online SCD approach that is pose-agnostic, label-free, and ensures multi-view consistency, while operating at over 10 FPS and achieving new state-of-the-art performance, surpassing even the best offline approaches. Our method introduces a new self-supervised fusion loss to infer scene changes from multiple cues and observations, PnP-based fast pose estimation against the reference scene, and a fast change-guided update strategy for the 3D Gaussian Splatting scene representation. Extensive experiments on complex real-world datasets demonstrate that our approach outperforms both online and offline baselines.", "date": "2025-11-15", "categories": []}
{"id": "2602.20943", "title": "UFO: Unifying Feed-Forward and Optimization-based Methods for Large Driving Scene Modeling", "abstract": "Dynamic driving scene reconstruction is critical for autonomous driving simulation and closed-loop learning. While recent feed-forward methods have shown promise for 3D reconstruction, they struggle with long-range driving sequences due to quadratic complexity in sequence length and challenges in modeling dynamic objects over extended durations. We propose UFO, a novel recurrent paradigm that combines the benefits of optimization-based and feed-forward methods for efficient long-range 4D reconstruction. Our approach maintains a 4D scene representation that is iteratively refined as new observations arrive, using a visibility-based filtering mechanism to select informative scene tokens and enable efficient processing of long sequences. For dynamic objects, we introduce an object pose-guided modeling approach that supports accurate long-range motion capture. Experiments on the Waymo Open Dataset demonstrate that our method significantly outperforms both per-scene optimization and existing feed-forward methods across various sequence lengths. Notably, our approach can reconstruct 16-second driving logs within 0.5 second while maintaining superior visual quality and geometric accuracy.", "date": "2026-02-24", "categories": []}
{"id": "2602.21186", "title": "Spa3R: Predictive Spatial Field Modeling for 3D Visual Reasoning", "abstract": "While Vision-Language Models (VLMs) exhibit exceptional 2D visual understanding, their ability to comprehend and reason about 3D space--a cornerstone of spatial intelligence--remains superficial. Current methodologies attempt to bridge this domain gap either by relying on explicit 3D modalities or by augmenting VLMs with partial, view-conditioned geometric priors. However, such approaches hinder scalability and ultimately burden the language model with the ill-posed task of implicitly reconstructing holistic 3D geometry from sparse cues. In this paper, we argue that spatial intelligence can emerge inherently from 2D vision alone, rather than being imposed via explicit spatial instruction tuning. To this end, we introduce Spa3R, a self-supervised framework that learns a unified, view-invariant spatial representation directly from unposed multi-view images. Spa3R is built upon the proposed Predictive Spatial Field Modeling (PSFM) paradigm, where Spa3R learns to synthesize feature fields for arbitrary unseen views conditioned on a compact latent representation, thereby internalizing a holistic and coherent understanding of the underlying 3D scene. We further integrate the pre-trained Spa3R Encoder into existing VLMs via a lightweight adapter to form Spa3-VLM, effectively grounding language reasoning in a global spatial context. Experiments on the challenging VSI-Bench demonstrate that Spa3-VLM achieves state-of-the-art accuracy of 58.6% on 3D VQA, significantly outperforming prior methods. These results highlight PSFM as a scalable path toward advancing spatial intelligence. Code is available at https://github.com/hustvl/Spa3R.", "date": "2026-02-24", "categories": []}
{"id": "2602.19823", "title": "Open-vocabulary 3D scene perception in industrial environments", "abstract": "Autonomous vision applications in production, intralogistics, or manufacturing environments require perception capabilities beyond a small, fixed set of classes. Recent open-vocabulary methods, leveraging 2D Vision-Language Foundation Models (VLFMs), target this task but often rely on class-agnostic segmentation models pre-trained on non-industrial datasets (e.g., household scenes). In this work, we first demonstrate that such models fail to generalize, performing poorly on common industrial objects. Therefore, we propose a training-free, open-vocabulary 3D perception pipeline that overcomes this limitation. Instead of using a pre-trained model to generate instance proposals, our method simply generates masks by merging pre-computed superpoints based on their semantic features. Following, we evaluate the domain-adapted VLFM \"IndustrialCLIP\" on a representative 3D industrial workshop scene for open-vocabulary querying. Our qualitative results demonstrate successful segmentation of industrial objects.", "date": "2026-02-23", "categories": []}
{"id": "2602.17134", "title": "B$^3$-Seg: Camera-Free, Training-Free 3DGS Segmentation via Analytic EIG and Beta-Bernoulli Bayesian Updates", "abstract": "Interactive 3D Gaussian Splatting (3DGS) segmentation is essential for real-time editing of pre-reconstructed assets in film and game production. However, existing methods rely on predefined camera viewpoints, ground-truth labels, or costly retraining, making them impractical for low-latency use. We propose B$^3$-Seg (Beta-Bernoulli Bayesian Segmentation for 3DGS), a fast and theoretically grounded method for open-vocabulary 3DGS segmentation under camera-free and training-free conditions. Our approach reformulates segmentation as sequential Beta-Bernoulli Bayesian updates and actively selects the next view via analytic Expected Information Gain (EIG). This Bayesian formulation guarantees the adaptive monotonicity and submodularity of EIG, which produces a greedy $(1{-}1/e)$ approximation to the optimal view sampling policy. Experiments on multiple datasets show that B$^3$-Seg achieves competitive results to high-cost supervised methods while operating end-to-end segmentation within a few seconds. The results demonstrate that B$^3$-Seg enables practical, interactive 3DGS segmentation with provable information efficiency.", "date": "2026-02-19", "categories": []}
{"id": "2602.16356", "title": "Articulated 3D Scene Graphs for Open-World Mobile Manipulation", "abstract": "Semantics has enabled 3D scene understanding and affordance-driven object interaction. However, robots operating in real-world environments face a critical limitation: they cannot anticipate how objects move. Long-horizon mobile manipulation requires closing the gap between semantics, geometry, and kinematics. In this work, we present MoMa-SG, a novel framework for building semantic-kinematic 3D scene graphs of articulated scenes containing a myriad of interactable objects. Given RGB-D sequences containing multiple object articulations, we temporally segment object interactions and infer object motion using occlusion-robust point tracking. We then lift point trajectories into 3D and estimate articulation models using a novel unified twist estimation formulation that robustly estimates revolute and prismatic joint parameters in a single optimization pass. Next, we associate objects with estimated articulations and detect contained objects by reasoning over parent-child relations at identified opening states. We also introduce the novel Arti4D-Semantic dataset, which uniquely combines hierarchical object semantics including parent-child relation labels with object axis annotations across 62 in-the-wild RGB-D sequences containing 600 object interactions and three distinct observation paradigms. We extensively evaluate the performance of MoMa-SG on two datasets and ablate key design choices of our approach. In addition, real-world experiments on both a quadruped and a mobile manipulator demonstrate that our semantic-kinematic scene graphs enable robust manipulation of articulated objects in everyday home environments. We provide code and data at: https://momasg.cs.uni-freiburg.de.", "date": "2026-02-18", "categories": []}
{"id": "2507.14841", "title": "Towards Geometric and Textural Consistency 3D Scene Generation via Single Image-guided Model Generation and Layout Optimization", "abstract": "In recent years, 3D generation has made great strides in both academia and industry. However, generating 3D scenes from a single RGB image remains a significant challenge, as current approaches often struggle to ensure both object generation quality and scene coherence in multi-object scenarios. To overcome these limitations, we propose a novel three-stage framework for 3D scene generation with explicit geometric representations and high-quality textural details via single image-guided model generation and spatial layout optimization. Our method begins with an image instance segmentation and inpainting phase, which recovers missing details of occluded objects in the input images, thereby achieving complete generation of foreground 3D assets. Subsequently, our approach captures the spatial geometry of reference image by constructing pseudo-stereo viewpoint for camera parameter estimation and scene depth inference, while employing a model selection strategy to ensure optimal alignment between the 3D assets generated in the previous step and the input. Finally, through model parameterization and minimization of the Chamfer distance between point clouds in 3D and 2D space, our approach optimizes layout parameters to produce an explicit 3D scene representation that maintains precise alignment with input guidance image. Extensive experiments on multi-object scene image sets have demonstrated that our approach not only outperforms state-of-the-art methods in terms of geometric accuracy and texture fidelity of individual generated 3D models, but also has significant advantages in scene layout synthesis.", "date": "2025-07-20", "categories": []}
{"id": "2602.15734", "title": "Language and Geometry Grounded Sparse Voxel Representations for Holistic Scene Understanding", "abstract": "Existing 3D open-vocabulary scene understanding methods mostly emphasize distilling language features from 2D foundation models into 3D feature fields, but largely overlook the synergy among scene appearance, semantics, and geometry. As a result, scene understanding often deviates from the underlying geometric structure of scenes and becomes decoupled from the reconstruction process. In this work, we propose a novel approach that leverages language and geometry grounded sparse voxel representations to comprehensively model appearance, semantics, and geometry within a unified framework. Specifically, we use 3D sparse voxels as primitives and employ an appearance field, a density field, a feature field, and a confidence field to holistically represent a 3D scene. To promote synergy among the appearance, density, and feature fields, we construct a feature modulation module and distill language features from a 2D foundation model into our 3D scene model. In addition, we integrate geometric distillation into feature field distillation to transfer geometric knowledge from a geometry foundation model to our 3D scene representations via depth correlation regularization and pattern consistency regularization. These components work together to synergistically model the appearance, semantics, and geometry of the 3D scene within a unified framework. Extensive experiments demonstrate that our approach achieves superior overall performance compared with state-of-the-art methods in holistic scene understanding and reconstruction.", "date": "2026-02-17", "categories": []}
{"id": "2505.10685", "title": "GaussianFormer3D: Multi-Modal Gaussian-based Semantic Occupancy Prediction with 3D Deformable Attention", "abstract": "3D semantic occupancy prediction is essential for achieving safe, reliable autonomous driving and robotic navigation. Compared to camera-only perception systems, multi-modal pipelines, especially LiDAR-camera fusion methods, can produce more accurate and fine-grained predictions. Although voxel-based scene representations are widely used for semantic occupancy prediction, 3D Gaussians have emerged as a continuous and significantly more compact alternative. In this work, we propose a multi-modal Gaussian-based semantic occupancy prediction framework utilizing 3D deformable attention, namely GaussianFormer3D. We introduce a voxel-to-Gaussian initialization strategy that provides 3D Gaussians with accurate geometry priors from LiDAR data, and design a LiDAR-guided 3D deformable attention mechanism to refine these Gaussians using LiDAR-camera fusion features in a lifted 3D space. Extensive experiments on real-world on-road and off-road autonomous driving datasets demonstrate that GaussianFormer3D achieves state-of-the-art prediction performance with reduced memory consumption and improved efficiency.", "date": "2025-05-15", "categories": []}
{"id": "2602.14021", "title": "Flow4R: Unifying 4D Reconstruction and Tracking with Scene Flow", "abstract": "Reconstructing and tracking dynamic 3D scenes remains a fundamental challenge in computer vision. Existing approaches often decouple geometry from motion: multi-view reconstruction methods assume static scenes, while dynamic tracking frameworks rely on explicit camera pose estimation or separate motion models. We propose Flow4R, a unified framework that treats camera-space scene flow as the central representation linking 3D structure, object motion, and camera motion. Flow4R predicts a minimal per-pixel property set-3D point position, scene flow, pose weight, and confidence-from two-view inputs using a Vision Transformer. This flow-centric formulation allows local geometry and bidirectional motion to be inferred symmetrically with a shared decoder in a single forward pass, without requiring explicit pose regressors or bundle adjustment. Trained jointly on static and dynamic datasets, Flow4R achieves state-of-the-art performance on 4D reconstruction and tracking tasks, demonstrating the effectiveness of the flow-central representation for spatiotemporal scene understanding.", "date": "2026-02-15", "categories": []}
{"id": "2602.04053", "title": "Seeing Through Clutter: Structured 3D Scene Reconstruction via Iterative Object Removal", "abstract": "We present SeeingThroughClutter, a method for reconstructing structured 3D representations from single images by segmenting and modeling objects individually. Prior approaches rely on intermediate tasks such as semantic segmentation and depth estimation, which often underperform in complex scenes, particularly in the presence of occlusion and clutter. We address this by introducing an iterative object removal and reconstruction pipeline that decomposes complex scenes into a sequence of simpler subtasks. Using VLMs as orchestrators, foreground objects are removed one at a time via detection, segmentation, object removal, and 3D fitting. We show that removing objects allows for cleaner segmentations of subsequent objects, even in highly occluded scenes. Our method requires no task-specific training and benefits directly from ongoing advances in foundation models. We demonstrate stateof-the-art robustness on 3D-Front and ADE20K datasets. Project Page: https://rioak.github.io/seeingthroughclutter/", "date": "2026-02-03", "categories": []}
{"id": "2509.05515", "title": "Visibility-Aware Language Aggregation for Open-Vocabulary Segmentation in 3D Gaussian Splatting", "abstract": "Recently, distilling open-vocabulary language features from 2D images into 3D Gaussians has attracted significant attention. Although existing methods achieve impressive language-based interactions of 3D scenes, we observe two fundamental issues: background Gaussians contributing negligibly to a rendered pixel get the same feature as the dominant foreground ones, and multi-view inconsistencies due to view-specific noise in language embeddings. We introduce Visibility-Aware Language Aggregation (VALA), a lightweight yet effective method that computes marginal contributions for each ray and applies a visibility-aware gate to retain only visible Gaussians. Moreover, we propose a streaming weighted geometric median in cosine space to merge noisy multi-view features. Our method yields a robust, view-consistent language feature embedding in a fast and memory-efficient manner. VALA improves open-vocabulary localization and segmentation across reference datasets, consistently surpassing existing works. More results are available at https://vala3d.github.io", "date": "2025-09-05", "categories": []}
{"id": "2602.11007", "title": "LaSSM: Efficient Semantic-Spatial Query Decoding via Local Aggregation and State Space Models for 3D Instance Segmentation", "abstract": "Query-based 3D scene instance segmentation from point clouds has attained notable performance. However, existing methods suffer from the query initialization dilemma due to the sparse nature of point clouds and rely on computationally intensive attention mechanisms in query decoders. We accordingly introduce LaSSM, prioritizing simplicity and efficiency while maintaining competitive performance. Specifically, we propose a hierarchical semantic-spatial query initializer to derive the query set from superpoints by considering both semantic cues and spatial distribution, achieving comprehensive scene coverage and accelerated convergence. We further present a coordinate-guided state space model (SSM) decoder that progressively refines queries. The novel decoder features a local aggregation scheme that restricts the model to focus on geometrically coherent regions and a spatial dual-path SSM block to capture underlying dependencies within the query set by integrating associated coordinates information. Our design enables efficient instance prediction, avoiding the incorporation of noisy information and reducing redundant computation. LaSSM ranks first place on the latest ScanNet++ V2 leaderboard, outperforming the previous best method by 2.5% mAP with only 1/3 FLOPs, demonstrating its superiority in challenging large-scale scene instance segmentation. LaSSM also achieves competitive performance on ScanNet, ScanNet200, S3DIS and ScanNet++ V1 benchmarks with less computational cost. Extensive ablation studies and qualitative results validate the effectiveness of our design. The code and weights are available at https://github.com/RayYoh/LaSSM.", "date": "2026-02-11", "categories": []}
{"id": "2602.04152", "title": "MA3DSG: Multi-Agent 3D Scene Graph Generation for Large-Scale Indoor Environments", "abstract": "Current 3D scene graph generation (3DSGG) approaches heavily rely on a single-agent assumption and small-scale environments, exhibiting limited scalability to real-world scenarios. In this work, we introduce Multi-Agent 3D Scene Graph Generation (MA3DSG) model, the first framework designed to tackle this scalability challenge using multiple agents. We develop a training-free graph alignment algorithm that efficiently merges partial query graphs from individual agents into a unified global scene graph. Leveraging extensive analysis and empirical insights, our approach enables conventional single-agent systems to operate collaboratively without requiring any learnable parameters. To rigorously evaluate 3DSGG performance, we propose MA3DSG-Bench-a benchmark that supports diverse agent configurations, domain sizes, and environmental conditions-providing a more general and extensible evaluation framework. This work lays a solid foundation for scalable, multi-agent 3DSGG research.", "date": "2026-02-04", "categories": []}
{"id": "2405.02162", "title": "Mapping the Unseen: Unified Promptable Panoptic Mapping with Dynamic Labeling using Foundation Models", "abstract": "Panoptic maps enable robots to reason about both geometry and semantics. However, open-vocabulary models repeatedly produce closely related labels that split panoptic entities and degrade volumetric consistency. The proposed UPPM advances open-world scene understanding by leveraging foundation models to introduce a panoptic Dynamic Descriptor that reconciles open-vocabulary labels with unified category structure and geometric size priors. The fusion for such dynamic descriptors is performed within a multi-resolution multi-TSDF map using language-guided open-vocabulary panoptic segmentation and semantic retrieval, resulting in a persistent and promptable panoptic map without additional model training. Based on our evaluation experiments, UPPM shows the best overall performance in terms of the map reconstruction accuracy and the panoptic segmentation quality. The ablation study investigates the contribution for each component of UPPM (custom NMS, blurry-frame filtering, and unified semantics) to the overall system performance. Consequently, UPPM preserves open-vocabulary interpretability while delivering strong geometric and panoptic accuracy.", "date": "2024-05-03", "categories": []}
{"id": "2602.03371", "title": "Multi-Resolution Alignment for Voxel Sparsity in Camera-Based 3D Semantic Scene Completion", "abstract": "Camera-based 3D semantic scene completion (SSC) offers a cost-effective solution for assessing the geometric occupancy and semantic labels of each voxel in the surrounding 3D scene with image inputs, providing a voxel-level scene perception foundation for the perception-prediction-planning autonomous driving systems. Although significant progress has been made in existing methods, their optimization rely solely on the supervision from voxel labels and face the challenge of voxel sparsity as a large portion of voxels in autonomous driving scenarios are empty, which limits both optimization efficiency and model performance. To address this issue, we propose a \\textit{Multi-Resolution Alignment (MRA)} approach to mitigate voxel sparsity in camera-based 3D semantic scene completion, which exploits the scene and instance level alignment across multi-resolution 3D features as auxiliary supervision. Specifically, we first propose the Multi-resolution View Transformer module, which projects 2D image features into multi-resolution 3D features and aligns them at the scene level through fusing discriminative seed features. Furthermore, we design the Cubic Semantic Anisotropy module to identify the instance-level semantic significance of each voxel, accounting for the semantic differences of a specific voxel against its neighboring voxels within a cubic area. Finally, we devise a Critical Distribution Alignment module, which selects critical voxels as instance-level anchors with the guidance of cubic semantic anisotropy, and applies a circulated loss for auxiliary supervision on the critical feature distribution consistency across different resolutions. The code is available at https://github.com/PKU-ICST-MIPL/MRA_TIP.", "date": "2026-02-03", "categories": []}
{"id": "2602.03809", "title": "Split&Splat: Zero-Shot Panoptic Segmentation via Explicit Instance Modeling and 3D Gaussian Splatting", "abstract": "3D Gaussian Splatting (GS) enables fast and high-quality scene reconstruction, but it lacks an object-consistent and semantically aware structure. We propose Split&Splat, a framework for panoptic scene reconstruction using 3DGS. Our approach explicitly models object instances. It first propagates instance masks across views using depth, thus producing view-consistent 2D masks. Each object is then reconstructed independently and merged back into the scene while refining its boundaries. Finally, instance-level semantic descriptors are embedded in the reconstructed objects, supporting various applications, including panoptic segmentation, object retrieval, and 3D editing. Unlike existing methods, Split&Splat tackles the problem by first segmenting the scene and then reconstructing each object individually. This design naturally supports downstream tasks and allows Split&Splat to achieve state-of-the-art performance on the ScanNetv2 segmentation benchmark.", "date": "2026-02-01", "categories": []}
{"id": "2602.00637", "title": "VIZOR: Viewpoint-Invariant Zero-Shot Scene Graph Generation for 3D Scene Reasoning", "abstract": "Scene understanding and reasoning has been a fundamental problem in 3D computer vision, requiring models to identify objects, their properties, and spatial or comparative relationships among the objects. Existing approaches enable this by creating scene graphs using multiple inputs such as 2D images, depth maps, object labels, and annotated relationships from specific reference view. However, these methods often struggle with generalization and produce inaccurate spatial relationships like \"left/right\", which become inconsistent across different viewpoints. To address these limitations, we propose Viewpoint-Invariant Zero-shot scene graph generation for 3D scene Reasoning (VIZOR). VIZOR is a training-free, end-to-end framework that constructs dense, viewpoint-invariant 3D scene graphs directly from raw 3D scenes. The generated scene graph is unambiguous, as spatial relationships are defined relative to each object's front-facing direction, making them consistent regardless of the reference view. Furthermore, it infers open-vocabulary relationships that describe spatial and proximity relationships among scene objects without requiring annotated training data. We conduct extensive quantitative and qualitative evaluations to assess the effectiveness of VIZOR in scene graph generation and downstream tasks, such as query-based object grounding. VIZOR outperforms state-of-the-art methods, showing clear improvements in scene graph generation and achieving 22% and 4.81% gains in zero-shot grounding accuracy on the Replica and Nr3D datasets, respectively.", "date": "2026-01-31", "categories": []}
{"id": "2511.07222", "title": "Omni-View: Unlocking How Generation Facilitates Understanding in Unified 3D Model based on Multiview images", "abstract": "This paper presents Omni-View, which extends the unified multimodal understanding and generation to 3D scenes based on multiview images, exploring the principle that \"generation facilitates understanding\". Consisting of understanding model, texture module, and geometry module, Omni-View jointly models scene understanding, novel view synthesis, and geometry estimation, enabling synergistic interaction between 3D scene understanding and generation tasks. By design, it leverages the spatiotemporal modeling capabilities of its texture module responsible for appearance synthesis, alongside the explicit geometric constraints provided by its dedicated geometry module, thereby enriching the model's holistic understanding of 3D scenes. Trained with a two-stage strategy, Omni-View achieves a state-of-the-art score of 55.4 on the VSI-Bench benchmark, outperforming existing specialized 3D understanding models, while simultaneously delivering strong performance in both novel view synthesis and 3D scene generation. The code and pretraiend models are open-sourced at https://github.com/AIDC-AI/Omni-View.", "date": "2025-11-10", "categories": []}
{"id": "2508.12216", "title": "Splat Feature Solver", "abstract": "Feature lifting has emerged as a crucial component in 3D scene understanding, enabling the attachment of rich image feature descriptors (e.g., DINO, CLIP) onto splat-based 3D representations. The core challenge lies in optimally assigning rich general attributes to 3D primitives while addressing the inconsistency issues from multi-view images. We present a unified, kernel- and feature-agnostic formulation of the feature lifting problem as a sparse linear inverse problem, which can be solved efficiently in closed form. Our approach admits a provable upper bound on the global optimal error under convex losses for delivering high quality lifted features. To address inconsistencies and noise in multi-view observations, we introduce two complementary regularization strategies to stabilize the solution and enhance semantic fidelity. Tikhonov Guidance enforces numerical stability through soft diagonal dominance, while Post-Lifting Aggregation filters noisy inputs via feature clustering. Extensive experiments demonstrate that our approach achieves state-of-the-art performance on open-vocabulary 3D segmentation benchmarks, outperforming training-based, grouping-based, and heuristic-forward baselines while producing lifted features in minutes. Our \\textbf{code} is available in the \\href{https://github.com/saliteta/splat-distiller/tree/main}{\\textcolor{blue}{GitHub}}. We provide additional \\href{https://splat-distiller.pages.dev/}{\\textcolor{blue}{website}} for more visualization, as well as the \\href{https://www.youtube.com/watch?v=CH-G5hbvArM}{\\textcolor{blue}{video}}.", "date": "2025-08-17", "categories": []}
{"id": "2506.08710", "title": "SceneSplat++: A Large Dataset and Comprehensive Benchmark for Language Gaussian Splatting", "abstract": "3D Gaussian Splatting (3DGS) serves as a highly performant and efficient encoding of scene geometry, appearance, and semantics. Moreover, grounding language in 3D scenes has proven to be an effective strategy for 3D scene understanding. Current Language Gaussian Splatting line of work fall into three main groups: (i) per-scene optimization-based, (ii) per-scene optimization-free, and (iii) generalizable approach. However, most of them are evaluated only on rendered 2D views of a handful of scenes and viewpoints close to the training views, limiting ability and insight into holistic 3D understanding. To address this gap, we propose the first large-scale benchmark that systematically assesses these three groups of methods directly in 3D space, evaluating on 1060 scenes across three indoor datasets and one outdoor dataset. Benchmark results demonstrate a clear advantage of the generalizable paradigm, particularly in relaxing the scene-specific limitation, enabling fast feed-forward inference on novel scenes, and achieving superior segmentation performance. We further introduce GaussianWorld-49K a carefully curated 3DGS dataset comprising around 49K diverse indoor and outdoor scenes obtained from multiple sources, with which we demonstrate the generalizable approach could harness strong data priors. Our codes, benchmark, and datasets are released at https://scenesplatpp.gaussianworld.ai/.", "date": "2025-06-10", "categories": []}
{"id": "2506.20671", "title": "IPFormer: Visual 3D Panoptic Scene Completion with Context-Adaptive Instance Proposals", "abstract": "Semantic Scene Completion (SSC) has emerged as a pivotal approach for jointly learning scene geometry and semantics, enabling downstream applications such as navigation in mobile robotics. The recent generalization to Panoptic Scene Completion (PSC) advances the SSC domain by integrating instance-level information, thereby enhancing object-level sensitivity in scene understanding. While PSC was introduced using LiDAR modality, methods based on camera images remain largely unexplored. Moreover, recent Transformer-based approaches utilize a fixed set of learned queries to reconstruct objects within the scene volume. Although these queries are typically updated with image context during training, they remain static at test time, limiting their ability to dynamically adapt specifically to the observed scene. To overcome these limitations, we propose IPFormer, the first method that leverages context-adaptive instance proposals at train and test time to address vision-based 3D Panoptic Scene Completion. Specifically, IPFormer adaptively initializes these queries as panoptic instance proposals derived from image context and further refines them through attention-based encoding and decoding to reason about semantic instance-voxel relationships. Extensive experimental results show that our approach achieves state-of-the-art in-domain performance, exhibits superior zero-shot generalization on out-of-domain data, and achieves a runtime reduction exceeding 14x. These results highlight our introduction of context-adaptive instance proposals as a pioneering effort in addressing vision-based 3D Panoptic Scene Completion. Code available at https://github.com/markus-42/ipformer.", "date": "2025-06-25", "categories": []}
{"id": "2509.15871", "title": "Zero-Shot Visual Grounding in 3D Gaussians via View Retrieval", "abstract": "3D Visual Grounding (3DVG) aims to locate objects in 3D scenes based on text prompts, which is essential for applications such as robotics. However, existing 3DVG methods encounter two main challenges: first, they struggle to handle the implicit representation of spatial textures in 3D Gaussian Splatting (3DGS), making per-scene training indispensable; second, they typically require larges amounts of labeled data for effective training. To this end, we propose \\underline{G}rounding via \\underline{V}iew \\underline{R}etrieval (GVR), a novel zero-shot visual grounding framework for 3DGS to transform 3DVG as a 2D retrieval task that leverages object-level view retrieval to collect grounding clues from multiple views, which not only avoids the costly process of 3D annotation, but also eliminates the need for per-scene training. Extensive experiments demonstrate that our method achieves state-of-the-art visual grounding performance while avoiding per-scene training, providing a solid foundation for zero-shot 3DVG research. Video demos can be found in https://github.com/leviome/GVR_demos.", "date": "2025-09-19", "categories": []}
{"id": "2503.18944", "title": "DINO in the Room: Leveraging 2D Foundation Models for 3D Segmentation", "abstract": "  Vision foundation models (VFMs) trained on large-scale image datasets provide high-quality features that have significantly advanced 2D visual recognition. However, their potential in 3D scene segmentation remains largely untapped, despite the common availability of 2D images alongside 3D point cloud datasets. While significant research has been dedicated to 2D-3D fusion, recent state-of-the-art 3D methods predominantly focus on 3D data, leaving the integration of VFMs into 3D models underexplored. In this work, we challenge this trend by introducing DITR, a generally applicable approach that extracts 2D foundation model features, projects them to 3D, and finally injects them into a 3D point cloud segmentation model. DITR achieves state-of-the-art results on both indoor and outdoor 3D semantic segmentation benchmarks. To enable the use of VFMs even when images are unavailable during inference, we additionally propose to pretrain 3D models by distilling 2D foundation models. By initializing the 3D backbone with knowledge distilled from 2D VFMs, we create a strong basis for downstream 3D segmentation tasks, ultimately boosting performance across various datasets.", "date": "2025-03-24", "categories": []}
{"id": "2601.15025", "title": "ExPrIS: Knowledge-Level Expectations as Priors for Object Interpretation from Sensor Data", "abstract": "While deep learning has significantly advanced robotic object recognition, purely data-driven approaches often lack semantic consistency and fail to leverage valuable, pre-existing knowledge about the environment. This report presents the ExPrIS project, which addresses this challenge by investigating how knowledge-level expectations can serve as to improve object interpretation from sensor data. Our approach is based on the incremental construction of a 3D Semantic Scene Graph (3DSSG). We integrate expectations from two sources: contextual priors from past observations and semantic knowledge from external graphs like ConceptNet. These are embedded into a heterogeneous Graph Neural Network (GNN) to create an expectation-biased inference process. This method moves beyond static, frame-by-frame analysis to enhance the robustness and consistency of scene understanding over time. The report details this architecture, its evaluation, and outlines its planned integration on a mobile robotic platform.", "date": "2026-01-21", "categories": []}
{"id": "2601.15250", "title": "FlowSSC: Universal Generative Monocular Semantic Scene Completion via One-Step Latent Diffusion", "abstract": "Semantic Scene Completion (SSC) from monocular RGB images is a fundamental yet challenging task due to the inherent ambiguity of inferring occluded 3D geometry from a single view. While feed-forward methods have made progress, they often struggle to generate plausible details in occluded regions and preserve the fundamental spatial relationships of objects. Such accurate generative reasoning capability for the entire 3D space is critical in real-world applications. In this paper, we present FlowSSC, the first generative framework applied directly to monocular semantic scene completion. FlowSSC treats the SSC task as a conditional generation problem and can seamlessly integrate with existing feed-forward SSC methods to significantly boost their performance. To achieve real-time inference without compromising quality, we introduce Shortcut Flow-matching that operates in a compact triplane latent space. Unlike standard diffusion models that require hundreds of steps, our method utilizes a shortcut mechanism to achieve high-fidelity generation in a single step, enabling practical deployment in autonomous systems. Extensive experiments on SemanticKITTI demonstrate that FlowSSC achieves state-of-the-art performance, significantly outperforming existing baselines.", "date": "2026-01-21", "categories": []}
{"id": "2411.03672", "title": "MetaSSC: Enhancing 3D Semantic Scene Completion for Autonomous Driving through Meta-Learning and Long-sequence Modeling", "abstract": "Semantic scene completion (SSC) is essential for achieving comprehensive perception in autonomous driving systems. However, existing SSC methods often overlook the high deployment costs in real-world applications. Traditional architectures, such as 3D Convolutional Neural Networks (3D CNNs) and self-attention mechanisms, face challenges in efficiently capturing long-range dependencies within 3D voxel grids, limiting their effectiveness. To address these issues, we introduce MetaSSC, a novel meta-learning-based framework for SSC that leverages deformable convolution, large-kernel attention, and the Mamba (D-LKA-M) model. Our approach begins with a voxel-based semantic segmentation (SS) pretraining task, aimed at exploring the semantics and geometry of incomplete regions while acquiring transferable meta-knowledge. Using simulated cooperative perception datasets, we supervise the perception training of a single vehicle using aggregated sensor data from multiple nearby connected autonomous vehicles (CAVs), generating richer and more comprehensive labels. This meta-knowledge is then adapted to the target domain through a dual-phase training strategy that does not add extra model parameters, enabling efficient deployment. To further enhance the model's capability in capturing long-sequence relationships within 3D voxel grids, we integrate Mamba blocks with deformable convolution and large-kernel attention into the backbone network. Extensive experiments demonstrate that MetaSSC achieves state-of-the-art performance, significantly outperforming competing models while also reducing deployment costs.", "date": "2024-11-06", "categories": []}
{"id": "2412.10231", "title": "SuperGSeg: Open-Vocabulary 3D Segmentation with Structured Super-Gaussians", "abstract": "3D Gaussian Splatting has recently gained traction for its efficient training and real-time rendering. While its vanilla representation is mainly designed for view synthesis, recent works extended it to scene understanding with language features. However, storing additional high-dimensional features per Gaussian for semantic information is memory-intensive, which limits their ability to segment and interpret challenging scenes. To this end, we introduce SuperGSeg, a novel approach that fosters cohesive, context-aware hierarchical scene representation by disentangling segmentation and language field distillation. SuperGSeg first employs neural 3D Gaussians to learn geometry, instance and hierarchical segmentation features from multi-view images with the aid of off-the-shelf 2D masks. These features are then leveraged to create a sparse set of \\acrlong{superg}s. \\acrlong{superg}s facilitate the lifting and distillation of 2D language features into 3D space. They enable hierarchical scene understanding with high-dimensional language feature rendering at moderate GPU memory costs. Extensive experiments demonstrate that SuperGSeg achieves remarkable performance on both open-vocabulary object selection and semantic segmentation tasks.", "date": "2024-12-13", "categories": []}
{"id": "2511.12498", "title": "Towards Temporal Fusion Beyond the Field of View for Camera-based Semantic Scene Completion", "abstract": "Recent camera-based 3D semantic scene completion (SSC) methods have increasingly explored leveraging temporal cues to enrich the features of the current frame. However, while these approaches primarily focus on enhancing in-frame regions, they often struggle to reconstruct critical out-of-frame areas near the sides of the ego-vehicle, although previous frames commonly contain valuable contextual information about these unseen regions. To address this limitation, we propose the Current-Centric Contextual 3D Fusion (C3DFusion) module, which generates hidden region-aware 3D feature geometry by explicitly aligning 3D-lifted point features from both current and historical frames. C3DFusion performs enhanced temporal fusion through two complementary techniques-historical context blurring and current-centric feature densification-which suppress noise from inaccurately warped historical point features by attenuating their scale, and enhance current point features by increasing their volumetric contribution. Simply integrated into standard SSC architectures, C3DFusion demonstrates strong effectiveness, significantly outperforming state-of-the-art methods on the SemanticKITTI and SSCBench-KITTI-360 datasets. Furthermore, it exhibits robust generalization, achieving notable performance gains when applied to other baseline models.", "date": "2025-11-16", "categories": []}
{"id": "2601.06474", "title": "SparseOccVLA: Bridging Occupancy and Vision-Language Models via Sparse Queries for Unified 4D Scene Understanding and Planning", "abstract": "In autonomous driving, Vision Language Models (VLMs) excel at high-level reasoning , whereas semantic occupancy provides fine-grained details. Despite significant progress in individual fields, there is still no method that can effectively integrate both paradigms. Conventional VLMs struggle with token explosion and limited spatiotemporal reasoning, while semantic occupancy provides a unified, explicit spatial representation but is too dense to integrate efficiently with VLMs. To address these challenges and bridge the gap between VLMs and occupancy, we propose SparseOccVLA, a novel vision-language-action model that unifies scene understanding, occupancy forecasting, and trajectory planning powered by sparse occupancy queries. Starting with a lightweight Sparse Occupancy Encoder, SparseOccVLA generates compact yet highly informative sparse occupancy queries that serve as the single bridge between vision and language. These queries are aligned into the language space and reasoned by the LLM for unified scene understanding and future occupancy forecasting. Furthermore, we introduce an LLM-guided Anchor-Diffusion Planner featuring decoupled anchor scoring and denoising, as well as cross-model trajectory-condition fusion. SparseOccVLA achieves a 7% relative improvement in CIDEr over the state-of-the-art on OmniDrive-nuScenes, a 0.5 increase in mIoU score on Occ3D-nuScenes, and sets state-of-the-art open-loop planning metric on nuScenes benchmark, demonstrating its strong holistic capability.", "date": "2026-01-10", "categories": []}
{"id": "2601.12683", "title": "GaussianTrimmer: Online Trimming Boundaries for 3DGS Segmentation", "abstract": "With the widespread application of 3D Gaussians in 3D scene representation, 3D scene segmentation methods based on 3D Gaussians have also gradually emerged. However, existing 3D Gaussian segmentation methods basically segment on the basis of Gaussian primitives. Due to the large variation range of the scale of 3D Gaussians, large-sized Gaussians that often span the foreground and background lead to jagged boundaries of segmented objects. To this end, we propose an online boundary trimming method, GaussianTrimmer, which is an efficient and plug-and-play post-processing method capable of trimming coarse boundaries for existing 3D Gaussian segmentation methods. Our method consists of two core steps: 1. Generating uniformly and well-covered virtual cameras; 2. Trimming Gaussian at the primitive level based on 2D segmentation results on virtual cameras. Extensive quantitative and qualitative experiments demonstrate that our method can improve the segmentation quality of existing 3D Gaussian segmentation methods as a plug-and-play method.", "date": "2026-01-19", "categories": []}
{"id": "2601.13132", "title": "GaussExplorer: 3D Gaussian Splatting for Embodied Exploration and Reasoning", "abstract": "We present GaussExplorer, a framework for embodied exploration and reasoning built on 3D Gaussian Splatting (3DGS). While prior approaches to language-embedded 3DGS have made meaningful progress in aligning simple text queries with Gaussian embeddings, they are generally optimized for relatively simple queries and struggle to interpret more complex, compositional language queries. Alternative studies based on object-centric RGB-D structured memories provide spatial grounding but are constrained by pre-fixed viewpoints. To address these issues, GaussExplorer introduces Vision-Language Models (VLMs) on top of 3DGS to enable question-driven exploration and reasoning within 3D scenes. We first identify pre-captured images that are most correlated with the query question, and subsequently adjust them into novel viewpoints to more accurately capture visual information for better reasoning by VLMs. Experiments show that ours outperforms existing methods on several benchmarks, demonstrating the effectiveness of integrating VLM-based reasoning with 3DGS for embodied tasks.", "date": "2026-01-19", "categories": []}
{"id": "2512.11574", "title": "Evaluating Foundation Models' 3D Understanding Through Multi-View Correspondence Analysis", "abstract": "Benchmarking 3D spatial understanding of foundation models is essential for real-world applications such as robotics and autonomous driving. Existing evaluations often rely on downstream fine-tuning with linear heads or task-specific decoders, making it difficult to isolate the intrinsic 3D reasoning ability of pre-trained encoders. In this work, we introduce a novel benchmark for in-context 3D scene understanding that requires no fine-tuning and directly probes the quality of dense visual features. Building on the Hummingbird framework, which evaluates in-context 2D scene understanding, we extend the setup to the 3D Multi-View ImageNet (MVImgNet) dataset. Given a set of images depicting objects at specific camera angles (keys), we benchmark the performance of segmenting novel views (queries) and report the scores in 4 categories of easy, medium, hard, and extreme based on the key-query view contrast. We benchmark 7 state-of-the-art foundation models and show that DINO-based encoders remain competitive across large viewpoint shifts. Our code is publicly available at https://github.com/ToyeshC/open-hummingbird-3d-eval.", "date": "2025-12-12", "categories": []}
{"id": "2601.11442", "title": "Map2Thought: Explicit 3D Spatial Reasoning via Metric Cognitive Maps", "abstract": "We propose Map2Thought, a framework that enables explicit and interpretable spatial reasoning for 3D VLMs. The framework is grounded in two key components: Metric Cognitive Map (Metric-CogMap) and Cognitive Chain-of-Thought (Cog-CoT). Metric-CogMap provides a unified spatial representation by integrating a discrete grid for relational reasoning with a continuous, metric-scale representation for precise geometric understanding. Building upon the Metric-CogMap, Cog-CoT performs explicit geometric reasoning through deterministic operations, including vector operations, bounding-box distances, and occlusion-aware appearance order cues, producing interpretable inference traces grounded in 3D structure. Experimental results show that Map2Thought enables explainable 3D understanding, achieving 59.9% accuracy using only half the supervision, closely matching the 60.9% baseline trained with the full dataset. It consistently outperforms state-of-the-art methods by 5.3%, 4.8%, and 4.0% under 10%, 25%, and 50% training subsets, respectively, on the VSI-Bench.", "date": "2026-01-16", "categories": []}
{"id": "2508.13601", "title": "Unleashing Semantic and Geometric Priors for 3D Scene Completion", "abstract": "Camera-based 3D semantic scene completion (SSC) provides dense geometric and semantic perception for autonomous driving and robotic navigation. However, existing methods rely on a coupled encoder to deliver both semantic and geometric priors, which forces the model to make a trade-off between conflicting demands and limits its overall performance. To tackle these challenges, we propose FoundationSSC, a novel framework that performs dual decoupling at both the source and pathway levels. At the source level, we introduce a foundation encoder that provides rich semantic feature priors for the semantic branch and high-fidelity stereo cost volumes for the geometric branch. At the pathway level, these priors are refined through specialised, decoupled pathways, yielding superior semantic context and depth distributions. Our dual-decoupling design produces disentangled and refined inputs, which are then utilised by a hybrid view transformation to generate complementary 3D features. Additionally, we introduce a novel Axis-Aware Fusion (AAF) module that addresses the often-overlooked challenge of fusing these features by anisotropically merging them into a unified representation. Extensive experiments demonstrate the advantages of FoundationSSC, achieving simultaneous improvements in both semantic and geometric metrics, surpassing prior bests by +0.23 mIoU and +2.03 IoU on SemanticKITTI. Additionally, we achieve state-of-the-art performance on SSCBench-KITTI-360, with 21.78 mIoU and 48.61 IoU.", "date": "2025-08-19", "categories": []}
{"id": "2601.09575", "title": "OpenVoxel: Training-Free Grouping and Captioning Voxels for Open-Vocabulary 3D Scene Understanding", "abstract": "We propose OpenVoxel, a training-free algorithm for grouping and captioning sparse voxels for the open-vocabulary 3D scene understanding tasks. Given the sparse voxel rasterization (SVR) model obtained from multi-view images of a 3D scene, our OpenVoxel is able to produce meaningful groups that describe different objects in the scene. Also, by leveraging powerful Vision Language Models (VLMs) and Multi-modal Large Language Models (MLLMs), our OpenVoxel successfully build an informative scene map by captioning each group, enabling further 3D scene understanding tasks such as open-vocabulary segmentation (OVS) or referring expression segmentation (RES). Unlike previous methods, our method is training-free and does not introduce embeddings from a CLIP/BERT text encoder. Instead, we directly proceed with text-to-text search using MLLMs. Through extensive experiments, our method demonstrates superior performance compared to recent studies, particularly in complex referring expression segmentation (RES) tasks. The code will be open.", "date": "2026-01-14", "categories": []}
{"id": "2601.08175", "title": "CogniMap3D: Cognitive 3D Mapping and Rapid Retrieval", "abstract": "We present CogniMap3D, a bioinspired framework for dynamic 3D scene understanding and reconstruction that emulates human cognitive processes. Our approach maintains a persistent memory bank of static scenes, enabling efficient spatial knowledge storage and rapid retrieval. CogniMap3D integrates three core capabilities: a multi-stage motion cue framework for identifying dynamic objects, a cognitive mapping system for storing, recalling, and updating static scenes across multiple visits, and a factor graph optimization strategy for refining camera poses. Given an image stream, our model identifies dynamic regions through motion cues with depth and camera pose priors, then matches static elements against its memory bank. When revisiting familiar locations, CogniMap3D retrieves stored scenes, relocates cameras, and updates memory with new observations. Evaluations on video depth estimation, camera pose reconstruction, and 3D mapping tasks demonstrate its state-of-the-art performance, while effectively supporting continuous scene understanding across extended sequences and multiple visits.", "date": "2026-01-13", "categories": []}
{"id": "2411.19290", "title": "TRASE: Tracking-free 4D Segmentation and Editing", "abstract": "Understanding dynamic 3D scenes is crucial for extended reality (XR) and autonomous driving. Incorporating semantic information into 3D reconstruction enables holistic scene representations, unlocking immersive and interactive applications. To this end, we introduce TRASE, a novel tracking-free 4D segmentation method for dynamic scene understanding. TRASE learns a 4D segmentation feature field in a weakly-supervised manner, leveraging a soft-mined contrastive learning objective guided by SAM masks. The resulting feature space is semantically coherent and well-separated, and final object-level segmentation is obtained via unsupervised clustering. This enables fast editing, such as object removal, composition, and style transfer, by directly manipulating the scene's Gaussians. We evaluate TRASE on five dynamic benchmarks, demonstrating state-of-the-art segmentation performance from unseen viewpoints and its effectiveness across various interactive editing tasks. Our project page is available at: https://yunjinli.github.io/project-sadg/", "date": "2024-11-28", "categories": []}
{"id": "2601.02905", "title": "LOST-3DSG: Lightweight Open-Vocabulary 3D Scene Graphs with Semantic Tracking in Dynamic Environments", "abstract": "Tracking objects that move within dynamic environments is a core challenge in robotics. Recent research has advanced this topic significantly; however, many existing approaches remain inefficient due to their reliance on heavy foundation models. To address this limitation, we propose LOST-3DSG, a lightweight open-vocabulary 3D scene graph designed to track dynamic objects in real-world environments. Our method adopts a semantic approach to entity tracking based on word2vec and sentence embeddings, enabling an open-vocabulary representation while avoiding the necessity of storing dense CLIP visual features. As a result, LOST-3DSG achieves superior performance compared to approaches that rely on high-dimensional visual embeddings. We evaluate our method through qualitative and quantitative experiments conducted in a real 3D environment using a TIAGo robot. The results demonstrate the effectiveness and efficiency of LOST-3DSG in dynamic object tracking. Code and supplementary material are publicly available on the project website at https://lab-rococo-sapienza.github.io/lost-3dsg/.", "date": "2026-01-06", "categories": []}
{"id": "2601.06415", "title": "Semantic Enrichment of CAD-Based Industrial Environments via Scene Graphs for Simulation and Reasoning", "abstract": "Utilizing functional elements in an industrial environment, such as displays and interactive valves, provide effective possibilities for robot training. When preparing simulations for robots or applications that involve high-level scene understanding, the simulation environment must be equally detailed. Although CAD files for such environments deliver an exact description of the geometry and visuals, they usually lack semantic, relational and functional information, thus limiting the simulation and training possibilities. A 3D scene graph can organize semantic, spatial and functional information by enriching the environment through a Large Vision-Language Model (LVLM). In this paper we present an offline approach to creating detailed 3D scene graphs from CAD environments. This will serve as a foundation to include the relations of functional and actionable elements, which then can be used for dynamic simulation and reasoning. Key results of this research include both quantitative results of the generated semantic labels as well as qualitative results of the scene graph, especially in hindsight of pipe structures and identified functional relations. All code, results and the environment will be made available at https://cad-scenegraph.github.io", "date": "2026-01-10", "categories": []}
{"id": "2601.06496", "title": "3D CoCa v2: Contrastive Learners with Test-Time Search for Generalizable Spatial Intelligence", "abstract": "Spatial intelligence refers to the ability to perceive, reason about, and describe objects and their relationships within three-dimensional environments, forming a foundation for embodied perception and scene understanding. 3D captioning aims to describe 3D scenes in natural language; however, it remains challenging due to the sparsity and irregularity of point clouds and, more critically, the weak grounding and limited out-of-distribution (OOD) generalization of existing captioners across drastically different environments, including indoor and outdoor 3D scenes. To address this challenge, we propose 3D CoCa v2, a generalizable 3D captioning framework that unifies contrastive vision-language learning with 3D caption generation and further improves robustness via test-time search (TTS) without updating the captioner parameters. 3D CoCa v2 builds on a frozen CLIP-based semantic prior, a spatially-aware 3D scene encoder for geometry, and a multimodal decoder jointly optimized with contrastive and captioning objectives, avoiding external detectors or handcrafted proposals. At inference, TTS produces diverse caption candidates and performs reward-guided selection using a compact scene summary. Experiments show improvements over 3D CoCa of +1.50 CIDEr@0.5IoU on ScanRefer and +1.61 CIDEr@0.5IoU on Nr3D, and +3.8 CIDEr@0.25 in zero-shot OOD evaluation on TOD3Cap. Code will be released at https://github.com/AIGeeksGroup/3DCoCav2.", "date": "2026-01-10", "categories": []}
{"id": "2508.13287", "title": "InnerGS: Internal Scenes Reconstruction and Segmentation via Factorized 3D Gaussian Splatting", "abstract": "3D Gaussian Splatting (3DGS) has recently gained popularity for efficient scene rendering by representing scenes as explicit sets of anisotropic 3D Gaussians. However, most existing work focuses primarily on modeling external surfaces. In this work, we target the reconstruction of internal scenes, which is crucial for applications that require a deep understanding of an object's interior. By directly modeling a continuous volumetric density through the inner 3D Gaussian distribution, our model effectively reconstructs smooth and detailed internal structures from sparse sliced data. Beyond high-fidelity reconstruction, we further demonstrate the framework's potential for downstream tasks such as segmentation. By integrating language features, we extend our approach to enable text-guided segmentation of medical scenes via natural language queries. Our approach eliminates the need for camera poses, is plug-and-play, and is inherently compatible with any data modalities. We provide cuda implementation at: https://github.com/Shuxin-Liang/InnerGS.", "date": "2025-08-18", "categories": []}
{"id": "2601.04860", "title": "DivAS: Interactive 3D Segmentation of NeRFs via Depth-Weighted Voxel Aggregation", "abstract": "Existing methods for segmenting Neural Radiance Fields (NeRFs) are often optimization-based, requiring slow per-scene training that sacrifices the zero-shot capabilities of 2D foundation models. We introduce DivAS (Depth-interactive Voxel Aggregation Segmentation), an optimization-free, fully interactive framework that addresses these limitations. Our method operates via a fast GUI-based workflow where 2D SAM masks, generated from user point prompts, are refined using NeRF-derived depth priors to improve geometric accuracy and foreground-background separation. The core of our contribution is a custom CUDA kernel that aggregates these refined multi-view masks into a unified 3D voxel grid in under 200ms, enabling real-time visual feedback. This optimization-free design eliminates the need for per-scene training. Experiments on Mip-NeRF 360{\\deg} and LLFF show that DivAS achieves segmentation quality comparable to optimization-based methods, while being 2-2.5x faster end-to-end, and up to an order of magnitude faster when excluding user prompting time.", "date": "2026-01-08", "categories": []}
{"id": "2512.13683", "title": "I-Scene: 3D Instance Models are Implicit Generalizable Spatial Learners", "abstract": "Generalization remains the central challenge for interactive 3D scene generation. Existing learning-based approaches ground spatial understanding in limited scene dataset, restricting generalization to new layouts. We instead reprogram a pre-trained 3D instance generator to act as a scene level learner, replacing dataset-bounded supervision with model-centric spatial supervision. This reprogramming unlocks the generator transferable spatial knowledge, enabling generalization to unseen layouts and novel object compositions. Remarkably, spatial reasoning still emerges even when the training scenes are randomly composed objects. This demonstrates that the generator's transferable scene prior provides a rich learning signal for inferring proximity, support, and symmetry from purely geometric cues. Replacing widely used canonical space, we instantiate this insight with a view-centric formulation of the scene space, yielding a fully feed-forward, generalizable scene generator that learns spatial relations directly from the instance model. Quantitative and qualitative results show that a 3D instance generator is an implicit spatial learner and reasoner, pointing toward foundation models for interactive 3D scene understanding and generation. Project page: https://luling06.github.io/I-Scene-project/", "date": "2025-12-15", "categories": []}
{"id": "2512.22351", "title": "VULCAN: Tool-Augmented Multi Agents for Iterative 3D Object Arrangement", "abstract": "Despite the remarkable progress of Multimodal Large Language Models (MLLMs) in 2D vision-language tasks, their application to complex 3D scene manipulation remains underexplored. In this paper, we bridge this critical gap by tackling three key challenges in 3D object arrangement task using MLLMs. First, to address the weak visual grounding of MLLMs, which struggle to link programmatic edits with precise 3D outcomes, we introduce an MCP-based API. This shifts the interaction from brittle raw code manipulation to more robust, function-level updates. Second, we augment the MLLM's 3D scene understanding with a suite of specialized visual tools to analyze scene state, gather spatial information, and validate action outcomes. This perceptual feedback loop is critical for closing the gap between language-based updates and precise 3D-aware manipulation. Third, to manage the iterative, error-prone updates, we propose a collaborative multi-agent framework with designated roles for planning, execution, and verification. This decomposition allows the system to robustly handle multi-step instructions and recover from intermediate errors. We demonstrate the effectiveness of our approach on a diverse set of 25 complex object arrangement tasks, where it significantly outperforms existing baselines. Website: vulcan-3d.github.io", "date": "2025-12-26", "categories": []}
{"id": "2506.08002", "title": "Aligning Text, Images, and 3D Structure Token-by-Token", "abstract": "Creating machines capable of understanding the world in 3D is essential in assisting designers that build and edit 3D environments and robots navigating and interacting within a three-dimensional space. Inspired by advances in language and image modeling, we investigate the potential of autoregressive models for a new modality: structured 3D scenes. To this end, we propose a unified LLM framework that aligns language, images, and 3D scenes and provide a detailed ''cookbook'' outlining critical design choices for achieving optimal training and performance addressing key questions related to data representation, modality-specific objectives, and more. We show how to tokenize complex 3D objects to incorporate into our structured 3D scene modality. We evaluate performance across four core 3D tasks -- rendering, recognition, instruction-following, and question-answering -- and four 3D datasets, synthetic and real-world. We show our model's effectiveness on reconstructing complete 3D scenes consisting of complex objects from a single image and on real-world 3D object recognition tasks. Project webpage: https://glab-caltech.github.io/kyvo/", "date": "2025-06-09", "categories": []}
{"id": "2601.02457", "title": "PatchAlign3D: Local Feature Alignment for Dense 3D Shape understanding", "abstract": "Current foundation models for 3D shapes excel at global tasks (retrieval, classification) but transfer poorly to local part-level reasoning. Recent approaches leverage vision and language foundation models to directly solve dense tasks through multi-view renderings and text queries. While promising, these pipelines require expensive inference over multiple renderings, depend heavily on large language-model (LLM) prompt engineering for captions, and fail to exploit the inherent 3D geometry of shapes. We address this gap by introducing an encoder-only 3D model that produces language-aligned patch-level features directly from point clouds. Our pre-training approach builds on existing data engines that generate part-annotated 3D shapes by pairing multi-view SAM regions with VLM captioning. Using this data, we train a point cloud transformer encoder in two stages: (1) distillation of dense 2D features from visual encoders such as DINOv2 into 3D patches, and (2) alignment of these patch embeddings with part-level text embeddings through a multi-positive contrastive objective. Our 3D encoder achieves zero-shot 3D part segmentation with fast single-pass inference without any test-time multi-view rendering, while significantly outperforming previous rendering-based and feed-forward approaches across several 3D part segmentation benchmarks. Project website: https://souhail-hadgi.github.io/patchalign3dsite/", "date": "2026-01-05", "categories": []}
{"id": "2506.04308", "title": "RoboRefer: Towards Spatial Referring with Reasoning in Vision-Language Models for Robotics", "abstract": "Spatial referring is a fundamental capability of embodied robots to interact with the 3D physical world. However, even with the powerful pretrained vision language models (VLMs), recent approaches are still not qualified to accurately understand the complex 3D scenes and dynamically reason about the instruction-indicated locations for interaction. To this end, we propose RoboRefer, a 3D-aware VLM that can first achieve precise spatial understanding by integrating a disentangled but dedicated depth encoder via supervised fine-tuning (SFT). Moreover, RoboRefer advances generalized multi-step spatial reasoning via reinforcement fine-tuning (RFT), with metric-sensitive process reward functions tailored for spatial referring tasks. To support SFT and RFT training, we introduce RefSpatial, a large-scale dataset of 20M QA pairs (2x prior), covering 31 spatial relations (vs. 15 prior) and supporting complex reasoning processes (up to 5 steps). In addition, we introduce RefSpatial-Bench, a challenging benchmark filling the gap in evaluating spatial referring with multi-step reasoning. Experiments show that SFT-trained RoboRefer achieves state-of-the-art spatial understanding, with an average success rate of 89.6%. RFT-trained RoboRefer further outperforms all other baselines by a large margin, even surpassing Gemini-2.5-Pro by 17.4% in average accuracy on RefSpatial-Bench. Notably, RoboRefer can be integrated with various control policies to execute long-horizon, dynamic tasks across diverse robots (e,g., UR5, G1 humanoid) in cluttered real-world scenes.", "date": "2025-06-04", "categories": []}
{"id": "2306.05242", "title": "Efficient Multi-Task Scene Analysis with RGB-D Transformers", "abstract": "Scene analysis is essential for enabling autonomous systems, such as mobile robots, to operate in real-world environments. However, obtaining a comprehensive understanding of the scene requires solving multiple tasks, such as panoptic segmentation, instance orientation estimation, and scene classification. Solving these tasks given limited computing and battery capabilities on mobile platforms is challenging. To address this challenge, we introduce an efficient multi-task scene analysis approach, called EMSAFormer, that uses an RGB-D Transformer-based encoder to simultaneously perform the aforementioned tasks. Our approach builds upon the previously published EMSANet. However, we show that the dual CNN-based encoder of EMSANet can be replaced with a single Transformer-based encoder. To achieve this, we investigate how information from both RGB and depth data can be effectively incorporated in a single encoder. To accelerate inference on robotic hardware, we provide a custom NVIDIA TensorRT extension enabling highly optimization for our EMSAFormer approach. Through extensive experiments on the commonly used indoor datasets NYUv2, SUNRGB-D, and ScanNet, we show that our approach achieves state-of-the-art performance while still enabling inference with up to 39.1 FPS on an NVIDIA Jetson AGX Orin 32 GB.", "date": "2023-06-08", "categories": []}
{"id": "2508.02261", "title": "SplatSSC: Decoupled Depth-Guided Gaussian Splatting for Semantic Scene Completion", "abstract": "Monocular 3D Semantic Scene Completion (SSC) is a challenging yet promising task that aims to infer dense geometric and semantic descriptions of a scene from a single image. While recent object-centric paradigms significantly improve efficiency by leveraging flexible 3D Gaussian primitives, they still rely heavily on a large number of randomly initialized primitives, which inevitably leads to 1) inefficient primitive initialization and 2) outlier primitives that introduce erroneous artifacts. In this paper, we propose SplatSSC, a novel framework that resolves these limitations with a depth-guided initialization strategy and a principled Gaussian aggregator. Instead of random initialization, SplatSSC utilizes a dedicated depth branch composed of a Group-wise Multi-scale Fusion (GMF) module, which integrates multi-scale image and depth features to generate a sparse yet representative set of initial Gaussian primitives. To mitigate noise from outlier primitives, we develop the Decoupled Gaussian Aggregator (DGA), which enhances robustness by decomposing geometric and semantic predictions during the Gaussian-to-voxel splatting process. Complemented with a specialized Probability Scale Loss, our method achieves state-of-the-art performance on the Occ-ScanNet dataset, outperforming prior approaches by over 6.3% in IoU and 4.1% in mIoU, while reducing both latency and memory cost by more than 9.3%.", "date": "2025-08-04", "categories": []}
{"id": "2510.23928", "title": "Adaptive Keyframe Selection for Scalable 3D Scene Reconstruction in Dynamic Environments", "abstract": "In this paper, we propose an adaptive keyframe selection method for improved 3D scene reconstruction in dynamic environments. The proposed method integrates two complementary modules: an error-based selection module utilizing photometric and structural similarity (SSIM) errors, and a momentum-based update module that dynamically adjusts keyframe selection thresholds according to scene motion dynamics. By dynamically curating the most informative frames, our approach addresses a key data bottleneck in real-time perception. This allows for the creation of high-quality 3D world representations from a compressed data stream, a critical step towards scalable robot learning and deployment in complex, dynamic environments. Experimental results demonstrate significant improvements over traditional static keyframe selection strategies, such as fixed temporal intervals or uniform frame skipping. These findings highlight a meaningful advancement toward adaptive perception systems that can dynamically respond to complex and evolving visual scenes. We evaluate our proposed adaptive keyframe selection module on two recent state-of-the-art 3D reconstruction networks, Spann3r and CUT3R, and observe consistent improvements in reconstruction quality across both frameworks. Furthermore, an extensive ablation study confirms the effectiveness of each individual component in our method, underlining their contribution to the overall performance gains.", "date": "2025-10-27", "categories": []}
{"id": "2512.19871", "title": "HyGE-Occ: Hybrid View-Transformation with 3D Gaussian and Edge Priors for 3D Panoptic Occupancy Prediction", "abstract": "3D Panoptic Occupancy Prediction aims to reconstruct a dense volumetric scene map by predicting the semantic class and instance identity of every occupied region in 3D space. Achieving such fine-grained 3D understanding requires precise geometric reasoning and spatially consistent scene representation across complex environments. However, existing approaches often struggle to maintain precise geometry and capture the precise spatial range of 3D instances critical for robust panoptic separation. To overcome these limitations, we introduce HyGE-Occ, a novel framework that leverages a hybrid view-transformation branch with 3D Gaussian and edge priors to enhance both geometric consistency and boundary awareness in 3D panoptic occupancy prediction. HyGE-Occ employs a hybrid view-transformation branch that fuses a continuous Gaussian-based depth representation with a discretized depth-bin formulation, producing BEV features with improved geometric consistency and structural coherence. In parallel, we extract edge maps from BEV features and use them as auxiliary information to learn edge cues. In our extensive experiments on the Occ3D-nuScenes dataset, HyGE-Occ outperforms existing work, demonstrating superior 3D geometric reasoning.", "date": "2025-12-22", "categories": []}
{"id": "2511.13031", "title": "Towards 3D Object-Centric Feature Learning for Semantic Scene Completion", "abstract": "Vision-based 3D Semantic Scene Completion (SSC) has received growing attention due to its potential in autonomous driving. While most existing approaches follow an ego-centric paradigm by aggregating and diffusing features over the entire scene, they often overlook fine-grained object-level details, leading to semantic and geometric ambiguities, especially in complex environments. To address this limitation, we propose Ocean, an object-centric prediction framework that decomposes the scene into individual object instances to enable more accurate semantic occupancy prediction. Specifically, we first employ a lightweight segmentation model, MobileSAM, to extract instance masks from the input image. Then, we introduce a 3D Semantic Group Attention module that leverages linear attention to aggregate object-centric features in 3D space. To handle segmentation errors and missing instances, we further design a Global Similarity-Guided Attention module that leverages segmentation features for global interaction. Finally, we propose an Instance-aware Local Diffusion module that improves instance features through a generative process and subsequently refines the scene representation in the BEV space. Extensive experiments on the SemanticKITTI and SSCBench-KITTI360 benchmarks demonstrate that Ocean achieves state-of-the-art performance, with mIoU scores of 17.40 and 20.28, respectively.", "date": "2025-11-17", "categories": []}
{"id": "2512.17955", "title": "A Modular Framework for Single-View 3D Reconstruction of Indoor Environments", "abstract": "We propose a modular framework for single-view indoor scene 3D reconstruction, where several core modules are powered by diffusion techniques. Traditional approaches for this task often struggle with the complex instance shapes and occlusions inherent in indoor environments. They frequently overshoot by attempting to predict 3D shapes directly from incomplete 2D images, which results in limited reconstruction quality. We aim to overcome this limitation by splitting the process into two steps: first, we employ diffusion-based techniques to predict the complete views of the room background and occluded indoor instances, then transform them into 3D. Our modular framework makes contributions to this field through the following components: an amodal completion module for restoring the full view of occluded instances, an inpainting model specifically trained to predict room layouts, a hybrid depth estimation technique that balances overall geometric accuracy with fine detail expressiveness, and a view-space alignment method that exploits both 2D and 3D cues to ensure precise placement of instances within the scene. This approach effectively reconstructs both foreground instances and the room background from a single image. Extensive experiments on the 3D-Front dataset demonstrate that our method outperforms current state-of-the-art (SOTA) approaches in terms of both visual quality and reconstruction accuracy. The framework holds promising potential for applications in interior design, real estate, and augmented reality.", "date": "2025-12-17", "categories": []}
{"id": "2512.16077", "title": "Auto-Vocabulary 3D Object Detection", "abstract": "Open-vocabulary 3D object detection methods are able to localize 3D boxes of classes unseen during training. Despite the name, existing methods rely on user-specified classes both at training and inference. We propose to study Auto-Vocabulary 3D Object Detection (AV3DOD), where the classes are automatically generated for the detected objects without any user input. To this end, we introduce Semantic Score (SS) to evaluate the quality of the generated class names. We then develop a novel framework, AV3DOD, which leverages 2D vision-language models (VLMs) to generate rich semantic candidates through image captioning, pseudo 3D box generation, and feature-space semantics expansion. AV3DOD achieves the state-of-the-art (SOTA) performance on both localization (mAP) and semantic quality (SS) on the ScanNetV2 and SUNRGB-D datasets. Notably, it surpasses the SOTA, CoDA, by 3.48 overall mAP and attains a 24.5% relative improvement in SS on ScanNetV2.", "date": "2025-12-18", "categories": []}
{"id": "2512.15621", "title": "OccSTeP: Benchmarking 4D Occupancy Spatio-Temporal Persistence", "abstract": "Autonomous driving requires a persistent understanding of 3D scenes that is robust to temporal disturbances and accounts for potential future actions. We introduce a new concept of 4D Occupancy Spatio-Temporal Persistence (OccSTeP), which aims to address two tasks: (1) reactive forecasting: ''what will happen next'' and (2) proactive forecasting: \"what would happen given a specific future action\". For the first time, we create a new OccSTeP benchmark with challenging scenarios (e.g., erroneous semantic labels and dropped frames). To address this task, we propose OccSTeP-WM, a tokenizer-free world model that maintains a dense voxel-based scene state and incrementally fuses spatio-temporal context over time. OccSTeP-WM leverages a linear-complexity attention backbone and a recurrent state-space module to capture long-range spatial dependencies while continually updating the scene memory with ego-motion compensation. This design enables online inference and robust performance even when historical sensor input is missing or noisy. Extensive experiments prove the effectiveness of the OccSTeP concept and our OccSTeP-WM, yielding an average semantic mIoU of 23.70% (+6.56% gain) and occupancy IoU of 35.89% (+9.26% gain). The data and code will be open source at https://github.com/FaterYU/OccSTeP.", "date": "2025-12-17", "categories": []}
{"id": "2512.13177", "title": "MMDrive: Interactive Scene Understanding Beyond Vision with Multi-representational Fusion", "abstract": "Vision-language models enable the understanding and reasoning of complex traffic scenarios through multi-source information fusion, establishing it as a core technology for autonomous driving. However, existing vision-language models are constrained by the image understanding paradigm in 2D plane, which restricts their capability to perceive 3D spatial information and perform deep semantic fusion, resulting in suboptimal performance in complex autonomous driving environments. This study proposes MMDrive, an multimodal vision-language model framework that extends traditional image understanding to a generalized 3D scene understanding framework. MMDrive incorporates three complementary modalities, including occupancy maps, LiDAR point clouds, and textual scene descriptions. To this end, it introduces two novel components for adaptive cross-modal fusion and key information extraction. Specifically, the Text-oriented Multimodal Modulator dynamically weights the contributions of each modality based on the semantic cues in the question, guiding context-aware feature integration. The Cross-Modal Abstractor employs learnable abstract tokens to generate compact, cross-modal summaries that highlight key regions and essential semantics. Comprehensive evaluations on the DriveLM and NuScenes-QA benchmarks demonstrate that MMDrive achieves significant performance gains over existing vision-language models for autonomous driving, with a BLEU-4 score of 54.56 and METEOR of 41.78 on DriveLM, and an accuracy score of 62.7% on NuScenes-QA. MMDrive effectively breaks the traditional image-only understanding barrier, enabling robust multimodal reasoning in complex driving environments and providing a new foundation for interpretable autonomous driving scene understanding.", "date": "2025-12-15", "categories": []}
{"id": "2512.14020", "title": "Deep Learning Perspective of Scene Understanding in Autonomous Robots", "abstract": "This paper provides a review of deep learning applications in scene understanding in autonomous robots, including innovations in object detection, semantic and instance segmentation, depth estimation, 3D reconstruction, and visual SLAM. It emphasizes how these techniques address limitations of traditional geometric models, improve depth perception in real time despite occlusions and textureless surfaces, and enhance semantic reasoning to understand the environment better. When these perception modules are integrated into dynamic and unstructured environments, they become more effective in decisionmaking, navigation and interaction. Lastly, the review outlines the existing problems and research directions to advance learning-based scene understanding of autonomous robots.", "date": "2025-12-16", "categories": []}
{"id": "2512.14126", "title": "Consistent Instance Field for Dynamic Scene Understanding", "abstract": "We introduce Consistent Instance Field, a continuous and probabilistic spatio-temporal representation for dynamic scene understanding. Unlike prior methods that rely on discrete tracking or view-dependent features, our approach disentangles visibility from persistent object identity by modeling each space-time point with an occupancy probability and a conditional instance distribution. To realize this, we introduce a novel instance-embedded representation based on deformable 3D Gaussians, which jointly encode radiance and semantic information and are learned directly from input RGB images and instance masks through differentiable rasterization. Furthermore, we introduce new mechanisms to calibrate per-Gaussian identities and resample Gaussians toward semantically active regions, ensuring consistent instance representations across space and time. Experiments on HyperNeRF and Neu3D datasets demonstrate that our method significantly outperforms state-of-the-art methods on novel-view panoptic segmentation and open-vocabulary 4D querying tasks.", "date": "2025-12-16", "categories": []}
{"id": "2512.12683", "title": "Quantum Implicit Neural Representations for 3D Scene Reconstruction and Novel View Synthesis", "abstract": "Implicit neural representations (INRs) have become a powerful paradigm for continuous signal modeling and 3D scene reconstruction, yet classical networks suffer from a well-known spectral bias that limits their ability to capture high-frequency details. Quantum Implicit Representation Networks (QIREN) mitigate this limitation by employing parameterized quantum circuits with inherent Fourier structures, enabling compact and expressive frequency modeling beyond classical MLPs. In this paper, we present Quantum Neural Radiance Fields (Q-NeRF), the first hybrid quantum-classical framework for neural radiance field rendering. Q-NeRF integrates QIREN modules into the Nerfacto backbone, preserving its efficient sampling, pose refinement, and volumetric rendering strategies while replacing selected density and radiance prediction components with quantum-enhanced counterparts. We systematically evaluate three hybrid configurations on standard multi-view indoor datasets, comparing them to classical baselines using PSNR, SSIM, and LPIPS metrics. Results show that hybrid quantum-classical models achieve competitive reconstruction quality under limited computational resources, with quantum modules particularly effective in representing fine-scale, view-dependent appearance. Although current implementations rely on quantum circuit simulators constrained to few-qubit regimes, the results highlight the potential of quantum encodings to alleviate spectral bias in implicit representations. Q-NeRF provides a foundational step toward scalable quantum-enabled 3D scene reconstruction and a baseline for future quantum neural rendering research.", "date": "2025-12-14", "categories": []}
{"id": "2512.12945", "title": "SLIM-VDB: A Real-Time 3D Probabilistic Semantic Mapping Framework", "abstract": "This paper introduces SLIM-VDB, a new lightweight semantic mapping system with probabilistic semantic fusion for closed-set or open-set dictionaries. Advances in data structures from the computer graphics community, such as OpenVDB, have demonstrated significantly improved computational and memory efficiency in volumetric scene representation. Although OpenVDB has been used for geometric mapping in robotics applications, semantic mapping for scene understanding with OpenVDB remains unexplored. In addition, existing semantic mapping systems lack support for integrating both fixed-category and open-language label predictions within a single framework. In this paper, we propose a novel 3D semantic mapping system that leverages the OpenVDB data structure and integrates a unified Bayesian update framework for both closed- and open-set semantic fusion. Our proposed framework, SLIM-VDB, achieves significant reduction in both memory and integration times compared to current state-of-the-art semantic mapping approaches, while maintaining comparable mapping accuracy. An open-source C++ codebase with a Python interface is available at https://github.com/umfieldrobotics/slim-vdb.", "date": "2025-12-15", "categories": []}
{"id": "2512.09062", "title": "SIP: Site in Pieces- A Dataset of Disaggregated Construction-Phase 3D Scans for Semantic Segmentation and Scene Understanding", "abstract": "Accurate 3D scene interpretation in active construction sites is essential for progress monitoring, safety assessment, and digital twin development. LiDAR is widely used in construction because it offers advantages over camera-based systems, performing reliably in cluttered and dynamically changing conditions. Yet most public datasets for 3D perception are derived from densely fused scans with uniform sampling and complete visibility, conditions that do not reflect real construction sites. Field data are often collected as isolated single-station LiDAR views, constrained by safety requirements, limited access, and ongoing operations. These factors lead to radial density decay, fragmented geometry, and view-dependent visibility-characteristics that remain underrepresented in existing datasets. This paper presents SIP, Site in Pieces, a dataset created to reflect the practical constraints of LiDAR acquisition during construction. SIP provides indoor and outdoor scenes captured with a terrestrial LiDAR scanner and annotated at the point level using a taxonomy tailored to construction environments: A. Built Environment, B. Construction Operations, and C. Site Surroundings. The dataset includes both structural components and slender temporary objects such as scaffolding, MEP piping, and scissor lifts, where sparsity caused by occlusion and fragmented geometry make segmentation particularly challenging. The scanning protocol, annotation workflow, and quality control procedures establish a consistent foundation for the dataset. SIP is openly available with a supporting Git repository, offering adaptable class configurations that streamline adoption within modern 3D deep learning frameworks. By providing field data that retain real-world sensing characteristics, SIP enables robust benchmarking and contributes to advancing construction-oriented 3D vision tasks.", "date": "2025-12-09", "categories": []}
{"id": "2512.09364", "title": "ASSIST-3D: Adapted Scene Synthesis for Class-Agnostic 3D Instance Segmentation", "abstract": "Class-agnostic 3D instance segmentation tackles the challenging task of segmenting all object instances, including previously unseen ones, without semantic class reliance. Current methods struggle with generalization due to the scarce annotated 3D scene data or noisy 2D segmentations. While synthetic data generation offers a promising solution, existing 3D scene synthesis methods fail to simultaneously satisfy geometry diversity, context complexity, and layout reasonability, each essential for this task. To address these needs, we propose an Adapted 3D Scene Synthesis pipeline for class-agnostic 3D Instance SegmenTation, termed as ASSIST-3D, to synthesize proper data for model generalization enhancement. Specifically, ASSIST-3D features three key innovations, including 1) Heterogeneous Object Selection from extensive 3D CAD asset collections, incorporating randomness in object sampling to maximize geometric and contextual diversity; 2) Scene Layout Generation through LLM-guided spatial reasoning combined with depth-first search for reasonable object placements; and 3) Realistic Point Cloud Construction via multi-view RGB-D image rendering and fusion from the synthetic scenes, closely mimicking real-world sensor data acquisition. Experiments on ScanNetV2, ScanNet++, and S3DIS benchmarks demonstrate that models trained with ASSIST-3D-generated data significantly outperform existing methods. Further comparisons underscore the superiority of our purpose-built pipeline over existing 3D scene synthesis approaches.", "date": "2025-12-10", "categories": []}
{"id": "2506.01946", "title": "3DRS: MLLMs Need 3D-Aware Representation Supervision for Scene Understanding", "abstract": "Recent advances in scene understanding have leveraged multimodal large language models (MLLMs) for 3D reasoning by capitalizing on their strong 2D pretraining. However, the lack of explicit 3D data during MLLM pretraining limits 3D representation capability. In this paper, we investigate the 3D-awareness of MLLMs by evaluating multi-view correspondence and reveal a strong positive correlation between the quality of 3D-aware representation and downstream task performance. Motivated by this, we propose 3DRS, a framework that enhances MLLM 3D representation learning by introducing supervision from pretrained 3D foundation models. Our approach aligns MLLM visual features with rich 3D knowledge distilled from 3D models, effectively improving scene understanding. Extensive experiments across multiple benchmarks and MLLMs -- including visual grounding, captioning, and question answering -- demonstrate consistent performance gains. Project page: https://visual-ai.github.io/3drs", "date": "2025-06-02", "categories": []}
{"id": "2507.14555", "title": "Descrip3D: Enhancing Large Language Model-based 3D Scene Understanding with Object-Level Text Descriptions", "abstract": "Understanding 3D scenes goes beyond simply recognizing objects; it requires reasoning about the spatial and semantic relationships between them. Current 3D scene-language models often struggle with this relational understanding, particularly when visual embeddings alone do not adequately convey the roles and interactions of objects. In this paper, we introduce Descrip3D, a novel and powerful framework that explicitly encodes the relationships between objects using natural language. Unlike previous methods that rely only on 2D and 3D embeddings, Descrip3D enhances each object with a textual description that captures both its intrinsic attributes and contextual relationships. These relational cues are incorporated into the model through a dual-level integration: embedding fusion and prompt-level injection. This allows for unified reasoning across various tasks such as grounding, captioning, and question answering, all without the need for task-specific heads or additional supervision. When evaluated on five benchmark datasets, including ScanRefer, Multi3DRefer, ScanQA, SQA3D, and Scan2Cap, Descrip3D consistently outperforms strong baseline models, demonstrating the effectiveness of language-guided relational representation for understanding complex indoor scenes. Our code and data are publicly available at https://github.com/jintangxue/Descrip3D.", "date": "2025-07-19", "categories": []}
{"id": "2509.24528", "title": "CORE-3D: Context-aware Open-vocabulary Retrieval by Embeddings in 3D", "abstract": "3D scene understanding is fundamental for embodied AI and robotics, supporting reliable perception for interaction and navigation. Recent approaches achieve zero-shot, open-vocabulary 3D semantic mapping by assigning embedding vectors to 2D class-agnostic masks generated via vision-language models (VLMs) and projecting these into 3D. However, these methods often produce fragmented masks and inaccurate semantic assignments due to the direct use of raw masks, limiting their effectiveness in complex environments. To address this, we leverage SemanticSAM with progressive granularity refinement to generate more accurate and numerous object-level masks, mitigating the over-segmentation commonly observed in mask generation models such as vanilla SAM, and improving downstream 3D semantic segmentation. To further enhance semantic context, we employ a context-aware CLIP encoding strategy that integrates multiple contextual views of each mask using empirically determined weighting, providing much richer visual context. We evaluate our approach on multiple 3D scene understanding tasks, including 3D semantic segmentation and object retrieval from language queries, across several benchmark datasets. Experimental results demonstrate significant improvements over existing methods, highlighting the effectiveness of our approach.", "date": "2025-09-29", "categories": []}
{"id": "2512.05988", "title": "VG3T: Visual Geometry Grounded Gaussian Transformer", "abstract": "Generating a coherent 3D scene representation from multi-view images is a fundamental yet challenging task. Existing methods often struggle with multi-view fusion, leading to fragmented 3D representations and sub-optimal performance. To address this, we introduce VG3T, a novel multi-view feed-forward network that predicts a 3D semantic occupancy via a 3D Gaussian representation. Unlike prior methods that infer Gaussians from single-view images, our model directly predicts a set of semantically attributed Gaussians in a joint, multi-view fashion. This novel approach overcomes the fragmentation and inconsistency inherent in view-by-view processing, offering a unified paradigm to represent both geometry and semantics. We also introduce two key components, Grid-Based Sampling and Positional Refinement, to mitigate the distance-dependent density bias common in pixel-aligned Gaussian initialization methods. Our VG3T shows a notable 1.7%p improvement in mIoU while using 46% fewer primitives than the previous state-of-the-art on the nuScenes benchmark, highlighting its superior efficiency and performance.", "date": "2025-11-28", "categories": []}
{"id": "2512.06882", "title": "Hierarchical Image-Guided 3D Point Cloud Segmentation in Industrial Scenes via Multi-View Bayesian Fusion", "abstract": "Reliable 3D segmentation is critical for understanding complex scenes with dense layouts and multi-scale objects, as commonly seen in industrial environments. In such scenarios, heavy occlusion weakens geometric boundaries between objects, and large differences in object scale will cause end-to-end models fail to capture both coarse and fine details accurately. Existing 3D point-based methods require costly annotations, while image-guided methods often suffer from semantic inconsistencies across views. To address these challenges, we propose a hierarchical image-guided 3D segmentation framework that progressively refines segmentation from instance-level to part-level. Instance segmentation involves rendering a top-view image and projecting SAM-generated masks prompted by YOLO-World back onto the 3D point cloud. Part-level segmentation is subsequently performed by rendering multi-view images of each instance obtained from the previous stage and applying the same 2D segmentation and back-projection process at each view, followed by Bayesian updating fusion to ensure semantic consistency across views. Experiments on real-world factory data demonstrate that our method effectively handles occlusion and structural complexity, achieving consistently high per-class mIoU scores. Additional evaluations on public dataset confirm the generalization ability of our framework, highlighting its robustness, annotation efficiency, and adaptability to diverse 3D environments.", "date": "2025-12-07", "categories": []}
{"id": "2512.07807", "title": "Lang3D-XL: Language Embedded 3D Gaussians for Large-scale Scenes", "abstract": "Embedding a language field in a 3D representation enables richer semantic understanding of spatial environments by linking geometry with descriptive meaning. This allows for a more intuitive human-computer interaction, enabling querying or editing scenes using natural language, and could potentially improve tasks like scene retrieval, navigation, and multimodal reasoning. While such capabilities could be transformative, in particular for large-scale scenes, we find that recent feature distillation approaches cannot effectively learn over massive Internet data due to challenges in semantic feature misalignment and inefficiency in memory and runtime. To this end, we propose a novel approach to address these challenges. First, we introduce extremely low-dimensional semantic bottleneck features as part of the underlying 3D Gaussian representation. These are processed by rendering and passing them through a multi-resolution, feature-based, hash encoder. This significantly improves efficiency both in runtime and GPU memory. Second, we introduce an Attenuated Downsampler module and propose several regularizations addressing the semantic misalignment of ground truth 2D features. We evaluate our method on the in-the-wild HolyScenes dataset and demonstrate that it surpasses existing approaches in both performance and efficiency.", "date": "2025-12-08", "categories": []}
{"id": "2405.05258", "title": "Multi-Modal Data-Efficient 3D Scene Understanding for Autonomous Driving", "abstract": "Efficient data utilization is crucial for advancing 3D scene understanding in autonomous driving, where reliance on heavily human-annotated LiDAR point clouds challenges fully supervised methods. Addressing this, our study extends into semi-supervised learning for LiDAR semantic segmentation, leveraging the intrinsic spatial priors of driving scenes and multi-sensor complements to augment the efficacy of unlabeled datasets. We introduce LaserMix++, an evolved framework that integrates laser beam manipulations from disparate LiDAR scans and incorporates LiDAR-camera correspondences to further assist data-efficient learning. Our framework is tailored to enhance 3D scene consistency regularization by incorporating multi-modality, including 1) multi-modal LaserMix operation for fine-grained cross-sensor interactions; 2) camera-to-LiDAR feature distillation that enhances LiDAR feature learning; and 3) language-driven knowledge guidance generating auxiliary supervisions using open-vocabulary models. The versatility of LaserMix++ enables applications across LiDAR representations, establishing it as a universally applicable solution. Our framework is rigorously validated through theoretical analysis and extensive experiments on popular driving perception datasets. Results demonstrate that LaserMix++ markedly outperforms fully supervised alternatives, achieving comparable accuracy with five times fewer annotations and significantly improving the supervised-only baselines. This substantial advancement underscores the potential of semi-supervised approaches in reducing the reliance on extensive labeled data in LiDAR-based 3D scene understanding systems.", "date": "2024-05-08", "categories": []}
{"id": "2512.05060", "title": "4DLangVGGT: 4D Language-Visual Geometry Grounded Transformer", "abstract": "Constructing 4D language fields is crucial for embodied AI, augmented/virtual reality, and 4D scene understanding, as they provide enriched semantic representations of dynamic environments and enable open-vocabulary querying in complex scenarios. However, existing approaches to 4D semantic field construction primarily rely on scene-specific Gaussian splatting, which requires per-scene optimization, exhibits limited generalization, and is difficult to scale to real-world applications. To address these limitations, we propose 4DLangVGGT, the first Transformer-based feed-forward unified framework for 4D language grounding, that jointly integrates geometric perception and language alignment within a single architecture. 4DLangVGGT has two key components: the 4D Visual Geometry Transformer, StreamVGGT, which captures spatio-temporal geometric representations of dynamic scenes; and the Semantic Bridging Decoder (SBD), which projects geometry-aware features into a language-aligned semantic space, thereby enhancing semantic interpretability while preserving structural fidelity. Unlike prior methods that depend on costly per-scene optimization, 4DLangVGGT can be jointly trained across multiple dynamic scenes and directly applied during inference, achieving both deployment efficiency and strong generalization. This design significantly improves the practicality of large-scale deployment and establishes a new paradigm for open-vocabulary 4D scene understanding. Experiments on HyperNeRF and Neu3D datasets demonstrate that our approach not only generalizes effectively but also achieves state-of-the-art performance, achieving up to 2% gains under per-scene training and 1% improvements under multi-scene training. Our code released in https://github.com/hustvl/4DLangVGGT", "date": "2025-12-04", "categories": []}
{"id": "2512.03284", "title": "SpatialReasoner: Active Perception for Large-Scale 3D Scene Understanding", "abstract": "Spatial reasoning in large-scale 3D environments remains challenging for current vision-language models, which are typically constrained to room-scale scenarios. We introduce H$^2$U3D (Holistic House Understanding in 3D), a 3D visual question answering dataset designed for house-scale scene understanding. H$^2$U3D features multi-floor environments spanning up to three floors and 10-20 rooms, covering more than 300 m$^2$. Through an automated annotation pipeline, it constructs hierarchical coarse-to-fine visual representations and generates diverse question-answer pairs with chain-of-thought annotations. We further propose SpatialReasoner, an active perception framework that autonomously invokes spatial tools to explore 3D scenes based on textual queries. SpatialReasoner is trained through a two-stage strategy: a supervised cold start followed by reinforcement learning with an adaptive exploration reward that promotes efficient exploration while discouraging redundant operations. Extensive experiments demonstrate that SpatialReasoner achieves state-of-the-art performance on H$^2$U3D, outperforming strong baselines including GPT-4o and Gemini-2.5-Pro. Notably, our method attains superior results while using only 3-4 images in total on average, compared to baselines requiring 16+ images, highlighting the effectiveness of our coarse-to-fine active exploration paradigm.", "date": "2025-12-02", "categories": []}
{"id": "2512.03601", "title": "Motion4D: Learning 3D-Consistent Motion and Semantics for 4D Scene Understanding", "abstract": "Recent advancements in foundation models for 2D vision have substantially improved the analysis of dynamic scenes from monocular videos. However, despite their strong generalization capabilities, these models often lack 3D consistency, a fundamental requirement for understanding scene geometry and motion, thereby causing severe spatial misalignment and temporal flickering in complex 3D environments. In this paper, we present Motion4D, a novel framework that addresses these challenges by integrating 2D priors from foundation models into a unified 4D Gaussian Splatting representation. Our method features a two-part iterative optimization framework: 1) Sequential optimization, which updates motion and semantic fields in consecutive stages to maintain local consistency, and 2) Global optimization, which jointly refines all attributes for long-term coherence. To enhance motion accuracy, we introduce a 3D confidence map that dynamically adjusts the motion priors, and an adaptive resampling process that inserts new Gaussians into under-represented regions based on per-pixel RGB and semantic errors. Furthermore, we enhance semantic coherence through an iterative refinement process that resolves semantic inconsistencies by alternately optimizing the semantic fields and updating prompts of SAM2. Extensive evaluations demonstrate that our Motion4D significantly outperforms both 2D foundation models and existing 3D-based approaches across diverse scene understanding tasks, including point-based tracking, video object segmentation, and novel view synthesis. Our code is available at https://hrzhou2.github.io/motion4d-web/.", "date": "2025-12-03", "categories": []}
{"id": "2507.22052", "title": "Ov3R: Open-Vocabulary Semantic 3D Reconstruction from RGB Videos", "abstract": "We present Ov3R, a novel framework for open-vocabulary semantic 3D reconstruction from RGB video streams, designed to advance Spatial AI. The system features two key components: CLIP3R, a CLIP-informed 3D reconstruction module that predicts dense point maps from overlapping clips while embedding object-level semantics; and 2D-3D OVS, a 2D-3D open-vocabulary semantic module that lifts 2D features into 3D by learning fused descriptors integrating spatial, geometric, and semantic cues. Unlike prior methods, Ov3R incorporates CLIP semantics directly into the reconstruction process, enabling globally consistent geometry and fine-grained semantic alignment. Our framework achieves state-of-the-art performance in both dense 3D reconstruction and open-vocabulary 3D segmentation, marking a step forward toward real-time, semantics-aware Spatial AI.", "date": "2025-07-29", "categories": []}
{"id": "2512.02450", "title": "HouseLayout3D: A Benchmark and Training-Free Baseline for 3D Layout Estimation in the Wild", "abstract": "Current 3D layout estimation models are primarily trained on synthetic datasets containing simple single room or single floor environments. As a consequence, they cannot natively handle large multi floor buildings and require scenes to be split into individual floors before processing, which removes global spatial context that is essential for reasoning about structures such as staircases that connect multiple levels. In this work, we introduce HouseLayout3D, a real world benchmark designed to support progress toward full building scale layout estimation, including multiple floors and architecturally intricate spaces. We also present MultiFloor3D, a simple training free baseline that leverages recent scene understanding methods and already outperforms existing 3D layout estimation models on both our benchmark and prior datasets, highlighting the need for further research in this direction. Data and code are available at: https://houselayout3d.github.io.", "date": "2025-12-02", "categories": []}
{"id": "2512.00226", "title": "DenseScan: Advancing 3D Scene Understanding with 2D Dense Annotation", "abstract": "3D understanding is a key capability for real-world AI assistance. High-quality data plays an important role in driving the development of the 3D understanding community. Current 3D scene understanding datasets often provide geometric and instance-level information, yet they lack the rich semantic annotations necessary for nuanced visual-language tasks.In this work, we introduce DenseScan, a novel dataset with detailed multi-level descriptions generated by an automated pipeline leveraging multi-view 2D images and multimodal large language models (MLLMs). Our approach enables dense captioning of scene elements, ensuring comprehensive object-level descriptions that capture context-sensitive details. Furthermore, we extend these annotations through scenario-based question generation, producing high-level queries that integrate object properties, spatial relationships, and scene context. By coupling geometric detail with semantic richness, DenseScan broadens the range of downstream tasks, from detailed visual-language navigation to interactive question answering. Experimental results demonstrate that our method significantly enhances object-level understanding and question-answering performance in 3D environments compared to traditional annotation pipelines. We release both the annotated dataset and our annotation pipeline to facilitate future research and applications in robotics, augmented reality, and beyond. Through DenseScan, we aim to catalyze new avenues in 3D scene understanding, allowing researchers and practitioners to tackle the complexities of real-world environments with richer, more contextually aware annotations.", "date": "2025-11-28", "categories": []}
{"id": "2512.00300", "title": "TGSFormer: Scalable Temporal Gaussian Splatting for Embodied Semantic Scene Completion", "abstract": "Embodied 3D Semantic Scene Completion (SSC) infers dense geometry and semantics from continuous egocentric observations. Most existing Gaussian-based methods rely on random initialization of many primitives within predefined spatial bounds, resulting in redundancy and poor scalability to unbounded scenes. Recent depth-guided approach alleviates this issue but remains local, suffering from latency and memory overhead as scale increases. To overcome these challenges, we propose TGSFormer, a scalable Temporal Gaussian Splatting framework for embodied SSC. It maintains a persistent Gaussian memory for temporal prediction, without relying on image coherence or frame caches. For temporal fusion, a Dual Temporal Encoder jointly processes current and historical Gaussian features through confidence-aware cross-attention. Subsequently, a Confidence-aware Voxel Fusion module merges overlapping primitives into voxel-aligned representations, regulating density and maintaining compactness. Extensive experiments demonstrate that TGSFormer achieves state-of-the-art results on both local and embodied SSC benchmarks, offering superior accuracy and scalability with significantly fewer primitives while maintaining consistent long-term scene integrity. The code will be released upon acceptance.", "date": "2025-11-29", "categories": []}
{"id": "2512.00565", "title": "Describe Anything Anywhere At Any Moment", "abstract": "Computer vision and robotics applications ranging from augmented reality to robot autonomy in large-scale environments require spatio-temporal memory frameworks that capture both geometric structure for accurate language-grounding as well as semantic detail. Existing methods face a tradeoff, where producing rich open-vocabulary descriptions comes at the expense of real-time performance when these descriptions have to be grounded in 3D. To address these challenges, we propose Describe Anything, Anywhere, at Any Moment (DAAAM), a novel spatio-temporal memory framework for large-scale and real-time 4D scene understanding. DAAAM introduces a novel optimization-based frontend to infer detailed semantic descriptions from localized captioning models, such as the Describe Anything Model (DAM), leveraging batch processing to speed up inference by an order of magnitude for online processing. It leverages such semantic understanding to build a hierarchical 4D scene graph (SG), which acts as an effective globally spatially and temporally consistent memory representation. DAAAM constructs 4D SGs with detailed, geometrically grounded descriptions while maintaining real-time performance. We show that DAAAM's 4D SG interfaces well with a tool-calling agent for inference and reasoning.\n  We thoroughly evaluate DAAAM in the complex task of spatio-temporal question answering on the NaVQA benchmark and show its generalization capabilities for sequential task grounding on the SG3D benchmark. We further curate an extended OC-NaVQA benchmark for large-scale and long-time evaluations. DAAAM achieves state-of-the-art results in both tasks, improving OC-NaVQA question accuracy by 53.6%, position errors by 21.9%, temporal errors by 21.6%, and SG3D task grounding accuracy by 27.8% over the most competitive baselines, respectively. We release our data and code open-source.", "date": "2025-11-29", "categories": []}
{"id": "2512.00944", "title": "Binary-Gaussian: Compact and Progressive Representation for 3D Gaussian Segmentation", "abstract": "3D Gaussian Splatting (3D-GS) has emerged as an efficient 3D representation and a promising foundation for semantic tasks like segmentation. However, existing 3D-GS-based segmentation methods typically rely on high-dimensional category features, which introduce substantial memory overhead. Moreover, fine-grained segmentation remains challenging due to label space congestion and the lack of stable multi-granularity control mechanisms. To address these limitations, we propose a coarse-to-fine binary encoding scheme for per-Gaussian category representation, which compresses each feature into a single integer via the binary-to-decimal mapping, drastically reducing memory usage. We further design a progressive training strategy that decomposes panoptic segmentation into a series of independent sub-tasks, reducing inter-class conflicts and thereby enhancing fine-grained segmentation capability. Additionally, we fine-tune opacity during segmentation training to address the incompatibility between photometric rendering and semantic segmentation, which often leads to foreground-background confusion. Extensive experiments on multiple benchmarks demonstrate that our method achieves state-of-the-art segmentation performance while significantly reducing memory consumption and accelerating inference.", "date": "2025-11-30", "categories": []}
{"id": "2512.01008", "title": "LISA-3D: Lifting Language-Image Segmentation to 3D via Multi-View Consistency", "abstract": "Text-driven 3D reconstruction demands a mask generator that simultaneously understands open-vocabulary instructions and remains consistent across viewpoints. We present LISA-3D, a two-stage framework that lifts language-image segmentation into 3D by retrofitting the instruction-following model LISA with geometry-aware Low-Rank Adaptation (LoRA) layers and reusing a frozen SAM-3D reconstructor. During training we exploit off-the-shelf RGB-D sequences and their camera poses to build a differentiable reprojection loss that enforces cross-view agreement without requiring any additional 3D-text supervision. The resulting masks are concatenated with RGB images to form RGBA prompts for SAM-3D, which outputs Gaussian splats or textured meshes without retraining. Across ScanRefer and Nr3D, LISA-3D improves language-to-3D accuracy by up to +15.6 points over single-view baselines while adapting only 11.6M parameters. The system is modular, data-efficient, and supports zero-shot deployment on unseen categories, providing a practical recipe for language-guided 3D content creation. Our code will be available at https://github.com/binisalegend/LISA-3D.", "date": "2025-11-30", "categories": []}
{"id": "2511.22961", "title": "HMR3D: Hierarchical Multimodal Representation for 3D Scene Understanding with Large Vision-Language Model", "abstract": "Recent advances in large vision-language models (VLMs) have shown significant promise for 3D scene understanding. Existing VLM-based approaches typically align 3D scene features with the VLM's embedding space. However, this implicit alignment often yields suboptimal performance due to the scarcity of 3D data and the inherent complexity of spatial relationships in 3D environments. To address these limitations, we propose a novel hierarchical multimodal representation for 3D scene reasoning that explicitly aligns with VLMs at the input space by leveraging both multi-view images and text descriptions. The text descriptions capture spatial relationships by referencing the 3D coordinates of detected objects, while the multi-view images include a top-down perspective and four directional views (forward, left, right, and backward), ensuring comprehensive scene coverage. Additionally, we introduce a hierarchical feature representation that aggregates patch-level image features into view-level and scene-level representations, enabling the model to reason over both local and global scene context. Experimental results on both situated 3D Q&A and general 3D Q&A benchmarks demonstrate the effectiveness of our approach.", "date": "2025-11-28", "categories": []}
{"id": "2503.22087", "title": "Stream and Query-guided Feature Aggregation for Efficient and Effective 3D Occupancy Prediction", "abstract": "3D occupancy prediction has become a key perception task in autonomous driving, as it enables comprehensive scene understanding. Recent methods enhance this understanding by incorporating spatiotemporal information through multi-frame fusion, but they suffer from a trade-off: dense voxel-based representations provide high accuracy at significant computational cost, whereas sparse representations improve efficiency but lose spatial detail. To mitigate this trade-off, we introduce DuOcc, which employs a dual aggregation strategy that retains dense voxel representations to preserve spatial fidelity while maintaining high efficiency. DuOcc consists of two key components: (i) Stream-based Voxel Aggregation, which recurrently accumulates voxel features over time and refines them to suppress warping-induced distortions, preserving a clear separation between occupied and free space. (ii) Query-guided Aggregation, which complements the limitations of voxel accumulation by selectively injecting instance-level query features into the voxel regions occupied by dynamic objects. Experiments on the widely used Occ3D-nuScenes and SurroundOcc datasets demonstrate that DuOcc achieves state-of-the-art performance in real-time settings, while reducing memory usage by over 40% compared to prior methods.", "date": "2025-03-28", "categories": []}
{"id": "2511.21317", "title": "HTTM: Head-wise Temporal Token Merging for Faster VGGT", "abstract": "The Visual Geometry Grounded Transformer (VGGT) marks a significant leap forward in 3D scene reconstruction, as it is the first model that directly infers all key 3D attributes (camera poses, depths, and dense geometry) jointly in one pass. However, this joint inference mechanism requires global attention layers that perform all-to-all attention computation on tokens from all views. For reconstruction of large scenes with long-sequence inputs, this causes a significant latency bottleneck. In this paper, we propose head-wise temporal merging (HTTM), a training-free 3D token merging method for accelerating VGGT. Existing merging techniques merge tokens uniformly across different attention heads, resulting in identical tokens in the layers' output, which hinders the model's representational ability. HTTM tackles this problem by merging tokens in multi-head granularity, which preserves the uniqueness of feature tokens after head concatenation. Additionally, this enables HTTM to leverage the spatial locality and temporal correspondence observed at the head level to achieve higher merging ratios with lower merging costs compared to existing methods. Thus, HTTM achieves up to 7x acceleration with negligible performance drops in a GPU-based inference.", "date": "2025-11-26", "categories": []}
{"id": "2408.11030", "title": "OpenScan: A Benchmark for Generalized Open-Vocabulary 3D Scene Understanding", "abstract": "Open-vocabulary 3D scene understanding (OV-3D) aims to localize and classify novel objects beyond the closed set of object classes. However, existing approaches and benchmarks primarily focus on the open vocabulary problem within the context of object classes, which is insufficient in providing a holistic evaluation to what extent a model understands the 3D scene. In this paper, we introduce a more challenging task called Generalized Open-Vocabulary 3D Scene Understanding (GOV-3D) to explore the open vocabulary problem beyond object classes. It encompasses an open and diverse set of generalized knowledge, expressed as linguistic queries of fine-grained and object-specific attributes. To this end, we contribute a new benchmark named \\textit{OpenScan}, which consists of 3D object attributes across eight representative linguistic aspects, including affordance, property, and material. We further evaluate state-of-the-art OV-3D methods on our OpenScan benchmark and discover that these methods struggle to comprehend the abstract vocabularies of the GOV-3D task, a challenge that cannot be addressed simply by scaling up object classes during training. We highlight the limitations of existing methodologies and explore promising directions to overcome the identified shortcomings.", "date": "2024-08-20", "categories": []}
{"id": "2511.10003", "title": "DBGroup: Dual-Branch Point Grouping for Weakly Supervised 3D Semantic Instance Segmentation", "abstract": "Weakly supervised 3D instance segmentation is essential for 3D scene understanding, especially as the growing scale of data and high annotation costs associated with fully supervised approaches. Existing methods primarily rely on two forms of weak supervision: one-thing-one-click annotations and bounding box annotations, both of which aim to reduce labeling efforts. However, these approaches still encounter limitations, including labor-intensive annotation processes, high complexity, and reliance on expert annotators. To address these challenges, we propose \\textbf{DBGroup}, a two-stage weakly supervised 3D instance segmentation framework that leverages scene-level annotations as a more efficient and scalable alternative. In the first stage, we introduce a Dual-Branch Point Grouping module to generate pseudo labels guided by semantic and mask cues extracted from multi-view images. To further improve label quality, we develop two refinement strategies: Granularity-Aware Instance Merging and Semantic Selection and Propagation. The second stage involves multi-round self-training on an end-to-end instance segmentation network using the refined pseudo-labels. Additionally, we introduce an Instance Mask Filter strategy to address inconsistencies within the pseudo labels. Extensive experiments demonstrate that DBGroup achieves competitive performance compared to sparse-point-level supervised 3D instance segmentation methods, while surpassing state-of-the-art scene-level supervised 3D semantic segmentation approaches. Code is available at https://github.com/liuxuexun/DBGroup.", "date": "2025-11-13", "categories": []}
{"id": "2511.20646", "title": "3D-Aware Multi-Task Learning with Cross-View Correlations for Dense Scene Understanding", "abstract": "This paper addresses the challenge of training a single network to jointly perform multiple dense prediction tasks, such as segmentation and depth estimation, i.e., multi-task learning (MTL). Current approaches mainly capture cross-task relations in the 2D image space, often leading to unstructured features lacking 3D-awareness. We argue that 3D-awareness is vital for modeling cross-task correlations essential for comprehensive scene understanding. We propose to address this problem by integrating correlations across views, i.e., cost volume, as geometric consistency in the MTL network. Specifically, we introduce a lightweight Cross-view Module (CvM), shared across tasks, to exchange information across views and capture cross-view correlations, integrated with a feature from MTL encoder for multi-task predictions. This module is architecture-agnostic and can be applied to both single and multi-view data. Extensive results on NYUv2 and PASCAL-Context demonstrate that our method effectively injects geometric consistency into existing MTL methods to improve performance.", "date": "2025-11-25", "categories": []}
{"id": "2501.01119", "title": "Leverage Cross-Attention for End-to-End Open-Vocabulary Panoptic Reconstruction", "abstract": "Open-vocabulary panoptic reconstruction offers comprehensive scene understanding, enabling advances in embodied robotics and photorealistic simulation. In this paper, we propose PanopticRecon++, an end-to-end method that formulates panoptic reconstruction through a novel cross-attention perspective. This perspective models the relationship between 3D instances (as queries) and the scene's 3D embedding field (as keys) through their attention map. Unlike existing methods that separate the optimization of queries and keys or overlook spatial proximity, PanopticRecon++ introduces learnable 3D Gaussians as instance queries. This formulation injects 3D spatial priors to preserve proximity while maintaining end-to-end optimizability. Moreover, this query formulation facilitates the alignment of 2D open-vocabulary instance IDs across frames by leveraging optimal linear assignment with instance masks rendered from the queries. Additionally, we ensure semantic-instance segmentation consistency by fusing query-based instance segmentation probabilities with semantic probabilities in a novel panoptic head supervised by a panoptic loss. During training, the number of instance query tokens dynamically adapts to match the number of objects. PanopticRecon++ shows competitive performance in terms of 3D and 2D segmentation and reconstruction performance on both simulation and real-world datasets, and demonstrates a user case as a robot simulator. Our project website is at: https://yuxuan1206.github.io/panopticrecon_pp/", "date": "2025-01-02", "categories": []}
{"id": "2511.17664", "title": "CubeletWorld: A New Abstraction for Scalable 3D Modeling", "abstract": "Modern cities produce vast streams of heterogeneous data, from infrastructure maps to mobility logs and satellite imagery. However, integrating these sources into coherent spatial models for planning and prediction remains a major challenge. Existing agent-centric methods often rely on direct environmental sensing, limiting scalability and raising privacy concerns. This paper introduces CubeletWorld, a novel framework for representing and analyzing urban environments through a discretized 3D grid of spatial units called cubelets. This abstraction enables privacy-preserving modeling by embedding diverse data signals, such as infrastructure, movement, or environmental indicators, into localized cubelet states. CubeletWorld supports downstream tasks such as planning, navigation, and occupancy prediction without requiring agent-driven sensing. To evaluate this paradigm, we propose the CubeletWorld State Prediction task, which involves predicting the cubelet state using a realistic dataset containing various urban elements like streets and buildings through this discretized representation. We explore a range of modified core models suitable for our setting and analyze challenges posed by increasing spatial granularity, specifically the issue of sparsity in representation and scalability of baselines. In contrast to existing 3D occupancy prediction models, our cubelet-centric approach focuses on inferring state at the spatial unit level, enabling greater generalizability across regions and improved privacy compliance. Our results demonstrate that CubeletWorld offers a flexible and extensible framework for learning from complex urban data, and it opens up new possibilities for scalable simulation and decision support in domains such as socio-demographic modeling, environmental monitoring, and emergency response. The code and datasets can be downloaded from here.", "date": "2025-11-21", "categories": []}
{"id": "2511.17904", "title": "CUS-GS: A Compact Unified Structured Gaussian Splatting Framework for Multimodal Scene Representation", "abstract": "Recent advances in Gaussian Splatting based 3D scene representation have shown two major trends: semantics-oriented approaches that focus on high-level understanding but lack explicit 3D geometry modeling, and structure-oriented approaches that capture spatial structures yet provide limited semantic abstraction. To bridge this gap, we present CUS-GS, a compact unified structured Gaussian Splatting representation, which connects multimodal semantic features with structured 3D geometry. Specifically, we design a voxelized anchor structure that constructs a spatial scaffold, while extracting multimodal semantic features from a set of foundation models (e.g., CLIP, DINOv2, SEEM). Moreover, we introduce a multimodal latent feature allocation mechanism to unify appearance, geometry, and semantics across heterogeneous feature spaces, ensuring a consistent representation across multiple foundation models. Finally, we propose a feature-aware significance evaluation strategy to dynamically guide anchor growing and pruning, effectively removing redundant or invalid anchors while maintaining semantic integrity. Extensive experiments show that CUS-GS achieves competitive performance compared to state-of-the-art methods using as few as 6M parameters - an order of magnitude smaller than the closest rival at 35M - highlighting the excellent trade off between performance and model efficiency of the proposed framework.", "date": "2025-11-22", "categories": []}
{"id": "2511.18416", "title": "4D-VGGT: A General Foundation Model with SpatioTemporal Awareness for Dynamic Scene Geometry Estimation", "abstract": "We investigate a challenging task of dynamic scene geometry estimation, which requires representing both spatial and temporal features. Typically, existing methods align the two features into a unified latent space to model scene geometry. However, this unified paradigm suffers from potential mismatched representation due to the heterogeneous nature between spatial and temporal features. In this work, we propose 4D-VGGT, a general foundation model with divide-and-conquer spatiotemporal representation for dynamic scene geometry. Our model is divided into three aspects: 1) Multi-setting input. We design an adaptive visual grid that supports input sequences with arbitrary numbers of views and time steps. 2) Multi-level representation. We propose a cross-view global fusion for spatial representation and a cross-time local fusion for temporal representation. 3) Multi-task prediction. We append multiple task-specific heads to spatiotemporal representations, enabling a comprehensive visual geometry estimation for dynamic scenes. Under this unified framework, these components enhance the feature discriminability and application universality of our model for dynamic scenes. In addition, we integrate multiple geometry datasets to train our model and conduct extensive experiments to verify the effectiveness of our method across various tasks on multiple dynamic scene geometry benchmarks.", "date": "2025-11-23", "categories": []}
{"id": "2511.16949", "title": "MobileOcc: A Human-Aware Semantic Occupancy Dataset for Mobile Robots", "abstract": "Dense 3D semantic occupancy perception is critical for mobile robots operating in pedestrian-rich environments, yet it remains underexplored compared to its application in autonomous driving. To address this gap, we present MobileOcc, a semantic occupancy dataset for mobile robots operating in crowded human environments. Our dataset is built using an annotation pipeline that incorporates static object occupancy annotations and a novel mesh optimization framework explicitly designed for human occupancy modeling. It reconstructs deformable human geometry from 2D images and subsequently refines and optimizes it using associated LiDAR point data. Using MobileOcc, we establish benchmarks for two tasks, i) Occupancy prediction and ii) Pedestrian velocity prediction, using different methods including monocular, stereo, and panoptic occupancy, with metrics and baseline implementations for reproducible comparison. Beyond occupancy prediction, we further assess our annotation method on 3D human pose estimation datasets. Results demonstrate that our method exhibits robust performance across different datasets.", "date": "2025-11-21", "categories": []}
{"id": "2507.06618", "title": "PointVDP: Learning View-Dependent Projection by Fireworks Rays for 3D Point Cloud Segmentation", "abstract": "In this paper, we propose view-dependent projection (VDP) to facilitate point cloud segmentation, designing efficient 3D-to-2D mapping that dynamically adapts to the spatial geometry from view variations. Existing projection-based methods leverage view-independent projection in complex scenes, relying on straight lines to generate direct rays or upward curves to reduce occlusions. However, their view independence provides projection rays that are limited to pre-defined parameters by human settings, restricting point awareness and failing to capture sufficient projection diversity across different view planes. Although multiple projections per view plane are commonly used to enhance spatial variety, the projected redundancy leads to excessive computational overhead and inefficiency in image processing. To address these limitations, we design a framework of VDP to generate data-driven projections from 3D point distributions, producing highly informative single-image inputs by predicting rays inspired by the adaptive behavior of fireworks. In addition, we construct color regularization to optimize the framework, which emphasizes essential features within semantic pixels and suppresses the non-semantic features within black pixels, thereby maximizing 2D space utilization in a projected image. As a result, our approach, PointVDP, develops lightweight projections in marginal computation costs. Experiments on S3DIS and ScanNet benchmarks show that our approach achieves competitive results, offering a resource-efficient solution for semantic understanding.", "date": "2025-07-09", "categories": []}
{"id": "2511.15288", "title": "Edge-Centric Relational Reasoning for 3D Scene Graph Prediction", "abstract": "3D scene graph prediction aims to abstract complex 3D environments into structured graphs consisting of objects and their pairwise relationships. Existing approaches typically adopt object-centric graph neural networks, where relation edge features are iteratively updated by aggregating messages from connected object nodes. However, this design inherently restricts relation representations to pairwise object context, making it difficult to capture high-order relational dependencies that are essential for accurate relation prediction. To address this limitation, we propose a Link-guided Edge-centric relational reasoning framework with Object-aware fusion, namely LEO, which enables progressive reasoning from relation-level context to object-level understanding. Specifically, LEO first predicts potential links between object pairs to suppress irrelevant edges, and then transforms the original scene graph into a line graph where each relation is treated as a node. A line graph neural network is applied to perform edge-centric relational reasoning to capture inter-relation context. The enriched relation features are subsequently integrated into the original object-centric graph to enhance object-level reasoning and improve relation prediction. Our framework is model-agnostic and can be integrated with any existing object-centric method. Experiments on the 3DSSG dataset with two competitive baselines show consistent improvements, highlighting the effectiveness of our edge-to-object reasoning paradigm.", "date": "2025-11-19", "categories": []}
{"id": "2511.15396", "title": "ShelfOcc: Native 3D Supervision beyond LiDAR for Vision-Based Occupancy Estimation", "abstract": "Recent progress in self- and weakly supervised occupancy estimation has largely relied on 2D projection or rendering-based supervision, which suffers from geometric inconsistencies and severe depth bleeding. We thus introduce ShelfOcc, a vision-only method that overcomes these limitations without relying on LiDAR. ShelfOcc brings supervision into native 3D space by generating metrically consistent semantic voxel labels from video, enabling true 3D supervision without any additional sensors or manual 3D annotations. While recent vision-based 3D geometry foundation models provide a promising source of prior knowledge, they do not work out of the box as a prediction due to sparse or noisy and inconsistent geometry, especially in dynamic driving scenes. Our method introduces a dedicated framework that mitigates these issues by filtering and accumulating static geometry consistently across frames, handling dynamic content and propagating semantic information into a stable voxel representation. This data-centric shift in supervision for weakly/shelf-supervised occupancy estimation allows the use of essentially any SOTA occupancy model architecture without relying on LiDAR data. We argue that such high-quality supervision is essential for robust occupancy learning and constitutes an important complementary avenue to architectural innovation. On the Occ3D-nuScenes benchmark, ShelfOcc substantially outperforms all previous weakly/shelf-supervised methods (up to a 34% relative improvement), establishing a new data-driven direction for LiDAR-free 3D scene understanding.", "date": "2025-11-19", "categories": []}
{"id": "2511.12977", "title": "ArtiWorld: LLM-Driven Articulation of 3D Objects in Scenes", "abstract": "Building interactive simulators and scalable robot-learning environments requires a large number of articulated assets. However, most existing 3D assets in simulation are rigid, and manually converting them into articulated objects is extremely labor- and cost-intensive. This raises a natural question: can we automatically identify articulable objects in a scene and convert them into articulated assets directly? In this paper, we present ArtiWorld, a scene-aware pipeline that localizes candidate articulable objects from textual scene descriptions and reconstructs executable URDF models that preserve the original geometry. At the core of this pipeline is Arti4URDF, which leverages 3D point cloud, prior knowledge of a large language model (LLM), and a URDF-oriented prompt design to rapidly convert rigid objects into interactive URDF-based articulated objects while maintaining their 3D shape. We evaluate ArtiWorld at three levels: 3D simulated objects, full 3D simulated scenes, and real-world scan scenes. Across all three settings, our method consistently outperforms existing approaches and achieves state-of-the-art performance, while preserving object geometry and correctly capturing object interactivity to produce usable URDF-based articulated models. This provides a practical path toward building interactive, robot-ready simulation environments directly from existing 3D assets. Code and data will be released.", "date": "2025-11-17", "categories": []}
{"id": "2511.11702", "title": "Task-Aware 3D Affordance Segmentation via 2D Guidance and Geometric Refinement", "abstract": "Understanding 3D scene-level affordances from natural language instructions is essential for enabling embodied agents to interact meaningfully in complex environments. However, this task remains challenging due to the need for semantic reasoning and spatial grounding. Existing methods mainly focus on object-level affordances or merely lift 2D predictions to 3D, neglecting rich geometric structure information in point clouds and incurring high computational costs. To address these limitations, we introduce Task-Aware 3D Scene-level Affordance segmentation (TASA), a novel geometry-optimized framework that jointly leverages 2D semantic cues and 3D geometric reasoning in a coarse-to-fine manner. To improve the affordance detection efficiency, TASA features a task-aware 2D affordance detection module to identify manipulable points from language and visual inputs, guiding the selection of task-relevant views. To fully exploit 3D geometric information, a 3D affordance refinement module is proposed to integrate 2D semantic priors with local 3D geometry, resulting in accurate and spatially coherent 3D affordance masks. Experiments on SceneFun3D demonstrate that TASA significantly outperforms the baselines in both accuracy and efficiency in scene-level affordance segmentation.", "date": "2025-11-12", "categories": []}
{"id": "2511.13047", "title": "DiffPixelFormer: Differential Pixel-Aware Transformer for RGB-D Indoor Scene Segmentation", "abstract": "Indoor semantic segmentation is fundamental to computer vision and robotics, supporting applications such as autonomous navigation, augmented reality, and smart environments. Although RGB-D fusion leverages complementary appearance and geometric cues, existing methods often depend on computationally intensive cross-attention mechanisms and insufficiently model intra- and inter-modal feature relationships, resulting in imprecise feature alignment and limited discriminative representation. To address these challenges, we propose DiffPixelFormer, a differential pixel-aware Transformer for RGB-D indoor scene segmentation that simultaneously enhances intra-modal representations and models inter-modal interactions. At its core, the Intra-Inter Modal Interaction Block (IIMIB) captures intra-modal long-range dependencies via self-attention and models inter-modal interactions with the Differential-Shared Inter-Modal (DSIM) module to disentangle modality-specific and shared cues, enabling fine-grained, pixel-level cross-modal alignment. Furthermore, a dynamic fusion strategy balances modality contributions and fully exploits RGB-D information according to scene characteristics. Extensive experiments on the SUN RGB-D and NYUDv2 benchmarks demonstrate that DiffPixelFormer-L achieves mIoU scores of 54.28% and 59.95%, outperforming DFormer-L by 1.78% and 2.75%, respectively. Code is available at https://github.com/gongyan1/DiffPixelFormer.", "date": "2025-11-17", "categories": []}
{"id": "2504.15888", "title": "MS-Occ: Multi-Stage LiDAR-Camera Fusion for 3D Semantic Occupancy Prediction", "abstract": "Accurate 3D semantic occupancy perception is essential for autonomous driving in complex environments with diverse and irregular objects. While vision-centric methods suffer from geometric inaccuracies, LiDAR-based approaches often lack rich semantic information. To address these limitations, MS-Occ, a novel multi-stage LiDAR-camera fusion framework which includes middle-stage fusion and late-stage fusion, is proposed, integrating LiDAR's geometric fidelity with camera-based semantic richness via hierarchical cross-modal fusion. The framework introduces innovations at two critical stages: (1) In the middle-stage feature fusion, the Gaussian-Geo module leverages Gaussian kernel rendering on sparse LiDAR depth maps to enhance 2D image features with dense geometric priors, and the Semantic-Aware module enriches LiDAR voxels with semantic context via deformable cross-attention; (2) In the late-stage voxel fusion, the Adaptive Fusion (AF) module dynamically balances voxel features across modalities, while the High Classification Confidence Voxel Fusion (HCCVF) module resolves semantic inconsistencies using self-attention-based refinement. Experiments on two large-scale benchmarks demonstrate state-of-the-art performance. On nuScenes-OpenOccupancy, MS-Occ achieves an Intersection over Union (IoU) of 32.1% and a mean IoU (mIoU) of 25.3%, surpassing the state-of-the-art by +0.7% IoU and +2.4% mIoU. Furthermore, on the SemanticKITTI benchmark, our method achieves a new state-of-the-art mIoU of 24.08%, robustly validating its generalization capabilities.Ablation studies further confirm the effectiveness of each individual module, highlighting substantial improvements in the perception of small objects and reinforcing the practical value of MS-Occ for safety-critical autonomous driving scenarios.", "date": "2025-04-22", "categories": []}
{"id": "2509.26091", "title": "Text-to-Scene with Large Reasoning Models", "abstract": "Prompt-driven scene synthesis allows users to generate complete 3D environments from textual descriptions. Current text-to-scene methods often struggle with complex geometries and object transformations, and tend to show weak adherence to complex instructions. We address these limitations by introducing Reason-3D, a text-to-scene model powered by large reasoning models (LRMs). Reason-3D integrates object retrieval using captions covering physical, functional, and contextual attributes. Reason-3D then places the selected objects based on implicit and explicit layout constraints, and refines their positions with collision-aware spatial reasoning. Evaluated on instructions ranging from simple to complex indoor configurations, Reason-3D significantly outperforms previous methods in human-rated visual fidelity, adherence to constraints, and asset retrieval quality. Beyond its contribution to the field of text-to-scene generation, our work showcases the advanced spatial reasoning abilities of modern LRMs. Additionally, we release the codebase to further the research in object retrieval and placement with LRMs.", "date": "2025-09-30", "categories": []}
{"id": "2511.07925", "title": "HD$^2$-SSC: High-Dimension High-Density Semantic Scene Completion for Autonomous Driving", "abstract": "Camera-based 3D semantic scene completion (SSC) plays a crucial role in autonomous driving, enabling voxelized 3D scene understanding for effective scene perception and decision-making. Existing SSC methods have shown efficacy in improving 3D scene representations, but suffer from the inherent input-output dimension gap and annotation-reality density gap, where the 2D planner view from input images with sparse annotated labels leads to inferior prediction of real-world dense occupancy with a 3D stereoscopic view. In light of this, we propose the corresponding High-Dimension High-Density Semantic Scene Completion (HD$^2$-SSC) framework with expanded pixel semantics and refined voxel occupancies. To bridge the dimension gap, a High-dimension Semantic Decoupling module is designed to expand 2D image features along a pseudo third dimension, decoupling coarse pixel semantics from occlusions, and then identify focal regions with fine semantics to enrich image features. To mitigate the density gap, a High-density Occupancy Refinement module is devised with a \"detect-and-refine\" architecture to leverage contextual geometric and semantic structures for enhanced semantic density with the completion of missing voxels and correction of erroneous ones. Extensive experiments and analyses on the SemanticKITTI and SSCBench-KITTI-360 datasets validate the effectiveness of our HD$^2$-SSC framework.", "date": "2025-11-11", "categories": []}
{"id": "2511.09883", "title": "HCC-3D: Hierarchical Compensatory Compression for 98% 3D Token Reduction in Vision-Language Models", "abstract": "3D understanding has drawn significant attention recently, leveraging Vision-Language Models (VLMs) to enable multi-modal reasoning between point cloud and text data. Current 3D-VLMs directly embed the 3D point clouds into 3D tokens, following large 2D-VLMs with powerful reasoning capabilities. However, this framework has a great computational cost limiting its application, where we identify that the bottleneck lies in processing all 3D tokens in the Large Language Model (LLM) part. This raises the question: how can we reduce the computational overhead introduced by 3D tokens while preserving the integrity of their essential information? To address this question, we introduce Hierarchical Compensatory Compression (HCC-3D) to efficiently compress 3D tokens while maintaining critical detail retention. Specifically, we first propose a global structure compression (GSC), in which we design global queries to compress all 3D tokens into a few key tokens while keeping overall structural information. Then, to compensate for the information loss in GSC, we further propose an adaptive detail mining (ADM) module that selectively recompresses salient but under-attended features through complementary scoring. Extensive experiments demonstrate that HCC-3D not only achieves extreme compression ratios (approximately 98%) compared to previous 3D-VLMs, but also achieves new state-of-the-art performance, showing the great improvements on both efficiency and performance.", "date": "2025-11-13", "categories": []}
{"id": "2411.19461", "title": "Robust Bayesian Scene Reconstruction with Retrieval-Augmented Priors for Precise Grasping and Planning", "abstract": "Constructing 3D representations of object geometry is critical for many robotics tasks, particularly manipulation problems. These representations must be built from potentially noisy partial observations. In this work, we focus on the problem of reconstructing a multi-object scene from a single RGBD image using a fixed camera. Traditional scene representation methods generally cannot infer the geometry of unobserved regions of the objects in the image. Attempts have been made to leverage deep learning to train on a dataset of known objects and representations, and then generalize to new observations. However, this can be brittle to noisy real-world observations and objects not contained in the dataset, and do not provide well-calibrated reconstruction confidences. We propose BRRP, a reconstruction method that leverages preexisting mesh datasets to build an informative prior during robust probabilistic reconstruction. We introduce the concept of a retrieval-augmented prior, where we retrieve relevant components of our prior distribution from a database of objects during inference. The resulting prior enables estimation of the geometry of occluded portions of the in-scene objects. Our method produces a distribution over object shape that can be used for reconstruction and measuring uncertainty. We evaluate our method in both simulated scenes and in the real world. We demonstrate the robustness of our method against deep learning-only approaches while being more accurate than a method without an informative prior. Through real-world experiments, we particularly highlight the capability of BRRP to enable successful dexterous manipulation in clutter.", "date": "2024-11-29", "categories": []}
{"id": "2509.11171", "title": "SPHERE: Semantic-PHysical Engaged REpresentation for 3D Semantic Scene Completion", "abstract": "Camera-based 3D Semantic Scene Completion (SSC) is a critical task in autonomous driving systems, assessing voxel-level geometry and semantics for holistic scene perception. While existing voxel-based and plane-based SSC methods have achieved considerable progress, they struggle to capture physical regularities for realistic geometric details. On the other hand, neural reconstruction methods like NeRF and 3DGS demonstrate superior physical awareness, but suffer from high computational cost and slow convergence when handling large-scale, complex autonomous driving scenes, leading to inferior semantic accuracy. To address these issues, we propose the Semantic-PHysical Engaged REpresentation (SPHERE) for camera-based SSC, which integrates voxel and Gaussian representations for joint exploitation of semantic and physical information. First, the Semantic-guided Gaussian Initialization (SGI) module leverages dual-branch 3D scene representations to locate focal voxels as anchors to guide efficient Gaussian initialization. Then, the Physical-aware Harmonics Enhancement (PHE) module incorporates semantic spherical harmonics to model physical-aware contextual details and promote semantic-geometry consistency through focal distribution alignment, generating SSC results with realistic details. Extensive experiments and analyses on the popular SemanticKITTI and SSCBench-KITTI-360 benchmarks validate the effectiveness of SPHERE. The code is available at https://github.com/PKU-ICST-MIPL/SPHERE_ACMMM2025.", "date": "2025-09-14", "categories": []}
{"id": "2511.07813", "title": "Sparse3DPR: Training-Free 3D Hierarchical Scene Parsing and Task-Adaptive Subgraph Reasoning from Sparse RGB Views", "abstract": "Recently, large language models (LLMs) have been explored widely for 3D scene understanding. Among them, training-free approaches are gaining attention for their flexibility and generalization over training-based methods. However, they typically struggle with accuracy and efficiency in practical deployment. To address the problems, we propose Sparse3DPR, a novel training-free framework for open-ended scene understanding, which leverages the reasoning capabilities of pre-trained LLMs and requires only sparse-view RGB inputs. Specifically, we introduce a hierarchical plane-enhanced scene graph that supports open vocabulary and adopts dominant planar structures as spatial anchors, which enables clearer reasoning chains and more reliable high-level inferences. Furthermore, we design a task-adaptive subgraph extraction method to filter query-irrelevant information dynamically, reducing contextual noise and improving 3D scene reasoning efficiency and accuracy. Experimental results demonstrate the superiority of Sparse3DPR, which achieves a 28.7% EM@1 improvement and a 78.2% speedup compared with ConceptGraphs on the Space3D-Bench. Moreover, Sparse3DPR obtains comparable performance to training-based methods on ScanQA, with additional real-world experiments confirming its robustness and generalization capability.", "date": "2025-11-11", "categories": []}
{"id": "2502.14520", "title": "Learning Temporal 3D Semantic Scene Completion via Optical Flow Guidance", "abstract": "3D Semantic Scene Completion (SSC) provides comprehensive scene geometry and semantics for autonomous driving perception, which is crucial for enabling accurate and reliable decision-making. However, existing SSC methods are limited to capturing sparse information from the current frame or naively stacking multi-frame temporal features, thereby failing to acquire effective scene context. These approaches ignore critical motion dynamics and struggle to achieve temporal consistency. To address the above challenges, we propose a novel temporal SSC method FlowScene: Learning Temporal 3D Semantic Scene Completion via Optical Flow Guidance. By leveraging optical flow, FlowScene can integrate motion, different viewpoints, occlusions, and other contextual cues, thereby significantly improving the accuracy of 3D scene completion. Specifically, our framework introduces two key components: (1) a Flow-Guided Temporal Aggregation module that aligns and aggregates temporal features using optical flow, capturing motion-aware context and deformable structures; and (2) an Occlusion-Guided Voxel Refinement module that injects occlusion masks and temporally aggregated features into 3D voxel space, adaptively refining voxel representations for explicit geometric modeling. Experimental results demonstrate that FlowScene achieves state-of-the-art performance on the SemanticKITTI and SSCBench-KITTI-360 benchmarks.", "date": "2025-02-20", "categories": []}
{"id": "2503.18135", "title": "MLLM-For3D: Adapting Multimodal Large Language Model for 3D Reasoning Segmentation", "abstract": "Reasoning segmentation aims to segment target objects in complex scenes based on human intent and spatial reasoning. While recent multimodal large language models (MLLMs) have demonstrated impressive 2D image reasoning segmentation, adapting these capabilities to 3D scenes remains underexplored. In this paper, we introduce MLLM-For3D, a simple yet effective framework that transfers knowledge from 2D MLLMs to 3D scene understanding. Specifically, we utilize MLLMs to generate multi-view pseudo segmentation masks and corresponding text embeddings, then unproject 2D masks into 3D space and align them with the text embeddings. The primary challenge lies in the absence of 3D context and spatial consistency across multiple views, causing the model to hallucinate objects that do not exist and fail to target objects consistently. Training the 3D model with such irrelevant objects leads to performance degradation. To address this, we introduce a spatial consistency strategy to enforce that segmentation masks remain coherent in the 3D space, effectively capturing the geometry of the scene. Moreover, we develop a Token-for-Query approach for multimodal semantic alignment, enabling consistent identification of the same object across different views. Extensive evaluations on various challenging indoor scene benchmarks demonstrate that, even without any labeled 3D training data, MLLM-For3D outperforms existing 3D reasoning segmentation methods, effectively interpreting user intent, understanding 3D scenes, and reasoning about spatial relationships.", "date": "2025-03-23", "categories": []}
{"id": "2511.05894", "title": "Open-World 3D Scene Graph Generation for Retrieval-Augmented Reasoning", "abstract": "Understanding 3D scenes in open-world settings poses fundamental challenges for vision and robotics, particularly due to the limitations of closed-vocabulary supervision and static annotations. To address this, we propose a unified framework for Open-World 3D Scene Graph Generation with Retrieval-Augmented Reasoning, which enables generalizable and interactive 3D scene understanding. Our method integrates Vision-Language Models (VLMs) with retrieval-based reasoning to support multimodal exploration and language-guided interaction. The framework comprises two key components: (1) a dynamic scene graph generation module that detects objects and infers semantic relationships without fixed label sets, and (2) a retrieval-augmented reasoning pipeline that encodes scene graphs into a vector database to support text/image-conditioned queries. We evaluate our method on 3DSSG and Replica benchmarks across four tasks-scene question answering, visual grounding, instance retrieval, and task planning-demonstrating robust generalization and superior performance in diverse environments. Our results highlight the effectiveness of combining open-vocabulary perception with retrieval-based reasoning for scalable 3D scene understanding.", "date": "2025-11-08", "categories": []}
{"id": "2511.06632", "title": "DIAL-GS: Dynamic Instance Aware Reconstruction for Label-free Street Scenes with 4D Gaussian Splatting", "abstract": "Urban scene reconstruction is critical for autonomous driving, enabling structured 3D representations for data synthesis and closed-loop testing. Supervised approaches rely on costly human annotations and lack scalability, while current self-supervised methods often confuse static and dynamic elements and fail to distinguish individual dynamic objects, limiting fine-grained editing. We propose DIAL-GS, a novel dynamic instance-aware reconstruction method for label-free street scenes with 4D Gaussian Splatting. We first accurately identify dynamic instances by exploiting appearance-position inconsistency between warped rendering and actual observation. Guided by instance-level dynamic perception, we employ instance-aware 4D Gaussians as the unified volumetric representation, realizing dynamic-adaptive and instance-aware reconstruction. Furthermore, we introduce a reciprocal mechanism through which identity and dynamics reinforce each other, enhancing both integrity and consistency. Experiments on urban driving scenarios show that DIAL-GS surpasses existing self-supervised baselines in reconstruction quality and instance-level editing, offering a concise yet powerful solution for urban scene modeling.", "date": "2025-11-10", "categories": []}
{"id": "2511.06744", "title": "PointCubeNet: 3D Part-level Reasoning with 3x3x3 Point Cloud Blocks", "abstract": "In this paper, we propose PointCubeNet, a novel multi-modal 3D understanding framework that achieves part-level reasoning without requiring any part annotations. PointCubeNet comprises global and local branches. The proposed local branch, structured into 3x3x3 local blocks, enables part-level analysis of point cloud sub-regions with the corresponding local text labels. Leveraging the proposed pseudo-labeling method and local loss function, PointCubeNet is effectively trained in an unsupervised manner. The experimental results demonstrate that understanding 3D object parts enhances the understanding of the overall 3D object. In addition, this is the first attempt to perform unsupervised 3D part-level reasoning and achieves reliable and meaningful results.", "date": "2025-11-10", "categories": []}
{"id": "2511.06908", "title": "Mono3DVG-EnSD: Enhanced Spatial-aware and Dimension-decoupled Text Encoding for Monocular 3D Visual Grounding", "abstract": "Monocular 3D Visual Grounding (Mono3DVG) is an emerging task that locates 3D objects in RGB images using text descriptions with geometric cues. However, existing methods face two key limitations. Firstly, they often over-rely on high-certainty keywords that explicitly identify the target object while neglecting critical spatial descriptions. Secondly, generalized textual features contain both 2D and 3D descriptive information, thereby capturing an additional dimension of details compared to singular 2D or 3D visual features. This characteristic leads to cross-dimensional interference when refining visual features under text guidance. To overcome these challenges, we propose Mono3DVG-EnSD, a novel framework that integrates two key components: the CLIP-Guided Lexical Certainty Adapter (CLIP-LCA) and the Dimension-Decoupled Module (D2M). The CLIP-LCA dynamically masks high-certainty keywords while retaining low-certainty implicit spatial descriptions, thereby forcing the model to develop a deeper understanding of spatial relationships in captions for object localization. Meanwhile, the D2M decouples dimension-specific (2D/3D) textual features from generalized textual features to guide corresponding visual features at same dimension, which mitigates cross-dimensional interference by ensuring dimensionally-consistent cross-modal interactions. Through comprehensive comparisons and ablation studies on the Mono3DRefer dataset, our method achieves state-of-the-art (SOTA) performance across all metrics. Notably, it improves the challenging Far(Acc@0.5) scenario by a significant +13.54%.", "date": "2025-11-10", "categories": []}
{"id": "2511.07007", "title": "TrueCity: Real and Simulated Urban Data for Cross-Domain 3D Scene Understanding", "abstract": "3D semantic scene understanding remains a long-standing challenge in the 3D computer vision community. One of the key issues pertains to limited real-world annotated data to facilitate generalizable models. The common practice to tackle this issue is to simulate new data. Although synthetic datasets offer scalability and perfect labels, their designer-crafted scenes fail to capture real-world complexity and sensor noise, resulting in a synthetic-to-real domain gap. Moreover, no benchmark provides synchronized real and simulated point clouds for segmentation-oriented domain shift analysis. We introduce TrueCity, the first urban semantic segmentation benchmark with cm-accurate annotated real-world point clouds, semantic 3D city models, and annotated simulated point clouds representing the same city. TrueCity proposes segmentation classes aligned with international 3D city modeling standards, enabling consistent evaluation of synthetic-to-real gap. Our extensive experiments on common baselines quantify domain shift and highlight strategies for exploiting synthetic data to enhance real-world 3D scene understanding. We are convinced that the TrueCity dataset will foster further development of sim-to-real gap quantification and enable generalizable data-driven models. The data, code, and 3D models are available online: https://tum-gis.github.io/TrueCity/", "date": "2025-11-10", "categories": []}
{"id": "2511.03970", "title": "Room Envelopes: A Synthetic Dataset for Indoor Layout Reconstruction from Images", "abstract": "Modern scene reconstruction methods are able to accurately recover 3D surfaces that are visible in one or more images. However, this leads to incomplete reconstructions, missing all occluded surfaces. While much progress has been made on reconstructing entire objects given partial observations using generative models, the structural elements of a scene, like the walls, floors and ceilings, have received less attention. We argue that these scene elements should be relatively easy to predict, since they are typically planar, repetitive and simple, and so less costly approaches may be suitable. In this work, we present a synthetic dataset -- Room Envelopes -- that facilitates progress on this task by providing a set of RGB images and two associated pointmaps for each image: one capturing the visible surface and one capturing the first surface once fittings and fixtures are removed, that is, the structural layout. As we show, this enables direct supervision for feed-forward monocular geometry estimators that predict both the first visible surface and the first layout surface. This confers an understanding of the scene's extent, as well as the shape and location of its objects.", "date": "2025-11-06", "categories": []}
{"id": "2506.07491", "title": "SpatialLM: Training Large Language Models for Structured Indoor Modeling", "abstract": "SpatialLM is a large language model designed to process 3D point cloud data and generate structured 3D scene understanding outputs. These outputs include architectural elements like walls, doors, windows, and oriented object boxes with their semantic categories. Unlike previous methods which exploit task-specific network designs, our model adheres to the standard multimodal LLM architecture and is fine-tuned directly from open-source LLMs.\n  To train SpatialLM, we collect a large-scale, high-quality synthetic dataset consisting of the point clouds of 12,328 indoor scenes (54,778 rooms) with ground-truth 3D annotations, and conduct a careful study on various modeling and training decisions. On public benchmarks, our model gives state-of-the-art performance in layout estimation and competitive results in 3D object detection. With that, we show a feasible path for enhancing the spatial understanding capabilities of modern LLMs for applications in augmented reality, embodied robotics, and more.", "date": "2025-06-09", "categories": []}
{"id": "2411.10509", "title": "TESGNN: Temporal Equivariant Scene Graph Neural Networks for Efficient and Robust Multi-View 3D Scene Understanding", "abstract": "Scene graphs have proven to be highly effective for various scene understanding tasks due to their compact and explicit representation of relational information. However, current methods often overlook the critical importance of preserving symmetry when generating scene graphs from 3D point clouds, which can lead to reduced accuracy and robustness, particularly when dealing with noisy, multi-view data. Furthermore, a major limitation of prior approaches is the lack of temporal modeling to capture time-dependent relationships among dynamically evolving entities in a scene. To address these challenges, we propose Temporal Equivariant Scene Graph Neural Network (TESGNN), consisting of two key components: (1) an Equivariant Scene Graph Neural Network (ESGNN), which extracts information from 3D point clouds to generate scene graph while preserving crucial symmetry properties, and (2) a Temporal Graph Matching Network, which fuses scene graphs generated by ESGNN across multiple time sequences into a unified global representation using an approximate graph-matching algorithm. Our combined architecture TESGNN shown to be effective compared to existing methods in scene graph generation, achieving higher accuracy and faster training convergence. Moreover, we show that leveraging the symmetry-preserving property produces a more stable and accurate global scene representation compared to existing approaches. Finally, it is computationally efficient and easily implementable using existing frameworks, making it well-suited for real-time applications in robotics and computer vision. This approach paves the way for more robust and scalable solutions to complex multi-view scene understanding challenges. Our source code is publicly available at: https://github.com/HySonLab/TESGraph", "date": "2024-11-15", "categories": []}
{"id": "2510.22706", "title": "IGGT: Instance-Grounded Geometry Transformer for Semantic 3D Reconstruction", "abstract": "Humans naturally perceive the geometric structure and semantic content of a 3D world as intertwined dimensions, enabling coherent and accurate understanding of complex scenes. However, most prior approaches prioritize training large geometry models for low-level 3D reconstruction and treat high-level spatial understanding in isolation, overlooking the crucial interplay between these two fundamental aspects of 3D-scene analysis, thereby limiting generalization and leading to poor performance in downstream 3D understanding tasks. Recent attempts have mitigated this issue by simply aligning 3D models with specific language models, thus restricting perception to the aligned model's capacity and limiting adaptability to downstream tasks. In this paper, we propose InstanceGrounded Geometry Transformer (IGGT), an end-to-end large unified transformer to unify the knowledge for both spatial reconstruction and instance-level contextual understanding. Specifically, we design a 3D-Consistent Contrastive Learning strategy that guides IGGT to encode a unified representation with geometric structures and instance-grounded clustering through only 2D visual inputs. This representation supports consistent lifting of 2D visual inputs into a coherent 3D scene with explicitly distinct object instances. To facilitate this task, we further construct InsScene-15K, a large-scale dataset with high-quality RGB images, poses, depth maps, and 3D-consistent instance-level mask annotations with a novel data curation pipeline.", "date": "2025-10-26", "categories": []}
{"id": "2508.12015", "title": "InstDrive: Instance-Aware 3D Gaussian Splatting for Driving Scenes", "abstract": "Reconstructing dynamic driving scenes from dashcam videos has attracted increasing attention due to its significance in autonomous driving and scene understanding. While recent advances have made impressive progress, most methods still unify all background elements into a single representation, hindering both instance-level understanding and flexible scene editing. Some approaches attempt to lift 2D segmentation into 3D space, but often rely on pre-processed instance IDs or complex pipelines to map continuous features to discrete identities. Moreover, these methods are typically designed for indoor scenes with rich viewpoints, making them less applicable to outdoor driving scenarios. In this paper, we present InstDrive, an instance-aware 3D Gaussian Splatting framework tailored for the interactive reconstruction of dynamic driving scene. We use masks generated by SAM as pseudo ground-truth to guide 2D feature learning via contrastive loss and pseudo-supervised objectives. At the 3D level, we introduce regularization to implicitly encode instance identities and enforce consistency through a voxel-based loss. A lightweight static codebook further bridges continuous features and discrete identities without requiring data pre-processing or complex optimization. Quantitative and qualitative experiments demonstrate the effectiveness of InstDrive, and to the best of our knowledge, it is the first framework to achieve 3D instance segmentation in dynamic, open-world driving scenes.More visualizations are available at our project page.", "date": "2025-08-16", "categories": []}
{"id": "2510.25146", "title": "EA3D: Online Open-World 3D Object Extraction from Streaming Videos", "abstract": "Current 3D scene understanding methods are limited by offline-collected multi-view data or pre-constructed 3D geometry. In this paper, we present ExtractAnything3D (EA3D), a unified online framework for open-world 3D object extraction that enables simultaneous geometric reconstruction and holistic scene understanding. Given a streaming video, EA3D dynamically interprets each frame using vision-language and 2D vision foundation encoders to extract object-level knowledge. This knowledge is integrated and embedded into a Gaussian feature map via a feed-forward online update strategy. We then iteratively estimate visual odometry from historical frames and incrementally update online Gaussian features with new observations. A recurrent joint optimization module directs the model's attention to regions of interest, simultaneously enhancing both geometric reconstruction and semantic understanding. Extensive experiments across diverse benchmarks and tasks, including photo-realistic rendering, semantic and instance segmentation, 3D bounding box and semantic occupancy estimation, and 3D mesh generation, demonstrate the effectiveness of EA3D. Our method establishes a unified and efficient framework for joint online 3D reconstruction and holistic scene understanding, enabling a broad range of downstream tasks.", "date": "2025-10-29", "categories": []}
{"id": "2509.20077", "title": "Queryable 3D Scene Representation: A Multi-Modal Framework for Semantic Reasoning and Robotic Task Planning", "abstract": "To enable robots to comprehend high-level human instructions and perform complex tasks, a key challenge lies in achieving comprehensive scene understanding: interpreting and interacting with the 3D environment in a meaningful way. This requires a smart map that fuses accurate geometric structure with rich, human-understandable semantics. To address this, we introduce the 3D Queryable Scene Representation (3D QSR), a novel framework built on multimedia data that unifies three complementary 3D representations: (1) 3D-consistent novel view rendering and segmentation from panoptic reconstruction, (2) precise geometry from 3D point clouds, and (3) structured, scalable organization via 3D scene graphs. Built on an object-centric design, the framework integrates with large vision-language models to enable semantic queryability by linking multimodal object embeddings, and supporting object-level retrieval of geometric, visual, and semantic information. The retrieved data are then loaded into a robotic task planner for downstream execution. We evaluate our approach through simulated robotic task planning scenarios in Unity, guided by abstract language instructions and using the indoor public dataset Replica. Furthermore, we apply it in a digital duplicate of a real wet lab environment to test QSR-supported robotic task planning for emergency response. The results demonstrate the framework's ability to facilitate scene understanding and integrate spatial and semantic reasoning, effectively translating high-level human instructions into precise robotic task planning in complex 3D environments.", "date": "2025-09-24", "categories": []}
{"id": "2509.23107", "title": "Open-Vocabulary Spatio-Temporal Scene Graph for Robot Perception and Teleoperation Planning", "abstract": "Teleoperation via natural-language reduces operator workload and enhances safety in high-risk or remote settings. However, in dynamic remote scenes, transmission latency during bidirectional communication creates gaps between remote perceived states and operator intent, leading to command misunderstanding and incorrect execution. To mitigate this, we introduce the Spatio-Temporal Open-Vocabulary Scene Graph (ST-OVSG), a representation that enriches open-vocabulary perception with temporal dynamics and lightweight latency annotations. ST-OVSG leverages LVLMs to construct open-vocabulary 3D object representations, and extends them into the temporal domain via Hungarian assignment with our temporal matching cost, yielding a unified spatio-temporal scene graph. A latency tag is embedded to enable LVLM planners to retrospectively query past scene states, thereby resolving local-remote state mismatches caused by transmission delays. To further reduce redundancy and highlight task-relevant cues, we propose a task-oriented subgraph filtering strategy that produces compact inputs for the planner. ST-OVSG generalizes to novel categories and enhances planning robustness against transmission latency without requiring fine-tuning. Experiments show that our method achieves 74 percent node accuracy on the Replica benchmark, outperforming ConceptGraph. Notably, in the latency-robustness experiment, the LVLM planner assisted by ST-OVSG achieved a planning success rate of 70.5 percent.", "date": "2025-09-27", "categories": []}
{"id": "2510.23184", "title": "Finding 3D Scene Analogies with Multimodal Foundation Models", "abstract": "Connecting current observations with prior experiences helps robots adapt and plan in new, unseen 3D environments. Recently, 3D scene analogies have been proposed to connect two 3D scenes, which are smooth maps that align scene regions with common spatial relationships. These maps enable detailed transfer of trajectories or waypoints, potentially supporting demonstration transfer for imitation learning or task plan transfer across scenes. However, existing methods for the task require additional training and fixed object vocabularies. In this work, we propose to use multimodal foundation models for finding 3D scene analogies in a zero-shot, open-vocabulary setting. Central to our approach is a hybrid neural representation of scenes that consists of a sparse graph based on vision-language model features and a feature field derived from 3D shape foundation models. 3D scene analogies are then found in a coarse-to-fine manner, by first aligning the graph and refining the correspondence with feature fields. Our method can establish accurate correspondences between complex scenes, and we showcase applications in trajectory and waypoint transfer.", "date": "2025-10-27", "categories": []}
{"id": "2510.21069", "title": "ZING-3D: Zero-shot Incremental 3D Scene Graphs via Vision-Language Models", "abstract": "Understanding and reasoning about complex 3D environments requires structured scene representations that capture not only objects but also their semantic and spatial relationships. While recent works on 3D scene graph generation have leveraged pretrained VLMs without task-specific fine-tuning, they are largely confined to single-view settings, fail to support incremental updates as new observations arrive and lack explicit geometric grounding in 3D space, all of which are essential for embodied scenarios. In this paper, we propose, ZING-3D, a framework that leverages the vast knowledge of pretrained foundation models to enable open-vocabulary recognition and generate a rich semantic representation of the scene in a zero-shot manner while also enabling incremental updates and geometric grounding in 3D space, making it suitable for downstream robotics applications. Our approach leverages VLM reasoning to generate a rich 2D scene graph, which is grounded in 3D using depth information. Nodes represent open-vocabulary objects with features, 3D locations, and semantic context, while edges capture spatial and semantic relations with inter-object distances. Our experiments on scenes from the Replica and HM3D dataset show that ZING-3D is effective at capturing spatial and relational knowledge without the need of task-specific training.", "date": "2025-10-24", "categories": []}
{"id": "2510.21441", "title": "OpenHype: Hyperbolic Embeddings for Hierarchical Open-Vocabulary Radiance Fields", "abstract": "Modeling the inherent hierarchical structure of 3D objects and 3D scenes is highly desirable, as it enables a more holistic understanding of environments for autonomous agents. Accomplishing this with implicit representations, such as Neural Radiance Fields, remains an unexplored challenge. Existing methods that explicitly model hierarchical structures often face significant limitations: they either require multiple rendering passes to capture embeddings at different levels of granularity, significantly increasing inference time, or rely on predefined, closed-set discrete hierarchies that generalize poorly to the diverse and nuanced structures encountered by agents in the real world. To address these challenges, we propose OpenHype, a novel approach that represents scene hierarchies using a continuous hyperbolic latent space. By leveraging the properties of hyperbolic geometry, OpenHype naturally encodes multi-scale relationships and enables smooth traversal of hierarchies through geodesic paths in latent space. Our method outperforms state-of-the-art approaches on standard benchmarks, demonstrating superior efficiency and adaptability in 3D scene understanding.", "date": "2025-10-24", "categories": []}
{"id": "2308.05515", "title": "Mono-hydra: Real-time 3D scene graph construction from monocular camera input with IMU", "abstract": "The ability of robots to autonomously navigate through 3D environments depends on their comprehension of spatial concepts, ranging from low-level geometry to high-level semantics, such as objects, places, and buildings. To enable such comprehension, 3D scene graphs have emerged as a robust tool for representing the environment as a layered graph of concepts and their relationships. However, building these representations using monocular vision systems in real-time remains a difficult task that has not been explored in depth. This paper puts forth a real-time spatial perception system Mono-Hydra, combining a monocular camera and an IMU sensor setup, focusing on indoor scenarios. However, the proposed approach is adaptable to outdoor applications, offering flexibility in its potential uses. The system employs a suite of deep learning algorithms to derive depth and semantics. It uses a robocentric visual-inertial odometry (VIO) algorithm based on square-root information, thereby ensuring consistent visual odometry with an IMU and a monocular camera. This system achieves sub-20 cm error in real-time processing at 15 fps, enabling real-time 3D scene graph construction using a laptop GPU (NVIDIA 3080). This enhances decision-making efficiency and effectiveness in simple camera setups, augmenting robotic system agility. We make Mono-Hydra publicly available at: https://github.com/UAV-Centre-ITC/Mono_Hydra", "date": "2023-08-10", "categories": []}
{"id": "2510.16272", "title": "Proactive Scene Decomposition and Reconstruction", "abstract": "Human behaviors are the major causes of scene dynamics and inherently contain rich cues regarding the dynamics. This paper formalizes a new task of proactive scene decomposition and reconstruction, an online approach that leverages human-object interactions to iteratively disassemble and reconstruct the environment. By observing these intentional interactions, we can dynamically refine the decomposition and reconstruction process, addressing inherent ambiguities in static object-level reconstruction. The proposed system effectively integrates multiple tasks in dynamic environments such as accurate camera and object pose estimation, instance decomposition, and online map updating, capitalizing on cues from human-object interactions in egocentric live streams for a flexible, progressive alternative to conventional object-level reconstruction methods. Aided by the Gaussian splatting technique, accurate and consistent dynamic scene modeling is achieved with photorealistic and efficient rendering. The efficacy is validated in multiple real-world scenarios with promising advantages.", "date": "2025-10-17", "categories": []}
{"id": "2510.17664", "title": "4DSegStreamer: Streaming 4D Panoptic Segmentation via Dual Threads", "abstract": "4D panoptic segmentation in a streaming setting is critical for highly dynamic environments, such as evacuating dense crowds and autonomous driving in complex scenarios, where real-time, fine-grained perception within a constrained time budget is essential. In this paper, we introduce 4DSegStreamer, a novel framework that employs a Dual-Thread System to efficiently process streaming frames. The framework is general and can be seamlessly integrated into existing 3D and 4D segmentation methods to enable real-time capability. It also demonstrates superior robustness compared to existing streaming perception approaches, particularly under high FPS conditions. The system consists of a predictive thread and an inference thread. The predictive thread leverages historical motion and geometric information to extract features and forecast future dynamics. The inference thread ensures timely prediction for incoming frames by aligning with the latest memory and compensating for ego-motion and dynamic object movements. We evaluate 4DSegStreamer on the indoor HOI4D dataset and the outdoor SemanticKITTI and nuScenes datasets. Comprehensive experiments demonstrate the effectiveness of our approach, particularly in accurately predicting dynamic objects in complex scenes.", "date": "2025-10-20", "categories": []}
{"id": "2509.20401", "title": "SGAligner++: Cross-Modal Language-Aided 3D Scene Graph Alignment", "abstract": "Aligning 3D scene graphs is a crucial initial step for several applications in robot navigation and embodied perception. Current methods in 3D scene graph alignment often rely on single-modality point cloud data and struggle with incomplete or noisy input. We introduce SGAligner++, a cross-modal, language-aided framework for 3D scene graph alignment. Our method addresses the challenge of aligning partially overlapping scene observations across heterogeneous modalities by learning a unified joint embedding space, enabling accurate alignment even under low-overlap conditions and sensor noise. By employing lightweight unimodal encoders and attention-based fusion, SGAligner++ enhances scene understanding for tasks such as visual localization, 3D reconstruction, and navigation, while ensuring scalability and minimal computational overhead. Extensive evaluations on real-world datasets demonstrate that SGAligner++ outperforms state-of-the-art methods by up to 40% on noisy real-world reconstructions, while enabling cross-modal generalization.", "date": "2025-09-23", "categories": []}
{"id": "2510.13800", "title": "Reasoning in Space via Grounding in the World", "abstract": "In this paper, we claim that 3D visual grounding is the cornerstone of spatial reasoning and introduce the Grounded-Spatial Reasoner (GS-Reasoner) to explore the effective spatial representations that bridge the gap between them. Existing 3D LLMs suffer from the absence of a unified 3D representation capable of jointly capturing semantic and geometric information. This deficiency is manifested either in poor performance on grounding or in an excessive reliance on external modules, ultimately hindering the seamless integration of grounding and spatial reasoning. To address this, we propose a simple yet effective dual-path pooling mechanism that tightly aligns geometric features with both semantic and positional cues, constructing a unified image patch-based 3D representation that encapsulates all essential information without increasing the number of input tokens. Leveraging this holistic representation, GS-Reasoner is the first 3D LLM that achieves autoregressive grounding entirely without external modules while delivering performance comparable to state-of-the-art models, establishing a unified and self-contained framework for 3D spatial reasoning. To further bridge grounding and spatial reasoning, we introduce the Grounded Chain-of-Thought (GCoT) dataset. This dataset is meticulously curated to include both 3D bounding box annotations for objects referenced in reasoning questions and step-by-step reasoning paths that integrate grounding as a core component of the problem-solving process. Extensive experiments demonstrate that GS-Reasoner achieves impressive results on 3D visual grounding, which in turn significantly enhances its spatial reasoning capabilities, leading to state-of-the-art performance.", "date": "2025-10-15", "categories": []}
{"id": "2507.07136", "title": "LangSplatV2: High-dimensional 3D Language Gaussian Splatting with 450+ FPS", "abstract": "In this paper, we introduce LangSplatV2, which achieves high-dimensional feature splatting at 476.2 FPS and 3D open-vocabulary text querying at 384.6 FPS for high-resolution images, providing a 42 $\\times$ speedup and a 47 $\\times$ boost over LangSplat respectively, along with improved query accuracy. LangSplat employs Gaussian Splatting to embed 2D CLIP language features into 3D, significantly enhancing speed and learning a precise 3D language field with SAM semantics. Such advancements in 3D language fields are crucial for applications that require language interaction within complex scenes. However, LangSplat does not yet achieve real-time inference performance (8.2 FPS), even with advanced A100 GPUs, severely limiting its broader application. In this paper, we first conduct a detailed time analysis of LangSplat, identifying the heavyweight decoder as the primary speed bottleneck. Our solution, LangSplatV2 assumes that each Gaussian acts as a sparse code within a global dictionary, leading to the learning of a 3D sparse coefficient field that entirely eliminates the need for a heavyweight decoder. By leveraging this sparsity, we further propose an efficient sparse coefficient splatting method with CUDA optimization, rendering high-dimensional feature maps at high quality while incurring only the time cost of splatting an ultra-low-dimensional feature. Our experimental results demonstrate that LangSplatV2 not only achieves better or competitive query accuracy but is also significantly faster. Codes and demos are available at our project page: https://langsplat-v2.github.io.", "date": "2025-07-09", "categories": []}
{"id": "2503.19764", "title": "OpenLex3D: A Tiered Evaluation Benchmark for Open-Vocabulary 3D Scene Representations", "abstract": "3D scene understanding has been transformed by open-vocabulary language models that enable interaction via natural language. However, at present the evaluation of these representations is limited to datasets with closed-set semantics that do not capture the richness of language. This work presents OpenLex3D, a dedicated benchmark for evaluating 3D open-vocabulary scene representations. OpenLex3D provides entirely new label annotations for scenes from Replica, ScanNet++, and HM3D, which capture real-world linguistic variability by introducing synonymical object categories and additional nuanced descriptions. Our label sets provide 13 times more labels per scene than the original datasets. By introducing an open-set 3D semantic segmentation task and an object retrieval task, we evaluate various existing 3D open-vocabulary methods on OpenLex3D, showcasing failure cases, and avenues for improvement. Our experiments provide insights on feature precision, segmentation, and downstream capabilities. The benchmark is publicly available at: https://openlex3d.github.io/.", "date": "2025-03-25", "categories": []}
{"id": "2510.11992", "title": "PanoTPS-Net: Panoramic Room Layout Estimation via Thin Plate Spline Transformation", "abstract": "Accurately estimating the 3D layout of rooms is a crucial task in computer vision, with potential applications in robotics, augmented reality, and interior design. This paper proposes a novel model, PanoTPS-Net, to estimate room layout from a single panorama image. Leveraging a Convolutional Neural Network (CNN) and incorporating a Thin Plate Spline (TPS) spatial transformation, the architecture of PanoTPS-Net is divided into two stages: First, a convolutional neural network extracts the high-level features from the input images, allowing the network to learn the spatial parameters of the TPS transformation. Second, the TPS spatial transformation layer is generated to warp a reference layout to the required layout based on the predicted parameters. This unique combination empowers the model to properly predict room layouts while also generalizing effectively to both cuboid and non-cuboid layouts. Extensive experiments on publicly available datasets and comparisons with state-of-the-art methods demonstrate the effectiveness of the proposed method. The results underscore the model's accuracy in room layout estimation and emphasize the compatibility between the TPS transformation and panorama images. The robustness of the model in handling both cuboid and non-cuboid room layout estimation is evident with a 3DIoU value of 85.49, 86.16, 81.76, and 91.98 on PanoContext, Stanford-2D3D, Matterport3DLayout, and ZInD datasets, respectively. The source code is available at: https://github.com/HatemHosam/PanoTPS_Net.", "date": "2025-10-13", "categories": []}
{"id": "2510.12182", "title": "BEEP3D: Box-Supervised End-to-End Pseudo-Mask Generation for 3D Instance Segmentation", "abstract": "3D instance segmentation is crucial for understanding complex 3D environments, yet fully supervised methods require dense point-level annotations, resulting in substantial annotation costs and labor overhead. To mitigate this, box-level annotations have been explored as a weaker but more scalable form of supervision. However, box annotations inherently introduce ambiguity in overlapping regions, making accurate point-to-instance assignment challenging. Recent methods address this ambiguity by generating pseudo-masks through training a dedicated pseudo-labeler in an additional training stage. However, such two-stage pipelines often increase overall training time and complexity, hinder end-to-end optimization. To overcome these challenges, we propose BEEP3D-Box-supervised End-to-End Pseudo-mask generation for 3D instance segmentation. BEEP3D adopts a student-teacher framework, where the teacher model serves as a pseudo-labeler and is updated by the student model via an Exponential Moving Average. To better guide the teacher model to generate precise pseudo-masks, we introduce an instance center-based query refinement that enhances position query localization and leverages features near instance centers. Additionally, we design two novel losses-query consistency loss and masked feature consistency loss-to align semantic and geometric signals between predictions and pseudo-masks. Extensive experiments on ScanNetV2 and S3DIS datasets demonstrate that BEEP3D achieves competitive or superior performance compared to state-of-the-art weakly supervised methods while remaining computationally efficient.", "date": "2025-10-14", "categories": []}
{"id": "2510.12362", "title": "CurriFlow: Curriculum-Guided Depth Fusion with Optical Flow-Based Temporal Alignment for 3D Semantic Scene Completion", "abstract": "Semantic Scene Completion (SSC) aims to infer complete 3D geometry and semantics from monocular images, serving as a crucial capability for camera-based perception in autonomous driving. However, existing SSC methods relying on temporal stacking or depth projection often lack explicit motion reasoning and struggle with occlusions and noisy depth supervision. We propose CurriFlow, a novel semantic occupancy prediction framework that integrates optical flow-based temporal alignment with curriculum-guided depth fusion. CurriFlow employs a multi-level fusion strategy to align segmentation, visual, and depth features across frames using pre-trained optical flow, thereby improving temporal consistency and dynamic object understanding. To enhance geometric robustness, a curriculum learning mechanism progressively transitions from sparse yet accurate LiDAR depth to dense but noisy stereo depth during training, ensuring stable optimization and seamless adaptation to real-world deployment. Furthermore, semantic priors from the Segment Anything Model (SAM) provide category-agnostic supervision, strengthening voxel-level semantic learning and spatial consistency. Experiments on the SemanticKITTI benchmark demonstrate that CurriFlow achieves state-of-the-art performance with a mean IoU of 16.9, validating the effectiveness of our motion-guided and curriculum-aware design for camera-based 3D semantic scene completion.", "date": "2025-10-14", "categories": []}
{"id": "2510.12749", "title": "SPORTS: Simultaneous Panoptic Odometry, Rendering, Tracking and Segmentation for Urban Scenes Understanding", "abstract": "The scene perception, understanding, and simulation are fundamental techniques for embodied-AI agents, while existing solutions are still prone to segmentation deficiency, dynamic objects' interference, sensor data sparsity, and view-limitation problems. This paper proposes a novel framework, named SPORTS, for holistic scene understanding via tightly integrating Video Panoptic Segmentation (VPS), Visual Odometry (VO), and Scene Rendering (SR) tasks into an iterative and unified perspective. Firstly, VPS designs an adaptive attention-based geometric fusion mechanism to align cross-frame features via enrolling the pose, depth, and optical flow modality, which automatically adjust feature maps for different decoding stages. And a post-matching strategy is integrated to improve identities tracking. In VO, panoptic segmentation results from VPS are combined with the optical flow map to improve the confidence estimation of dynamic objects, which enhances the accuracy of the camera pose estimation and completeness of the depth map generation via the learning-based paradigm. Furthermore, the point-based rendering of SR is beneficial from VO, transforming sparse point clouds into neural fields to synthesize high-fidelity RGB views and twin panoptic views. Extensive experiments on three public datasets demonstrate that our attention-based feature fusion outperforms most existing state-of-the-art methods on the odometry, tracking, segmentation, and novel view synthesis tasks.", "date": "2025-10-14", "categories": []}
{"id": "2510.11509", "title": "Situat3DChange: Situated 3D Change Understanding Dataset for Multimodal Large Language Model", "abstract": "Physical environments and circumstances are fundamentally dynamic, yet current 3D datasets and evaluation benchmarks tend to concentrate on either dynamic scenarios or dynamic situations in isolation, resulting in incomplete comprehension. To overcome these constraints, we introduce Situat3DChange, an extensive dataset supporting three situation-aware change understanding tasks following the perception-action model: 121K question-answer pairs, 36K change descriptions for perception tasks, and 17K rearrangement instructions for the action task. To construct this large-scale dataset, Situat3DChange leverages 11K human observations of environmental changes to establish shared mental models and shared situational awareness for human-AI collaboration. These observations, enriched with egocentric and allocentric perspectives as well as categorical and coordinate spatial relations, are integrated using an LLM to support understanding of situated changes. To address the challenge of comparing pairs of point clouds from the same scene with minor changes, we propose SCReasoner, an efficient 3D MLLM approach that enables effective point cloud comparison with minimal parameter overhead and no additional tokens required for the language decoder. Comprehensive evaluation on Situat3DChange tasks highlights both the progress and limitations of MLLMs in dynamic scene and situation understanding. Additional experiments on data scaling and cross-domain transfer demonstrate the task-agnostic effectiveness of using Situat3DChange as a training dataset for MLLMs.", "date": "2025-10-13", "categories": []}
{"id": "2510.11649", "title": "PhySIC: Physically Plausible 3D Human-Scene Interaction and Contact from a Single Image", "abstract": "Reconstructing metrically accurate humans and their surrounding scenes from a single image is crucial for virtual reality, robotics, and comprehensive 3D scene understanding. However, existing methods struggle with depth ambiguity, occlusions, and physically inconsistent contacts. To address these challenges, we introduce PhySIC, a framework for physically plausible Human-Scene Interaction and Contact reconstruction. PhySIC recovers metrically consistent SMPL-X human meshes, dense scene surfaces, and vertex-level contact maps within a shared coordinate frame from a single RGB image. Starting from coarse monocular depth and body estimates, PhySIC performs occlusion-aware inpainting, fuses visible depth with unscaled geometry for a robust metric scaffold, and synthesizes missing support surfaces like floors. A confidence-weighted optimization refines body pose, camera parameters, and global scale by jointly enforcing depth alignment, contact priors, interpenetration avoidance, and 2D reprojection consistency. Explicit occlusion masking safeguards invisible regions against implausible configurations. PhySIC is efficient, requiring only 9 seconds for joint human-scene optimization and under 27 seconds end-to-end. It naturally handles multiple humans, enabling reconstruction of diverse interactions. Empirically, PhySIC outperforms single-image baselines, reducing mean per-vertex scene error from 641 mm to 227 mm, halving PA-MPJPE to 42 mm, and improving contact F1 from 0.09 to 0.51. Qualitative results show realistic foot-floor interactions, natural seating, and plausible reconstructions of heavily occluded furniture. By converting a single image into a physically plausible 3D human-scene pair, PhySIC advances scalable 3D scene understanding. Our implementation is publicly available at https://yuxuan-xue.com/physic.", "date": "2025-10-13", "categories": []}
{"id": "2505.23044", "title": "SpatialSplat: Efficient Semantic 3D from Sparse Unposed Images", "abstract": "A major breakthrough in 3D reconstruction is the feedforward paradigm to generate pixel-wise 3D points or Gaussian primitives from sparse, unposed images. To further incorporate semantics while avoiding the significant memory and storage costs of high-dimensional semantic features, existing methods extend this paradigm by associating each primitive with a compressed semantic feature vector. However, these methods have two major limitations: (a) the naively compressed feature compromises expressiveness, affecting the model's ability to capture fine-grained semantics, and (b) the pixel-wise primitive prediction introduces redundancy in overlapping areas, causing unnecessary memory overhead. To this end, we introduce \\textbf{SpatialSplat}, a feedforward framework that produces redundancy-aware Gaussians and capitalizes on a dual-field semantic representation. Particularly, with the insight that primitives within the same instance exhibit high semantic consistency, we decompose the semantic representation into a coarse feature field that encodes uncompressed semantics with minimal primitives, and a fine-grained yet low-dimensional feature field that captures detailed inter-instance relationships. Moreover, we propose a selective Gaussian mechanism, which retains only essential Gaussians in the scene, effectively eliminating redundant primitives. Our proposed Spatialsplat learns accurate semantic information and detailed instances prior with more compact 3D Gaussians, making semantic 3D reconstruction more applicable. We conduct extensive experiments to evaluate our method, demonstrating a remarkable 60\\% reduction in scene representation parameters while achieving superior performance over state-of-the-art methods. The code is available at https://github.com/shengyuuu/SpatialSplat.git", "date": "2025-05-29", "categories": []}
{"id": "2411.15043", "title": "Open-Vocabulary Online Semantic Mapping for SLAM", "abstract": "This paper presents an Open-Vocabulary Online 3D semantic mapping pipeline, that we denote by its acronym OVO. Given a sequence of posed RGB-D frames, we detect and track 3D segments, which we describe using CLIP vectors. These are computed from the viewpoints where they are observed by a novel CLIP merging method. Notably, our OVO has a significantly lower computational and memory footprint than offline baselines, while also showing better segmentation metrics than offline and online ones. Along with superior segmentation performance, we also show experimental results of our mapping contributions integrated with two different full SLAM backbones (Gaussian-SLAM and ORB-SLAM2), being the first ones using a neural network to merge CLIP descriptors and demonstrating end-to-end open-vocabulary online 3D mapping with loop closure.", "date": "2024-11-22", "categories": []}
{"id": "2510.05560", "title": "HoloScene: Simulation-Ready Interactive 3D Worlds from a Single Video", "abstract": "Digitizing the physical world into accurate simulation-ready virtual environments offers significant opportunities in a variety of fields such as augmented and virtual reality, gaming, and robotics. However, current 3D reconstruction and scene-understanding methods commonly fall short in one or more critical aspects, such as geometry completeness, object interactivity, physical plausibility, photorealistic rendering, or realistic physical properties for reliable dynamic simulation. To address these limitations, we introduce HoloScene, a novel interactive 3D reconstruction framework that simultaneously achieves these requirements. HoloScene leverages a comprehensive interactive scene-graph representation, encoding object geometry, appearance, and physical properties alongside hierarchical and inter-object relationships. Reconstruction is formulated as an energy-based optimization problem, integrating observational data, physical constraints, and generative priors into a unified, coherent objective. Optimization is efficiently performed via a hybrid approach combining sampling-based exploration with gradient-based refinement. The resulting digital twins exhibit complete and precise geometry, physical stability, and realistic rendering from novel viewpoints. Evaluations conducted on multiple benchmark datasets demonstrate superior performance, while practical use-cases in interactive gaming and real-time digital-twin manipulation illustrate HoloScene's broad applicability and effectiveness. Project page: https://xiahongchi.github.io/HoloScene.", "date": "2025-10-07", "categories": []}
{"id": "2505.05288", "title": "PlaceIt3D: Language-Guided Object Placement in Real 3D Scenes", "abstract": "We introduce the novel task of Language-Guided Object Placement in Real 3D Scenes. Our model is given a 3D scene's point cloud, a 3D asset, and a textual prompt broadly describing where the 3D asset should be placed. The task here is to find a valid placement for the 3D asset that respects the prompt. Compared with other language-guided localization tasks in 3D scenes such as grounding, this task has specific challenges: it is ambiguous because it has multiple valid solutions, and it requires reasoning about 3D geometric relationships and free space. We inaugurate this task by proposing a new benchmark and evaluation protocol. We also introduce a new dataset for training 3D LLMs on this task, as well as the first method to serve as a non-trivial baseline. We believe that this challenging task and our new benchmark could become part of the suite of benchmarks used to evaluate and compare generalist 3D LLM models.", "date": "2025-05-08", "categories": []}
{"id": "2509.06266", "title": "Spatial Reasoning with Vision-Language Models in Ego-Centric Multi-View Scenes", "abstract": "Understanding 3D spatial relationships remains a major limitation of current Vision-Language Models (VLMs). Prior work has addressed this issue by creating spatial question-answering (QA) datasets based on single images or indoor videos. However, real-world embodied AI agents such as robots and self-driving cars typically rely on ego-centric, multi-view observations. To this end, we introduce Ego3D-Bench, a new benchmark designed to evaluate the spatial reasoning abilities of VLMs using ego-centric, multi-view outdoor data. Ego3D-Bench comprises over 8,600 QA pairs, created with significant involvement from human annotators to ensure quality and diversity. We benchmark 16 SOTA VLMs, including GPT-4o, Gemini1.5-Pro, InternVL3, and Qwen2.5-VL. Our results reveal a notable performance gap between human level scores and VLM performance, highlighting that current VLMs still fall short of human level spatial understanding. To bridge this gap, we propose Ego3D-VLM, a post-training framework that enhances 3D spatial reasoning of VLMs. Ego3D-VLM generates cognitive map based on estimated global 3D coordinates, resulting in 12% average improvement on multi-choice QA and 56% average improvement on absolute distance estimation. Ego3D-VLM is modular and can be integrated with any existing VLM. Together, Ego3D-Bench and Ego3D-VLM offer valuable tools for advancing toward human level spatial understanding in real-world, multi-view environments.", "date": "2025-09-08", "categories": []}
{"id": "2509.24288", "title": "ASIA: Adaptive 3D Segmentation using Few Image Annotations", "abstract": "We introduce ASIA (Adaptive 3D Segmentation using few Image Annotations), a novel framework that enables segmentation of possibly non-semantic and non-text-describable \"parts\" in 3D. Our segmentation is controllable through a few user-annotated in-the-wild images, which are easier to collect than multi-view images, less demanding to annotate than 3D models, and more precise than potentially ambiguous text descriptions. Our method leverages the rich priors of text-to-image diffusion models, such as Stable Diffusion (SD), to transfer segmentations from image space to 3D, even when the annotated and target objects differ significantly in geometry or structure. During training, we optimize a text token for each segment and fine-tune our model with a novel cross-view part correspondence loss. At inference, we segment multi-view renderings of the 3D mesh, fuse the labels in UV-space via voting, refine them with our novel Noise Optimization technique, and finally map the UV-labels back onto the mesh. ASIA provides a practical and generalizable solution for both semantic and non-semantic 3D segmentation tasks, outperforming existing methods by a noticeable margin in both quantitative and qualitative evaluations.", "date": "2025-09-29", "categories": []}
{"id": "2509.24966", "title": "Social 3D Scene Graphs: Modeling Human Actions and Relations for Interactive Service Robots", "abstract": "Understanding how people interact with their surroundings and each other is essential for enabling robots to act in socially compliant and context-aware ways. While 3D Scene Graphs have emerged as a powerful semantic representation for scene understanding, existing approaches largely ignore humans in the scene, also due to the lack of annotated human-environment relationships. Moreover, existing methods typically capture only open-vocabulary relations from single image frames, which limits their ability to model long-range interactions beyond the observed content. We introduce Social 3D Scene Graphs, an augmented 3D Scene Graph representation that captures humans, their attributes, activities and relationships in the environment, both local and remote, using an open-vocabulary framework. Furthermore, we introduce a new benchmark consisting of synthetic environments with comprehensive human-scene relationship annotations and diverse types of queries for evaluating social scene understanding in 3D. The experiments demonstrate that our representation improves human activity prediction and reasoning about human-environment relations, paving the way toward socially intelligent robots.", "date": "2025-09-29", "categories": []}
{"id": "2503.21767", "title": "Semantic Consistent Language Gaussian Splatting for Point-Level Open-vocabulary Querying", "abstract": "Open-vocabulary 3D scene understanding is crucial for robotics applications, such as natural language-driven manipulation, human-robot interaction, and autonomous navigation. Existing methods for querying 3D Gaussian Splatting often struggle with inconsistent 2D mask supervision and lack a robust 3D point-level retrieval mechanism. In this work, (i) we present a novel point-level querying framework that performs tracking on segmentation masks to establish a semantically consistent ground-truth for distilling the language Gaussians; (ii) we introduce a GT-anchored querying approach that first retrieves the distilled ground-truth and subsequently uses the ground-truth to query the individual Gaussians. Extensive experiments on three benchmark datasets demonstrate that the proposed method outperforms state-of-the-art performance. Our method achieves an mIoU improvement of +4.14, +20.42, and +1.7 on the LERF, 3D-OVS, and Replica datasets. These results validate our framework as a promising step toward open-vocabulary understanding in real-world robotic systems.", "date": "2025-03-27", "categories": []}
{"id": "2507.02705", "title": "SIU3R: Simultaneous Scene Understanding and 3D Reconstruction Beyond Feature Alignment", "abstract": "Simultaneous understanding and 3D reconstruction plays an important role in developing end-to-end embodied intelligent systems. To achieve this, recent approaches resort to 2D-to-3D feature alignment paradigm, which leads to limited 3D understanding capability and potential semantic information loss. In light of this, we propose SIU3R, the first alignment-free framework for generalizable simultaneous understanding and 3D reconstruction from unposed images. Specifically, SIU3R bridges reconstruction and understanding tasks via pixel-aligned 3D representation, and unifies multiple understanding (segmentation) tasks into a set of unified learnable queries, enabling native 3D understanding without the need of alignment with 2D models. To encourage collaboration between the two tasks with shared representation, we further conduct in-depth analyses of their mutual benefits, and propose two lightweight modules to facilitate their interaction. Extensive experiments demonstrate that our method achieves state-of-the-art performance not only on the individual tasks of 3D reconstruction and understanding, but also on the task of simultaneous understanding and 3D reconstruction, highlighting the advantages of our alignment-free framework and the effectiveness of the mutual benefit designs. Project page: https://insomniaaac.github.io/siu3r/", "date": "2025-07-03", "categories": []}
{"id": "2509.21388", "title": "TUN3D: Towards Real-World Scene Understanding from Unposed Images", "abstract": "Layout estimation and 3D object detection are two fundamental tasks in indoor scene understanding. When combined, they enable the creation of a compact yet semantically rich spatial representation of a scene. Existing approaches typically rely on point cloud input, which poses a major limitation since most consumer cameras lack depth sensors and visual-only data remains far more common. We address this issue with TUN3D, the first method that tackles joint layout estimation and 3D object detection in real scans, given multi-view images as input, and does not require ground-truth camera poses or depth supervision. Our approach builds on a lightweight sparse-convolutional backbone and employs two dedicated heads: one for 3D object detection and one for layout estimation, leveraging a novel and effective parametric wall representation. Extensive experiments show that TUN3D achieves state-of-the-art performance across three challenging scene understanding benchmarks: (i) using ground-truth point clouds, (ii) using posed images, and (iii) using unposed images. While performing on par with specialized 3D object detection methods, TUN3D significantly advances layout estimation, setting a new benchmark in holistic indoor scene understanding. Code is available at https://github.com/col14m/tun3d .", "date": "2025-09-23", "categories": []}
{"id": "2509.19973", "title": "OmniScene: Attention-Augmented Multimodal 4D Scene Understanding for Autonomous Driving", "abstract": "Human vision is capable of transforming two-dimensional observations into an egocentric three-dimensional scene understanding, which underpins the ability to translate complex scenes and exhibit adaptive behaviors. This capability, however, remains lacking in current autonomous driving systems, where mainstream approaches primarily rely on depth-based 3D reconstruction rather than true scene understanding. To address this limitation, we propose a novel human-like framework called OmniScene. First, we introduce the OmniScene Vision-Language Model (OmniVLM), a vision-language framework that integrates multi-view and temporal perception for holistic 4D scene understanding. Then, harnessing a teacher-student OmniVLM architecture and knowledge distillation, we embed textual representations into 3D instance features for semantic supervision, enriching feature learning, and explicitly capturing human-like attentional semantics. These feature representations are further aligned with human driving behaviors, forming a more human-like perception-understanding-action architecture. In addition, we propose a Hierarchical Fusion Strategy (HFS) to address imbalances in modality contributions during multimodal integration. Our approach adaptively calibrates the relative significance of geometric and semantic features at multiple abstraction levels, enabling the synergistic use of complementary cues from visual and textual modalities. This learnable dynamic fusion enables a more nuanced and effective exploitation of heterogeneous information. We evaluate OmniScene comprehensively on the nuScenes dataset, benchmarking it against over ten state-of-the-art models across various tasks. Our approach consistently achieves superior results, establishing new benchmarks in perception, prediction, planning, and visual question answering.", "date": "2025-09-24", "categories": []}
{"id": "2506.21188", "title": "GroundFlow: A Plug-in Module for Temporal Reasoning on 3D Point Cloud Sequential Grounding", "abstract": "Sequential grounding in 3D point clouds (SG3D) refers to locating sequences of objects by following text instructions for a daily activity with detailed steps. Current 3D visual grounding (3DVG) methods treat text instructions with multiple steps as a whole, without extracting useful temporal information from each step. However, the instructions in SG3D often contain pronouns such as \"it\", \"here\" and \"the same\" to make language expressions concise. This requires grounding methods to understand the context and retrieve relevant information from previous steps to correctly locate object sequences. Due to the lack of an effective module for collecting related historical information, state-of-the-art 3DVG methods face significant challenges in adapting to the SG3D task. To fill this gap, we propose GroundFlow -- a plug-in module for temporal reasoning on 3D point cloud sequential grounding. Firstly, we demonstrate that integrating GroundFlow improves the task accuracy of 3DVG baseline methods by a large margin (+7.5\\% and +10.2\\%) in the SG3D benchmark, even outperforming a 3D large language model pre-trained on various datasets. Furthermore, we selectively extract both short-term and long-term step information based on its relevance to the current instruction, enabling GroundFlow to take a comprehensive view of historical information and maintain its temporal understanding advantage as step counts increase. Overall, our work introduces temporal reasoning capabilities to existing 3DVG models and achieves state-of-the-art performance in the SG3D benchmark across five datasets.", "date": "2025-06-26", "categories": []}
{"id": "2509.16721", "title": "Text-Scene: A Scene-to-Language Parsing Framework for 3D Scene Understanding", "abstract": "Enabling agents to understand and interact with complex 3D scenes is a fundamental challenge for embodied artificial intelligence systems. While Multimodal Large Language Models (MLLMs) have achieved significant progress in 2D image understanding, extending such capabilities to 3D scenes remains difficult: 1) 3D environment involves richer concepts such as spatial relationships, affordances, physics, layout, and so on, 2) the absence of large-scale 3D vision-language datasets has posed a significant obstacle. In this paper, we introduce Text-Scene, a framework that automatically parses 3D scenes into textual descriptions for scene understanding. Given a 3D scene, our model identifies object attributes and spatial relationships, and then generates a coherent summary of the whole scene, bridging the gap between 3D observation and language without requiring human-in-the-loop intervention. By leveraging both geometric analysis and MLLMs, Text-Scene produces descriptions that are accurate, detailed, and human-interpretable, capturing object-level details and global-level context. Experimental results on benchmarks demonstrate that our textual parses can faithfully represent 3D scenes and benefit downstream tasks. To evaluate the reasoning capability of MLLMs, we present InPlan3D, a comprehensive benchmark for 3D task planning, consisting of 3174 long-term planning tasks across 636 indoor scenes. We emphasize clarity and accessibility in our approach, aiming to make 3D scene content understandable through language. Code and datasets will be released.", "date": "2025-09-20", "categories": []}
{"id": "2509.17462", "title": "MAESTRO: Task-Relevant Optimization via Adaptive Feature Enhancement and Suppression for Multi-task 3D Perception", "abstract": "The goal of multi-task learning is to learn to conduct multiple tasks simultaneously based on a shared data representation. While this approach can improve learning efficiency, it may also cause performance degradation due to task conflicts that arise when optimizing the model for different objectives. To address this challenge, we introduce MAESTRO, a structured framework designed to generate task-specific features and mitigate feature interference in multi-task 3D perception, including 3D object detection, bird's-eye view (BEV) map segmentation, and 3D occupancy prediction. MAESTRO comprises three components: the Class-wise Prototype Generator (CPG), the Task-Specific Feature Generator (TSFG), and the Scene Prototype Aggregator (SPA). CPG groups class categories into foreground and background groups and generates group-wise prototypes. The foreground and background prototypes are assigned to the 3D object detection task and the map segmentation task, respectively, while both are assigned to the 3D occupancy prediction task. TSFG leverages these prototype groups to retain task-relevant features while suppressing irrelevant features, thereby enhancing the performance for each task. SPA enhances the prototype groups assigned for 3D occupancy prediction by utilizing the information produced by the 3D object detection head and the map segmentation head. Extensive experiments on the nuScenes and Occ3D benchmarks demonstrate that MAESTRO consistently outperforms existing methods across 3D object detection, BEV map segmentation, and 3D occupancy prediction tasks.", "date": "2025-09-22", "categories": []}
{"id": "2509.15924", "title": "Sparse Multiview Open-Vocabulary 3D Detection", "abstract": "The ability to interpret and comprehend a 3D scene is essential for many vision and robotics systems. In numerous applications, this involves 3D object detection, i.e.~identifying the location and dimensions of objects belonging to a specific category, typically represented as bounding boxes. This has traditionally been solved by training to detect a fixed set of categories, which limits its use. In this work, we investigate open-vocabulary 3D object detection in the challenging yet practical sparse-view setting, where only a limited number of posed RGB images are available as input. Our approach is training-free, relying on pre-trained, off-the-shelf 2D foundation models instead of employing computationally expensive 3D feature fusion or requiring 3D-specific learning. By lifting 2D detections and directly optimizing 3D proposals for featuremetric consistency across views, we fully leverage the extensive training data available in 2D compared to 3D. Through standard benchmarks, we demonstrate that this simple pipeline establishes a powerful baseline, performing competitively with state-of-the-art techniques in densely sampled scenarios while significantly outperforming them in the sparse-view setting.", "date": "2025-09-19", "categories": []}
{"id": "2412.18131", "title": "UniPLV: Towards Label-Efficient Open-World 3D Scene Understanding by Regional Visual Language Supervision", "abstract": "Open-world 3D scene understanding is a critical challenge that involves recognizing and distinguishing diverse objects and categories from 3D data, such as point clouds, without relying on manual annotations. Traditional methods struggle with this open-world task, especially due to the limitations of constructing extensive point cloud-text pairs and handling multimodal data effectively. In response to these challenges, we present UniPLV, a robust framework that unifies point clouds, images, and text within a single learning paradigm for comprehensive 3D scene understanding. UniPLV leverages images as a bridge to co-embed 3D points with pre-aligned images and text in a shared feature space, eliminating the need for labor-intensive point cloud-text pair crafting. Our framework achieves precise multimodal alignment through two innovative strategies: (i) Logit and feature distillation modules between images and point clouds to enhance feature coherence; (ii) A vision-point matching module that implicitly corrects 3D semantic predictions affected by projection inaccuracies from points to pixels. To further boost performance, we implement four task-specific losses alongside a two-stage training strategy. Extensive experiments demonstrate that UniPLV significantly surpasses state-of-the-art methods, with average improvements of 15.6% and 14.8% in semantic segmentation for Base-Annotated and Annotation-Free tasks, respectively. These results underscore UniPLV's efficacy in pushing the boundaries of open-world 3D scene understanding. We will release the code to support future research and development.", "date": "2024-12-24", "categories": []}
{"id": "2509.12938", "title": "Beyond Averages: Open-Vocabulary 3D Scene Understanding with Gaussian Splatting and Bag of Embeddings", "abstract": "Novel view synthesis has seen significant advancements with 3D Gaussian Splatting (3DGS), enabling real-time photorealistic rendering. However, the inherent fuzziness of Gaussian Splatting presents challenges for 3D scene understanding, restricting its broader applications in AR/VR and robotics. While recent works attempt to learn semantics via 2D foundation model distillation, they inherit fundamental limitations: alpha blending averages semantics across objects, making 3D-level understanding impossible. We propose a paradigm-shifting alternative that bypasses differentiable rendering for semantics entirely. Our key insight is to leverage predecomposed object-level Gaussians and represent each object through multiview CLIP feature aggregation, creating comprehensive \"bags of embeddings\" that holistically describe objects. This allows: (1) accurate open-vocabulary object retrieval by comparing text queries to object-level (not Gaussian-level) embeddings, and (2) seamless task adaptation: propagating object IDs to pixels for 2D segmentation or to Gaussians for 3D extraction. Experiments demonstrate that our method effectively overcomes the challenges of 3D open-vocabulary object extraction while remaining comparable to state-of-the-art performance in 2D open-vocabulary segmentation, ensuring minimal compromise.", "date": "2025-09-16", "categories": []}
{"id": "2504.00763", "title": "UnIRe: Unsupervised Instance Decomposition for Dynamic Urban Scene Reconstruction", "abstract": "Reconstructing and decomposing dynamic urban scenes is crucial for autonomous driving, urban planning, and scene editing. However, existing methods fail to perform instance-aware decomposition without manual annotations, which is crucial for instance-level scene editing.We propose UnIRe, a 3D Gaussian Splatting (3DGS) based approach that decomposes a scene into a static background and individual dynamic instances using only RGB images and LiDAR point clouds. At its core, we introduce 4D superpoints, a novel representation that clusters multi-frame LiDAR points in 4D space, enabling unsupervised instance separation based on spatiotemporal correlations. These 4D superpoints serve as the foundation for our decomposed 4D initialization, i.e., providing spatial and temporal initialization to train a dynamic 3DGS for arbitrary dynamic classes without requiring bounding boxes or object templates.Furthermore, we introduce a smoothness regularization strategy in both 2D and 3D space, further improving the temporal stability.Experiments on benchmark datasets show that our method outperforms existing methods in decomposed dynamic scene reconstruction while enabling accurate and flexible instance-level editing, making it a practical solution for real-world applications.", "date": "2025-04-01", "categories": []}
{"id": "2509.11097", "title": "3DAeroRelief: The first 3D Benchmark UAV Dataset for Post-Disaster Assessment", "abstract": "Timely assessment of structural damage is critical for disaster response and recovery. However, most prior work in natural disaster analysis relies on 2D imagery, which lacks depth, suffers from occlusions, and provides limited spatial context. 3D semantic segmentation offers a richer alternative, but existing 3D benchmarks focus mainly on urban or indoor scenes, with little attention to disaster-affected areas. To address this gap, we present 3DAeroRelief--the first 3D benchmark dataset specifically designed for post-disaster assessment. Collected using low-cost unmanned aerial vehicles (UAVs) over hurricane-damaged regions, the dataset features dense 3D point clouds reconstructed via Structure-from-Motion and Multi-View Stereo techniques. Semantic annotations were produced through manual 2D labeling and projected into 3D space. Unlike existing datasets, 3DAeroRelief captures 3D large-scale outdoor environments with fine-grained structural damage in real-world disaster contexts. UAVs enable affordable, flexible, and safe data collection in hazardous areas, making them particularly well-suited for emergency scenarios. To demonstrate the utility of 3DAeroRelief, we evaluate several state-of-the-art 3D segmentation models on the dataset to highlight both the challenges and opportunities of 3D scene understanding in disaster response. Our dataset serves as a valuable resource for advancing robust 3D vision systems in real-world applications for post-disaster scenarios.", "date": "2025-09-14", "categories": []}
{"id": "2509.11895", "title": "Integrating Prior Observations for Incremental 3D Scene Graph Prediction", "abstract": "3D semantic scene graphs (3DSSG) provide compact structured representations of environments by explicitly modeling objects, attributes, and relationships. While 3DSSGs have shown promise in robotics and embodied AI, many existing methods rely mainly on sensor data, not integrating further information from semantically rich environments. Additionally, most methods assume access to complete scene reconstructions, limiting their applicability in real-world, incremental settings. This paper introduces a novel heterogeneous graph model for incremental 3DSSG prediction that integrates additional, multi-modal information, such as prior observations, directly into the message-passing process. Utilizing multiple layers, the model flexibly incorporates global and local scene representations without requiring specialized modules or full scene reconstructions. We evaluate our approach on the 3DSSG dataset, showing that GNNs enriched with multi-modal information such as semantic embeddings (e.g., CLIP) and prior observations offer a scalable and generalizable solution for complex, real-world environments. The full source code of the presented architecture will be made available at https://github.com/m4renz/incremental-scene-graph-prediction.", "date": "2025-09-15", "categories": []}
{"id": "2411.07725", "title": "ALOcc: Adaptive Lifting-Based 3D Semantic Occupancy and Cost Volume-Based Flow Predictions", "abstract": "3D semantic occupancy and flow prediction are fundamental to spatiotemporal scene understanding. This paper proposes a vision-based framework with three targeted improvements. First, we introduce an occlusion-aware adaptive lifting mechanism incorporating depth denoising. This enhances the robustness of 2D-to-3D feature transformation while mitigating reliance on depth priors. Second, we enforce 3D-2D semantic consistency via jointly optimized prototypes, using confidence- and category-aware sampling to address the long-tail classes problem. Third, to streamline joint prediction, we devise a BEV-centric cost volume to explicitly correlate semantic and flow features, supervised by a hybrid classification-regression scheme that handles diverse motion scales. Our purely convolutional architecture establishes new SOTA performance on multiple benchmarks for both semantic occupancy and joint occupancy semantic-flow prediction. We also present a family of models offering a spectrum of efficiency-performance trade-offs. Our real-time version exceeds all existing real-time methods in speed and accuracy, ensuring its practical viability.", "date": "2024-11-12", "categories": []}
{"id": "2503.13111", "title": "MM-Spatial: Exploring 3D Spatial Understanding in Multimodal LLMs", "abstract": "Multimodal large language models (MLLMs) excel at 2D visual understanding but remain limited in their ability to reason about 3D space. In this work, we leverage large-scale high-quality 3D scene data with open-set annotations to introduce 1) a novel supervised fine-tuning dataset and 2) a new evaluation benchmark, focused on indoor scenes. Our Cubify Anything VQA (CA-VQA) data covers diverse spatial tasks including spatial relationship prediction, metric size and distance estimation, and 3D grounding. We show that CA-VQA enables us to train MM-Spatial, a strong generalist MLLM that also achieves state-of-the-art performance on 3D spatial understanding benchmarks, including our own. We show how incorporating metric depth and multi-view inputs (provided in CA-VQA) can further improve 3D understanding, and demonstrate that data alone allows our model to achieve depth perception capabilities comparable to dedicated monocular depth estimation models.", "date": "2025-03-17", "categories": []}
{"id": "2509.03635", "title": "Reg3D: Reconstructive Geometry Instruction Tuning for 3D Scene Understanding", "abstract": "The rapid development of Large Multimodal Models (LMMs) has led to remarkable progress in 2D visual understanding; however, extending these capabilities to 3D scene understanding remains a significant challenge. Existing approaches predominantly rely on text-only supervision, which fails to provide the geometric constraints required for learning robust 3D spatial representations. In this paper, we introduce Reg3D, a novel Reconstructive Geometry Instruction Tuning framework that addresses this limitation by incorporating geometry-aware supervision directly into the training process. Our key insight is that effective 3D understanding necessitates reconstructing underlying geometric structures rather than merely describing them. Unlike existing methods that inject 3D information solely at the input level, Reg3D adopts a dual-supervision paradigm that leverages 3D geometric information both as input and as explicit learning targets. Specifically, we design complementary object-level and frame-level reconstruction tasks within a dual-encoder architecture, enforcing geometric consistency to encourage the development of spatial reasoning capabilities. Extensive experiments on ScanQA, Scan2Cap, ScanRefer, and SQA3D demonstrate that Reg3D delivers substantial performance improvements, establishing a new training paradigm for spatially aware multimodal models.", "date": "2025-09-03", "categories": []}
{"id": "2502.01401", "title": "Language-to-Space Programming for Training-Free 3D Visual Grounding", "abstract": "3D visual grounding (3DVG) is challenging due to the need to understand 3D spatial relations. While supervised approaches have achieved superior performance, they are constrained by the scarcity and high annotation costs of 3D vision-language datasets. Training-free approaches based on LLMs/VLMs eliminate the need for large-scale training data, but they either incur prohibitive grounding time and token costs or have unsatisfactory accuracy. To address the challenges, we introduce a novel method for training-free 3D visual grounding, namely Language-to-Space Programming (LaSP). LaSP introduces LLM-generated codes to analyze 3D spatial relations among objects, along with a pipeline that evaluates and optimizes the codes automatically. Experimental results demonstrate that LaSP achieves 52.9% accuracy on the Nr3D benchmark, ranking among the best training-free methods. Moreover, it substantially reduces the grounding time and token costs, offering a balanced trade-off between performance and efficiency.", "date": "2025-02-03", "categories": []}
{"id": "2508.19699", "title": "LabelGS: Label-Aware 3D Gaussian Splatting for 3D Scene Segmentation", "abstract": "3D Gaussian Splatting (3DGS) has emerged as a novel explicit representation for 3D scenes, offering both high-fidelity reconstruction and efficient rendering. However, 3DGS lacks 3D segmentation ability, which limits its applicability in tasks that require scene understanding. The identification and isolating of specific object components is crucial. To address this limitation, we propose Label-aware 3D Gaussian Splatting (LabelGS), a method that augments the Gaussian representation with object label.LabelGS introduces cross-view consistent semantic masks for 3D Gaussians and employs a novel Occlusion Analysis Model to avoid overfitting occlusion during optimization, Main Gaussian Labeling model to lift 2D semantic prior to 3D Gaussian and Gaussian Projection Filter to avoid Gaussian label conflict. Our approach achieves effective decoupling of Gaussian representations and refines the 3DGS optimization process through a random region sampling strategy, significantly improving efficiency. Extensive experiments demonstrate that LabelGS outperforms previous state-of-the-art methods, including Feature-3DGS, in the 3D scene segmentation task. Notably, LabelGS achieves a remarkable 22X speedup in training compared to Feature-3DGS, at a resolution of 1440X1080. Our code will be at https://github.com/garrisonz/LabelGS.", "date": "2025-08-27", "categories": []}
{"id": "2508.17437", "title": "Pixie: Fast and Generalizable Supervised Learning of 3D Physics from Pixels", "abstract": "Inferring the physical properties of 3D scenes from visual information is a critical yet challenging task for creating interactive and realistic virtual worlds. While humans intuitively grasp material characteristics such as elasticity or stiffness, existing methods often rely on slow, per-scene optimization, limiting their generalizability and application. To address this problem, we introduce PIXIE, a novel method that trains a generalizable neural network to predict physical properties across multiple scenes from 3D visual features purely using supervised losses. Once trained, our feed-forward network can perform fast inference of plausible material fields, which coupled with a learned static scene representation like Gaussian Splatting enables realistic physics simulation under external forces. To facilitate this research, we also collected PIXIEVERSE, one of the largest known datasets of paired 3D assets and physic material annotations. Extensive evaluations demonstrate that PIXIE is about 1.46-4.39x better and orders of magnitude faster than test-time optimization methods. By leveraging pretrained visual features like CLIP, our method can also zero-shot generalize to real-world scenes despite only ever been trained on synthetic data. https://pixie-3d.github.io/", "date": "2025-08-20", "categories": []}
{"id": "2508.19165", "title": "Dual Enhancement on 3D Vision-Language Perception for Monocular 3D Visual Grounding", "abstract": "Monocular 3D visual grounding is a novel task that aims to locate 3D objects in RGB images using text descriptions with explicit geometry information. Despite the inclusion of geometry details in the text, we observe that the text embeddings are sensitive to the magnitude of numerical values but largely ignore the associated measurement units. For example, simply equidistant mapping the length with unit \"meter\" to \"decimeters\" or \"centimeters\" leads to severe performance degradation, even though the physical length remains equivalent. This observation signifies the weak 3D comprehension of pre-trained language model, which generates misguiding text features to hinder 3D perception. Therefore, we propose to enhance the 3D perception of model on text embeddings and geometry features with two simple and effective methods. Firstly, we introduce a pre-processing method named 3D-text Enhancement (3DTE), which enhances the comprehension of mapping relationships between different units by augmenting the diversity of distance descriptors in text queries. Next, we propose a Text-Guided Geometry Enhancement (TGE) module to further enhance the 3D-text information by projecting the basic text features into geometrically consistent space. These 3D-enhanced text features are then leveraged to precisely guide the attention of geometry features. We evaluate the proposed method through extensive comparisons and ablation studies on the Mono3DRefer dataset. Experimental results demonstrate substantial improvements over previous methods, achieving new state-of-the-art results with a notable accuracy gain of 11.94\\% in the \"Far\" scenario. Our code will be made publicly available.", "date": "2025-08-26", "categories": []}
{"id": "2412.04380", "title": "EmbodiedOcc: Embodied 3D Occupancy Prediction for Vision-based Online Scene Understanding", "abstract": "3D occupancy prediction provides a comprehensive description of the surrounding scenes and has become an essential task for 3D perception. Most existing methods focus on offline perception from one or a few views and cannot be applied to embodied agents that demand to gradually perceive the scene through progressive embodied exploration. In this paper, we formulate an embodied 3D occupancy prediction task to target this practical scenario and propose a Gaussian-based EmbodiedOcc framework to accomplish it. We initialize the global scene with uniform 3D semantic Gaussians and progressively update local regions observed by the embodied agent. For each update, we extract semantic and structural features from the observed image and efficiently incorporate them via deformable cross-attention to refine the regional Gaussians. Finally, we employ Gaussian-to-voxel splatting to obtain the global 3D occupancy from the updated 3D Gaussians. Our EmbodiedOcc assumes an unknown (i.e., uniformly distributed) environment and maintains an explicit global memory of it with 3D Gaussians. It gradually gains knowledge through the local refinement of regional Gaussians, which is consistent with how humans understand new scenes through embodied exploration. We reorganize an EmbodiedOcc-ScanNet benchmark based on local annotations to facilitate the evaluation of the embodied 3D occupancy prediction task. Our EmbodiedOcc outperforms existing methods by a large margin and accomplishes the embodied occupancy prediction with high accuracy and efficiency. Code: https://github.com/YkiWu/EmbodiedOcc.", "date": "2024-12-05", "categories": []}
{"id": "2502.17288", "title": "GaussianFlowOcc: Sparse and Weakly Supervised Occupancy Estimation using Gaussian Splatting and Temporal Flow", "abstract": "Occupancy estimation has become a prominent task in 3D computer vision, particularly within the autonomous driving community. In this paper, we present a novel approach to occupancy estimation, termed GaussianFlowOcc, which is inspired by Gaussian Splatting and replaces traditional dense voxel grids with a sparse 3D Gaussian representation. Our efficient model architecture based on a Gaussian Transformer significantly reduces computational and memory requirements by eliminating the need for expensive 3D convolutions used with inefficient voxel-based representations that predominantly represent empty 3D spaces. GaussianFlowOcc effectively captures scene dynamics by estimating temporal flow for each Gaussian during the overall network training process, offering a straightforward solution to a complex problem that is often neglected by existing methods. Moreover, GaussianFlowOcc is designed for scalability, as it employs weak supervision and does not require costly dense 3D voxel annotations based on additional data (e.g., LiDAR). Through extensive experimentation, we demonstrate that GaussianFlowOcc significantly outperforms all previous methods for weakly supervised occupancy estimation on the nuScenes dataset while featuring an inference speed that is 50 times faster than current SOTA.", "date": "2025-02-24", "categories": []}
{"id": "2508.14278", "title": "GALA: Guided Attention with Language Alignment for Open Vocabulary Gaussian Splatting", "abstract": "3D scene reconstruction and understanding have gained increasing popularity, yet existing methods still struggle to capture fine-grained, language-aware 3D representations from 2D images. In this paper, we present GALA, a novel framework for open-vocabulary 3D scene understanding with 3D Gaussian Splatting (3DGS). GALA distills a scene-specific 3D instance feature field via self-supervised contrastive learning. To extend to generalized language feature fields, we introduce the core contribution of GALA, a cross-attention module with two learnable codebooks that encode view-independent semantic embeddings. This design not only ensures intra-instance feature similarity but also supports seamless 2D and 3D open-vocabulary queries. It reduces memory consumption by avoiding per-Gaussian high-dimensional feature learning. Extensive experiments on real-world datasets demonstrate GALA's remarkable open-vocabulary performance on both 2D and 3D.", "date": "2025-08-19", "categories": []}
{"id": "2508.05064", "title": "A Study of the Framework and Real-World Applications of Language Embedding for 3D Scene Understanding", "abstract": "Gaussian Splatting has rapidly emerged as a transformative technique for real-time 3D scene representation, offering a highly efficient and expressive alternative to Neural Radiance Fields (NeRF). Its ability to render complex scenes with high fidelity has enabled progress across domains such as scene reconstruction, robotics, and interactive content creation. More recently, the integration of Large Language Models (LLMs) and language embeddings into Gaussian Splatting pipelines has opened new possibilities for text-conditioned generation, editing, and semantic scene understanding. Despite these advances, a comprehensive overview of this emerging intersection has been lacking. This survey presents a structured review of current research efforts that combine language guidance with 3D Gaussian Splatting, detailing theoretical foundations, integration strategies, and real-world use cases. We highlight key limitations such as computational bottlenecks, generalizability, and the scarcity of semantically annotated 3D Gaussian data and outline open challenges and future directions for advancing language-guided 3D scene understanding using Gaussian Splatting.", "date": "2025-08-07", "categories": []}
{"id": "2412.08331", "title": "SLGaussian: Fast Language Gaussian Splatting in Sparse Views", "abstract": "3D semantic field learning is crucial for applications like autonomous navigation, AR/VR, and robotics, where accurate comprehension of 3D scenes from limited viewpoints is essential. Existing methods struggle under sparse view conditions, relying on inefficient per-scene multi-view optimizations, which are impractical for many real-world tasks. To address this, we propose SLGaussian, a feed-forward method for constructing 3D semantic fields from sparse viewpoints, allowing direct inference of 3DGS-based scenes. By ensuring consistent SAM segmentations through video tracking and using low-dimensional indexing for high-dimensional CLIP features, SLGaussian efficiently embeds language information in 3D space, offering a robust solution for accurate 3D scene understanding under sparse view conditions. In experiments on two-view sparse 3D object querying and segmentation in the LERF and 3D-OVS datasets, SLGaussian outperforms existing methods in chosen IoU, Localization Accuracy, and mIoU. Moreover, our model achieves scene inference in under 30 seconds and open-vocabulary querying in just 0.011 seconds per query.", "date": "2024-12-11", "categories": []}
{"id": "2505.08124", "title": "SLAG: Scalable Language-Augmented Gaussian Splatting", "abstract": "Language-augmented scene representations hold great promise for large-scale robotics applications such as search-and-rescue, smart cities, and mining. Many of these scenarios are time-sensitive, requiring rapid scene encoding while also being data-intensive, necessitating scalable solutions. Deploying these representations on robots with limited computational resources further adds to the challenge. To address this, we introduce SLAG, a multi-GPU framework for language-augmented Gaussian splatting that enhances the speed and scalability of embedding large scenes. Our method integrates 2D visual-language model features into 3D scenes using SAM and CLIP. Unlike prior approaches, SLAG eliminates the need for a loss function to compute per-Gaussian language embeddings. Instead, it derives embeddings from 3D Gaussian scene parameters via a normalized weighted average, enabling highly parallelized scene encoding. Additionally, we introduce a vector database for efficient embedding storage and retrieval. Our experiments show that SLAG achieves an 18 times speedup in embedding computation on a 16-GPU setup compared to OpenGaussian, while preserving embedding quality on the ScanNet and LERF datasets. For more details, visit our project website: https://slag-project.github.io/.", "date": "2025-05-12", "categories": []}
{"id": "2508.10935", "title": "HQ-OV3D: A High Box Quality Open-World 3D Detection Framework based on Diffision Model", "abstract": "Traditional closed-set 3D detection frameworks fail to meet the demands of open-world applications like autonomous driving. Existing open-vocabulary 3D detection methods typically adopt a two-stage pipeline consisting of pseudo-label generation followed by semantic alignment. While vision-language models (VLMs) recently have dramatically improved the semantic accuracy of pseudo-labels, their geometric quality, particularly bounding box precision, remains commonly neglected. To address this issue, we propose a High Box Quality Open-Vocabulary 3D Detection (HQ-OV3D) framework, dedicated to generate and refine high-quality pseudo-labels for open-vocabulary classes. The framework comprises two key components: an Intra-Modality Cross-Validated (IMCV) Proposal Generator that utilizes cross-modality geometric consistency to generate high-quality initial 3D proposals, and an Annotated-Class Assisted (ACA) Denoiser that progressively refines 3D proposals by leveraging geometric priors from annotated categories through a DDIM-based denoising mechanism. Compared to the state-of-the-art method, training with pseudo-labels generated by our approach achieves a 7.37% improvement in mAP on novel classes, demonstrating the superior quality of the pseudo-labels produced by our framework. HQ-OV3D can serve not only as a strong standalone open-vocabulary 3D detector but also as a plug-in high-quality pseudo-label generator for existing open-vocabulary detection or annotation pipelines.", "date": "2025-08-12", "categories": []}
{"id": "2505.04058", "title": "LSVG: Language-Guided Scene Graphs with 2D-Assisted Multi-Modal Encoding for 3D Visual Grounding", "abstract": "3D visual grounding aims to localize the unique target described by natural languages in 3D scenes. The significant gap between 3D and language modalities makes it a notable challenge to distinguish multiple similar objects through the described spatial relationships. Current methods attempt to achieve cross-modal understanding in complex scenes via a target-centered learning mechanism, ignoring the modeling of referred objects. We propose a novel 3D visual grounding framework that constructs language-guided scene graphs with referred object discrimination to improve relational perception. The framework incorporates a dual-branch visual encoder that leverages pre-trained 2D semantics to enhance and supervise the multi-modal 3D encoding. Furthermore, we employ graph attention to promote relationship-oriented information fusion in cross-modal interaction. The learned object representations and scene graph structure enable effective alignment between 3D visual content and textual descriptions. Experimental results on popular benchmarks demonstrate our superior performance compared to state-of-the-art methods, especially in handling the challenges of multiple similar distractors.", "date": "2025-05-07", "categories": []}
{"id": "2508.09626", "title": "Semantic-aware DropSplat: Adaptive Pruning of Redundant Gaussians for 3D Aerial-View Segmentation", "abstract": "In the task of 3D Aerial-view Scene Semantic Segmentation (3D-AVS-SS), traditional methods struggle to address semantic ambiguity caused by scale variations and structural occlusions in aerial images. This limits their segmentation accuracy and consistency. To tackle these challenges, we propose a novel 3D-AVS-SS approach named SAD-Splat. Our method introduces a Gaussian point drop module, which integrates semantic confidence estimation with a learnable sparsity mechanism based on the Hard Concrete distribution. This module effectively eliminates redundant and semantically ambiguous Gaussian points, enhancing both segmentation performance and representation compactness. Furthermore, SAD-Splat incorporates a high-confidence pseudo-label generation pipeline. It leverages 2D foundation models to enhance supervision when ground-truth labels are limited, thereby further improving segmentation accuracy. To advance research in this domain, we introduce a challenging benchmark dataset: 3D Aerial Semantic (3D-AS), which encompasses diverse real-world aerial scenes with sparse annotations. Experimental results demonstrate that SAD-Splat achieves an excellent balance between segmentation accuracy and representation compactness. It offers an efficient and scalable solution for 3D aerial scene understanding.", "date": "2025-08-13", "categories": []}
{"id": "2506.18785", "title": "SWA-SOP: Spatially-aware Window Attention for Semantic Occupancy Prediction in Autonomous Driving", "abstract": "Perception systems in autonomous driving rely on sensors such as LiDAR and cameras to perceive the 3D environment. However, due to occlusions and data sparsity, these sensors often fail to capture complete information. Semantic Occupancy Prediction (SOP) addresses this challenge by inferring both occupancy and semantics of unobserved regions. Existing transformer-based SOP methods lack explicit modeling of spatial structure in attention computation, resulting in limited geometric awareness and poor performance in sparse or occluded areas. To this end, we propose Spatially-aware Window Attention (SWA), a novel mechanism that incorporates local spatial context into attention. SWA significantly improves scene completion and achieves state-of-the-art results on LiDAR-based SOP benchmarks. We further validate its generality by integrating SWA into a camera-based SOP pipeline, where it also yields consistent gains across modalities.", "date": "2025-06-23", "categories": []}
{"id": "2408.11747", "title": "OE3DIS: Open-Ended 3D Point Cloud Instance Segmentation", "abstract": "Open-Vocab 3D Instance Segmentation methods (OV-3DIS) have recently demonstrated their ability to generalize to unseen objects. However, these methods still depend on predefined class names during testing, restricting the autonomy of agents. To mitigate this constraint, we propose a novel problem termed Open-Ended 3D Instance Segmentation (OE-3DIS), which eliminates the necessity for predefined class names during testing. Moreover, we contribute a comprehensive set of strong baselines, derived from OV-3DIS approaches and leveraging 2D Multimodal Large Language Models. To assess the performance of our OE-3DIS system, we introduce a novel Open-Ended score, evaluating both the semantic and geometric quality of predicted masks and their associated class names, alongside the standard AP score. Our approach demonstrates significant performance improvements over the baselines on the ScanNet200 and ScanNet++ datasets. Remarkably, our method surpasses the performance of Open3DIS, the current state-of-the-art method in OV-3DIS, even in the absence of ground-truth object class names.", "date": "2024-08-21", "categories": []}
{"id": "2411.14594", "title": "Solving Zero-Shot 3D Visual Grounding as Constraint Satisfaction Problems", "abstract": "3D visual grounding (3DVG) aims to locate objects in a 3D scene with natural language descriptions. Supervised methods have achieved decent accuracy, but have a closed vocabulary and limited language understanding ability. Zero-shot methods utilize large language models (LLMs) to handle natural language descriptions, where the LLM either produces grounding results directly or generates programs that compute results (symbolically). In this work, we propose a zero-shot method that reformulates the 3DVG task as a Constraint Satisfaction Problem (CSP), where the variables and constraints represent objects and their spatial relations, respectively. This allows a global symbolic reasoning of all relevant objects, producing grounding results of both the target and anchor objects. Moreover, we demonstrate the flexibility of our framework by handling negation- and counting-based queries with only minor extra coding efforts. Our system, Constraint Satisfaction Visual Grounding (CSVG), has been extensively evaluated on the public datasets ScanRefer and Nr3D datasets using only open-source LLMs. Results show the effectiveness of CSVG and superior grounding accuracy over current state-of-the-art zero-shot 3DVG methods with improvements of $+7.0\\%$ (Acc@0.5 score) and $+11.2\\%$ on the ScanRefer and Nr3D datasets, respectively. The code of our system is available at https://asig-x.github.io/csvg_web.", "date": "2024-11-21", "categories": []}
{"id": "2503.07909", "title": "FunGraph: Functionality Aware 3D Scene Graphs for Language-Prompted Scene Interaction", "abstract": "The concept of 3D scene graphs is increasingly recognized as a powerful semantic and hierarchical representation of the environment. Current approaches often address this at a coarse, object-level resolution. In contrast, our goal is to develop a representation that enables robots to directly interact with their environment by identifying both the location of functional interactive elements and how these can be used. To achieve this, we focus on detecting and storing objects at a finer resolution, focusing on affordance-relevant parts. The primary challenge lies in the scarcity of data that extends beyond instance-level detection and the inherent difficulty of capturing detailed object features using robotic sensors. We leverage currently available 3D resources to generate 2D data and train a detector, which is then used to augment the standard 3D scene graph generation pipeline. Through our experiments, we demonstrate that our approach achieves functional element segmentation comparable to state-of-the-art 3D models and that our augmentation enables task-driven affordance grounding with higher accuracy than the current solutions. See our project page at https://fungraph.github.io.", "date": "2025-03-10", "categories": []}
{"id": "2507.19993", "title": "FROSS: Faster-than-Real-Time Online 3D Semantic Scene Graph Generation from RGB-D Images", "abstract": "The ability to abstract complex 3D environments into simplified and structured representations is crucial across various domains. 3D semantic scene graphs (SSGs) achieve this by representing objects as nodes and their interrelationships as edges, facilitating high-level scene understanding. Existing methods for 3D SSG generation, however, face significant challenges, including high computational demands and non-incremental processing that hinder their suitability for real-time open-world applications. To address this issue, we propose FROSS (Faster-than-Real-Time Online 3D Semantic Scene Graph Generation), an innovative approach for online and faster-than-real-time 3D SSG generation that leverages the direct lifting of 2D scene graphs to 3D space and represents objects as 3D Gaussian distributions. This framework eliminates the dependency on precise and computationally-intensive point cloud processing. Furthermore, we extend the Replica dataset with inter-object relationship annotations, creating the ReplicaSSG dataset for comprehensive evaluation of FROSS. The experimental results from evaluations on ReplicaSSG and 3DSSG datasets show that FROSS can achieve superior performance while operating significantly faster than prior 3D SSG generation methods. Our implementation and dataset are publicly available at https://github.com/Howardkhh/FROSS.", "date": "2025-07-26", "categories": []}
{"id": "2508.06546", "title": "Statistical Confidence Rescoring for Robust 3D Scene Graph Generation from Multi-View Images", "abstract": "Modern 3D semantic scene graph estimation methods utilize ground truth 3D annotations to accurately predict target objects, predicates, and relationships. In the absence of given 3D ground truth representations, we explore leveraging only multi-view RGB images to tackle this task. To attain robust features for accurate scene graph estimation, we must overcome the noisy reconstructed pseudo point-based geometry from predicted depth maps and reduce the amount of background noise present in multi-view image features. The key is to enrich node and edge features with accurate semantic and spatial information and through neighboring relations. We obtain semantic masks to guide feature aggregation to filter background features and design a novel method to incorporate neighboring node information to aid robustness of our scene graph estimates. Furthermore, we leverage on explicit statistical priors calculated from the training summary statistics to refine node and edge predictions based on their one-hop neighborhood. Our experiments show that our method outperforms current methods purely using multi-view images as the initial input. Our project page is available at https://qixun1.github.io/projects/SCRSSG.", "date": "2025-08-05", "categories": []}
{"id": "2502.00342", "title": "Embodied Intelligence for 3D Understanding: A Survey on 3D Scene Question Answering", "abstract": "3D Scene Question Answering (3D SQA) represents an interdisciplinary task that integrates 3D visual perception and natural language processing, empowering intelligent agents to comprehend and interact with complex 3D environments. Recent advances in large multimodal modelling have driven the creation of diverse datasets and spurred the development of instruction-tuning and zero-shot methods for 3D SQA. However, this rapid progress introduces challenges, particularly in achieving unified analysis and comparison across datasets and baselines. In this survey, we provide the first comprehensive and systematic review of 3D SQA. We organize existing work from three perspectives: datasets, methodologies, and evaluation metrics. Beyond basic categorization, we identify shared architectural patterns across methods. Our survey further synthesizes core limitations and discusses how current trends, such as instruction tuning, multimodal alignment, and zero-shot, can shape future developments. Finally, we propose a range of promising research directions covering dataset construction, task generalization, interaction modeling, and unified evaluation protocols. This work aims to serve as a foundation for future research and foster progress toward more generalizable and intelligent 3D SQA systems.", "date": "2025-02-01", "categories": []}
{"id": "2506.01109", "title": "CountingFruit: Language-Guided 3D Fruit Counting with Semantic Gaussian Splatting", "abstract": "Accurate 3D fruit counting in orchards is challenging due to heavy occlusion, semantic ambiguity between fruits and surrounding structures, and the high computational cost of volumetric reconstruction. Existing pipelines often rely on multi-view 2D segmentation and dense volumetric sampling, which lead to accumulated fusion errors and slow inference. We introduce FruitLangGS, a language-guided 3D fruit counting framework that reconstructs orchard-scale scenes using an adaptive-density Gaussian Splatting pipeline with radius-aware pruning and tile-based rasterization, enabling scalable 3D representation. During inference, compressed CLIP-aligned semantic vectors embedded in each Gaussian are filtered via a dual-threshold cosine similarity mechanism, retrieving Gaussians relevant to target prompts while suppressing common distractors (e.g., foliage), without requiring retraining or image-space masks. The selected Gaussians are then sampled into dense point clouds and clustered geometrically to estimate fruit instances, remaining robust under severe occlusion and viewpoint variation. Experiments on nine different orchard-scale datasets demonstrate that FruitLangGS consistently outperforms existing pipelines in instance counting recall, avoiding multi-view segmentation fusion errors and achieving up to 99.7% recall on Pfuji-Size_Orch2018 orchard dataset. Ablation studies further confirm that language-conditioned semantic embedding and dual-threshold prompt filtering are essential for suppressing distractors and improving counting accuracy under heavy occlusion. Beyond fruit counting, the same framework enables prompt-driven 3D semantic retrieval without retraining, highlighting the potential of language-guided 3D perception for scalable agricultural scene understanding.", "date": "2025-06-01", "categories": []}
{"id": "2412.18450", "title": "3DGraphLLM: Combining Semantic Graphs and Large Language Models for 3D Scene Understanding", "abstract": "A 3D scene graph represents a compact scene model by capturing both the objects present and the semantic relationships between them, making it a promising structure for robotic applications. To effectively interact with users, an embodied intelligent agent should be able to answer a wide range of natural language queries about the surrounding 3D environment. Large Language Models (LLMs) are beneficial solutions for user-robot interaction due to their natural language understanding and reasoning abilities. Recent methods for learning scene representations have shown that adapting these representations to the 3D world can significantly improve the quality of LLM responses. However, existing methods typically rely only on geometric information, such as object coordinates, and overlook the rich semantic relationships between objects. In this work, we propose 3DGraphLLM, a method for constructing a learnable representation of a 3D scene graph that explicitly incorporates semantic relationships. This representation is used as input to LLMs for performing 3D vision-language tasks. In our experiments on popular ScanRefer, Multi3DRefer, ScanQA, Sqa3D, and Scan2cap datasets, we demonstrate that our approach outperforms baselines that do not leverage semantic relationships between objects. The code is publicly available at https://github.com/CognitiveAISystems/3DGraphLLM.", "date": "2024-12-24", "categories": []}
{"id": "2508.04659", "title": "PixCuboid: Room Layout Estimation from Multi-view Featuremetric Alignment", "abstract": "Coarse room layout estimation provides important geometric cues for many downstream tasks. Current state-of-the-art methods are predominantly based on single views and often assume panoramic images. We introduce PixCuboid, an optimization-based approach for cuboid-shaped room layout estimation, which is based on multi-view alignment of dense deep features. By training with the optimization end-to-end, we learn feature maps that yield large convergence basins and smooth loss landscapes in the alignment. This allows us to initialize the room layout using simple heuristics.\n  For the evaluation we propose two new benchmarks based on ScanNet++ and 2D-3D-Semantics, with manually verified ground truth 3D cuboids. In thorough experiments we validate our approach and significantly outperform the competition. Finally, while our network is trained with single cuboids, the flexibility of the optimization-based approach allow us to easily extend to multi-room estimation, e.g. larger apartments or offices. Code and model weights are available at https://github.com/ghanning/PixCuboid.", "date": "2025-08-06", "categories": []}
{"id": "2508.03227", "title": "Trace3D: Consistent Segmentation Lifting via Gaussian Instance Tracing", "abstract": "We address the challenge of lifting 2D visual segmentation to 3D in Gaussian Splatting. Existing methods often suffer from inconsistent 2D masks across viewpoints and produce noisy segmentation boundaries as they neglect these semantic cues to refine the learned Gaussians. To overcome this, we introduce Gaussian Instance Tracing (GIT), which augments the standard Gaussian representation with an instance weight matrix across input views. Leveraging the inherent consistency of Gaussians in 3D, we use this matrix to identify and correct 2D segmentation inconsistencies. Furthermore, since each Gaussian ideally corresponds to a single object, we propose a GIT-guided adaptive density control mechanism to split and prune ambiguous Gaussians during training, resulting in sharper and more coherent 2D and 3D segmentation boundaries. Experimental results show that our method extracts clean 3D assets and consistently improves 3D segmentation in both online (e.g., self-prompting) and offline (e.g., contrastive lifting) settings, enabling applications such as hierarchical segmentation, object extraction, and scene editing.", "date": "2025-08-05", "categories": []}
{"id": "2406.00830", "title": "Collaborative Novel Object Discovery and Box-Guided Cross-Modal Alignment for Open-Vocabulary 3D Object Detection", "abstract": "Open-vocabulary 3D Object Detection (OV-3DDet) addresses the detection of objects from an arbitrary list of novel categories in 3D scenes, which remains a very challenging problem. In this work, we propose CoDAv2, a unified framework designed to innovatively tackle both the localization and classification of novel 3D objects, under the condition of limited base categories. For localization, the proposed 3D Novel Object Discovery (3D-NOD) strategy utilizes 3D geometries and 2D open-vocabulary semantic priors to discover pseudo labels for novel objects during training. 3D-NOD is further extended with an Enrichment strategy that significantly enriches the novel object distribution in the training scenes, and then enhances the model's ability to localize more novel objects. The 3D-NOD with Enrichment is termed 3D-NODE. For classification, the Discovery-driven Cross-modal Alignment (DCMA) module aligns features from 3D point clouds and 2D/textual modalities, employing both class-agnostic and class-specific alignments that are iteratively refined to handle the expanding vocabulary of objects. Besides, 2D box guidance boosts the classification accuracy against complex background noises, which is coined as Box-DCMA. Extensive evaluation demonstrates the superiority of CoDAv2. CoDAv2 outperforms the best-performing method by a large margin (AP_Novel of 9.17 vs. 3.61 on SUN-RGBD and 9.12 vs. 3.74 on ScanNetv2). Source code and pre-trained models are available at the GitHub project page.", "date": "2024-06-02", "categories": []}
{"id": "2409.18092", "title": "DiffSSC: Semantic LiDAR Scan Completion using Denoising Diffusion Probabilistic Models", "abstract": "Perception systems play a crucial role in autonomous driving, incorporating multiple sensors and corresponding computer vision algorithms. 3D LiDAR sensors are widely used to capture sparse point clouds of the vehicle's surroundings. However, such systems struggle to perceive occluded areas and gaps in the scene due to the sparsity of these point clouds and their lack of semantics. To address these challenges, Semantic Scene Completion (SSC) jointly predicts unobserved geometry and semantics in the scene given raw LiDAR measurements, aiming for a more complete scene representation. Building on promising results of diffusion models in image generation and super-resolution tasks, we propose their extension to SSC by implementing the noising and denoising diffusion processes in the point and semantic spaces individually. To control the generation, we employ semantic LiDAR point clouds as conditional input and design local and global regularization losses to stabilize the denoising process. We evaluate our approach on autonomous driving datasets, and it achieves state-of-the-art performance for SSC, surpassing most existing methods.", "date": "2024-09-26", "categories": []}
{"id": "2411.13927", "title": "Multimodal 3D Reasoning Segmentation with Complex Scenes", "abstract": "The recent development in multimodal learning has greatly advanced the research in 3D scene understanding in various real-world tasks such as embodied AI. However, most existing studies are facing two common challenges: 1) they are short of reasoning ability for interaction and interpretation of human intentions and 2) they focus on scenarios with single-category objects and over-simplified textual descriptions and neglect multi-object scenarios with complicated spatial relations among objects. We address the above challenges by proposing a 3D reasoning segmentation task for reasoning segmentation with multiple objects in scenes. The task allows producing 3D segmentation masks and detailed textual explanations as enriched by 3D spatial relations among objects. To this end, we create ReasonSeg3D, a large-scale and high-quality benchmark that integrates 3D segmentation masks and 3D spatial relations with generated question-answer pairs. In addition, we design MORE3D, a novel 3D reasoning network that works with queries of multiple objects and is tailored for 3D scene understanding. MORE3D learns detailed explanations on 3D relations and employs them to capture spatial information of objects and reason textual outputs. Extensive experiments show that MORE3D excels in reasoning and segmenting complex multi-object 3D scenes. In addition, the created ReasonSeg3D offers a valuable platform for future exploration of 3D reasoning segmentation. The data and code will be released.", "date": "2024-11-21", "categories": []}
{"id": "2502.04981", "title": "AutoOcc: Automatic Open-Ended Semantic Occupancy Annotation via Vision-Language Guided Gaussian Splatting", "abstract": "Obtaining high-quality 3D semantic occupancy from raw sensor data remains an essential yet challenging task, often requiring extensive manual labeling. In this work, we propose AutoOcc, a vision-centric automated pipeline for open-ended semantic occupancy annotation that integrates differentiable Gaussian splatting guided by vision-language models. We formulate the open-ended semantic 3D occupancy reconstruction task to automatically generate scene occupancy by combining attention maps from vision-language models and foundation vision models. We devise semantic-aware Gaussians as intermediate geometric descriptors and propose a cumulative Gaussian-to-voxel splatting algorithm that enables effective and efficient occupancy annotation. Our framework outperforms existing automated occupancy annotation methods without human labels. AutoOcc also enables open-ended semantic occupancy auto-labeling, achieving robust performance in both static and dynamically complex scenarios.", "date": "2025-02-07", "categories": []}
{"id": "2507.21045", "title": "Reconstructing 4D Spatial Intelligence: A Survey", "abstract": "Reconstructing 4D spatial intelligence from visual observations has long been a central yet challenging task in computer vision, with broad real-world applications. These range from entertainment domains like movies, where the focus is often on reconstructing fundamental visual elements, to embodied AI, which emphasizes interaction modeling and physical realism. Fueled by rapid advances in 3D representations and deep learning architectures, the field has evolved quickly, outpacing the scope of previous surveys. Additionally, existing surveys rarely offer a comprehensive analysis of the hierarchical structure of 4D scene reconstruction. To address this gap, we present a new perspective that organizes existing methods into five progressive levels of 4D spatial intelligence: (1) Level 1 -- reconstruction of low-level 3D attributes (e.g., depth, pose, and point maps); (2) Level 2 -- reconstruction of 3D scene components (e.g., objects, humans, structures); (3) Level 3 -- reconstruction of 4D dynamic scenes; (4) Level 4 -- modeling of interactions among scene components; and (5) Level 5 -- incorporation of physical laws and constraints. We conclude the survey by discussing the key challenges at each level and highlighting promising directions for advancing toward even richer levels of 4D spatial intelligence. To track ongoing developments, we maintain an up-to-date project page: https://github.com/yukangcao/Awesome-4D-Spatial-Intelligence.", "date": "2025-07-28", "categories": []}
{"id": "2508.01150", "title": "OpenGS-Fusion: Open-Vocabulary Dense Mapping with Hybrid 3D Gaussian Splatting for Refined Object-Level Understanding", "abstract": "Recent advancements in 3D scene understanding have made significant strides in enabling interaction with scenes using open-vocabulary queries, particularly for VR/AR and robotic applications. Nevertheless, existing methods are hindered by rigid offline pipelines and the inability to provide precise 3D object-level understanding given open-ended queries. In this paper, we present OpenGS-Fusion, an innovative open-vocabulary dense mapping framework that improves semantic modeling and refines object-level understanding. OpenGS-Fusion combines 3D Gaussian representation with a Truncated Signed Distance Field to facilitate lossless fusion of semantic features on-the-fly. Furthermore, we introduce a novel multimodal language-guided approach named MLLM-Assisted Adaptive Thresholding, which refines the segmentation of 3D objects by adaptively adjusting similarity thresholds, achieving an improvement 17\\% in 3D mIoU compared to the fixed threshold strategy. Extensive experiments demonstrate that our method outperforms existing methods in 3D object understanding and scene reconstruction quality, as well as showcasing its effectiveness in language-guided scene interaction. The code is available at https://young-bit.github.io/opengs-fusion.github.io/ .", "date": "2025-08-02", "categories": []}
{"id": "2508.01740", "title": "AG$^2$aussian: Anchor-Graph Structured Gaussian Splatting for Instance-Level 3D Scene Understanding and Editing", "abstract": "3D Gaussian Splatting (3DGS) has witnessed exponential adoption across diverse applications, driving a critical need for semantic-aware 3D Gaussian representations to enable scene understanding and editing tasks. Existing approaches typically attach semantic features to a collection of free Gaussians and distill the features via differentiable rendering, leading to noisy segmentation and a messy selection of Gaussians. In this paper, we introduce AG$^2$aussian, a novel framework that leverages an anchor-graph structure to organize semantic features and regulate Gaussian primitives. Our anchor-graph structure not only promotes compact and instance-aware Gaussian distributions, but also facilitates graph-based propagation, achieving a clean and accurate instance-level Gaussian selection. Extensive validation across four applications, i.e. interactive click-based query, open-vocabulary text-driven query, object removal editing, and physics simulation, demonstrates the advantages of our approach and its benefits to various applications. The experiments and ablation studies further evaluate the effectiveness of the key designs of our approach.", "date": "2025-08-03", "categories": []}
{"id": "2503.11806", "title": "Human-in-the-Loop Local Corrections of 3D Scene Layouts via Infilling", "abstract": "We present a novel human-in-the-loop approach to estimate 3D scene layout that uses human feedback from an egocentric standpoint. We study this approach through introduction of a novel local correction task, where users identify local errors and prompt a model to automatically correct them. Building on SceneScript, a state-of-the-art framework for 3D scene layout estimation that leverages structured language, we propose a solution that structures this problem as \"infilling\", a task studied in natural language processing. We train a multi-task version of SceneScript that maintains performance on global predictions while significantly improving its local correction ability. We integrate this into a human-in-the-loop system, enabling a user to iteratively refine scene layout estimates via a low-friction \"one-click fix'' workflow. Our system enables the final refined layout to diverge from the training distribution, allowing for more accurate modelling of complex layouts.", "date": "2025-03-14", "categories": []}
{"id": "2503.15897", "title": "Learning 3D Scene Analogies with Neural Contextual Scene Maps", "abstract": "Understanding scene contexts is crucial for machines to perform tasks and adapt prior knowledge in unseen or noisy 3D environments. As data-driven learning is intractable to comprehensively encapsulate diverse ranges of layouts and open spaces, we propose teaching machines to identify relational commonalities in 3D spaces. Instead of focusing on point-wise or object-wise representations, we introduce 3D scene analogies, which are smooth maps between 3D scene regions that align spatial relationships. Unlike well-studied single instance-level maps, these scene-level maps smoothly link large scene regions, potentially enabling unique applications in trajectory transfer in AR/VR, long demonstration transfer for imitation learning, and context-aware object rearrangement. To find 3D scene analogies, we propose neural contextual scene maps, which extract descriptor fields summarizing semantic and geometric contexts, and holistically align them in a coarse-to-fine manner for map estimation. This approach reduces reliance on individual feature points, making it robust to input noise or shape variations. Experiments demonstrate the effectiveness of our approach in identifying scene analogies and transferring trajectories or object placements in diverse indoor scenes, indicating its potential for robotics and AR/VR applications. Project page including the code is available through this link: https://82magnolia.github.io/3d_scene_analogies/.", "date": "2025-03-20", "categories": []}
{"id": "2503.17856", "title": "ClaraVid: A Holistic Scene Reconstruction Benchmark From Aerial Perspective With Delentropy-Based Complexity Profiling", "abstract": "The development of aerial holistic scene understanding algorithms is hindered by the scarcity of comprehensive datasets that enable both semantic and geometric reconstruction. While synthetic datasets offer an alternative, existing options exhibit task-specific limitations, unrealistic scene compositions, and rendering artifacts that compromise real-world applicability. We introduce ClaraVid, a synthetic aerial dataset specifically designed to overcome these limitations. Comprising 16,917 high-resolution images captured at 4032x3024 from multiple viewpoints across diverse landscapes, ClaraVid provides dense depth maps, panoptic segmentation, sparse point clouds, and dynamic object masks, while mitigating common rendering artifacts. To further advance neural reconstruction, we introduce the Delentropic Scene Profile (DSP), a novel complexity metric derived from differential entropy analysis, designed to quantitatively assess scene difficulty and inform reconstruction tasks. Utilizing DSP, we systematically benchmark neural reconstruction methods, uncovering a consistent, measurable correlation between scene complexity and reconstruction accuracy. Empirical results indicate that higher delentropy strongly correlates with increased reconstruction errors, validating DSP as a reliable complexity prior. The data and code are available on the project page at https://rdbch.github.io/claravid/", "date": "2025-03-22", "categories": []}
{"id": "2507.23134", "title": "Details Matter for Indoor Open-vocabulary 3D Instance Segmentation", "abstract": "Unlike closed-vocabulary 3D instance segmentation that is often trained end-to-end, open-vocabulary 3D instance segmentation (OV-3DIS) often leverages vision-language models (VLMs) to generate 3D instance proposals and classify them. While various concepts have been proposed from existing research, we observe that these individual concepts are not mutually exclusive but complementary. In this paper, we propose a new state-of-the-art solution for OV-3DIS by carefully designing a recipe to combine the concepts together and refining them to address key challenges. Our solution follows the two-stage scheme: 3D proposal generation and instance classification. We employ robust 3D tracking-based proposal aggregation to generate 3D proposals and remove overlapped or partial proposals by iterative merging/removal. For the classification stage, we replace the standard CLIP model with Alpha-CLIP, which incorporates object masks as an alpha channel to reduce background noise and obtain object-centric representation. Additionally, we introduce the standardized maximum similarity (SMS) score to normalize text-to-proposal similarity, effectively filtering out false positives and boosting precision. Our framework achieves state-of-the-art performance on ScanNet200 and S3DIS across all AP and AR metrics, even surpassing an end-to-end closed-vocabulary method.", "date": "2025-07-30", "categories": []}
{"id": "2411.16072", "title": "Language Driven Occupancy Prediction", "abstract": "We introduce LOcc, an effective and generalizable framework for open-vocabulary occupancy (OVO) prediction. Previous approaches typically supervise the networks through coarse voxel-to-text correspondences via image features as intermediates or noisy and sparse correspondences from voxel-based model-view projections. To alleviate the inaccurate supervision, we propose a semantic transitive labeling pipeline to generate dense and fine-grained 3D language occupancy ground truth. Our pipeline presents a feasible way to dig into the valuable semantic information of images, transferring text labels from images to LiDAR point clouds and ultimately to voxels, to establish precise voxel-to-text correspondences. By replacing the original prediction head of supervised occupancy models with a geometry head for binary occupancy states and a language head for language features, LOcc effectively uses the generated language ground truth to guide the learning of 3D language volume. Through extensive experiments, we demonstrate that our transitive semantic labeling pipeline can produce more accurate pseudo-labeled ground truth, diminishing labor-intensive human annotations. Additionally, we validate LOcc across various architectures, where all models consistently outperform state-of-the-art zero-shot occupancy prediction approaches on the Occ3D-nuScenes dataset.", "date": "2024-11-25", "categories": []}
{"id": "2507.22668", "title": "Graph-Guided Dual-Level Augmentation for 3D Scene Segmentation", "abstract": "3D point cloud segmentation aims to assign semantic labels to individual points in a scene for fine-grained spatial understanding. Existing methods typically adopt data augmentation to alleviate the burden of large-scale annotation. However, most augmentation strategies only focus on local transformations or semantic recomposition, lacking the consideration of global structural dependencies within scenes. To address this limitation, we propose a graph-guided data augmentation framework with dual-level constraints for realistic 3D scene synthesis. Our method learns object relationship statistics from real-world data to construct guiding graphs for scene generation. Local-level constraints enforce geometric plausibility and semantic consistency between objects, while global-level constraints maintain the topological structure of the scene by aligning the generated layout with the guiding graph. Extensive experiments on indoor and outdoor datasets demonstrate that our framework generates diverse and high-quality augmented scenes, leading to consistent improvements in point cloud segmentation performance across various models.", "date": "2025-07-30", "categories": []}
{"id": "2507.22825", "title": "DepR: Depth Guided Single-view Scene Reconstruction with Instance-level Diffusion", "abstract": "We propose DepR, a depth-guided single-view scene reconstruction framework that integrates instance-level diffusion within a compositional paradigm. Instead of reconstructing the entire scene holistically, DepR generates individual objects and subsequently composes them into a coherent 3D layout. Unlike previous methods that use depth solely for object layout estimation during inference and therefore fail to fully exploit its rich geometric information, DepR leverages depth throughout both training and inference. Specifically, we introduce depth-guided conditioning to effectively encode shape priors into diffusion models. During inference, depth further guides DDIM sampling and layout optimization, enhancing alignment between the reconstruction and the input image. Despite being trained on limited synthetic data, DepR achieves state-of-the-art performance and demonstrates strong generalization in single-view scene reconstruction, as shown through evaluations on both synthetic and real-world datasets.", "date": "2025-07-30", "categories": []}
{"id": "2506.13629", "title": "FreeQ-Graph: Free-form Querying with Semantic Consistent Scene Graph for 3D Scene Understanding", "abstract": "Semantic querying in complex 3D scenes through free-form language presents a significant challenge. Existing 3D scene understanding methods use large-scale training data and CLIP to align text queries with 3D semantic features. However, their reliance on predefined vocabulary priors from training data hinders free-form semantic querying. Besides, recent advanced methods rely on LLMs for scene understanding but lack comprehensive 3D scene-level information and often overlook the potential inconsistencies in LLM-generated outputs. In our paper, we propose FreeQ-Graph, which enables Free-form Querying with a semantic consistent scene Graph for 3D scene understanding. The core idea is to encode free-form queries from a complete and accurate 3D scene graph without predefined vocabularies, and to align them with 3D consistent semantic labels, which accomplished through three key steps. We initiate by constructing a complete and accurate 3D scene graph that maps free-form objects and their relations through LLM and LVLM guidance, entirely free from training data or predefined priors. Most importantly, we align graph nodes with accurate semantic labels by leveraging 3D semantic aligned features from merged superpoints, enhancing 3D semantic consistency. To enable free-form semantic querying, we then design an LLM-based reasoning algorithm that combines scene-level and object-level information to intricate reasoning. We conducted extensive experiments on 3D semantic grounding, segmentation, and complex querying tasks, while also validating the accuracy of graph generation. Experiments on 6 datasets show that our model excels in both complex free-form semantic queries and intricate relational reasoning.", "date": "2025-06-16", "categories": []}
{"id": "2504.09540", "title": "EmbodiedOcc++: Boosting Embodied 3D Occupancy Prediction with Plane Regularization and Uncertainty Sampler", "abstract": "Online 3D occupancy prediction provides a comprehensive spatial understanding of embodied environments. While the innovative EmbodiedOcc framework utilizes 3D semantic Gaussians for progressive indoor occupancy prediction, it overlooks the geometric characteristics of indoor environments, which are primarily characterized by planar structures. This paper introduces EmbodiedOcc++, enhancing the original framework with two key innovations: a Geometry-guided Refinement Module (GRM) that constrains Gaussian updates through plane regularization, along with a Semantic-aware Uncertainty Sampler (SUS) that enables more effective updates in overlapping regions between consecutive frames. GRM regularizes the position update to align with surface normals. It determines the adaptive regularization weight using curvature-based and depth-based constraints, allowing semantic Gaussians to align accurately with planar surfaces while adapting in complex regions. To effectively improve geometric consistency from different views, SUS adaptively selects proper Gaussians to update. Comprehensive experiments on the EmbodiedOcc-ScanNet benchmark demonstrate that EmbodiedOcc++ achieves state-of-the-art performance across different settings. Our method demonstrates improved edge accuracy and retains more geometric details while ensuring computational efficiency, which is essential for online embodied perception. The code will be released at: https://github.com/PKUHaoWang/EmbodiedOcc2.", "date": "2025-04-13", "categories": []}
{"id": "2507.05211", "title": "All in One: Visual-Description-Guided Unified Point Cloud Segmentation", "abstract": "Unified segmentation of 3D point clouds is crucial for scene understanding, but is hindered by its sparse structure, limited annotations, and the challenge of distinguishing fine-grained object classes in complex environments. Existing methods often struggle to capture rich semantic and contextual information due to limited supervision and a lack of diverse multimodal cues, leading to suboptimal differentiation of classes and instances. To address these challenges, we propose VDG-Uni3DSeg, a novel framework that integrates pre-trained vision-language models (e.g., CLIP) and large language models (LLMs) to enhance 3D segmentation. By leveraging LLM-generated textual descriptions and reference images from the internet, our method incorporates rich multimodal cues, facilitating fine-grained class and instance separation. We further design a Semantic-Visual Contrastive Loss to align point features with multimodal queries and a Spatial Enhanced Module to model scene-wide relationships efficiently. Operating within a closed-set paradigm that utilizes multimodal knowledge generated offline, VDG-Uni3DSeg achieves state-of-the-art results in semantic, instance, and panoptic segmentation, offering a scalable and practical solution for 3D understanding. Our code is available at https://github.com/Hanzy1996/VDG-Uni3DSeg.", "date": "2025-07-07", "categories": []}
{"id": "2507.06230", "title": "Feed-Forward SceneDINO for Unsupervised Semantic Scene Completion", "abstract": "Semantic scene completion (SSC) aims to infer both the 3D geometry and semantics of a scene from single images. In contrast to prior work on SSC that heavily relies on expensive ground-truth annotations, we approach SSC in an unsupervised setting. Our novel method, SceneDINO, adapts techniques from self-supervised representation learning and 2D unsupervised scene understanding to SSC. Our training exclusively utilizes multi-view consistency self-supervision without any form of semantic or geometric ground truth. Given a single input image, SceneDINO infers the 3D geometry and expressive 3D DINO features in a feed-forward manner. Through a novel 3D feature distillation approach, we obtain unsupervised 3D semantics. In both 3D and 2D unsupervised scene understanding, SceneDINO reaches state-of-the-art segmentation accuracy. Linear probing our 3D features matches the segmentation accuracy of a current supervised SSC approach. Additionally, we showcase the domain generalization and multi-view consistency of SceneDINO, taking the first steps towards a strong foundation for single image 3D scene understanding.", "date": "2025-07-08", "categories": []}
{"id": "2507.19188", "title": "VisHall3D: Monocular Semantic Scene Completion from Reconstructing the Visible Regions to Hallucinating the Invisible Regions", "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.\n  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.", "date": "2025-07-25", "categories": []}
{"id": "2507.17661", "title": "Monocular Semantic Scene Completion via Masked Recurrent Networks", "abstract": "Monocular Semantic Scene Completion (MSSC) aims to predict the voxel-wise occupancy and semantic category from a single-view RGB image. Existing methods adopt a single-stage framework that aims to simultaneously achieve visible region segmentation and occluded region hallucination, while also being affected by inaccurate depth estimation. Such methods often achieve suboptimal performance, especially in complex scenes. We propose a novel two-stage framework that decomposes MSSC into coarse MSSC followed by the Masked Recurrent Network. Specifically, we propose the Masked Sparse Gated Recurrent Unit (MS-GRU) which concentrates on the occupied regions by the proposed mask updating mechanism, and a sparse GRU design is proposed to reduce the computation cost. Additionally, we propose the distance attention projection to reduce projection errors by assigning different attention scores according to the distance to the observed surface. Experimental results demonstrate that our proposed unified framework, MonoMRN, effectively supports both indoor and outdoor scenes and achieves state-of-the-art performance on the NYUv2 and SemanticKITTI datasets. Furthermore, we conduct robustness analysis under various disturbances, highlighting the role of the Masked Recurrent Network in enhancing the model's resilience to such challenges. The source code is publicly available.", "date": "2025-07-23", "categories": []}
{"id": "2507.08416", "title": "InstaScene: Towards Complete 3D Instance Decomposition and Reconstruction from Cluttered Scenes", "abstract": "Humans can naturally identify and mentally complete occluded objects in cluttered environments. However, imparting similar cognitive ability to robotics remains challenging even with advanced reconstruction techniques, which models scenes as undifferentiated wholes and fails to recognize complete object from partial observations. In this paper, we propose InstaScene, a new paradigm towards holistic 3D perception of complex scenes with a primary goal: decomposing arbitrary instances while ensuring complete reconstruction. To achieve precise decomposition, we develop a novel spatial contrastive learning by tracing rasterization of each instance across views, significantly enhancing semantic supervision in cluttered scenes. To overcome incompleteness from limited observations, we introduce in-situ generation that harnesses valuable observations and geometric cues, effectively guiding 3D generative models to reconstruct complete instances that seamlessly align with the real world. Experiments on scene decomposition and object completion across complex real-world and synthetic scenes demonstrate that our method achieves superior decomposition accuracy while producing geometrically faithful and visually intact objects.", "date": "2025-07-11", "categories": []}
{"id": "2507.15454", "title": "ObjectGS: Object-aware Scene Reconstruction and Scene Understanding via Gaussian Splatting", "abstract": "3D Gaussian Splatting is renowned for its high-fidelity reconstructions and real-time novel view synthesis, yet its lack of semantic understanding limits object-level perception. In this work, we propose ObjectGS, an object-aware framework that unifies 3D scene reconstruction with semantic understanding. Instead of treating the scene as a unified whole, ObjectGS models individual objects as local anchors that generate neural Gaussians and share object IDs, enabling precise object-level reconstruction. During training, we dynamically grow or prune these anchors and optimize their features, while a one-hot ID encoding with a classification loss enforces clear semantic constraints. We show through extensive experiments that ObjectGS not only outperforms state-of-the-art methods on open-vocabulary and panoptic segmentation tasks, but also integrates seamlessly with applications like mesh extraction and scene editing. Project page: https://ruijiezhu94.github.io/ObjectGS_page", "date": "2025-07-21", "categories": []}
{"id": "2507.13801", "title": "One Step Closer: Creating the Future to Boost Monocular Semantic Scene Completion", "abstract": "In recent years, visual 3D Semantic Scene Completion (SSC) has emerged as a critical perception task for autonomous driving due to its ability to infer complete 3D scene layouts and semantics from single 2D images. However, in real-world traffic scenarios, a significant portion of the scene remains occluded or outside the camera's field of view -- a fundamental challenge that existing monocular SSC methods fail to address adequately. To overcome these limitations, we propose Creating the Future SSC (CF-SSC), a novel temporal SSC framework that leverages pseudo-future frame prediction to expand the model's effective perceptual range. Our approach combines poses and depths to establish accurate 3D correspondences, enabling geometrically-consistent fusion of past, present, and predicted future frames in 3D space. Unlike conventional methods that rely on simple feature stacking, our 3D-aware architecture achieves more robust scene completion by explicitly modeling spatial-temporal relationships. Comprehensive experiments on SemanticKITTI and SSCBench-KITTI-360 benchmarks demonstrate state-of-the-art performance, validating the effectiveness of our approach, highlighting our method's ability to improve occlusion reasoning and 3D scene completion accuracy.", "date": "2025-07-18", "categories": []}
{"id": "2507.12795", "title": "City-VLM: Towards Multidomain Perception Scene Understanding via Multimodal Incomplete Learning", "abstract": "Scene understanding enables intelligent agents to interpret and comprehend their environment. While existing large vision-language models (LVLMs) for scene understanding have primarily focused on indoor household tasks, they face two significant limitations when applied to outdoor large-scale scene understanding. First, outdoor scenarios typically encompass larger-scale environments observed through various sensors from multiple viewpoints (e.g., bird view and terrestrial view), while existing indoor LVLMs mainly analyze single visual modalities within building-scale contexts from humanoid viewpoints. Second, existing LVLMs suffer from missing multidomain perception outdoor data and struggle to effectively integrate 2D and 3D visual information. To address the aforementioned limitations, we build the first multidomain perception outdoor scene understanding dataset, named \\textbf{\\underline{SVM-City}}, deriving from multi\\textbf{\\underline{S}}cale scenarios with multi\\textbf{\\underline{V}}iew and multi\\textbf{\\underline{M}}odal instruction tuning data. It contains $420$k images and $4, 811$M point clouds with $567$k question-answering pairs from vehicles, low-altitude drones, high-altitude aerial planes, and satellite. To effectively fuse the multimodal data in the absence of one modality, we introduce incomplete multimodal learning to model outdoor scene understanding and design the LVLM named \\textbf{\\underline{City-VLM}}. Multimodal fusion is realized by constructing a joint probabilistic distribution space rather than implementing directly explicit fusion operations (e.g., concatenation). Experimental results on three typical outdoor scene understanding tasks show City-VLM achieves $18.14 \\%$ performance surpassing existing LVLMs in question-answering tasks averagely. Our method demonstrates pragmatic and generalization performance across multiple outdoor scenes.", "date": "2025-07-17", "categories": []}
{"id": "2507.12916", "title": "Argus: Leveraging Multiview Images for Improved 3-D Scene Understanding With Large Language Models", "abstract": "Advancements in foundation models have made it possible to conduct applications in various downstream tasks. Especially, the new era has witnessed a remarkable capability to extend Large Language Models (LLMs) for tackling tasks of 3D scene understanding. Current methods rely heavily on 3D point clouds, but the 3D point cloud reconstruction of an indoor scene often results in information loss. Some textureless planes or repetitive patterns are prone to omission and manifest as voids within the reconstructed 3D point clouds. Besides, objects with complex structures tend to introduce distortion of details caused by misalignments between the captured images and the dense reconstructed point clouds. 2D multi-view images present visual consistency with 3D point clouds and provide more detailed representations of scene components, which can naturally compensate for these deficiencies. Based on these insights, we propose Argus, a novel 3D multimodal framework that leverages multi-view images for enhanced 3D scene understanding with LLMs. In general, Argus can be treated as a 3D Large Multimodal Foundation Model (3D-LMM) since it takes various modalities as input(text instructions, 2D multi-view images, and 3D point clouds) and expands the capability of LLMs to tackle 3D tasks. Argus involves fusing and integrating multi-view images and camera poses into view-as-scene features, which interact with the 3D features to create comprehensive and detailed 3D-aware scene embeddings. Our approach compensates for the information loss while reconstructing 3D point clouds and helps LLMs better understand the 3D world. Extensive experiments demonstrate that our method outperforms existing 3D-LMMs in various downstream tasks.", "date": "2025-07-17", "categories": []}
{"id": "2403.06403", "title": "PointSeg: A Training-Free Paradigm for 3D Scene Segmentation via Foundation Models", "abstract": "Recent success of vision foundation models have shown promising performance for the 2D perception tasks. However, it is difficult to train a 3D foundation network directly due to the limited dataset and it remains under explored whether existing foundation models can be lifted to 3D space seamlessly. In this paper, we present PointSeg, a novel training-free paradigm that leverages off-the-shelf vision foundation models to address 3D scene perception tasks. PointSeg can segment anything in 3D scene by acquiring accurate 3D prompts to align their corresponding pixels across frames. Concretely, we design a two-branch prompts learning structure to construct the 3D point-box prompts pairs, combining with the bidirectional matching strategy for accurate point and proposal prompts generation. Then, we perform the iterative post-refinement adaptively when cooperated with different vision foundation models. Moreover, we design a affinity-aware merging algorithm to improve the final ensemble masks. PointSeg demonstrates impressive segmentation performance across various datasets, all without training. Specifically, our approach significantly surpasses the state-of-the-art specialist training-free model by 14.1$\\%$, 12.3$\\%$, and 12.6$\\%$ mAP on ScanNet, ScanNet++, and KITTI-360 datasets, respectively. On top of that, PointSeg can incorporate with various foundation models and even surpasses the specialist training-based methods by 3.4$\\%$-5.4$\\%$ mAP across various datasets, serving as an effective generalist model.", "date": "2024-03-11", "categories": []}
{"id": "2503.12955", "title": "HIS-GPT: Towards 3D Human-In-Scene Multimodal Understanding", "abstract": "We propose a new task to benchmark human-in-scene understanding for embodied agents: Human-In-Scene Question Answering (HIS-QA). Given a human motion within a 3D scene, HIS-QA requires the agent to comprehend human states and behaviors, reason about its surrounding environment, and answer human-related questions within the scene. To support this new task, we present HIS-Bench, a multimodal benchmark that systematically evaluates HIS understanding across a broad spectrum, from basic perception to commonsense reasoning and planning. Our evaluation of various vision-language models on HIS-Bench reveals significant limitations in their ability to handle HIS-QA tasks. To this end, we propose HIS-GPT, the first foundation model for HIS understanding. HIS-GPT integrates 3D scene context and human motion dynamics into large language models while incorporating specialized mechanisms to capture human-scene interactions. Extensive experiments demonstrate that HIS-GPT sets a new state-of-the-art on HIS-QA tasks. We hope this work inspires future research on human behavior analysis in 3D scenes, advancing embodied AI and world models. The codes and data: https://github.com/ZJHTerry18/HumanInScene.", "date": "2025-03-17", "categories": []}
{"id": "2507.12123", "title": "Open-Vocabulary Indoor Object Grounding with 3D Hierarchical Scene Graph", "abstract": "We propose OVIGo-3DHSG method - Open-Vocabulary Indoor Grounding of objects using 3D Hierarchical Scene Graph. OVIGo-3DHSG represents an extensive indoor environment over a Hierarchical Scene Graph derived from sequences of RGB-D frames utilizing a set of open-vocabulary foundation models and sensor data processing. The hierarchical representation explicitly models spatial relations across floors, rooms, locations, and objects. To effectively address complex queries involving spatial reference to other objects, we integrate the hierarchical scene graph with a Large Language Model for multistep reasoning. This integration leverages inter-layer (e.g., room-to-object) and intra-layer (e.g., object-to-object) connections, enhancing spatial contextual understanding. We investigate the semantic and geometry accuracy of hierarchical representation on Habitat Matterport 3D Semantic multi-floor scenes. Our approach demonstrates efficient scene comprehension and robust object grounding compared to existing methods. Overall OVIGo-3DHSG demonstrates strong potential for applications requiring spatial reasoning and understanding of indoor environments. Related materials can be found at https://github.com/linukc/OVIGo-3DHSG.", "date": "2025-07-16", "categories": []}
{"id": "2408.11447", "title": "GaussianOcc: Fully Self-supervised and Efficient 3D Occupancy Estimation with Gaussian Splatting", "abstract": "We introduce GaussianOcc, a systematic method that investigates the two usages of Gaussian splatting for fully self-supervised and efficient 3D occupancy estimation in surround views. First, traditional methods for self-supervised 3D occupancy estimation still require ground truth 6D poses from sensors during training. To address this limitation, we propose Gaussian Splatting for Projection (GSP) module to provide accurate scale information for fully self-supervised training from adjacent view projection. Additionally, existing methods rely on volume rendering for final 3D voxel representation learning using 2D signals (depth maps, semantic maps), which is both time-consuming and less effective. We propose Gaussian Splatting from Voxel space (GSV) to leverage the fast rendering properties of Gaussian splatting. As a result, the proposed GaussianOcc method enables fully self-supervised (no ground truth pose) 3D occupancy estimation in competitive performance with low computational cost (2.7 times faster in training and 5 times faster in rendering). The relevant code is available in https://github.com/GANWANSHUI/GaussianOcc.git.", "date": "2024-08-21", "categories": []}
{"id": "2411.09145", "title": "Self-Supervised Monocular 4D Scene Reconstruction for Egocentric Videos", "abstract": "Egocentric videos provide valuable insights into human interactions with the physical world, which has sparked growing interest in the computer vision and robotics communities. A critical challenge in fully understanding the geometry and dynamics of egocentric videos is dense scene reconstruction. However, the lack of high-quality labeled datasets in this field has hindered the effectiveness of current supervised learning methods. In this work, we aim to address this issue by exploring an self-supervised dynamic scene reconstruction approach. We introduce EgoMono4D, a novel model that unifies the estimation of multiple variables necessary for Egocentric Monocular 4D reconstruction, including camera intrinsic, camera poses, and video depth, all within a fast feed-forward framework. Starting from pretrained single-frame depth and intrinsic estimation model, we extend it with camera poses estimation and align multi-frame results on large-scale unlabeled egocentric videos. We evaluate EgoMono4D in both in-domain and zero-shot generalization settings, achieving superior performance in dense pointclouds sequence reconstruction compared to all baselines. EgoMono4D represents the first attempt to apply self-supervised learning for pointclouds sequence reconstruction to the label-scarce egocentric field, enabling fast, dense, and generalizable reconstruction. The interactable visualization, code and trained models are released https://egomono4d.github.io/", "date": "2024-11-14", "categories": []}
{"id": "2505.11868", "title": "MonoMobility: Zero-Shot 3D Mobility Analysis from Monocular Videos", "abstract": "Accurately analyzing the motion parts and their motion attributes in dynamic environments is crucial for advancing key areas such as embodied intelligence. Addressing the limitations of existing methods that rely on dense multi-view images or detailed part-level annotations, we propose an innovative framework that can analyze 3D mobility from monocular videos in a zero-shot manner. This framework can precisely parse motion parts and motion attributes only using a monocular video, completely eliminating the need for annotated training data. Specifically, our method first constructs the scene geometry and roughly analyzes the motion parts and their initial motion attributes combining depth estimation, optical flow analysis and point cloud registration method, then employs 2D Gaussian splatting for scene representation. Building on this, we introduce an end-to-end dynamic scene optimization algorithm specifically designed for articulated objects, refining the initial analysis results to ensure the system can handle 'rotation', 'translation', and even complex movements ('rotation+translation'), demonstrating high flexibility and versatility. To validate the robustness and wide applicability of our method, we created a comprehensive dataset comprising both simulated and real-world scenarios. Experimental results show that our framework can effectively analyze articulated object motions in an annotation-free manner, showcasing its significant potential in future embodied intelligence applications.", "date": "2025-05-17", "categories": []}
{"id": "2507.08555", "title": "Disentangling Instance and Scene Contexts for 3D Semantic Scene Completion", "abstract": "3D Semantic Scene Completion (SSC) has gained increasing attention due to its pivotal role in 3D perception. Recent advancements have primarily focused on refining voxel-level features to construct 3D scenes. However, treating voxels as the basic interaction units inherently limits the utilization of class-level information, which is proven critical for enhancing the granularity of completion results. To address this, we propose \\textbf{D}isentangling Instance and Scene Contexts (DISC), a novel dual-stream paradigm that enhances learning for both instance and scene categories through separated optimization. Specifically, we replace voxel queries with discriminative class queries, which incorporate class-specific geometric and semantic priors. Additionally, we exploit the intrinsic properties of classes to design specialized decoding modules, facilitating targeted interactions and efficient class-level information flow. Experimental results demonstrate that DISC achieves state-of-the-art (SOTA) performance on both SemanticKITTI and SSCBench-KITTI-360 benchmarks, with mIoU scores of 17.35 and 20.55, respectively. Remarkably, DISC even outperforms multi-frame SOTA methods using only single-frame input and significantly improves instance category performance, surpassing both single-frame and multi-frame SOTA instance mIoU by 17.9\\% and 11.9\\%, respectively, on the SemanticKITTI hidden test. The code is available at https://github.com/Enyu-Liu/DISC.", "date": "2025-07-11", "categories": []}
{"id": "2507.07395", "title": "Seg-Wild: Interactive Segmentation based on 3D Gaussian Splatting for Unconstrained Image Collections", "abstract": "Reconstructing and segmenting scenes from unconstrained photo collections obtained from the Internet is a novel but challenging task. Unconstrained photo collections are easier to get than well-captured photo collections. These unconstrained images suffer from inconsistent lighting and transient occlusions, which makes segmentation challenging. Previous segmentation methods cannot address transient occlusions or accurately restore the scene's lighting conditions. Therefore, we propose Seg-Wild, an interactive segmentation method based on 3D Gaussian Splatting for unconstrained image collections, suitable for in-the-wild scenes. We integrate multi-dimensional feature embeddings for each 3D Gaussian and calculate the feature similarity between the feature embeddings and the segmentation target to achieve interactive segmentation in the 3D scene. Additionally, we introduce the Spiky 3D Gaussian Cutter (SGC) to smooth abnormal 3D Gaussians. We project the 3D Gaussians onto a 2D plane and calculate the ratio of 3D Gaussians that need to be cut using the SAM mask. We also designed a benchmark to evaluate segmentation quality in in-the-wild scenes. Experimental results demonstrate that compared to previous methods, Seg-Wild achieves better segmentation results and reconstruction quality. Our code will be available at https://github.com/Sugar0725/Seg-Wild.", "date": "2025-07-10", "categories": []}
{"id": "2507.07781", "title": "SURPRISE3D: A Dataset for Spatial Understanding and Reasoning in Complex 3D Scenes", "abstract": "The integration of language and 3D perception is critical for embodied AI and robotic systems to perceive, understand, and interact with the physical world. Spatial reasoning, a key capability for understanding spatial relationships between objects, remains underexplored in current 3D vision-language research. Existing datasets often mix semantic cues (e.g., object name) with spatial context, leading models to rely on superficial shortcuts rather than genuinely interpreting spatial relationships. To address this gap, we introduce S\\textsc{urprise}3D, a novel dataset designed to evaluate language-guided spatial reasoning segmentation in complex 3D scenes. S\\textsc{urprise}3D consists of more than 200k vision language pairs across 900+ detailed indoor scenes from ScanNet++ v2, including more than 2.8k unique object classes. The dataset contains 89k+ human-annotated spatial queries deliberately crafted without object name, thereby mitigating shortcut biases in spatial understanding. These queries comprehensively cover various spatial reasoning skills, such as relative position, narrative perspective, parametric perspective, and absolute distance reasoning. Initial benchmarks demonstrate significant challenges for current state-of-the-art expert 3D visual grounding methods and 3D-LLMs, underscoring the necessity of our dataset and the accompanying 3D Spatial Reasoning Segmentation (3D-SRS) benchmark suite. S\\textsc{urprise}3D and 3D-SRS aim to facilitate advancements in spatially aware AI, paving the way for effective embodied interaction and robotic planning. The code and datasets can be found in https://github.com/liziwennba/SUPRISE.", "date": "2025-07-10", "categories": []}
{"id": "2410.11564", "title": "PAVLM: Advancing Point Cloud based Affordance Understanding Via Vision-Language Model", "abstract": "Affordance understanding, the task of identifying actionable regions on 3D objects, plays a vital role in allowing robotic systems to engage with and operate within the physical world. Although Visual Language Models (VLMs) have excelled in high-level reasoning and long-horizon planning for robotic manipulation, they still fall short in grasping the nuanced physical properties required for effective human-robot interaction. In this paper, we introduce PAVLM (Point cloud Affordance Vision-Language Model), an innovative framework that utilizes the extensive multimodal knowledge embedded in pre-trained language models to enhance 3D affordance understanding of point cloud. PAVLM integrates a geometric-guided propagation module with hidden embeddings from large language models (LLMs) to enrich visual semantics. On the language side, we prompt Llama-3.1 models to generate refined context-aware text, augmenting the instructional input with deeper semantic cues. Experimental results on the 3D-AffordanceNet benchmark demonstrate that PAVLM outperforms baseline methods for both full and partial point clouds, particularly excelling in its generalization to novel open-world affordance tasks of 3D objects. For more information, visit our project site: pavlm-source.github.io.", "date": "2024-10-15", "categories": []}
{"id": "2507.02813", "title": "LangScene-X: Reconstruct Generalizable 3D Language-Embedded Scenes with TriMap Video Diffusion", "abstract": "Recovering 3D structures with open-vocabulary scene understanding from 2D images is a fundamental but daunting task. Recent developments have achieved this by performing per-scene optimization with embedded language information. However, they heavily rely on the calibrated dense-view reconstruction paradigm, thereby suffering from severe rendering artifacts and implausible semantic synthesis when limited views are available. In this paper, we introduce a novel generative framework, coined LangScene-X, to unify and generate 3D consistent multi-modality information for reconstruction and understanding. Powered by the generative capability of creating more consistent novel observations, we can build generalizable 3D language-embedded scenes from only sparse views. Specifically, we first train a TriMap video diffusion model that can generate appearance (RGBs), geometry (normals), and semantics (segmentation maps) from sparse inputs through progressive knowledge integration. Furthermore, we propose a Language Quantized Compressor (LQC), trained on large-scale image datasets, to efficiently encode language embeddings, enabling cross-scene generalization without per-scene retraining. Finally, we reconstruct the language surface fields by aligning language information onto the surface of 3D scenes, enabling open-ended language queries. Extensive experiments on real-world data demonstrate the superiority of our LangScene-X over state-of-the-art methods in terms of quality and generalizability. Project Page: https://liuff19.github.io/LangScene-X.", "date": "2025-07-03", "categories": []}
{"id": "2412.01398", "title": "Articulate3D: Holistic Understanding of 3D Scenes as Universal Scene Description", "abstract": "3D scene understanding is a long-standing challenge in computer vision and a key component in enabling mixed reality, wearable computing, and embodied AI. Providing a solution to these applications requires a multifaceted approach that covers scene-centric, object-centric, as well as interaction-centric capabilities. While there exist numerous datasets and algorithms approaching the former two problems, the task of understanding interactable and articulated objects is underrepresented and only partly covered in the research field. In this work, we address this shortcoming by introducing: (1) Articulate3D, an expertly curated 3D dataset featuring high-quality manual annotations on 280 indoor scenes. Articulate3D provides 8 types of annotations for articulated objects, covering parts and detailed motion information, all stored in a standardized scene representation format designed for scalable 3D content creation, exchange and seamless integration into simulation environments. (2) USDNet, a novel unified framework capable of simultaneously predicting part segmentation along with a full specification of motion attributes for articulated objects. We evaluate USDNet on Articulate3D as well as two existing datasets, demonstrating the advantage of our unified dense prediction approach. Furthermore, we highlight the value of Articulate3D through cross-dataset and cross-domain evaluations and showcase its applicability in downstream tasks such as scene editing through LLM prompting and robotic policy training for articulated object manipulation. We provide open access to our dataset, benchmark, and method's source code.", "date": "2024-12-02", "categories": []}
{"id": "2506.23309", "title": "SurgTPGS: Semantic 3D Surgical Scene Understanding with Text Promptable Gaussian Splatting", "abstract": "In contemporary surgical research and practice, accurately comprehending 3D surgical scenes with text-promptable capabilities is particularly crucial for surgical planning and real-time intra-operative guidance, where precisely identifying and interacting with surgical tools and anatomical structures is paramount. However, existing works focus on surgical vision-language model (VLM), 3D reconstruction, and segmentation separately, lacking support for real-time text-promptable 3D queries. In this paper, we present SurgTPGS, a novel text-promptable Gaussian Splatting method to fill this gap. We introduce a 3D semantics feature learning strategy incorporating the Segment Anything model and state-of-the-art vision-language models. We extract the segmented language features for 3D surgical scene reconstruction, enabling a more in-depth understanding of the complex surgical environment. We also propose semantic-aware deformation tracking to capture the seamless deformation of semantic features, providing a more precise reconstruction for both texture and semantic features. Furthermore, we present semantic region-aware optimization, which utilizes regional-based semantic information to supervise the training, particularly promoting the reconstruction quality and semantic smoothness. We conduct comprehensive experiments on two real-world surgical datasets to demonstrate the superiority of SurgTPGS over state-of-the-art methods, highlighting its potential to revolutionize surgical practices. SurgTPGS paves the way for developing next-generation intelligent surgical systems by enhancing surgical precision and safety. Our code is available at: https://github.com/lastbasket/SurgTPGS.", "date": "2025-06-29", "categories": []}
{"id": "2411.18066", "title": "GLS: Geometry-aware 3D Language Gaussian Splatting", "abstract": "Recently, 3D Gaussian Splatting (3DGS) has achieved impressive performance on indoor surface reconstruction and 3D open-vocabulary segmentation. This paper presents GLS, a unified framework of 3D surface reconstruction and open-vocabulary segmentation based on 3DGS. GLS extends two fields by improving their sharpness and smoothness. For indoor surface reconstruction, we introduce surface normal prior as a geometric cue to guide the rendered normal, and use the normal error to optimize the rendered depth. For 3D open-vocabulary segmentation, we employ 2D CLIP features to guide instance features and enhance the surface smoothness, then utilize DEVA masks to maintain their view consistency. Extensive experiments demonstrate the effectiveness of jointly optimizing surface reconstruction and 3D open-vocabulary segmentation, where GLS surpasses state-of-the-art approaches of each task on MuSHRoom, ScanNet++ and LERF-OVS datasets. Project webpage: https://jiaxiongq.github.io/GLS_ProjectPage.", "date": "2024-11-27", "categories": []}
{"id": "2506.22593", "title": "Pixels-to-Graph: Real-time Integration of Building Information Models and Scene Graphs for Semantic-Geometric Human-Robot Understanding", "abstract": "Autonomous robots are increasingly playing key roles as support platforms for human operators in high-risk, dangerous applications. To accomplish challenging tasks, an efficient human-robot cooperation and understanding is required. While typically robotic planning leverages 3D geometric information, human operators are accustomed to a high-level compact representation of the environment, like top-down 2D maps representing the Building Information Model (BIM). 3D scene graphs have emerged as a powerful tool to bridge the gap between human readable 2D BIM and the robot 3D maps. In this work, we introduce Pixels-to-Graph (Pix2G), a novel lightweight method to generate structured scene graphs from image pixels and LiDAR maps in real-time for the autonomous exploration of unknown environments on resource-constrained robot platforms. To satisfy onboard compute constraints, the framework is designed to perform all operation on CPU only. The method output are a de-noised 2D top-down environment map and a structure-segmented 3D pointcloud which are seamlessly connected using a multi-layer graph abstracting information from object-level up to the building-level. The proposed method is quantitatively and qualitatively evaluated during real-world experiments performed using the NASA JPL NeBula-Spot legged robot to autonomously explore and map cluttered garage and urban office like environments in real-time.", "date": "2025-06-27", "categories": []}
{"id": "2506.22799", "title": "VoteSplat: Hough Voting Gaussian Splatting for 3D Scene Understanding", "abstract": "3D Gaussian Splatting (3DGS) has become horsepower in high-quality, real-time rendering for novel view synthesis of 3D scenes. However, existing methods focus primarily on geometric and appearance modeling, lacking deeper scene understanding while also incurring high training costs that complicate the originally streamlined differentiable rendering pipeline. To this end, we propose VoteSplat, a novel 3D scene understanding framework that integrates Hough voting with 3DGS. Specifically, Segment Anything Model (SAM) is utilized for instance segmentation, extracting objects, and generating 2D vote maps. We then embed spatial offset vectors into Gaussian primitives. These offsets construct 3D spatial votes by associating them with 2D image votes, while depth distortion constraints refine localization along the depth axis. For open-vocabulary object localization, VoteSplat maps 2D image semantics to 3D point clouds via voting points, reducing training costs associated with high-dimensional CLIP features while preserving semantic unambiguity. Extensive experiments demonstrate effectiveness of VoteSplat in open-vocabulary 3D instance localization, 3D point cloud understanding, click-based 3D object localization, hierarchical segmentation, and ablation studies. Our code is available at https://sy-ja.github.io/votesplat/", "date": "2025-06-28", "categories": []}
{"id": "2506.22817", "title": "Unleashing the Multi-View Fusion Potential: Noise Correction in VLM for Open-Vocabulary 3D Scene Understanding", "abstract": "Recent open-vocabulary 3D scene understanding approaches mainly focus on training 3D networks through contrastive learning with point-text pairs or by distilling 2D features into 3D models via point-pixel alignment. While these methods show considerable performance in benchmarks with limited vocabularies, they struggle to handle diverse object categories as the limited amount of 3D data upbound training strong open-vocabulary 3d models. We observe that 2D multi-view fusion methods take precedence in understanding diverse concepts in 3D scenes. However, inherent noises in vision-language models lead multi-view fusion to sub-optimal performance. To this end, we introduce MVOV3D, a novel approach aimed at unleashing the potential of 2D multi-view fusion for open-vocabulary 3D scene understanding. We focus on reducing the inherent noises without training, thereby preserving the generalizability while enhancing open-world capabilities. Specifically, MVOV3D improves multi-view 2D features by leveraging precise region-level image features and text features encoded by CLIP encoders and incorporates 3D geometric priors to optimize multi-view fusion. Extensive experiments on various datasets demonstrate the effectiveness of our method. Notably, our MVOV3D achieves a new record with 14.7% mIoU on ScanNet200 and 16.2% mIoU on Matterport160 for challenge open-vocabulary semantic segmentation, outperforming current leading trained 3D networks by a significant margin.", "date": "2025-06-28", "categories": []}
{"id": "2506.23120", "title": "Enhancing Spatial Reasoning in Multimodal Large Language Models through Reasoning-based Segmentation", "abstract": "Recent advances in point cloud perception have demonstrated remarkable progress in scene understanding through vision-language alignment leveraging large language models (LLMs). However, existing methods may still encounter challenges in handling complex instructions that require accurate spatial reasoning, even if the 3D point cloud data provides detailed spatial cues such as size and position for identifying the targets. To tackle this issue, we propose Relevant Reasoning Segmentation (R$^2$S), a reasoning-based segmentation framework. The framework emulates human cognitive processes by decomposing spatial reasoning into two sequential stages: first identifying relevant elements, then processing instructions guided by their associated visual priors. Furthermore, acknowledging the inadequacy of existing datasets in complex reasoning tasks, we introduce 3D ReasonSeg, a reasoning-based segmentation dataset comprising 25,185 training samples and 3,966 validation samples with precise annotations. Both quantitative and qualitative experiments demonstrate that the R$^2$S and 3D ReasonSeg effectively endow 3D point cloud perception with stronger spatial reasoning capabilities, and we hope that they can serve as a new baseline and benchmark for future work.", "date": "2025-06-29", "categories": []}
{"id": "2506.23352", "title": "GeoProg3D: Compositional Visual Reasoning for City-Scale 3D Language Fields", "abstract": "The advancement of 3D language fields has enabled intuitive interactions with 3D scenes via natural language. However, existing approaches are typically limited to small-scale environments, lacking the scalability and compositional reasoning capabilities necessary for large, complex urban settings. To overcome these limitations, we propose GeoProg3D, a visual programming framework that enables natural language-driven interactions with city-scale high-fidelity 3D scenes. GeoProg3D consists of two key components: (i) a Geography-aware City-scale 3D Language Field (GCLF) that leverages a memory-efficient hierarchical 3D model to handle large-scale data, integrated with geographic information for efficiently filtering vast urban spaces using directional cues, distance measurements, elevation data, and landmark references; and (ii) Geographical Vision APIs (GV-APIs), specialized geographic vision tools such as area segmentation and object detection. Our framework employs large language models (LLMs) as reasoning engines to dynamically combine GV-APIs and operate GCLF, effectively supporting diverse geographic vision tasks. To assess performance in city-scale reasoning, we introduce GeoEval3D, a comprehensive benchmark dataset containing 952 query-answer pairs across five challenging tasks: grounding, spatial reasoning, comparison, counting, and measurement. Experiments demonstrate that GeoProg3D significantly outperforms existing 3D language fields and vision-language models across multiple tasks. To our knowledge, GeoProg3D is the first framework enabling compositional geographic reasoning in high-fidelity city-scale 3D environments via natural language. The code is available at https://snskysk.github.io/GeoProg3D/.", "date": "2025-06-29", "categories": []}
{"id": "2503.13188", "title": "3D Hierarchical Panoptic Segmentation in Real Orchard Environments Across Different Sensors", "abstract": "Crop yield estimation is a relevant problem in agriculture, because an accurate yield estimate can support farmers' decisions on harvesting or precision intervention. Robots can help to automate this process. To do so, they need to be able to perceive the surrounding environment to identify target objects such as trees and plants. In this paper, we introduce a novel approach to address the problem of hierarchical panoptic segmentation of apple orchards on 3D data from different sensors. Our approach is able to simultaneously provide semantic segmentation, instance segmentation of trunks and fruits, and instance segmentation of trees (a trunk with its fruits). This allows us to identify relevant information such as individual plants, fruits, and trunks, and capture the relationship among them, such as precisely estimate the number of fruits associated to each tree in an orchard. To efficiently evaluate our approach for hierarchical panoptic segmentation, we provide a dataset designed specifically for this task. Our dataset is recorded in Bonn, Germany, in a real apple orchard with a variety of sensors, spanning from a terrestrial laser scanner to a RGB-D camera mounted on different robots platforms. The experiments show that our approach surpasses state-of-the-art approaches in 3D panoptic segmentation in the agricultural domain, while also providing full hierarchical panoptic segmentation. Our dataset is publicly available at https://www.ipb.uni-bonn.de/data/hops/. The open-source implementation of our approach is available at https://github.com/PRBonn/hapt3D.", "date": "2025-03-17", "categories": []}
{"id": "2506.20979", "title": "3D Scene-Camera Representation with Joint Camera Photometric Optimization", "abstract": "Representing scenes from multi-view images is a crucial task in computer vision with extensive applications. However, inherent photometric distortions in the camera imaging can significantly degrade image quality. Without accounting for these distortions, the 3D scene representation may inadvertently incorporate erroneous information unrelated to the scene, diminishing the quality of the representation. In this paper, we propose a novel 3D scene-camera representation with joint camera photometric optimization. By introducing internal and external photometric model, we propose a full photometric model and corresponding camera representation. Based on simultaneously optimizing the parameters of the camera representation, the proposed method effectively separates scene-unrelated information from the 3D scene representation. Additionally, during the optimization of the photometric parameters, we introduce a depth regularization to prevent the 3D scene representation from fitting scene-unrelated information. By incorporating the camera model as part of the mapping process, the proposed method constructs a complete map that includes both the scene radiance field and the camera photometric model. Experimental results demonstrate that the proposed method can achieve high-quality 3D scene representations, even under conditions of imaging degradation, such as vignetting and dirt.", "date": "2025-06-26", "categories": []}
{"id": "2506.19331", "title": "Segment Any 3D-Part in a Scene from a Sentence", "abstract": "This paper aims to achieve the segmentation of any 3D part in a scene based on natural language descriptions, extending beyond traditional object-level 3D scene understanding and addressing both data and methodological challenges. Due to the expensive acquisition and annotation burden, existing datasets and methods are predominantly limited to object-level comprehension. To overcome the limitations of data and annotation availability, we introduce the 3D-PU dataset, the first large-scale 3D dataset with dense part annotations, created through an innovative and cost-effective method for constructing synthetic 3D scenes with fine-grained part-level annotations, paving the way for advanced 3D-part scene understanding. On the methodological side, we propose OpenPart3D, a 3D-input-only framework to effectively tackle the challenges of part-level segmentation. Extensive experiments demonstrate the superiority of our approach in open-vocabulary 3D scene understanding tasks at the part level, with strong generalization capabilities across various 3D scene datasets.", "date": "2025-06-24", "categories": []}
{"id": "2506.17545", "title": "Scene-R1: Video-Grounded Large Language Models for 3D Scene Reasoning without 3D Annotations", "abstract": "Currently, utilizing large language models to understand the 3D world is becoming popular. Yet existing 3D-aware LLMs act as black boxes: they output bounding boxes or textual answers without revealing how those decisions are made, and they still rely on pre-trained 3D detectors to supply object proposals. We introduce Scene-R1, a video-grounded framework that learns to reason about 3D scenes without any point-wise 3D instance supervision by pairing reinforcement-learning-driven reasoning with a two-stage grounding pipeline. In the temporal grounding stage, we explicitly reason about the video and select the video snippets most relevant to an open-ended query. In the subsequent image grounding stage, we analyze the image and predict the 2D bounding box. After that, we track the object using SAM2 to produce pixel-accurate masks in RGB frames, and project them back into 3D, thereby eliminating the need for 3D detector-based proposals while capturing fine geometry and material cues. Scene-R1 can also adapt to the 3D visual question answering task to answer free-form questions directly from video. Our training pipeline only needs task-level 2D boxes or textual labels without dense 3D point-wise labels. Scene-R1 surpasses existing open-vocabulary baselines on multiple datasets, while delivering transparent, step-by-step rationales. These results show that reinforcement-learning-based reasoning combined with RGB-D video alone offers a practical, annotation-efficient route to trustworthy 3D scene understanding.", "date": "2025-06-21", "categories": []}
{"id": "2506.18266", "title": "YouTube-Occ: Learning Indoor 3D Semantic Occupancy Prediction from YouTube Videos", "abstract": "3D semantic occupancy prediction in the past was considered to require precise geometric relationships in order to enable effective training. However, in complex indoor environments, the large-scale and widespread collection of data, along with the necessity for fine-grained annotations, becomes impractical due to the complexity of data acquisition setups and privacy concerns. In this paper, we demonstrate that 3D spatially-accurate training can be achieved using only indoor Internet data, without the need for any pre-knowledge of intrinsic or extrinsic camera parameters. In our framework, we collect a web dataset, YouTube-Occ, which comprises house tour videos from YouTube, providing abundant real house scenes for 3D representation learning. Upon on this web dataset, we establish a fully self-supervised model to leverage accessible 2D prior knowledge for reaching powerful 3D indoor perception. Specifically, we harness the advantages of the prosperous vision foundation models, distilling the 2D region-level knowledge into the occupancy network by grouping the similar pixels into superpixels. Experimental results show that our method achieves state-of-the-art zero-shot performance on two popular benchmarks (NYUv2 and OccScanNet", "date": "2025-06-23", "categories": []}
{"id": "2503.00513", "title": "Inst3D-LMM: Instance-Aware 3D Scene Understanding with Multi-modal Instruction Tuning", "abstract": "Despite encouraging progress in 3D scene understanding, it remains challenging to develop an effective Large Multi-modal Model (LMM) that is capable of understanding and reasoning in complex 3D environments. Most previous methods typically encode 3D point and 2D image features separately, neglecting interactions between 2D semantics and 3D object properties, as well as the spatial relationships within the 3D environment. This limitation not only hinders comprehensive representations of 3D scene, but also compromises training and inference efficiency. To address these challenges, we propose a unified Instance-aware 3D Large Multi-modal Model (Inst3D-LMM) to deal with multiple 3D scene understanding tasks simultaneously. To obtain the fine-grained instance-level visual tokens, we first introduce a novel Multi-view Cross-Modal Fusion (MCMF) module to inject the multi-view 2D semantics into their corresponding 3D geometric features. For scene-level relation-aware tokens, we further present a 3D Instance Spatial Relation (3D-ISR) module to capture the intricate pairwise spatial relationships among objects. Additionally, we perform end-to-end multi-task instruction tuning simultaneously without the subsequent task-specific fine-tuning. Extensive experiments demonstrate that our approach outperforms the state-of-the-art methods across 3D scene understanding, reasoning and grounding tasks. Source code is available at https://github.com/hanxunyu/Inst3D-LMM", "date": "2025-03-01", "categories": []}
{"id": "2506.09565", "title": "SemanticSplat: Feed-Forward 3D Scene Understanding with Language-Aware Gaussian Fields", "abstract": "Holistic 3D scene understanding, which jointly models geometry, appearance, and semantics, is crucial for applications like augmented reality and robotic interaction. Existing feed-forward 3D scene understanding methods (e.g., LSM) are limited to extracting language-based semantics from scenes, failing to achieve holistic scene comprehension. Additionally, they suffer from low-quality geometry reconstruction and noisy artifacts. In contrast, per-scene optimization methods rely on dense input views, which reduces practicality and increases complexity during deployment. In this paper, we propose SemanticSplat, a feed-forward semantic-aware 3D reconstruction method, which unifies 3D Gaussians with latent semantic attributes for joint geometry-appearance-semantics modeling. To predict the semantic anisotropic Gaussians, SemanticSplat fuses diverse feature fields (e.g., LSeg, SAM) with a cost volume representation that stores cross-view feature similarities, enhancing coherent and accurate scene comprehension. Leveraging a two-stage distillation framework, SemanticSplat reconstructs a holistic multi-modal semantic feature field from sparse-view images. Experiments demonstrate the effectiveness of our method for 3D scene understanding tasks like promptable and open-vocabulary segmentation. Video results are available at https://semanticsplat.github.io.", "date": "2025-06-11", "categories": []}
{"id": "2506.12009", "title": "Affogato: Learning Open-Vocabulary Affordance Grounding with Automated Data Generation at Scale", "abstract": "Affordance grounding-localizing object regions based on natural language descriptions of interactions-is a critical challenge for enabling intelligent agents to understand and interact with their environments. However, this task remains challenging due to the need for fine-grained part-level localization, the ambiguity arising from multiple valid interaction regions, and the scarcity of large-scale datasets. In this work, we introduce Affogato, a large-scale benchmark comprising 150K instances, annotated with open-vocabulary text descriptions and corresponding 3D affordance heatmaps across a diverse set of objects and interactions. Building on this benchmark, we develop simple yet effective vision-language models that leverage pretrained part-aware vision backbones and a text-conditional heatmap decoder. Our models trained with the Affogato dataset achieve promising performance on the existing 2D and 3D benchmarks, and notably, exhibit effectiveness in open-vocabulary cross-domain generalization. The Affogato dataset is shared in public: https://huggingface.co/datasets/project-affogato/affogato", "date": "2025-06-13", "categories": []}
{"id": "2506.10980", "title": "InstaInpaint: Instant 3D-Scene Inpainting with Masked Large Reconstruction Model", "abstract": "Recent advances in 3D scene reconstruction enable real-time viewing in virtual and augmented reality. To support interactive operations for better immersiveness, such as moving or editing objects, 3D scene inpainting methods are proposed to repair or complete the altered geometry. However, current approaches rely on lengthy and computationally intensive optimization, making them impractical for real-time or online applications. We propose InstaInpaint, a reference-based feed-forward framework that produces 3D-scene inpainting from a 2D inpainting proposal within 0.4 seconds. We develop a self-supervised masked-finetuning strategy to enable training of our custom large reconstruction model (LRM) on the large-scale dataset. Through extensive experiments, we analyze and identify several key designs that improve generalization, textural consistency, and geometric correctness. InstaInpaint achieves a 1000x speed-up from prior methods while maintaining a state-of-the-art performance across two standard benchmarks. Moreover, we show that InstaInpaint generalizes well to flexible downstream applications such as object insertion and multi-region inpainting. More video results are available at our project page: https://dhmbb2.github.io/InstaInpaint_page/.", "date": "2025-06-12", "categories": []}
{"id": "2506.08777", "title": "Gaussian2Scene: 3D Scene Representation Learning via Self-supervised Learning with 3D Gaussian Splatting", "abstract": "Self-supervised learning (SSL) for point cloud pre-training has become a cornerstone for many 3D vision tasks, enabling effective learning from large-scale unannotated data. At the scene level, existing SSL methods often incorporate volume rendering into the pre-training framework, using RGB-D images as reconstruction signals to facilitate cross-modal learning. This strategy promotes alignment between 2D and 3D modalities and enables the model to benefit from rich visual cues in the RGB-D inputs. However, these approaches are limited by their reliance on implicit scene representations and high memory demands. Furthermore, since their reconstruction objectives are applied only in 2D space, they often fail to capture underlying 3D geometric structures. To address these challenges, we propose Gaussian2Scene, a novel scene-level SSL framework that leverages the efficiency and explicit nature of 3D Gaussian Splatting (3DGS) for pre-training. The use of 3DGS not only alleviates the computational burden associated with volume rendering but also supports direct 3D scene reconstruction, thereby enhancing the geometric understanding of the backbone network. Our approach follows a progressive two-stage training strategy. In the first stage, a dual-branch masked autoencoder learns both 2D and 3D scene representations. In the second stage, we initialize training with reconstructed point clouds and further supervise learning using the geometric locations of Gaussian primitives and rendered RGB images. This process reinforces both geometric and cross-modal learning. We demonstrate the effectiveness of Gaussian2Scene across several downstream 3D object detection tasks, showing consistent improvements over existing pre-training methods.", "date": "2025-06-10", "categories": []}
{"id": "2506.09378", "title": "UniForward: Unified 3D Scene and Semantic Field Reconstruction via Feed-Forward Gaussian Splatting from Only Sparse-View Images", "abstract": "We propose a feed-forward Gaussian Splatting model that unifies 3D scene and semantic field reconstruction. Combining 3D scenes with semantic fields facilitates the perception and understanding of the surrounding environment. However, key challenges include embedding semantics into 3D representations, achieving generalizable real-time reconstruction, and ensuring practical applicability by using only images as input without camera parameters or ground truth depth. To this end, we propose UniForward, a feed-forward model to predict 3D Gaussians with anisotropic semantic features from only uncalibrated and unposed sparse-view images. To enable the unified representation of the 3D scene and semantic field, we embed semantic features into 3D Gaussians and predict them through a dual-branch decoupled decoder. During training, we propose a loss-guided view sampler to sample views from easy to hard, eliminating the need for ground truth depth or masks required by previous methods and stabilizing the training process. The whole model can be trained end-to-end using a photometric loss and a distillation loss that leverages semantic features from a pre-trained 2D semantic model. At the inference stage, our UniForward can reconstruct 3D scenes and the corresponding semantic fields in real time from only sparse-view images. The reconstructed 3D scenes achieve high-quality rendering, and the reconstructed 3D semantic field enables the rendering of view-consistent semantic features from arbitrary views, which can be further decoded into dense segmentation masks in an open-vocabulary manner. Experiments on novel view synthesis and novel view segmentation demonstrate that our method achieves state-of-the-art performances for unifying 3D scene and semantic field reconstruction.", "date": "2025-06-11", "categories": []}
{"id": "2406.09401", "title": "MMScan: A Multi-Modal 3D Scene Dataset with Hierarchical Grounded Language Annotations", "abstract": "With the emergence of LLMs and their integration with other data modalities, multi-modal 3D perception attracts more attention due to its connectivity to the physical world and makes rapid progress. However, limited by existing datasets, previous works mainly focus on understanding object properties or inter-object spatial relationships in a 3D scene. To tackle this problem, this paper builds the first largest ever multi-modal 3D scene dataset and benchmark with hierarchical grounded language annotations, MMScan. It is constructed based on a top-down logic, from region to object level, from a single target to inter-target relationships, covering holistic aspects of spatial and attribute understanding. The overall pipeline incorporates powerful VLMs via carefully designed prompts to initialize the annotations efficiently and further involve humans' correction in the loop to ensure the annotations are natural, correct, and comprehensive. Built upon existing 3D scanning data, the resulting multi-modal 3D dataset encompasses 1.4M meta-annotated captions on 109k objects and 7.7k regions as well as over 3.04M diverse samples for 3D visual grounding and question-answering benchmarks. We evaluate representative baselines on our benchmarks, analyze their capabilities in different aspects, and showcase the key problems to be addressed in the future. Furthermore, we use this high-quality dataset to train state-of-the-art 3D visual grounding and LLMs and obtain remarkable performance improvement both on existing benchmarks and in-the-wild evaluation. Codes, datasets, and benchmarks will be available at https://github.com/OpenRobotLab/EmbodiedScan.", "date": "2024-06-13", "categories": []}
{"id": "2506.06822", "title": "Hi-LSplat: Hierarchical 3D Language Gaussian Splatting", "abstract": "Modeling 3D language fields with Gaussian Splatting for open-ended language queries has recently garnered increasing attention. However, recent 3DGS-based models leverage view-dependent 2D foundation models to refine 3D semantics but lack a unified 3D representation, leading to view inconsistencies. Additionally, inherent open-vocabulary challenges cause inconsistencies in object and relational descriptions, impeding hierarchical semantic understanding. In this paper, we propose Hi-LSplat, a view-consistent Hierarchical Language Gaussian Splatting work for 3D open-vocabulary querying. To achieve view-consistent 3D hierarchical semantics, we first lift 2D features to 3D features by constructing a 3D hierarchical semantic tree with layered instance clustering, which addresses the view inconsistency issue caused by 2D semantic features. Besides, we introduce instance-wise and part-wise contrastive losses to capture all-sided hierarchical semantic representations. Notably, we construct two hierarchical semantic datasets to better assess the model's ability to distinguish different semantic levels. Extensive experiments highlight our method's superiority in 3D open-vocabulary segmentation and localization. Its strong performance on hierarchical semantic datasets underscores its ability to capture complex hierarchical semantics within 3D scenes.", "date": "2025-06-07", "categories": []}
{"id": "2506.06995", "title": "Technical Report for ICRA 2025 GOOSE 3D Semantic Segmentation Challenge: Adaptive Point Cloud Understanding for Heterogeneous Robotic Systems", "abstract": "This technical report presents the implementation details of the winning solution for the ICRA 2025 GOOSE 3D Semantic Segmentation Challenge. This challenge focuses on semantic segmentation of 3D point clouds from diverse unstructured outdoor environments collected from multiple robotic platforms. This problem was addressed by implementing Point Prompt Tuning (PPT) integrated with Point Transformer v3 (PTv3) backbone, enabling adaptive processing of heterogeneous LiDAR data through platform-specific conditioning and cross-dataset class alignment strategies. The model is trained without requiring additional external data. As a result, this approach achieved substantial performance improvements with mIoU increases of up to 22.59% on challenging platforms compared to the baseline PTv3 model, demonstrating the effectiveness of adaptive point cloud understanding for field robotics applications.", "date": "2025-06-08", "categories": []}
{"id": "2506.07697", "title": "OpenSplat3D: Open-Vocabulary 3D Instance Segmentation using Gaussian Splatting", "abstract": "3D Gaussian Splatting (3DGS) has emerged as a powerful representation for neural scene reconstruction, offering high-quality novel view synthesis while maintaining computational efficiency. In this paper, we extend the capabilities of 3DGS beyond pure scene representation by introducing an approach for open-vocabulary 3D instance segmentation without requiring manual labeling, termed OpenSplat3D. Our method leverages feature-splatting techniques to associate semantic information with individual Gaussians, enabling fine-grained scene understanding. We incorporate Segment Anything Model instance masks with a contrastive loss formulation as guidance for the instance features to achieve accurate instance-level segmentation. Furthermore, we utilize language embeddings of a vision-language model, allowing for flexible, text-driven instance identification. This combination enables our system to identify and segment arbitrary objects in 3D scenes based on natural language descriptions. We show results on LERF-mask and LERF-OVS as well as the full ScanNet++ validation set, demonstrating the effectiveness of our approach.", "date": "2025-06-09", "categories": []}
{"id": "2506.05338", "title": "Defurnishing with X-Ray Vision: Joint Removal of Furniture from Panoramas and Mesh", "abstract": "We present a pipeline for generating defurnished replicas of indoor spaces represented as textured meshes and corresponding multi-view panoramic images. To achieve this, we first segment and remove furniture from the mesh representation, extend planes, and fill holes, obtaining a simplified defurnished mesh (SDM). This SDM acts as an ``X-ray'' of the scene's underlying structure, guiding the defurnishing process. We extract Canny edges from depth and normal images rendered from the SDM. We then use these as a guide to remove the furniture from panorama images via ControlNet inpainting. This control signal ensures the availability of global geometric information that may be hidden from a particular panoramic view by the furniture being removed. The inpainted panoramas are used to texture the mesh. We show that our approach produces higher quality assets than methods that rely on neural radiance fields, which tend to produce blurry low-resolution images, or RGB-D inpainting, which is highly susceptible to hallucinations.", "date": "2025-06-05", "categories": []}
{"id": "2506.05473", "title": "S2GO: Streaming Sparse Gaussian Occupancy Prediction", "abstract": "Despite the demonstrated efficiency and performance of sparse query-based representations for perception, state-of-the-art 3D occupancy prediction methods still rely on voxel-based or dense Gaussian-based 3D representations. However, dense representations are slow, and they lack flexibility in capturing the temporal dynamics of driving scenes. Distinct from prior work, we instead summarize the scene into a compact set of 3D queries which are propagated through time in an online, streaming fashion. These queries are then decoded into semantic Gaussians at each timestep. We couple our framework with a denoising rendering objective to guide the queries and their constituent Gaussians in effectively capturing scene geometry. Owing to its efficient, query-based representation, S2GO achieves state-of-the-art performance on the nuScenes and KITTI occupancy benchmarks, outperforming prior art (e.g., GaussianWorld) by 1.5 IoU with 5.9x faster inference.", "date": "2025-06-05", "categories": []}
{"id": "2506.05563", "title": "VoxelSplat: Dynamic Gaussian Splatting as an Effective Loss for Occupancy and Flow Prediction", "abstract": "Recent advancements in camera-based occupancy prediction have focused on the simultaneous prediction of 3D semantics and scene flow, a task that presents significant challenges due to specific difficulties, e.g., occlusions and unbalanced dynamic environments. In this paper, we analyze these challenges and their underlying causes. To address them, we propose a novel regularization framework called VoxelSplat. This framework leverages recent developments in 3D Gaussian Splatting to enhance model performance in two key ways: (i) Enhanced Semantics Supervision through 2D Projection: During training, our method decodes sparse semantic 3D Gaussians from 3D representations and projects them onto the 2D camera view. This provides additional supervision signals in the camera-visible space, allowing 2D labels to improve the learning of 3D semantics. (ii) Scene Flow Learning: Our framework uses the predicted scene flow to model the motion of Gaussians, and is thus able to learn the scene flow of moving objects in a self-supervised manner using the labels of adjacent frames. Our method can be seamlessly integrated into various existing occupancy models, enhancing performance without increasing inference time. Extensive experiments on benchmark datasets demonstrate the effectiveness of VoxelSplat in improving the accuracy of both semantic occupancy and scene flow estimation. The project page and codes are available at https://zzy816.github.io/VoxelSplat-Demo/.", "date": "2025-06-05", "categories": []}
{"id": "2506.05689", "title": "Pts3D-LLM: Studying the Impact of Token Structure for 3D Scene Understanding With Large Language Models", "abstract": "Effectively representing 3D scenes for Multimodal Large Language Models (MLLMs) is crucial yet challenging. Existing approaches commonly only rely on 2D image features and use varied tokenization approaches. This work presents a rigorous study of 3D token structures, systematically comparing video-based and point-based representations while maintaining consistent model backbones and parameters. We propose a novel approach that enriches visual tokens by incorporating 3D point cloud features from a Sonata pretrained Point Transformer V3 encoder. Our experiments demonstrate that merging explicit 3D features significantly boosts performance. Furthermore, we show that point-based token structures can rival video-based ones when the points are cleverly sampled and ordered. Our best models from both structures achieve state-of-the-art results on multiple 3D understanding benchmarks. We emphasize our analysis of token structures as a key contribution, alongside transparent reporting of results averaged over multiple seeds, a practice we believe is vital for robust progress in the field.", "date": "2025-06-06", "categories": []}
{"id": "2503.18052", "title": "SceneSplat: Gaussian Splatting-based Scene Understanding with Vision-Language Pretraining", "abstract": "Recognizing arbitrary or previously unseen categories is essential for comprehensive real-world 3D scene understanding. Currently, all existing methods rely on 2D or textual modalities during training or together at inference. This highlights the clear absence of a model capable of processing 3D data alone for learning semantics end-to-end, along with the necessary data to train such a model. Meanwhile, 3D Gaussian Splatting (3DGS) has emerged as the de facto standard for 3D scene representation across various vision tasks. However, effectively integrating semantic reasoning into 3DGS in a generalizable manner remains an open challenge. To address these limitations, we introduce SceneSplat, to our knowledge the first large-scale 3D indoor scene understanding approach that operates natively on 3DGS. Furthermore, we propose a self-supervised learning scheme that unlocks rich 3D feature learning from unlabeled scenes. To power the proposed methods, we introduce SceneSplat-7K, the first large-scale 3DGS dataset for indoor scenes, comprising 7916 scenes derived from seven established datasets, such as ScanNet and Matterport3D. Generating SceneSplat-7K required computational resources equivalent to 150 GPU days on an L4 GPU, enabling standardized benchmarking for 3DGS-based reasoning for indoor scenes. Our exhaustive experiments on SceneSplat-7K demonstrate the significant benefit of the proposed method over the established baselines.", "date": "2025-03-23", "categories": []}
{"id": "2506.00083", "title": "Hi-Dyna Graph: Hierarchical Dynamic Scene Graph for Robotic Autonomy in Human-Centric Environments", "abstract": "Autonomous operation of service robotics in human-centric scenes remains challenging due to the need for understanding of changing environments and context-aware decision-making. While existing approaches like topological maps offer efficient spatial priors, they fail to model transient object relationships, whereas dense neural representations (e.g., NeRF) incur prohibitive computational costs. Inspired by the hierarchical scene representation and video scene graph generation works, we propose Hi-Dyna Graph, a hierarchical dynamic scene graph architecture that integrates persistent global layouts with localized dynamic semantics for embodied robotic autonomy. Our framework constructs a global topological graph from posed RGB-D inputs, encoding room-scale connectivity and large static objects (e.g., furniture), while environmental and egocentric cameras populate dynamic subgraphs with object position relations and human-object interaction patterns. A hybrid architecture is conducted by anchoring these subgraphs to the global topology using semantic and spatial constraints, enabling seamless updates as the environment evolves. An agent powered by large language models (LLMs) is employed to interpret the unified graph, infer latent task triggers, and generate executable instructions grounded in robotic affordances. We conduct complex experiments to demonstrate Hi-Dyna Grap's superior scene representation effectiveness. Real-world deployments validate the system's practicality with a mobile manipulator: robotics autonomously complete complex tasks with no further training or complex rewarding in a dynamic scene as cafeteria assistant. See https://anonymous.4open.science/r/Hi-Dyna-Graph-B326 for video demonstration and more details.", "date": "2025-05-30", "categories": []}
{"id": "2506.01379", "title": "RadarSplat: Radar Gaussian Splatting for High-Fidelity Data Synthesis and 3D Reconstruction of Autonomous Driving Scenes", "abstract": "High-Fidelity 3D scene reconstruction plays a crucial role in autonomous driving by enabling novel data generation from existing datasets. This allows simulating safety-critical scenarios and augmenting training datasets without incurring further data collection costs. While recent advances in radiance fields have demonstrated promising results in 3D reconstruction and sensor data synthesis using cameras and LiDAR, their potential for radar remains largely unexplored. Radar is crucial for autonomous driving due to its robustness in adverse weather conditions like rain, fog, and snow, where optical sensors often struggle. Although the state-of-the-art radar-based neural representation shows promise for 3D driving scene reconstruction, it performs poorly in scenarios with significant radar noise, including receiver saturation and multipath reflection. Moreover, it is limited to synthesizing preprocessed, noise-excluded radar images, failing to address realistic radar data synthesis. To address these limitations, this paper proposes RadarSplat, which integrates Gaussian Splatting with novel radar noise modeling to enable realistic radar data synthesis and enhanced 3D reconstruction. Compared to the state-of-the-art, RadarSplat achieves superior radar image synthesis (+3.4 PSNR / 2.6x SSIM) and improved geometric reconstruction (-40% RMSE / 1.5x Accuracy), demonstrating its effectiveness in generating high-fidelity radar data and scene reconstruction. A project page is available at https://umautobots.github.io/radarsplat.", "date": "2025-06-02", "categories": []}
{"id": "2505.24746", "title": "Tackling View-Dependent Semantics in 3D Language Gaussian Splatting", "abstract": "Recent advancements in 3D Gaussian Splatting (3D-GS) enable high-quality 3D scene reconstruction from RGB images. Many studies extend this paradigm for language-driven open-vocabulary scene understanding. However, most of them simply project 2D semantic features onto 3D Gaussians and overlook a fundamental gap between 2D and 3D understanding: a 3D object may exhibit various semantics from different viewpoints--a phenomenon we term view-dependent semantics. To address this challenge, we propose LaGa (Language Gaussians), which establishes cross-view semantic connections by decomposing the 3D scene into objects. Then, it constructs view-aggregated semantic representations by clustering semantic descriptors and reweighting them based on multi-view semantics. Extensive experiments demonstrate that LaGa effectively captures key information from view-dependent semantics, enabling a more comprehensive understanding of 3D scenes. Notably, under the same settings, LaGa achieves a significant improvement of +18.7% mIoU over the previous SOTA on the LERF-OVS dataset. Our code is available at: https://github.com/SJTU-DeepVisionLab/LaGa.", "date": "2025-05-30", "categories": []}
{"id": "2411.16310", "title": "Functionality understanding and segmentation in 3D scenes", "abstract": "Understanding functionalities in 3D scenes involves interpreting natural language descriptions to locate functional interactive objects, such as handles and buttons, in a 3D environment. Functionality understanding is highly challenging, as it requires both world knowledge to interpret language and spatial perception to identify fine-grained objects. For example, given a task like 'turn on the ceiling light', an embodied AI agent must infer that it needs to locate the light switch, even though the switch is not explicitly mentioned in the task description. To date, no dedicated methods have been developed for this problem. In this paper, we introduce Fun3DU, the first approach designed for functionality understanding in 3D scenes. Fun3DU uses a language model to parse the task description through Chain-of-Thought reasoning in order to identify the object of interest. The identified object is segmented across multiple views of the captured scene by using a vision and language model. The segmentation results from each view are lifted in 3D and aggregated into the point cloud using geometric information. Fun3DU is training-free, relying entirely on pre-trained models. We evaluate Fun3DU on SceneFun3D, the most recent and only dataset to benchmark this task, which comprises over 3000 task descriptions on 230 scenes. Our method significantly outperforms state-of-the-art open-vocabulary 3D segmentation approaches. Project page: https://tev-fbk.github.io/fun3du/", "date": "2024-11-25", "categories": []}
{"id": "2502.00954", "title": "Hypo3D: Exploring Hypothetical Reasoning in 3D", "abstract": "The rise of vision-language foundation models marks an advancement in bridging the gap between human and machine capabilities in 3D scene reasoning. Existing 3D reasoning benchmarks assume real-time scene accessibility, which is impractical due to the high cost of frequent scene updates. To this end, we introduce Hypothetical 3D Reasoning, namely Hypo3D, a benchmark designed to evaluate models' ability to reason without access to real-time scene data. Models need to imagine the scene state based on a provided change description before reasoning. Hypo3D is formulated as a 3D Visual Question Answering (VQA) benchmark, comprising 7,727 context changes across 700 indoor scenes, resulting in 14,885 question-answer pairs. An anchor-based world frame is established for all scenes, ensuring consistent reference to a global frame for directional terms in context changes and QAs. Extensive experiments show that state-of-the-art foundation models struggle to reason in hypothetically changed scenes. This reveals a substantial performance gap compared to humans, particularly in scenarios involving movement changes and directional reasoning. Even when the context change is irrelevant to the question, models often incorrectly adjust their answers. Project website: https://matchlab-imperial.github.io/Hypo3D/", "date": "2025-02-02", "categories": []}
{"id": "2505.22429", "title": "Zero-Shot 3D Visual Grounding from Vision-Language Models", "abstract": "3D Visual Grounding (3DVG) seeks to locate target objects in 3D scenes using natural language descriptions, enabling downstream applications such as augmented reality and robotics. Existing approaches typically rely on labeled 3D data and predefined categories, limiting scalability to open-world settings. We present SeeGround, a zero-shot 3DVG framework that leverages 2D Vision-Language Models (VLMs) to bypass the need for 3D-specific training. To bridge the modality gap, we introduce a hybrid input format that pairs query-aligned rendered views with spatially enriched textual descriptions. Our framework incorporates two core components: a Perspective Adaptation Module that dynamically selects optimal viewpoints based on the query, and a Fusion Alignment Module that integrates visual and spatial signals to enhance localization precision. Extensive evaluations on ScanRefer and Nr3D confirm that SeeGround achieves substantial improvements over existing zero-shot baselines -- outperforming them by 7.7% and 7.1%, respectively -- and even rivals fully supervised alternatives, demonstrating strong generalization under challenging conditions.", "date": "2025-05-28", "categories": []}
{"id": "2505.20610", "title": "OmniIndoor3D: Comprehensive Indoor 3D Reconstruction", "abstract": "We propose a novel framework for comprehensive indoor 3D reconstruction using Gaussian representations, called OmniIndoor3D. This framework enables accurate appearance, geometry, and panoptic reconstruction of diverse indoor scenes captured by a consumer-level RGB-D camera. Since 3DGS is primarily optimized for photorealistic rendering, it lacks the precise geometry critical for high-quality panoptic reconstruction. Therefore, OmniIndoor3D first combines multiple RGB-D images to create a coarse 3D reconstruction, which is then used to initialize the 3D Gaussians and guide the 3DGS training. To decouple the optimization conflict between appearance and geometry, we introduce a lightweight MLP that adjusts the geometric properties of 3D Gaussians. The introduced lightweight MLP serves as a low-pass filter for geometry reconstruction and significantly reduces noise in indoor scenes. To improve the distribution of Gaussian primitives, we propose a densification strategy guided by panoptic priors to encourage smoothness on planar surfaces. Through the joint optimization of appearance, geometry, and panoptic reconstruction, OmniIndoor3D provides comprehensive 3D indoor scene understanding, which facilitates accurate and robust robotic navigation. We perform thorough evaluations across multiple datasets, and OmniIndoor3D achieves state-of-the-art results in appearance, geometry, and panoptic reconstruction. We believe our work bridges a critical gap in indoor 3D reconstruction. The code will be released at: https://ucwxb.github.io/OmniIndoor3D/", "date": "2025-05-27", "categories": []}
{"id": "2503.22436", "title": "NuGrounding: A Multi-View 3D Visual Grounding Framework in Autonomous Driving", "abstract": "Multi-view 3D visual grounding is critical for autonomous driving vehicles to interpret natural languages and localize target objects in complex environments. However, existing datasets and methods suffer from coarse-grained language instructions, and inadequate integration of 3D geometric reasoning with linguistic comprehension. To this end, we introduce NuGrounding, the first large-scale benchmark for multi-view 3D visual grounding in autonomous driving. We present a Hierarchy of Grounding (HoG) method to construct NuGrounding to generate hierarchical multi-level instructions, ensuring comprehensive coverage of human instruction patterns. To tackle this challenging dataset, we propose a novel paradigm that seamlessly combines instruction comprehension abilities of multi-modal LLMs (MLLMs) with precise localization abilities of specialist detection models. Our approach introduces two decoupled task tokens and a context query to aggregate 3D geometric information and semantic instructions, followed by a fusion decoder to refine spatial-semantic feature fusion for precise localization. Extensive experiments demonstrate that our method significantly outperforms the baselines adapted from representative 3D scene understanding methods by a significant margin and achieves 0.59 in precision and 0.64 in recall, with improvements of 50.8% and 54.7%.", "date": "2025-03-28", "categories": []}
{"id": "2505.19154", "title": "FHGS: Feature-Homogenized Gaussian Splatting", "abstract": "Scene understanding based on 3D Gaussian Splatting (3DGS) has recently achieved notable advances. Although 3DGS related methods have efficient rendering capabilities, they fail to address the inherent contradiction between the anisotropic color representation of gaussian primitives and the isotropic requirements of semantic features, leading to insufficient cross-view feature consistency. To overcome the limitation, we proposes $\\textit{FHGS}$ (Feature-Homogenized Gaussian Splatting), a novel 3D feature fusion framework inspired by physical models, which can achieve high-precision mapping of arbitrary 2D features from pre-trained models to 3D scenes while preserving the real-time rendering efficiency of 3DGS. Specifically, our $\\textit{FHGS}$ introduces the following innovations: Firstly, a universal feature fusion architecture is proposed, enabling robust embedding of large-scale pre-trained models' semantic features (e.g., SAM, CLIP) into sparse 3D structures. Secondly, a non-differentiable feature fusion mechanism is introduced, which enables semantic features to exhibit viewpoint independent isotropic distributions. This fundamentally balances the anisotropic rendering of gaussian primitives and the isotropic expression of features; Thirdly, a dual-driven optimization strategy inspired by electric potential fields is proposed, which combines external supervision from semantic feature fields with internal primitive clustering guidance. This mechanism enables synergistic optimization of global semantic alignment and local structural consistency. More interactive results can be accessed on: https://fhgs.cuastro.org/.", "date": "2025-05-25", "categories": []}
{"id": "2505.17402", "title": "From Flight to Insight: Semantic 3D Reconstruction for Aerial Inspection via Gaussian Splatting and Language-Guided Segmentation", "abstract": "High-fidelity 3D reconstruction is critical for aerial inspection tasks such as infrastructure monitoring, structural assessment, and environmental surveying. While traditional photogrammetry techniques enable geometric modeling, they lack semantic interpretability, limiting their effectiveness for automated inspection workflows. Recent advances in neural rendering and 3D Gaussian Splatting (3DGS) offer efficient, photorealistic reconstructions but similarly lack scene-level understanding.\n  In this work, we present a UAV-based pipeline that extends Feature-3DGS for language-guided 3D segmentation. We leverage LSeg-based feature fields with CLIP embeddings to generate heatmaps in response to language prompts. These are thresholded to produce rough segmentations, and the highest-scoring point is then used as a prompt to SAM or SAM2 for refined 2D segmentation on novel view renderings. Our results highlight the strengths and limitations of various feature field backbones (CLIP-LSeg, SAM, SAM2) in capturing meaningful structure in large-scale outdoor environments. We demonstrate that this hybrid approach enables flexible, language-driven interaction with photorealistic 3D reconstructions, opening new possibilities for semantic aerial inspection and scene understanding.", "date": "2025-05-23", "categories": []}
{"id": "2408.15966", "title": "More Text, Less Point: Towards 3D Data-Efficient Point-Language Understanding", "abstract": "Enabling Large Language Models (LLMs) to comprehend the 3D physical world remains a significant challenge. Due to the lack of large-scale 3D-text pair datasets, the success of LLMs has yet to be replicated in 3D understanding. In this paper, we rethink this issue and propose a new task: 3D Data-Efficient Point-Language Understanding. The goal is to enable LLMs to achieve robust 3D object understanding with minimal 3D point cloud and text data pairs. To address this task, we introduce GreenPLM, which leverages more text data to compensate for the lack of 3D data. First, inspired by using CLIP to align images and text, we utilize a pre-trained point cloud-text encoder to map the 3D point cloud space to the text space. This mapping leaves us to seamlessly connect the text space with LLMs. Once the point-text-LLM connection is established, we further enhance text-LLM alignment by expanding the intermediate text space, thereby reducing the reliance on 3D point cloud data. Specifically, we generate 6M free-text descriptions of 3D objects, and design a three-stage training strategy to help LLMs better explore the intrinsic connections between different modalities. To achieve efficient modality alignment, we design a zero-parameter cross-attention module for token pooling. Extensive experimental results show that GreenPLM requires only 12% of the 3D training data used by existing state-of-the-art models to achieve superior 3D understanding. Remarkably, GreenPLM also achieves competitive performance using text-only data. The code and weights are available at: https://github.com/TangYuan96/GreenPLM.", "date": "2024-08-28", "categories": []}
{"id": "2505.16399", "title": "Sketchy Bounding-box Supervision for 3D Instance Segmentation", "abstract": "Bounding box supervision has gained considerable attention in weakly supervised 3D instance segmentation. While this approach alleviates the need for extensive point-level annotations, obtaining accurate bounding boxes in practical applications remains challenging. To this end, we explore the inaccurate bounding box, named sketchy bounding box, which is imitated through perturbing ground truth bounding box by adding scaling, translation, and rotation. In this paper, we propose Sketchy-3DIS, a novel weakly 3D instance segmentation framework, which jointly learns pseudo labeler and segmentator to improve the performance under the sketchy bounding-box supervisions. Specifically, we first propose an adaptive box-to-point pseudo labeler that adaptively learns to assign points located in the overlapped parts between two sketchy bounding boxes to the correct instance, resulting in compact and pure pseudo instance labels. Then, we present a coarse-to-fine instance segmentator that first predicts coarse instances from the entire point cloud and then learns fine instances based on the region of coarse instances. Finally, by using the pseudo instance labels to supervise the instance segmentator, we can gradually generate high-quality instances through joint training. Extensive experiments show that our method achieves state-of-the-art performance on both the ScanNetV2 and S3DIS benchmarks, and even outperforms several fully supervised methods using sketchy bounding boxes. Code is available at https://github.com/dengq7/Sketchy-3DIS.", "date": "2025-05-22", "categories": []}
{"id": "2505.15232", "title": "DC-Scene: Data-Centric Learning for 3D Scene Understanding", "abstract": "3D scene understanding plays a fundamental role in vision applications such as robotics, autonomous driving, and augmented reality. However, advancing learning-based 3D scene understanding remains challenging due to two key limitations: (1) the large scale and complexity of 3D scenes lead to higher computational costs and slower training compared to 2D counterparts; and (2) high-quality annotated 3D datasets are significantly scarcer than those available for 2D vision. These challenges underscore the need for more efficient learning paradigms. In this work, we propose DC-Scene, a data-centric framework tailored for 3D scene understanding, which emphasizes enhancing data quality and training efficiency. Specifically, we introduce a CLIP-driven dual-indicator quality (DIQ) filter, combining vision-language alignment scores with caption-loss perplexity, along with a curriculum scheduler that progressively expands the training pool from the top 25% to 75% of scene-caption pairs. This strategy filters out noisy samples and significantly reduces dependence on large-scale labeled 3D data. Extensive experiments on ScanRefer and Nr3D demonstrate that DC-Scene achieves state-of-the-art performance (86.1 CIDEr with the top-75% subset vs. 85.4 with the full dataset) while reducing training cost by approximately two-thirds, confirming that a compact set of high-quality samples can outperform exhaustive training. Code will be available at https://github.com/AIGeeksGroup/DC-Scene.", "date": "2025-05-21", "categories": []}
{"id": "2505.15373", "title": "RAZER: Robust Accelerated Zero-Shot 3D Open-Vocabulary Panoptic Reconstruction with Spatio-Temporal Aggregation", "abstract": "Mapping and understanding complex 3D environments is fundamental to how autonomous systems perceive and interact with the physical world, requiring both precise geometric reconstruction and rich semantic comprehension. While existing 3D semantic mapping systems excel at reconstructing and identifying predefined object instances, they lack the flexibility to efficiently build semantic maps with open-vocabulary during online operation. Although recent vision-language models have enabled open-vocabulary object recognition in 2D images, they haven't yet bridged the gap to 3D spatial understanding. The critical challenge lies in developing a training-free unified system that can simultaneously construct accurate 3D maps while maintaining semantic consistency and supporting natural language interactions in real time. In this paper, we develop a zero-shot framework that seamlessly integrates GPU-accelerated geometric reconstruction with open-vocabulary vision-language models through online instance-level semantic embedding fusion, guided by hierarchical object association with spatial indexing. Our training-free system achieves superior performance through incremental processing and unified geometric-semantic updates, while robustly handling 2D segmentation inconsistencies. The proposed general-purpose 3D scene understanding framework can be used for various tasks including zero-shot 3D instance retrieval, segmentation, and object detection to reason about previously unseen objects and interpret natural language queries. The project page is available at https://razer-3d.github.io.", "date": "2025-05-21", "categories": []}
{"id": "2403.08504", "title": "Offboard Occupancy Refinement with Hybrid Propagation for Autonomous Driving", "abstract": "Vision-based occupancy prediction, also known as 3D Semantic Scene Completion (SSC), presents a significant challenge in computer vision. Previous methods, confined to onboard processing, struggle with simultaneous geometric and semantic estimation, continuity across varying viewpoints, and single-view occlusion. Our paper introduces OccFiner, a novel offboard framework designed to enhance the accuracy of vision-based occupancy predictions. OccFiner operates in two hybrid phases: 1) a multi-to-multi local propagation network that implicitly aligns and processes multiple local frames for correcting onboard model errors and consistently enhancing occupancy accuracy across all distances. 2) the region-centric global propagation, focuses on refining labels using explicit multi-view geometry and integrating sensor bias, particularly for increasing the accuracy of distant occupied voxels. Extensive experiments demonstrate that OccFiner improves both geometric and semantic accuracy across various types of coarse occupancy, setting a new state-of-the-art performance on the SemanticKITTI dataset. Notably, OccFiner significantly boosts the performance of vision-based SSC models, achieving accuracy levels competitive with established LiDAR-based onboard SSC methods. Furthermore, OccFiner is the first to achieve automatic annotation of SSC in a purely vision-based approach. Quantitative experiments prove that OccFiner successfully facilitates occupancy data loop-closure in autonomous driving. Additionally, we quantitatively and qualitatively validate the superiority of the offboard approach on city-level SSC static maps. The source code will be made publicly available at https://github.com/MasterHow/OccFiner.", "date": "2024-03-13", "categories": []}
{"id": "2411.19551", "title": "Bootstraping Clustering of Gaussians for View-consistent 3D Scene Understanding", "abstract": "Injecting semantics into 3D Gaussian Splatting (3DGS) has recently garnered significant attention. While current approaches typically distill 3D semantic features from 2D foundational models (e.g., CLIP and SAM) to facilitate novel view segmentation and semantic understanding, their heavy reliance on 2D supervision can undermine cross-view semantic consistency and necessitate complex data preparation processes, therefore hindering view-consistent scene understanding. In this work, we present FreeGS, an unsupervised semantic-embedded 3DGS framework that achieves view-consistent 3D scene understanding without the need for 2D labels. Instead of directly learning semantic features, we introduce the IDentity-coupled Semantic Field (IDSF) into 3DGS, which captures both semantic representations and view-consistent instance indices for each Gaussian. We optimize IDSF with a two-step alternating strategy: semantics help to extract coherent instances in 3D space, while the resulting instances regularize the injection of stable semantics from 2D space. Additionally, we adopt a 2D-3D joint contrastive loss to enhance the complementarity between view-consistent 3D geometry and rich semantics during the bootstrapping process, enabling FreeGS to uniformly perform tasks such as novel-view semantic segmentation, object selection, and 3D object detection. Extensive experiments on LERF-Mask, 3D-OVS, and ScanNet datasets demonstrate that FreeGS performs comparably to state-of-the-art methods while avoiding the complex data preprocessing workload. Our code is publicly available at https://github.com/wb014/FreeGS.", "date": "2024-11-29", "categories": []}
{"id": "2504.13580", "title": "Leveraging Automatic CAD Annotations for Supervised Learning in 3D Scene Understanding", "abstract": "High-level 3D scene understanding is essential in many applications. However, the challenges of generating accurate 3D annotations make development of deep learning models difficult. We turn to recent advancements in automatic retrieval of synthetic CAD models, and show that data generated by such methods can be used as high-quality ground truth for training supervised deep learning models. More exactly, we employ a pipeline akin to the one previously used to automatically annotate objects in ScanNet scenes with their 9D poses and CAD models. This time, we apply it to the recent ScanNet++ v1 dataset, which previously lacked such annotations. Our findings demonstrate that it is not only possible to train deep learning models on these automatically-obtained annotations but that the resulting models outperform those trained on manually annotated data. We validate this on two distinct tasks: point cloud completion and single-view CAD model retrieval and alignment. Our results underscore the potential of automatic 3D annotations to enhance model performance while significantly reducing annotation costs. To support future research in 3D scene understanding, we will release our annotations, which we call SCANnotate++, along with our trained models.", "date": "2025-04-18", "categories": []}
{"id": "2505.09562", "title": "Camera-Only 3D Panoptic Scene Completion for Autonomous Driving through Differentiable Object Shapes", "abstract": "Autonomous vehicles need a complete map of their surroundings to plan and act. This has sparked research into the tasks of 3D occupancy prediction, 3D scene completion, and 3D panoptic scene completion, which predict a dense map of the ego vehicle's surroundings as a voxel grid. Scene completion extends occupancy prediction by predicting occluded regions of the voxel grid, and panoptic scene completion further extends this task by also distinguishing object instances within the same class; both aspects are crucial for path planning and decision-making. However, 3D panoptic scene completion is currently underexplored. This work introduces a novel framework for 3D panoptic scene completion that extends existing 3D semantic scene completion models. We propose an Object Module and Panoptic Module that can easily be integrated with 3D occupancy and scene completion methods presented in the literature. Our approach leverages the available annotations in occupancy benchmarks, allowing individual object shapes to be learned as a differentiable problem. The code is available at https://github.com/nicolamarinello/OffsetOcc .", "date": "2025-05-14", "categories": []}
{"id": "2505.05512", "title": "Occupancy World Model for Robots", "abstract": "Understanding and forecasting the scene evolutions deeply affect the exploration and decision of embodied agents. While traditional methods simulate scene evolutions through trajectory prediction of potential instances, current works use the occupancy world model as a generative framework for describing fine-grained overall scene dynamics. However, existing methods cluster on the outdoor structured road scenes, while ignoring the exploration of forecasting 3D occupancy scene evolutions for robots in indoor scenes. In this work, we explore a new framework for learning the scene evolutions of observed fine-grained occupancy and propose an occupancy world model based on the combined spatio-temporal receptive field and guided autoregressive transformer to forecast the scene evolutions, called RoboOccWorld. We propose the Conditional Causal State Attention (CCSA), which utilizes camera poses of next state as conditions to guide the autoregressive transformer to adapt and understand the indoor robotics scenarios. In order to effectively exploit the spatio-temporal cues from historical observations, Hybrid Spatio-Temporal Aggregation (HSTA) is proposed to obtain the combined spatio-temporal receptive field based on multi-scale spatio-temporal windows. In addition, we restructure the OccWorld-ScanNet benchmark based on local annotations to facilitate the evaluation of the indoor 3D occupancy scene evolution prediction task. Experimental results demonstrate that our RoboOccWorld outperforms state-of-the-art methods in indoor 3D occupancy scene evolution prediction task. The code will be released soon.", "date": "2025-05-07", "categories": []}
{"id": "2409.03757", "title": "Lexicon3D: Probing Visual Foundation Models for Complex 3D Scene\n  Understanding", "abstract": "  Complex 3D scene understanding has gained increasing attention, with scene\nencoding strategies playing a crucial role in this success. However, the\noptimal scene encoding strategies for various scenarios remain unclear,\nparticularly compared to their image-based counterparts. To address this issue,\nwe present a comprehensive study that probes various visual encoding models for\n3D scene understanding, identifying the strengths and limitations of each model\nacross different scenarios. Our evaluation spans seven vision foundation\nencoders, including image-based, video-based, and 3D foundation models. We\nevaluate these models in four tasks: Vision-Language Scene Reasoning, Visual\nGrounding, Segmentation, and Registration, each focusing on different aspects\nof scene understanding. Our evaluations yield key findings: DINOv2 demonstrates\nsuperior performance, video models excel in object-level tasks, diffusion\nmodels benefit geometric tasks, and language-pretrained models show unexpected\nlimitations in language-related tasks. These insights challenge some\nconventional understandings, provide novel perspectives on leveraging visual\nfoundation models, and highlight the need for more flexible encoder selection\nin future vision-language and scene-understanding tasks. Code:\nhttps://github.com/YunzeMan/Lexicon3D\n", "date": "2024-09-05", "categories": []}
{"id": "2505.04911", "title": "SpatialPrompting: Keyframe-driven Zero-Shot Spatial Reasoning with\n  Off-the-Shelf Multimodal Large Language Models", "abstract": "  This study introduces SpatialPrompting, a novel framework that harnesses the\nemergent reasoning capabilities of off-the-shelf multimodal large language\nmodels to achieve zero-shot spatial reasoning in three-dimensional (3D)\nenvironments. Unlike existing methods that rely on expensive 3D-specific\nfine-tuning with specialized 3D inputs such as point clouds or voxel-based\nfeatures, SpatialPrompting employs a keyframe-driven prompt generation\nstrategy. This framework uses metrics such as vision-language similarity,\nMahalanobis distance, field of view, and image sharpness to select a diverse\nand informative set of keyframes from image sequences and then integrates them\nwith corresponding camera pose data to effectively abstract spatial\nrelationships and infer complex 3D structures. The proposed framework not only\nestablishes a new paradigm for flexible spatial reasoning that utilizes\nintuitive visual and positional cues but also achieves state-of-the-art\nzero-shot performance on benchmark datasets, such as ScanQA and SQA3D, across\nseveral metrics. The proposed method effectively eliminates the need for\nspecialized 3D inputs and fine-tuning, offering a simpler and more scalable\nalternative to conventional approaches.\n", "date": "2025-05-08", "categories": []}
{"id": "2505.05474", "title": "3D Scene Generation: A Survey", "abstract": "  3D scene generation seeks to synthesize spatially structured, semantically\nmeaningful, and photorealistic environments for applications such as immersive\nmedia, robotics, autonomous driving, and embodied AI. Early methods based on\nprocedural rules offered scalability but limited diversity. Recent advances in\ndeep generative models (e.g., GANs, diffusion models) and 3D representations\n(e.g., NeRF, 3D Gaussians) have enabled the learning of real-world scene\ndistributions, improving fidelity, diversity, and view consistency. Recent\nadvances like diffusion models bridge 3D scene synthesis and photorealism by\nreframing generation as image or video synthesis problems. This survey provides\na systematic overview of state-of-the-art approaches, organizing them into four\nparadigms: procedural generation, neural 3D-based generation, image-based\ngeneration, and video-based generation. We analyze their technical foundations,\ntrade-offs, and representative results, and review commonly used datasets,\nevaluation protocols, and downstream applications. We conclude by discussing\nkey challenges in generation capacity, 3D representation, data and annotations,\nand evaluation, and outline promising directions including higher fidelity,\nphysics-aware and interactive generation, and unified perception-generation\nmodels. This review organizes recent advances in 3D scene generation and\nhighlights promising directions at the intersection of generative AI, 3D\nvision, and embodied intelligence. To track ongoing developments, we maintain\nan up-to-date project page:\nhttps://github.com/hzxie/Awesome-3D-Scene-Generation.\n", "date": "2025-05-08", "categories": []}
{"id": "2505.03300", "title": "3D Can Be Explored In 2D: Pseudo-Label Generation for LiDAR Point Clouds\n  Using Sensor-Intensity-Based 2D Semantic Segmentation", "abstract": "  Semantic segmentation of 3D LiDAR point clouds, essential for autonomous\ndriving and infrastructure management, is best achieved by supervised learning,\nwhich demands extensive annotated datasets and faces the problem of domain\nshifts. We introduce a new 3D semantic segmentation pipeline that leverages\naligned scenes and state-of-the-art 2D segmentation methods, avoiding the need\nfor direct 3D annotation or reliance on additional modalities such as camera\nimages at inference time. Our approach generates 2D views from LiDAR scans\ncolored by sensor intensity and applies 2D semantic segmentation to these views\nusing a camera-domain pretrained model. The segmented 2D outputs are then\nback-projected onto the 3D points, with a simple voting-based estimator that\nmerges the labels associated to each 3D point. Our main contribution is a\nglobal pipeline for 3D semantic segmentation requiring no prior 3D annotation\nand not other modality for inference, which can be used for pseudo-label\ngeneration. We conduct a thorough ablation study and demonstrate the potential\nof the generated pseudo-labels for the Unsupervised Domain Adaptation task.\n", "date": "2025-05-06", "categories": []}
{"id": "2503.06222", "title": "Vision-based 3D Semantic Scene Completion via Capture Dynamic\n  Representations", "abstract": "  The vision-based semantic scene completion task aims to predict dense\ngeometric and semantic 3D scene representations from 2D images. However, the\npresence of dynamic objects in the scene seriously affects the accuracy of the\nmodel inferring 3D structures from 2D images. Existing methods simply stack\nmultiple frames of image input to increase dense scene semantic information,\nbut ignore the fact that dynamic objects and non-texture areas violate\nmulti-view consistency and matching reliability. To address these issues, we\npropose a novel method, CDScene: Vision-based Robust Semantic Scene Completion\nvia Capturing Dynamic Representations. First, we leverage a multimodal\nlarge-scale model to extract 2D explicit semantics and align them into 3D\nspace. Second, we exploit the characteristics of monocular and stereo depth to\ndecouple scene information into dynamic and static features. The dynamic\nfeatures contain structural relationships around dynamic objects, and the\nstatic features contain dense contextual spatial information. Finally, we\ndesign a dynamic-static adaptive fusion module to effectively extract and\naggregate complementary features, achieving robust and accurate semantic scene\ncompletion in autonomous driving scenarios. Extensive experimental results on\nthe SemanticKITTI, SSCBench-KITTI360, and SemanticKITTI-C datasets demonstrate\nthe superiority and robustness of CDScene over existing state-of-the-art\nmethods.\n", "date": "2025-03-08", "categories": []}
{"id": "2505.02405", "title": "Estimating Commonsense Scene Composition on Belief Scene Graphs", "abstract": "  This work establishes the concept of commonsense scene composition, with a\nfocus on extending Belief Scene Graphs by estimating the spatial distribution\nof unseen objects. Specifically, the commonsense scene composition capability\nrefers to the understanding of the spatial relationships among related objects\nin the scene, which in this article is modeled as a joint probability\ndistribution for all possible locations of the semantic object class. The\nproposed framework includes two variants of a Correlation Information (CECI)\nmodel for learning probability distributions: (i) a baseline approach based on\na Graph Convolutional Network, and (ii) a neuro-symbolic extension that\nintegrates a spatial ontology based on Large Language Models (LLMs).\nFurthermore, this article provides a detailed description of the dataset\ngeneration process for such tasks. Finally, the framework has been validated\nthrough multiple runs on simulated data, as well as in a real-world indoor\nenvironment, demonstrating its ability to spatially interpret scenes across\ndifferent room types.\n", "date": "2025-05-05", "categories": []}
{"id": "2504.19266", "title": "OpenFusion++: An Open-vocabulary Real-time Scene Understanding System", "abstract": "  Real-time open-vocabulary scene understanding is essential for efficient 3D\nperception in applications such as vision-language navigation, embodied\nintelligence, and augmented reality. However, existing methods suffer from\nimprecise instance segmentation, static semantic updates, and limited handling\nof complex queries. To address these issues, we present OpenFusion++, a\nTSDF-based real-time 3D semantic-geometric reconstruction system. Our approach\nrefines 3D point clouds by fusing confidence maps from foundational models,\ndynamically updates global semantic labels via an adaptive cache based on\ninstance area, and employs a dual-path encoding framework that integrates\nobject attributes with environmental context for precise query responses.\nExperiments on the ICL, Replica, ScanNet, and ScanNet++ datasets demonstrate\nthat OpenFusion++ significantly outperforms the baseline in both semantic\naccuracy and query responsiveness.\n", "date": "2025-04-27", "categories": []}
{"id": "2504.19500", "title": "Masked Point-Entity Contrast for Open-Vocabulary 3D Scene Understanding", "abstract": "  Open-vocabulary 3D scene understanding is pivotal for enhancing physical\nintelligence, as it enables embodied agents to interpret and interact\ndynamically within real-world environments. This paper introduces MPEC, a novel\nMasked Point-Entity Contrastive learning method for open-vocabulary 3D semantic\nsegmentation that leverages both 3D entity-language alignment and point-entity\nconsistency across different point cloud views to foster entity-specific\nfeature representations. Our method improves semantic discrimination and\nenhances the differentiation of unique instances, achieving state-of-the-art\nresults on ScanNet for open-vocabulary 3D semantic segmentation and\ndemonstrating superior zero-shot scene understanding capabilities. Extensive\nfine-tuning experiments on 8 datasets, spanning from low-level perception to\nhigh-level reasoning tasks, showcase the potential of learned 3D features,\ndriving consistent performance gains across varied 3D scene understanding\ntasks. Project website: https://mpec-3d.github.io/\n", "date": "2025-04-28", "categories": []}
{"id": "2504.19749", "title": "STCOcc: Sparse Spatial-Temporal Cascade Renovation for 3D Occupancy and\n  Scene Flow Prediction", "abstract": "  3D occupancy and scene flow offer a detailed and dynamic representation of 3D\nscene. Recognizing the sparsity and complexity of 3D space, previous\nvision-centric methods have employed implicit learning-based approaches to\nmodel spatial and temporal information. However, these approaches struggle to\ncapture local details and diminish the model's spatial discriminative ability.\nTo address these challenges, we propose a novel explicit state-based modeling\nmethod designed to leverage the occupied state to renovate the 3D features.\nSpecifically, we propose a sparse occlusion-aware attention mechanism,\nintegrated with a cascade refinement strategy, which accurately renovates 3D\nfeatures with the guidance of occupied state information. Additionally, we\nintroduce a novel method for modeling long-term dynamic interactions, which\nreduces computational costs and preserves spatial information. Compared to the\nprevious state-of-the-art methods, our efficient explicit renovation strategy\nnot only delivers superior performance in terms of RayIoU and mAVE for\noccupancy and scene flow prediction but also markedly reduces GPU memory usage\nduring training, bringing it down to 8.7GB. Our code is available on\nhttps://github.com/lzzzzzm/STCOcc\n", "date": "2025-04-28", "categories": []}
{"id": "2504.18380", "title": "Spatial Reasoner: A 3D Inference Pipeline for XR Applications", "abstract": "  Modern extended reality XR systems provide rich analysis of image data and\nfusion of sensor input and demand AR/VR applications that can reason about 3D\nscenes in a semantic manner. We present a spatial reasoning framework that\nbridges geometric facts with symbolic predicates and relations to handle key\ntasks such as determining how 3D objects are arranged among each other ('on',\n'behind', 'near', etc.). Its foundation relies on oriented 3D bounding box\nrepresentations, enhanced by a comprehensive set of spatial predicates, ranging\nfrom topology and connectivity to directionality and orientation, expressed in\na formalism related to natural language. The derived predicates form a spatial\nknowledge graph and, in combination with a pipeline-based inference model,\nenable spatial queries and dynamic rule evaluation. Implementations for client-\nand server-side processing demonstrate the framework's capability to\nefficiently translate geometric data into actionable knowledge, ensuring\nscalable and technology-independent spatial reasoning in complex 3D\nenvironments. The Spatial Reasoner framework is fostering the creation of\nspatial ontologies, and seamlessly integrates with and therefore enriches\nmachine learning, natural language processing, and rule systems in XR\napplications.\n", "date": "2025-04-25", "categories": []}
{"id": "2501.01163", "title": "3D-LLaVA: Towards Generalist 3D LMMs with Omni Superpoint Transformer", "abstract": "  Current 3D Large Multimodal Models (3D LMMs) have shown tremendous potential\nin 3D-vision-based dialogue and reasoning. However, how to further enhance 3D\nLMMs to achieve fine-grained scene understanding and facilitate flexible\nhuman-agent interaction remains a challenging problem. In this work, we\nintroduce 3D-LLaVA, a simple yet highly powerful 3D LMM designed to act as an\nintelligent assistant in comprehending, reasoning, and interacting with the 3D\nworld. Unlike existing top-performing methods that rely on complicated\npipelines-such as offline multi-view feature extraction or additional\ntask-specific heads-3D-LLaVA adopts a minimalist design with integrated\narchitecture and only takes point clouds as input. At the core of 3D-LLaVA is a\nnew Omni Superpoint Transformer (OST), which integrates three functionalities:\n(1) a visual feature selector that converts and selects visual tokens, (2) a\nvisual prompt encoder that embeds interactive visual prompts into the visual\ntoken space, and (3) a referring mask decoder that produces 3D masks based on\ntext description. This versatile OST is empowered by the hybrid pretraining to\nobtain perception priors and leveraged as the visual connector that bridges the\n3D data to the LLM. After performing unified instruction tuning, our 3D-LLaVA\nreports impressive results on various benchmarks.\n", "date": "2025-01-02", "categories": []}
{"id": "2404.06836", "title": "O2V-Mapping: Online Open-Vocabulary Mapping with Neural Implicit\n  Representation", "abstract": "  Online construction of open-ended language scenes is crucial for robotic\napplications, where open-vocabulary interactive scene understanding is\nrequired. Recently, neural implicit representation has provided a promising\ndirection for online interactive mapping. However, implementing open-vocabulary\nscene understanding capability into online neural implicit mapping still faces\nthree challenges: lack of local scene updating ability, blurry spatial\nhierarchical semantic segmentation and difficulty in maintaining multi-view\nconsistency. To this end, we proposed O2V-mapping, which utilizes voxel-based\nlanguage and geometric features to create an open-vocabulary field, thus\nallowing for local updates during online training process. Additionally, we\nleverage a foundational model for image segmentation to extract language\nfeatures on object-level entities, achieving clear segmentation boundaries and\nhierarchical semantic features. For the purpose of preserving consistency in 3D\nobject properties across different viewpoints, we propose a spatial adaptive\nvoxel adjustment mechanism and a multi-view weight selection method. Extensive\nexperiments on open-vocabulary object localization and semantic segmentation\ndemonstrate that O2V-mapping achieves online construction of language scenes\nwhile enhancing accuracy, outperforming the previous SOTA method.\n", "date": "2024-04-10", "categories": []}
{"id": "2405.06945", "title": "Direct Learning of Mesh and Appearance via 3D Gaussian Splatting", "abstract": "  Accurately reconstructing a 3D scene including explicit geometry information\nis both attractive and challenging. Geometry reconstruction can benefit from\nincorporating differentiable appearance models, such as Neural Radiance Fields\nand 3D Gaussian Splatting (3DGS). However, existing methods encounter\nefficiency issues due to indirect geometry learning and the paradigm of\nseparately modeling geometry and surface appearance. In this work, we propose a\nlearnable scene model that incorporates 3DGS with an explicit geometry\nrepresentation, namely a mesh. Our model learns the mesh and appearance in an\nend-to-end manner, where we bind 3D Gaussians to the mesh faces and perform\ndifferentiable rendering of 3DGS to obtain photometric supervision. The model\ncreates an effective information pathway to supervise the learning of both 3DGS\nand mesh. Experimental results demonstrate that the learned scene model not\nonly improves efficiency and rendering quality but also enables manipulation\nvia the explicit mesh. In addition, our model has a unique advantage in\nadapting to scene updates, thanks to the end-to-end learning of both mesh and\nappearance.\n", "date": "2024-05-11", "categories": []}
{"id": "2504.12959", "title": "Rethinking Temporal Fusion with a Unified Gradient Descent View for 3D\n  Semantic Occupancy Prediction", "abstract": "  We present GDFusion, a temporal fusion method for vision-based 3D semantic\noccupancy prediction (VisionOcc). GDFusion opens up the underexplored aspects\nof temporal fusion within the VisionOcc framework, focusing on both temporal\ncues and fusion strategies. It systematically examines the entire VisionOcc\npipeline, identifying three fundamental yet previously overlooked temporal\ncues: scene-level consistency, motion calibration, and geometric\ncomplementation. These cues capture diverse facets of temporal evolution and\nmake distinct contributions across various modules in the VisionOcc framework.\nTo effectively fuse temporal signals across heterogeneous representations, we\npropose a novel fusion strategy by reinterpreting the formulation of vanilla\nRNNs. This reinterpretation leverages gradient descent on features to unify the\nintegration of diverse temporal information, seamlessly embedding the proposed\ntemporal cues into the network. Extensive experiments on nuScenes demonstrate\nthat GDFusion significantly outperforms established baselines. Notably, on\nOcc3D benchmark, it achieves 1.4\\%-4.8\\% mIoU improvements and reduces memory\nconsumption by 27\\%-72\\%.\n", "date": "2025-04-17", "categories": []}
{"id": "2504.13579", "title": "HDBFormer: Efficient RGB-D Semantic Segmentation with A Heterogeneous\n  Dual-Branch Framework", "abstract": "  In RGB-D semantic segmentation for indoor scenes, a key challenge is\neffectively integrating the rich color information from RGB images with the\nspatial distance information from depth images. However, most existing methods\noverlook the inherent differences in how RGB and depth images express\ninformation. Properly distinguishing the processing of RGB and depth images is\nessential to fully exploiting their unique and significant characteristics. To\naddress this, we propose a novel heterogeneous dual-branch framework called\nHDBFormer, specifically designed to handle these modality differences. For RGB\nimages, which contain rich detail, we employ both a basic and detail encoder to\nextract local and global features. For the simpler depth images, we propose\nLDFormer, a lightweight hierarchical encoder that efficiently extracts depth\nfeatures with fewer parameters. Additionally, we introduce the Modality\nInformation Interaction Module (MIIM), which combines transformers with large\nkernel convolutions to interact global and local information across modalities\nefficiently. Extensive experiments show that HDBFormer achieves\nstate-of-the-art performance on the NYUDepthv2 and SUN-RGBD datasets. The code\nis available at: https://github.com/Weishuobin/HDBFormer.\n", "date": "2025-04-18", "categories": []}
{"id": "2504.13590", "title": "HAECcity: Open-Vocabulary Scene Understanding of City-Scale Point Clouds\n  with Superpoint Graph Clustering", "abstract": "  Traditional 3D scene understanding techniques are generally predicated on\nhand-annotated label sets, but in recent years a new class of open-vocabulary\n3D scene understanding techniques has emerged. Despite the success of this\nparadigm on small scenes, existing approaches cannot scale efficiently to\ncity-scale 3D datasets. In this paper, we present Hierarchical vocab-Agnostic\nExpert Clustering (HAEC), after the latin word for 'these', a superpoint graph\nclustering based approach which utilizes a novel mixture of experts graph\ntransformer for its backbone. We administer this highly scalable approach to\nthe first application of open-vocabulary scene understanding on the SensatUrban\ncity-scale dataset. We also demonstrate a synthetic labeling pipeline which is\nderived entirely from the raw point clouds with no hand-annotation. Our\ntechnique can help unlock complex operations on dense urban 3D scenes and open\na new path forward in the processing of digital twins.\n", "date": "2025-04-18", "categories": []}
{"id": "2501.17547", "title": "Towards Training-Free Open-World Classification with 3D Generative\n  Models", "abstract": "  3D open-world classification is a challenging yet essential task in dynamic\nand unstructured real-world scenarios, requiring both open-category and\nopen-pose recognition. To address these challenges, recent wisdom often takes\nsophisticated 2D pre-trained models to provide enriched and stable\nrepresentations. However, these methods largely rely on how 3D objects can be\nprojected into 2D space, which is unfortunately not well solved, and thus\nsignificantly limits their performance. Unlike these present efforts, in this\npaper we make a pioneering exploration of 3D generative models for 3D\nopen-world classification. Drawing on abundant prior knowledge from 3D\ngenerative models, we additionally craft a rotation-invariant feature\nextractor. This innovative synergy endows our pipeline with the advantages of\nbeing training-free, open-category, and pose-invariant, thus well suited to 3D\nopen-world classification. Extensive experiments on benchmark datasets\ndemonstrate the potential of generative models in 3D open-world classification,\nachieving state-of-the-art performance on ModelNet10 and McGill with 32.0% and\n8.7% overall accuracy improvement, respectively.\n", "date": "2025-01-29", "categories": []}
{"id": "2504.13153", "title": "Training-Free Hierarchical Scene Understanding for Gaussian Splatting\n  with Superpoint Graphs", "abstract": "  Bridging natural language and 3D geometry is a crucial step toward flexible,\nlanguage-driven scene understanding. While recent advances in 3D Gaussian\nSplatting (3DGS) have enabled fast and high-quality scene reconstruction,\nresearch has also explored incorporating open-vocabulary understanding into\n3DGS. However, most existing methods require iterative optimization over\nper-view 2D semantic feature maps, which not only results in inefficiencies but\nalso leads to inconsistent 3D semantics across views. To address these\nlimitations, we introduce a training-free framework that constructs a\nsuperpoint graph directly from Gaussian primitives. The superpoint graph\npartitions the scene into spatially compact and semantically coherent regions,\nforming view-consistent 3D entities and providing a structured foundation for\nopen-vocabulary understanding. Based on the graph structure, we design an\nefficient reprojection strategy that lifts 2D semantic features onto the\nsuperpoints, avoiding costly multi-view iterative training. The resulting\nrepresentation ensures strong 3D semantic coherence and naturally supports\nhierarchical understanding, enabling both coarse- and fine-grained\nopen-vocabulary perception within a unified semantic field. Extensive\nexperiments demonstrate that our method achieves state-of-the-art\nopen-vocabulary segmentation performance, with semantic field reconstruction\ncompleted over $30\\times$ faster. Our code will be available at\nhttps://github.com/Atrovast/THGS.\n", "date": "2025-04-17", "categories": []}
{"id": "2405.07847", "title": "SceneFactory: A Workflow-centric and Unified Framework for Incremental\n  Scene Modeling", "abstract": "  We present SceneFactory, a workflow-centric and unified framework for\nincremental scene modeling, that conveniently supports a wide range of\napplications, such as (unposed and/or uncalibrated) multi-view depth\nestimation, LiDAR completion, (dense) RGB-D/RGB-L/Mono/Depth-only\nreconstruction and SLAM. The workflow-centric design uses multiple blocks as\nthe basis for constructing different production lines. The supported\napplications, i.e., productions avoid redundancy in their designs. Thus, the\nfocus is placed on each block itself for independent expansion. To support all\ninput combinations, our implementation consists of four building blocks that\nform SceneFactory: (1) tracking, (2) flexion, (3) depth estimation, and (4)\nscene reconstruction. The tracking block is based on Mono SLAM and is extended\nto support RGB-D and RGB-LiDAR (RGB-L) inputs. Flexion is used to convert the\ndepth image (untrackable) into a trackable image. For general-purpose depth\nestimation, we propose an unposed \\& uncalibrated multi-view depth estimation\nmodel (U$^2$-MVD) to estimate dense geometry. U$^2$-MVD exploits dense bundle\nadjustment to solve for poses, intrinsics, and inverse depth. A semantic-aware\nScaleCov step is then introduced to complete the multi-view depth. Relying on\nU$^2$-MVD, SceneFactory both supports user-friendly 3D creation (with just\nimages) and bridges the applications of Dense RGB-D and Dense Mono. For\nhigh-quality surface and color reconstruction, we propose Dual-purpose\nMulti-resolutional Neural Points (DM-NPs) for the first surface accessible\nSurface Color Field design, where we introduce Improved Point Rasterization\n(IPR) for point cloud based surface query. ...\n", "date": "2024-05-13", "categories": []}
{"id": "2504.11893", "title": "CAGS: Open-Vocabulary 3D Scene Understanding with Context-Aware Gaussian\n  Splatting", "abstract": "  Open-vocabulary 3D scene understanding is crucial for applications requiring\nnatural language-driven spatial interpretation, such as robotics and augmented\nreality. While 3D Gaussian Splatting (3DGS) offers a powerful representation\nfor scene reconstruction, integrating it with open-vocabulary frameworks\nreveals a key challenge: cross-view granularity inconsistency. This issue,\nstemming from 2D segmentation methods like SAM, results in inconsistent object\nsegmentations across views (e.g., a \"coffee set\" segmented as a single entity\nin one view but as \"cup + coffee + spoon\" in another). Existing 3DGS-based\nmethods often rely on isolated per-Gaussian feature learning, neglecting the\nspatial context needed for cohesive object reasoning, leading to fragmented\nrepresentations. We propose Context-Aware Gaussian Splatting (CAGS), a novel\nframework that addresses this challenge by incorporating spatial context into\n3DGS. CAGS constructs local graphs to propagate contextual features across\nGaussians, reducing noise from inconsistent granularity, employs mask-centric\ncontrastive learning to smooth SAM-derived features across views, and leverages\na precomputation strategy to reduce computational cost by precomputing\nneighborhood relationships, enabling efficient training in large-scale scenes.\nBy integrating spatial context, CAGS significantly improves 3D instance\nsegmentation and reduces fragmentation errors on datasets like LERF-OVS and\nScanNet, enabling robust language-guided 3D scene understanding.\n", "date": "2025-04-16", "categories": []}
{"id": "2504.12264", "title": "Towards Learning to Complete Anything in Lidar", "abstract": "  We propose CAL (Complete Anything in Lidar) for Lidar-based shape-completion\nin-the-wild. This is closely related to Lidar-based semantic/panoptic scene\ncompletion. However, contemporary methods can only complete and recognize\nobjects from a closed vocabulary labeled in existing Lidar datasets. Different\nto that, our zero-shot approach leverages the temporal context from multi-modal\nsensor sequences to mine object shapes and semantic features of observed\nobjects. These are then distilled into a Lidar-only instance-level completion\nand recognition model. Although we only mine partial shape completions, we find\nthat our distilled model learns to infer full object shapes from multiple such\npartial observations across the dataset. We show that our model can be prompted\non standard benchmarks for Semantic and Panoptic Scene Completion, localize\nobjects as (amodal) 3D bounding boxes, and recognize objects beyond fixed class\nvocabularies. Our project page is\nhttps://research.nvidia.com/labs/dvl/projects/complete-anything-lidar\n", "date": "2025-04-16", "categories": []}
{"id": "2411.19235", "title": "InstanceGaussian: Appearance-Semantic Joint Gaussian Representation for\n  3D Instance-Level Perception", "abstract": "  3D scene understanding has become an essential area of research with\napplications in autonomous driving, robotics, and augmented reality. Recently,\n3D Gaussian Splatting (3DGS) has emerged as a powerful approach, combining\nexplicit modeling with neural adaptability to provide efficient and detailed\nscene representations. However, three major challenges remain in leveraging\n3DGS for scene understanding: 1) an imbalance between appearance and semantics,\nwhere dense Gaussian usage for fine-grained texture modeling does not align\nwith the minimal requirements for semantic attributes; 2) inconsistencies\nbetween appearance and semantics, as purely appearance-based Gaussians often\nmisrepresent object boundaries; and 3) reliance on top-down instance\nsegmentation methods, which struggle with uneven category distributions,\nleading to over- or under-segmentation. In this work, we propose\nInstanceGaussian, a method that jointly learns appearance and semantic features\nwhile adaptively aggregating instances. Our contributions include: i) a novel\nSemantic-Scaffold-GS representation balancing appearance and semantics to\nimprove feature representations and boundary delineation; ii) a progressive\nappearance-semantic joint training strategy to enhance stability and\nsegmentation accuracy; and iii) a bottom-up, category-agnostic instance\naggregation approach that addresses segmentation challenges through farthest\npoint sampling and connected component analysis. Our approach achieves\nstate-of-the-art performance in category-agnostic, open-vocabulary 3D\npoint-level segmentation, highlighting the effectiveness of the proposed\nrepresentation and training strategies. Project page:\nhttps://lhj-git.github.io/InstanceGaussian/\n", "date": "2024-11-28", "categories": []}
{"id": "2502.02548", "title": "Mosaic3D: Foundation Dataset and Model for Open-Vocabulary 3D\n  Segmentation", "abstract": "  We tackle open-vocabulary 3D scene understanding by introducing a novel data\ngeneration pipeline and training framework. Our method addresses three critical\nrequirements for effective training: precise 3D region segmentation,\ncomprehensive textual descriptions, and sufficient dataset scale. By leveraging\nstate-of-the-art open-vocabulary image segmentation models and region-aware\nVision-Language Models, we develop an automatic pipeline that generates\nhigh-quality 3D mask-text pairs. Applying this pipeline to multiple 3D scene\ndatasets, we create Mosaic3D-5.6M, a dataset of over 30K annotated scenes with\n5.6M mask-text pairs, significantly larger than existing datasets. Building\nupon this data, we propose Mosaic3D, a foundation model combining a 3D encoder\ntrained with contrastive learning and a lightweight mask decoder for\nopen-vocabulary 3D semantic and instance segmentation. Our approach achieves\nstate-of-the-art results on open-vocabulary 3D semantic and instance\nsegmentation tasks including ScanNet200, Matterport3D, and ScanNet++, with\nablation studies validating the effectiveness of our large-scale training data.\n", "date": "2025-02-04", "categories": []}
{"id": "2504.09328", "title": "Text To 3D Object Generation For Scalable Room Assembly", "abstract": "  Modern machine learning models for scene understanding, such as depth\nestimation and object tracking, rely on large, high-quality datasets that mimic\nreal-world deployment scenarios. To address data scarcity, we propose an\nend-to-end system for synthetic data generation for scalable, high-quality, and\ncustomizable 3D indoor scenes. By integrating and adapting text-to-image and\nmulti-view diffusion models with Neural Radiance Field-based meshing, this\nsystem generates highfidelity 3D object assets from text prompts and\nincorporates them into pre-defined floor plans using a rendering tool. By\nintroducing novel loss functions and training strategies into existing methods,\nthe system supports on-demand scene generation, aiming to alleviate the\nscarcity of current available data, generally manually crafted by artists. This\nsystem advances the role of synthetic data in addressing machine learning\ntraining limitations, enabling more robust and generalizable models for\nreal-world applications.\n", "date": "2025-04-12", "categories": []}
{"id": "2504.09518", "title": "3D CoCa: Contrastive Learners are 3D Captioners", "abstract": "  3D captioning, which aims to describe the content of 3D scenes in natural\nlanguage, remains highly challenging due to the inherent sparsity of point\nclouds and weak cross-modal alignment in existing methods. To address these\nchallenges, we propose 3D CoCa, a novel unified framework that seamlessly\ncombines contrastive vision-language learning with 3D caption generation in a\nsingle architecture. Our approach leverages a frozen CLIP vision-language\nbackbone to provide rich semantic priors, a spatially-aware 3D scene encoder to\ncapture geometric context, and a multi-modal decoder to generate descriptive\ncaptions. Unlike prior two-stage methods that rely on explicit object\nproposals, 3D CoCa jointly optimizes contrastive and captioning objectives in a\nshared feature space, eliminating the need for external detectors or\nhandcrafted proposals. This joint training paradigm yields stronger spatial\nreasoning and richer semantic grounding by aligning 3D and textual\nrepresentations. Extensive experiments on the ScanRefer and Nr3D benchmarks\ndemonstrate that 3D CoCa significantly outperforms current state-of-the-arts by\n10.2% and 5.76% in CIDEr at 0.5IoU, respectively. Code will be available at\nhttps://github.com/AIGeeksGroup/3DCoCa.\n", "date": "2025-04-13", "categories": []}
{"id": "2504.09623", "title": "Ges3ViG: Incorporating Pointing Gestures into Language-Based 3D Visual\n  Grounding for Embodied Reference Understanding", "abstract": "  3-Dimensional Embodied Reference Understanding (3D-ERU) combines a language\ndescription and an accompanying pointing gesture to identify the most relevant\ntarget object in a 3D scene. Although prior work has explored pure\nlanguage-based 3D grounding, there has been limited exploration of 3D-ERU,\nwhich also incorporates human pointing gestures. To address this gap, we\nintroduce a data augmentation framework-Imputer, and use it to curate a new\nbenchmark dataset-ImputeRefer for 3D-ERU, by incorporating human pointing\ngestures into existing 3D scene datasets that only contain language\ninstructions. We also propose Ges3ViG, a novel model for 3D-ERU that achieves\n~30% improvement in accuracy as compared to other 3D-ERU models and ~9%\ncompared to other purely language-based 3D grounding models. Our code and\ndataset are available at https://github.com/AtharvMane/Ges3ViG.\n", "date": "2025-04-13", "categories": []}
{"id": "2504.10048", "title": "Multi-Object Grounding via Hierarchical Contrastive Siamese Transformers", "abstract": "  Multi-object grounding in 3D scenes involves localizing multiple objects\nbased on natural language input. While previous work has primarily focused on\nsingle-object grounding, real-world scenarios often demand the localization of\nseveral objects. To tackle this challenge, we propose Hierarchical Contrastive\nSiamese Transformers (H-COST), which employs a Hierarchical Processing strategy\nto progressively refine object localization, enhancing the understanding of\ncomplex language instructions. Additionally, we introduce a Contrastive Siamese\nTransformer framework, where two networks with the identical structure are\nused: one auxiliary network processes robust object relations from ground-truth\nlabels to guide and enhance the second network, the reference network, which\noperates on segmented point-cloud data. This contrastive mechanism strengthens\nthe model' s semantic understanding and significantly enhances its ability to\nprocess complex point-cloud data. Our approach outperforms previous\nstate-of-the-art methods by 9.5% on challenging multi-object grounding\nbenchmarks.\n", "date": "2025-04-14", "categories": []}
{"id": "2504.10106", "title": "SoccerNet-v3D: Leveraging Sports Broadcast Replays for 3D Scene\n  Understanding", "abstract": "  Sports video analysis is a key domain in computer vision, enabling detailed\nspatial understanding through multi-view correspondences. In this work, we\nintroduce SoccerNet-v3D and ISSIA-3D, two enhanced and scalable datasets\ndesigned for 3D scene understanding in soccer broadcast analysis. These\ndatasets extend SoccerNet-v3 and ISSIA by incorporating field-line-based camera\ncalibration and multi-view synchronization, enabling 3D object localization\nthrough triangulation. We propose a monocular 3D ball localization task built\nupon the triangulation of ground-truth 2D ball annotations, along with several\ncalibration and reprojection metrics to assess annotation quality on demand.\nAdditionally, we present a single-image 3D ball localization method as a\nbaseline, leveraging camera calibration and ball size priors to estimate the\nball's position from a monocular viewpoint. To further refine 2D annotations,\nwe introduce a bounding box optimization technique that ensures alignment with\nthe 3D scene representation. Our proposed datasets establish new benchmarks for\n3D soccer scene understanding, enhancing both spatial and temporal analysis in\nsports analytics. Finally, we provide code to facilitate access to our\nannotations and the generation pipelines for the datasets.\n", "date": "2025-04-14", "categories": []}
{"id": "2504.06553", "title": "ASHiTA: Automatic Scene-grounded HIerarchical Task Analysis", "abstract": "  While recent work in scene reconstruction and understanding has made strides\nin grounding natural language to physical 3D environments, it is still\nchallenging to ground abstract, high-level instructions to a 3D scene.\nHigh-level instructions might not explicitly invoke semantic elements in the\nscene, and even the process of breaking a high-level task into a set of more\nconcrete subtasks, a process called hierarchical task analysis, is\nenvironment-dependent. In this work, we propose ASHiTA, the first framework\nthat generates a task hierarchy grounded to a 3D scene graph by breaking down\nhigh-level tasks into grounded subtasks. ASHiTA alternates LLM-assisted\nhierarchical task analysis, to generate the task breakdown, with task-driven 3D\nscene graph construction to generate a suitable representation of the\nenvironment. Our experiments show that ASHiTA performs significantly better\nthan LLM baselines in breaking down high-level tasks into environment-dependent\nsubtasks and is additionally able to achieve grounding performance comparable\nto state-of-the-art methods.\n", "date": "2025-04-09", "categories": []}
{"id": "2504.08581", "title": "FMLGS: Fast Multilevel Language Embedded Gaussians for Part-level\n  Interactive Agents", "abstract": "  The semantically interactive radiance field has long been a promising\nbackbone for 3D real-world applications, such as embodied AI to achieve scene\nunderstanding and manipulation. However, multi-granularity interaction remains\na challenging task due to the ambiguity of language and degraded quality when\nit comes to queries upon object components. In this work, we present FMLGS, an\napproach that supports part-level open-vocabulary query within 3D Gaussian\nSplatting (3DGS). We propose an efficient pipeline for building and querying\nconsistent object- and part-level semantics based on Segment Anything Model 2\n(SAM2). We designed a semantic deviation strategy to solve the problem of\nlanguage ambiguity among object parts, which interpolates the semantic features\nof fine-grained targets for enriched information. Once trained, we can query\nboth objects and their describable parts using natural language. Comparisons\nwith other state-of-the-art methods prove that our method can not only better\nlocate specified part-level targets, but also achieve first-place performance\nconcerning both speed and accuracy, where FMLGS is 98 x faster than LERF, 4 x\nfaster than LangSplat and 2.5 x faster than LEGaussians. Meanwhile, we further\nintegrate FMLGS as a virtual agent that can interactively navigate through 3D\nscenes, locate targets, and respond to user demands through a chat interface,\nwhich demonstrates the potential of our work to be further expanded and applied\nin the future.\n", "date": "2025-04-11", "categories": []}
{"id": "2503.16825", "title": "SGFormer: Satellite-Ground Fusion for 3D Semantic Scene Completion", "abstract": "  Recently, camera-based solutions have been extensively explored for scene\nsemantic completion (SSC). Despite their success in visible areas, existing\nmethods struggle to capture complete scene semantics due to frequent visual\nocclusions. To address this limitation, this paper presents the first\nsatellite-ground cooperative SSC framework, i.e., SGFormer, exploring the\npotential of satellite-ground image pairs in the SSC task. Specifically, we\npropose a dual-branch architecture that encodes orthogonal satellite and ground\nviews in parallel, unifying them into a common domain. Additionally, we design\na ground-view guidance strategy that corrects satellite image biases during\nfeature encoding, addressing misalignment between satellite and ground views.\nMoreover, we develop an adaptive weighting strategy that balances contributions\nfrom satellite and ground views. Experiments demonstrate that SGFormer\noutperforms the state of the art on SemanticKITTI and SSCBench-KITTI-360\ndatasets. Our code is available on https://github.com/gxytcrc/SGFormer.\n", "date": "2025-03-21", "categories": []}
{"id": "2504.07524", "title": "DGOcc: Depth-aware Global Query-based Network for Monocular 3D Occupancy\n  Prediction", "abstract": "  Monocular 3D occupancy prediction, aiming to predict the occupancy and\nsemantics within interesting regions of 3D scenes from only 2D images, has\ngarnered increasing attention recently for its vital role in 3D scene\nunderstanding. Predicting the 3D occupancy of large-scale outdoor scenes from\n2D images is ill-posed and resource-intensive. In this paper, we present\n\\textbf{DGOcc}, a \\textbf{D}epth-aware \\textbf{G}lobal query-based network for\nmonocular 3D \\textbf{Occ}upancy prediction. We first explore prior depth maps\nto extract depth context features that provide explicit geometric information\nfor the occupancy network. Then, in order to fully exploit the depth context\nfeatures, we propose a Global Query-based (GQ) Module. The cooperation of\nattention mechanisms and scale-aware operations facilitates the feature\ninteraction between images and 3D voxels. Moreover, a Hierarchical Supervision\nStrategy (HSS) is designed to avoid upsampling the high-dimension 3D voxel\nfeatures to full resolution, which mitigates GPU memory utilization and time\ncost. Extensive experiments on SemanticKITTI and SSCBench-KITTI-360 datasets\ndemonstrate that the proposed method achieves the best performance on monocular\nsemantic occupancy prediction while reducing GPU and time overhead.\n", "date": "2025-04-10", "categories": []}
{"id": "2504.06719", "title": "Masked Scene Modeling: Narrowing the Gap Between Supervised and\n  Self-Supervised Learning in 3D Scene Understanding", "abstract": "  Self-supervised learning has transformed 2D computer vision by enabling\nmodels trained on large, unannotated datasets to provide versatile\noff-the-shelf features that perform similarly to models trained with labels.\nHowever, in 3D scene understanding, self-supervised methods are typically only\nused as a weight initialization step for task-specific fine-tuning, limiting\ntheir utility for general-purpose feature extraction. This paper addresses this\nshortcoming by proposing a robust evaluation protocol specifically designed to\nassess the quality of self-supervised features for 3D scene understanding. Our\nprotocol uses multi-resolution feature sampling of hierarchical models to\ncreate rich point-level representations that capture the semantic capabilities\nof the model and, hence, are suitable for evaluation with linear probing and\nnearest-neighbor methods. Furthermore, we introduce the first self-supervised\nmodel that performs similarly to supervised models when only off-the-shelf\nfeatures are used in a linear probing setup. In particular, our model is\ntrained natively in 3D with a novel self-supervised approach based on a Masked\nScene Modeling objective, which reconstructs deep features of masked patches in\na bottom-up manner and is specifically tailored to hierarchical 3D models. Our\nexperiments not only demonstrate that our method achieves competitive\nperformance to supervised models, but also surpasses existing self-supervised\napproaches by a large margin. The model and training code can be found at our\nGithub repository (https://github.com/phermosilla/msm).\n", "date": "2025-04-09", "categories": []}
{"id": "2504.06827", "title": "IAAO: Interactive Affordance Learning for Articulated Objects in 3D\n  Environments", "abstract": "  This work presents IAAO, a novel framework that builds an explicit 3D model\nfor intelligent agents to gain understanding of articulated objects in their\nenvironment through interaction. Unlike prior methods that rely on\ntask-specific networks and assumptions about movable parts, our IAAO leverages\nlarge foundation models to estimate interactive affordances and part\narticulations in three stages. We first build hierarchical features and label\nfields for each object state using 3D Gaussian Splatting (3DGS) by distilling\nmask features and view-consistent labels from multi-view images. We then\nperform object- and part-level queries on the 3D Gaussian primitives to\nidentify static and articulated elements, estimating global transformations and\nlocal articulation parameters along with affordances. Finally, scenes from\ndifferent states are merged and refined based on the estimated transformations,\nenabling robust affordance-based interaction and manipulation of objects.\nExperimental results demonstrate the effectiveness of our method.\n", "date": "2025-04-09", "categories": []}
{"id": "2504.06994", "title": "RayFronts: Open-Set Semantic Ray Frontiers for Online Scene\n  Understanding and Exploration", "abstract": "  Open-set semantic mapping is crucial for open-world robots. Current mapping\napproaches either are limited by the depth range or only map beyond-range\nentities in constrained settings, where overall they fail to combine\nwithin-range and beyond-range observations. Furthermore, these methods make a\ntrade-off between fine-grained semantics and efficiency. We introduce\nRayFronts, a unified representation that enables both dense and beyond-range\nefficient semantic mapping. RayFronts encodes task-agnostic open-set semantics\nto both in-range voxels and beyond-range rays encoded at map boundaries,\nempowering the robot to reduce search volumes significantly and make informed\ndecisions both within & beyond sensory range, while running at 8.84 Hz on an\nOrin AGX. Benchmarking the within-range semantics shows that RayFronts's\nfine-grained image encoding provides 1.34x zero-shot 3D semantic segmentation\nperformance while improving throughput by 16.5x. Traditionally, online mapping\nperformance is entangled with other system components, complicating evaluation.\nWe propose a planner-agnostic evaluation framework that captures the utility\nfor online beyond-range search and exploration, and show RayFronts reduces\nsearch volume 2.2x more efficiently than the closest online baselines.\n", "date": "2025-04-09", "categories": []}
{"id": "2504.05751", "title": "InvNeRF-Seg: Fine-Tuning a Pre-Trained NeRF for 3D Object Segmentation", "abstract": "  Neural Radiance Fields (NeRF) have been widely adopted for reconstructing\nhigh quality 3D point clouds from 2D RGB images. However, the segmentation of\nthese reconstructed 3D scenes is more essential for downstream tasks such as\nobject counting, size estimation, and scene understanding. While segmentation\non raw 3D point clouds using deep learning requires labor intensive and\ntime-consuming manual annotation, directly training NeRF on binary masks also\nfails due to the absence of color and shading cues essential for geometry\nlearning. We propose Invariant NeRF for Segmentation (InvNeRFSeg), a two step,\nzero change fine tuning strategy for 3D segmentation. We first train a standard\nNeRF on RGB images and then fine tune it using 2D segmentation masks without\naltering either the model architecture or loss function. This approach produces\nhigher quality, cleaner segmented point clouds directly from the refined\nradiance field with minimal computational overhead or complexity. Field density\nanalysis reveals consistent semantic refinement: densities of object regions\nincrease while background densities are suppressed, ensuring clean and\ninterpretable segmentations. We demonstrate InvNeRFSegs superior performance\nover both SA3D and FruitNeRF on both synthetic fruit and real world soybean\ndatasets. This approach effectively extends 2D segmentation to high quality 3D\nsegmentation.\n", "date": "2025-04-08", "categories": []}
{"id": "2504.05882", "title": "Turin3D: Evaluating Adaptation Strategies under Label Scarcity in Urban\n  LiDAR Segmentation with Semi-Supervised Techniques", "abstract": "  3D semantic segmentation plays a critical role in urban modelling, enabling\ndetailed understanding and mapping of city environments. In this paper, we\nintroduce Turin3D: a new aerial LiDAR dataset for point cloud semantic\nsegmentation covering an area of around 1.43 km2 in the city centre of Turin\nwith almost 70M points. We describe the data collection process and compare\nTurin3D with others previously proposed in the literature. We did not fully\nannotate the dataset due to the complexity and time-consuming nature of the\nprocess; however, a manual annotation process was performed on the validation\nand test sets, to enable a reliable evaluation of the proposed techniques. We\nfirst benchmark the performances of several point cloud semantic segmentation\nmodels, trained on the existing datasets, when tested on Turin3D, and then\nimprove their performances by applying a semi-supervised learning technique\nleveraging the unlabelled training set. The dataset will be publicly available\nto support research in outdoor point cloud segmentation, with particular\nrelevance for self-supervised and semi-supervised learning approaches given the\nabsence of ground truth annotations for the training set.\n", "date": "2025-04-08", "categories": []}
{"id": "2502.15011", "title": "CrossOver: 3D Scene Cross-Modal Alignment", "abstract": "  Multi-modal 3D object understanding has gained significant attention, yet\ncurrent approaches often assume complete data availability and rigid alignment\nacross all modalities. We present CrossOver, a novel framework for cross-modal\n3D scene understanding via flexible, scene-level modality alignment. Unlike\ntraditional methods that require aligned modality data for every object\ninstance, CrossOver learns a unified, modality-agnostic embedding space for\nscenes by aligning modalities -- RGB images, point clouds, CAD models,\nfloorplans, and text descriptions -- with relaxed constraints and without\nexplicit object semantics. Leveraging dimensionality-specific encoders, a\nmulti-stage training pipeline, and emergent cross-modal behaviors, CrossOver\nsupports robust scene retrieval and object localization, even with missing\nmodalities. Evaluations on ScanNet and 3RScan datasets show its superior\nperformance across diverse metrics, highlighting the adaptability for\nreal-world applications in 3D scene understanding.\n", "date": "2025-02-20", "categories": []}
{"id": "2504.03875", "title": "3D Scene Understanding Through Local Random Access Sequence Modeling", "abstract": "  3D scene understanding from single images is a pivotal problem in computer\nvision with numerous downstream applications in graphics, augmented reality,\nand robotics. While diffusion-based modeling approaches have shown promise,\nthey often struggle to maintain object and scene consistency, especially in\ncomplex real-world scenarios. To address these limitations, we propose an\nautoregressive generative approach called Local Random Access Sequence (LRAS)\nmodeling, which uses local patch quantization and randomly ordered sequence\ngeneration. By utilizing optical flow as an intermediate representation for 3D\nscene editing, our experiments demonstrate that LRAS achieves state-of-the-art\nnovel view synthesis and 3D object manipulation capabilities. Furthermore, we\nshow that our framework naturally extends to self-supervised depth estimation\nthrough a simple modification of the sequence design. By achieving strong\nperformance on multiple 3D scene understanding tasks, LRAS provides a unified\nand effective framework for building the next generation of 3D vision models.\n", "date": "2025-04-04", "categories": []}
{"id": "2504.04744", "title": "Grounding 3D Object Affordance with Language Instructions, Visual\n  Observations and Interactions", "abstract": "  Grounding 3D object affordance is a task that locates objects in 3D space\nwhere they can be manipulated, which links perception and action for embodied\nintelligence. For example, for an intelligent robot, it is necessary to\naccurately ground the affordance of an object and grasp it according to human\ninstructions. In this paper, we introduce a novel task that grounds 3D object\naffordance based on language instructions, visual observations and\ninteractions, which is inspired by cognitive science. We collect an Affordance\nGrounding dataset with Points, Images and Language instructions (AGPIL) to\nsupport the proposed task. In the 3D physical world, due to observation\norientation, object rotation, or spatial occlusion, we can only get a partial\nobservation of the object. So this dataset includes affordance estimations of\nobjects from full-view, partial-view, and rotation-view perspectives. To\naccomplish this task, we propose LMAffordance3D, the first multi-modal,\nlanguage-guided 3D affordance grounding network, which applies a\nvision-language model to fuse 2D and 3D spatial features with semantic\nfeatures. Comprehensive experiments on AGPIL demonstrate the effectiveness and\nsuperiority of our method on this task, even in unseen experimental settings.\nOur project is available at https://sites.google.com/view/lmaffordance3d.\n", "date": "2025-04-07", "categories": []}
{"id": "2411.17735", "title": "3D-Mem: 3D Scene Memory for Embodied Exploration and Reasoning", "abstract": "  Constructing compact and informative 3D scene representations is essential\nfor effective embodied exploration and reasoning, especially in complex\nenvironments over extended periods. Existing representations, such as\nobject-centric 3D scene graphs, oversimplify spatial relationships by modeling\nscenes as isolated objects with restrictive textual relationships, making it\ndifficult to address queries requiring nuanced spatial understanding. Moreover,\nthese representations lack natural mechanisms for active exploration and memory\nmanagement, hindering their application to lifelong autonomy. In this work, we\npropose 3D-Mem, a novel 3D scene memory framework for embodied agents. 3D-Mem\nemploys informative multi-view images, termed Memory Snapshots, to represent\nthe scene and capture rich visual information of explored regions. It further\nintegrates frontier-based exploration by introducing Frontier\nSnapshots-glimpses of unexplored areas-enabling agents to make informed\ndecisions by considering both known and potential new information. To support\nlifelong memory in active exploration settings, we present an incremental\nconstruction pipeline for 3D-Mem, as well as a memory retrieval technique for\nmemory management. Experimental results on three benchmarks demonstrate that\n3D-Mem significantly enhances agents' exploration and reasoning capabilities in\n3D environments, highlighting its potential for advancing applications in\nembodied AI.\n", "date": "2024-11-23", "categories": []}
{"id": "2504.03129", "title": "GraphSeg: Segmented 3D Representations via Graph Edge Addition and\n  Contraction", "abstract": "  Robots operating in unstructured environments often require accurate and\nconsistent object-level representations. This typically requires segmenting\nindividual objects from the robot's surroundings. While recent large models\nsuch as Segment Anything (SAM) offer strong performance in 2D image\nsegmentation. These advances do not translate directly to performance in the\nphysical 3D world, where they often over-segment objects and fail to produce\nconsistent mask correspondences across views. In this paper, we present\nGraphSeg, a framework for generating consistent 3D object segmentations from a\nsparse set of 2D images of the environment without any depth information.\nGraphSeg adds edges to graphs and constructs dual correspondence graphs: one\nfrom 2D pixel-level similarities and one from inferred 3D structure. We\nformulate segmentation as a problem of edge addition, then subsequent graph\ncontraction, which merges multiple 2D masks into unified object-level\nsegmentations. We can then leverage \\emph{3D foundation models} to produce\nsegmented 3D representations. GraphSeg achieves robust segmentation with\nsignificantly fewer images and greater accuracy than prior methods. We\ndemonstrate state-of-the-art performance on tabletop scenes and show that\nGraphSeg enables improved performance on downstream robotic manipulation tasks.\nCode available at https://github.com/tomtang502/graphseg.git.\n", "date": "2025-04-04", "categories": []}
{"id": "2408.09928", "title": "Enforcing View-Consistency in Class-Agnostic 3D Segmentation Fields", "abstract": "  Radiance Fields have become a powerful tool for modeling 3D scenes from\nmultiple images. However, they remain difficult to segment into semantically\nmeaningful regions. Some methods work well using 2D semantic masks, but they\ngeneralize poorly to class-agnostic segmentations. More recent methods\ncircumvent this issue by using contrastive learning to optimize a\nhigh-dimensional 3D feature field instead. However, recovering a segmentation\nthen requires clustering and fine-tuning the associated hyperparameters. In\ncontrast, we aim to identify the necessary changes in segmentation field\nmethods to directly learn a segmentation field while being robust to\ninconsistent class-agnostic masks, successfully decomposing the scene into a\nset of objects of any class.\n  By introducing an additional spatial regularization term and restricting the\nfield to a limited number of competing object slots against which masks are\nmatched, a meaningful object representation emerges that best explains the 2D\nsupervision. Our experiments demonstrate the ability of our method to generate\n3D panoptic segmentations on complex scenes, and extract high-quality 3D assets\nfrom radiance fields that can then be used in virtual 3D environments.\n", "date": "2024-08-19", "categories": []}
{"id": "2504.02161", "title": "Preference-Driven Active 3D Scene Representation for Robotic Inspection\n  in Nuclear Decommissioning", "abstract": "  Active 3D scene representation is pivotal in modern robotics applications,\nincluding remote inspection, manipulation, and telepresence. Traditional\nmethods primarily optimize geometric fidelity or rendering accuracy, but often\noverlook operator-specific objectives, such as safety-critical coverage or\ntask-driven viewpoints. This limitation leads to suboptimal viewpoint\nselection, particularly in constrained environments such as nuclear\ndecommissioning. To bridge this gap, we introduce a novel framework that\nintegrates expert operator preferences into the active 3D scene representation\npipeline. Specifically, we employ Reinforcement Learning from Human Feedback\n(RLHF) to guide robotic path planning, reshaping the reward function based on\nexpert input. To capture operator-specific priorities, we conduct interactive\nchoice experiments that evaluate user preferences in 3D scene representation.\nWe validate our framework using a UR3e robotic arm for reactor tile inspection\nin a nuclear decommissioning scenario. Compared to baseline methods, our\napproach enhances scene representation while optimizing trajectory efficiency.\nThe RLHF-based policy consistently outperforms random selection, prioritizing\ntask-critical details. By unifying explicit 3D geometric modeling with implicit\nhuman-in-the-loop optimization, this work establishes a foundation for\nadaptive, safety-critical robotic perception systems, paving the way for\nenhanced automation in nuclear decommissioning, remote maintenance, and other\nhigh-risk environments.\n", "date": "2025-04-02", "categories": []}
{"id": "2411.16308", "title": "An End-to-End Robust Point Cloud Semantic Segmentation Network with\n  Single-Step Conditional Diffusion Models", "abstract": "  Existing conditional Denoising Diffusion Probabilistic Models (DDPMs) with a\nNoise-Conditional Framework (NCF) remain challenging for 3D scene understanding\ntasks, as the complex geometric details in scenes increase the difficulty of\nfitting the gradients of the data distribution (the scores) from semantic\nlabels. This also results in longer training and inference time for DDPMs\ncompared to non-DDPMs. From a different perspective, we delve deeply into the\nmodel paradigm dominated by the Conditional Network. In this paper, we propose\nan end-to-end robust semantic Segmentation Network based on a Conditional-Noise\nFramework (CNF) of DDPMs, named CDSegNet. Specifically, CDSegNet models the\nNoise Network (NN) as a learnable noise-feature generator. This enables the\nConditional Network (CN) to understand 3D scene semantics under multi-level\nfeature perturbations, enhancing the generalization in unseen scenes.\nMeanwhile, benefiting from the noise system of DDPMs, CDSegNet exhibits strong\nnoise and sparsity robustness in experiments. Moreover, thanks to CNF, CDSegNet\ncan generate the semantic labels in a single-step inference like non-DDPMs, due\nto avoiding directly fitting the scores from semantic labels in the dominant\nnetwork of CDSegNet. On public indoor and outdoor benchmarks, CDSegNet\nsignificantly outperforms existing methods, achieving state-of-the-art\nperformance.\n", "date": "2024-11-25", "categories": []}
{"id": "2503.22420", "title": "Unveiling the Mist over 3D Vision-Language Understanding: Object-centric\n  Evaluation with Chain-of-Analysis", "abstract": "  Existing 3D vision-language (3D-VL) benchmarks fall short in evaluating 3D-VL\nmodels, creating a \"mist\" that obscures rigorous insights into model\ncapabilities and 3D-VL tasks. This mist persists due to three key limitations.\nFirst, flawed test data, like ambiguous referential text in the grounding task,\ncan yield incorrect and unreliable test results. Second, oversimplified metrics\nsuch as simply averaging accuracy per question answering (QA) pair, cannot\nreveal true model capability due to their vulnerability to language variations.\nThird, existing benchmarks isolate the grounding and QA tasks, disregarding the\nunderlying coherence that QA should be based on solid grounding capabilities.\nTo unveil the \"mist\", we propose Beacon3D, a benchmark for 3D-VL grounding and\nQA tasks, delivering a perspective shift in the evaluation of 3D-VL\nunderstanding. Beacon3D features (i) high-quality test data with precise and\nnatural language, (ii) object-centric evaluation with multiple tests per object\nto ensure robustness, and (iii) a novel chain-of-analysis paradigm to address\nlanguage robustness and model performance coherence across grounding and QA.\nOur evaluation of state-of-the-art 3D-VL models on Beacon3D reveals that (i)\nobject-centric evaluation elicits true model performance and particularly weak\ngeneralization in QA; (ii) grounding-QA coherence remains fragile in current\n3D-VL models, and (iii) incorporating large language models (LLMs) to 3D-VL\nmodels, though as a prevalent practice, hinders grounding capabilities and has\nyet to elevate QA capabilities. We hope Beacon3D and our comprehensive analysis\ncould benefit the 3D-VL community towards faithful developments.\n", "date": "2025-03-28", "categories": []}
{"id": "2411.19626", "title": "GREAT: Geometry-Intention Collaborative Inference for Open-Vocabulary 3D\n  Object Affordance Grounding", "abstract": "  Open-Vocabulary 3D object affordance grounding aims to anticipate ``action\npossibilities'' regions on 3D objects with arbitrary instructions, which is\ncrucial for robots to generically perceive real scenarios and respond to\noperational changes. Existing methods focus on combining images or languages\nthat depict interactions with 3D geometries to introduce external interaction\npriors. However, they are still vulnerable to a limited semantic space by\nfailing to leverage implied invariant geometries and potential interaction\nintentions. Normally, humans address complex tasks through multi-step reasoning\nand respond to diverse situations by leveraging associative and analogical\nthinking. In light of this, we propose GREAT (GeometRy-intEntion collAboraTive\ninference) for Open-Vocabulary 3D Object Affordance Grounding, a novel\nframework that mines the object invariant geometry attributes and performs\nanalogically reason in potential interaction scenarios to form affordance\nknowledge, fully combining the knowledge with both geometries and visual\ncontents to ground 3D object affordance. Besides, we introduce the Point Image\nAffordance Dataset v2 (PIADv2), the largest 3D object affordance dataset at\npresent to support the task. Extensive experiments demonstrate the\neffectiveness and superiority of GREAT. The code and dataset are available at\nhttps://yawen-shao.github.io/GREAT/.\n", "date": "2024-11-29", "categories": []}
{"id": "2503.23024", "title": "Empowering Large Language Models with 3D Situation Awareness", "abstract": "  Driven by the great success of Large Language Models (LLMs) in the 2D image\ndomain, their applications in 3D scene understanding has emerged as a new\ntrend. A key difference between 3D and 2D is that the situation of an\negocentric observer in 3D scenes can change, resulting in different\ndescriptions (e.g., ''left\" or ''right\"). However, current LLM-based methods\noverlook the egocentric perspective and simply use datasets from a global\nviewpoint. To address this issue, we propose a novel approach to automatically\ngenerate a situation-aware dataset by leveraging the scanning trajectory during\ndata collection and utilizing Vision-Language Models (VLMs) to produce\nhigh-quality captions and question-answer pairs. Furthermore, we introduce a\nsituation grounding module to explicitly predict the position and orientation\nof observer's viewpoint, thereby enabling LLMs to ground situation description\nin 3D scenes. We evaluate our approach on several benchmarks, demonstrating\nthat our method effectively enhances the 3D situational awareness of LLMs while\nsignificantly expanding existing datasets and reducing manual effort.\n", "date": "2025-03-29", "categories": []}
{"id": "2503.16707", "title": "Cross-Modal and Uncertainty-Aware Agglomeration for Open-Vocabulary 3D\n  Scene Understanding", "abstract": "  The lack of a large-scale 3D-text corpus has led recent works to distill\nopen-vocabulary knowledge from vision-language models (VLMs). However, these\nmethods typically rely on a single VLM to align the feature spaces of 3D models\nwithin a common language space, which limits the potential of 3D models to\nleverage the diverse spatial and semantic capabilities encapsulated in various\nfoundation models. In this paper, we propose Cross-modal and Uncertainty-aware\nAgglomeration for Open-vocabulary 3D Scene Understanding dubbed CUA-O3D, the\nfirst model to integrate multiple foundation models-such as CLIP, DINOv2, and\nStable Diffusion-into 3D scene understanding. We further introduce a\ndeterministic uncertainty estimation to adaptively distill and harmonize the\nheterogeneous 2D feature embeddings from these models. Our method addresses two\nkey challenges: (1) incorporating semantic priors from VLMs alongside the\ngeometric knowledge of spatially-aware vision foundation models, and (2) using\na novel deterministic uncertainty estimation to capture model-specific\nuncertainties across diverse semantic and geometric sensitivities, helping to\nreconcile heterogeneous representations during training. Extensive experiments\non ScanNetV2 and Matterport3D demonstrate that our method not only advances\nopen-vocabulary segmentation but also achieves robust cross-domain alignment\nand competitive spatial perception capabilities. The code will be available at:\nhttps://github.com/TyroneLi/CUA_O3D.\n", "date": "2025-03-20", "categories": []}
{"id": "2412.00493", "title": "Video-3D LLM: Learning Position-Aware Video Representation for 3D Scene\n  Understanding", "abstract": "  The rapid advancement of Multimodal Large Language Models (MLLMs) has\nsignificantly impacted various multimodal tasks. However, these models face\nchallenges in tasks that require spatial understanding within 3D environments.\nEfforts to enhance MLLMs, such as incorporating point cloud features, have been\nmade, yet a considerable gap remains between the models' learned\nrepresentations and the inherent complexity of 3D scenes. This discrepancy\nlargely stems from the training of MLLMs on predominantly 2D data, which\nrestricts their effectiveness in comprehending 3D spaces. To address this\nissue, in this paper, we propose a novel generalist model, i.e., Video-3D LLM,\nfor 3D scene understanding. By treating 3D scenes as dynamic videos and\nincorporating 3D position encoding into these representations, our Video-3D LLM\naligns video representations with real-world spatial contexts more accurately.\nIn addition, we have implemented a maximum coverage sampling technique to\noptimize the trade-off between computational cost and performance. Extensive\nexperiments demonstrate that our model achieves state-of-the-art performance on\nseveral 3D scene understanding benchmarks, including ScanRefer, Multi3DRefer,\nScan2Cap, ScanQA, and SQA3D.\n", "date": "2024-11-30", "categories": []}
{"id": "2503.21562", "title": "uLayout: Unified Room Layout Estimation for Perspective and Panoramic\n  Images", "abstract": "  We present uLayout, a unified model for estimating room layout geometries\nfrom both perspective and panoramic images, whereas traditional solutions\nrequire different model designs for each image type. The key idea of our\nsolution is to unify both domains into the equirectangular projection,\nparticularly, allocating perspective images into the most suitable latitude\ncoordinate to effectively exploit both domains seamlessly. To address the\nField-of-View (FoV) difference between the input domains, we design uLayout\nwith a shared feature extractor with an extra 1D-Convolution layer to condition\neach domain input differently. This conditioning allows us to efficiently\nformulate a column-wise feature regression problem regardless of the FoV input.\nThis simple yet effective approach achieves competitive performance with\ncurrent state-of-the-art solutions and shows for the first time a single\nend-to-end model for both domains. Extensive experiments in the real-world\ndatasets, LSUN, Matterport3D, PanoContext, and Stanford 2D-3D evidence the\ncontribution of our approach. Code is available at\nhttps://github.com/JonathanLee112/uLayout.\n", "date": "2025-03-27", "categories": []}
{"id": "2411.16129", "title": "Three Cars Approaching within 100m! Enhancing Distant Geometry by\n  Tri-Axis Voxel Scanning for Camera-based Semantic Scene Completion", "abstract": "  Camera-based Semantic Scene Completion (SSC) is gaining attentions in the 3D\nperception field. However, properties such as perspective and occlusion lead to\nthe underestimation of the geometry in distant regions, posing a critical issue\nfor safety-focused autonomous driving systems. To tackle this, we propose\nScanSSC, a novel camera-based SSC model composed of a Scan Module and Scan\nLoss, both designed to enhance distant scenes by leveraging context from\nnear-viewpoint scenes. The Scan Module uses axis-wise masked attention, where\neach axis employing a near-to-far cascade masking that enables distant voxels\nto capture relationships with preceding voxels. In addition, the Scan Loss\ncomputes the cross-entropy along each axis between cumulative logits and\ncorresponding class distributions in a near-to-far direction, thereby\npropagating rich context-aware signals to distant voxels. Leveraging the\nsynergy between these components, ScanSSC achieves state-of-the-art\nperformance, with IoUs of 44.54 and 48.29, and mIoUs of 17.40 and 20.14 on the\nSemanticKITTI and SSCBench-KITTI-360 benchmarks.\n", "date": "2024-11-25", "categories": []}
{"id": "2412.13652", "title": "RelationField: Relate Anything in Radiance Fields", "abstract": "  Neural radiance fields are an emerging 3D scene representation and recently\neven been extended to learn features for scene understanding by distilling\nopen-vocabulary features from vision-language models. However, current method\nprimarily focus on object-centric representations, supporting object\nsegmentation or detection, while understanding semantic relationships between\nobjects remains largely unexplored. To address this gap, we propose\nRelationField, the first method to extract inter-object relationships directly\nfrom neural radiance fields. RelationField represents relationships between\nobjects as pairs of rays within a neural radiance field, effectively extending\nits formulation to include implicit relationship queries. To teach\nRelationField complex, open-vocabulary relationships, relationship knowledge is\ndistilled from multi-modal LLMs. To evaluate RelationField, we solve\nopen-vocabulary 3D scene graph generation tasks and relationship-guided\ninstance segmentation, achieving state-of-the-art performance in both tasks.\nSee the project website at https://relationfield.github.io.\n", "date": "2024-12-18", "categories": []}
{"id": "2503.19199", "title": "Open-Vocabulary Functional 3D Scene Graphs for Real-World Indoor Spaces", "abstract": "  We introduce the task of predicting functional 3D scene graphs for real-world\nindoor environments from posed RGB-D images. Unlike traditional 3D scene graphs\nthat focus on spatial relationships of objects, functional 3D scene graphs\ncapture objects, interactive elements, and their functional relationships. Due\nto the lack of training data, we leverage foundation models, including visual\nlanguage models (VLMs) and large language models (LLMs), to encode functional\nknowledge. We evaluate our approach on an extended SceneFun3D dataset and a\nnewly collected dataset, FunGraph3D, both annotated with functional 3D scene\ngraphs. Our method significantly outperforms adapted baselines, including\nOpen3DSG and ConceptGraph, demonstrating its effectiveness in modeling complex\nscene functionalities. We also demonstrate downstream applications such as 3D\nquestion answering and robotic manipulation using functional 3D scene graphs.\nSee our project page at https://openfungraph.github.io\n", "date": "2025-03-24", "categories": []}
{"id": "2412.13193", "title": "GaussTR: Foundation Model-Aligned Gaussian Transformer for\n  Self-Supervised 3D Spatial Understanding", "abstract": "  3D Semantic Occupancy Prediction is fundamental for spatial understanding,\nyet existing approaches face challenges in scalability and generalization due\nto their reliance on extensive labeled data and computationally intensive\nvoxel-wise representations. In this paper, we introduce GaussTR, a novel\nGaussian-based Transformer framework that unifies sparse 3D modeling with\nfoundation model alignment through Gaussian representations to advance 3D\nspatial understanding. GaussTR predicts sparse sets of Gaussians in a\nfeed-forward manner to represent 3D scenes. By splatting the Gaussians into 2D\nviews and aligning the rendered features with foundation models, GaussTR\nfacilitates self-supervised 3D representation learning and enables\nopen-vocabulary semantic occupancy prediction without requiring explicit\nannotations. Empirical experiments on the Occ3D-nuScenes dataset demonstrate\nGaussTR's state-of-the-art zero-shot performance of 12.27 mIoU, along with a\n40% reduction in training time. These results highlight the efficacy of GaussTR\nfor scalable and holistic 3D spatial understanding, with promising implications\nin autonomous driving and embodied agents. The code is available at\nhttps://github.com/hustvl/GaussTR.\n", "date": "2024-12-17", "categories": []}
{"id": "2503.14912", "title": "Deep Polycuboid Fitting for Compact 3D Representation of Indoor Scenes", "abstract": "  This paper presents a novel framework for compactly representing a 3D indoor\nscene using a set of polycuboids through a deep learning-based fitting method.\nIndoor scenes mainly consist of man-made objects, such as furniture, which\noften exhibit rectilinear geometry. This property allows indoor scenes to be\nrepresented using combinations of polycuboids, providing a compact\nrepresentation that benefits downstream applications like furniture\nrearrangement. Our framework takes a noisy point cloud as input and first\ndetects six types of cuboid faces using a transformer network. Then, a graph\nneural network is used to validate the spatial relationships of the detected\nfaces to form potential polycuboids. Finally, each polycuboid instance is\nreconstructed by forming a set of boxes based on the aggregated face labels. To\ntrain our networks, we introduce a synthetic dataset encompassing a diverse\nrange of cuboid and polycuboid shapes that reflect the characteristics of\nindoor scenes. Our framework generalizes well to real-world indoor scene\ndatasets, including Replica, ScanNet, and scenes captured with an iPhone. The\nversatility of our method is demonstrated through practical applications, such\nas virtual room tours and scene editing.\n", "date": "2025-03-19", "categories": []}
{"id": "2503.18073", "title": "PanopticSplatting: End-to-End Panoptic Gaussian Splatting", "abstract": "  Open-vocabulary panoptic reconstruction is a challenging task for\nsimultaneous scene reconstruction and understanding. Recently, methods have\nbeen proposed for 3D scene understanding based on Gaussian splatting. However,\nthese methods are multi-staged, suffering from the accumulated errors and the\ndependence of hand-designed components. To streamline the pipeline and achieve\nglobal optimization, we propose PanopticSplatting, an end-to-end system for\nopen-vocabulary panoptic reconstruction. Our method introduces query-guided\nGaussian segmentation with local cross attention, lifting 2D instance masks\nwithout cross-frame association in an end-to-end way. The local cross attention\nwithin view frustum effectively reduces the training memory, making our model\nmore accessible to large scenes with more Gaussians and objects. In addition,\nto address the challenge of noisy labels in 2D pseudo masks, we propose label\nblending to promote consistent 3D segmentation with less noisy floaters, as\nwell as label warping on 2D predictions which enhances multi-view coherence and\nsegmentation accuracy. Our method demonstrates strong performances in 3D scene\npanoptic reconstruction on the ScanNet-V2 and ScanNet++ datasets, compared with\nboth NeRF-based and Gaussian-based panoptic reconstruction methods. Moreover,\nPanopticSplatting can be easily generalized to numerous variants of Gaussian\nsplatting, and we demonstrate its robustness on different Gaussian base models.\n", "date": "2025-03-23", "categories": []}
{"id": "2503.18107", "title": "PanoGS: Gaussian-based Panoptic Segmentation for 3D Open Vocabulary\n  Scene Understanding", "abstract": "  Recently, 3D Gaussian Splatting (3DGS) has shown encouraging performance for\nopen vocabulary scene understanding tasks. However, previous methods cannot\ndistinguish 3D instance-level information, which usually predicts a heatmap\nbetween the scene feature and text query. In this paper, we propose PanoGS, a\nnovel and effective 3D panoptic open vocabulary scene understanding approach.\nTechnically, to learn accurate 3D language features that can scale to large\nindoor scenarios, we adopt the pyramid tri-plane to model the latent continuous\nparametric feature space and use a 3D feature decoder to regress the multi-view\nfused 2D feature cloud. Besides, we propose language-guided graph cuts that\nsynergistically leverage reconstructed geometry and learned language cues to\ngroup 3D Gaussian primitives into a set of super-primitives. To obtain 3D\nconsistent instance, we perform graph clustering based segmentation with\nSAM-guided edge affinity computation between different super-primitives.\nExtensive experiments on widely used datasets show better or more competitive\nperformance on 3D panoptic open vocabulary scene understanding. Project page:\n\\href{https://zju3dv.github.io/panogs}{https://zju3dv.github.io/panogs}.\n", "date": "2025-03-23", "categories": []}
{"id": "2412.01550", "title": "SeqAfford: Sequential 3D Affordance Reasoning via Multimodal Large\n  Language Model", "abstract": "  3D affordance segmentation aims to link human instructions to touchable\nregions of 3D objects for embodied manipulations. Existing efforts typically\nadhere to single-object, single-affordance paradigms, where each affordance\ntype or explicit instruction strictly corresponds to a specific affordance\nregion and are unable to handle long-horizon tasks. Such a paradigm cannot\nactively reason about complex user intentions that often imply sequential\naffordances. In this paper, we introduce the Sequential 3D Affordance Reasoning\ntask, which extends the traditional paradigm by reasoning from cumbersome user\nintentions and then decomposing them into a series of segmentation maps. Toward\nthis, we construct the first instruction-based affordance segmentation\nbenchmark that includes reasoning over both single and sequential affordances,\ncomprising 180K instruction-point cloud pairs. Based on the benchmark, we\npropose our model, SeqAfford, to unlock the 3D multi-modal large language model\nwith additional affordance segmentation abilities, which ensures reasoning with\nworld knowledge and fine-grained affordance grounding in a cohesive framework.\nWe further introduce a multi-granular language-point integration module to\nendow 3D dense prediction. Extensive experimental evaluations show that our\nmodel excels over well-established methods and exhibits open-world\ngeneralization with sequential reasoning abilities.\n", "date": "2024-12-02", "categories": []}
{"id": "2503.17044", "title": "ExCap3D: Expressive 3D Scene Understanding via Object Captioning with\n  Varying Detail", "abstract": "  Generating text descriptions of objects in 3D indoor scenes is an important\nbuilding block of embodied understanding. Existing methods do this by\ndescribing objects at a single level of detail, which often does not capture\nfine-grained details such as varying textures, materials, and shapes of the\nparts of objects. We propose the task of expressive 3D captioning: given an\ninput 3D scene, describe objects at multiple levels of detail: a high-level\nobject description, and a low-level description of the properties of its parts.\nTo produce such captions, we present ExCap3D, an expressive 3D captioning model\nwhich takes as input a 3D scan, and for each detected object in the scan,\ngenerates a fine-grained collective description of the parts of the object,\nalong with an object-level description conditioned on the part-level\ndescription. We design ExCap3D to encourage semantic consistency between the\ngenerated text descriptions, as well as textual similarity in the latent space,\nto further increase the quality of the generated captions. To enable this task,\nwe generated the ExCap3D Dataset by leveraging a visual-language model (VLM)\nfor multi-view captioning. The ExCap3D Dataset contains captions on the\nScanNet++ dataset with varying levels of detail, comprising 190k text\ndescriptions of 34k 3D objects in 947 indoor scenes. Our experiments show that\nthe object- and part-level of detail captions generated by ExCap3D are of\nhigher quality than those produced by state-of-the-art methods, with a Cider\nscore improvement of 17% and 124% for object- and part-level details\nrespectively. Our code, dataset and models will be made publicly available.\n", "date": "2025-03-21", "categories": []}
{"id": "2410.13924", "title": "ARKit LabelMaker: A New Scale for Indoor 3D Scene Understanding", "abstract": "  Neural network performance scales with both model size and data volume, as\nshown in both language and image processing. This requires scaling-friendly\narchitectures and large datasets. While transformers have been adapted for 3D\nvision, a `GPT-moment' remains elusive due to limited training data. We\nintroduce ARKit LabelMaker, a large-scale real-world 3D dataset with dense\nsemantic annotation that is more than three times larger than prior largest\ndataset. Specifically, we extend ARKitScenes with automatically generated dense\n3D labels using an extended LabelMaker pipeline, tailored for large-scale\npre-training. Training on our dataset improves accuracy across architectures,\nachieving state-of-the-art 3D semantic segmentation scores on ScanNet and\nScanNet200, with notable gains on tail classes. Our code is available at\nhttps://labelmaker.org and our dataset at\nhttps://huggingface.co/datasets/labelmaker/arkit_labelmaker.\n", "date": "2024-10-17", "categories": []}
{"id": "2503.15712", "title": "SPNeRF: Open Vocabulary 3D Neural Scene Segmentation with Superpoints", "abstract": "  Open-vocabulary segmentation, powered by large visual-language models like\nCLIP, has expanded 2D segmentation capabilities beyond fixed classes predefined\nby the dataset, enabling zero-shot understanding across diverse scenes.\nExtending these capabilities to 3D segmentation introduces challenges, as\nCLIP's image-based embeddings often lack the geometric detail necessary for 3D\nscene segmentation. Recent methods tend to address this by introducing\nadditional segmentation models or replacing CLIP with variations trained on\nsegmentation data, which lead to redundancy or loss on CLIP's general language\ncapabilities. To overcome this limitation, we introduce SPNeRF, a NeRF based\nzero-shot 3D segmentation approach that leverages geometric priors. We\nintegrate geometric primitives derived from the 3D scene into NeRF training to\nproduce primitive-wise CLIP features, avoiding the ambiguity of point-wise\nfeatures. Additionally, we propose a primitive-based merging mechanism enhanced\nwith affinity scores. Without relying on additional segmentation models, our\nmethod further explores CLIP's capability for 3D segmentation and achieves\nnotable improvements over original LERF.\n", "date": "2025-03-19", "categories": []}
{"id": "2503.16399", "title": "SA-Occ: Satellite-Assisted 3D Occupancy Prediction in Real World", "abstract": "  Existing vision-based 3D occupancy prediction methods are inherently limited\nin accuracy due to their exclusive reliance on street-view imagery, neglecting\nthe potential benefits of incorporating satellite views. We propose SA-Occ, the\nfirst Satellite-Assisted 3D occupancy prediction model, which leverages GPS &\nIMU to integrate historical yet readily available satellite imagery into\nreal-time applications, effectively mitigating limitations of ego-vehicle\nperceptions, involving occlusions and degraded performance in distant regions.\nTo address the core challenges of cross-view perception, we propose: 1)\nDynamic-Decoupling Fusion, which resolves inconsistencies in dynamic regions\ncaused by the temporal asynchrony between satellite and street views; 2)\n3D-Proj Guidance, a module that enhances 3D feature extraction from inherently\n2D satellite imagery; and 3) Uniform Sampling Alignment, which aligns the\nsampling density between street and satellite views. Evaluated on\nOcc3D-nuScenes, SA-Occ achieves state-of-the-art performance, especially among\nsingle-frame methods, with a 39.05% mIoU (a 6.97% improvement), while incurring\nonly 6.93 ms of additional latency per frame. Our code and newly curated\ndataset are available at https://github.com/chenchen235/SA-Occ.\n", "date": "2025-03-20", "categories": []}
{"id": "2503.16413", "title": "M3: 3D-Spatial MultiModal Memory", "abstract": "  We present 3D Spatial MultiModal Memory (M3), a multimodal memory system\ndesigned to retain information about medium-sized static scenes through video\nsources for visual perception. By integrating 3D Gaussian Splatting techniques\nwith foundation models, M3 builds a multimodal memory capable of rendering\nfeature representations across granularities, encompassing a wide range of\nknowledge. In our exploration, we identify two key challenges in previous works\non feature splatting: (1) computational constraints in storing high-dimensional\nfeatures for each Gaussian primitive, and (2) misalignment or information loss\nbetween distilled features and foundation model features. To address these\nchallenges, we propose M3 with key components of principal scene components and\nGaussian memory attention, enabling efficient training and inference. To\nvalidate M3, we conduct comprehensive quantitative evaluations of feature\nsimilarity and downstream tasks, as well as qualitative visualizations to\nhighlight the pixel trace of Gaussian memory attention. Our approach\nencompasses a diverse range of foundation models, including vision-language\nmodels (VLMs), perception models, and large multimodal and language models\n(LMMs/LLMs). Furthermore, to demonstrate real-world applicability, we deploy\nM3's feature field in indoor scenes on a quadruped robot. Notably, we claim\nthat M3 is the first work to address the core compression challenges in 3D\nfeature distillation.\n", "date": "2025-03-20", "categories": []}
{"id": "2503.15019", "title": "Learning 4D Panoptic Scene Graph Generation from Rich 2D Visual Scene", "abstract": "  The latest emerged 4D Panoptic Scene Graph (4D-PSG) provides an advanced-ever\nrepresentation for comprehensively modeling the dynamic 4D visual real world.\nUnfortunately, current pioneering 4D-PSG research can primarily suffer from\ndata scarcity issues severely, as well as the resulting out-of-vocabulary\nproblems; also, the pipeline nature of the benchmark generation method can lead\nto suboptimal performance. To address these challenges, this paper investigates\na novel framework for 4D-PSG generation that leverages rich 2D visual scene\nannotations to enhance 4D scene learning. First, we introduce a 4D Large\nLanguage Model (4D-LLM) integrated with a 3D mask decoder for end-to-end\ngeneration of 4D-PSG. A chained SG inference mechanism is further designed to\nexploit LLMs' open-vocabulary capabilities to infer accurate and comprehensive\nobject and relation labels iteratively. Most importantly, we propose a 2D-to-4D\nvisual scene transfer learning framework, where a spatial-temporal scene\ntranscending strategy effectively transfers dimension-invariant features from\nabundant 2D SG annotations to 4D scenes, effectively compensating for data\nscarcity in 4D-PSG. Extensive experiments on the benchmark data demonstrate\nthat we strikingly outperform baseline models by a large margin, highlighting\nthe effectiveness of our method.\n", "date": "2025-03-19", "categories": []}
{"id": "2503.15091", "title": "Intelligent Spatial Perception by Building Hierarchical 3D Scene Graphs\n  for Indoor Scenarios with the Help of LLMs", "abstract": "  This paper addresses the high demand in advanced intelligent robot navigation\nfor a more holistic understanding of spatial environments, by introducing a\nnovel system that harnesses the capabilities of Large Language Models (LLMs) to\nconstruct hierarchical 3D Scene Graphs (3DSGs) for indoor scenarios. The\nproposed framework constructs 3DSGs consisting of a fundamental layer with rich\nmetric-semantic information, an object layer featuring precise point-cloud\nrepresentation of object nodes as well as visual descriptors, and higher layers\nof room, floor, and building nodes. Thanks to the innovative application of\nLLMs, not only object nodes but also nodes of higher layers, e.g., room nodes,\nare annotated in an intelligent and accurate manner. A polling mechanism for\nroom classification using LLMs is proposed to enhance the accuracy and\nreliability of the room node annotation. Thorough numerical experiments\ndemonstrate the system's ability to integrate semantic descriptions with\ngeometric data, creating an accurate and comprehensive representation of the\nenvironment instrumental for context-aware navigation and task planning.\n", "date": "2025-03-19", "categories": []}
{"id": "2503.15185", "title": "3D Occupancy Prediction with Low-Resolution Queries via Prototype-aware\n  View Transformation", "abstract": "  The resolution of voxel queries significantly influences the quality of view\ntransformation in camera-based 3D occupancy prediction. However, computational\nconstraints and the practical necessity for real-time deployment require\nsmaller query resolutions, which inevitably leads to an information loss.\nTherefore, it is essential to encode and preserve rich visual details within\nlimited query sizes while ensuring a comprehensive representation of 3D\noccupancy. To this end, we introduce ProtoOcc, a novel occupancy network that\nleverages prototypes of clustered image segments in view transformation to\nenhance low-resolution context. In particular, the mapping of 2D prototypes\nonto 3D voxel queries encodes high-level visual geometries and complements the\nloss of spatial information from reduced query resolutions. Additionally, we\ndesign a multi-perspective decoding strategy to efficiently disentangle the\ndensely compressed visual cues into a high-dimensional 3D occupancy scene.\nExperimental results on both Occ3D and SemanticKITTI benchmarks demonstrate the\neffectiveness of the proposed method, showing clear improvements over the\nbaselines. More importantly, ProtoOcc achieves competitive performance against\nthe baselines even with 75\\% reduced voxel resolution.\n", "date": "2025-03-19", "categories": []}
{"id": "2503.14029", "title": "Rethinking End-to-End 2D to 3D Scene Segmentation in Gaussian Splatting", "abstract": "  Lifting multi-view 2D instance segmentation to a radiance field has proven to\nbe effective to enhance 3D understanding. Existing methods rely on direct\nmatching for end-to-end lifting, yielding inferior results; or employ a\ntwo-stage solution constrained by complex pre- or post-processing. In this\nwork, we design a new end-to-end object-aware lifting approach, named\nUnified-Lift that provides accurate 3D segmentation based on the 3D Gaussian\nrepresentation. To start, we augment each Gaussian point with an additional\nGaussian-level feature learned using a contrastive loss to encode instance\ninformation. Importantly, we introduce a learnable object-level codebook to\naccount for individual objects in the scene for an explicit object-level\nunderstanding and associate the encoded object-level features with the\nGaussian-level point features for segmentation predictions. While promising,\nachieving effective codebook learning is non-trivial and a naive solution leads\nto degraded performance. Therefore, we formulate the association learning\nmodule and the noisy label filtering module for effective and robust codebook\nlearning. We conduct experiments on three benchmarks: LERF-Masked, Replica, and\nMessy Rooms datasets. Both qualitative and quantitative results manifest that\nour Unified-Lift clearly outperforms existing methods in terms of segmentation\nquality and time efficiency. The code is publicly available at\n\\href{https://github.com/Runsong123/Unified-Lift}{https://github.com/Runsong123/Unified-Lift}.\n", "date": "2025-03-18", "categories": []}
{"id": "2411.19492", "title": "Diorama: Unleashing Zero-shot Single-view 3D Indoor Scene Modeling", "abstract": "  Reconstructing structured 3D scenes from RGB images using CAD objects unlocks\nefficient and compact scene representations that maintain compositionality and\ninteractability. Existing works propose training-heavy methods relying on\neither expensive yet inaccurate real-world annotations or controllable yet\nmonotonous synthetic data that do not generalize well to unseen objects or\ndomains. We present Diorama, the first zero-shot open-world system that\nholistically models 3D scenes from single-view RGB observations without\nrequiring end-to-end training or human annotations. We show the feasibility of\nour approach by decomposing the problem into subtasks and introduce robust,\ngeneralizable solutions to each: architecture reconstruction, 3D shape\nretrieval, object pose estimation, and scene layout optimization. We evaluate\nour system on both synthetic and real-world data to show we significantly\noutperform baselines from prior work. We also demonstrate generalization to\ninternet images and the text-to-scene task.\n", "date": "2024-11-29", "categories": []}
{"id": "2503.12369", "title": "L2COcc: Lightweight Camera-Centric Semantic Scene Completion via\n  Distillation of LiDAR Model", "abstract": "  Semantic Scene Completion (SSC) constitutes a pivotal element in autonomous\ndriving perception systems, tasked with inferring the 3D semantic occupancy of\na scene from sensory data. To improve accuracy, prior research has implemented\nvarious computationally demanding and memory-intensive 3D operations, imposing\nsignificant computational requirements on the platform during training and\ntesting. This paper proposes L2COcc, a lightweight camera-centric SSC framework\nthat also accommodates LiDAR inputs. With our proposed efficient voxel\ntransformer (EVT) and cross-modal knowledge modules, including feature\nsimilarity distillation (FSD), TPV distillation (TPVD) and prediction alignment\ndistillation (PAD), our method substantially reduce computational burden while\nmaintaining high accuracy. The experimental evaluations demonstrate that our\nproposed method surpasses the current state-of-the-art vision-based SSC methods\nregarding accuracy on both the SemanticKITTI and SSCBench-KITTI-360 benchmarks,\nrespectively. Additionally, our method is more lightweight, exhibiting a\nreduction in both memory consumption and inference time by over 23% compared to\nthe current state-of-the-arts method. Code is available at our project\npage:https://studyingfufu.github.io/L2COcc/.\n", "date": "2025-03-16", "categories": []}
{"id": "2503.11345", "title": "EgoSplat: Open-Vocabulary Egocentric Scene Understanding with Language\n  Embedded 3D Gaussian Splatting", "abstract": "  Egocentric scenes exhibit frequent occlusions, varied viewpoints, and dynamic\ninteractions compared to typical scene understanding tasks. Occlusions and\nvaried viewpoints can lead to multi-view semantic inconsistencies, while\ndynamic objects may act as transient distractors, introducing artifacts into\nsemantic feature modeling. To address these challenges, we propose EgoSplat, a\nlanguage-embedded 3D Gaussian Splatting framework for open-vocabulary\negocentric scene understanding. A multi-view consistent instance feature\naggregation method is designed to leverage the segmentation and tracking\ncapabilities of SAM2 to selectively aggregate complementary features across\nviews for each instance, ensuring precise semantic representation of scenes.\nAdditionally, an instance-aware spatial-temporal transient prediction module is\nconstructed to improve spatial integrity and temporal continuity in predictions\nby incorporating spatial-temporal associations across multi-view instances,\neffectively reducing artifacts in the semantic reconstruction of egocentric\nscenes. EgoSplat achieves state-of-the-art performance in both localization and\nsegmentation tasks on two datasets, outperforming existing methods with a 8.2%\nimprovement in localization accuracy and a 3.7% improvement in segmentation\nmIoU on the ADT dataset, and setting a new benchmark in open-vocabulary\negocentric scene understanding. The code will be made publicly available.\n", "date": "2025-03-14", "categories": []}
{"id": "2503.11651", "title": "VGGT: Visual Geometry Grounded Transformer", "abstract": "  We present VGGT, a feed-forward neural network that directly infers all key\n3D attributes of a scene, including camera parameters, point maps, depth maps,\nand 3D point tracks, from one, a few, or hundreds of its views. This approach\nis a step forward in 3D computer vision, where models have typically been\nconstrained to and specialized for single tasks. It is also simple and\nefficient, reconstructing images in under one second, and still outperforming\nalternatives that require post-processing with visual geometry optimization\ntechniques. The network achieves state-of-the-art results in multiple 3D tasks,\nincluding camera parameter estimation, multi-view depth estimation, dense point\ncloud reconstruction, and 3D point tracking. We also show that using pretrained\nVGGT as a feature backbone significantly enhances downstream tasks, such as\nnon-rigid point tracking and feed-forward novel view synthesis. Code and models\nare publicly available at https://github.com/facebookresearch/vggt.\n", "date": "2025-03-14", "categories": []}
{"id": "2411.13753", "title": "FAST-Splat: Fast, Ambiguity-Free Semantics Transfer in Gaussian\n  Splatting", "abstract": "  We present FAST-Splat for fast, ambiguity-free semantic Gaussian Splatting,\nwhich seeks to address the main limitations of existing semantic Gaussian\nSplatting methods, namely: slow training and rendering speeds; high memory\nusage; and ambiguous semantic object localization. We take a bottom-up approach\nin deriving FAST-Splat, dismantling the limitations of closed-set semantic\ndistillation to enable open-set (open-vocabulary) semantic distillation.\nUltimately, this key approach enables FAST-Splat to provide precise semantic\nobject localization results, even when prompted with ambiguous user-provided\nnatural-language queries. Further, by exploiting the explicit form of the\nGaussian Splatting scene representation to the fullest extent, FAST-Splat\nretains the remarkable training and rendering speeds of Gaussian Splatting.\nPrecisely, while existing semantic Gaussian Splatting methods distill semantics\ninto a separate neural field or utilize neural models for dimensionality\nreduction, FAST-Splat directly augments each Gaussian with specific semantic\ncodes, preserving the training, rendering, and memory-usage advantages of\nGaussian Splatting over neural field methods. These Gaussian-specific semantic\ncodes, together with a hash-table, enable semantic similarity to be measured\nwith open-vocabulary user prompts and further enable FAST-Splat to respond with\nunambiguous semantic object labels and $3$D masks, unlike prior methods. In\nexperiments, we demonstrate that FAST-Splat is 6x to 8x faster to train,\nachieves between 18x to 51x faster rendering speeds, and requires about 6x\nsmaller GPU memory, compared to the best-competing semantic Gaussian Splatting\nmethods. Further, FAST-Splat achieves relatively similar or better semantic\nsegmentation performance compared to existing methods. After the review period,\nwe will provide links to the project website and the codebase.\n", "date": "2024-11-20", "categories": []}
{"id": "2408.09859", "title": "OccMamba: Semantic Occupancy Prediction with State Space Models", "abstract": "  Training deep learning models for semantic occupancy prediction is\nchallenging due to factors such as a large number of occupancy cells, severe\nocclusion, limited visual cues, complicated driving scenarios, etc. Recent\nmethods often adopt transformer-based architectures given their strong\ncapability in learning input-conditioned weights and long-range relationships.\nHowever, transformer-based networks are notorious for their quadratic\ncomputation complexity, seriously undermining their efficacy and deployment in\nsemantic occupancy prediction. Inspired by the global modeling and linear\ncomputation complexity of the Mamba architecture, we present the first\nMamba-based network for semantic occupancy prediction, termed OccMamba.\nSpecifically, we first design the hierarchical Mamba module and local context\nprocessor to better aggregate global and local contextual information,\nrespectively. Besides, to relieve the inherent domain gap between the\nlinguistic and 3D domains, we present a simple yet effective 3D-to-1D\nreordering scheme, i.e., height-prioritized 2D Hilbert expansion. It can\nmaximally retain the spatial structure of 3D voxels as well as facilitate the\nprocessing of Mamba blocks. Endowed with the aforementioned designs, our\nOccMamba is capable of directly and efficiently processing large volumes of\ndense scene grids, achieving state-of-the-art performance across three\nprevalent occupancy prediction benchmarks, including OpenOccupancy,\nSemanticKITTI, and SemanticPOSS. Notably, on OpenOccupancy, our OccMamba\noutperforms the previous state-of-the-art Co-Occ by 5.1% IoU and 4.3% mIoU,\nrespectively. Our implementation is open-sourced and available at:\nhttps://github.com/USTCLH/OccMamba.\n", "date": "2024-08-19", "categories": []}
{"id": "2411.19162", "title": "Lost & Found: Tracking Changes from Egocentric Observations in 3D\n  Dynamic Scene Graphs", "abstract": "  Recent approaches have successfully focused on the segmentation of static\nreconstructions, thereby equipping downstream applications with semantic 3D\nunderstanding. However, the world in which we live is dynamic, characterized by\nnumerous interactions between the environment and humans or robotic agents.\nStatic semantic maps are unable to capture this information, and the naive\nsolution of rescanning the environment after every change is both costly and\nineffective in tracking e.g. objects being stored away in drawers. With Lost &\nFound we present an approach that addresses this limitation. Based solely on\negocentric recordings with corresponding hand position and camera pose\nestimates, we are able to track the 6DoF poses of the moving object within the\ndetected interaction interval. These changes are applied online to a\ntransformable scene graph that captures object-level relations. Compared to\nstate-of-the-art object pose trackers, our approach is more reliable in\nhandling the challenging egocentric viewpoint and the lack of depth\ninformation. It outperforms the second-best approach by 34% and 56% for\ntranslational and orientational error, respectively, and produces visibly\nsmoother 6DoF object trajectories. In addition, we illustrate how the acquired\ninteraction information in the dynamic scene graph can be employed in the\ncontext of robotic applications that would otherwise be unfeasible: We show how\nour method allows to command a mobile manipulator through teach & repeat, and\nhow information about prior interaction allows a mobile manipulator to retrieve\nan object hidden in a drawer. Code, videos and corresponding data are\naccessible at https://behretj.github.io/LostAndFound.\n", "date": "2024-11-28", "categories": []}
{"id": "2503.07739", "title": "SIRE: SE(3) Intrinsic Rigidity Embeddings", "abstract": "  Motion serves as a powerful cue for scene perception and understanding by\nseparating independently moving surfaces and organizing the physical world into\ndistinct entities. We introduce SIRE, a self-supervised method for motion\ndiscovery of objects and dynamic scene reconstruction from casual scenes by\nlearning intrinsic rigidity embeddings from videos. Our method trains an image\nencoder to estimate scene rigidity and geometry, supervised by a simple 4D\nreconstruction loss: a least-squares solver uses the estimated geometry and\nrigidity to lift 2D point track trajectories into SE(3) tracks, which are\nsimply re-projected back to 2D and compared against the original 2D\ntrajectories for supervision. Crucially, our framework is fully end-to-end\ndifferentiable and can be optimized either on video datasets to learn\ngeneralizable image priors, or even on a single video to capture scene-specific\nstructure - highlighting strong data efficiency. We demonstrate the\neffectiveness of our rigidity embeddings and geometry across multiple settings,\nincluding downstream object segmentation, SE(3) rigid motion estimation, and\nself-supervised depth estimation. Our findings suggest that SIRE can learn\nstrong geometry and motion rigidity priors from video data, with minimal\nsupervision.\n", "date": "2025-03-10", "categories": []}
{"id": "2503.08471", "title": "TrackOcc: Camera-based 4D Panoptic Occupancy Tracking", "abstract": "  Comprehensive and consistent dynamic scene understanding from camera input is\nessential for advanced autonomous systems. Traditional camera-based perception\ntasks like 3D object tracking and semantic occupancy prediction lack either\nspatial comprehensiveness or temporal consistency. In this work, we introduce a\nbrand-new task, Camera-based 4D Panoptic Occupancy Tracking, which\nsimultaneously addresses panoptic occupancy segmentation and object tracking\nfrom camera-only input. Furthermore, we propose TrackOcc, a cutting-edge\napproach that processes image inputs in a streaming, end-to-end manner with 4D\npanoptic queries to address the proposed task. Leveraging the\nlocalization-aware loss, TrackOcc enhances the accuracy of 4D panoptic\noccupancy tracking without bells and whistles. Experimental results demonstrate\nthat our method achieves state-of-the-art performance on the Waymo dataset. The\nsource code will be released at https://github.com/Tsinghua-MARS-Lab/TrackOcc.\n", "date": "2025-03-11", "categories": []}
{"id": "2503.08512", "title": "SAS: Segment Any 3D Scene with Integrated 2D Priors", "abstract": "  The open vocabulary capability of 3D models is increasingly valued, as\ntraditional methods with models trained with fixed categories fail to recognize\nunseen objects in complex dynamic 3D scenes. In this paper, we propose a simple\nyet effective approach, SAS, to integrate the open vocabulary capability of\nmultiple 2D models and migrate it to 3D domain. Specifically, we first propose\nModel Alignment via Text to map different 2D models into the same embedding\nspace using text as a bridge. Then, we propose Annotation-Free Model Capability\nConstruction to explicitly quantify the 2D model's capability of recognizing\ndifferent categories using diffusion models. Following this, point cloud\nfeatures from different 2D models are fused with the guide of constructed model\ncapabilities. Finally, the integrated 2D open vocabulary capability is\ntransferred to 3D domain through feature distillation. SAS outperforms previous\nmethods by a large margin across multiple datasets, including ScanNet v2,\nMatterport3D, and nuScenes, while its generalizability is further validated on\ndownstream tasks, e.g., gaussian segmentation and instance segmentation.\n", "date": "2025-03-11", "categories": []}
{"id": "2409.12518", "title": "Hier-SLAM: Scaling-up Semantics in SLAM with a Hierarchically\n  Categorical Gaussian Splatting", "abstract": "  We propose Hier-SLAM, a semantic 3D Gaussian Splatting SLAM method featuring\na novel hierarchical categorical representation, which enables accurate global\n3D semantic mapping, scaling-up capability, and explicit semantic label\nprediction in the 3D world. The parameter usage in semantic SLAM systems\nincreases significantly with the growing complexity of the environment, making\nit particularly challenging and costly for scene understanding. To address this\nproblem, we introduce a novel hierarchical representation that encodes semantic\ninformation in a compact form into 3D Gaussian Splatting, leveraging the\ncapabilities of large language models (LLMs). We further introduce a novel\nsemantic loss designed to optimize hierarchical semantic information through\nboth inter-level and cross-level optimization. Furthermore, we enhance the\nwhole SLAM system, resulting in improved tracking and mapping performance. Our\n\\MethodName{} outperforms existing dense SLAM methods in both mapping and\ntracking accuracy, while achieving a 2x operation speed-up. Additionally, it\nachieves on-par semantic rendering performance compared to existing methods\nwhile significantly reducing storage and training time requirements. Rendering\nFPS impressively reaches 2,000 with semantic information and 3,000 without it.\nMost notably, it showcases the capability of handling the complex real-world\nscene with more than 500 semantic classes, highlighting its valuable scaling-up\ncapability. The open-source code is available at\nhttps://github.com/LeeBY68/Hier-SLAM\n", "date": "2024-09-19", "categories": []}
{"id": "2411.11466", "title": "MGNiceNet: Unified Monocular Geometric Scene Understanding", "abstract": "  Monocular geometric scene understanding combines panoptic segmentation and\nself-supervised depth estimation, focusing on real-time application in\nautonomous vehicles. We introduce MGNiceNet, a unified approach that uses a\nlinked kernel formulation for panoptic segmentation and self-supervised depth\nestimation. MGNiceNet is based on the state-of-the-art real-time panoptic\nsegmentation method RT-K-Net and extends the architecture to cover both\npanoptic segmentation and self-supervised monocular depth estimation. To this\nend, we introduce a tightly coupled self-supervised depth estimation predictor\nthat explicitly uses information from the panoptic path for depth prediction.\nFurthermore, we introduce a panoptic-guided motion masking method to improve\ndepth estimation without relying on video panoptic segmentation annotations. We\nevaluate our method on two popular autonomous driving datasets, Cityscapes and\nKITTI. Our model shows state-of-the-art results compared to other real-time\nmethods and closes the gap to computationally more demanding methods. Source\ncode and trained models are available at\nhttps://github.com/markusschoen/MGNiceNet.\n", "date": "2024-11-18", "categories": []}
{"id": "2412.01807", "title": "Occam's LGS: An Efficient Approach for Language Gaussian Splatting", "abstract": "  TL;DR: Gaussian Splatting is a widely adopted approach for 3D scene\nrepresentation, offering efficient, high-quality reconstruction and rendering.\nA key reason for its success is the simplicity of representing scenes with sets\nof Gaussians, making it interpretable and adaptable. To enhance understanding\nbeyond visual representation, recent approaches extend Gaussian Splatting with\nsemantic vision-language features, enabling open-set tasks. Typically, these\nlanguage features are aggregated from multiple 2D views, however, existing\nmethods rely on cumbersome techniques, resulting in high computational costs\nand longer training times.\n  In this work, we show that the complicated pipelines for language 3D Gaussian\nSplatting are simply unnecessary. Instead, we follow a probabilistic\nformulation of Language Gaussian Splatting and apply Occam's razor to the task\nat hand, leading to a highly efficient weighted multi-view feature aggregation\ntechnique. Doing so offers us state-of-the-art results with a speed-up of two\norders of magnitude without any compression, allowing for easy scene\nmanipulation. Project Page: https://insait-institute.github.io/OccamLGS/\n", "date": "2024-12-02", "categories": []}
{"id": "2412.13654", "title": "GAGS: Granularity-Aware Feature Distillation for Language Gaussian\n  Splatting", "abstract": "  3D open-vocabulary scene understanding, which accurately perceives complex\nsemantic properties of objects in space, has gained significant attention in\nrecent years. In this paper, we propose GAGS, a framework that distills 2D CLIP\nfeatures into 3D Gaussian splatting, enabling open-vocabulary queries for\nrenderings on arbitrary viewpoints. The main challenge of distilling 2D\nfeatures for 3D fields lies in the multiview inconsistency of extracted 2D\nfeatures, which provides unstable supervision for the 3D feature field. GAGS\naddresses this challenge with two novel strategies. First, GAGS associates the\nprompt point density of SAM with the camera distances, which significantly\nimproves the multiview consistency of segmentation results. Second, GAGS\nfurther decodes a granularity factor to guide the distillation process and this\ngranularity factor can be learned in a unsupervised manner to only select the\nmultiview consistent 2D features in the distillation process. Experimental\nresults on two datasets demonstrate significant performance and stability\nimprovements of GAGS in visual grounding and semantic segmentation, with an\ninference speed 2$\\times$ faster than baseline methods. The code and additional\nresults are available at https://pz0826.github.io/GAGS-Webpage/ .\n", "date": "2024-12-18", "categories": []}
{"id": "2503.06014", "title": "Towards Ambiguity-Free Spatial Foundation Model: Rethinking and\n  Decoupling Depth Ambiguity", "abstract": "  Depth ambiguity is a fundamental challenge in spatial scene understanding,\nespecially in transparent scenes where single-depth estimates fail to capture\nfull 3D structure. Existing models, limited to deterministic predictions,\noverlook real-world multi-layer depth. To address this, we introduce a paradigm\nshift from single-prediction to multi-hypothesis spatial foundation models. We\nfirst present \\texttt{MD-3k}, a benchmark exposing depth biases in expert and\nfoundational models through multi-layer spatial relationship labels and new\nmetrics. To resolve depth ambiguity, we propose Laplacian Visual Prompting\n(LVP), a training-free spectral prompting technique that extracts hidden depth\nfrom pre-trained models via Laplacian-transformed RGB inputs. By integrating\nLVP-inferred depth with standard RGB-based estimates, our approach elicits\nmulti-layer depth without model retraining. Extensive experiments validate the\neffectiveness of LVP in zero-shot multi-layer depth estimation, unlocking more\nrobust and comprehensive geometry-conditioned visual generation, 3D-grounded\nspatial reasoning, and temporally consistent video-level depth inference. Our\nbenchmark and code will be available at\nhttps://github.com/Xiaohao-Xu/Ambiguity-in-Space.\n", "date": "2025-03-08", "categories": []}
{"id": "2503.06219", "title": "VLScene: Vision-Language Guidance Distillation for Camera-Based 3D\n  Semantic Scene Completion", "abstract": "  Camera-based 3D semantic scene completion (SSC) provides dense geometric and\nsemantic perception for autonomous driving. However, images provide limited\ninformation making the model susceptible to geometric ambiguity caused by\nocclusion and perspective distortion. Existing methods often lack explicit\nsemantic modeling between objects, limiting their perception of 3D semantic\ncontext. To address these challenges, we propose a novel method VLScene:\nVision-Language Guidance Distillation for Camera-based 3D Semantic Scene\nCompletion. The key insight is to use the vision-language model to introduce\nhigh-level semantic priors to provide the object spatial context required for\n3D scene understanding. Specifically, we design a vision-language guidance\ndistillation process to enhance image features, which can effectively capture\nsemantic knowledge from the surrounding environment and improve spatial context\nreasoning. In addition, we introduce a geometric-semantic sparse awareness\nmechanism to propagate geometric structures in the neighborhood and enhance\nsemantic information through contextual sparse interactions. Experimental\nresults demonstrate that VLScene achieves rank-1st performance on challenging\nbenchmarks--SemanticKITTI and SSCBench-KITTI-360, yielding remarkably mIoU\nscores of 17.52 and 19.10, respectively.\n", "date": "2025-03-08", "categories": []}
{"id": "2503.06469", "title": "Vector Quantized Feature Fields for Fast 3D Semantic Lifting", "abstract": "  We generalize lifting to semantic lifting by incorporating per-view masks\nthat indicate relevant pixels for lifting tasks. These masks are determined by\nquerying corresponding multiscale pixel-aligned feature maps, which are derived\nfrom scene representations such as distilled feature fields and feature point\nclouds. However, storing per-view feature maps rendered from distilled feature\nfields is impractical, and feature point clouds are expensive to store and\nquery. To enable lightweight on-demand retrieval of pixel-aligned relevance\nmasks, we introduce the Vector-Quantized Feature Field. We demonstrate the\neffectiveness of the Vector-Quantized Feature Field on complex indoor and\noutdoor scenes. Semantic lifting, when paired with a Vector-Quantized Feature\nField, can unlock a myriad of applications in scene representation and embodied\nintelligence. Specifically, we showcase how our method enables text-driven\nlocalized scene editing and significantly improves the efficiency of embodied\nquestion answering.\n", "date": "2025-03-09", "categories": []}
{"id": "2503.06569", "title": "Global-Aware Monocular Semantic Scene Completion with State Space Models", "abstract": "  Monocular Semantic Scene Completion (MonoSSC) reconstructs and interprets 3D\nenvironments from a single image, enabling diverse real-world applications.\nHowever, existing methods are often constrained by the local receptive field of\nConvolutional Neural Networks (CNNs), making it challenging to handle the\nnon-uniform distribution of projected points (Fig. \\ref{fig:perspective}) and\neffectively reconstruct missing information caused by the 3D-to-2D projection.\nIn this work, we introduce GA-MonoSSC, a hybrid architecture for MonoSSC that\neffectively captures global context in both the 2D image domain and 3D space.\nSpecifically, we propose a Dual-Head Multi-Modality Encoder, which leverages a\nTransformer architecture to capture spatial relationships across all features\nin the 2D image domain, enabling more comprehensive 2D feature extraction.\nAdditionally, we introduce the Frustum Mamba Decoder, built on the State Space\nModel (SSM), to efficiently capture long-range dependencies in 3D space.\nFurthermore, we propose a frustum reordering strategy within the Frustum Mamba\nDecoder to mitigate feature discontinuities in the reordered voxel sequence,\nensuring better alignment with the scan mechanism of the State Space Model\n(SSM) for improved 3D representation learning. We conduct extensive experiments\non the widely used Occ-ScanNet and NYUv2 datasets, demonstrating that our\nproposed method achieves state-of-the-art performance, validating its\neffectiveness. The code will be released upon acceptance.\n", "date": "2025-03-09", "categories": []}
{"id": "2503.07125", "title": "Learning A Zero-shot Occupancy Network from Vision Foundation Models via\n  Self-supervised Adaptation", "abstract": "  Estimating the 3D world from 2D monocular images is a fundamental yet\nchallenging task due to the labour-intensive nature of 3D annotations. To\nsimplify label acquisition, this work proposes a novel approach that bridges 2D\nvision foundation models (VFMs) with 3D tasks by decoupling 3D supervision into\nan ensemble of image-level primitives, e.g., semantic and geometric components.\nAs a key motivator, we leverage the zero-shot capabilities of vision-language\nmodels for image semantics. However, due to the notorious ill-posed problem -\nmultiple distinct 3D scenes can produce identical 2D projections, directly\ninferring metric depth from a monocular image in a zero-shot manner is\nunsuitable. In contrast, 2D VFMs provide promising sources of relative depth,\nwhich theoretically aligns with metric depth when properly scaled and offset.\nThus, we adapt the relative depth derived from VFMs into metric depth by\noptimising the scale and offset using temporal consistency, also known as novel\nview synthesis, without access to ground-truth metric depth. Consequently, we\nproject the semantics into 3D space using the reconstructed metric depth,\nthereby providing 3D supervision. Extensive experiments on nuScenes and\nSemanticKITTI demonstrate the effectiveness of our framework. For instance, the\nproposed method surpasses the current state-of-the-art by 3.34% mIoU on\nnuScenes for voxel occupancy prediction.\n", "date": "2025-03-10", "categories": []}
{"id": "2503.07593", "title": "Hierarchical Cross-Modal Alignment for Open-Vocabulary 3D Object\n  Detection", "abstract": "  Open-vocabulary 3D object detection (OV-3DOD) aims at localizing and\nclassifying novel objects beyond closed sets. The recent success of\nvision-language models (VLMs) has demonstrated their remarkable capabilities to\nunderstand open vocabularies. Existing works that leverage VLMs for 3D object\ndetection (3DOD) generally resort to representations that lose the rich scene\ncontext required for 3D perception. To address this problem, we propose in this\npaper a hierarchical framework, named HCMA, to simultaneously learn local\nobject and global scene information for OV-3DOD. Specifically, we first design\na Hierarchical Data Integration (HDI) approach to obtain coarse-to-fine\n3D-image-text data, which is fed into a VLM to extract object-centric\nknowledge. To facilitate the association of feature hierarchies, we then\npropose an Interactive Cross-Modal Alignment (ICMA) strategy to establish\neffective intra-level and inter-level feature connections. To better align\nfeatures across different levels, we further propose an Object-Focusing Context\nAdjustment (OFCA) module to refine multi-level features by emphasizing\nobject-related features. Extensive experiments demonstrate that the proposed\nmethod outperforms SOTA methods on the existing OV-3DOD benchmarks. It also\nachieves promising OV-3DOD results even without any 3D annotations.\n", "date": "2025-03-10", "categories": []}
{"id": "2503.05086", "title": "Fake It To Make It: Virtual Multiviews to Enhance Monocular Indoor\n  Semantic Scene Completion", "abstract": "  Monocular Indoor Semantic Scene Completion (SSC) aims to reconstruct a 3D\nsemantic occupancy map from a single RGB image of an indoor scene, inferring\nspatial layout and object categories from 2D image cues. The challenge of this\ntask arises from the depth, scale, and shape ambiguities that emerge when\ntransforming a 2D image into 3D space, particularly within the complex and\noften heavily occluded environments of indoor scenes. Current SSC methods often\nstruggle with these ambiguities, resulting in distorted or missing object\nrepresentations. To overcome these limitations, we introduce an innovative\napproach that leverages novel view synthesis and multiview fusion.\nSpecifically, we demonstrate how virtual cameras can be placed around the scene\nto emulate multiview inputs that enhance contextual scene information. We also\nintroduce a Multiview Fusion Adaptor (MVFA) to effectively combine the\nmultiview 3D scene predictions into a unified 3D semantic occupancy map.\nFinally, we identify and study the inherent limitation of generative techniques\nwhen applied to SSC, specifically the Novelty-Consistency tradeoff. Our system,\nGenFuSE, demonstrates IoU score improvements of up to 2.8% for Scene Completion\nand 4.9% for Semantic Scene Completion when integrated with existing SSC\nnetworks on the NYUv2 dataset. This work introduces GenFuSE as a standard\nframework for advancing monocular SSC with synthesized inputs.\n", "date": "2025-03-07", "categories": []}
{"id": "2503.04034", "title": "GaussianGraph: 3D Gaussian-based Scene Graph Generation for Open-world\n  Scene Understanding", "abstract": "  Recent advancements in 3D Gaussian Splatting(3DGS) have significantly\nimproved semantic scene understanding, enabling natural language queries to\nlocalize objects within a scene. However, existing methods primarily focus on\nembedding compressed CLIP features to 3D Gaussians, suffering from low object\nsegmentation accuracy and lack spatial reasoning capabilities. To address these\nlimitations, we propose GaussianGraph, a novel framework that enhances\n3DGS-based scene understanding by integrating adaptive semantic clustering and\nscene graph generation. We introduce a \"Control-Follow\" clustering strategy,\nwhich dynamically adapts to scene scale and feature distribution, avoiding\nfeature compression and significantly improving segmentation accuracy.\nAdditionally, we enrich scene representation by integrating object attributes\nand spatial relations extracted from 2D foundation models. To address\ninaccuracies in spatial relationships, we propose 3D correction modules that\nfilter implausible relations through spatial consistency verification, ensuring\nreliable scene graph construction. Extensive experiments on three datasets\ndemonstrate that GaussianGraph outperforms state-of-the-art methods in both\nsemantic segmentation and object grounding tasks, providing a robust solution\nfor complex scene understanding and interaction.\n", "date": "2025-03-06", "categories": []}
{"id": "2502.20041", "title": "3D-AffordanceLLM: Harnessing Large Language Models for Open-Vocabulary\n  Affordance Detection in 3D Worlds", "abstract": "  3D Affordance detection is a challenging problem with broad applications on\nvarious robotic tasks. Existing methods typically formulate the detection\nparadigm as a label-based semantic segmentation task. This paradigm relies on\npredefined labels and lacks the ability to comprehend complex natural language,\nresulting in limited generalization in open-world scene. To address these\nlimitations, we reformulate the traditional affordance detection paradigm into\n\\textit{Instruction Reasoning Affordance Segmentation} (IRAS) task. This task\nis designed to output a affordance mask region given a query reasoning text,\nwhich avoids fixed categories of input labels. We accordingly propose the\n\\textit{3D-AffordanceLLM} (3D-ADLLM), a framework designed for reasoning\naffordance detection in 3D open-scene. Specifically, 3D-ADLLM introduces large\nlanguage models (LLMs) to 3D affordance perception with a custom-designed\ndecoder for generating affordance masks, thus achieving open-world reasoning\naffordance detection. In addition, given the scarcity of 3D affordance datasets\nfor training large models, we seek to extract knowledge from general\nsegmentation data and transfer it to affordance detection. Thus, we propose a\nmulti-stage training strategy that begins with a novel pre-training task, i.e.,\n\\textit{Referring Object Part Segmentation}~(ROPS). This stage is designed to\nequip the model with general recognition and segmentation capabilities at the\nobject-part level. Then followed by fine-tuning with the IRAS task, 3D-ADLLM\nobtains the reasoning ability for affordance detection. In summary, 3D-ADLLM\nleverages the rich world knowledge and human-object interaction reasoning\nability of LLMs, achieving approximately an 8\\% improvement in mIoU on\nopen-vocabulary affordance detection tasks.\n", "date": "2025-02-27", "categories": []}
{"id": "2403.17010", "title": "Calib3D: Calibrating Model Preferences for Reliable 3D Scene\n  Understanding", "abstract": "  Safety-critical 3D scene understanding tasks necessitate not only accurate\nbut also confident predictions from 3D perception models. This study introduces\nCalib3D, a pioneering effort to benchmark and scrutinize the reliability of 3D\nscene understanding models from an uncertainty estimation viewpoint. We\ncomprehensively evaluate 28 state-of-the-art models across 10 diverse 3D\ndatasets, uncovering insightful phenomena that cope with both the aleatoric and\nepistemic uncertainties in 3D scene understanding. We discover that despite\nachieving impressive levels of accuracy, existing models frequently fail to\nprovide reliable uncertainty estimates -- a pitfall that critically undermines\ntheir applicability in safety-sensitive contexts. Through extensive analysis of\nkey factors such as network capacity, LiDAR representations, rasterization\nresolutions, and 3D data augmentation techniques, we correlate these aspects\ndirectly with the model calibration efficacy. Furthermore, we introduce DeptS,\na novel depth-aware scaling approach aimed at enhancing 3D model calibration.\nExtensive experiments across a wide range of configurations validate the\nsuperiority of our method. We hope this work could serve as a cornerstone for\nfostering reliable 3D scene understanding. Code and benchmark toolkit are\npublicly available.\n", "date": "2024-03-25", "categories": []}
{"id": "2410.03878", "title": "SPARTUN3D: Situated Spatial Understanding of 3D World in Large Language\n  Models", "abstract": "  Integrating the 3D world into large language models (3D-based LLMs) has been\na promising research direction for 3D scene understanding. However, current\n3D-based LLMs fall short in situated understanding due to two key limitations:\n1) existing 3D datasets are constructed from a global perspective of the 3D\nscenes and lack situated context. 2) the architectures of existing 3D-based\nLLMs lack explicit alignment between the spatial representations of 3D scenes\nand natural language, limiting their performance in tasks requiring precise\nspatial reasoning. We address these issues by introducing a scalable situated\n3D dataset, named Spartun3D, that incorporates various situated spatial\nreasoning tasks. Furthermore, we propose Spartun3D-LLM, built on an existing\n3D-based LLM but integrated with a novel situated spatial alignment module,\naiming to enhance the alignment between 3D visual representations and their\ncorresponding textual descriptions. Experimental results demonstrate that both\nour proposed dataset and alignment module significantly enhance the situated\nspatial understanding of 3D-based LLMs.\n", "date": "2024-10-04", "categories": []}
{"id": "2410.11019", "title": "ET-Former: Efficient Triplane Deformable Attention for 3D Semantic Scene\n  Completion From Monocular Camera", "abstract": "  We introduce ET-Former, a novel end-to-end algorithm for semantic scene\ncompletion using a single monocular camera. Our approach generates a semantic\noccupancy map from single RGB observation while simultaneously providing\nuncertainty estimates for semantic predictions. By designing a triplane-based\ndeformable attention mechanism, our approach improves geometric understanding\nof the scene than other SOTA approaches and reduces noise in semantic\npredictions. Additionally, through the use of a Conditional Variational\nAutoEncoder (CVAE), we estimate the uncertainties of these predictions. The\ngenerated semantic and uncertainty maps will help formulate navigation\nstrategies that facilitate safe and permissible decision making in the future.\nEvaluated on the Semantic-KITTI dataset, ET-Former achieves the highest\nIntersection over Union (IoU) and mean IoU (mIoU) scores while maintaining the\nlowest GPU memory usage, surpassing state-of-the-art (SOTA) methods. It\nimproves the SOTA scores of IoU from 44.71 to 51.49 and mIoU from 15.04 to\n16.30 on SeamnticKITTI test, with a notably low training memory consumption of\n10.9 GB. Project page: https://github.com/jingGM/ET-Former.git.\n", "date": "2024-10-14", "categories": []}
{"id": "2503.00726", "title": "Enhancing Monocular 3D Scene Completion with Diffusion Model", "abstract": "  3D scene reconstruction is essential for applications in virtual reality,\nrobotics, and autonomous driving, enabling machines to understand and interact\nwith complex environments. Traditional 3D Gaussian Splatting techniques rely on\nimages captured from multiple viewpoints to achieve optimal performance, but\nthis dependence limits their use in scenarios where only a single image is\navailable. In this work, we introduce FlashDreamer, a novel approach for\nreconstructing a complete 3D scene from a single image, significantly reducing\nthe need for multi-view inputs. Our approach leverages a pre-trained\nvision-language model to generate descriptive prompts for the scene, guiding a\ndiffusion model to produce images from various perspectives, which are then\nfused to form a cohesive 3D reconstruction. Extensive experiments show that our\nmethod effectively and robustly expands single-image inputs into a\ncomprehensive 3D scene, extending monocular 3D reconstruction capabilities\nwithout further training. Our code is available\nhttps://github.com/CharlieSong1999/FlashDreamer/tree/main.\n", "date": "2025-03-02", "categories": []}
{"id": "2503.01646", "title": "OpenGS-SLAM: Open-Set Dense Semantic SLAM with 3D Gaussian Splatting for\n  Object-Level Scene Understanding", "abstract": "  Recent advancements in 3D Gaussian Splatting have significantly improved the\nefficiency and quality of dense semantic SLAM. However, previous methods are\ngenerally constrained by limited-category pre-trained classifiers and implicit\nsemantic representation, which hinder their performance in open-set scenarios\nand restrict 3D object-level scene understanding. To address these issues, we\npropose OpenGS-SLAM, an innovative framework that utilizes 3D Gaussian\nrepresentation to perform dense semantic SLAM in open-set environments. Our\nsystem integrates explicit semantic labels derived from 2D foundational models\ninto the 3D Gaussian framework, facilitating robust 3D object-level scene\nunderstanding. We introduce Gaussian Voting Splatting to enable fast 2D label\nmap rendering and scene updating. Additionally, we propose a Confidence-based\n2D Label Consensus method to ensure consistent labeling across multiple views.\nFurthermore, we employ a Segmentation Counter Pruning strategy to improve the\naccuracy of semantic scene representation. Extensive experiments on both\nsynthetic and real-world datasets demonstrate the effectiveness of our method\nin scene understanding, tracking, and mapping, achieving 10 times faster\nsemantic rendering and 2 times lower storage costs compared to existing\nmethods. Project page: https://young-bit.github.io/opengs-github.github.io/.\n", "date": "2025-03-03", "categories": []}
{"id": "2406.04138", "title": "The 3D-PC: a benchmark for visual perspective taking in humans and\n  machines", "abstract": "  Visual perspective taking (VPT) is the ability to perceive and reason about\nthe perspectives of others. It is an essential feature of human intelligence,\nwhich develops over the first decade of life and requires an ability to process\nthe 3D structure of visual scenes. A growing number of reports have indicated\nthat deep neural networks (DNNs) become capable of analyzing 3D scenes after\ntraining on large image datasets. We investigated if this emergent ability for\n3D analysis in DNNs is sufficient for VPT with the 3D perception challenge\n(3D-PC): a novel benchmark for 3D perception in humans and DNNs. The 3D-PC is\ncomprised of three 3D-analysis tasks posed within natural scene images: 1. a\nsimple test of object depth order, 2. a basic VPT task (VPT-basic), and 3.\nanother version of VPT (VPT-Strategy) designed to limit the effectiveness of\n\"shortcut\" visual strategies. We tested human participants (N=33) and linearly\nprobed or text-prompted over 300 DNNs on the challenge and found that nearly\nall of the DNNs approached or exceeded human accuracy in analyzing object depth\norder. Surprisingly, DNN accuracy on this task correlated with their object\nrecognition performance. In contrast, there was an extraordinary gap between\nDNNs and humans on VPT-basic. Humans were nearly perfect, whereas most DNNs\nwere near chance. Fine-tuning DNNs on VPT-basic brought them close to human\nperformance, but they, unlike humans, dropped back to chance when tested on\nVPT-Strategy. Our challenge demonstrates that the training routines and\narchitectures of today's DNNs are well-suited for learning basic 3D properties\nof scenes and objects but are ill-suited for reasoning about these properties\nas humans do. We release our 3D-PC datasets and code to help bridge this gap in\n3D perception between humans and machines.\n", "date": "2024-06-06", "categories": []}
{"id": "2412.08774", "title": "ProtoOcc: Accurate, Efficient 3D Occupancy Prediction Using Dual Branch\n  Encoder-Prototype Query Decoder", "abstract": "  In this paper, we introduce ProtoOcc, a novel 3D occupancy prediction model\ndesigned to predict the occupancy states and semantic classes of 3D voxels\nthrough a deep semantic understanding of scenes. ProtoOcc consists of two main\ncomponents: the Dual Branch Encoder (DBE) and the Prototype Query Decoder\n(PQD). The DBE produces a new 3D voxel representation by combining 3D voxel and\nBEV representations across multiple scales through a dual branch structure.\nThis design enhances both performance and computational efficiency by providing\na large receptive field for the BEV representation while maintaining a smaller\nreceptive field for the voxel representation. The PQD introduces Prototype\nQueries to accelerate the decoding process. Scene-Adaptive Prototypes are\nderived from the 3D voxel features of input sample, while Scene-Agnostic\nPrototypes are computed by applying Scene-Adaptive Prototypes to an Exponential\nMoving Average during the training phase. By using these prototype-based\nqueries for decoding, we can directly predict 3D occupancy in a single step,\neliminating the need for iterative Transformer decoding. Additionally, we\npropose the Robust Prototype Learning, which injects noise into prototype\ngeneration process and trains the model to denoise during the training phase.\nProtoOcc achieves state-of-the-art performance with 45.02% mIoU on the\nOcc3D-nuScenes benchmark. For single-frame method, it reaches 39.56% mIoU with\nan inference speed of 12.83 FPS on an NVIDIA RTX 3090. Our code can be found at\nhttps://github.com/SPA-junghokim/ProtoOcc.\n", "date": "2024-12-11", "categories": []}
{"id": "2403.16539", "title": "Data-Efficient 3D Visual Grounding via Order-Aware Referring", "abstract": "  3D visual grounding aims to identify the target object within a 3D point\ncloud scene referred to by a natural language description. Previous works\nusually require significant data relating to point color and their descriptions\nto exploit the corresponding complicated verbo-visual relations. In our work,\nwe introduce Vigor, a novel Data-Efficient 3D Visual Grounding framework via\nOrder-aware Referring. Vigor leverages LLM to produce a desirable referential\norder from the input description for 3D visual grounding. With the proposed\nstacked object-referring blocks, the predicted anchor objects in the above\norder allow one to locate the target object progressively without supervision\non the identities of anchor objects or exact relations between anchor/target\nobjects. In addition, we present an order-aware warm-up training strategy,\nwhich augments referential orders for pre-training the visual grounding\nframework. This allows us to better capture the complex verbo-visual relations\nand benefit the desirable data-efficient learning scheme. Experimental results\non the NR3D and ScanRefer datasets demonstrate our superiority in low-resource\nscenarios. In particular, Vigor surpasses current state-of-the-art frameworks\nby 9.3% and 7.6% grounding accuracy under 1% data and 10% data settings on the\nNR3D dataset, respectively. Our code is publicly available at\nhttps://github.com/tony10101105/Vigor.\n", "date": "2024-03-25", "categories": []}
{"id": "2412.05596", "title": "TB-HSU: Hierarchical 3D Scene Understanding with Contextual Affordances", "abstract": "  The concept of function and affordance is a critical aspect of 3D scene\nunderstanding and supports task-oriented objectives. In this work, we develop a\nmodel that learns to structure and vary functional affordance across a 3D\nhierarchical scene graph representing the spatial organization of a scene. The\nvarying functional affordance is designed to integrate with the varying spatial\ncontext of the graph. More specifically, we develop an algorithm that learns to\nconstruct a 3D hierarchical scene graph (3DHSG) that captures the spatial\norganization of the scene. Starting from segmented object point clouds and\nobject semantic labels, we develop a 3DHSG with a top node that identifies the\nroom label, child nodes that define local spatial regions inside the room with\nregion-specific affordances, and grand-child nodes indicating object locations\nand object-specific affordances. To support this work, we create a custom 3DHSG\ndataset that provides ground truth data for local spatial regions with\nregion-specific affordances and also object-specific affordances for each\nobject. We employ a transformer-based model to learn the 3DHSG. We use a\nmulti-task learning framework that learns both room classification and learns\nto define spatial regions within the room with region-specific affordances. Our\nwork improves on the performance of state-of-the-art baseline models and shows\none approach for applying transformer models to 3D scene understanding and the\ngeneration of 3DHSGs that capture the spatial organization of a room. The code\nand dataset are publicly available.\n", "date": "2024-12-07", "categories": []}
{"id": "2502.16303", "title": "Pointmap Association and Piecewise-Plane Constraint for Consistent and\n  Compact 3D Gaussian Segmentation Field", "abstract": "  Achieving a consistent and compact 3D segmentation field is crucial for\nmaintaining semantic coherence across views and accurately representing scene\nstructures. Previous 3D scene segmentation methods rely on video segmentation\nmodels to address inconsistencies across views, but the absence of spatial\ninformation often leads to object misassociation when object temporarily\ndisappear and reappear. Furthermore, in the process of 3D scene reconstruction,\nsegmentation and optimization are often treated as separate tasks. As a result,\noptimization typically lacks awareness of semantic category information, which\ncan result in floaters with ambiguous segmentation. To address these\nchallenges, we introduce CCGS, a method designed to achieve both view\nconsistent 2D segmentation and a compact 3D Gaussian segmentation field. CCGS\nincorporates pointmap association and a piecewise-plane constraint. First, we\nestablish pixel correspondence between adjacent images by minimizing the\nEuclidean distance between their pointmaps. We then redefine object mask\noverlap accordingly. The Hungarian algorithm is employed to optimize mask\nassociation by minimizing the total matching cost, while allowing for partial\nmatches. To further enhance compactness, the piecewise-plane constraint\nrestricts point displacement within local planes during optimization, thereby\npreserving structural integrity. Experimental results on ScanNet and Replica\ndatasets demonstrate that CCGS outperforms existing methods in both 2D panoptic\nsegmentation and 3D Gaussian segmentation.\n", "date": "2025-02-22", "categories": []}
{"id": "2502.16652", "title": "Dr. Splat: Directly Referring 3D Gaussian Splatting via Direct Language\n  Embedding Registration", "abstract": "  We introduce Dr. Splat, a novel approach for open-vocabulary 3D scene\nunderstanding leveraging 3D Gaussian Splatting. Unlike existing\nlanguage-embedded 3DGS methods, which rely on a rendering process, our method\ndirectly associates language-aligned CLIP embeddings with 3D Gaussians for\nholistic 3D scene understanding. The key of our method is a language feature\nregistration technique where CLIP embeddings are assigned to the dominant\nGaussians intersected by each pixel-ray. Moreover, we integrate Product\nQuantization (PQ) trained on general large-scale image data to compactly\nrepresent embeddings without per-scene optimization. Experiments demonstrate\nthat our approach significantly outperforms existing approaches in 3D\nperception benchmarks, such as open-vocabulary 3D semantic segmentation, 3D\nobject localization, and 3D object selection tasks. For video results, please\nvisit : https://drsplat.github.io/\n", "date": "2025-02-23", "categories": []}
{"id": "2408.07416", "title": "Rethinking Open-Vocabulary Segmentation of Radiance Fields in 3D Space", "abstract": "  Understanding the 3D semantics of a scene is a fundamental problem for\nvarious scenarios such as embodied agents. While NeRFs and 3DGS excel at\nnovel-view synthesis, previous methods for understanding their semantics have\nbeen limited to incomplete 3D understanding: their segmentation results are\nrendered as 2D masks that do not represent the entire 3D space. To address this\nlimitation, we redefine the problem to segment the 3D volume and propose the\nfollowing methods for better 3D understanding. We directly supervise the 3D\npoints to train the language embedding field, unlike previous methods that\nanchor supervision at 2D pixels. We transfer the learned language field to\n3DGS, achieving the first real-time rendering speed without sacrificing\ntraining time or accuracy. Lastly, we introduce a 3D querying and evaluation\nprotocol for assessing the reconstructed geometry and semantics together. Code,\ncheckpoints, and annotations are available at the project page.\n", "date": "2024-08-14", "categories": []}
{"id": "2312.00663", "title": "Generalized Robot 3D Vision-Language Model with Fast Rendering and\n  Pre-Training Vision-Language Alignment", "abstract": "  Deep neural network models have achieved remarkable progress in 3D scene\nunderstanding while trained in the closed-set setting and with full labels.\nHowever, the major bottleneck is that these models do not have the capacity to\nrecognize any unseen novel classes beyond the training categories in diverse\nreal-world applications. Therefore, we are in urgent need of a framework that\ncan simultaneously be applicable to both 3D point cloud segmentation and\ndetection, particularly in the circumstances where the labels are rather\nscarce. This work presents a generalized and straightforward framework for\ndealing with 3D scene understanding when the labeled scenes are quite limited.\nTo extract knowledge for novel categories from the pre-trained vision-language\nmodels, we propose a hierarchical feature-aligned pre-training and knowledge\ndistillation strategy to extract and distill meaningful information from\nlarge-scale vision-language models, which helps benefit the open-vocabulary\nscene understanding tasks. To encourage latent instance discrimination and to\nguarantee efficiency, we propose the unsupervised region-level semantic\ncontrastive learning scheme for point clouds, using confident predictions of\nthe neural network to discriminate the intermediate feature embeddings at\nmultiple stages. In the limited reconstruction case, our proposed approach,\ntermed WS3D++, ranks 1st on the large-scale ScanNet benchmark on both the task\nof semantic segmentation and instance segmentation. Extensive experiments with\nboth indoor and outdoor scenes demonstrated the effectiveness of our approach\nin both data-efficient learning and open-world few-shot learning. The code is\nmade publicly available at:\nhttps://drive.google.com/drive/folders/1M58V-PtR8DBEwD296zJkNg_m2qq-MTAP?usp=sharing.\n", "date": "2023-12-01", "categories": []}
{"id": "2405.13989", "title": "TS40K: a 3D Point Cloud Dataset of Rural Terrain and Electrical\n  Transmission System", "abstract": "  Research on supervised learning algorithms in 3D scene understanding has\nrisen in prominence and witness great increases in performance across several\ndatasets. The leading force of this research is the problem of autonomous\ndriving followed by indoor scene segmentation. However, openly available 3D\ndata on these tasks mainly focuses on urban scenarios. In this paper, we\npropose TS40K, a 3D point cloud dataset that encompasses more than 40,000 Km on\nelectrical transmission systems situated in European rural terrain. This is not\nonly a novel problem for the research community that can aid in the high-risk\nmission of power-grid inspection, but it also offers 3D point clouds with\ndistinct characteristics from those in self-driving and indoor 3D data, such as\nhigh point-density and no occlusion. In our dataset, each 3D point is labeled\nwith 1 out of 22 annotated classes. We evaluate the performance of\nstate-of-the-art methods on our dataset concerning 3D semantic segmentation and\n3D object detection. Finally, we provide a comprehensive analysis of the\nresults along with key challenges such as using labels that were not originally\nintended for learning tasks.\n", "date": "2024-05-22", "categories": []}
{"id": "2402.07376", "title": "Unsupervised Discovery of Object-Centric Neural Fields", "abstract": "  We study inferring 3D object-centric scene representations from a single\nimage. While recent methods have shown potential in unsupervised 3D object\ndiscovery from simple synthetic images, they fail to generalize to real-world\nscenes with visually rich and diverse objects. This limitation stems from their\nobject representations, which entangle objects' intrinsic attributes like shape\nand appearance with extrinsic, viewer-centric properties such as their 3D\nlocation. To address this bottleneck, we propose Unsupervised discovery of\nObject-Centric neural Fields (uOCF). uOCF focuses on learning the intrinsics of\nobjects and models the extrinsics separately. Our approach significantly\nimproves systematic generalization, thus enabling unsupervised learning of\nhigh-fidelity object-centric scene representations from sparse real-world\nimages. To evaluate our approach, we collect three new datasets, including two\nreal kitchen environments. Extensive experiments show that uOCF enables\nunsupervised discovery of visually rich objects from a single real image,\nallowing applications such as 3D object segmentation and scene manipulation.\nNotably, uOCF demonstrates zero-shot generalization to unseen objects from a\nsingle real image. Project page: https://red-fairy.github.io/uOCF/\n", "date": "2024-02-12", "categories": []}
{"id": "2409.19987", "title": "OccRWKV: Rethinking Efficient 3D Semantic Occupancy Prediction with\n  Linear Complexity", "abstract": "  3D semantic occupancy prediction networks have demonstrated remarkable\ncapabilities in reconstructing the geometric and semantic structure of 3D\nscenes, providing crucial information for robot navigation and autonomous\ndriving systems. However, due to their large overhead from dense network\nstructure designs, existing networks face challenges balancing accuracy and\nlatency. In this paper, we introduce OccRWKV, an efficient semantic occupancy\nnetwork inspired by Receptance Weighted Key Value (RWKV). OccRWKV separates\nsemantics, occupancy prediction, and feature fusion into distinct branches,\neach incorporating Sem-RWKV and Geo-RWKV blocks. These blocks are designed to\ncapture long-range dependencies, enabling the network to learn domain-specific\nrepresentation (i.e., semantics and geometry), which enhances prediction\naccuracy. Leveraging the sparse nature of real-world 3D occupancy, we reduce\ncomputational overhead by projecting features into the bird's-eye view (BEV)\nspace and propose a BEV-RWKV block for efficient feature enhancement and\nfusion. This enables real-time inference at 22.2 FPS without compromising\nperformance. Experiments demonstrate that OccRWKV outperforms the\nstate-of-the-art methods on the SemanticKITTI dataset, achieving a mIoU of 25.1\nwhile being 20 times faster than the best baseline, Co-Occ, making it suitable\nfor real-time deployment on robots to enhance autonomous navigation efficiency.\nCode and video are available on our project page:\nhttps://jmwang0117.github.io/OccRWKV/.\n", "date": "2024-09-30", "categories": []}
{"id": "2407.20908", "title": "Dynamic Scene Understanding through Object-Centric Voxelization and\n  Neural Rendering", "abstract": "  Learning object-centric representations from unsupervised videos is\nchallenging. Unlike most previous approaches that focus on decomposing 2D\nimages, we present a 3D generative model named DynaVol-S for dynamic scenes\nthat enables object-centric learning within a differentiable volume rendering\nframework. The key idea is to perform object-centric voxelization to capture\nthe 3D nature of the scene, which infers per-object occupancy probabilities at\nindividual spatial locations. These voxel features evolve through a\ncanonical-space deformation function and are optimized in an inverse rendering\npipeline with a compositional NeRF. Additionally, our approach integrates 2D\nsemantic features to create 3D semantic grids, representing the scene through\nmultiple disentangled voxel grids. DynaVol-S significantly outperforms existing\nmodels in both novel view synthesis and unsupervised decomposition tasks for\ndynamic scenes. By jointly considering geometric structures and semantic\nfeatures, it effectively addresses challenging real-world scenarios involving\ncomplex object interactions. Furthermore, once trained, the explicitly\nmeaningful voxel features enable additional capabilities that 2D scene\ndecomposition methods cannot achieve, such as novel scene generation through\nediting geometric shapes or manipulating the motion trajectories of objects.\n", "date": "2024-07-30", "categories": []}
{"id": "2405.12821", "title": "Talk2Radar: Bridging Natural Language with 4D mmWave Radar for 3D\n  Referring Expression Comprehension", "abstract": "  Embodied perception is essential for intelligent vehicles and robots in\ninteractive environmental understanding. However, these advancements primarily\nfocus on vision, with limited attention given to using 3D modeling sensors,\nrestricting a comprehensive understanding of objects in response to prompts\ncontaining qualitative and quantitative queries. Recently, as a promising\nautomotive sensor with affordable cost, 4D millimeter-wave radars provide\ndenser point clouds than conventional radars and perceive both semantic and\nphysical characteristics of objects, thereby enhancing the reliability of\nperception systems. To foster the development of natural language-driven\ncontext understanding in radar scenes for 3D visual grounding, we construct the\nfirst dataset, Talk2Radar, which bridges these two modalities for 3D Referring\nExpression Comprehension (REC). Talk2Radar contains 8,682 referring prompt\nsamples with 20,558 referred objects. Moreover, we propose a novel model,\nT-RadarNet, for 3D REC on point clouds, achieving State-Of-The-Art (SOTA)\nperformance on the Talk2Radar dataset compared to counterparts. Deformable-FPN\nand Gated Graph Fusion are meticulously designed for efficient point cloud\nfeature modeling and cross-modal fusion between radar and text features,\nrespectively. Comprehensive experiments provide deep insights into radar-based\n3D REC. We release our project at https://github.com/GuanRunwei/Talk2Radar.\n", "date": "2024-05-21", "categories": []}
{"id": "2502.03901", "title": "LeAP: Consistent multi-domain 3D labeling using Foundation Models", "abstract": "  Availability of datasets is a strong driver for research on 3D semantic\nunderstanding, and whilst obtaining unlabeled 3D point cloud data is\nstraightforward, manually annotating this data with semantic labels is\ntime-consuming and costly. Recently, Vision Foundation Models (VFMs) enable\nopen-set semantic segmentation on camera images, potentially aiding automatic\nlabeling. However,VFMs for 3D data have been limited to adaptations of 2D\nmodels, which can introduce inconsistencies to 3D labels. This work introduces\nLabel Any Pointcloud (LeAP), leveraging 2D VFMs to automatically label 3D data\nwith any set of classes in any kind of application whilst ensuring label\nconsistency. Using a Bayesian update, point labels are combined into voxels to\nimprove spatio-temporal consistency. A novel 3D Consistency Network (3D-CN)\nexploits 3D information to further improve label quality. Through various\nexperiments, we show that our method can generate high-quality 3D semantic\nlabels across diverse fields without any manual labeling. Further, models\nadapted to new domains using our labels show up to a 34.2 mIoU increase in\nsemantic segmentation tasks.\n", "date": "2025-02-06", "categories": []}
{"id": "2502.04139", "title": "Beyond the Final Layer: Hierarchical Query Fusion Transformer with\n  Agent-Interpolation Initialization for 3D Instance Segmentation", "abstract": "  3D instance segmentation aims to predict a set of object instances in a scene\nand represent them as binary foreground masks with corresponding semantic\nlabels. Currently, transformer-based methods are gaining increasing attention\ndue to their elegant pipelines, reduced manual selection of geometric\nproperties, and superior performance. However, transformer-based methods fail\nto simultaneously maintain strong position and content information during query\ninitialization. Additionally, due to supervision at each decoder layer, there\nexists a phenomenon of object disappearance with the deepening of layers. To\novercome these hurdles, we introduce Beyond the Final Layer: Hierarchical Query\nFusion Transformer with Agent-Interpolation Initialization for 3D Instance\nSegmentation (BFL). Specifically, an Agent-Interpolation Initialization Module\nis designed to generate resilient queries capable of achieving a balance\nbetween foreground coverage and content learning. Additionally, a Hierarchical\nQuery Fusion Decoder is designed to retain low overlap queries, mitigating the\ndecrease in recall with the deepening of layers. Extensive experiments on\nScanNetV2, ScanNet200, ScanNet++ and S3DIS datasets demonstrate the superior\nperformance of BFL.\n", "date": "2025-02-06", "categories": []}
{"id": "2312.00860", "title": "Segment Any 3D Gaussians", "abstract": "  This paper presents SAGA (Segment Any 3D GAussians), a highly efficient 3D\npromptable segmentation method based on 3D Gaussian Splatting (3D-GS). Given 2D\nvisual prompts as input, SAGA can segment the corresponding 3D target\nrepresented by 3D Gaussians within 4 ms. This is achieved by attaching an\nscale-gated affinity feature to each 3D Gaussian to endow it a new property\ntowards multi-granularity segmentation. Specifically, a scale-aware contrastive\ntraining strategy is proposed for the scale-gated affinity feature learning. It\n1) distills the segmentation capability of the Segment Anything Model (SAM)\nfrom 2D masks into the affinity features and 2) employs a soft scale gate\nmechanism to deal with multi-granularity ambiguity in 3D segmentation through\nadjusting the magnitude of each feature channel according to a specified 3D\nphysical scale. Evaluations demonstrate that SAGA achieves real-time\nmulti-granularity segmentation with quality comparable to state-of-the-art\nmethods. As one of the first methods addressing promptable segmentation in\n3D-GS, the simplicity and effectiveness of SAGA pave the way for future\nadvancements in this field. Our code will be released.\n", "date": "2023-12-01", "categories": []}
{"id": "2502.02334", "title": "Event-aided Semantic Scene Completion", "abstract": "  Autonomous driving systems rely on robust 3D scene understanding. Recent\nadvances in Semantic Scene Completion (SSC) for autonomous driving underscore\nthe limitations of RGB-based approaches, which struggle under motion blur, poor\nlighting, and adverse weather. Event cameras, offering high dynamic range and\nlow latency, address these challenges by providing asynchronous data that\ncomplements RGB inputs. We present DSEC-SSC, the first real-world benchmark\nspecifically designed for event-aided SSC, which includes a novel 4D labeling\npipeline for generating dense, visibility-aware labels that adapt dynamically\nto object motion. Our proposed RGB-Event fusion framework, EvSSC, introduces an\nEvent-aided Lifting Module (ELM) that effectively bridges 2D RGB-Event features\nto 3D space, enhancing view transformation and the robustness of 3D volume\nconstruction across SSC models. Extensive experiments on DSEC-SSC and simulated\nSemanticKITTI-E demonstrate that EvSSC is adaptable to both transformer-based\nand LSS-based SSC architectures. Notably, evaluations on SemanticKITTI-C\ndemonstrate that EvSSC achieves consistently improved prediction accuracy\nacross five degradation modes and both In-domain and Out-of-domain settings,\nachieving up to a 52.5% relative improvement in mIoU when the image sensor\npartially fails. Additionally, we quantitatively and qualitatively validate the\nsuperiority of EvSSC under motion blur and extreme weather conditions, where\nautonomous driving is challenged. The established datasets and our codebase\nwill be made publicly at https://github.com/Pandapan01/EvSSC.\n", "date": "2025-02-04", "categories": []}
{"id": "2412.01292", "title": "LSceneLLM: Enhancing Large 3D Scene Understanding Using Adaptive Visual\n  Preferences", "abstract": "  Research on 3D Vision-Language Models (3D-VLMs) is gaining increasing\nattention, which is crucial for developing embodied AI within 3D scenes, such\nas visual navigation and embodied question answering. Due to the high density\nof visual features, especially in large 3D scenes, accurately locating\ntask-relevant visual information is challenging. Existing works attempt to\nsegment all objects and consider their features as scene representations.\nHowever, these task-agnostic object features include much redundant information\nand missing details for the task-relevant area. To tackle these problems, we\npropose LSceneLLM, an adaptive framework that automatically identifies\ntask-relevant areas by leveraging LLM's visual preference for different tasks,\nfollowed by a plug-and-play scene magnifier module to capture fine-grained\ndetails in focused areas. Specifically, a dense token selector examines the\nattention map of LLM to identify visual preferences for the instruction input.\nIt then magnifies fine-grained details of the focusing area. An adaptive\nself-attention module is leveraged to fuse the coarse-grained and selected\nfine-grained visual information. To comprehensively evaluate the large scene\nunderstanding ability of 3D-VLMs, we further introduce a cross-room\nunderstanding benchmark, XR-Scene, which contains a series of large scene\nunderstanding tasks including XR-QA, XR-EmbodiedPlanning, and XR-SceneCaption.\nExperiments show that our method surpasses existing methods on both large scene\nunderstanding and existing scene understanding benchmarks. Plunging our scene\nmagnifier module into the existing 3D-VLMs also brings significant improvement.\n", "date": "2024-12-02", "categories": []}
{"id": "2502.00173", "title": "Lifting by Gaussians: A Simple, Fast and Flexible Method for 3D Instance\n  Segmentation", "abstract": "  We introduce Lifting By Gaussians (LBG), a novel approach for open-world\ninstance segmentation of 3D Gaussian Splatted Radiance Fields (3DGS). Recently,\n3DGS Fields have emerged as a highly efficient and explicit alternative to\nNeural Field-based methods for high-quality Novel View Synthesis. Our 3D\ninstance segmentation method directly lifts 2D segmentation masks from SAM\n(alternately FastSAM, etc.), together with features from CLIP and DINOv2,\ndirectly fusing them onto 3DGS (or similar Gaussian radiance fields such as\n2DGS). Unlike previous approaches, LBG requires no per-scene training, allowing\nit to operate seamlessly on any existing 3DGS reconstruction. Our approach is\nnot only an order of magnitude faster and simpler than existing approaches; it\nis also highly modular, enabling 3D semantic segmentation of existing 3DGS\nfields without requiring a specific parametrization of the 3D Gaussians.\nFurthermore, our technique achieves superior semantic segmentation for 2D\nsemantic novel view synthesis and 3D asset extraction results while maintaining\nflexibility and efficiency. We further introduce a novel approach to evaluate\nindividually segmented 3D assets from 3D radiance field segmentation methods.\n", "date": "2025-01-31", "categories": []}
{"id": "2501.19084", "title": "Laser: Efficient Language-Guided Segmentation in Neural Radiance Fields", "abstract": "  In this work, we propose a method that leverages CLIP feature distillation,\nachieving efficient 3D segmentation through language guidance. Unlike previous\nmethods that rely on multi-scale CLIP features and are limited by processing\nspeed and storage requirements, our approach aims to streamline the workflow by\ndirectly and effectively distilling dense CLIP features, thereby achieving\nprecise segmentation of 3D scenes using text. To achieve this, we introduce an\nadapter module and mitigate the noise issue in the dense CLIP feature\ndistillation process through a self-cross-training strategy. Moreover, to\nenhance the accuracy of segmentation edges, this work presents a low-rank\ntransient query attention mechanism. To ensure the consistency of segmentation\nfor similar colors under different viewpoints, we convert the segmentation task\ninto a classification task through label volume, which significantly improves\nthe consistency of segmentation in color-similar areas. We also propose a\nsimplified text augmentation strategy to alleviate the issue of ambiguity in\nthe correspondence between CLIP features and text. Extensive experimental\nresults show that our method surpasses current state-of-the-art technologies in\nboth training speed and performance. Our code is available on:\nhttps://github.com/xingy038/Laser.git.\n", "date": "2025-01-31", "categories": []}
{"id": "2306.05978", "title": "3D objects and scenes classification, recognition, segmentation, and\n  reconstruction using 3D point cloud data: A review", "abstract": "  Three-dimensional (3D) point cloud analysis has become one of the attractive\nsubjects in realistic imaging and machine visions due to its simplicity,\nflexibility and powerful capacity of visualization. Actually, the\nrepresentation of scenes and buildings using 3D shapes and formats leveraged\nmany applications among which automatic driving, scenes and objects\nreconstruction, etc. Nevertheless, working with this emerging type of data has\nbeen a challenging task for objects representation, scenes recognition,\nsegmentation, and reconstruction. In this regard, a significant effort has\nrecently been devoted to developing novel strategies, using different\ntechniques such as deep learning models. To that end, we present in this paper\na comprehensive review of existing tasks on 3D point cloud: a well-defined\ntaxonomy of existing techniques is performed based on the nature of the adopted\nalgorithms, application scenarios, and main objectives. Various tasks performed\non 3D point could data are investigated, including objects and scenes\ndetection, recognition, segmentation and reconstruction. In addition, we\nintroduce a list of used datasets, we discuss respective evaluation metrics and\nwe compare the performance of existing solutions to better inform the\nstate-of-the-art and identify their limitations and strengths. Lastly, we\nelaborate on current challenges facing the subject of technology and future\ntrends attracting considerable interest, which could be a starting point for\nupcoming research studies\n", "date": "2023-06-09", "categories": []}
{"id": "2409.18431", "title": "Search3D: Hierarchical Open-Vocabulary 3D Segmentation", "abstract": "  Open-vocabulary 3D segmentation enables exploration of 3D spaces using\nfree-form text descriptions. Existing methods for open-vocabulary 3D instance\nsegmentation primarily focus on identifying object-level instances but struggle\nwith finer-grained scene entities such as object parts, or regions described by\ngeneric attributes. In this work, we introduce Search3D, an approach to\nconstruct hierarchical open-vocabulary 3D scene representations, enabling 3D\nsearch at multiple levels of granularity: fine-grained object parts, entire\nobjects, or regions described by attributes like materials. Unlike prior\nmethods, Search3D shifts towards a more flexible open-vocabulary 3D search\nparadigm, moving beyond explicit object-centric queries. For systematic\nevaluation, we further contribute a scene-scale open-vocabulary 3D part\nsegmentation benchmark based on MultiScan, along with a set of open-vocabulary\nfine-grained part annotations on ScanNet++. Search3D outperforms baselines in\nscene-scale open-vocabulary 3D part segmentation, while maintaining strong\nperformance in segmenting 3D objects and materials. Our project page is\nhttp://search3d-segmentation.github.io.\n", "date": "2024-09-27", "categories": []}
{"id": "2501.12387", "title": "Continuous 3D Perception Model with Persistent State", "abstract": "  We present a unified framework capable of solving a broad range of 3D tasks.\nOur approach features a stateful recurrent model that continuously updates its\nstate representation with each new observation. Given a stream of images, this\nevolving state can be used to generate metric-scale pointmaps (per-pixel 3D\npoints) for each new input in an online fashion. These pointmaps reside within\na common coordinate system, and can be accumulated into a coherent, dense scene\nreconstruction that updates as new images arrive. Our model, called CUT3R\n(Continuous Updating Transformer for 3D Reconstruction), captures rich priors\nof real-world scenes: not only can it predict accurate pointmaps from image\nobservations, but it can also infer unseen regions of the scene by probing at\nvirtual, unobserved views. Our method is simple yet highly flexible, naturally\naccepting varying lengths of images that may be either video streams or\nunordered photo collections, containing both static and dynamic content. We\nevaluate our method on various 3D/4D tasks and demonstrate competitive or\nstate-of-the-art performance in each. Project Page: https://cut3r.github.io/\n", "date": "2025-01-21", "categories": []}
{"id": "2411.02104", "title": "Deep Learning on 3D Semantic Segmentation: A Detailed Review", "abstract": "  In this paper an exhaustive review and comprehensive analysis of recent and\nformer deep learning methods in 3D Semantic Segmentation (3DSS) is presented.\nIn the related literature, the taxonomy scheme used for the classification of\nthe 3DSS deep learning methods is ambiguous. Based on the taxonomy schemes of 9\nexisting review papers, a new taxonomy scheme of the 3DSS deep learning methods\nis proposed, aiming to standardize it and improve the comparability and clarity\nacross related studies. Furthermore, an extensive overview of the available\n3DSS indoor and outdoor datasets is provided along with their links. The core\npart of the review is the detailed presentation of recent and former 3DSS deep\nlearning methods and their classification using the proposed taxonomy scheme\nalong with their GitHub repositories. Additionally, a brief but informative\nanalysis of the evaluation metrics and loss functions used in 3DSS is included.\nFinally, a fruitful discussion of the examined 3DSS methods and datasets, is\npresented to foster new research directions and applications in the field of\n3DSS. Supplementary, to this review a GitHub repository is provided\n(https://github.com/thobet/Deep-Learning-on-3D-Semantic-Segmentation-a-\nDetailed-Review) including a quick classification of over 400 3DSS methods,\nusing the proposed taxonomy scheme.\n", "date": "2024-11-04", "categories": []}
{"id": "2501.09428", "title": "AugRefer: Advancing 3D Visual Grounding via Cross-Modal Augmentation and\n  Spatial Relation-based Referring", "abstract": "  3D visual grounding (3DVG), which aims to correlate a natural language\ndescription with the target object within a 3D scene, is a significant yet\nchallenging task. Despite recent advancements in this domain, existing\napproaches commonly encounter a shortage: a limited amount and diversity of\ntext3D pairs available for training. Moreover, they fall short in effectively\nleveraging different contextual clues (e.g., rich spatial relations within the\n3D visual space) for grounding. To address these limitations, we propose\nAugRefer, a novel approach for advancing 3D visual grounding. AugRefer\nintroduces cross-modal augmentation designed to extensively generate diverse\ntext-3D pairs by placing objects into 3D scenes and creating accurate and\nsemantically rich descriptions using foundation models. Notably, the resulting\npairs can be utilized by any existing 3DVG methods for enriching their training\ndata. Additionally, AugRefer presents a language-spatial adaptive decoder that\neffectively adapts the potential referring objects based on the language\ndescription and various 3D spatial relations. Extensive experiments on three\nbenchmark datasets clearly validate the effectiveness of AugRefer.\n", "date": "2025-01-16", "categories": []}
{"id": "2501.07819", "title": "3UR-LLM: An End-to-End Multimodal Large Language Model for 3D Scene\n  Understanding", "abstract": "  Multi-modal Large Language Models (MLLMs) exhibit impressive capabilities in\n2D tasks, yet encounter challenges in discerning the spatial positions,\ninterrelations, and causal logic in scenes when transitioning from 2D to 3D\nrepresentations. We find that the limitations mainly lie in: i) the high\nannotation cost restricting the scale-up of volumes of 3D scene data, and ii)\nthe lack of a straightforward and effective way to perceive 3D information\nwhich results in prolonged training durations and complicates the streamlined\nframework. To this end, we develop pipeline based on open-source 2D MLLMs and\nLLMs to generate high-quality 3D-text pairs and construct 3DS-160K , to enhance\nthe pre-training process. Leveraging this high-quality pre-training data, we\nintroduce the 3UR-LLM model, an end-to-end 3D MLLM designed for precise\ninterpretation of 3D scenes, showcasing exceptional capability in navigating\nthe complexities of the physical world. 3UR-LLM directly receives 3D point\ncloud as input and project 3D features fused with text instructions into a\nmanageable set of tokens. Considering the computation burden derived from these\nhybrid tokens, we design a 3D compressor module to cohesively compress the 3D\nspatial cues and textual narrative. 3UR-LLM achieves promising performance with\nrespect to the previous SOTAs, for instance, 3UR-LLM exceeds its counterparts\nby 7.1\\% CIDEr on ScanQA, while utilizing fewer training resources. The code\nand model weights for 3UR-LLM and the 3DS-160K benchmark are available at\n3UR-LLM.\n", "date": "2025-01-14", "categories": []}
{"id": "2501.07260", "title": "Skip Mamba Diffusion for Monocular 3D Semantic Scene Completion", "abstract": "  3D semantic scene completion is critical for multiple downstream tasks in\nautonomous systems. It estimates missing geometric and semantic information in\nthe acquired scene data. Due to the challenging real-world conditions, this\ntask usually demands complex models that process multi-modal data to achieve\nacceptable performance. We propose a unique neural model, leveraging advances\nfrom the state space and diffusion generative modeling to achieve remarkable 3D\nsemantic scene completion performance with monocular image input. Our technique\nprocesses the data in the conditioned latent space of a variational autoencoder\nwhere diffusion modeling is carried out with an innovative state space\ntechnique. A key component of our neural network is the proposed Skimba (Skip\nMamba) denoiser, which is adept at efficiently processing long-sequence data.\nThe Skimba diffusion model is integral to our 3D scene completion network,\nincorporating a triple Mamba structure, dimensional decomposition residuals and\nvarying dilations along three directions. We also adopt a variant of this\nnetwork for the subsequent semantic segmentation stage of our method. Extensive\nevaluation on the standard SemanticKITTI and SSCBench-KITTI360 datasets show\nthat our approach not only outperforms other monocular techniques by a large\nmargin, it also achieves competitive performance against stereo methods. The\ncode is available at https://github.com/xrkong/skimba\n", "date": "2025-01-13", "categories": []}
{"id": "2412.11210", "title": "ViPOcc: Leveraging Visual Priors from Vision Foundation Models for\n  Single-View 3D Occupancy Prediction", "abstract": "  Inferring the 3D structure of a scene from a single image is an ill-posed and\nchallenging problem in the field of vision-centric autonomous driving. Existing\nmethods usually employ neural radiance fields to produce voxelized 3D\noccupancy, lacking instance-level semantic reasoning and temporal photometric\nconsistency. In this paper, we propose ViPOcc, which leverages the visual\npriors from vision foundation models (VFMs) for fine-grained 3D occupancy\nprediction. Unlike previous works that solely employ volume rendering for RGB\nand depth image reconstruction, we introduce a metric depth estimation branch,\nin which an inverse depth alignment module is proposed to bridge the domain gap\nin depth distribution between VFM predictions and the ground truth. The\nrecovered metric depth is then utilized in temporal photometric alignment and\nspatial geometric alignment to ensure accurate and consistent 3D occupancy\nprediction. Additionally, we also propose a semantic-guided non-overlapping\nGaussian mixture sampler for efficient, instance-aware ray sampling, which\naddresses the redundant and imbalanced sampling issue that still exists in\nprevious state-of-the-art methods. Extensive experiments demonstrate the\nsuperior performance of ViPOcc in both 3D occupancy prediction and depth\nestimation tasks on the KITTI-360 and KITTI Raw datasets. Our code is available\nat: \\url{https://mias.group/ViPOcc}.\n", "date": "2024-12-15", "categories": []}
{"id": "2408.11559", "title": "Semi-supervised 3D Semantic Scene Completion with 2D Vision Foundation\n  Model Guidance", "abstract": "  Accurate prediction of 3D semantic occupancy from 2D visual images is vital\nin enabling autonomous agents to comprehend their surroundings for planning and\nnavigation. State-of-the-art methods typically employ fully supervised\napproaches, necessitating a huge labeled dataset acquired through expensive\nLiDAR sensors and meticulous voxel-wise labeling by human annotators. The\nresource-intensive nature of this annotating process significantly hampers the\napplication and scalability of these methods. We introduce a novel\nsemi-supervised framework to alleviate the dependency on densely annotated\ndata. Our approach leverages 2D foundation models to generate essential 3D\nscene geometric and semantic cues, facilitating a more efficient training\nprocess. Our framework exhibits notable properties: (1) Generalizability,\napplicable to various 3D semantic scene completion approaches, including 2D-3D\nlifting and 3D-2D transformer methods. (2) Effectiveness, as demonstrated\nthrough experiments on SemanticKITTI and NYUv2, wherein our method achieves up\nto 85% of the fully-supervised performance using only 10% labeled data. This\napproach not only reduces the cost and labor associated with data annotation\nbut also demonstrates the potential for broader adoption in camera-based\nsystems for 3D semantic occupancy prediction.\n", "date": "2024-08-21", "categories": []}
{"id": "2411.10616", "title": "Voxel-Aggregated Feature Synthesis: Efficient Dense Mapping for\n  Simulated 3D Reasoning", "abstract": "  We address the issue of the exploding computational requirements of recent\nState-of-the-art (SOTA) open set multimodel 3D mapping (dense 3D mapping)\nalgorithms and present Voxel-Aggregated Feature Synthesis (VAFS), a novel\napproach to dense 3D mapping in simulation. Dense 3D mapping involves\nsegmenting and embedding sequential RGBD frames which are then fused into 3D.\nThis leads to redundant computation as the differences between frames are small\nbut all are individually segmented and embedded. This makes dense 3D mapping\nimpractical for research involving embodied agents in which the environment,\nand thus the mapping, must be modified with regularity. VAFS drastically\nreduces this computation by using the segmented point cloud computed by a\nsimulator's physics engine and synthesizing views of each region. This reduces\nthe number of features to embed from the number of captured RGBD frames to the\nnumber of objects in the scene, effectively allowing a \"ground truth\" semantic\nmap to be computed an order of magnitude faster than traditional methods. We\ntest the resulting representation by assessing the IoU scores of semantic\nqueries for different objects in the simulated scene, and find that VAFS\nexceeds the accuracy and speed of prior dense 3D mapping techniques.\n", "date": "2024-11-15", "categories": []}
{"id": "2311.11371", "title": "SOccDPT: Semi-Supervised 3D Semantic Occupancy from Dense Prediction\n  Transformers trained under memory constraints", "abstract": "  We present SOccDPT, a memory-efficient approach for 3D semantic occupancy\nprediction from monocular image input using dense prediction transformers. To\naddress the limitations of existing methods trained on structured traffic\ndatasets, we train our model on unstructured datasets including the Indian\nDriving Dataset and Bengaluru Driving Dataset. Our semi-supervised training\npipeline allows SOccDPT to learn from datasets with limited labels by reducing\nthe requirement for manual labelling by substituting it with pseudo-ground\ntruth labels to produce our Bengaluru Semantic Occupancy Dataset. This broader\ntraining enhances our model's ability to handle unstructured traffic scenarios\neffectively. To overcome memory limitations during training, we introduce\npatch-wise training where we select a subset of parameters to train each epoch,\nreducing memory usage during auto-grad graph construction. In the context of\nunstructured traffic and memory-constrained training and inference, SOccDPT\noutperforms existing disparity estimation approaches as shown by the RMSE score\nof 9.1473, achieves a semantic segmentation IoU score of 46.02% and operates at\na competitive frequency of 69.47 Hz. We make our code and semantic occupancy\ndataset public.\n", "date": "2023-11-19", "categories": []}
{"id": "2408.03516", "title": "Query3D: LLM-Powered Open-Vocabulary Scene Segmentation with Language\n  Embedded 3D Gaussian", "abstract": "  This paper introduces a novel method for open-vocabulary 3D scene querying in\nautonomous driving by combining Language Embedded 3D Gaussians with Large\nLanguage Models (LLMs). We propose utilizing LLMs to generate both contextually\ncanonical phrases and helping positive words for enhanced segmentation and\nscene interpretation. Our method leverages GPT-3.5 Turbo as an expert model to\ncreate a high-quality text dataset, which we then use to fine-tune smaller,\nmore efficient LLMs for on-device deployment. Our comprehensive evaluation on\nthe WayveScenes101 dataset demonstrates that LLM-guided segmentation\nsignificantly outperforms traditional approaches based on predefined canonical\nphrases. Notably, our fine-tuned smaller models achieve performance comparable\nto larger expert models while maintaining faster inference times. Through\nablation studies, we discover that the effectiveness of helping positive words\ncorrelates with model scale, with larger models better equipped to leverage\nadditional semantic information. This work represents a significant advancement\ntowards more efficient, context-aware autonomous driving systems, effectively\nbridging 3D scene representation with high-level semantic querying while\nmaintaining practical deployment considerations.\n", "date": "2024-08-07", "categories": []}
{"id": "2501.02519", "title": "Layout2Scene: 3D Semantic Layout Guided Scene Generation via Geometry\n  and Appearance Diffusion Priors", "abstract": "  3D scene generation conditioned on text prompts has significantly progressed\ndue to the development of 2D diffusion generation models. However, the textual\ndescription of 3D scenes is inherently inaccurate and lacks fine-grained\ncontrol during training, leading to implausible scene generation. As an\nintuitive and feasible solution, the 3D layout allows for precise specification\nof object locations within the scene. To this end, we present a text-to-scene\ngeneration method (namely, Layout2Scene) using additional semantic layout as\nthe prompt to inject precise control of 3D object positions. Specifically, we\nfirst introduce a scene hybrid representation to decouple objects and\nbackgrounds, which is initialized via a pre-trained text-to-3D model. Then, we\npropose a two-stage scheme to optimize the geometry and appearance of the\ninitialized scene separately. To fully leverage 2D diffusion priors in geometry\nand appearance generation, we introduce a semantic-guided geometry diffusion\nmodel and a semantic-geometry guided diffusion model which are finetuned on a\nscene dataset. Extensive experiments demonstrate that our method can generate\nmore plausible and realistic scenes as compared to state-of-the-art approaches.\nFurthermore, the generated scene allows for flexible yet precise editing,\nthereby facilitating multiple downstream applications.\n", "date": "2025-01-05", "categories": []}
{"id": "2501.00352", "title": "PanoSLAM: Panoptic 3D Scene Reconstruction via Gaussian SLAM", "abstract": "  Understanding geometric, semantic, and instance information in 3D scenes from\nsequential video data is essential for applications in robotics and augmented\nreality. However, existing Simultaneous Localization and Mapping (SLAM) methods\ngenerally focus on either geometric or semantic reconstruction. In this paper,\nwe introduce PanoSLAM, the first SLAM system to integrate geometric\nreconstruction, 3D semantic segmentation, and 3D instance segmentation within a\nunified framework. Our approach builds upon 3D Gaussian Splatting, modified\nwith several critical components to enable efficient rendering of depth, color,\nsemantic, and instance information from arbitrary viewpoints. To achieve\npanoptic 3D scene reconstruction from sequential RGB-D videos, we propose an\nonline Spatial-Temporal Lifting (STL) module that transfers 2D panoptic\npredictions from vision models into 3D Gaussian representations. This STL\nmodule addresses the challenges of label noise and inconsistencies in 2D\npredictions by refining the pseudo labels across multi-view inputs, creating a\ncoherent 3D representation that enhances segmentation accuracy. Our experiments\nshow that PanoSLAM outperforms recent semantic SLAM methods in both mapping and\ntracking accuracy. For the first time, it achieves panoptic 3D reconstruction\nof open-world environments directly from the RGB-D video.\n(https://github.com/runnanchen/PanoSLAM)\n", "date": "2024-12-31", "categories": []}
{"id": "2412.20480", "title": "MR-Occ: Efficient Camera-LiDAR 3D Semantic Occupancy Prediction Using\n  Hierarchical Multi-Resolution Voxel Representation", "abstract": "  Accurate 3D perception is essential for understanding the environment in\nautonomous driving. Recent advancements in 3D semantic occupancy prediction\nhave leveraged camera-LiDAR fusion to improve robustness and accuracy. However,\ncurrent methods allocate computational resources uniformly across all voxels,\nleading to inefficiency, and they also fail to adequately address occlusions,\nresulting in reduced accuracy in challenging scenarios. We propose MR-Occ, a\nnovel approach for camera-LiDAR fusion-based 3D semantic occupancy prediction,\naddressing these challenges through three key components: Hierarchical Voxel\nFeature Refinement (HVFR), Multi-scale Occupancy Decoder (MOD), and Pixel to\nVoxel Fusion Network (PVF-Net). HVFR improves performance by enhancing features\nfor critical voxels, reducing computational cost. MOD introduces an `occluded'\nclass to better handle regions obscured from sensor view, improving accuracy.\nPVF-Net leverages densified LiDAR features to effectively fuse camera and LiDAR\ndata through a deformable attention mechanism. Extensive experiments\ndemonstrate that MR-Occ achieves state-of-the-art performance on the\nnuScenes-Occupancy dataset, surpassing previous approaches by +5.2% in IoU and\n+5.3% in mIoU while using fewer parameters and FLOPs. Moreover, MR-Occ\ndemonstrates superior performance on the SemanticKITTI dataset, further\nvalidating its effectiveness and generalizability across diverse 3D semantic\noccupancy benchmarks.\n", "date": "2024-12-29", "categories": []}
{"id": "2405.15826", "title": "3D Learnable Supertoken Transformer for LiDAR Point Cloud Scene\n  Segmentation", "abstract": "  3D Transformers have achieved great success in point cloud understanding and\nrepresentation. However, there is still considerable scope for further\ndevelopment in effective and efficient Transformers for large-scale LiDAR point\ncloud scene segmentation. This paper proposes a novel 3D Transformer framework,\nnamed 3D Learnable Supertoken Transformer (3DLST). The key contributions are\nsummarized as follows. Firstly, we introduce the first Dynamic Supertoken\nOptimization (DSO) block for efficient token clustering and aggregating, where\nthe learnable supertoken definition avoids the time-consuming pre-processing of\ntraditional superpoint generation. Since the learnable supertokens can be\ndynamically optimized by multi-level deep features during network learning,\nthey are tailored to the semantic homogeneity-aware token clustering. Secondly,\nan efficient Cross-Attention-guided Upsampling (CAU) block is proposed for\ntoken reconstruction from optimized supertokens. Thirdly, the 3DLST is equipped\nwith a novel W-net architecture instead of the common U-net design, which is\nmore suitable for Transformer-based feature learning. The SOTA performance on\nthree challenging LiDAR datasets (airborne MultiSpectral LiDAR (MS-LiDAR)\n(89.3% of the average F1 score), DALES (80.2% of mIoU), and Toronto-3D dataset\n(80.4% of mIoU)) demonstrate the superiority of 3DLST and its strong\nadaptability to various LiDAR point cloud data (airborne MS-LiDAR, aerial\nLiDAR, and vehicle-mounted LiDAR data). Furthermore, 3DLST also achieves\nsatisfactory results in terms of algorithm efficiency, which is up to 5x faster\nthan previous best-performing methods.\n", "date": "2024-05-23", "categories": []}
{"id": "2412.16932", "title": "GSemSplat: Generalizable Semantic 3D Gaussian Splatting from\n  Uncalibrated Image Pairs", "abstract": "  Modeling and understanding the 3D world is crucial for various applications,\nfrom augmented reality to robotic navigation. Recent advancements based on 3D\nGaussian Splatting have integrated semantic information from multi-view images\ninto Gaussian primitives. However, these methods typically require costly\nper-scene optimization from dense calibrated images, limiting their\npracticality. In this paper, we consider the new task of generalizable 3D\nsemantic field modeling from sparse, uncalibrated image pairs. Building upon\nthe Splatt3R architecture, we introduce GSemSplat, a framework that learns\nopen-vocabulary semantic representations linked to 3D Gaussians without the\nneed for per-scene optimization, dense image collections or calibration. To\nensure effective and reliable learning of semantic features in 3D space, we\nemploy a dual-feature approach that leverages both region-specific and\ncontext-aware semantic features as supervision in the 2D space. This allows us\nto capitalize on their complementary strengths. Experimental results on the\nScanNet++ dataset demonstrate the effectiveness and superiority of our approach\ncompared to the traditional scene-specific method. We hope our work will\ninspire more research into generalizable 3D understanding.\n", "date": "2024-12-22", "categories": []}
{"id": "2410.22306", "title": "Multi-Object 3D Grounding with Dynamic Modules and Language-Informed\n  Spatial Attention", "abstract": "  Multi-object 3D Grounding involves locating 3D boxes based on a given query\nphrase from a point cloud. It is a challenging and significant task with\nnumerous applications in visual understanding, human-computer interaction, and\nrobotics. To tackle this challenge, we introduce D-LISA, a two-stage approach\nincorporating three innovations. First, a dynamic vision module that enables a\nvariable and learnable number of box proposals. Second, a dynamic camera\npositioning that extracts features for each proposal. Third, a\nlanguage-informed spatial attention module that better reasons over the\nproposals to output the final prediction. Empirically, experiments show that\nour method outperforms the state-of-the-art methods on multi-object 3D\ngrounding by 12.8% (absolute) and is competitive in single-object 3D grounding.\n", "date": "2024-10-29", "categories": []}
{"id": "2407.01220", "title": "Fast and Efficient: Mask Neural Fields for 3D Scene Segmentation", "abstract": "  Understanding 3D scenes is a crucial challenge in computer vision research\nwith applications spanning multiple domains. Recent advancements in distilling\n2D vision-language foundation models into neural fields, like NeRF and 3DGS,\nenable open-vocabulary segmentation of 3D scenes from 2D multi-view images\nwithout the need for precise 3D annotations. However, while effective, these\nmethods typically rely on the per-pixel distillation of high-dimensional CLIP\nfeatures, introducing ambiguity and necessitating complex regularization\nstrategies, which adds inefficiency during training. This paper presents\nMaskField, which enables efficient 3D open-vocabulary segmentation with neural\nfields from a novel perspective. Unlike previous methods, MaskField decomposes\nthe distillation of mask and semantic features from foundation models by\nformulating a mask feature field and queries. MaskField overcomes ambiguous\nobject boundaries by naturally introducing SAM segmented object shapes without\nextra regularization during training. By circumventing the direct handling of\ndense high-dimensional CLIP features during training, MaskField is particularly\ncompatible with explicit scene representations like 3DGS. Our extensive\nexperiments show that MaskField not only surpasses prior state-of-the-art\nmethods but also achieves remarkably fast convergence. We hope that MaskField\nwill inspire further exploration into how neural fields can be trained to\ncomprehend 3D scenes from 2D models.\n", "date": "2024-07-01", "categories": []}
{"id": "2412.14837", "title": "ObjVariantEnsemble: Advancing Point Cloud LLM Evaluation in Challenging\n  Scenes with Subtly Distinguished Objects", "abstract": "  3D scene understanding is an important task, and there has been a recent\nsurge of research interest in aligning 3D representations of point clouds with\ntext to empower embodied AI. However, due to the lack of comprehensive 3D\nbenchmarks, the capabilities of 3D models in real-world scenes, particularly\nthose that are challenging with subtly distinguished objects, remain\ninsufficiently investigated. To facilitate a more thorough evaluation of 3D\nmodels' capabilities, we propose a scheme, ObjVariantEnsemble, to\nsystematically introduce more scenes with specified object classes, colors,\nshapes, quantities, and spatial relationships to meet model evaluation needs.\nMore importantly, we intentionally construct scenes with similar objects to a\ncertain degree and design an LLM-VLM-cooperated annotator to capture key\ndistinctions as annotations. The resultant benchmark can better challenge 3D\nmodels, reveal their shortcomings in understanding, and potentially aid in the\nfurther development of 3D models.\n", "date": "2024-12-19", "categories": []}
{"id": "2404.11064", "title": "Rethinking 3D Dense Caption and Visual Grounding in A Unified Framework\n  through Prompt-based Localization", "abstract": "  3D Visual Grounding (3DVG) and 3D Dense Captioning (3DDC) are two crucial\ntasks in various 3D applications, which require both shared and complementary\ninformation in localization and visual-language relationships. Therefore,\nexisting approaches adopt the two-stage \"detect-then-describe/discriminate\"\npipeline, which relies heavily on the performance of the detector, resulting in\nsuboptimal performance. Inspired by DETR, we propose a unified framework,\n3DGCTR, to jointly solve these two distinct but closely related tasks in an\nend-to-end fashion. The key idea is to reconsider the prompt-based localization\nability of the 3DVG model. In this way, the 3DVG model with a well-designed\nprompt as input can assist the 3DDC task by extracting localization information\nfrom the prompt. In terms of implementation, we integrate a Lightweight Caption\nHead into the existing 3DVG network with a Caption Text Prompt as a connection,\neffectively harnessing the existing 3DVG model's inherent localization\ncapacity, thereby boosting 3DDC capability. This integration facilitates\nsimultaneous multi-task training on both tasks, mutually enhancing their\nperformance. Extensive experimental results demonstrate the effectiveness of\nthis approach. Specifically, on the ScanRefer dataset, 3DGCTR surpasses the\nstate-of-the-art 3DDC method by 4.3% in CIDEr@0.5IoU in MLE training and\nimproves upon the SOTA 3DVG method by 3.16% in Acc@0.25IoU. The codes are at\nhttps://github.com/Leon1207/3DGCTR.\n", "date": "2024-04-17", "categories": []}
{"id": "2412.07616", "title": "PVP: Polar Representation Boost for 3D Semantic Occupancy Prediction", "abstract": "  Recently, polar coordinate-based representations have shown promise for 3D\nperceptual tasks. Compared to Cartesian methods, polar grids provide a viable\nalternative, offering better detail preservation in nearby spaces while\ncovering larger areas. However, they face feature distortion due to non-uniform\ndivision. To address these issues, we introduce the Polar Voxel Occupancy\nPredictor (PVP), a novel 3D multi-modal predictor that operates in polar\ncoordinates. PVP features two key design elements to overcome distortion: a\nGlobal Represent Propagation (GRP) module that integrates global spatial data\ninto 3D volumes, and a Plane Decomposed Convolution (PD-Conv) that simplifies\n3D distortions into 2D convolutions. These innovations enable PVP to outperform\nexisting methods, achieving significant improvements in mIoU and IoU metrics on\nthe OpenOccupancy dataset.\n", "date": "2024-12-10", "categories": []}
{"id": "2412.13569", "title": "Multi-View Pedestrian Occupancy Prediction with a Novel Synthetic\n  Dataset", "abstract": "  We address an advanced challenge of predicting pedestrian occupancy as an\nextension of multi-view pedestrian detection in urban traffic. To support this,\nwe have created a new synthetic dataset called MVP-Occ, designed for dense\npedestrian scenarios in large-scale scenes. Our dataset provides detailed\nrepresentations of pedestrians using voxel structures, accompanied by rich\nsemantic scene understanding labels, facilitating visual navigation and\ninsights into pedestrian spatial information. Furthermore, we present a robust\nbaseline model, termed OmniOcc, capable of predicting both the voxel occupancy\nstate and panoptic labels for the entire scene from multi-view images. Through\nin-depth analysis, we identify and evaluate the key elements of our proposed\nmodel, highlighting their specific contributions and importance.\n", "date": "2024-12-18", "categories": []}
{"id": "2412.11241", "title": "Volumetric Mapping with Panoptic Refinement via Kernel Density\n  Estimation for Mobile Robots", "abstract": "  Reconstructing three-dimensional (3D) scenes with semantic understanding is\nvital in many robotic applications. Robots need to identify which objects,\nalong with their positions and shapes, to manipulate them precisely with given\ntasks. Mobile robots, especially, usually use lightweight networks to segment\nobjects on RGB images and then localize them via depth maps; however, they\noften encounter out-of-distribution scenarios where masks over-cover the\nobjects. In this paper, we address the problem of panoptic segmentation quality\nin 3D scene reconstruction by refining segmentation errors using non-parametric\nstatistical methods. To enhance mask precision, we map the predicted masks into\na depth frame to estimate their distribution via kernel densities. The outliers\nin depth perception are then rejected without the need for additional\nparameters in an adaptive manner to out-of-distribution scenarios, followed by\n3D reconstruction using projective signed distance functions (SDFs). We\nvalidate our method on a synthetic dataset, which shows improvements in both\nquantitative and qualitative results for panoptic mapping. Through real-world\ntesting, the results furthermore show our method's capability to be deployed on\na real-robot system. Our source code is available at:\nhttps://github.com/mkhangg/refined panoptic mapping.\n", "date": "2024-12-15", "categories": []}
{"id": "2412.11258", "title": "GaussianProperty: Integrating Physical Properties to 3D Gaussians with\n  LMMs", "abstract": "  Estimating physical properties for visual data is a crucial task in computer\nvision, graphics, and robotics, underpinning applications such as augmented\nreality, physical simulation, and robotic grasping. However, this area remains\nunder-explored due to the inherent ambiguities in physical property estimation.\nTo address these challenges, we introduce GaussianProperty, a training-free\nframework that assigns physical properties of materials to 3D Gaussians.\nSpecifically, we integrate the segmentation capability of SAM with the\nrecognition capability of GPT-4V(ision) to formulate a global-local physical\nproperty reasoning module for 2D images. Then we project the physical\nproperties from multi-view 2D images to 3D Gaussians using a voting strategy.\nWe demonstrate that 3D Gaussians with physical property annotations enable\napplications in physics-based dynamic simulation and robotic grasping. For\nphysics-based dynamic simulation, we leverage the Material Point Method (MPM)\nfor realistic dynamic simulation. For robot grasping, we develop a grasping\nforce prediction strategy that estimates a safe force range required for object\ngrasping based on the estimated physical properties. Extensive experiments on\nmaterial segmentation, physics-based dynamic simulation, and robotic grasping\nvalidate the effectiveness of our proposed method, highlighting its crucial\nrole in understanding physical properties from visual data. Online demo, code,\nmore cases and annotated datasets are available on\n\\href{https://Gaussian-Property.github.io}{this https URL}.\n", "date": "2024-12-15", "categories": []}
{"id": "2302.03531", "title": "Structured Generative Models for Scene Understanding", "abstract": "  This position paper argues for the use of \\emph{structured generative models}\n(SGMs) for the understanding of static scenes. This requires the reconstruction\nof a 3D scene from an input image (or a set of multi-view images), whereby the\ncontents of the image(s) are causally explained in terms of models of\ninstantiated objects, each with their own type, shape, appearance and pose,\nalong with global variables like scene lighting and camera parameters. This\napproach also requires scene models which account for the co-occurrences and\ninter-relationships of objects in a scene. The SGM approach has the merits that\nit is compositional and generative, which lead to interpretability and\neditability. \\\\\\\\ To pursue the SGM agenda, we need models for objects and\nscenes, and approaches to carry out inference. We first review models for\nobjects, which include ``things'' (object categories that have a well defined\nshape), and ``stuff'' (categories which have amorphous spatial extent). We then\nmove on to review \\emph{scene models} which describe the inter-relationships of\nobjects. Perhaps the most challenging problem for SGMs is \\emph{inference} of\nthe objects, lighting and camera parameters, and scene inter-relationships from\ninput consisting of a single or multiple images. We conclude with a discussion\nof issues that need addressing to advance the SGM agenda.\n", "date": "2023-02-07", "categories": []}
{"id": "2412.10294", "title": "Coherent 3D Scene Diffusion From a Single RGB Image", "abstract": "  We present a novel diffusion-based approach for coherent 3D scene\nreconstruction from a single RGB image. Our method utilizes an\nimage-conditioned 3D scene diffusion model to simultaneously denoise the 3D\nposes and geometries of all objects within the scene. Motivated by the\nill-posed nature of the task and to obtain consistent scene reconstruction\nresults, we learn a generative scene prior by conditioning on all scene objects\nsimultaneously to capture the scene context and by allowing the model to learn\ninter-object relationships throughout the diffusion process. We further propose\nan efficient surface alignment loss to facilitate training even in the absence\nof full ground-truth annotation, which is common in publicly available\ndatasets. This loss leverages an expressive shape representation, which enables\ndirect point sampling from intermediate shape predictions. By framing the task\nof single RGB image 3D scene reconstruction as a conditional diffusion process,\nour approach surpasses current state-of-the-art methods, achieving a 12.04%\nimprovement in AP3D on SUN RGB-D and a 13.43% increase in F-Score on Pix3D.\n", "date": "2024-12-13", "categories": []}
{"id": "2404.06442", "title": "QueSTMaps: Queryable Semantic Topological Maps for 3D Scene\n  Understanding", "abstract": "  Robotic tasks such as planning and navigation require a hierarchical semantic\nunderstanding of a scene, which could include multiple floors and rooms.\nCurrent methods primarily focus on object segmentation for 3D scene\nunderstanding. However, such methods struggle to segment out topological\nregions like \"kitchen\" in the scene. In this work, we introduce a two-step\npipeline to solve this problem. First, we extract a topological map, i.e.,\nfloorplan of the indoor scene using a novel multi-channel occupancy\nrepresentation. Then, we generate CLIP-aligned features and semantic labels for\nevery room instance based on the objects it contains using a self-attention\ntransformer. Our language-topology alignment supports natural language\nquerying, e.g., a \"place to cook\" locates the \"kitchen\". We outperform the\ncurrent state-of-the-art on room segmentation by ~20% and room classification\nby ~12%. Our detailed qualitative analysis and ablation studies provide\ninsights into the problem of joint structural and semantic 3D scene\nunderstanding. Project Page: quest-maps.github.io\n", "date": "2024-04-09", "categories": []}
{"id": "2406.02058", "title": "OpenGaussian: Towards Point-Level 3D Gaussian-based Open Vocabulary\n  Understanding", "abstract": "  This paper introduces OpenGaussian, a method based on 3D Gaussian Splatting\n(3DGS) capable of 3D point-level open vocabulary understanding. Our primary\nmotivation stems from observing that existing 3DGS-based open vocabulary\nmethods mainly focus on 2D pixel-level parsing. These methods struggle with 3D\npoint-level tasks due to weak feature expressiveness and inaccurate 2D-3D\nfeature associations. To ensure robust feature presentation and 3D point-level\nunderstanding, we first employ SAM masks without cross-frame associations to\ntrain instance features with 3D consistency. These features exhibit both\nintra-object consistency and inter-object distinction. Then, we propose a\ntwo-stage codebook to discretize these features from coarse to fine levels. At\nthe coarse level, we consider the positional information of 3D points to\nachieve location-based clustering, which is then refined at the fine level.\nFinally, we introduce an instance-level 3D-2D feature association method that\nlinks 3D points to 2D masks, which are further associated with 2D CLIP\nfeatures. Extensive experiments, including open vocabulary-based 3D object\nselection, 3D point cloud understanding, click-based 3D object selection, and\nablation studies, demonstrate the effectiveness of our proposed method. The\nsource code is available at our project page:\nhttps://3d-aigc.github.io/OpenGaussian\n", "date": "2024-06-04", "categories": []}
{"id": "2410.07987", "title": "A transition towards virtual representations of visual scenes", "abstract": "  Visual scene understanding is a fundamental task in computer vision that aims\nto extract meaningful information from visual data. It traditionally involves\ndisjoint and specialized algorithms for different tasks that are tailored for\nspecific application scenarios. This can be cumbersome when designing complex\nsystems that include processing of visual and semantic data extracted from\nvisual scenes, which is even more noticeable nowadays with the influx of\napplications for virtual or augmented reality. When designing a system that\nemploys automatic visual scene understanding to enable a precise and\nsemantically coherent description of the underlying scene, which can be used to\nfuel a visualization component with 3D virtual synthesis, the lack of\nflexibility and unified frameworks become more prominent. To alleviate this\nissue and its inherent problems, we propose an architecture that addresses the\nchallenges of visual scene understanding and description towards a 3D virtual\nsynthesis that enables an adaptable, unified and coherent solution.\nFurthermore, we expose how our proposition can be of use into multiple\napplication areas. Additionally, we also present a proof of concept system that\nemploys our architecture to further prove its usability in practice.\n", "date": "2024-10-10", "categories": []}
{"id": "2412.04384", "title": "GaussianFormer-2: Probabilistic Gaussian Superposition for Efficient 3D\n  Occupancy Prediction", "abstract": "  3D semantic occupancy prediction is an important task for robust\nvision-centric autonomous driving, which predicts fine-grained geometry and\nsemantics of the surrounding scene. Most existing methods leverage dense\ngrid-based scene representations, overlooking the spatial sparsity of the\ndriving scenes. Although 3D semantic Gaussian serves as an object-centric\nsparse alternative, most of the Gaussians still describe the empty region with\nlow efficiency. To address this, we propose a probabilistic Gaussian\nsuperposition model which interprets each Gaussian as a probability\ndistribution of its neighborhood being occupied and conforms to probabilistic\nmultiplication to derive the overall geometry. Furthermore, we adopt the exact\nGaussian mixture model for semantics calculation to avoid unnecessary\noverlapping of Gaussians. To effectively initialize Gaussians in non-empty\nregion, we design a distribution-based initialization module which learns the\npixel-aligned occupancy distribution instead of the depth of surfaces. We\nconduct extensive experiments on nuScenes and KITTI-360 datasets and our\nGaussianFormer-2 achieves state-of-the-art performance with high efficiency.\nCode: https://github.com/huang-yh/GaussianFormer.\n", "date": "2024-12-05", "categories": []}
{"id": "2407.14108", "title": "GaussianBeV: 3D Gaussian Representation meets Perception Models for BeV\n  Segmentation", "abstract": "  The Bird's-eye View (BeV) representation is widely used for 3D perception\nfrom multi-view camera images. It allows to merge features from different\ncameras into a common space, providing a unified representation of the 3D\nscene. The key component is the view transformer, which transforms image views\ninto the BeV. However, actual view transformer methods based on geometry or\ncross-attention do not provide a sufficiently detailed representation of the\nscene, as they use a sub-sampling of the 3D space that is non-optimal for\nmodeling the fine structures of the environment. In this paper, we propose\nGaussianBeV, a novel method for transforming image features to BeV by finely\nrepresenting the scene using a set of 3D gaussians located and oriented in 3D\nspace. This representation is then splattered to produce the BeV feature map by\nadapting recent advances in 3D representation rendering based on gaussian\nsplatting. GaussianBeV is the first approach to use this 3D gaussian modeling\nand 3D scene rendering process online, i.e. without optimizing it on a specific\nscene and directly integrated into a single stage model for BeV scene\nunderstanding. Experiments show that the proposed representation is highly\neffective and place GaussianBeV as the new state-of-the-art on the BeV semantic\nsegmentation task on the nuScenes dataset.\n", "date": "2024-07-19", "categories": []}
{"id": "2412.02245", "title": "SparseLGS: Sparse View Language Embedded Gaussian Splatting", "abstract": "  Recently, several studies have combined Gaussian Splatting to obtain scene\nrepresentations with language embeddings for open-vocabulary 3D scene\nunderstanding. While these methods perform well, they essentially require very\ndense multi-view inputs, limiting their applicability in real-world scenarios.\nIn this work, we propose SparseLGS to address the challenge of 3D scene\nunderstanding with pose-free and sparse view input images. Our method leverages\na learning-based dense stereo model to handle pose-free and sparse inputs, and\na three-step region matching approach to address the multi-view semantic\ninconsistency problem, which is especially important for sparse inputs.\nDifferent from directly learning high-dimensional CLIP features, we extract\nlow-dimensional information and build bijections to avoid excessive learning\nand storage costs. We introduce a reconstruction loss during semantic training\nto improve Gaussian positions and shapes. To the best of our knowledge, we are\nthe first to address the 3D semantic field problem with sparse pose-free\ninputs. Experimental results show that SparseLGS achieves comparable quality\nwhen reconstructing semantic fields with fewer inputs (3-4 views) compared to\nprevious SOTA methods with dense input. Besides, when using the same sparse\ninput, SparseLGS leads significantly in quality and heavily improves the\ncomputation speed (5$\\times$speedup). Project page:\nhttps://ustc3dv.github.io/SparseLGS\n", "date": "2024-12-03", "categories": []}
{"id": "2412.02449", "title": "BYE: Build Your Encoder with One Sequence of Exploration Data for\n  Long-Term Dynamic Scene Understanding", "abstract": "  Dynamic scene understanding remains a persistent challenge in robotic\napplications. Early dynamic mapping methods focused on mitigating the negative\ninfluence of short-term dynamic objects on camera motion estimation by masking\nor tracking specific categories, which often fall short in adapting to\nlong-term scene changes. Recent efforts address object association in long-term\ndynamic environments using neural networks trained on synthetic datasets, but\nthey still rely on predefined object shapes and categories. Other methods\nincorporate visual, geometric, or semantic heuristics for the association but\noften lack robustness. In this work, we introduce BYE, a class-agnostic,\nper-scene point cloud encoder that removes the need for predefined categories,\nshape priors, or extensive association datasets. Trained on only a single\nsequence of exploration data, BYE can efficiently perform object association in\ndynamically changing scenes. We further propose an ensembling scheme combining\nthe semantic strengths of Vision Language Models (VLMs) with the scene-specific\nexpertise of BYE, achieving a 7% improvement and a 95% success rate in object\nassociation tasks. Code and dataset are available at\nhttps://byencoder.github.io.\n", "date": "2024-12-03", "categories": []}
{"id": "2403.03077", "title": "MiKASA: Multi-Key-Anchor & Scene-Aware Transformer for 3D Visual\n  Grounding", "abstract": "  3D visual grounding involves matching natural language descriptions with\ntheir corresponding objects in 3D spaces. Existing methods often face\nchallenges with accuracy in object recognition and struggle in interpreting\ncomplex linguistic queries, particularly with descriptions that involve\nmultiple anchors or are view-dependent. In response, we present the MiKASA\n(Multi-Key-Anchor Scene-Aware) Transformer. Our novel end-to-end trained model\nintegrates a self-attention-based scene-aware object encoder and an original\nmulti-key-anchor technique, enhancing object recognition accuracy and the\nunderstanding of spatial relationships. Furthermore, MiKASA improves the\nexplainability of decision-making, facilitating error diagnosis. Our model\nachieves the highest overall accuracy in the Referit3D challenge for both the\nSr3D and Nr3D datasets, particularly excelling by a large margin in categories\nthat require viewpoint-dependent descriptions.\n", "date": "2024-03-05", "categories": []}
{"id": "2412.00734", "title": "ChatSplat: 3D Conversational Gaussian Splatting", "abstract": "  Humans naturally interact with their 3D surroundings using language, and\nmodeling 3D language fields for scene understanding and interaction has gained\ngrowing interest. This paper introduces ChatSplat, a system that constructs a\n3D language field, enabling rich chat-based interaction within 3D space. Unlike\nexisting methods that primarily use CLIP-derived language features focused\nsolely on segmentation, ChatSplat facilitates interaction on three levels:\nobjects, views, and the entire 3D scene. For view-level interaction, we\ndesigned an encoder that encodes the rendered feature map of each view into\ntokens, which are then processed by a large language model (LLM) for\nconversation. At the scene level, ChatSplat combines multi-view tokens,\nenabling interactions that consider the entire scene. For object-level\ninteraction, ChatSplat uses a patch-wise language embedding, unlike LangSplat's\npixel-wise language embedding that implicitly includes mask and embedding.\nHere, we explicitly decouple the language embedding into separate mask and\nfeature map representations, allowing more flexible object-level interaction.\nTo address the challenge of learning 3D Gaussians posed by the complex and\ndiverse distribution of language embeddings used in the LLM, we introduce a\nlearnable normalization technique to standardize these embeddings, facilitating\neffective learning. Extensive experimental results demonstrate that ChatSplat\nsupports multi-level interactions -- object, view, and scene -- within 3D\nspace, enhancing both understanding and engagement.\n", "date": "2024-12-01", "categories": []}
{"id": "2412.01431", "title": "Semantic Scene Completion with Multi-Feature Data Balancing Network", "abstract": "  Semantic Scene Completion (SSC) is a critical task in computer vision, that\nutilized in applications such as virtual reality (VR). SSC aims to construct\ndetailed 3D models from partial views by transforming a single 2D image into a\n3D representation, assigning each voxel a semantic label. The main challenge\nlies in completing 3D volumes with limited information, compounded by data\nimbalance, inter-class ambiguity, and intra-class diversity in indoor scenes.\nTo address this, we propose the Multi-Feature Data Balancing Network (MDBNet),\na dual-head model for RGB and depth data (F-TSDF) inputs. Our hybrid\nencoder-decoder architecture with identity transformation in a pre-activation\nresidual module (ITRM) effectively manages diverse signals within F-TSDF. We\nevaluate RGB feature fusion strategies and use a combined loss function cross\nentropy for 2D RGB features and weighted cross-entropy for 3D SSC predictions.\nMDBNet results surpass comparable state-of-the-art (SOTA) methods on NYU\ndatasets, demonstrating the effectiveness of our approach.\n", "date": "2024-12-02", "categories": []}
{"id": "2412.01539", "title": "The Bare Necessities: Designing Simple, Effective Open-Vocabulary Scene\n  Graphs", "abstract": "  3D open-vocabulary scene graph methods are a promising map representation for\nembodied agents, however many current approaches are computationally expensive.\nIn this paper, we reexamine the critical design choices established in previous\nworks to optimize both efficiency and performance. We propose a general scene\ngraph framework and conduct three studies that focus on image pre-processing,\nfeature fusion, and feature selection. Our findings reveal that commonly used\nimage pre-processing techniques provide minimal performance improvement while\ntripling computation (on a per object view basis). We also show that averaging\nfeature labels across different views significantly degrades performance. We\nstudy alternative feature selection strategies that enhance performance without\nadding unnecessary computational costs. Based on our findings, we introduce a\ncomputationally balanced approach for 3D point cloud segmentation with\nper-object features. The approach matches state-of-the-art classification\naccuracy while achieving a threefold reduction in computation.\n", "date": "2024-12-02", "categories": []}
{"id": "2411.18666", "title": "3D Scene Graph Guided Vision-Language Pre-training", "abstract": "  3D vision-language (VL) reasoning has gained significant attention due to its\npotential to bridge the 3D physical world with natural language descriptions.\nExisting approaches typically follow task-specific, highly specialized\nparadigms. Therefore, these methods focus on a limited range of reasoning\nsub-tasks and rely heavily on the hand-crafted modules and auxiliary losses.\nThis highlights the need for a simpler, unified and general-purpose model. In\nthis paper, we leverage the inherent connection between 3D scene graphs and\nnatural language, proposing a 3D scene graph-guided vision-language\npre-training (VLP) framework. Our approach utilizes modality encoders, graph\nconvolutional layers and cross-attention layers to learn universal\nrepresentations that adapt to a variety of 3D VL reasoning tasks, thereby\neliminating the need for task-specific designs. The pre-training objectives\ninclude: 1) Scene graph-guided contrastive learning, which leverages the strong\ncorrelation between 3D scene graphs and natural language to align 3D objects\nwith textual features at various fine-grained levels; and 2) Masked modality\nlearning, which uses cross-modality information to reconstruct masked words and\n3D objects. Instead of directly reconstructing the 3D point clouds of masked\nobjects, we use position clues to predict their semantic categories. Extensive\nexperiments demonstrate that our pre-training model, when fine-tuned on several\ndownstream tasks, achieves performance comparable to or better than existing\nmethods in tasks such as 3D visual grounding, 3D dense captioning, and 3D\nquestion answering.\n", "date": "2024-11-27", "categories": []}
{"id": "2411.18807", "title": "Reconstructing Animals and the Wild", "abstract": "  The idea of 3D reconstruction as scene understanding is foundational in\ncomputer vision. Reconstructing 3D scenes from 2D visual observations requires\nstrong priors to disambiguate structure. Much work has been focused on the\nanthropocentric, which, characterized by smooth surfaces, coherent normals, and\nregular edges, allows for the integration of strong geometric inductive biases.\nHere, we consider a more challenging problem where such assumptions do not\nhold: the reconstruction of natural scenes containing trees, bushes, boulders,\nand animals. While numerous works have attempted to tackle the problem of\nreconstructing animals in the wild, they have focused solely on the animal,\nneglecting environmental context. This limits their usefulness for analysis\ntasks, as animals exist inherently within the 3D world, and information is lost\nwhen environmental factors are disregarded. We propose a method to reconstruct\nnatural scenes from single images. We base our approach on recent advances\nleveraging the strong world priors ingrained in Large Language Models and train\nan autoregressive model to decode a CLIP embedding into a structured\ncompositional scene representation, encompassing both animals and the wild\n(RAW). To enable this, we propose a synthetic dataset comprising one million\nimages and thousands of assets. Our approach, having been trained solely on\nsynthetic data, generalizes to the task of reconstructing animals and their\nenvironments in real-world images. We will release our dataset and code to\nencourage future research at https://raw.is.tue.mpg.de/\n", "date": "2024-11-27", "categories": []}
{"id": "2407.03263", "title": "A Unified Framework for 3D Scene Understanding", "abstract": "  We propose UniSeg3D, a unified 3D scene understanding framework that achieves\npanoptic, semantic, instance, interactive, referring, and open-vocabulary\nsegmentation tasks within a single model. Most previous 3D segmentation\napproaches are typically tailored to a specific task, limiting their\nunderstanding of 3D scenes to a task-specific perspective. In contrast, the\nproposed method unifies six tasks into unified representations processed by the\nsame Transformer. It facilitates inter-task knowledge sharing, thereby\npromoting comprehensive 3D scene understanding. To take advantage of multi-task\nunification, we enhance performance by establishing explicit inter-task\nassociations. Specifically, we design knowledge distillation and contrastive\nlearning methods to transfer task-specific knowledge across different tasks.\nExperiments on three benchmarks, including ScanNet20, ScanRefer, and\nScanNet200, demonstrate that the UniSeg3D consistently outperforms current SOTA\nmethods, even those specialized for individual tasks. We hope UniSeg3D can\nserve as a solid unified baseline and inspire future work. Code and models are\navailable at https://github.com/dk-liang/UniSeg3D.\n", "date": "2024-07-03", "categories": []}
{"id": "2311.17084", "title": "DepthSSC: Monocular 3D Semantic Scene Completion via Depth-Spatial\n  Alignment and Voxel Adaptation", "abstract": "  The task of 3D semantic scene completion using monocular cameras is gaining\nsignificant attention in the field of autonomous driving. This task aims to\npredict the occupancy status and semantic labels of each voxel in a 3D scene\nfrom partial image inputs. Despite numerous existing methods, many face\nchallenges such as inaccurately predicting object shapes and misclassifying\nobject boundaries. To address these issues, we propose DepthSSC, an advanced\nmethod for semantic scene completion using only monocular cameras. DepthSSC\nintegrates the Spatial Transformation Graph Fusion (ST-GF) module with\nGeometric-Aware Voxelization (GAV), enabling dynamic adjustment of voxel\nresolution to accommodate the geometric complexity of 3D space. This ensures\nprecise alignment between spatial and depth information, effectively mitigating\nissues such as object boundary distortion and incorrect depth perception found\nin previous methods. Evaluations on the SemanticKITTI and SSCBench-KITTI-360\ndataset demonstrate that DepthSSC not only captures intricate 3D structural\ndetails effectively but also achieves state-of-the-art performance.\n", "date": "2023-11-28", "categories": []}
{"id": "2403.07563", "title": "Learning Generalizable Feature Fields for Mobile Manipulation", "abstract": "  An open problem in mobile manipulation is how to represent objects and scenes\nin a unified manner so that robots can use both for navigation and\nmanipulation. The latter requires capturing intricate geometry while\nunderstanding fine-grained semantics, whereas the former involves capturing the\ncomplexity inherent at an expansive physical scale. In this work, we present\nGeFF (Generalizable Feature Fields), a scene-level generalizable neural feature\nfield that acts as a unified representation for both navigation and\nmanipulation that performs in real-time. To do so, we treat generative novel\nview synthesis as a pre-training task, and then align the resulting rich scene\npriors with natural language via CLIP feature distillation. We demonstrate the\neffectiveness of this approach by deploying GeFF on a quadrupedal robot\nequipped with a manipulator. We quantitatively evaluate GeFF's ability for\nopen-vocabulary object-/part-level manipulation and show that GeFF outperforms\npoint-based baselines in runtime and storage-accuracy trade-offs, with\nqualitative examples of semantics-aware navigation and articulated object\nmanipulation.\n", "date": "2024-03-12", "categories": []}
{"id": "2403.16888", "title": "Towards Balanced RGB-TSDF Fusion for Consistent Semantic Scene\n  Completion by 3D RGB Feature Completion and a Classwise Entropy Loss Function", "abstract": "  Semantic Scene Completion (SSC) aims to jointly infer semantics and\noccupancies of 3D scenes. Truncated Signed Distance Function (TSDF), a 3D\nencoding of depth, has been a common input for SSC. Furthermore, RGB-TSDF\nfusion, seems promising since these two modalities provide color and geometry\ninformation, respectively. Nevertheless, RGB-TSDF fusion has been considered\nnontrivial and commonly-used naive addition will result in inconsistent\nresults. We argue that the inconsistency comes from the sparsity of RGB\nfeatures upon projecting into 3D space, while TSDF features are dense, leading\nto imbalanced feature maps when summed up. To address this RGB-TSDF\ndistribution difference, we propose a two-stage network with a 3D RGB feature\ncompletion module that completes RGB features with meaningful values for\noccluded areas. Moreover, we propose an effective classwise entropy loss\nfunction to punish inconsistency. Extensive experiments on public datasets\nverify that our method achieves state-of-the-art performance among methods that\ndo not adopt extra data.\n", "date": "2024-03-25", "categories": []}
{"id": "2403.19797", "title": "Efficient 3D Instance Mapping and Localization with Neural Fields", "abstract": "  We tackle the problem of learning an implicit scene representation for 3D\ninstance segmentation from a sequence of posed RGB images. Towards this, we\nintroduce 3DIML, a novel framework that efficiently learns a neural label field\nwhich can render 3D instance segmentation masks from novel viewpoints. Opposed\nto prior art that optimizes a neural field in a self-supervised manner,\nrequiring complicated training procedures and loss function design, 3DIML\nleverages a two-phase process. The first phase, InstanceMap, takes as input 2D\nsegmentation masks of the image sequence generated by a frontend instance\nsegmentation model, and associates corresponding masks across images to 3D\nlabels. These almost 3D-consistent pseudolabel masks are then used in the\nsecond phase, InstanceLift, to supervise the training of a neural label field,\nwhich interpolates regions missed by InstanceMap and resolves ambiguities.\nAdditionally, we introduce InstanceLoc, which enables near realtime\nlocalization of instance masks given a trained neural label field. We evaluate\n3DIML on sequences from the Replica and ScanNet datasets and demonstrate its\neffectiveness under mild assumptions for the image sequences. We achieve a\nlarge practical speedup over existing implicit scene representation methods\nwith comparable quality, showcasing its potential to facilitate faster and more\neffective 3D scene understanding.\n", "date": "2024-03-28", "categories": []}
{"id": "2411.15657", "title": "Training an Open-Vocabulary Monocular 3D Object Detection Model without\n  3D Data", "abstract": "  Open-vocabulary 3D object detection has recently attracted considerable\nattention due to its broad applications in autonomous driving and robotics,\nwhich aims to effectively recognize novel classes in previously unseen domains.\nHowever, existing point cloud-based open-vocabulary 3D detection models are\nlimited by their high deployment costs. In this work, we propose a novel\nopen-vocabulary monocular 3D object detection framework, dubbed OVM3D-Det,\nwhich trains detectors using only RGB images, making it both cost-effective and\nscalable to publicly available data. Unlike traditional methods, OVM3D-Det does\nnot require high-precision LiDAR or 3D sensor data for either input or\ngenerating 3D bounding boxes. Instead, it employs open-vocabulary 2D models and\npseudo-LiDAR to automatically label 3D objects in RGB images, fostering the\nlearning of open-vocabulary monocular 3D detectors. However, training 3D models\nwith labels directly derived from pseudo-LiDAR is inadequate due to imprecise\nboxes estimated from noisy point clouds and severely occluded objects. To\naddress these issues, we introduce two innovative designs: adaptive\npseudo-LiDAR erosion and bounding box refinement with prior knowledge from\nlarge language models. These techniques effectively calibrate the 3D labels and\nenable RGB-only training for 3D detectors. Extensive experiments demonstrate\nthe superiority of OVM3D-Det over baselines in both indoor and outdoor\nscenarios. The code will be released.\n", "date": "2024-11-23", "categories": []}
{"id": "2410.11187", "title": "Multiview Scene Graph", "abstract": "  A proper scene representation is central to the pursuit of spatial\nintelligence where agents can robustly reconstruct and efficiently understand\n3D scenes. A scene representation is either metric, such as landmark maps in 3D\nreconstruction, 3D bounding boxes in object detection, or voxel grids in\noccupancy prediction, or topological, such as pose graphs with loop closures in\nSLAM or visibility graphs in SfM. In this work, we propose to build Multiview\nScene Graphs (MSG) from unposed images, representing a scene topologically with\ninterconnected place and object nodes. The task of building MSG is challenging\nfor existing representation learning methods since it needs to jointly address\nboth visual place recognition, object detection, and object association from\nimages with limited fields of view and potentially large viewpoint changes. To\nevaluate any method tackling this task, we developed an MSG dataset and\nannotation based on a public 3D dataset. We also propose an evaluation metric\nbased on the intersection-over-union score of MSG edges. Moreover, we develop a\nnovel baseline method built on mainstream pretrained vision models, combining\nvisual place recognition and object association into one Transformer decoder\narchitecture. Experiments demonstrate that our method has superior performance\ncompared to existing relevant baselines.\n", "date": "2024-10-15", "categories": []}
{"id": "2410.15674", "title": "TALoS: Enhancing Semantic Scene Completion via Test-time Adaptation on\n  the Line of Sight", "abstract": "  Semantic Scene Completion (SSC) aims to perform geometric completion and\nsemantic segmentation simultaneously. Despite the promising results achieved by\nexisting studies, the inherently ill-posed nature of the task presents\nsignificant challenges in diverse driving scenarios. This paper introduces\nTALoS, a novel test-time adaptation approach for SSC that excavates the\ninformation available in driving environments. Specifically, we focus on that\nobservations made at a certain moment can serve as Ground Truth (GT) for scene\ncompletion at another moment. Given the characteristics of the LiDAR sensor, an\nobservation of an object at a certain location confirms both 1) the occupation\nof that location and 2) the absence of obstacles along the line of sight from\nthe LiDAR to that point. TALoS utilizes these observations to obtain\nself-supervision about occupancy and emptiness, guiding the model to adapt to\nthe scene in test time. In a similar manner, we aggregate reliable SSC\npredictions among multiple moments and leverage them as semantic pseudo-GT for\nadaptation. Further, to leverage future observations that are not accessible at\nthe current time, we present a dual optimization scheme using the model in\nwhich the update is delayed until the future observation is available.\nEvaluations on the SemanticKITTI validation and test sets demonstrate that\nTALoS significantly improves the performance of the pre-trained SSC model. Our\ncode is available at https://github.com/blue-531/TALoS.\n", "date": "2024-10-21", "categories": []}
{"id": "2411.13251", "title": "BelHouse3D: A Benchmark Dataset for Assessing Occlusion Robustness in 3D\n  Point Cloud Semantic Segmentation", "abstract": "  Large-scale 2D datasets have been instrumental in advancing machine learning;\nhowever, progress in 3D vision tasks has been relatively slow. This disparity\nis largely due to the limited availability of 3D benchmarking datasets. In\nparticular, creating real-world point cloud datasets for indoor scene semantic\nsegmentation presents considerable challenges, including data collection within\nconfined spaces and the costly, often inaccurate process of per-point labeling\nto generate ground truths. While synthetic datasets address some of these\nchallenges, they often fail to replicate real-world conditions, particularly\nthe occlusions that occur in point clouds collected from real environments.\nExisting 3D benchmarking datasets typically evaluate deep learning models under\nthe assumption that training and test data are independently and identically\ndistributed (IID), which affects the models' usability for real-world point\ncloud segmentation. To address these challenges, we introduce the BelHouse3D\ndataset, a new synthetic point cloud dataset designed for 3D indoor scene\nsemantic segmentation. This dataset is constructed using real-world references\nfrom 32 houses in Belgium, ensuring that the synthetic data closely aligns with\nreal-world conditions. Additionally, we include a test set with data occlusion\nto simulate out-of-distribution (OOD) scenarios, reflecting the occlusions\ncommonly encountered in real-world point clouds. We evaluate popular\npoint-based semantic segmentation methods using our OOD setting and present a\nbenchmark. We believe that BelHouse3D and its OOD setting will advance research\nin 3D point cloud semantic segmentation for indoor scenes, providing valuable\ninsights for the development of more generalizable models.\n", "date": "2024-11-20", "categories": []}
{"id": "2210.05663", "title": "CLIP-Fields: Weakly Supervised Semantic Fields for Robotic Memory", "abstract": "  We propose CLIP-Fields, an implicit scene model that can be used for a\nvariety of tasks, such as segmentation, instance identification, semantic\nsearch over space, and view localization. CLIP-Fields learns a mapping from\nspatial locations to semantic embedding vectors. Importantly, we show that this\nmapping can be trained with supervision coming only from web-image and web-text\ntrained models such as CLIP, Detic, and Sentence-BERT; and thus uses no direct\nhuman supervision. When compared to baselines like Mask-RCNN, our method\noutperforms on few-shot instance identification or semantic segmentation on the\nHM3D dataset with only a fraction of the examples. Finally, we show that using\nCLIP-Fields as a scene memory, robots can perform semantic navigation in\nreal-world environments. Our code and demonstration videos are available here:\nhttps://mahis.life/clip-fields\n", "date": "2022-10-11", "categories": []}
{"id": "2405.10370", "title": "Grounded 3D-LLM with Referent Tokens", "abstract": "  Prior studies on 3D scene understanding have primarily developed specialized\nmodels for specific tasks or required task-specific fine-tuning. In this study,\nwe propose Grounded 3D-LLM, which explores the potential of 3D large\nmulti-modal models (3D LMMs) to consolidate various 3D vision tasks within a\nunified generative framework. The model uses scene referent tokens as special\nnoun phrases to reference 3D scenes, enabling it to handle sequences that\ninterleave 3D and textual data. Per-task instruction-following templates are\nemployed to ensure natural and diversity in translating 3D vision tasks into\nlanguage formats. To facilitate the use of referent tokens in subsequent\nlanguage modeling, we provide a large-scale, automatically curated grounded\nscene-text dataset with over 1 million phrase-to-region correspondences and\nintroduce Contrastive Language-Scene Pre-training (CLASP) to perform\nphrase-level scene-text alignment using this data. Our comprehensive evaluation\ncovers open-ended tasks like dense captioning and 3D question answering,\nalongside close-ended tasks such as object detection and language grounding.\nExperiments across multiple 3D benchmarks reveal the leading performance and\nthe broad applicability of Grounded 3D-LLM. Code and datasets are available at\nthe https://groundedscenellm.github.io/grounded_3d-llm.github.io.\n", "date": "2024-05-16", "categories": []}
{"id": "2409.02389", "title": "Multi-modal Situated Reasoning in 3D Scenes", "abstract": "  Situation awareness is essential for understanding and reasoning about 3D\nscenes in embodied AI agents. However, existing datasets and benchmarks for\nsituated understanding are limited in data modality, diversity, scale, and task\nscope. To address these limitations, we propose Multi-modal Situated Question\nAnswering (MSQA), a large-scale multi-modal situated reasoning dataset,\nscalably collected leveraging 3D scene graphs and vision-language models (VLMs)\nacross a diverse range of real-world 3D scenes. MSQA includes 251K situated\nquestion-answering pairs across 9 distinct question categories, covering\ncomplex scenarios within 3D scenes. We introduce a novel interleaved\nmulti-modal input setting in our benchmark to provide text, image, and point\ncloud for situation and question description, resolving ambiguity in previous\nsingle-modality convention (e.g., text). Additionally, we devise the\nMulti-modal Situated Next-step Navigation (MSNN) benchmark to evaluate models'\nsituated reasoning for navigation. Comprehensive evaluations on MSQA and MSNN\nhighlight the limitations of existing vision-language models and underscore the\nimportance of handling multi-modal interleaved inputs and situation modeling.\nExperiments on data scaling and cross-domain transfer further demonstrate the\nefficacy of leveraging MSQA as a pre-training dataset for developing more\npowerful situated reasoning models.\n", "date": "2024-09-04", "categories": []}
{"id": "2411.07555", "title": "GaussianCut: Interactive segmentation via graph cut for 3D Gaussian\n  Splatting", "abstract": "  We introduce GaussianCut, a new method for interactive multiview segmentation\nof scenes represented as 3D Gaussians. Our approach allows for selecting the\nobjects to be segmented by interacting with a single view. It accepts intuitive\nuser input, such as point clicks, coarse scribbles, or text. Using 3D Gaussian\nSplatting (3DGS) as the underlying scene representation simplifies the\nextraction of objects of interest which are considered to be a subset of the\nscene's Gaussians. Our key idea is to represent the scene as a graph and use\nthe graph-cut algorithm to minimize an energy function to effectively partition\nthe Gaussians into foreground and background. To achieve this, we construct a\ngraph based on scene Gaussians and devise a segmentation-aligned energy\nfunction on the graph to combine user inputs with scene properties. To obtain\nan initial coarse segmentation, we leverage 2D image/video segmentation models\nand further refine these coarse estimates using our graph construction. Our\nempirical evaluations show the adaptability of GaussianCut across a diverse set\nof scenes. GaussianCut achieves competitive performance with state-of-the-art\napproaches for 3D segmentation without requiring any additional\nsegmentation-aware training.\n", "date": "2024-11-12", "categories": []}
{"id": "2411.01781", "title": "MSTA3D: Multi-scale Twin-attention for 3D Instance Segmentation", "abstract": "  Recently, transformer-based techniques incorporating superpoints have become\nprevalent in 3D instance segmentation. However, they often encounter an\nover-segmentation problem, especially noticeable with large objects.\nAdditionally, unreliable mask predictions stemming from superpoint mask\nprediction further compound this issue. To address these challenges, we propose\na novel framework called MSTA3D. It leverages multi-scale feature\nrepresentation and introduces a twin-attention mechanism to effectively capture\nthem. Furthermore, MSTA3D integrates a box query with a box regularizer,\noffering a complementary spatial constraint alongside semantic queries.\nExperimental evaluations on ScanNetV2, ScanNet200 and S3DIS datasets\ndemonstrate that our approach surpasses state-of-the-art 3D instance\nsegmentation methods.\n", "date": "2024-11-04", "categories": []}
{"id": "2411.04963", "title": "VAIR: Visuo-Acoustic Implicit Representations for Low-Cost, Multi-Modal\n  Transparent Surface Reconstruction in Indoor Scenes", "abstract": "  Mobile robots operating indoors must be prepared to navigate challenging\nscenes that contain transparent surfaces. This paper proposes a novel method\nfor the fusion of acoustic and visual sensing modalities through implicit\nneural representations to enable dense reconstruction of transparent surfaces\nin indoor scenes. We propose a novel model that leverages generative latent\noptimization to learn an implicit representation of indoor scenes consisting of\ntransparent surfaces. We demonstrate that we can query the implicit\nrepresentation to enable volumetric rendering in image space or 3D geometry\nreconstruction (point clouds or mesh) with transparent surface prediction. We\nevaluate our method's effectiveness qualitatively and quantitatively on a new\ndataset collected using a custom, low-cost sensing platform featuring RGB-D\ncameras and ultrasonic sensors. Our method exhibits significant improvement\nover state-of-the-art for transparent surface reconstruction.\n", "date": "2024-11-07", "categories": []}
{"id": "2407.02077", "title": "Hierarchical Temporal Context Learning for Camera-based Semantic Scene\n  Completion", "abstract": "  Camera-based 3D semantic scene completion (SSC) is pivotal for predicting\ncomplicated 3D layouts with limited 2D image observations. The existing\nmainstream solutions generally leverage temporal information by roughly\nstacking history frames to supplement the current frame, such straightforward\ntemporal modeling inevitably diminishes valid clues and increases learning\ndifficulty. To address this problem, we present HTCL, a novel Hierarchical\nTemporal Context Learning paradigm for improving camera-based semantic scene\ncompletion. The primary innovation of this work involves decomposing temporal\ncontext learning into two hierarchical steps: (a) cross-frame affinity\nmeasurement and (b) affinity-based dynamic refinement. Firstly, to separate\ncritical relevant context from redundant information, we introduce the pattern\naffinity with scale-aware isolation and multiple independent learners for\nfine-grained contextual correspondence modeling. Subsequently, to dynamically\ncompensate for incomplete observations, we adaptively refine the feature\nsampling locations based on initially identified locations with high affinity\nand their neighboring relevant regions. Our method ranks $1^{st}$ on the\nSemanticKITTI benchmark and even surpasses LiDAR-based methods in terms of mIoU\non the OpenOccupancy benchmark. Our code is available on\nhttps://github.com/Arlo0o/HTCL.\n", "date": "2024-07-02", "categories": []}
{"id": "2411.03819", "title": "SA3DIP: Segment Any 3D Instance with Potential 3D Priors", "abstract": "  The proliferation of 2D foundation models has sparked research into adapting\nthem for open-world 3D instance segmentation. Recent methods introduce a\nparadigm that leverages superpoints as geometric primitives and incorporates 2D\nmulti-view masks from Segment Anything model (SAM) as merging guidance,\nachieving outstanding zero-shot instance segmentation results. However, the\nlimited use of 3D priors restricts the segmentation performance. Previous\nmethods calculate the 3D superpoints solely based on estimated normal from\nspatial coordinates, resulting in under-segmentation for instances with similar\ngeometry. Besides, the heavy reliance on SAM and hand-crafted algorithms in 2D\nspace suffers from over-segmentation due to SAM's inherent part-level\nsegmentation tendency. To address these issues, we propose SA3DIP, a novel\nmethod for Segmenting Any 3D Instances via exploiting potential 3D Priors.\nSpecifically, on one hand, we generate complementary 3D primitives based on\nboth geometric and textural priors, which reduces the initial errors that\naccumulate in subsequent procedures. On the other hand, we introduce\nsupplemental constraints from the 3D space by using a 3D detector to guide a\nfurther merging process. Furthermore, we notice a considerable portion of\nlow-quality ground truth annotations in ScanNetV2 benchmark, which affect the\nfair evaluations. Thus, we present ScanNetV2-INS with complete ground truth\nlabels and supplement additional instances for 3D class-agnostic instance\nsegmentation. Experimental evaluations on various 2D-3D datasets demonstrate\nthe effectiveness and robustness of our approach. Our code and proposed\nScanNetV2-INS dataset are available HERE.\n", "date": "2024-11-06", "categories": []}
{"id": "2411.02547", "title": "Modeling Uncertainty in 3D Gaussian Splatting through Continuous\n  Semantic Splatting", "abstract": "  In this paper, we present a novel algorithm for probabilistically updating\nand rasterizing semantic maps within 3D Gaussian Splatting (3D-GS). Although\nprevious methods have introduced algorithms which learn to rasterize features\nin 3D-GS for enhanced scene understanding, 3D-GS can fail without warning which\npresents a challenge for safety-critical robotic applications. To address this\ngap, we propose a method which advances the literature of continuous semantic\nmapping from voxels to ellipsoids, combining the precise structure of 3D-GS\nwith the ability to quantify uncertainty of probabilistic robotic maps. Given a\nset of images, our algorithm performs a probabilistic semantic update directly\non the 3D ellipsoids to obtain an expectation and variance through the use of\nconjugate priors. We also propose a probabilistic rasterization which returns\nper-pixel segmentation predictions with quantifiable uncertainty. We compare\nour method with similar probabilistic voxel-based methods to verify our\nextension to 3D ellipsoids, and perform ablation studies on uncertainty\nquantification and temporal smoothing.\n", "date": "2024-11-04", "categories": []}
{"id": "2406.12095", "title": "DistillNeRF: Perceiving 3D Scenes from Single-Glance Images by\n  Distilling Neural Fields and Foundation Model Features", "abstract": "  We propose DistillNeRF, a self-supervised learning framework addressing the\nchallenge of understanding 3D environments from limited 2D observations in\noutdoor autonomous driving scenes. Our method is a generalizable feedforward\nmodel that predicts a rich neural scene representation from sparse,\nsingle-frame multi-view camera inputs with limited view overlap, and is trained\nself-supervised with differentiable rendering to reconstruct RGB, depth, or\nfeature images. Our first insight is to exploit per-scene optimized Neural\nRadiance Fields (NeRFs) by generating dense depth and virtual camera targets\nfrom them, which helps our model to learn enhanced 3D geometry from sparse\nnon-overlapping image inputs. Second, to learn a semantically rich 3D\nrepresentation, we propose distilling features from pre-trained 2D foundation\nmodels, such as CLIP or DINOv2, thereby enabling various downstream tasks\nwithout the need for costly 3D human annotations. To leverage these two\ninsights, we introduce a novel model architecture with a two-stage\nlift-splat-shoot encoder and a parameterized sparse hierarchical voxel\nrepresentation. Experimental results on the NuScenes and Waymo NOTR datasets\ndemonstrate that DistillNeRF significantly outperforms existing comparable\nstate-of-the-art self-supervised methods for scene reconstruction, novel view\nsynthesis, and depth estimation; and it allows for competitive zero-shot 3D\nsemantic occupancy prediction, as well as open-world scene understanding\nthrough distilled foundation model features. Demos and code will be available\nat https://distillnerf.github.io/.\n", "date": "2024-06-17", "categories": []}
{"id": "2410.18956", "title": "Large Spatial Model: End-to-end Unposed Images to Semantic 3D", "abstract": "  Reconstructing and understanding 3D structures from a limited number of\nimages is a well-established problem in computer vision. Traditional methods\nusually break this task into multiple subtasks, each requiring complex\ntransformations between different data representations. For instance, dense\nreconstruction through Structure-from-Motion (SfM) involves converting images\ninto key points, optimizing camera parameters, and estimating structures.\nAfterward, accurate sparse reconstructions are required for further dense\nmodeling, which is subsequently fed into task-specific neural networks. This\nmulti-step process results in considerable processing time and increased\nengineering complexity.\n  In this work, we present the Large Spatial Model (LSM), which processes\nunposed RGB images directly into semantic radiance fields. LSM simultaneously\nestimates geometry, appearance, and semantics in a single feed-forward\noperation, and it can generate versatile label maps by interacting with\nlanguage at novel viewpoints. Leveraging a Transformer-based architecture, LSM\nintegrates global geometry through pixel-aligned point maps. To enhance spatial\nattribute regression, we incorporate local context aggregation with multi-scale\nfusion, improving the accuracy of fine local details. To tackle the scarcity of\nlabeled 3D semantic data and enable natural language-driven scene manipulation,\nwe incorporate a pre-trained 2D language-based segmentation model into a\n3D-consistent semantic feature field. An efficient decoder then parameterizes a\nset of semantic anisotropic Gaussians, facilitating supervised end-to-end\nlearning. Extensive experiments across various tasks show that LSM unifies\nmultiple 3D vision tasks directly from unposed images, achieving real-time\nsemantic 3D reconstruction for the first time.\n", "date": "2024-10-24", "categories": []}
{"id": "2410.22626", "title": "Symbolic Graph Inference for Compound Scene Understanding", "abstract": "  Scene understanding is a fundamental capability needed in many domains,\nranging from question-answering to robotics. Unlike recent end-to-end\napproaches that must explicitly learn varying compositions of the same scene,\nour method reasons over their constituent objects and analyzes their\narrangement to infer a scene's meaning. We propose a novel approach that\nreasons over a scene's scene- and knowledge-graph, capturing spatial\ninformation while being able to utilize general domain knowledge in a joint\ngraph search. Empirically, we demonstrate the feasibility of our method on the\nADE20K dataset and compare it to current scene understanding approaches.\n", "date": "2024-10-30", "categories": []}
{"id": "2406.10527", "title": "Panoptic-FlashOcc: An Efficient Baseline to Marry Semantic Occupancy\n  with Panoptic via Instance Center", "abstract": "  Panoptic occupancy poses a novel challenge by aiming to integrate instance\noccupancy and semantic occupancy within a unified framework. However, there is\nstill a lack of efficient solutions for panoptic occupancy. In this paper, we\npropose Panoptic-FlashOcc, a straightforward yet robust 2D feature framework\nthat enables realtime panoptic occupancy. Building upon the lightweight design\nof FlashOcc, our approach simultaneously learns semantic occupancy and\nclass-aware instance clustering in a single network, these outputs are jointly\nincorporated through panoptic occupancy procession for panoptic occupancy. This\napproach effectively addresses the drawbacks of high memory and computation\nrequirements associated with three-dimensional voxel-level representations.\nWith its straightforward and efficient design that facilitates easy deployment,\nPanoptic-FlashOcc demonstrates remarkable achievements in panoptic occupancy\nprediction. On the Occ3D-nuScenes benchmark, it achieves exceptional\nperformance, with 38.5 RayIoU and 29.1 mIoU for semantic occupancy, operating\nat a rapid speed of 43.9 FPS. Furthermore, it attains a notable score of 16.0\nRayPQ for panoptic occupancy, accompanied by a fast inference speed of 30.2\nFPS. These results surpass the performance of existing methodologies in terms\nof both speed and accuracy. The source code and trained models can be found at\nthe following github repository: https://github.com/Yzichen/FlashOCC.\n", "date": "2024-06-15", "categories": []}
{"id": "2405.13675", "title": "Context and Geometry Aware Voxel Transformer for Semantic Scene\n  Completion", "abstract": "  Vision-based Semantic Scene Completion (SSC) has gained much attention due to\nits widespread applications in various 3D perception tasks. Existing\nsparse-to-dense approaches typically employ shared context-independent queries\nacross various input images, which fails to capture distinctions among them as\nthe focal regions of different inputs vary and may result in undirected feature\naggregation of cross-attention. Additionally, the absence of depth information\nmay lead to points projected onto the image plane sharing the same 2D position\nor similar sampling points in the feature map, resulting in depth ambiguity. In\nthis paper, we present a novel context and geometry aware voxel transformer. It\nutilizes a context aware query generator to initialize context-dependent\nqueries tailored to individual input images, effectively capturing their unique\ncharacteristics and aggregating information within the region of interest.\nFurthermore, it extend deformable cross-attention from 2D to 3D pixel space,\nenabling the differentiation of points with similar image coordinates based on\ntheir depth coordinates. Building upon this module, we introduce a neural\nnetwork named CGFormer to achieve semantic scene completion. Simultaneously,\nCGFormer leverages multiple 3D representations (i.e., voxel and TPV) to boost\nthe semantic and geometric representation abilities of the transformed 3D\nvolume from both local and global perspectives. Experimental results\ndemonstrate that CGFormer achieves state-of-the-art performance on the\nSemanticKITTI and SSCBench-KITTI-360 benchmarks, attaining a mIoU of 16.87 and\n20.05, as well as an IoU of 45.99 and 48.07, respectively. Remarkably, CGFormer\neven outperforms approaches employing temporal images as inputs or much larger\nimage backbone networks.\n", "date": "2024-05-22", "categories": []}
{"id": "2404.07989", "title": "Any2Point: Empowering Any-modality Large Models for Efficient 3D\n  Understanding", "abstract": "  Large foundation models have recently emerged as a prominent focus of\ninterest, attaining superior performance in widespread scenarios. Due to the\nscarcity of 3D data, many efforts have been made to adapt pre-trained\ntransformers from vision to 3D domains. However, such 2D-to-3D approaches are\nstill limited, due to the potential loss of spatial geometries and high\ncomputation cost. More importantly, their frameworks are mainly designed for 2D\nmodels, lacking a general any-to-3D paradigm. In this paper, we introduce\nAny2Point, a parameter-efficient method to empower any-modality large models\n(vision, language, audio) for 3D understanding. Given a frozen transformer from\nany source modality, we propose a 3D-to-any (1D or 2D) virtual projection\nstrategy that correlates the input 3D points to the original 1D or 2D positions\nwithin the source modality. This mechanism enables us to assign each 3D token\nwith a positional encoding paired with the pre-trained model, which avoids 3D\ngeometry loss caused by the true projection and better motivates the\ntransformer for 3D learning with 1D/2D positional priors. Then, within each\ntransformer block, we insert an any-to-3D guided adapter module for\nparameter-efficient fine-tuning. The adapter incorporates prior spatial\nknowledge from the source modality to guide the local feature aggregation of 3D\ntokens, compelling the semantic adaption of any-modality transformers. We\nconduct extensive experiments to showcase the effectiveness and efficiency of\nour method. Code and models are released at\nhttps://github.com/Ivan-Tang-3D/Any2Point.\n", "date": "2024-04-11", "categories": []}
{"id": "2410.12158", "title": "SAM-Guided Masked Token Prediction for 3D Scene Understanding", "abstract": "  Foundation models have significantly enhanced 2D task performance, and recent\nworks like Bridge3D have successfully applied these models to improve 3D scene\nunderstanding through knowledge distillation, marking considerable\nadvancements. Nonetheless, challenges such as the misalignment between 2D and\n3D representations and the persistent long-tail distribution in 3D datasets\nstill restrict the effectiveness of knowledge distillation from 2D to 3D using\nfoundation models. To tackle these issues, we introduce a novel SAM-guided\ntokenization method that seamlessly aligns 3D transformer structures with\nregion-level knowledge distillation, replacing the traditional KNN-based\ntokenization techniques. Additionally, we implement a group-balanced\nre-weighting strategy to effectively address the long-tail problem in knowledge\ndistillation. Furthermore, inspired by the recent success of masked feature\nprediction, our framework incorporates a two-stage masked token prediction\nprocess in which the student model predicts both the global embeddings and the\ntoken-wise local embeddings derived from the teacher models trained in the\nfirst stage. Our methodology has been validated across multiple datasets,\nincluding SUN RGB-D, ScanNet, and S3DIS, for tasks like 3D object detection and\nsemantic segmentation. The results demonstrate significant improvements over\ncurrent State-of-the-art self-supervised methods, establishing new benchmarks\nin this field.\n", "date": "2024-10-16", "categories": []}
{"id": "2410.13860", "title": "VLM-Grounder: A VLM Agent for Zero-Shot 3D Visual Grounding", "abstract": "  3D visual grounding is crucial for robots, requiring integration of natural\nlanguage and 3D scene understanding. Traditional methods depending on\nsupervised learning with 3D point clouds are limited by scarce datasets.\nRecently zero-shot methods leveraging LLMs have been proposed to address the\ndata issue. While effective, these methods only use object-centric information,\nlimiting their ability to handle complex queries. In this work, we present\nVLM-Grounder, a novel framework using vision-language models (VLMs) for\nzero-shot 3D visual grounding based solely on 2D images. VLM-Grounder\ndynamically stitches image sequences, employs a grounding and feedback scheme\nto find the target object, and uses a multi-view ensemble projection to\naccurately estimate 3D bounding boxes. Experiments on ScanRefer and Nr3D\ndatasets show VLM-Grounder outperforms previous zero-shot methods, achieving\n51.6% Acc@0.25 on ScanRefer and 48.0% Acc on Nr3D, without relying on 3D\ngeometry or object priors. Codes are available at\nhttps://github.com/OpenRobotLab/VLM-Grounder .\n", "date": "2024-10-17", "categories": []}
{"id": "2410.11228", "title": "TEOcc: Radar-camera Multi-modal Occupancy Prediction via Temporal\n  Enhancement", "abstract": "  As a novel 3D scene representation, semantic occupancy has gained much\nattention in autonomous driving. However, existing occupancy prediction methods\nmainly focus on designing better occupancy representations, such as\ntri-perspective view or neural radiance fields, while ignoring the advantages\nof using long-temporal information. In this paper, we propose a radar-camera\nmulti-modal temporal enhanced occupancy prediction network, dubbed TEOcc. Our\nmethod is inspired by the success of utilizing temporal information in 3D\nobject detection. Specifically, we introduce a temporal enhancement branch to\nlearn temporal occupancy prediction. In this branch, we randomly discard the\nt-k input frame of the multi-view camera and predict its 3D occupancy by\nlong-term and short-term temporal decoders separately with the information from\nother adjacent frames and multi-modal inputs. Besides, to reduce computational\ncosts and incorporate multi-modal inputs, we specially designed 3D\nconvolutional layers for long-term and short-term temporal decoders.\nFurthermore, since the lightweight occupancy prediction head is a dense\nclassification head, we propose to use a shared occupancy prediction head for\nthe temporal enhancement and main branches. It is worth noting that the\ntemporal enhancement branch is only performed during training and is discarded\nduring inference. Experiment results demonstrate that TEOcc achieves\nstate-of-the-art occupancy prediction on nuScenes benchmarks. In addition, the\nproposed temporal enhancement branch is a plug-and-play module that can be\neasily integrated into existing occupancy prediction methods to improve the\nperformance of occupancy prediction. The code and models will be released at\nhttps://github.com/VDIGPKU/TEOcc.\n", "date": "2024-10-15", "categories": []}
{"id": "2410.10659", "title": "PCF-Lift: Panoptic Lifting by Probabilistic Contrastive Fusion", "abstract": "  Panoptic lifting is an effective technique to address the 3D panoptic\nsegmentation task by unprojecting 2D panoptic segmentations from multi-views to\n3D scene. However, the quality of its results largely depends on the 2D\nsegmentations, which could be noisy and error-prone, so its performance often\ndrops significantly for complex scenes. In this work, we design a new pipeline\ncoined PCF-Lift based on our Probabilis-tic Contrastive Fusion (PCF) to learn\nand embed probabilistic features throughout our pipeline to actively consider\ninaccurate segmentations and inconsistent instance IDs. Technical-wise, we\nfirst model the probabilistic feature embeddings through multivariate Gaussian\ndistributions. To fuse the probabilistic features, we incorporate the\nprobability product kernel into the contrastive loss formulation and design a\ncross-view constraint to enhance the feature consistency across different\nviews. For the inference, we introduce a new probabilistic clustering method to\neffectively associate prototype features with the underlying 3D object\ninstances for the generation of consistent panoptic segmentation results.\nFurther, we provide a theoretical analysis to justify the superiority of the\nproposed probabilistic solution. By conducting extensive experiments, our\nPCF-lift not only significantly outperforms the state-of-the-art methods on\nwidely used benchmarks including the ScanNet dataset and the challenging Messy\nRoom dataset (4.4% improvement of scene-level PQ), but also demonstrates strong\nrobustness when incorporating various 2D segmentation models or different\nlevels of hand-crafted noise.\n", "date": "2024-10-14", "categories": []}
{"id": "2410.08365", "title": "Are We Ready for Real-Time LiDAR Semantic Segmentation in Autonomous\n  Driving?", "abstract": "  Within a perception framework for autonomous mobile and robotic systems,\nsemantic analysis of 3D point clouds typically generated by LiDARs is key to\nnumerous applications, such as object detection and recognition, and scene\nreconstruction. Scene semantic segmentation can be achieved by directly\nintegrating 3D spatial data with specialized deep neural networks. Although\nthis type of data provides rich geometric information regarding the surrounding\nenvironment, it also presents numerous challenges: its unstructured and sparse\nnature, its unpredictable size, and its demanding computational requirements.\nThese characteristics hinder the real-time semantic analysis, particularly on\nresource-constrained hardware architectures that constitute the main\ncomputational components of numerous robotic applications. Therefore, in this\npaper, we investigate various 3D semantic segmentation methodologies and\nanalyze their performance and capabilities for resource-constrained inference\non embedded NVIDIA Jetson platforms. We evaluate them for a fair comparison\nthrough a standardized training protocol and data augmentations, providing\nbenchmark results on the Jetson AGX Orin and AGX Xavier series for two\nlarge-scale outdoor datasets: SemanticKITTI and nuScenes.\n", "date": "2024-10-10", "categories": []}
{"id": "2410.04529", "title": "In-Place Panoptic Radiance Field Segmentation with Perceptual Prior for\n  3D Scene Understanding", "abstract": "  Accurate 3D scene representation and panoptic understanding are essential for\napplications such as virtual reality, robotics, and autonomous driving.\nHowever, challenges persist with existing methods, including precise 2D-to-3D\nmapping, handling complex scene characteristics like boundary ambiguity and\nvarying scales, and mitigating noise in panoptic pseudo-labels. This paper\nintroduces a novel perceptual-prior-guided 3D scene representation and panoptic\nunderstanding method, which reformulates panoptic understanding within neural\nradiance fields as a linear assignment problem involving 2D semantics and\ninstance recognition. Perceptual information from pre-trained 2D panoptic\nsegmentation models is incorporated as prior guidance, thereby synchronizing\nthe learning processes of appearance, geometry, and panoptic understanding\nwithin neural radiance fields. An implicit scene representation and\nunderstanding model is developed to enhance generalization across indoor and\noutdoor scenes by extending the scale-encoded cascaded grids within a\nreparameterized domain distillation framework. This model effectively manages\ncomplex scene attributes and generates 3D-consistent scene representations and\npanoptic understanding outcomes for various scenes. Experiments and ablation\nstudies under challenging conditions, including synthetic and real-world\nscenes, demonstrate the proposed method's effectiveness in enhancing 3D scene\nrepresentation and panoptic segmentation accuracy.\n", "date": "2024-10-06", "categories": []}
{"id": "2410.04652", "title": "Multimodal 3D Fusion and In-Situ Learning for Spatially Aware AI", "abstract": "  Seamless integration of virtual and physical worlds in augmented reality\nbenefits from the system semantically \"understanding\" the physical environment.\nAR research has long focused on the potential of context awareness,\ndemonstrating novel capabilities that leverage the semantics in the 3D\nenvironment for various object-level interactions. Meanwhile, the computer\nvision community has made leaps in neural vision-language understanding to\nenhance environment perception for autonomous tasks. In this work, we introduce\na multimodal 3D object representation that unifies both semantic and linguistic\nknowledge with the geometric representation, enabling user-guided machine\nlearning involving physical objects. We first present a fast multimodal 3D\nreconstruction pipeline that brings linguistic understanding to AR by fusing\nCLIP vision-language features into the environment and object models. We then\npropose \"in-situ\" machine learning, which, in conjunction with the multimodal\nrepresentation, enables new tools and interfaces for users to interact with\nphysical spaces and objects in a spatially and linguistically meaningful\nmanner. We demonstrate the usefulness of the proposed system through two\nreal-world AR applications on Magic Leap 2: a) spatial search in physical\nenvironments with natural language and b) an intelligent inventory system that\ntracks object changes over time. We also make our full implementation and demo\ndata available at (https://github.com/cy-xu/spatially_aware_AI) to encourage\nfurther exploration and research in spatially aware AI.\n", "date": "2024-10-06", "categories": []}
{"id": "2410.02352", "title": "ProtoSeg: A Prototype-Based Point Cloud Instance Segmentation Method", "abstract": "  3D instance segmentation is crucial for obtaining an understanding of a point\ncloud scene. This paper presents a novel neural network architecture for\nperforming instance segmentation on 3D point clouds. We propose to jointly\nlearn coefficients and prototypes in parallel which can be combined to obtain\nthe instance predictions. The coefficients are computed using an overcomplete\nset of sampled points with a novel multi-scale module, dubbed dilated point\ninception. As the set of obtained instance mask predictions is overcomplete, we\nemploy a non-maximum suppression algorithm to retrieve the final predictions.\nThis approach allows to omit the time-expensive clustering step and leads to a\nmore stable inference time. The proposed method is not only 28% faster than the\nstate-of-the-art, it also exhibits the lowest standard deviation. Our\nexperiments have shown that the standard deviation of the inference time is\nonly 1.0% of the total time while it ranges between 10.8 and 53.1% for the\nstate-of-the-art methods. Lastly, our method outperforms the state-of-the-art\nboth on S3DIS-blocks (4.9% in mRec on Fold-5) and PartNet (2.0% on average in\nmAP).\n", "date": "2024-10-03", "categories": []}
{"id": "2306.09001", "title": "SSCBench: A Large-Scale 3D Semantic Scene Completion Benchmark for\n  Autonomous Driving", "abstract": "  Monocular scene understanding is a foundational component of autonomous\nsystems. Within the spectrum of monocular perception topics, one crucial and\nuseful task for holistic 3D scene understanding is semantic scene completion\n(SSC), which jointly completes semantic information and geometric details from\nRGB input. However, progress in SSC, particularly in large-scale street views,\nis hindered by the scarcity of high-quality datasets. To address this issue, we\nintroduce SSCBench, a comprehensive benchmark that integrates scenes from\nwidely used automotive datasets (e.g., KITTI-360, nuScenes, and Waymo).\nSSCBench follows an established setup and format in the community, facilitating\nthe easy exploration of SSC methods in various street views. We benchmark\nmodels using monocular, trinocular, and point cloud input to assess the\nperformance gap resulting from sensor coverage and modality. Moreover, we have\nunified semantic labels across diverse datasets to simplify cross-domain\ngeneralization testing. We commit to including more datasets and SSC models to\ndrive further advancements in this field.\n", "date": "2023-06-15", "categories": []}
{"id": "2312.05752", "title": "Camera-based 3D Semantic Scene Completion with Sparse Guidance Network", "abstract": "  Semantic scene completion (SSC) aims to predict the semantic occupancy of\neach voxel in the entire 3D scene from limited observations, which is an\nemerging and critical task for autonomous driving. Recently, many studies have\nturned to camera-based SSC solutions due to the richer visual cues and\ncost-effectiveness of cameras. However, existing methods usually rely on\nsophisticated and heavy 3D models to process the lifted 3D features directly,\nwhich are not discriminative enough for clear segmentation boundaries. In this\npaper, we adopt the dense-sparse-dense design and propose a one-stage\ncamera-based SSC framework, termed SGN, to propagate semantics from the\nsemantic-aware seed voxels to the whole scene based on spatial geometry cues.\nFirstly, to exploit depth-aware context and dynamically select sparse seed\nvoxels, we redesign the sparse voxel proposal network to process points\ngenerated by depth prediction directly with the coarse-to-fine paradigm.\nFurthermore, by designing hybrid guidance (sparse semantic and geometry\nguidance) and effective voxel aggregation for spatial geometry cues, we enhance\nthe feature separation between different categories and expedite the\nconvergence of semantic propagation. Finally, we devise the multi-scale\nsemantic propagation module for flexible receptive fields while reducing the\ncomputation resources. Extensive experimental results on the SemanticKITTI and\nSSCBench-KITTI-360 datasets demonstrate the superiority of our SGN over\nexisting state-of-the-art methods. And even our lightweight version SGN-L\nachieves notable scores of 14.80\\% mIoU and 45.45\\% IoU on SeamnticKITTI\nvalidation with only 12.5 M parameters and 7.16 G training memory. Code is\navailable at https://github.com/Jieqianyu/SGN.\n", "date": "2023-12-10", "categories": []}
{"id": "2312.08168", "title": "Chat-Scene: Bridging 3D Scene and Large Language Models with Object\n  Identifiers", "abstract": "  Recent advancements in 3D Large Language Models (LLMs) have demonstrated\npromising capabilities for 3D scene understanding. However, previous methods\nexhibit deficiencies in general referencing and grounding capabilities for\nintricate scene comprehension. In this paper, we introduce the use of object\nidentifiers and object-centric representations to interact with scenes at the\nobject level. Specifically, we decompose the input 3D scene into a set of\nobject proposals, each assigned a unique identifier token, which enables\nefficient object referencing and grounding during user-assistant interactions.\nGiven the scarcity of scene-language data, we model the scene embeddings as a\nsequence of explicit object-level embeddings, derived from semantic-rich 2D or\n3D representations. By employing object identifiers, we transform diverse 3D\nscene-language tasks into a unified question-answering format, facilitating\njoint training without the need for additional task-specific heads. With\nminimal fine-tuning on all downstream tasks, our model significantly\noutperforms existing methods on benchmarks including ScanRefer, Multi3DRefer,\nScan2Cap, ScanQA, and SQA3D.\n", "date": "2023-12-13", "categories": []}
{"id": "2403.17846", "title": "Hierarchical Open-Vocabulary 3D Scene Graphs for Language-Grounded Robot\n  Navigation", "abstract": "  Recent open-vocabulary robot mapping methods enrich dense geometric maps with\npre-trained visual-language features. While these maps allow for the prediction\nof point-wise saliency maps when queried for a certain language concept,\nlarge-scale environments and abstract queries beyond the object level still\npose a considerable hurdle, ultimately limiting language-grounded robotic\nnavigation. In this work, we present HOV-SG, a hierarchical open-vocabulary 3D\nscene graph mapping approach for language-grounded robot navigation. Leveraging\nopen-vocabulary vision foundation models, we first obtain state-of-the-art\nopen-vocabulary segment-level maps in 3D and subsequently construct a 3D scene\ngraph hierarchy consisting of floor, room, and object concepts, each enriched\nwith open-vocabulary features. Our approach is able to represent multi-story\nbuildings and allows robotic traversal of those using a cross-floor Voronoi\ngraph. HOV-SG is evaluated on three distinct datasets and surpasses previous\nbaselines in open-vocabulary semantic accuracy on the object, room, and floor\nlevel while producing a 75% reduction in representation size compared to dense\nopen-vocabulary maps. In order to prove the efficacy and generalization\ncapabilities of HOV-SG, we showcase successful long-horizon\nlanguage-conditioned robot navigation within real-world multi-storage\nenvironments. We provide code and trial video data at http://hovsg.github.io/.\n", "date": "2024-03-26", "categories": []}
{"id": "2404.16538", "title": "OpenDlign: Open-World Point Cloud Understanding with Depth-Aligned\n  Images", "abstract": "  Recent open-world 3D representation learning methods using Vision-Language\nModels (VLMs) to align 3D point cloud with image-text information have shown\nsuperior 3D zero-shot performance. However, CAD-rendered images for this\nalignment often lack realism and texture variation, compromising alignment\nrobustness. Moreover, the volume discrepancy between 3D and 2D pretraining\ndatasets highlights the need for effective strategies to transfer the\nrepresentational abilities of VLMs to 3D learning. In this paper, we present\nOpenDlign, a novel open-world 3D model using depth-aligned images generated\nfrom a diffusion model for robust multimodal alignment. These images exhibit\ngreater texture diversity than CAD renderings due to the stochastic nature of\nthe diffusion model. By refining the depth map projection pipeline and\ndesigning depth-specific prompts, OpenDlign leverages rich knowledge in\npre-trained VLM for 3D representation learning with streamlined fine-tuning.\nOur experiments show that OpenDlign achieves high zero-shot and few-shot\nperformance on diverse 3D tasks, despite only fine-tuning 6 million parameters\non a limited ShapeNet dataset. In zero-shot classification, OpenDlign surpasses\nprevious models by 8.0% on ModelNet40 and 16.4% on OmniObject3D. Additionally,\nusing depth-aligned images for multimodal alignment consistently enhances the\nperformance of other state-of-the-art models.\n", "date": "2024-04-25", "categories": []}
{"id": "2409.19650", "title": "Grounding 3D Scene Affordance From Egocentric Interactions", "abstract": "  Grounding 3D scene affordance aims to locate interactive regions in 3D\nenvironments, which is crucial for embodied agents to interact intelligently\nwith their surroundings. Most existing approaches achieve this by mapping\nsemantics to 3D instances based on static geometric structure and visual\nappearance. This passive strategy limits the agent's ability to actively\nperceive and engage with the environment, making it reliant on predefined\nsemantic instructions. In contrast, humans develop complex interaction skills\nby observing and imitating how others interact with their surroundings. To\nempower the model with such abilities, we introduce a novel task: grounding 3D\nscene affordance from egocentric interactions, where the goal is to identify\nthe corresponding affordance regions in a 3D scene based on an egocentric video\nof an interaction. This task faces the challenges of spatial complexity and\nalignment complexity across multiple sources. To address these challenges, we\npropose the Egocentric Interaction-driven 3D Scene Affordance Grounding\n(Ego-SAG) framework, which utilizes interaction intent to guide the model in\nfocusing on interaction-relevant sub-regions and aligns affordance features\nfrom different sources through a bidirectional query decoder mechanism.\nFurthermore, we introduce the Egocentric Video-3D Scene Affordance Dataset\n(VSAD), covering a wide range of common interaction types and diverse 3D\nenvironments to support this task. Extensive experiments on VSAD validate both\nthe feasibility of the proposed task and the effectiveness of our approach.\n", "date": "2024-09-29", "categories": []}
{"id": "2409.13430", "title": "CVT-Occ: Cost Volume Temporal Fusion for 3D Occupancy Prediction", "abstract": "  Vision-based 3D occupancy prediction is significantly challenged by the\ninherent limitations of monocular vision in depth estimation. This paper\nintroduces CVT-Occ, a novel approach that leverages temporal fusion through the\ngeometric correspondence of voxels over time to improve the accuracy of 3D\noccupancy predictions. By sampling points along the line of sight of each voxel\nand integrating the features of these points from historical frames, we\nconstruct a cost volume feature map that refines current volume features for\nimproved prediction outcomes. Our method takes advantage of parallax cues from\nhistorical observations and employs a data-driven approach to learn the cost\nvolume. We validate the effectiveness of CVT-Occ through rigorous experiments\non the Occ3D-Waymo dataset, where it outperforms state-of-the-art methods in 3D\noccupancy prediction with minimal additional computational cost. The code is\nreleased at \\url{https://github.com/Tsinghua-MARS-Lab/CVT-Occ}.\n", "date": "2024-09-20", "categories": []}
{"id": "2402.04648", "title": "OV-NeRF: Open-vocabulary Neural Radiance Fields with Vision and Language\n  Foundation Models for 3D Semantic Understanding", "abstract": "  The development of Neural Radiance Fields (NeRFs) has provided a potent\nrepresentation for encapsulating the geometric and appearance characteristics\nof 3D scenes. Enhancing the capabilities of NeRFs in open-vocabulary 3D\nsemantic perception tasks has been a recent focus. However, current methods\nthat extract semantics directly from Contrastive Language-Image Pretraining\n(CLIP) for semantic field learning encounter difficulties due to noisy and\nview-inconsistent semantics provided by CLIP. To tackle these limitations, we\npropose OV-NeRF, which exploits the potential of pre-trained vision and\nlanguage foundation models to enhance semantic field learning through proposed\nsingle-view and cross-view strategies. First, from the single-view perspective,\nwe introduce Region Semantic Ranking (RSR) regularization by leveraging 2D mask\nproposals derived from Segment Anything (SAM) to rectify the noisy semantics of\neach training view, facilitating accurate semantic field learning. Second, from\nthe cross-view perspective, we propose a Cross-view Self-enhancement (CSE)\nstrategy to address the challenge raised by view-inconsistent semantics. Rather\nthan invariably utilizing the 2D inconsistent semantics from CLIP, CSE\nleverages the 3D consistent semantics generated from the well-trained semantic\nfield itself for semantic field training, aiming to reduce ambiguity and\nenhance overall semantic consistency across different views. Extensive\nexperiments validate our OV-NeRF outperforms current state-of-the-art methods,\nachieving a significant improvement of 20.31% and 18.42% in mIoU metric on\nReplica and ScanNet, respectively. Furthermore, our approach exhibits\nconsistent superior results across various CLIP configurations, further\nverifying its robustness. Project page: https://github.com/pcl3dv/OV-NeRF.\n", "date": "2024-02-07", "categories": []}
{"id": "2409.01807", "title": "EPRecon: An Efficient Framework for Real-Time Panoptic 3D Reconstruction\n  from Monocular Video", "abstract": "  Panoptic 3D reconstruction from a monocular video is a fundamental perceptual\ntask in robotic scene understanding. However, existing efforts suffer from\ninefficiency in terms of inference speed and accuracy, limiting their practical\napplicability. We present EPRecon, an efficient real-time panoptic 3D\nreconstruction framework. Current volumetric-based reconstruction methods\nusually utilize multi-view depth map fusion to obtain scene depth priors, which\nis time-consuming and poses challenges to real-time scene reconstruction. To\naddress this issue, we propose a lightweight module to directly estimate scene\ndepth priors in a 3D volume for reconstruction quality improvement by\ngenerating occupancy probabilities of all voxels. In addition, compared with\nexisting panoptic segmentation methods, EPRecon extracts panoptic features from\nboth voxel features and corresponding image features, obtaining more detailed\nand comprehensive instance-level semantic information and achieving more\naccurate segmentation results. Experimental results on the ScanNetV2 dataset\ndemonstrate the superiority of EPRecon over current state-of-the-art methods in\nterms of both panoptic 3D reconstruction quality and real-time inference. Code\nis available at https://github.com/zhen6618/EPRecon.\n", "date": "2024-09-03", "categories": []}
{"id": "2409.11681", "title": "Gradient-Driven 3D Segmentation and Affordance Transfer in Gaussian\n  Splatting Using 2D Masks", "abstract": "  3D Gaussian Splatting has emerged as a powerful 3D scene representation\ntechnique, capturing fine details with high efficiency. In this paper, we\nintroduce a novel voting-based method that extends 2D segmentation models to 3D\nGaussian splats. Our approach leverages masked gradients, where gradients are\nfiltered by input 2D masks, and these gradients are used as votes to achieve\naccurate segmentation. As a byproduct, we discovered that inference-time\ngradients can also be used to prune Gaussians, resulting in up to 21%\ncompression. Additionally, we explore few-shot affordance transfer, allowing\nannotations from 2D images to be effectively transferred onto 3D Gaussian\nsplats. The robust yet straightforward mathematical formulation underlying this\napproach makes it a highly effective tool for numerous downstream applications,\nsuch as augmented reality (AR), object editing, and robotics. The project code\nand additional resources are available at\nhttps://jojijoseph.github.io/3dgs-segmentation.\n", "date": "2024-09-18", "categories": []}
{"id": "2406.07037", "title": "PanoSSC: Exploring Monocular Panoptic 3D Scene Reconstruction for\n  Autonomous Driving", "abstract": "  Vision-centric occupancy networks, which represent the surrounding\nenvironment with uniform voxels with semantics, have become a new trend for\nsafe driving of camera-only autonomous driving perception systems, as they are\nable to detect obstacles regardless of their shape and occlusion. Modern\noccupancy networks mainly focus on reconstructing visible voxels from object\nsurfaces with voxel-wise semantic prediction. Usually, they suffer from\ninconsistent predictions of one object and mixed predictions for adjacent\nobjects. These confusions may harm the safety of downstream planning modules.\nTo this end, we investigate panoptic segmentation on 3D voxel scenarios and\npropose an instance-aware occupancy network, PanoSSC. We predict foreground\nobjects and backgrounds separately and merge both in post-processing. For\nforeground instance grouping, we propose a novel 3D instance mask decoder that\ncan efficiently extract individual objects. we unify geometric reconstruction,\n3D semantic segmentation, and 3D instance segmentation into PanoSSC framework\nand propose new metrics for evaluating panoptic voxels. Extensive experiments\nshow that our method achieves competitive results on SemanticKITTI semantic\nscene completion benchmark.\n", "date": "2024-06-11", "categories": []}
{"id": "2409.12008", "title": "Panoptic-Depth Forecasting", "abstract": "  Forecasting the semantics and 3D structure of scenes is essential for robots\nto navigate and plan actions safely. Recent methods have explored semantic and\npanoptic scene forecasting; however, they do not consider the geometry of the\nscene. In this work, we propose the panoptic-depth forecasting task for jointly\npredicting the panoptic segmentation and depth maps of unobserved future\nframes, from monocular camera images. To facilitate this work, we extend the\npopular KITTI-360 and Cityscapes benchmarks by computing depth maps from LiDAR\npoint clouds and leveraging sequential labeled data. We also introduce a\nsuitable evaluation metric that quantifies both the panoptic quality and depth\nestimation accuracy of forecasts in a coherent manner. Furthermore, we present\ntwo baselines and propose the novel PDcast architecture that learns rich\nspatio-temporal representations by incorporating a transformer-based encoder, a\nforecasting module, and task-specific decoders to predict future panoptic-depth\noutputs. Extensive evaluations demonstrate the effectiveness of PDcast across\ntwo datasets and three forecasting tasks, consistently addressing the primary\nchallenges. We make the code publicly available at\nhttps://pdcast.cs.uni-freiburg.de.\n", "date": "2024-09-18", "categories": []}
{"id": "2407.14279", "title": "OpenSU3D: Open World 3D Scene Understanding using Foundation Models", "abstract": "  In this paper, we present a novel, scalable approach for constructing open\nset, instance-level 3D scene representations, advancing open world\nunderstanding of 3D environments. Existing methods require pre-constructed 3D\nscenes and face scalability issues due to per-point feature vector learning,\nlimiting their efficacy with complex queries. Our method overcomes these\nlimitations by incrementally building instance-level 3D scene representations\nusing 2D foundation models, efficiently aggregating instance-level details such\nas masks, feature vectors, names, and captions. We introduce fusion schemes for\nfeature vectors to enhance their contextual knowledge and performance on\ncomplex queries. Additionally, we explore large language models for robust\nautomatic annotation and spatial reasoning tasks. We evaluate our proposed\napproach on multiple scenes from ScanNet and Replica datasets demonstrating\nzero-shot generalization capabilities, exceeding current state-of-the-art\nmethods in open world 3D scene understanding.\n", "date": "2024-07-19", "categories": []}
{"id": "2409.10350", "title": "Point2Graph: An End-to-end Point Cloud-based 3D Open-Vocabulary Scene\n  Graph for Robot Navigation", "abstract": "  Current open-vocabulary scene graph generation algorithms highly rely on both\n3D scene point cloud data and posed RGB-D images and thus have limited\napplications in scenarios where RGB-D images or camera poses are not readily\navailable. To solve this problem, we propose Point2Graph, a novel end-to-end\npoint cloud-based 3D open-vocabulary scene graph generation framework in which\nthe requirement of posed RGB-D image series is eliminated. This hierarchical\nframework contains room and object detection/segmentation and open-vocabulary\nclassification. For the room layer, we leverage the advantage of merging the\ngeometry-based border detection algorithm with the learning-based region\ndetection to segment rooms and create a \"Snap-Lookup\" framework for\nopen-vocabulary room classification. In addition, we create an end-to-end\npipeline for the object layer to detect and classify 3D objects based solely on\n3D point cloud data. Our evaluation results show that our framework can\noutperform the current state-of-the-art (SOTA) open-vocabulary object and room\nsegmentation and classification algorithm on widely used real-scene datasets.\n", "date": "2024-09-16", "categories": []}
{"id": "2409.06625", "title": "Towards Localizing Structural Elements: Merging Geometrical Detection\n  with Semantic Verification in RGB-D Data", "abstract": "  RGB-D cameras supply rich and dense visual and spatial information for\nvarious robotics tasks such as scene understanding, map reconstruction, and\nlocalization. Integrating depth and visual information can aid robots in\nlocalization and element mapping, advancing applications like 3D scene graph\ngeneration and Visual Simultaneous Localization and Mapping (VSLAM). While\npoint cloud data containing such information is primarily used for enhanced\nscene understanding, exploiting their potential to capture and represent rich\nsemantic information has yet to be adequately targeted. This paper presents a\nreal-time pipeline for localizing building components, including wall and\nground surfaces, by integrating geometric calculations for pure 3D plane\ndetection followed by validating their semantic category using point cloud data\nfrom RGB-D cameras. It has a parallel multi-thread architecture to precisely\nestimate poses and equations of all the planes detected in the environment,\nfilters the ones forming the map structure using a panoptic segmentation\nvalidation, and keeps only the validated building components. Incorporating the\nproposed method into a VSLAM framework confirmed that constraining the map with\nthe detected environment-driven semantic elements can improve scene\nunderstanding and map reconstruction accuracy. It can also ensure\n(re-)association of these detected components into a unified 3D scene graph,\nbridging the gap between geometric accuracy and semantic understanding.\nAdditionally, the pipeline allows for the detection of potential higher-level\nstructural entities, such as rooms, by identifying the relationships between\nbuilding components based on their layout.\n", "date": "2024-09-10", "categories": []}
{"id": "2405.15580", "title": "Open-Vocabulary SAM3D: Towards Training-free Open-Vocabulary 3D Scene\n  Understanding", "abstract": "  Open-vocabulary 3D scene understanding presents a significant challenge in\nthe field. Recent works have sought to transfer knowledge embedded in\nvision-language models from 2D to 3D domains. However, these approaches often\nrequire prior knowledge from specific 3D scene datasets, limiting their\napplicability in open-world scenarios. The Segment Anything Model (SAM) has\ndemonstrated remarkable zero-shot segmentation capabilities, prompting us to\ninvestigate its potential for comprehending 3D scenes without training. In this\npaper, we introduce OV-SAM3D, a training-free method that contains a universal\nframework for understanding open-vocabulary 3D scenes. This framework is\ndesigned to perform understanding tasks for any 3D scene without requiring\nprior knowledge of the scene. Specifically, our method is composed of two key\nsub-modules: First, we initiate the process by generating superpoints as the\ninitial 3D prompts and refine these prompts using segment masks derived from\nSAM. Moreover, we then integrate a specially designed overlapping score table\nwith open tags from the Recognize Anything Model (RAM) to produce final 3D\ninstances with open-world labels. Empirical evaluations on the ScanNet200 and\nnuScenes datasets demonstrate that our approach surpasses existing\nopen-vocabulary methods in unknown open-world environments.\n", "date": "2024-05-24", "categories": []}
{"id": "2407.14491", "title": "PD-APE: A Parallel Decoding Framework with Adaptive Position Encoding\n  for 3D Visual Grounding", "abstract": "  3D visual grounding aims to identify objects in 3D point cloud scenes that\nmatch specific natural language descriptions. This requires the model to not\nonly focus on the target object itself but also to consider the surrounding\nenvironment to determine whether the descriptions are met. Most previous works\nattempt to accomplish both tasks within the same module, which can easily lead\nto a distraction of attention. To this end, we propose PD-APE, a dual-branch\ndecoding framework that separately decodes target object attributes and\nsurrounding layouts. Specifically, in the target object branch, the decoder\nprocesses text tokens that describe features of the target object (e.g.,\ncategory and color), guiding the queries to pay attention to the target object\nitself. In the surrounding branch, the queries align with other text tokens\nthat carry surrounding environment information, making the attention maps\naccurately capture the layout described in the text. Benefiting from the\nproposed dual-branch design, the queries are allowed to focus on points\nrelevant to each branch's specific objective. Moreover, we design an adaptive\nposition encoding method for each branch respectively. In the target object\nbranch, the position encoding relies on the relative positions between seed\npoints and predicted 3D boxes. In the surrounding branch, the attention map is\nadditionally guided by the confidence between visual and text features,\nenabling the queries to focus on points that have valuable layout information.\nExtensive experiments demonstrate that we surpass the state-of-the-art on two\nwidely adopted 3D visual grounding datasets, ScanRefer and Nr3D.\n", "date": "2024-07-19", "categories": []}
{"id": "2409.01662", "title": "Efficiently Expanding Receptive Fields: Local Split Attention and\n  Parallel Aggregation for Enhanced Large-scale Point Cloud Semantic\n  Segmentation", "abstract": "  Expanding the receptive field in a deep learning model for large-scale 3D\npoint cloud segmentation is an effective technique for capturing rich\ncontextual information, which consequently enhances the network's ability to\nlearn meaningful features. However, this often leads to increased computational\ncomplexity and risk of overfitting, challenging the efficiency and\neffectiveness of the learning paradigm. To address these limitations, we\npropose the Local Split Attention Pooling (LSAP) mechanism to effectively\nexpand the receptive field through a series of local split operations, thus\nfacilitating the acquisition of broader contextual knowledge. Concurrently, it\noptimizes the computational workload associated with attention-pooling layers\nto ensure a more streamlined processing workflow. Based on LSAP, a Parallel\nAggregation Enhancement (PAE) module is introduced to enable parallel\nprocessing of data using both 2D and 3D neighboring information to further\nenhance contextual representations within the network. In light of the\naforementioned designs, we put forth a novel framework, designated as LSNet,\nfor large-scale point cloud semantic segmentation. Extensive evaluations\ndemonstrated the efficacy of seamlessly integrating the proposed PAE module\ninto existing frameworks, yielding significant improvements in mean\nintersection over union (mIoU) metrics, with a notable increase of up to 11%.\nFurthermore, LSNet demonstrated superior performance compared to\nstate-of-the-art semantic segmentation networks on three benchmark datasets,\nincluding S3DIS, Toronto3D, and SensatUrban. It is noteworthy that our method\nachieved a substantial speedup of approximately 38.8% compared to those\nemploying similar-sized receptive fields, which serves to highlight both its\ncomputational efficiency and practical utility in real-world large-scale\nscenes.\n", "date": "2024-09-03", "categories": []}
{"id": "2403.14628", "title": "Zero-Shot Multi-Object Scene Completion", "abstract": "  We present a 3D scene completion method that recovers the complete geometry\nof multiple unseen objects in complex scenes from a single RGB-D image. Despite\nnotable advancements in single-object 3D shape completion, high-quality\nreconstructions in highly cluttered real-world multi-object scenes remains a\nchallenge. To address this issue, we propose OctMAE, an architecture that\nleverages an Octree U-Net and a latent 3D MAE to achieve high-quality and near\nreal-time multi-object scene completion through both local and global geometric\nreasoning. Because a naive 3D MAE can be computationally intractable and memory\nintensive even in the latent space, we introduce a novel occlusion masking\nstrategy and adopt 3D rotary embeddings, which significantly improves the\nruntime and scene completion quality. To generalize to a wide range of objects\nin diverse scenes, we create a large-scale photorealistic dataset, featuring a\ndiverse set of 12K 3D object models from the Objaverse dataset which are\nrendered in multi-object scenes with physics-based positioning. Our method\noutperforms the current state-of-the-art on both synthetic and real-world\ndatasets and demonstrates a strong zero-shot capability.\n", "date": "2024-03-21", "categories": []}
{"id": "2405.18033", "title": "RT-GS2: Real-Time Generalizable Semantic Segmentation for 3D Gaussian\n  Representations of Radiance Fields", "abstract": "  Gaussian Splatting has revolutionized the world of novel view synthesis by\nachieving high rendering performance in real-time. Recently, studies have\nfocused on enriching these 3D representations with semantic information for\ndownstream tasks. In this paper, we introduce RT-GS2, the first generalizable\nsemantic segmentation method employing Gaussian Splatting. While existing\nGaussian Splatting-based approaches rely on scene-specific training, RT-GS2\ndemonstrates the ability to generalize to unseen scenes. Our method adopts a\nnew approach by first extracting view-independent 3D Gaussian features in a\nself-supervised manner, followed by a novel View-Dependent / View-Independent\n(VDVI) feature fusion to enhance semantic consistency over different views.\nExtensive experimentation on three different datasets showcases RT-GS2's\nsuperiority over the state-of-the-art methods in semantic segmentation quality,\nexemplified by a 8.01% increase in mIoU on the Replica dataset. Moreover, our\nmethod achieves real-time performance of 27.03 FPS, marking an astonishing 901\ntimes speedup compared to existing approaches. This work represents a\nsignificant advancement in the field by introducing, to the best of our\nknowledge, the first real-time generalizable semantic segmentation method for\n3D Gaussian representations of radiance fields.\n", "date": "2024-05-28", "categories": []}
{"id": "2312.16268", "title": "360 Layout Estimation via Orthogonal Planes Disentanglement and\n  Multi-view Geometric Consistency Perception", "abstract": "  Existing panoramic layout estimation solutions tend to recover room\nboundaries from a vertically compressed sequence, yielding imprecise results as\nthe compression process often muddles the semantics between various planes.\nBesides, these data-driven approaches impose an urgent demand for massive data\nannotations, which are laborious and time-consuming. For the first problem, we\npropose an orthogonal plane disentanglement network (termed DOPNet) to\ndistinguish ambiguous semantics. DOPNet consists of three modules that are\nintegrated to deliver distortion-free, semantics-clean, and detail-sharp\ndisentangled representations, which benefit the subsequent layout recovery. For\nthe second problem, we present an unsupervised adaptation technique tailored\nfor horizon-depth and ratio representations. Concretely, we introduce an\noptimization strategy for decision-level layout analysis and a 1D cost volume\nconstruction method for feature-level multi-view aggregation, both of which are\ndesigned to fully exploit the geometric consistency across multiple\nperspectives. The optimizer provides a reliable set of pseudo-labels for\nnetwork training, while the 1D cost volume enriches each view with\ncomprehensive scene information derived from other perspectives. Extensive\nexperiments demonstrate that our solution outperforms other SoTA models on both\nmonocular layout estimation and multi-view layout estimation tasks. Cobe can be\navailable at https://github.com/zhijieshen-bjtu/MV-DOPNet.\n", "date": "2023-12-26", "categories": []}
{"id": "2403.16318", "title": "AutoInst: Automatic Instance-Based Segmentation of LiDAR 3D Scans", "abstract": "  Recently, progress in acquisition equipment such as LiDAR sensors has enabled\nsensing increasingly spacious outdoor 3D environments. Making sense of such 3D\nacquisitions requires fine-grained scene understanding, such as constructing\ninstance-based 3D scene segmentations. Commonly, a neural network is trained\nfor this task; however, this requires access to a large, densely annotated\ndataset, which is widely known to be challenging to obtain. To address this\nissue, in this work we propose to predict instance segmentations for 3D scenes\nin an unsupervised way, without relying on ground-truth annotations. To this\nend, we construct a learning framework consisting of two components: (1) a\npseudo-annotation scheme for generating initial unsupervised pseudo-labels; and\n(2) a self-training algorithm for instance segmentation to fit robust, accurate\ninstances from initial noisy proposals. To enable generating 3D instance mask\nproposals, we construct a weighted proxy-graph by connecting 3D points with\nedges integrating multi-modal image- and point-based self-supervised features,\nand perform graph-cuts to isolate individual pseudo-instances. We then build on\na state-of-the-art point-based architecture and train a 3D instance\nsegmentation model, resulting in significant refinement of initial proposals.\nTo scale to arbitrary complexity 3D scenes, we design our algorithm to operate\non local 3D point chunks and construct a merging step to generate scene-level\ninstance segmentations. Experiments on the challenging SemanticKITTI benchmark\ndemonstrate the potential of our approach, where it attains 13.3% higher\nAverage Precision and 9.1% higher F1 score compared to the best-performing\nbaseline. The code will be made publicly available at\nhttps://github.com/artonson/autoinst.\n", "date": "2024-03-24", "categories": []}
{"id": "2408.16768", "title": "SAM2Point: Segment Any 3D as Videos in Zero-shot and Promptable Manners", "abstract": "  We introduce SAM2Point, a preliminary exploration adapting Segment Anything\nModel 2 (SAM 2) for zero-shot and promptable 3D segmentation. SAM2Point\ninterprets any 3D data as a series of multi-directional videos, and leverages\nSAM 2 for 3D-space segmentation, without further training or 2D-3D projection.\nOur framework supports various prompt types, including 3D points, boxes, and\nmasks, and can generalize across diverse scenarios, such as 3D objects, indoor\nscenes, outdoor environments, and raw sparse LiDAR. Demonstrations on multiple\n3D datasets, e.g., Objaverse, S3DIS, ScanNet, Semantic3D, and KITTI, highlight\nthe robust generalization capabilities of SAM2Point. To our best knowledge, we\npresent the most faithful implementation of SAM in 3D, which may serve as a\nstarting point for future research in promptable 3D segmentation. Online Demo:\nhttps://huggingface.co/spaces/ZiyuG/SAM2Point . Code:\nhttps://github.com/ZiyuGuo99/SAM2Point .\n", "date": "2024-08-29", "categories": []}
{"id": "2408.15608", "title": "Geometry-guided Feature Learning and Fusion for Indoor Scene\n  Reconstruction", "abstract": "  In addition to color and textural information, geometry provides important\ncues for 3D scene reconstruction. However, current reconstruction methods only\ninclude geometry at the feature level thus not fully exploiting the geometric\ninformation.\n  In contrast, this paper proposes a novel geometry integration mechanism for\n3D scene reconstruction. Our approach incorporates 3D geometry at three levels,\ni.e. feature learning, feature fusion, and network supervision. First,\ngeometry-guided feature learning encodes geometric priors to contain\nview-dependent information. Second, a geometry-guided adaptive feature fusion\nis introduced which utilizes the geometric priors as a guidance to adaptively\ngenerate weights for multiple views. Third, at the supervision level, taking\nthe consistency between 2D and 3D normals into account, a consistent 3D normal\nloss is designed to add local constraints.\n  Large-scale experiments are conducted on the ScanNet dataset, showing that\nvolumetric methods with our geometry integration mechanism outperform\nstate-of-the-art methods quantitatively as well as qualitatively. Volumetric\nmethods with ours also show good generalization on the 7-Scenes and TUM RGB-D\ndatasets.\n", "date": "2024-08-28", "categories": []}
{"id": "2408.15813", "title": "DQFormer: Towards Unified LiDAR Panoptic Segmentation with Decoupled\n  Queries", "abstract": "  LiDAR panoptic segmentation, which jointly performs instance and semantic\nsegmentation for things and stuff classes, plays a fundamental role in LiDAR\nperception tasks. While most existing methods explicitly separate these two\nsegmentation tasks and utilize different branches (i.e., semantic and instance\nbranches), some recent methods have embraced the query-based paradigm to unify\nLiDAR panoptic segmentation. However, the distinct spatial distribution and\ninherent characteristics of objects(things) and their surroundings(stuff) in 3D\nscenes lead to challenges, including the mutual competition of things/stuff and\nthe ambiguity of classification/segmentation. In this paper, we propose\ndecoupling things/stuff queries according to their intrinsic properties for\nindividual decoding and disentangling classification/segmentation to mitigate\nambiguity. To this end, we propose a novel framework dubbed DQFormer to\nimplement semantic and instance segmentation in a unified workflow.\nSpecifically, we design a decoupled query generator to propose informative\nqueries with semantics by localizing things/stuff positions and fusing\nmulti-level BEV embeddings. Moreover, a query-oriented mask decoder is\nintroduced to decode corresponding segmentation masks by performing masked\ncross-attention between queries and mask embeddings. Finally, the decoded masks\nare combined with the semantics of the queries to produce panoptic results.\nExtensive experiments on nuScenes and SemanticKITTI datasets demonstrate the\nsuperiority of our DQFormer framework.\n", "date": "2024-08-28", "categories": []}
{"id": "2408.13936", "title": "OpenNav: Efficient Open Vocabulary 3D Object Detection for Smart\n  Wheelchair Navigation", "abstract": "  Open vocabulary 3D object detection (OV3D) allows precise and extensible\nobject recognition crucial for adapting to diverse environments encountered in\nassistive robotics. This paper presents OpenNav, a zero-shot 3D object\ndetection pipeline based on RGB-D images for smart wheelchairs. Our pipeline\nintegrates an open-vocabulary 2D object detector with a mask generator for\nsemantic segmentation, followed by depth isolation and point cloud construction\nto create 3D bounding boxes. The smart wheelchair exploits these 3D bounding\nboxes to identify potential targets and navigate safely. We demonstrate\nOpenNav's performance through experiments on the Replica dataset and we report\npreliminary results with a real wheelchair. OpenNav improves state-of-the-art\nsignificantly on the Replica dataset at mAP25 (+9pts) and mAP50 (+5pts) with\nmarginal improvement at mAP. The code is publicly available at this link:\nhttps://github.com/EasyWalk-PRIN/OpenNav.\n", "date": "2024-08-25", "categories": []}
{"id": "2403.15624", "title": "Semantic Gaussians: Open-Vocabulary Scene Understanding with 3D Gaussian\n  Splatting", "abstract": "  Open-vocabulary 3D scene understanding presents a significant challenge in\ncomputer vision, with wide-ranging applications in embodied agents and\naugmented reality systems. Existing methods adopt neurel rendering methods as\n3D representations and jointly optimize color and semantic features to achieve\nrendering and scene understanding simultaneously. In this paper, we introduce\nSemantic Gaussians, a novel open-vocabulary scene understanding approach based\non 3D Gaussian Splatting. Our key idea is to distill knowledge from 2D\npre-trained models to 3D Gaussians. Unlike existing methods, we design a\nversatile projection approach that maps various 2D semantic features from\npre-trained image encoders into a novel semantic component of 3D Gaussians,\nwhich is based on spatial relationship and need no additional training. We\nfurther build a 3D semantic network that directly predicts the semantic\ncomponent from raw 3D Gaussians for fast inference. The quantitative results on\nScanNet segmentation and LERF object localization demonstates the superior\nperformance of our method. Additionally, we explore several applications of\nSemantic Gaussians including object part segmentation, instance segmentation,\nscene editing, and spatiotemporal segmentation with better qualitative results\nover 2D and 3D baselines, highlighting its versatility and effectiveness on\nsupporting diverse downstream tasks.\n", "date": "2024-03-22", "categories": []}
{"id": "2403.08512", "title": "MergeOcc: Bridge the Domain Gap between Different LiDARs for Robust\n  Occupancy Prediction", "abstract": "  LiDAR-based 3D occupancy prediction evolved rapidly alongside the emergence\nof large datasets. Nevertheless, the potential of existing diverse datasets\nremains underutilized as they kick in individually. Models trained on a\nspecific dataset often suffer considerable performance degradation when\ndeployed to real-world scenarios or datasets involving disparate LiDARs. This\npaper aims to develop a generalized model called MergeOcc, to simultaneously\nhandle different LiDARs by leveraging multiple datasets. The gaps among LiDAR\ndatasets primarily manifest in geometric disparities and semantic\ninconsistencies. Thus, MergeOcc incorporates a novel model featuring a\ngeometric realignment module and a semantic label mapping module to enable\nmultiple datasets training (MDT). The effectiveness of MergeOcc is validated\nthrough experiments on two prominent datasets for autonomous vehicles:\nOpenOccupancy-nuScenes and SemanticKITTI. The results demonstrate its enhanced\nrobustness and remarkable performance across both types of LiDARs,\noutperforming several SOTA multi-modality methods. Notably, despite using an\nidentical model architecture and hyper-parameter set, MergeOcc can\nsignificantly surpass the baseline due to its exposure to more diverse data.\nMergeOcc is considered the first cross-dataset 3D occupancy prediction pipeline\nthat effectively bridges the domain gap for seamless deployment across\nheterogeneous platforms.\n", "date": "2024-03-13", "categories": []}
{"id": "2408.09104", "title": "HybridOcc: NeRF Enhanced Transformer-based Multi-Camera 3D Occupancy\n  Prediction", "abstract": "  Vision-based 3D semantic scene completion (SSC) describes autonomous driving\nscenes through 3D volume representations. However, the occlusion of invisible\nvoxels by scene surfaces poses challenges to current SSC methods in\nhallucinating refined 3D geometry. This paper proposes HybridOcc, a hybrid 3D\nvolume query proposal method generated by Transformer framework and NeRF\nrepresentation and refined in a coarse-to-fine SSC prediction framework.\nHybridOcc aggregates contextual features through the Transformer paradigm based\non hybrid query proposals while combining it with NeRF representation to obtain\ndepth supervision. The Transformer branch contains multiple scales and uses\nspatial cross-attention for 2D to 3D transformation. The newly designed NeRF\nbranch implicitly infers scene occupancy through volume rendering, including\nvisible and invisible voxels, and explicitly captures scene depth rather than\ngenerating RGB color. Furthermore, we present an innovative occupancy-aware ray\nsampling method to orient the SSC task instead of focusing on the scene\nsurface, further improving the overall performance. Extensive experiments on\nnuScenes and SemanticKITTI datasets demonstrate the effectiveness of our\nHybridOcc on the SSC task.\n", "date": "2024-08-17", "categories": []}
{"id": "2408.07648", "title": "See It All: Contextualized Late Aggregation for 3D Dense Captioning", "abstract": "  3D dense captioning is a task to localize objects in a 3D scene and generate\ndescriptive sentences for each object. Recent approaches in 3D dense captioning\nhave adopted transformer encoder-decoder frameworks from object detection to\nbuild an end-to-end pipeline without hand-crafted components. However, these\napproaches struggle with contradicting objectives where a single query\nattention has to simultaneously view both the tightly localized object regions\nand contextual environment. To overcome this challenge, we introduce SIA\n(See-It-All), a transformer pipeline that engages in 3D dense captioning with a\nnovel paradigm called late aggregation. SIA simultaneously decodes two sets of\nqueries-context query and instance query. The instance query focuses on\nlocalization and object attribute descriptions, while the context query\nversatilely captures the region-of-interest of relationships between multiple\nobjects or with the global scene, then aggregated afterwards (i.e., late\naggregation) via simple distance-based measures. To further enhance the quality\nof contextualized caption generation, we design a novel aggregator to generate\na fully informed caption based on the surrounding context, the global\nenvironment, and object instances. Extensive experiments on two of the most\nwidely-used 3D dense captioning datasets demonstrate that our proposed method\nachieves a significant improvement over prior methods.\n", "date": "2024-08-14", "categories": []}
{"id": "2408.06662", "title": "Bi-directional Contextual Attention for 3D Dense Captioning", "abstract": "  3D dense captioning is a task involving the localization of objects and the\ngeneration of descriptions for each object in a 3D scene. Recent approaches\nhave attempted to incorporate contextual information by modeling relationships\nwith object pairs or aggregating the nearest neighbor features of an object.\nHowever, the contextual information constructed in these scenarios is limited\nin two aspects: first, objects have multiple positional relationships that\nexist across the entire global scene, not only near the object itself. Second,\nit faces with contradicting objectives--where localization and attribute\ndescriptions are generated better with tight localization, while descriptions\ninvolving global positional relations are generated better with contextualized\nfeatures of the global scene. To overcome this challenge, we introduce BiCA, a\ntransformer encoder-decoder pipeline that engages in 3D dense captioning for\neach object with Bi-directional Contextual Attention. Leveraging parallelly\ndecoded instance queries for objects and context queries for non-object\ncontexts, BiCA generates object-aware contexts, where the contexts relevant to\neach object is summarized, and context-aware objects, where the objects\nrelevant to the summarized object-aware contexts are aggregated. This extension\nrelieves previous methods from the contradicting objectives, enhancing both\nlocalization performance and enabling the aggregation of contextual features\nthroughout the global scene; thus improving caption generation performance\nsimultaneously. Extensive experiments on two of the most widely-used 3D dense\ncaptioning datasets demonstrate that our proposed method achieves a significant\nimprovement over prior methods.\n", "date": "2024-08-13", "categories": []}
{"id": "2408.06926", "title": "SceneGPT: A Language Model for 3D Scene Understanding", "abstract": "  Building models that can understand and reason about 3D scenes is difficult\nowing to the lack of data sources for 3D supervised training and large-scale\ntraining regimes. In this work we ask - How can the knowledge in a pre-trained\nlanguage model be leveraged for 3D scene understanding without any 3D\npre-training. The aim of this work is to establish whether pre-trained LLMs\npossess priors/knowledge required for reasoning in 3D space and how can we\nprompt them such that they can be used for general purpose spatial reasoning\nand object understanding in 3D. To this end, we present SceneGPT, an LLM based\nscene understanding system which can perform 3D spatial reasoning without\ntraining or explicit 3D supervision. The key components of our framework are -\n1) a 3D scene graph, that serves as scene representation, encoding the objects\nin the scene and their spatial relationships 2) a pre-trained LLM that can be\nadapted with in context learning for 3D spatial reasoning. We evaluate our\nframework qualitatively on object and scene understanding tasks including\nobject semantics, physical properties and affordances (object-level) and\nspatial understanding (scene-level).\n", "date": "2024-08-13", "categories": []}
{"id": "2403.11796", "title": "OpenOcc: Open Vocabulary 3D Scene Reconstruction via Occupancy\n  Representation", "abstract": "  3D reconstruction has been widely used in autonomous navigation fields of\nmobile robotics. However, the former research can only provide the basic\ngeometry structure without the capability of open-world scene understanding,\nlimiting advanced tasks like human interaction and visual navigation. Moreover,\ntraditional 3D scene understanding approaches rely on expensive labeled 3D\ndatasets to train a model for a single task with supervision. Thus, geometric\nreconstruction with zero-shot scene understanding i.e. Open vocabulary 3D\nUnderstanding and Reconstruction, is crucial for the future development of\nmobile robots. In this paper, we propose OpenOcc, a novel framework unifying\nthe 3D scene reconstruction and open vocabulary understanding with neural\nradiance fields. We model the geometric structure of the scene with occupancy\nrepresentation and distill the pre-trained open vocabulary model into a 3D\nlanguage field via volume rendering for zero-shot inference. Furthermore, a\nnovel semantic-aware confidence propagation (SCP) method has been proposed to\nrelieve the issue of language field representation degeneracy caused by\ninconsistent measurements in distilled features. Experimental results show that\nour approach achieves competitive performance in 3D scene understanding tasks,\nespecially for small and long-tail objects.\n", "date": "2024-03-18", "categories": []}
{"id": "2408.03338", "title": "InLUT3D: Challenging real indoor dataset for point cloud analysis", "abstract": "  In this paper, we introduce the InLUT3D point cloud dataset, a comprehensive\nresource designed to advance the field of scene understanding in indoor\nenvironments. The dataset covers diverse spaces within the W7 faculty buildings\nof Lodz University of Technology, characterised by high-resolution laser-based\npoint clouds and manual labelling. Alongside the dataset, we propose metrics\nand benchmarking guidelines essential for ensuring trustworthy and reproducible\nresults in algorithm evaluation. We anticipate that the introduction of the\nInLUT3D dataset and its associated benchmarks will catalyse future advancements\nin 3D scene understanding, facilitating methodological rigour and inspiring new\napproaches in the field.\n", "date": "2024-07-22", "categories": []}
{"id": "2404.16845", "title": "HaLo-NeRF: Learning Geometry-Guided Semantics for Exploring\n  Unconstrained Photo Collections", "abstract": "  Internet image collections containing photos captured by crowds of\nphotographers show promise for enabling digital exploration of large-scale\ntourist landmarks. However, prior works focus primarily on geometric\nreconstruction and visualization, neglecting the key role of language in\nproviding a semantic interface for navigation and fine-grained understanding.\nIn constrained 3D domains, recent methods have leveraged vision-and-language\nmodels as a strong prior of 2D visual semantics. While these models display an\nexcellent understanding of broad visual semantics, they struggle with\nunconstrained photo collections depicting such tourist landmarks, as they lack\nexpert knowledge of the architectural domain. In this work, we present a\nlocalization system that connects neural representations of scenes depicting\nlarge-scale landmarks with text describing a semantic region within the scene,\nby harnessing the power of SOTA vision-and-language models with adaptations for\nunderstanding landmark scene semantics. To bolster such models with\nfine-grained knowledge, we leverage large-scale Internet data containing images\nof similar landmarks along with weakly-related textual information. Our\napproach is built upon the premise that images physically grounded in space can\nprovide a powerful supervision signal for localizing new concepts, whose\nsemantics may be unlocked from Internet textual metadata with large language\nmodels. We use correspondences between views of scenes to bootstrap spatial\nunderstanding of these semantics, providing guidance for 3D-compatible\nsegmentation that ultimately lifts to a volumetric scene representation. Our\nresults show that HaLo-NeRF can accurately localize a variety of semantic\nconcepts related to architectural landmarks, surpassing the results of other 3D\nmodels as well as strong 2D segmentation baselines. Our project page is at\nhttps://tau-vailab.github.io/HaLo-NeRF/.\n", "date": "2024-02-14", "categories": []}
{"id": "2407.20664", "title": "3D-GRES: Generalized 3D Referring Expression Segmentation", "abstract": "  3D Referring Expression Segmentation (3D-RES) is dedicated to segmenting a\nspecific instance within a 3D space based on a natural language description.\nHowever, current approaches are limited to segmenting a single target,\nrestricting the versatility of the task. To overcome this limitation, we\nintroduce Generalized 3D Referring Expression Segmentation (3D-GRES), which\nextends the capability to segment any number of instances based on natural\nlanguage instructions. In addressing this broader task, we propose the\nMulti-Query Decoupled Interaction Network (MDIN), designed to break down\nmulti-object segmentation tasks into simpler, individual segmentations. MDIN\ncomprises two fundamental components: Text-driven Sparse Queries (TSQ) and\nMulti-object Decoupling Optimization (MDO). TSQ generates sparse point cloud\nfeatures distributed over key targets as the initialization for queries.\nMeanwhile, MDO is tasked with assigning each target in multi-object scenarios\nto different queries while maintaining their semantic consistency. To adapt to\nthis new task, we build a new dataset, namely Multi3DRes. Our comprehensive\nevaluations on this dataset demonstrate substantial enhancements over existing\nmodels, thus charting a new path for intricate multi-object 3D scene\ncomprehension. The benchmark and code are available at\nhttps://github.com/sosppxo/MDIN.\n", "date": "2024-07-30", "categories": []}
{"id": "2401.05925", "title": "Learning Segmented 3D Gaussians via Efficient Feature Unprojection for\n  Zero-shot Neural Scene Segmentation", "abstract": "  Zero-shot neural scene segmentation, which reconstructs 3D neural\nsegmentation field without manual annotations, serves as an effective way for\nscene understanding. However, existing models, especially the efficient 3D\nGaussian-based methods, struggle to produce compact segmentation results. This\nissue stems primarily from their redundant learnable attributes assigned on\nindividual Gaussians, leading to a lack of robustness against the\n3D-inconsistencies in zero-shot generated raw labels. To address this problem,\nour work, named Compact Segmented 3D Gaussians (CoSegGaussians), proposes the\nFeature Unprojection and Fusion module as the segmentation field, which\nutilizes a shallow decoder generalizable for all Gaussians based on high-level\nfeatures. Specifically, leveraging the learned Gaussian geometric parameters,\nsemantic-aware image-based features are introduced into the scene via our\nunprojection technique. The lifted features, together with spatial information,\nare fed into the multi-scale aggregation decoder to generate segmentation\nidentities for all Gaussians. Furthermore, we design CoSeg Loss to boost model\nrobustness against 3D-inconsistent noises. Experimental results show that our\nmodel surpasses baselines on zero-shot semantic segmentation task, improving by\n~10% mIoU over the best baseline. Code and more results will be available at\nhttps://David-Dou.github.io/CoSegGaussians.\n", "date": "2024-01-11", "categories": []}
{"id": "2403.10997", "title": "N2F2: Hierarchical Scene Understanding with Nested Neural Feature Fields", "abstract": "  Understanding complex scenes at multiple levels of abstraction remains a\nformidable challenge in computer vision. To address this, we introduce Nested\nNeural Feature Fields (N2F2), a novel approach that employs hierarchical\nsupervision to learn a single feature field, wherein different dimensions\nwithin the same high-dimensional feature encode scene properties at varying\ngranularities. Our method allows for a flexible definition of hierarchies,\ntailored to either the physical dimensions or semantics or both, thereby\nenabling a comprehensive and nuanced understanding of scenes. We leverage a 2D\nclass-agnostic segmentation model to provide semantically meaningful pixel\ngroupings at arbitrary scales in the image space, and query the CLIP\nvision-encoder to obtain language-aligned embeddings for each of these\nsegments. Our proposed hierarchical supervision method then assigns different\nnested dimensions of the feature field to distill the CLIP embeddings using\ndeferred volumetric rendering at varying physical scales, creating a\ncoarse-to-fine representation. Extensive experiments show that our approach\noutperforms the state-of-the-art feature field distillation methods on tasks\nsuch as open-vocabulary 3D segmentation and localization, demonstrating the\neffectiveness of the learned nested feature field.\n", "date": "2024-03-16", "categories": []}
{"id": "2405.17596", "title": "GOI: Find 3D Gaussians of Interest with an Optimizable Open-vocabulary\n  Semantic-space Hyperplane", "abstract": "  3D open-vocabulary scene understanding, crucial for advancing augmented\nreality and robotic applications, involves interpreting and locating specific\nregions within a 3D space as directed by natural language instructions. To this\nend, we introduce GOI, a framework that integrates semantic features from 2D\nvision-language foundation models into 3D Gaussian Splatting (3DGS) and\nidentifies 3D Gaussians of Interest using an Optimizable Semantic-space\nHyperplane. Our approach includes an efficient compression method that utilizes\nscene priors to condense noisy high-dimensional semantic features into compact\nlow-dimensional vectors, which are subsequently embedded in 3DGS. During the\nopen-vocabulary querying process, we adopt a distinct approach compared to\nexisting methods, which depend on a manually set fixed empirical threshold to\nselect regions based on their semantic feature distance to the query text\nembedding. This traditional approach often lacks universal accuracy, leading to\nchallenges in precisely identifying specific target areas. Instead, our method\ntreats the feature selection process as a hyperplane division within the\nfeature space, retaining only those features that are highly relevant to the\nquery. We leverage off-the-shelf 2D Referring Expression Segmentation (RES)\nmodels to fine-tune the semantic-space hyperplane, enabling a more precise\ndistinction between target regions and others. This fine-tuning substantially\nimproves the accuracy of open-vocabulary queries, ensuring the precise\nlocalization of pertinent 3D Gaussians. Extensive experiments demonstrate GOI's\nsuperiority over previous state-of-the-art methods. Our project page is\navailable at https://quyans.github.io/GOI-Hyperplane/ .\n", "date": "2024-05-27", "categories": []}
{"id": "2405.11442", "title": "Unifying 3D Vision-Language Understanding via Promptable Queries", "abstract": "  A unified model for 3D vision-language (3D-VL) understanding is expected to\ntake various scene representations and perform a wide range of tasks in a 3D\nscene. However, a considerable gap exists between existing methods and such a\nunified model, due to the independent application of representation and\ninsufficient exploration of 3D multi-task training. In this paper, we introduce\nPQ3D, a unified model capable of using Promptable Queries to tackle a wide\nrange of 3D-VL tasks, from low-level instance segmentation to high-level\nreasoning and planning. This is achieved through three key innovations: (1)\nunifying various 3D scene representations (i.e., voxels, point clouds,\nmulti-view images) into a shared 3D coordinate space by segment-level grouping,\n(2) an attention-based query decoder for task-specific information retrieval\nguided by prompts, and (3) universal output heads for different tasks to\nsupport multi-task training. Tested across ten diverse 3D-VL datasets, PQ3D\ndemonstrates impressive performance on these tasks, setting new records on most\nbenchmarks. Particularly, PQ3D improves the state-of-the-art on ScanNet200 by\n4.9% (AP25), ScanRefer by 5.4% (acc@0.5), Multi3DRefer by 11.7% (F1@0.5), and\nScan2Cap by 13.4% (CIDEr@0.5). Moreover, PQ3D supports flexible inference with\nindividual or combined forms of available 3D representations, e.g., solely\nvoxel input.\n", "date": "2024-05-19", "categories": []}
{"id": "2403.18118", "title": "EgoLifter: Open-world 3D Segmentation for Egocentric Perception", "abstract": "  In this paper we present EgoLifter, a novel system that can automatically\nsegment scenes captured from egocentric sensors into a complete decomposition\nof individual 3D objects. The system is specifically designed for egocentric\ndata where scenes contain hundreds of objects captured from natural\n(non-scanning) motion. EgoLifter adopts 3D Gaussians as the underlying\nrepresentation of 3D scenes and objects and uses segmentation masks from the\nSegment Anything Model (SAM) as weak supervision to learn flexible and\npromptable definitions of object instances free of any specific object\ntaxonomy. To handle the challenge of dynamic objects in ego-centric videos, we\ndesign a transient prediction module that learns to filter out dynamic objects\nin the 3D reconstruction. The result is a fully automatic pipeline that is able\nto reconstruct 3D object instances as collections of 3D Gaussians that\ncollectively compose the entire scene. We created a new benchmark on the Aria\nDigital Twin dataset that quantitatively demonstrates its state-of-the-art\nperformance in open-world 3D segmentation from natural egocentric input. We run\nEgoLifter on various egocentric activity datasets which shows the promise of\nthe method for 3D egocentric perception at scale.\n", "date": "2024-03-26", "categories": []}
{"id": "2403.19580", "title": "OV-Uni3DETR: Towards Unified Open-Vocabulary 3D Object Detection via\n  Cycle-Modality Propagation", "abstract": "  In the current state of 3D object detection research, the severe scarcity of\nannotated 3D data, substantial disparities across different data modalities,\nand the absence of a unified architecture, have impeded the progress towards\nthe goal of universality. In this paper, we propose \\textbf{OV-Uni3DETR}, a\nunified open-vocabulary 3D detector via cycle-modality propagation. Compared\nwith existing 3D detectors, OV-Uni3DETR offers distinct advantages: 1)\nOpen-vocabulary 3D detection: During training, it leverages various accessible\ndata, especially extensive 2D detection images, to boost training diversity.\nDuring inference, it can detect both seen and unseen classes. 2) Modality\nunifying: It seamlessly accommodates input data from any given modality,\neffectively addressing scenarios involving disparate modalities or missing\nsensor information, thereby supporting test-time modality switching. 3) Scene\nunifying: It provides a unified multi-modal model architecture for diverse\nscenes collected by distinct sensors. Specifically, we propose the\ncycle-modality propagation, aimed at propagating knowledge bridging 2D and 3D\nmodalities, to support the aforementioned functionalities. 2D semantic\nknowledge from large-vocabulary learning guides novel class discovery in the 3D\ndomain, and 3D geometric knowledge provides localization supervision for 2D\ndetection images. OV-Uni3DETR achieves the state-of-the-art performance on\nvarious scenarios, surpassing existing methods by more than 6\\% on average. Its\nperformance using only RGB images is on par with or even surpasses that of\nprevious point cloud based methods. Code and pre-trained models will be\nreleased later.\n", "date": "2024-03-28", "categories": []}
{"id": "2406.05785", "title": "A Survey on Text-guided 3D Visual Grounding: Elements, Recent Advances,\n  and Future Directions", "abstract": "  Text-guided 3D visual grounding (T-3DVG), which aims to locate a specific\nobject that semantically corresponds to a language query from a complicated 3D\nscene, has drawn increasing attention in the 3D research community over the\npast few years. Compared to 2D visual grounding, this task presents great\npotential and challenges due to its closer proximity to the real world and the\ncomplexity of data collection and 3D point cloud source processing. In this\nsurvey, we attempt to provide a comprehensive overview of the T-3DVG progress,\nincluding its fundamental elements, recent research advances, and future\nresearch directions. To the best of our knowledge, this is the first systematic\nsurvey on the T-3DVG task. Specifically, we first provide a general structure\nof the T-3DVG pipeline with detailed components in a tutorial style, presenting\na complete background overview. Then, we summarize the existing T-3DVG\napproaches into different categories and analyze their strengths and\nweaknesses. We also present the benchmark datasets and evaluation metrics to\nassess their performances. Finally, we discuss the potential limitations of\nexisting T-3DVG and share some insights on several promising research\ndirections. The latest papers are continually collected at\nhttps://github.com/liudaizong/Awesome-3D-Visual-Grounding.\n", "date": "2024-06-09", "categories": []}
{"id": "2407.15041", "title": "Self-training Room Layout Estimation via Geometry-aware Ray-casting", "abstract": "  In this paper, we introduce a novel geometry-aware self-training framework\nfor room layout estimation models on unseen scenes with unlabeled data. Our\napproach utilizes a ray-casting formulation to aggregate multiple estimates\nfrom different viewing positions, enabling the computation of reliable\npseudo-labels for self-training. In particular, our ray-casting approach\nenforces multi-view consistency along all ray directions and prioritizes\nspatial proximity to the camera view for geometry reasoning. As a result, our\ngeometry-aware pseudo-labels effectively handle complex room geometries and\noccluded walls without relying on assumptions such as Manhattan World or planar\nroom walls. Evaluation on publicly available datasets, including synthetic and\nreal-world scenarios, demonstrates significant improvements in current\nstate-of-the-art layout models without using any human annotation.\n", "date": "2024-07-21", "categories": []}
{"id": "2312.17118", "title": "Fully Sparse 3D Occupancy Prediction", "abstract": "  Occupancy prediction plays a pivotal role in autonomous driving. Previous\nmethods typically construct dense 3D volumes, neglecting the inherent sparsity\nof the scene and suffering from high computational costs. To bridge the gap, we\nintroduce a novel fully sparse occupancy network, termed SparseOcc. SparseOcc\ninitially reconstructs a sparse 3D representation from camera-only inputs and\nsubsequently predicts semantic/instance occupancy from the 3D sparse\nrepresentation by sparse queries. A mask-guided sparse sampling is designed to\nenable sparse queries to interact with 2D features in a fully sparse manner,\nthereby circumventing costly dense features or global attention. Additionally,\nwe design a thoughtful ray-based evaluation metric, namely RayIoU, to solve the\ninconsistency penalty along the depth axis raised in traditional voxel-level\nmIoU criteria. SparseOcc demonstrates its effectiveness by achieving a RayIoU\nof 34.0, while maintaining a real-time inference speed of 17.3 FPS, with 7\nhistory frames inputs. By incorporating more preceding frames to 15, SparseOcc\ncontinuously improves its performance to 35.1 RayIoU without bells and\nwhistles.\n", "date": "2023-12-28", "categories": []}
{"id": "2405.19678", "title": "View-Consistent Hierarchical 3D Segmentation Using Ultrametric Feature\n  Fields", "abstract": "  Large-scale vision foundation models such as Segment Anything (SAM)\ndemonstrate impressive performance in zero-shot image segmentation at multiple\nlevels of granularity. However, these zero-shot predictions are rarely\n3D-consistent. As the camera viewpoint changes in a scene, so do the\nsegmentation predictions, as well as the characterizations of \"coarse\" or\n\"fine\" granularity. In this work, we address the challenging task of lifting\nmulti-granular and view-inconsistent image segmentations into a hierarchical\nand 3D-consistent representation. We learn a novel feature field within a\nNeural Radiance Field (NeRF) representing a 3D scene, whose segmentation\nstructure can be revealed at different scales by simply using different\nthresholds on feature distance. Our key idea is to learn an ultrametric feature\nspace, which unlike a Euclidean space, exhibits transitivity in distance-based\ngrouping, naturally leading to a hierarchical clustering. Put together, our\nmethod takes view-inconsistent multi-granularity 2D segmentations as input and\nproduces a hierarchy of 3D-consistent segmentations as output. We evaluate our\nmethod and several baselines on synthetic datasets with multi-view images and\nmulti-granular segmentation, showcasing improved accuracy and\nviewpoint-consistency. We additionally provide qualitative examples of our\nmodel's 3D hierarchical segmentations in real world scenes. The code and\ndataset are available at https://github.com/hardyho/ultrametric_feature_fields\n", "date": "2024-05-30", "categories": []}
{"id": "2407.13362", "title": "Open Vocabulary 3D Scene Understanding via Geometry Guided\n  Self-Distillation", "abstract": "  The scarcity of large-scale 3D-text paired data poses a great challenge on\nopen vocabulary 3D scene understanding, and hence it is popular to leverage\ninternet-scale 2D data and transfer their open vocabulary capabilities to 3D\nmodels through knowledge distillation. However, the existing distillation-based\n3D scene understanding approaches rely on the representation capacity of 2D\nmodels, disregarding the exploration of geometric priors and inherent\nrepresentational advantages offered by 3D data. In this paper, we propose an\neffective approach, namely Geometry Guided Self-Distillation (GGSD), to learn\nsuperior 3D representations from 2D pre-trained models. Specifically, we first\ndesign a geometry guided distillation module to distill knowledge from 2D\nmodels, and then leverage the 3D geometric priors to alleviate the inherent\nnoise in 2D models and enhance the representation learning process. Due to the\nadvantages of 3D representation, the performance of the distilled 3D student\nmodel can significantly surpass that of the 2D teacher model. This motivates us\nto further leverage the representation advantages of 3D data through\nself-distillation. As a result, our proposed GGSD approach outperforms the\nexisting open vocabulary 3D scene understanding methods by a large margin, as\ndemonstrated by our experiments on both indoor and outdoor benchmark datasets.\n", "date": "2024-07-18", "categories": []}
{"id": "2407.13642", "title": "Open-Vocabulary 3D Semantic Segmentation with Text-to-Image Diffusion\n  Models", "abstract": "  In this paper, we investigate the use of diffusion models which are\npre-trained on large-scale image-caption pairs for open-vocabulary 3D semantic\nunderstanding. We propose a novel method, namely Diff2Scene, which leverages\nfrozen representations from text-image generative models, along with\nsalient-aware and geometric-aware masks, for open-vocabulary 3D semantic\nsegmentation and visual grounding tasks. Diff2Scene gets rid of any labeled 3D\ndata and effectively identifies objects, appearances, materials, locations and\ntheir compositions in 3D scenes. We show that it outperforms competitive\nbaselines and achieves significant improvements over state-of-the-art methods.\nIn particular, Diff2Scene improves the state-of-the-art method on ScanNet200 by\n12%.\n", "date": "2024-07-18", "categories": []}
{"id": "2407.11730", "title": "Monocular Occupancy Prediction for Scalable Indoor Scenes", "abstract": "  Camera-based 3D occupancy prediction has recently garnered increasing\nattention in outdoor driving scenes. However, research in indoor scenes remains\nrelatively unexplored. The core differences in indoor scenes lie in the\ncomplexity of scene scale and the variance in object size. In this paper, we\npropose a novel method, named ISO, for predicting indoor scene occupancy using\nmonocular images. ISO harnesses the advantages of a pretrained depth model to\nachieve accurate depth predictions. Furthermore, we introduce the Dual Feature\nLine of Sight Projection (D-FLoSP) module within ISO, which enhances the\nlearning of 3D voxel features. To foster further research in this domain, we\nintroduce Occ-ScanNet, a large-scale occupancy benchmark for indoor scenes.\nWith a dataset size 40 times larger than the NYUv2 dataset, it facilitates\nfuture scalable research in indoor scene analysis. Experimental results on both\nNYUv2 and Occ-ScanNet demonstrate that our method achieves state-of-the-art\nperformance. The dataset and code are made publicly at\nhttps://github.com/hongxiaoy/ISO.git.\n", "date": "2024-07-16", "categories": []}
{"id": "2407.11793", "title": "Click-Gaussian: Interactive Segmentation to Any 3D Gaussians", "abstract": "  Interactive segmentation of 3D Gaussians opens a great opportunity for\nreal-time manipulation of 3D scenes thanks to the real-time rendering\ncapability of 3D Gaussian Splatting. However, the current methods suffer from\ntime-consuming post-processing to deal with noisy segmentation output. Also,\nthey struggle to provide detailed segmentation, which is important for\nfine-grained manipulation of 3D scenes. In this study, we propose\nClick-Gaussian, which learns distinguishable feature fields of two-level\ngranularity, facilitating segmentation without time-consuming post-processing.\nWe delve into challenges stemming from inconsistently learned feature fields\nresulting from 2D segmentation obtained independently from a 3D scene. 3D\nsegmentation accuracy deteriorates when 2D segmentation results across the\nviews, primary cues for 3D segmentation, are in conflict. To overcome these\nissues, we propose Global Feature-guided Learning (GFL). GFL constructs the\nclusters of global feature candidates from noisy 2D segments across the views,\nwhich smooths out noises when training the features of 3D Gaussians. Our method\nruns in 10 ms per click, 15 to 130 times as fast as the previous methods, while\nalso significantly improving segmentation accuracy. Our project page is\navailable at https://seokhunchoi.github.io/Click-Gaussian\n", "date": "2024-07-16", "categories": []}
{"id": "2407.09781", "title": "Dense Multimodal Alignment for Open-Vocabulary 3D Scene Understanding", "abstract": "  Recent vision-language pre-training models have exhibited remarkable\ngeneralization ability in zero-shot recognition tasks. Previous open-vocabulary\n3D scene understanding methods mostly focus on training 3D models using either\nimage or text supervision while neglecting the collective strength of all\nmodalities. In this work, we propose a Dense Multimodal Alignment (DMA)\nframework to densely co-embed different modalities into a common space for\nmaximizing their synergistic benefits. Instead of extracting coarse view- or\nregion-level text prompts, we leverage large vision-language models to extract\ncomplete category information and scalable scene descriptions to build the text\nmodality, and take image modality as the bridge to build dense point-pixel-text\nassociations. Besides, in order to enhance the generalization ability of the 2D\nmodel for downstream 3D tasks without compromising the open-vocabulary\ncapability, we employ a dual-path integration approach to combine frozen CLIP\nvisual features and learnable mask features. Extensive experiments show that\nour DMA method produces highly competitive open-vocabulary segmentation\nperformance on various indoor and outdoor tasks.\n", "date": "2024-07-13", "categories": []}
{"id": "2407.10084", "title": "Part2Object: Hierarchical Unsupervised 3D Instance Segmentation", "abstract": "  Unsupervised 3D instance segmentation aims to segment objects from a 3D point\ncloud without any annotations. Existing methods face the challenge of either\ntoo loose or too tight clustering, leading to under-segmentation or\nover-segmentation. To address this issue, we propose Part2Object, hierarchical\nclustering with object guidance. Part2Object employs multi-layer clustering\nfrom points to object parts and objects, allowing objects to manifest at any\nlayer. Additionally, it extracts and utilizes 3D objectness priors from\ntemporally consecutive 2D RGB frames to guide the clustering process. Moreover,\nwe propose Hi-Mask3D to support hierarchical 3D object part and instance\nsegmentation. By training Hi-Mask3D on the objects and object parts extracted\nfrom Part2Object, we achieve consistent and superior performance compared to\nstate-of-the-art models in various settings, including unsupervised instance\nsegmentation, data-efficient fine-tuning, and cross-dataset generalization.\nCode is release at https://github.com/ChengShiest/Part2Object\n", "date": "2024-07-14", "categories": []}
{"id": "2407.10200", "title": "Shape2Scene: 3D Scene Representation Learning Through Pre-training on\n  Shape Data", "abstract": "  Current 3D self-supervised learning methods of 3D scenes face a data desert\nissue, resulting from the time-consuming and expensive collecting process of 3D\nscene data. Conversely, 3D shape datasets are easier to collect. Despite this,\nexisting pre-training strategies on shape data offer limited potential for 3D\nscene understanding due to significant disparities in point quantities. To\ntackle these challenges, we propose Shape2Scene (S2S), a novel method that\nlearns representations of large-scale 3D scenes from 3D shape data. We first\ndesign multiscale and high-resolution backbones for shape and scene level 3D\ntasks, i.e., MH-P (point-based) and MH-V (voxel-based). MH-P/V establishes\ndirect paths to highresolution features that capture deep semantic information\nacross multiple scales. This pivotal nature makes them suitable for a wide\nrange of 3D downstream tasks that tightly rely on high-resolution features. We\nthen employ a Shape-to-Scene strategy (S2SS) to amalgamate points from various\nshapes, creating a random pseudo scene (comprising multiple objects) for\ntraining data, mitigating disparities between shapes and scenes. Finally, a\npoint-point contrastive loss (PPC) is applied for the pre-training of MH-P/V.\nIn PPC, the inherent correspondence (i.e., point pairs) is naturally obtained\nin S2SS. Extensive experiments have demonstrated the transferability of 3D\nrepresentations learned by MH-P/V across shape-level and scene-level 3D tasks.\nMH-P achieves notable performance on well-known point cloud datasets (93.8% OA\non ScanObjectNN and 87.6% instance mIoU on ShapeNetPart). MH-V also achieves\npromising performance in 3D semantic segmentation and 3D object detection.\n", "date": "2024-07-14", "categories": []}
{"id": "2407.10687", "title": "FRI-Net: Floorplan Reconstruction via Room-wise Implicit Representation", "abstract": "  In this paper, we introduce a novel method called FRI-Net for 2D floorplan\nreconstruction from 3D point cloud. Existing methods typically rely on corner\nregression or box regression, which lack consideration for the global shapes of\nrooms. To address these issues, we propose a novel approach using a room-wise\nimplicit representation with structural regularization to characterize the\nshapes of rooms in floorplans. By incorporating geometric priors of room\nlayouts in floorplans into our training strategy, the generated room polygons\nare more geometrically regular. We have conducted experiments on two\nchallenging datasets, Structured3D and SceneCAD. Our method demonstrates\nimproved performance compared to state-of-the-art methods, validating the\neffectiveness of our proposed representation for floorplan reconstruction.\n", "date": "2024-07-15", "categories": []}
{"id": "2407.10743", "title": "Scaling 3D Reasoning with LMMs to Large Robot Mission Environments Using\n  Datagraphs", "abstract": "  This paper addresses the challenge of scaling Large Multimodal Models (LMMs)\nto expansive 3D environments. Solving this open problem is especially relevant\nfor robot deployment in many first-responder scenarios, such as\nsearch-and-rescue missions that cover vast spaces. The use of LMMs in these\nsettings is currently hampered by the strict context windows that limit the\nLMM's input size. We therefore introduce a novel approach that utilizes a\ndatagraph structure, which allows the LMM to iteratively query smaller sections\nof a large environment. Using the datagraph in conjunction with graph traversal\nalgorithms, we can prioritize the most relevant locations to the query, thereby\nimproving the scalability of 3D scene language tasks. We illustrate the\ndatagraph using 3D scenes, but these can be easily substituted by other dense\nmodalities that represent the environment, such as pointclouds or Gaussian\nsplats. We demonstrate the potential to use the datagraph for two 3D scene\nlanguage task use cases, in a search-and-rescue mission example.\n", "date": "2024-07-15", "categories": []}
{"id": "2403.13556", "title": "Find n' Propagate: Open-Vocabulary 3D Object Detection in Urban\n  Environments", "abstract": "  In this work, we tackle the limitations of current LiDAR-based 3D object\ndetection systems, which are hindered by a restricted class vocabulary and the\nhigh costs associated with annotating new object classes. Our exploration of\nopen-vocabulary (OV) learning in urban environments aims to capture novel\ninstances using pre-trained vision-language models (VLMs) with multi-sensor\ndata. We design and benchmark a set of four potential solutions as baselines,\ncategorizing them into either top-down or bottom-up approaches based on their\ninput data strategies. While effective, these methods exhibit certain\nlimitations, such as missing novel objects in 3D box estimation or applying\nrigorous priors, leading to biases towards objects near the camera or of\nrectangular geometries. To overcome these limitations, we introduce a universal\n\\textsc{Find n' Propagate} approach for 3D OV tasks, aimed at maximizing the\nrecall of novel objects and propagating this detection capability to more\ndistant areas thereby progressively capturing more. In particular, we utilize a\ngreedy box seeker to search against 3D novel boxes of varying orientations and\ndepth in each generated frustum and ensure the reliability of newly identified\nboxes by cross alignment and density ranker. Additionally, the inherent bias\ntowards camera-proximal objects is alleviated by the proposed remote simulator,\nwhich randomly diversifies pseudo-labeled novel instances in the self-training\nprocess, combined with the fusion of base samples in the memory bank. Extensive\nexperiments demonstrate a 53% improvement in novel recall across diverse OV\nsettings, VLMs, and 3D detectors. Notably, we achieve up to a 3.97-fold\nincrease in Average Precision (AP) for novel object classes. The source code is\nmade available at https://github.com/djamahl99/findnpropagate.\n", "date": "2024-03-20", "categories": []}
{"id": "2405.04299", "title": "ViewFormer: Exploring Spatiotemporal Modeling for Multi-View 3D\n  Occupancy Perception via View-Guided Transformers", "abstract": "  3D occupancy, an advanced perception technology for driving scenarios,\nrepresents the entire scene without distinguishing between foreground and\nbackground by quantifying the physical space into a grid map. The widely\nadopted projection-first deformable attention, efficient in transforming image\nfeatures into 3D representations, encounters challenges in aggregating\nmulti-view features due to sensor deployment constraints. To address this\nissue, we propose our learning-first view attention mechanism for effective\nmulti-view feature aggregation. Moreover, we showcase the scalability of our\nview attention across diverse multi-view 3D tasks, including map construction\nand 3D object detection. Leveraging the proposed view attention as well as an\nadditional multi-frame streaming temporal attention, we introduce ViewFormer, a\nvision-centric transformer-based framework for spatiotemporal feature\naggregation. To further explore occupancy-level flow representation, we present\nFlowOcc3D, a benchmark built on top of existing high-quality datasets.\nQualitative and quantitative analyses on this benchmark reveal the potential to\nrepresent fine-grained dynamic scenes. Extensive experiments show that our\napproach significantly outperforms prior state-of-the-art methods. The codes\nare available at \\url{https://github.com/ViewFormerOcc/ViewFormer-Occ}.\n", "date": "2024-05-07", "categories": []}
{"id": "2407.04504", "title": "Segment Any 4D Gaussians", "abstract": "  Modeling, understanding, and reconstructing the real world are crucial in\nXR/VR. Recently, 3D Gaussian Splatting (3D-GS) methods have shown remarkable\nsuccess in modeling and understanding 3D scenes. Similarly, various 4D\nrepresentations have demonstrated the ability to capture the dynamics of the 4D\nworld. However, there is a dearth of research focusing on segmentation within\n4D representations. In this paper, we propose Segment Any 4D Gaussians (SA4D),\none of the first frameworks to segment anything in the 4D digital world based\non 4D Gaussians. In SA4D, an efficient temporal identity feature field is\nintroduced to handle Gaussian drifting, with the potential to learn precise\nidentity features from noisy and sparse input. Additionally, a 4D segmentation\nrefinement process is proposed to remove artifacts. Our SA4D achieves precise,\nhigh-quality segmentation within seconds in 4D Gaussians and shows the ability\nto remove, recolor, compose, and render high-quality anything masks. More demos\nare available at: https://jsxzs.github.io/sa4d/.\n", "date": "2024-07-05", "categories": []}
{"id": "2312.00732", "title": "Gaussian Grouping: Segment and Edit Anything in 3D Scenes", "abstract": "  The recent Gaussian Splatting achieves high-quality and real-time novel-view\nsynthesis of the 3D scenes. However, it is solely concentrated on the\nappearance and geometry modeling, while lacking in fine-grained object-level\nscene understanding. To address this issue, we propose Gaussian Grouping, which\nextends Gaussian Splatting to jointly reconstruct and segment anything in\nopen-world 3D scenes. We augment each Gaussian with a compact Identity\nEncoding, allowing the Gaussians to be grouped according to their object\ninstance or stuff membership in the 3D scene. Instead of resorting to expensive\n3D labels, we supervise the Identity Encodings during the differentiable\nrendering by leveraging the 2D mask predictions by Segment Anything Model\n(SAM), along with introduced 3D spatial consistency regularization. Compared to\nthe implicit NeRF representation, we show that the discrete and grouped 3D\nGaussians can reconstruct, segment and edit anything in 3D with high visual\nquality, fine granularity and efficiency. Based on Gaussian Grouping, we\nfurther propose a local Gaussian Editing scheme, which shows efficacy in\nversatile scene editing applications, including 3D object removal, inpainting,\ncolorization, style transfer and scene recomposition. Our code and models are\nat https://github.com/lkeab/gaussian-grouping.\n", "date": "2023-12-01", "categories": []}
{"id": "2407.05615", "title": "OSN: Infinite Representations of Dynamic 3D Scenes from Monocular Videos", "abstract": "  It has long been challenging to recover the underlying dynamic 3D scene\nrepresentations from a monocular RGB video. Existing works formulate this\nproblem into finding a single most plausible solution by adding various\nconstraints such as depth priors and strong geometry constraints, ignoring the\nfact that there could be infinitely many 3D scene representations corresponding\nto a single dynamic video. In this paper, we aim to learn all plausible 3D\nscene configurations that match the input video, instead of just inferring a\nspecific one. To achieve this ambitious goal, we introduce a new framework,\ncalled OSN. The key to our approach is a simple yet innovative object scale\nnetwork together with a joint optimization module to learn an accurate scale\nrange for every dynamic 3D object. This allows us to sample as many faithful 3D\nscene configurations as possible. Extensive experiments show that our method\nsurpasses all baselines and achieves superior accuracy in dynamic novel view\nsynthesis on multiple synthetic and real-world datasets. Most notably, our\nmethod demonstrates a clear advantage in learning fine-grained 3D scene\ngeometry. Our code and data are available at https://github.com/vLAR-group/OSN\n", "date": "2024-07-08", "categories": []}
{"id": "2407.00609", "title": "ESGNN: Towards Equivariant Scene Graph Neural Network for 3D Scene\n  Understanding", "abstract": "  Scene graphs have been proven to be useful for various scene understanding\ntasks due to their compact and explicit nature. However, existing approaches\noften neglect the importance of maintaining the symmetry-preserving property\nwhen generating scene graphs from 3D point clouds. This oversight can diminish\nthe accuracy and robustness of the resulting scene graphs, especially when\nhandling noisy, multi-view 3D data. This work, to the best of our knowledge, is\nthe first to implement an Equivariant Graph Neural Network in semantic scene\ngraph generation from 3D point clouds for scene understanding. Our proposed\nmethod, ESGNN, outperforms existing state-of-the-art approaches, demonstrating\na significant improvement in scene estimation with faster convergence. ESGNN\ndemands low computational resources and is easy to implement from available\nframeworks, paving the way for real-time applications such as robotics and\ncomputer vision.\n", "date": "2024-06-30", "categories": []}
{"id": "2407.01349", "title": "PanopticRecon: Leverage Open-vocabulary Instance Segmentation for\n  Zero-shot Panoptic Reconstruction", "abstract": "  Panoptic reconstruction is a challenging task in 3D scene understanding.\nHowever, most existing methods heavily rely on pre-trained semantic\nsegmentation models and known 3D object bounding boxes for 3D panoptic\nsegmentation, which is not available for in-the-wild scenes. In this paper, we\npropose a novel zero-shot panoptic reconstruction method from RGB-D images of\nscenes. For zero-shot segmentation, we leverage open-vocabulary instance\nsegmentation, but it has to face partial labeling and instance association\nchallenges. We tackle both challenges by propagating partial labels with the\naid of dense generalized features and building a 3D instance graph for\nassociating 2D instance IDs. Specifically, we exploit partial labels to learn a\nclassifier for generalized semantic features to provide complete labels for\nscenes with dense distilled features. Moreover, we formulate instance\nassociation as a 3D instance graph segmentation problem, allowing us to fully\nutilize the scene geometry prior and all 2D instance masks to infer global\nunique pseudo 3D instance ID. Our method outperforms state-of-the-art methods\non the indoor dataset ScanNet V2 and the outdoor dataset KITTI-360,\ndemonstrating the effectiveness of our graph segmentation method and\nreconstruction network.\n", "date": "2024-07-01", "categories": []}
{"id": "2406.07544", "title": "Situational Awareness Matters in 3D Vision Language Reasoning", "abstract": "  Being able to carry out complicated vision language reasoning tasks in 3D\nspace represents a significant milestone in developing household robots and\nhuman-centered embodied AI. In this work, we demonstrate that a critical and\ndistinct challenge in 3D vision language reasoning is situational awareness,\nwhich incorporates two key components: (1) The autonomous agent grounds its\nself-location based on a language prompt. (2) The agent answers open-ended\nquestions from the perspective of its calculated position. To address this\nchallenge, we introduce SIG3D, an end-to-end Situation-Grounded model for 3D\nvision language reasoning. We tokenize the 3D scene into sparse voxel\nrepresentation and propose a language-grounded situation estimator, followed by\na situated question answering module. Experiments on the SQA3D and ScanQA\ndatasets show that SIG3D outperforms state-of-the-art models in situation\nestimation and question answering by a large margin (e.g., an enhancement of\nover 30% on situation estimation accuracy). Subsequent analysis corroborates\nour architectural design choices, explores the distinct functions of visual and\ntextual tokens, and highlights the importance of situational awareness in the\ndomain of 3D question answering.\n", "date": "2024-06-11", "categories": []}
{"id": "2406.08907", "title": "Dual Attribute-Spatial Relation Alignment for 3D Visual Grounding", "abstract": "  3D visual grounding is an emerging research area dedicated to making\nconnections between the 3D physical world and natural language, which is\ncrucial for achieving embodied intelligence. In this paper, we propose DASANet,\na Dual Attribute-Spatial relation Alignment Network that separately models and\naligns object attributes and spatial relation features between language and 3D\nvision modalities. We decompose both the language and 3D point cloud input into\ntwo separate parts and design a dual-branch attention module to separately\nmodel the decomposed inputs while preserving global context in\nattribute-spatial feature fusion by cross attentions. Our DASANet achieves the\nhighest grounding accuracy 65.1% on the Nr3D dataset, 1.3% higher than the best\ncompetitor. Besides, the visualization of the two branches proves that our\nmethod is efficient and highly interpretable.\n", "date": "2024-06-13", "categories": []}
{"id": "2406.08960", "title": "AirPlanes: Accurate Plane Estimation via 3D-Consistent Embeddings", "abstract": "  Extracting planes from a 3D scene is useful for downstream tasks in robotics\nand augmented reality. In this paper we tackle the problem of estimating the\nplanar surfaces in a scene from posed images. Our first finding is that a\nsurprisingly competitive baseline results from combining popular clustering\nalgorithms with recent improvements in 3D geometry estimation. However, such\npurely geometric methods are understandably oblivious to plane semantics, which\nare crucial to discerning distinct planes. To overcome this limitation, we\npropose a method that predicts multi-view consistent plane embeddings that\ncomplement geometry when clustering points into planes. We show through\nextensive evaluation on the ScanNetV2 dataset that our new method outperforms\nexisting approaches and our strong geometric baseline for the task of plane\nestimation.\n", "date": "2024-06-13", "categories": []}
{"id": "2406.08009", "title": "OpenObj: Open-Vocabulary Object-Level Neural Radiance Fields with\n  Fine-Grained Understanding", "abstract": "  In recent years, there has been a surge of interest in open-vocabulary 3D\nscene reconstruction facilitated by visual language models (VLMs), which\nshowcase remarkable capabilities in open-set retrieval. However, existing\nmethods face some limitations: they either focus on learning point-wise\nfeatures, resulting in blurry semantic understanding, or solely tackle\nobject-level reconstruction, thereby overlooking the intricate details of the\nobject's interior. To address these challenges, we introduce OpenObj, an\ninnovative approach to build open-vocabulary object-level Neural Radiance\nFields (NeRF) with fine-grained understanding. In essence, OpenObj establishes\na robust framework for efficient and watertight scene modeling and\ncomprehension at the object-level. Moreover, we incorporate part-level features\ninto the neural fields, enabling a nuanced representation of object interiors.\nThis approach captures object-level instances while maintaining a fine-grained\nunderstanding. The results on multiple datasets demonstrate that OpenObj\nachieves superior performance in zero-shot semantic segmentation and retrieval\ntasks. Additionally, OpenObj supports real-world robotics tasks at multiple\nscales, including global movement and local manipulation.\n", "date": "2024-06-12", "categories": []}
{"id": "2212.05040", "title": "Cross-Domain Synthetic-to-Real In-the-Wild Depth and Normal Estimation\n  for 3D Scene Understanding", "abstract": "  We present a cross-domain inference technique that learns from synthetic data\nto estimate depth and normals for in-the-wild omnidirectional 3D scenes\nencountered in real-world uncontrolled settings. To this end, we introduce\nUBotNet, an architecture that combines UNet and Bottleneck Transformer elements\nto predict consistent scene normals and depth. We also introduce the\nOmniHorizon synthetic dataset containing 24,335 omnidirectional images that\nrepresent a wide variety of outdoor environments, including buildings, streets,\nand diverse vegetation. This dataset is generated from expansive, lifelike\nvirtual spaces and encompasses dynamic scene elements, such as changing\nlighting conditions, different times of day, pedestrians, and vehicles. Our\nexperiments show that UBotNet achieves significantly improved accuracy in depth\nestimation and normal estimation compared to existing models. Lastly, we\nvalidate cross-domain synthetic-to-real depth and normal estimation on real\noutdoor images using UBotNet trained solely on our synthetic OmniHorizon\ndataset, demonstrating the potential of both the synthetic dataset and the\nproposed network for real-world scene understanding applications.\n", "date": "2022-12-09", "categories": []}
{"id": "2403.19589", "title": "TOD3Cap: Towards 3D Dense Captioning in Outdoor Scenes", "abstract": "  3D dense captioning stands as a cornerstone in achieving a comprehensive\nunderstanding of 3D scenes through natural language. It has recently witnessed\nremarkable achievements, particularly in indoor settings. However, the\nexploration of 3D dense captioning in outdoor scenes is hindered by two major\nchallenges: 1) the domain gap between indoor and outdoor scenes, such as\ndynamics and sparse visual inputs, makes it difficult to directly adapt\nexisting indoor methods; 2) the lack of data with comprehensive box-caption\npair annotations specifically tailored for outdoor scenes. To this end, we\nintroduce the new task of outdoor 3D dense captioning. As input, we assume a\nLiDAR point cloud and a set of RGB images captured by the panoramic camera rig.\nThe expected output is a set of object boxes with captions. To tackle this\ntask, we propose the TOD3Cap network, which leverages the BEV representation to\ngenerate object box proposals and integrates Relation Q-Former with\nLLaMA-Adapter to generate rich captions for these objects. We also introduce\nthe TOD3Cap dataset, the largest one to our knowledge for 3D dense captioning\nin outdoor scenes, which contains 2.3M descriptions of 64.3K outdoor objects\nfrom 850 scenes. Notably, our TOD3Cap network can effectively localize and\ncaption 3D objects in outdoor scenes, which outperforms baseline methods by a\nsignificant margin (+9.6 CiDEr@0.5IoU). Code, data, and models are publicly\navailable at https://github.com/jxbbb/TOD3Cap.\n", "date": "2024-03-28", "categories": []}
{"id": "2302.03640", "title": "SSR-2D: Semantic 3D Scene Reconstruction from 2D Images", "abstract": "  Most deep learning approaches to comprehensive semantic modeling of 3D indoor\nspaces require costly dense annotations in the 3D domain. In this work, we\nexplore a central 3D scene modeling task, namely, semantic scene reconstruction\nwithout using any 3D annotations. The key idea of our approach is to design a\ntrainable model that employs both incomplete 3D reconstructions and their\ncorresponding source RGB-D images, fusing cross-domain features into volumetric\nembeddings to predict complete 3D geometry, color, and semantics with only 2D\nlabeling which can be either manual or machine-generated. Our key technical\ninnovation is to leverage differentiable rendering of color and semantics to\nbridge 2D observations and unknown 3D space, using the observed RGB images and\n2D semantics as supervision, respectively. We additionally develop a learning\npipeline and corresponding method to enable learning from imperfect predicted\n2D labels, which could be additionally acquired by synthesizing in an augmented\nset of virtual training views complementing the original real captures,\nenabling more efficient self-supervision loop for semantics. As a result, our\nend-to-end trainable solution jointly addresses geometry completion,\ncolorization, and semantic mapping from limited RGB-D images, without relying\non any 3D ground-truth information. Our method achieves the state-of-the-art\nperformance of semantic scene completion on two large-scale benchmark datasets\nMatterPort3D and ScanNet, surpasses baselines even with costly 3D annotations\nin predicting both geometry and semantics. To our knowledge, our method is also\nthe first 2D-driven method addressing completion and semantic segmentation of\nreal-world 3D scans simultaneously.\n", "date": "2023-02-07", "categories": []}
{"id": "2405.20364", "title": "Learning 3D Robotics Perception using Inductive Priors", "abstract": "  Recent advances in deep learning have led to a data-centric intelligence i.e.\nartificially intelligent models unlocking the potential to ingest a large\namount of data and be really good at performing digital tasks such as\ntext-to-image generation, machine-human conversation, and image recognition.\nThis thesis covers the topic of learning with structured inductive bias and\npriors to design approaches and algorithms unlocking the potential of\nprinciple-centric intelligence. Prior knowledge (priors for short), often\navailable in terms of past experience as well as assumptions of how the world\nworks, helps the autonomous agent generalize better and adapt their behavior\nbased on past experience. In this thesis, I demonstrate the use of prior\nknowledge in three different robotics perception problems. 1. object-centric 3D\nreconstruction, 2. vision and language for decision-making, and 3. 3D scene\nunderstanding. To solve these challenging problems, I propose various sources\nof prior knowledge including 1. geometry and appearance priors from synthetic\ndata, 2. modularity and semantic map priors and 3. semantic, structural, and\ncontextual priors. I study these priors for solving robotics 3D perception\ntasks and propose ways to efficiently encode them in deep learning models. Some\npriors are used to warm-start the network for transfer learning, others are\nused as hard constraints to restrict the action space of robotics agents. While\nclassical techniques are brittle and fail to generalize to unseen scenarios and\ndata-centric approaches require a large amount of labeled data, this thesis\naims to build intelligent agents which require very-less real-world data or\ndata acquired only from simulation to generalize to highly dynamic and\ncluttered environments in novel simulations (i.e. sim2sim) or real-world unseen\nenvironments (i.e. sim2real) for a holistic scene understanding of the 3D\nworld.\n", "date": "2024-05-30", "categories": []}
{"id": "2405.19321", "title": "DGD: Dynamic 3D Gaussians Distillation", "abstract": "  We tackle the task of learning dynamic 3D semantic radiance fields given a\nsingle monocular video as input. Our learned semantic radiance field captures\nper-point semantics as well as color and geometric properties for a dynamic 3D\nscene, enabling the generation of novel views and their corresponding\nsemantics. This enables the segmentation and tracking of a diverse set of 3D\nsemantic entities, specified using a simple and intuitive interface that\nincludes a user click or a text prompt. To this end, we present DGD, a unified\n3D representation for both the appearance and semantics of a dynamic 3D scene,\nbuilding upon the recently proposed dynamic 3D Gaussians representation. Our\nrepresentation is optimized over time with both color and semantic information.\nKey to our method is the joint optimization of the appearance and semantic\nattributes, which jointly affect the geometric properties of the scene. We\nevaluate our approach in its ability to enable dense semantic 3D object\ntracking and demonstrate high-quality results that are fast to render, for a\ndiverse set of scenes. Our project webpage is available on\nhttps://isaaclabe.github.io/DGD-Website/\n", "date": "2024-05-29", "categories": []}
{"id": "2312.02158", "title": "PaSCo: Urban 3D Panoptic Scene Completion with Uncertainty Awareness", "abstract": "  We propose the task of Panoptic Scene Completion (PSC) which extends the\nrecently popular Semantic Scene Completion (SSC) task with instance-level\ninformation to produce a richer understanding of the 3D scene. Our PSC proposal\nutilizes a hybrid mask-based technique on the non-empty voxels from sparse\nmulti-scale completions. Whereas the SSC literature overlooks uncertainty which\nis critical for robotics applications, we instead propose an efficient\nensembling to estimate both voxel-wise and instance-wise uncertainties along\nPSC. This is achieved by building on a multi-input multi-output (MIMO)\nstrategy, while improving performance and yielding better uncertainty for\nlittle additional compute. Additionally, we introduce a technique to aggregate\npermutation-invariant mask predictions. Our experiments demonstrate that our\nmethod surpasses all baselines in both Panoptic Scene Completion and\nuncertainty estimation on three large-scale autonomous driving datasets. Our\ncode and data are available at https://astra-vision.github.io/PaSCo .\n", "date": "2023-12-04", "categories": []}
{"id": "2309.08504", "title": "OccupancyDETR: Using DETR for Mixed Dense-sparse 3D Occupancy Prediction", "abstract": "  Visual-based 3D semantic occupancy perception is a key technology for\nrobotics, including autonomous vehicles, offering an enhanced understanding of\nthe environment by 3D. This approach, however, typically requires more\ncomputational resources than BEV or 2D methods. We propose a novel 3D semantic\noccupancy perception method, OccupancyDETR, which utilizes a DETR-like object\ndetection, a mixed dense-sparse 3D occupancy decoder. Our approach\ndistinguishes between foreground and background within a scene. Initially,\nforeground objects are detected using the DETR-like object detection.\nSubsequently, queries for both foreground and background objects are fed into\nthe mixed dense-sparse 3D occupancy decoder, performing upsampling in dense and\nsparse methods, respectively. Finally, a MaskFormer is utilized to infer the\nsemantics of the background voxels. Our approach strikes a balance between\nefficiency and accuracy, achieving faster inference times, lower resource\nconsumption, and improved performance for small object detection. We\ndemonstrate the effectiveness of our proposed method on the SemanticKITTI\ndataset, showcasing an mIoU of 14 and a processing speed of 10 FPS, thereby\npresenting a promising solution for real-time 3D semantic occupancy perception.\n", "date": "2023-09-15", "categories": []}
{"id": "2405.07201", "title": "Building a Strong Pre-Training Baseline for Universal 3D Large-Scale\n  Perception", "abstract": "  An effective pre-training framework with universal 3D representations is\nextremely desired in perceiving large-scale dynamic scenes. However,\nestablishing such an ideal framework that is both task-generic and\nlabel-efficient poses a challenge in unifying the representation of the same\nprimitive across diverse scenes. The current contrastive 3D pre-training\nmethods typically follow a frame-level consistency, which focuses on the 2D-3D\nrelationships in each detached image. Such inconsiderate consistency greatly\nhampers the promising path of reaching an universal pre-training framework: (1)\nThe cross-scene semantic self-conflict, i.e., the intense collision between\nprimitive segments of the same semantics from different scenes; (2) Lacking a\nglobally unified bond that pushes the cross-scene semantic consistency into 3D\nrepresentation learning. To address above challenges, we propose a CSC\nframework that puts a scene-level semantic consistency in the heart, bridging\nthe connection of the similar semantic segments across various scenes. To\nachieve this goal, we combine the coherent semantic cues provided by the vision\nfoundation model and the knowledge-rich cross-scene prototypes derived from the\ncomplementary multi-modality information. These allow us to train a universal\n3D pre-training model that facilitates various downstream tasks with less\nfine-tuning efforts. Empirically, we achieve consistent improvements over SOTA\npre-training approaches in semantic segmentation (+1.4% mIoU), object detection\n(+1.0% mAP), and panoptic segmentation (+3.0% PQ) using their task-specific 3D\nnetwork on nuScenes. Code is released at https://github.com/chenhaomingbob/CSC,\nhoping to inspire future research.\n", "date": "2024-05-12", "categories": []}
{"id": "2303.13959", "title": "Bridging Stereo Geometry and BEV Representation with Reliable Mutual\n  Interaction for Semantic Scene Completion", "abstract": "  3D semantic scene completion (SSC) is an ill-posed perception task that\nrequires inferring a dense 3D scene from limited observations. Previous\ncamera-based methods struggle to predict accurate semantic scenes due to\ninherent geometric ambiguity and incomplete observations. In this paper, we\nresort to stereo matching technique and bird's-eye-view (BEV) representation\nlearning to address such issues in SSC. Complementary to each other, stereo\nmatching mitigates geometric ambiguity with epipolar constraint while BEV\nrepresentation enhances the hallucination ability for invisible regions with\nglobal semantic context. However, due to the inherent representation gap\nbetween stereo geometry and BEV features, it is non-trivial to bridge them for\ndense prediction task of SSC. Therefore, we further develop a unified\noccupancy-based framework dubbed BRGScene, which effectively bridges these two\nrepresentations with dense 3D volumes for reliable semantic scene completion.\nSpecifically, we design a novel Mutual Interactive Ensemble (MIE) block for\npixel-level reliable aggregation of stereo geometry and BEV features. Within\nthe MIE block, a Bi-directional Reliable Interaction (BRI) module, enhanced\nwith confidence re-weighting, is employed to encourage fine-grained interaction\nthrough mutual guidance. Besides, a Dual Volume Ensemble (DVE) module is\nintroduced to facilitate complementary aggregation through channel-wise\nrecalibration and multi-group voting. Our method outperforms all published\ncamera-based methods on SemanticKITTI for semantic scene completion. Our code\nis available on https://github.com/Arlo0o/StereoScene.\n", "date": "2023-03-24", "categories": []}
{"id": "2304.00962", "title": "RegionPLC: Regional Point-Language Contrastive Learning for Open-World\n  3D Scene Understanding", "abstract": "  We propose a lightweight and scalable Regional Point-Language Contrastive\nlearning framework, namely \\textbf{RegionPLC}, for open-world 3D scene\nunderstanding, aiming to identify and recognize open-set objects and\ncategories. Specifically, based on our empirical studies, we introduce a\n3D-aware SFusion strategy that fuses 3D vision-language pairs derived from\nmultiple 2D foundation models, yielding high-quality, dense region-level\nlanguage descriptions without human 3D annotations. Subsequently, we devise a\nregion-aware point-discriminative contrastive learning objective to enable\nrobust and effective 3D learning from dense regional language supervision. We\ncarry out extensive experiments on ScanNet, ScanNet200, and nuScenes datasets,\nand our model outperforms prior 3D open-world scene understanding approaches by\nan average of 17.2\\% and 9.1\\% for semantic and instance segmentation,\nrespectively, while maintaining greater scalability and lower resource demands.\nFurthermore, our method has the flexibility to be effortlessly integrated with\nlanguage models to enable open-ended grounded 3D reasoning without extra\ntask-specific training. Code is available at https://github.com/CVMI-Lab/PLA.\n", "date": "2023-04-03", "categories": []}
{"id": "2401.01970", "title": "FMGS: Foundation Model Embedded 3D Gaussian Splatting for Holistic 3D\n  Scene Understanding", "abstract": "  Precisely perceiving the geometric and semantic properties of real-world 3D\nobjects is crucial for the continued evolution of augmented reality and robotic\napplications. To this end, we present Foundation Model Embedded Gaussian\nSplatting (FMGS), which incorporates vision-language embeddings of foundation\nmodels into 3D Gaussian Splatting (GS). The key contribution of this work is an\nefficient method to reconstruct and represent 3D vision-language models. This\nis achieved by distilling feature maps generated from image-based foundation\nmodels into those rendered from our 3D model. To ensure high-quality rendering\nand fast training, we introduce a novel scene representation by integrating\nstrengths from both GS and multi-resolution hash encodings (MHE). Our effective\ntraining procedure also introduces a pixel alignment loss that makes the\nrendered feature distance of the same semantic entities close, following the\npixel-level semantic boundaries. Our results demonstrate remarkable multi-view\nsemantic consistency, facilitating diverse downstream tasks, beating\nstate-of-the-art methods by 10.2 percent on open-vocabulary language-based\nobject detection, despite that we are 851X faster for inference. This research\nexplores the intersection of vision, language, and 3D scene representation,\npaving the way for enhanced scene understanding in uncontrolled real-world\nenvironments. We plan to release the code on the project page.\n", "date": "2024-01-03", "categories": []}
{"id": "2405.03388", "title": "3D LiDAR Mapping in Dynamic Environments Using a 4D Implicit Neural\n  Representation", "abstract": "  Building accurate maps is a key building block to enable reliable\nlocalization, planning, and navigation of autonomous vehicles. We propose a\nnovel approach for building accurate maps of dynamic environments utilizing a\nsequence of LiDAR scans. To this end, we propose encoding the 4D scene into a\nnovel spatio-temporal implicit neural map representation by fitting a\ntime-dependent truncated signed distance function to each point. Using our\nrepresentation, we extract the static map by filtering the dynamic parts. Our\nneural representation is based on sparse feature grids, a globally shared\ndecoder, and time-dependent basis functions, which we jointly optimize in an\nunsupervised fashion. To learn this representation from a sequence of LiDAR\nscans, we design a simple yet efficient loss function to supervise the map\noptimization in a piecewise way. We evaluate our approach on various scenes\ncontaining moving objects in terms of the reconstruction quality of static maps\nand the segmentation of dynamic point clouds. The experimental results\ndemonstrate that our method is capable of removing the dynamic part of the\ninput point clouds while reconstructing accurate and complete 3D maps,\noutperforming several state-of-the-art methods. Codes are available at:\nhttps://github.com/PRBonn/4dNDF\n", "date": "2024-05-06", "categories": []}
{"id": "2306.08068", "title": "DORSal: Diffusion for Object-centric Representations of Scenes et al", "abstract": "  Recent progress in 3D scene understanding enables scalable learning of\nrepresentations across large datasets of diverse scenes. As a consequence,\ngeneralization to unseen scenes and objects, rendering novel views from just a\nsingle or a handful of input images, and controllable scene generation that\nsupports editing, is now possible. However, training jointly on a large number\nof scenes typically compromises rendering quality when compared to single-scene\noptimized models such as NeRFs. In this paper, we leverage recent progress in\ndiffusion models to equip 3D scene representation learning models with the\nability to render high-fidelity novel views, while retaining benefits such as\nobject-level scene editing to a large degree. In particular, we propose DORSal,\nwhich adapts a video diffusion architecture for 3D scene generation conditioned\non frozen object-centric slot-based representations of scenes. On both complex\nsynthetic multi-object scenes and on the real-world large-scale Street View\ndataset, we show that DORSal enables scalable neural rendering of 3D scenes\nwith object-level editing and improves upon existing approaches.\n", "date": "2023-06-13", "categories": []}
{"id": "2303.14541", "title": "UnScene3D: Unsupervised 3D Instance Segmentation for Indoor Scenes", "abstract": "  3D instance segmentation is fundamental to geometric understanding of the\nworld around us. Existing methods for instance segmentation of 3D scenes rely\non supervision from expensive, manual 3D annotations. We propose UnScene3D, the\nfirst fully unsupervised 3D learning approach for class-agnostic 3D instance\nsegmentation of indoor scans. UnScene3D first generates pseudo masks by\nleveraging self-supervised color and geometry features to find potential object\nregions. We operate on a basis of geometric oversegmentation, enabling\nefficient representation and learning on high-resolution 3D data. The coarse\nproposals are then refined through self-training our model on its predictions.\nOur approach improves over state-of-the-art unsupervised 3D instance\nsegmentation methods by more than 300% Average Precision score, demonstrating\neffective instance segmentation even in challenging, cluttered 3D scenes.\n", "date": "2023-03-25", "categories": []}
{"id": "2312.11713", "title": "Indoor and Outdoor 3D Scene Graph Generation via Language-Enabled\n  Spatial Ontologies", "abstract": "  This paper proposes an approach to build 3D scene graphs in arbitrary indoor\nand outdoor environments. Such extension is challenging; the hierarchy of\nconcepts that describe an outdoor environment is more complex than for indoors,\nand manually defining such hierarchy is time-consuming and does not scale.\nFurthermore, the lack of training data prevents the straightforward application\nof learning-based tools used in indoor settings. To address these challenges,\nwe propose two novel extensions. First, we develop methods to build a spatial\nontology defining concepts and relations relevant for indoor and outdoor robot\noperation. In particular, we use a Large Language Model (LLM) to build such an\nontology, thus largely reducing the amount of manual effort required. Second,\nwe leverage the spatial ontology for 3D scene graph construction using Logic\nTensor Networks (LTN) to add logical rules, or axioms (e.g., \"a beach contains\nsand\"), which provide additional supervisory signals at training time thus\nreducing the need for labelled data, providing better predictions, and even\nallowing predicting concepts unseen at training time. We test our approach in a\nvariety of datasets, including indoor, rural, and coastal environments, and\nshow that it leads to a significant increase in the quality of the 3D scene\ngraph generation with sparsely annotated data.\n", "date": "2023-12-18", "categories": []}
{"id": "2404.13842", "title": "On Support Relations Inference and Scene Hierarchy Graph Construction\n  from Point Cloud in Clustered Environments", "abstract": "  Over the years, scene understanding has attracted a growing interest in\ncomputer vision, providing the semantic and physical scene information\nnecessary for robots to complete some particular tasks autonomously. In 3D\nscenes, rich spatial geometric and topological information are often ignored by\nRGB-based approaches for scene understanding. In this study, we develop a\nbottom-up approach for scene understanding that infers support relations\nbetween objects from a point cloud. Our approach utilizes the spatial topology\ninformation of the plane pairs in the scene, consisting of three major steps.\n1) Detection of pairwise spatial configuration: dividing primitive pairs into\nlocal support connection and local inner connection; 2) primitive\nclassification: a combinatorial optimization method applied to classify\nprimitives; and 3) support relations inference and hierarchy graph\nconstruction: bottom-up support relations inference and scene hierarchy graph\nconstruction containing primitive level and object level. Through experiments,\nwe demonstrate that the algorithm achieves excellent performance in primitive\nclassification and support relations inference. Additionally, we show that the\nscene hierarchy graph contains rich geometric and topological information of\nobjects, and it possesses great scalability for scene understanding.\n", "date": "2024-04-22", "categories": []}
{"id": "2401.11395", "title": "UniM-OV3D: Uni-Modality Open-Vocabulary 3D Scene Understanding with\n  Fine-Grained Feature Representation", "abstract": "  3D open-vocabulary scene understanding aims to recognize arbitrary novel\ncategories beyond the base label space. However, existing works not only fail\nto fully utilize all the available modal information in the 3D domain but also\nlack sufficient granularity in representing the features of each modality. In\nthis paper, we propose a unified multimodal 3D open-vocabulary scene\nunderstanding network, namely UniM-OV3D, which aligns point clouds with image,\nlanguage and depth. To better integrate global and local features of the point\nclouds, we design a hierarchical point cloud feature extraction module that\nlearns comprehensive fine-grained feature representations. Further, to\nfacilitate the learning of coarse-to-fine point-semantic representations from\ncaptions, we propose the utilization of hierarchical 3D caption pairs,\ncapitalizing on geometric constraints across various viewpoints of 3D scenes.\nExtensive experimental results demonstrate the effectiveness and superiority of\nour method in open-vocabulary semantic and instance segmentation, which\nachieves state-of-the-art performance on both indoor and outdoor benchmarks\nsuch as ScanNet, ScanNet200, S3IDS and nuScenes. Code is available at\nhttps://github.com/hithqd/UniM-OV3D.\n", "date": "2024-01-21", "categories": []}
{"id": "2404.12784", "title": "Contrastive Gaussian Clustering: Weakly Supervised 3D Scene Segmentation", "abstract": "  We introduce Contrastive Gaussian Clustering, a novel approach capable of\nprovide segmentation masks from any viewpoint and of enabling 3D segmentation\nof the scene. Recent works in novel-view synthesis have shown how to model the\nappearance of a scene via a cloud of 3D Gaussians, and how to generate accurate\nimages from a given viewpoint by projecting on it the Gaussians before $\\alpha$\nblending their color. Following this example, we train a model to include also\na segmentation feature vector for each Gaussian. These can then be used for 3D\nscene segmentation, by clustering Gaussians according to their feature vectors;\nand to generate 2D segmentation masks, by projecting the Gaussians on a plane\nand $\\alpha$ blending over their segmentation features. Using a combination of\ncontrastive learning and spatial regularization, our method can be trained on\ninconsistent 2D segmentation masks, and still learn to generate segmentation\nmasks consistent across all views. Moreover, the resulting model is extremely\naccurate, improving the IoU accuracy of the predicted masks by $+8\\%$ over the\nstate of the art. Code and trained models will be released soon.\n", "date": "2024-04-19", "categories": []}
{"id": "2404.13044", "title": "Unified Scene Representation and Reconstruction for 3D Large Language\n  Models", "abstract": "  Enabling Large Language Models (LLMs) to interact with 3D environments is\nchallenging. Existing approaches extract point clouds either from ground truth\n(GT) geometry or 3D scenes reconstructed by auxiliary models. Text-image\naligned 2D features from CLIP are then lifted to point clouds, which serve as\ninputs for LLMs. However, this solution lacks the establishment of 3D\npoint-to-point connections, leading to a deficiency of spatial structure\ninformation. Concurrently, the absence of integration and unification between\nthe geometric and semantic representations of the scene culminates in a\ndiminished level of 3D scene understanding. In this paper, we demonstrate the\nimportance of having a unified scene representation and reconstruction\nframework, which is essential for LLMs in 3D scenes. Specifically, we introduce\nUni3DR^2 extracts 3D geometric and semantic aware representation features via\nthe frozen pre-trained 2D foundation models (e.g., CLIP and SAM) and a\nmulti-scale aggregate 3D decoder. Our learned 3D representations not only\ncontribute to the reconstruction process but also provide valuable knowledge\nfor LLMs. Experimental results validate that our Uni3DR^2 yields convincing\ngains over the baseline on the 3D reconstruction dataset ScanNet (increasing\nF-Score by +1.8\\%). When applied to LLMs, our Uni3DR^2-LLM exhibits superior\nperformance over the baseline on the 3D vision-language understanding dataset\nScanQA (increasing BLEU-1 by +4.0\\% and +4.2\\% on the val set and test set,\nrespectively). Furthermore, it outperforms the state-of-the-art method that\nuses additional GT point clouds on both ScanQA and 3DMV-VQA.\n", "date": "2024-04-19", "categories": []}
{"id": "2404.11958", "title": "Not All Voxels Are Equal: Hardness-Aware Semantic Scene Completion with\n  Self-Distillation", "abstract": "  Semantic scene completion, also known as semantic occupancy prediction, can\nprovide dense geometric and semantic information for autonomous vehicles, which\nattracts the increasing attention of both academia and industry. Unfortunately,\nexisting methods usually formulate this task as a voxel-wise classification\nproblem and treat each voxel equally in 3D space during training. As the hard\nvoxels have not been paid enough attention, the performance in some challenging\nregions is limited. The 3D dense space typically contains a large number of\nempty voxels, which are easy to learn but require amounts of computation due to\nhandling all the voxels uniformly for the existing models. Furthermore, the\nvoxels in the boundary region are more challenging to differentiate than those\nin the interior. In this paper, we propose HASSC approach to train the semantic\nscene completion model with hardness-aware design. The global hardness from the\nnetwork optimization process is defined for dynamical hard voxel selection.\nThen, the local hardness with geometric anisotropy is adopted for voxel-wise\nrefinement. Besides, self-distillation strategy is introduced to make training\nprocess stable and consistent. Extensive experiments show that our HASSC scheme\ncan effectively promote the accuracy of the baseline model without incurring\nthe extra inference cost. Source code is available at:\nhttps://github.com/songw-zju/HASSC.\n", "date": "2024-04-18", "categories": []}
{"id": "2312.05889", "title": "SuperPrimitive: Scene Reconstruction at a Primitive Level", "abstract": "  Joint camera pose and dense geometry estimation from a set of images or a\nmonocular video remains a challenging problem due to its computational\ncomplexity and inherent visual ambiguities. Most dense incremental\nreconstruction systems operate directly on image pixels and solve for their 3D\npositions using multi-view geometry cues. Such pixel-level approaches suffer\nfrom ambiguities or violations of multi-view consistency (e.g. caused by\ntextureless or specular surfaces).\n  We address this issue with a new image representation which we call a\nSuperPrimitive. SuperPrimitives are obtained by splitting images into\nsemantically correlated local regions and enhancing them with estimated surface\nnormal directions, both of which are predicted by state-of-the-art single image\nneural networks. This provides a local geometry estimate per SuperPrimitive,\nwhile their relative positions are adjusted based on multi-view observations.\n  We demonstrate the versatility of our new representation by addressing three\n3D reconstruction tasks: depth completion, few-view structure from motion, and\nmonocular dense visual odometry.\n", "date": "2023-12-10", "categories": []}
{"id": "2304.12308", "title": "Segment Anything in 3D with Radiance Fields", "abstract": "  The Segment Anything Model (SAM) emerges as a powerful vision foundation\nmodel to generate high-quality 2D segmentation results. This paper aims to\ngeneralize SAM to segment 3D objects. Rather than replicating the data\nacquisition and annotation procedure which is costly in 3D, we design an\nefficient solution, leveraging the radiance field as a cheap and off-the-shelf\nprior that connects multi-view 2D images to the 3D space. We refer to the\nproposed solution as SA3D, short for Segment Anything in 3D. With SA3D, the\nuser is only required to provide a 2D segmentation prompt (e.g., rough points)\nfor the target object in a single view, which is used to generate its\ncorresponding 2D mask with SAM. Next, SA3D alternately performs mask inverse\nrendering and cross-view self-prompting across various views to iteratively\nrefine the 3D mask of the target object. For one view, mask inverse rendering\nprojects the 2D mask obtained by SAM into the 3D space with guidance of the\ndensity distribution learned by the radiance field for 3D mask refinement;\nThen, cross-view self-prompting extracts reliable prompts automatically as the\ninput to SAM from the rendered 2D mask of the inaccurate 3D mask for a new\nview. We show in experiments that SA3D adapts to various scenes and achieves 3D\nsegmentation within seconds. Our research reveals a potential methodology to\nlift the ability of a 2D segmentation model to 3D. Our code is available at\nhttps://github.com/Jumpat/SegmentAnythingin3D.\n", "date": "2023-04-24", "categories": []}
{"id": "2312.02244", "title": "Geometrically-driven Aggregation for Zero-shot 3D Point Cloud\n  Understanding", "abstract": "  Zero-shot 3D point cloud understanding can be achieved via 2D Vision-Language\nModels (VLMs). Existing strategies directly map Vision-Language Models from 2D\npixels of rendered or captured views to 3D points, overlooking the inherent and\nexpressible point cloud geometric structure. Geometrically similar or close\nregions can be exploited for bolstering point cloud understanding as they are\nlikely to share semantic information. To this end, we introduce the first\ntraining-free aggregation technique that leverages the point cloud's 3D\ngeometric structure to improve the quality of the transferred Vision-Language\nModels. Our approach operates iteratively, performing local-to-global\naggregation based on geometric and semantic point-level reasoning. We benchmark\nour approach on three downstream tasks, including classification, part\nsegmentation, and semantic segmentation, with a variety of datasets\nrepresenting both synthetic/real-world, and indoor/outdoor scenarios. Our\napproach achieves new state-of-the-art results in all benchmarks. Our approach\noperates iteratively, performing local-to-global aggregation based on geometric\nand semantic point-level reasoning. Code and dataset are available at\nhttps://luigiriz.github.io/geoze-website/\n", "date": "2023-12-04", "categories": []}
{"id": "2404.09993", "title": "No More Ambiguity in 360{\\deg} Room Layout via Bi-Layout Estimation", "abstract": "  Inherent ambiguity in layout annotations poses significant challenges to\ndeveloping accurate 360{\\deg} room layout estimation models. To address this\nissue, we propose a novel Bi-Layout model capable of predicting two distinct\nlayout types. One stops at ambiguous regions, while the other extends to\nencompass all visible areas. Our model employs two global context embeddings,\nwhere each embedding is designed to capture specific contextual information for\neach layout type. With our novel feature guidance module, the image feature\nretrieves relevant context from these embeddings, generating layout-aware\nfeatures for precise bi-layout predictions. A unique property of our Bi-Layout\nmodel is its ability to inherently detect ambiguous regions by comparing the\ntwo predictions. To circumvent the need for manual correction of ambiguous\nannotations during testing, we also introduce a new metric for disambiguating\nground truth layouts. Our method demonstrates superior performance on benchmark\ndatasets, notably outperforming leading approaches. Specifically, on the\nMatterportLayout dataset, it improves 3DIoU from 81.70% to 82.57% across the\nfull test set and notably from 54.80% to 59.97% in subsets with significant\nambiguity. Project page: https://liagm.github.io/Bi_Layout/\n", "date": "2024-04-15", "categories": []}
{"id": "2404.07933", "title": "Boosting Self-Supervision for Single-View Scene Completion via Knowledge\n  Distillation", "abstract": "  Inferring scene geometry from images via Structure from Motion is a\nlong-standing and fundamental problem in computer vision. While classical\napproaches and, more recently, depth map predictions only focus on the visible\nparts of a scene, the task of scene completion aims to reason about geometry\neven in occluded regions. With the popularity of neural radiance fields\n(NeRFs), implicit representations also became popular for scene completion by\npredicting so-called density fields. Unlike explicit approaches. e.g.\nvoxel-based methods, density fields also allow for accurate depth prediction\nand novel-view synthesis via image-based rendering. In this work, we propose to\nfuse the scene reconstruction from multiple images and distill this knowledge\ninto a more accurate single-view scene reconstruction. To this end, we propose\nMulti-View Behind the Scenes (MVBTS) to fuse density fields from multiple posed\nimages, trained fully self-supervised only from image data. Using knowledge\ndistillation, we use MVBTS to train a single-view scene completion network via\ndirect supervision called KDBTS. It achieves state-of-the-art performance on\noccupancy prediction, especially in occluded regions.\n", "date": "2024-04-11", "categories": []}
{"id": "2311.06694", "title": "Which One? Leveraging Context Between Objects and Multiple Views for\n  Language Grounding", "abstract": "  When connecting objects and their language referents in an embodied 3D\nenvironment, it is important to note that: (1) an object can be better\ncharacterized by leveraging comparative information between itself and other\nobjects, and (2) an object's appearance can vary with camera position. As such,\nwe present the Multi-view Approach to Grounding in Context (MAGiC), which\nselects an object referent based on language that distinguishes between two\nsimilar objects. By pragmatically reasoning over both objects and across\nmultiple views of those objects, MAGiC improves over the state-of-the-art model\non the SNARE object reference task with a relative error reduction of 12.9\\%\n(representing an absolute improvement of 2.7\\%). Ablation studies show that\nreasoning jointly over object referent candidates and multiple views of each\nobject both contribute to improved accuracy. Code:\nhttps://github.com/rcorona/magic_snare/\n", "date": "2023-11-12", "categories": []}
{"id": "2403.04198", "title": "CN-RMA: Combined Network with Ray Marching Aggregation for 3D Indoors\n  Object Detection from Multi-view Images", "abstract": "  This paper introduces CN-RMA, a novel approach for 3D indoor object detection\nfrom multi-view images. We observe the key challenge as the ambiguity of image\nand 3D correspondence without explicit geometry to provide occlusion\ninformation. To address this issue, CN-RMA leverages the synergy of 3D\nreconstruction networks and 3D object detection networks, where the\nreconstruction network provides a rough Truncated Signed Distance Function\n(TSDF) and guides image features to vote to 3D space correctly in an end-to-end\nmanner. Specifically, we associate weights to sampled points of each ray\nthrough ray marching, representing the contribution of a pixel in an image to\ncorresponding 3D locations. Such weights are determined by the predicted signed\ndistances so that image features vote only to regions near the reconstructed\nsurface. Our method achieves state-of-the-art performance in 3D object\ndetection from multi-view images, as measured by mAP@0.25 and mAP@0.5 on the\nScanNet and ARKitScenes datasets. The code and models are released at\nhttps://github.com/SerCharles/CN-RMA.\n", "date": "2024-03-07", "categories": []}
{"id": "2311.11863", "title": "GP-NeRF: Generalized Perception NeRF for Context-Aware 3D Scene\n  Understanding", "abstract": "  Applying NeRF to downstream perception tasks for scene understanding and\nrepresentation is becoming increasingly popular. Most existing methods treat\nsemantic prediction as an additional rendering task, \\textit{i.e.}, the \"label\nrendering\" task, to build semantic NeRFs. However, by rendering\nsemantic/instance labels per pixel without considering the contextual\ninformation of the rendered image, these methods usually suffer from unclear\nboundary segmentation and abnormal segmentation of pixels within an object. To\nsolve this problem, we propose Generalized Perception NeRF (GP-NeRF), a novel\npipeline that makes the widely used segmentation model and NeRF work compatibly\nunder a unified framework, for facilitating context-aware 3D scene perception.\nTo accomplish this goal, we introduce transformers to aggregate radiance as\nwell as semantic embedding fields jointly for novel views and facilitate the\njoint volumetric rendering of both fields. In addition, we propose two\nself-distillation mechanisms, i.e., the Semantic Distill Loss and the\nDepth-Guided Semantic Distill Loss, to enhance the discrimination and quality\nof the semantic field and the maintenance of geometric consistency. In\nevaluation, we conduct experimental comparisons under two perception tasks\n(\\textit{i.e.} semantic and instance segmentation) using both synthetic and\nreal-world datasets. Notably, our method outperforms SOTA approaches by 6.94\\%,\n11.76\\%, and 8.47\\% on generalized semantic segmentation, finetuning semantic\nsegmentation, and instance segmentation, respectively.\n", "date": "2023-11-20", "categories": []}
{"id": "2312.01531", "title": "SANeRF-HQ: Segment Anything for NeRF in High Quality", "abstract": "  Recently, the Segment Anything Model (SAM) has showcased remarkable\ncapabilities of zero-shot segmentation, while NeRF (Neural Radiance Fields) has\ngained popularity as a method for various 3D problems beyond novel view\nsynthesis. Though there exist initial attempts to incorporate these two methods\ninto 3D segmentation, they face the challenge of accurately and consistently\nsegmenting objects in complex scenarios. In this paper, we introduce the\nSegment Anything for NeRF in High Quality (SANeRF-HQ) to achieve high-quality\n3D segmentation of any target object in a given scene. SANeRF-HQ utilizes SAM\nfor open-world object segmentation guided by user-supplied prompts, while\nleveraging NeRF to aggregate information from different viewpoints. To overcome\nthe aforementioned challenges, we employ density field and RGB similarity to\nenhance the accuracy of segmentation boundary during the aggregation.\nEmphasizing on segmentation accuracy, we evaluate our method on multiple NeRF\ndatasets where high-quality ground-truths are available or manually annotated.\nSANeRF-HQ shows a significant quality improvement over state-of-the-art methods\nin NeRF object segmentation, provides higher flexibility for object\nlocalization, and enables more consistent object segmentation across multiple\nviews. Results and code are available at the project site:\nhttps://lyclyc52.github.io/SANeRF-HQ/.\n", "date": "2023-12-03", "categories": []}
{"id": "2404.05163", "title": "Semantic Flow: Learning Semantic Field of Dynamic Scenes from Monocular\n  Videos", "abstract": "  In this work, we pioneer Semantic Flow, a neural semantic representation of\ndynamic scenes from monocular videos. In contrast to previous NeRF methods that\nreconstruct dynamic scenes from the colors and volume densities of individual\npoints, Semantic Flow learns semantics from continuous flows that contain rich\n3D motion information. As there is 2D-to-3D ambiguity problem in the viewing\ndirection when extracting 3D flow features from 2D video frames, we consider\nthe volume densities as opacity priors that describe the contributions of flow\nfeatures to the semantics on the frames. More specifically, we first learn a\nflow network to predict flows in the dynamic scene, and propose a flow feature\naggregation module to extract flow features from video frames. Then, we propose\na flow attention module to extract motion information from flow features, which\nis followed by a semantic network to output semantic logits of flows. We\nintegrate the logits with volume densities in the viewing direction to\nsupervise the flow features with semantic labels on video frames. Experimental\nresults show that our model is able to learn from multiple dynamic scenes and\nsupports a series of new tasks such as instance-level scene editing, semantic\ncompletions, dynamic scene tracking and semantic adaption on novel scenes.\nCodes are available at https://github.com/tianfr/Semantic-Flow/.\n", "date": "2024-04-08", "categories": []}
{"id": "2404.04050", "title": "No Time to Train: Empowering Non-Parametric Networks for Few-shot 3D\n  Scene Segmentation", "abstract": "  To reduce the reliance on large-scale datasets, recent works in 3D\nsegmentation resort to few-shot learning. Current 3D few-shot segmentation\nmethods first pre-train models on 'seen' classes, and then evaluate their\ngeneralization performance on 'unseen' classes. However, the prior pre-training\nstage not only introduces excessive time overhead but also incurs a significant\ndomain gap on 'unseen' classes. To tackle these issues, we propose a\nNon-parametric Network for few-shot 3D Segmentation, Seg-NN, and its Parametric\nvariant, Seg-PN. Without training, Seg-NN extracts dense representations by\nhand-crafted filters and achieves comparable performance to existing parametric\nmodels. Due to the elimination of pre-training, Seg-NN can alleviate the domain\ngap issue and save a substantial amount of time. Based on Seg-NN, Seg-PN only\nrequires training a lightweight QUEry-Support Transferring (QUEST) module,\nwhich enhances the interaction between the support set and query set.\nExperiments suggest that Seg-PN outperforms previous state-of-the-art method by\n+4.19% and +7.71% mIoU on S3DIS and ScanNet datasets respectively, while\nreducing training time by -90%, indicating its effectiveness and efficiency.\n", "date": "2024-04-05", "categories": []}
{"id": "2404.03650", "title": "OpenNeRF: Open Set 3D Neural Scene Segmentation with Pixel-Wise Features\n  and Rendered Novel Views", "abstract": "  Large visual-language models (VLMs), like CLIP, enable open-set image\nsegmentation to segment arbitrary concepts from an image in a zero-shot manner.\nThis goes beyond the traditional closed-set assumption, i.e., where models can\nonly segment classes from a pre-defined training set. More recently, first\nworks on open-set segmentation in 3D scenes have appeared in the literature.\nThese methods are heavily influenced by closed-set 3D convolutional approaches\nthat process point clouds or polygon meshes. However, these 3D scene\nrepresentations do not align well with the image-based nature of the\nvisual-language models. Indeed, point cloud and 3D meshes typically have a\nlower resolution than images and the reconstructed 3D scene geometry might not\nproject well to the underlying 2D image sequences used to compute pixel-aligned\nCLIP features. To address these challenges, we propose OpenNeRF which naturally\noperates on posed images and directly encodes the VLM features within the NeRF.\nThis is similar in spirit to LERF, however our work shows that using pixel-wise\nVLM features (instead of global CLIP features) results in an overall less\ncomplex architecture without the need for additional DINO regularization. Our\nOpenNeRF further leverages NeRF's ability to render novel views and extract\nopen-set VLM features from areas that are not well observed in the initial\nposed images. For 3D point cloud segmentation on the Replica dataset, OpenNeRF\noutperforms recent open-vocabulary methods such as LERF and OpenScene by at\nleast +4.9 mIoU.\n", "date": "2024-04-04", "categories": []}
{"id": "2404.03658", "title": "Know Your Neighbors: Improving Single-View Reconstruction via Spatial\n  Vision-Language Reasoning", "abstract": "  Recovering the 3D scene geometry from a single view is a fundamental yet\nill-posed problem in computer vision. While classical depth estimation methods\ninfer only a 2.5D scene representation limited to the image plane, recent\napproaches based on radiance fields reconstruct a full 3D representation.\nHowever, these methods still struggle with occluded regions since inferring\ngeometry without visual observation requires (i) semantic knowledge of the\nsurroundings, and (ii) reasoning about spatial context. We propose KYN, a novel\nmethod for single-view scene reconstruction that reasons about semantic and\nspatial context to predict each point's density. We introduce a vision-language\nmodulation module to enrich point features with fine-grained semantic\ninformation. We aggregate point representations across the scene through a\nlanguage-guided spatial attention mechanism to yield per-point density\npredictions aware of the 3D semantic context. We show that KYN improves 3D\nshape recovery compared to predicting density for each 3D point in isolation.\nWe achieve state-of-the-art results in scene and object reconstruction on\nKITTI-360, and show improved zero-shot generalization compared to prior work.\nProject page: https://ruili3.github.io/kyn.\n", "date": "2024-04-04", "categories": []}
{"id": "2404.02527", "title": "Weakly-Supervised 3D Scene Graph Generation via Visual-Linguistic\n  Assisted Pseudo-labeling", "abstract": "  Learning to build 3D scene graphs is essential for real-world perception in a\nstructured and rich fashion. However, previous 3D scene graph generation\nmethods utilize a fully supervised learning manner and require a large amount\nof entity-level annotation data of objects and relations, which is extremely\nresource-consuming and tedious to obtain. To tackle this problem, we propose\n3D-VLAP, a weakly-supervised 3D scene graph generation method via\nVisual-Linguistic Assisted Pseudo-labeling. Specifically, our 3D-VLAP exploits\nthe superior ability of current large-scale visual-linguistic models to align\nthe semantics between texts and 2D images, as well as the naturally existing\ncorrespondences between 2D images and 3D point clouds, and thus implicitly\nconstructs correspondences between texts and 3D point clouds. First, we\nestablish the positional correspondence from 3D point clouds to 2D images via\ncamera intrinsic and extrinsic parameters, thereby achieving alignment of 3D\npoint clouds and 2D images. Subsequently, a large-scale cross-modal\nvisual-linguistic model is employed to indirectly align 3D instances with the\ntextual category labels of objects by matching 2D images with object category\nlabels. The pseudo labels for objects and relations are then produced for\n3D-VLAP model training by calculating the similarity between visual embeddings\nand textual category embeddings of objects and relations encoded by the\nvisual-linguistic model, respectively. Ultimately, we design an edge\nself-attention based graph neural network to generate scene graphs of 3D point\ncloud scenes. Extensive experiments demonstrate that our 3D-VLAP achieves\ncomparable results with current advanced fully supervised methods, meanwhile\nsignificantly alleviating the pressure of data annotation.\n", "date": "2024-04-03", "categories": []}
{"id": "2404.02157", "title": "Segment Any 3D Object with Language", "abstract": "  In this paper, we investigate Open-Vocabulary 3D Instance Segmentation\n(OV-3DIS) with free-form language instructions. Earlier works that rely on only\nannotated base categories for training suffer from limited generalization to\nunseen novel categories. Recent works mitigate poor generalizability to novel\ncategories by generating class-agnostic masks or projecting generalized masks\nfrom 2D to 3D, but disregard semantic or geometry information, leading to\nsub-optimal performance. Instead, generating generalizable but semantic-related\nmasks directly from 3D point clouds would result in superior outcomes. In this\npaper, we introduce Segment any 3D Object with LanguagE (SOLE), which is a\nsemantic and geometric-aware visual-language learning framework with strong\ngeneralizability by generating semantic-related masks directly from 3D point\nclouds. Specifically, we propose a multimodal fusion network to incorporate\nmultimodal semantics in both backbone and decoder. In addition, to align the 3D\nsegmentation model with various language instructions and enhance the mask\nquality, we introduce three types of multimodal associations as supervision.\nOur SOLE outperforms previous methods by a large margin on ScanNetv2,\nScanNet200, and Replica benchmarks, and the results are even close to the\nfully-supervised counterpart despite the absence of class annotations in the\ntraining. Furthermore, extensive qualitative results demonstrate the\nversatility of our SOLE to language instructions.\n", "date": "2024-04-02", "categories": []}
{"id": "2402.12259", "title": "Open3DSG: Open-Vocabulary 3D Scene Graphs from Point Clouds with\n  Queryable Objects and Open-Set Relationships", "abstract": "  Current approaches for 3D scene graph prediction rely on labeled datasets to\ntrain models for a fixed set of known object classes and relationship\ncategories. We present Open3DSG, an alternative approach to learn 3D scene\ngraph prediction in an open world without requiring labeled scene graph data.\nWe co-embed the features from a 3D scene graph prediction backbone with the\nfeature space of powerful open world 2D vision language foundation models. This\nenables us to predict 3D scene graphs from 3D point clouds in a zero-shot\nmanner by querying object classes from an open vocabulary and predicting the\ninter-object relationships from a grounded LLM with scene graph features and\nqueried object classes as context. Open3DSG is the first 3D point cloud method\nto predict not only explicit open-vocabulary object classes, but also open-set\nrelationships that are not limited to a predefined label set, making it\npossible to express rare as well as specific objects and relationships in the\npredicted 3D scene graph. Our experiments show that Open3DSG is effective at\npredicting arbitrary object classes as well as their complex inter-object\nrelationships describing spatial, supportive, semantic and comparative\nrelationships.\n", "date": "2024-02-19", "categories": []}
{"id": "2404.00931", "title": "GOV-NeSF: Generalizable Open-Vocabulary Neural Semantic Fields", "abstract": "  Recent advancements in vision-language foundation models have significantly\nenhanced open-vocabulary 3D scene understanding. However, the generalizability\nof existing methods is constrained due to their framework designs and their\nreliance on 3D data. We address this limitation by introducing Generalizable\nOpen-Vocabulary Neural Semantic Fields (GOV-NeSF), a novel approach offering a\ngeneralizable implicit representation of 3D scenes with open-vocabulary\nsemantics. We aggregate the geometry-aware features using a cost volume, and\npropose a Multi-view Joint Fusion module to aggregate multi-view features\nthrough a cross-view attention mechanism, which effectively predicts\nview-specific blending weights for both colors and open-vocabulary features.\nRemarkably, our GOV-NeSF exhibits state-of-the-art performance in both 2D and\n3D open-vocabulary semantic segmentation, eliminating the need for ground truth\nsemantic labels or depth priors, and effectively generalize across scenes and\ndatasets without fine-tuning.\n", "date": "2024-04-01", "categories": []}
{"id": "2404.01248", "title": "Scalable Scene Modeling from Perspective Imaging: Physics-based\n  Appearance and Geometry Inference", "abstract": "  3D scene modeling techniques serve as the bedrocks in the geospatial\nengineering and computer science, which drives many applications ranging from\nautomated driving, terrain mapping, navigation, virtual, augmented, mixed, and\nextended reality (for gaming and movie industry etc.). This dissertation\npresents a fraction of contributions that advances 3D scene modeling to its\nstate of the art, in the aspects of both appearance and geometry modeling. In\ncontrast to the prevailing deep learning methods, as a core contribution, this\nthesis aims to develop algorithms that follow first principles, where\nsophisticated physic-based models are introduced alongside with simpler\nlearning and inference tasks. The outcomes of these algorithms yield processes\nthat can consume much larger volume of data for highly accurate reconstructing\n3D scenes at a scale without losing methodological generality, which are not\npossible by contemporary complex-model based deep learning methods.\nSpecifically, the dissertation introduces three novel methodologies that\naddress the challenges of inferring appearance and geometry through\nphysics-based modeling.\n  Overall, the research encapsulated in this dissertation marks a series of\nmethodological triumphs in the processing of complex datasets. By navigating\nthe confluence of deep learning, computational geometry, and photogrammetry,\nthis work lays down a robust framework for future exploration and practical\napplication in the rapidly evolving field of 3D scene reconstruction. The\noutcomes of these studies are evidenced through rigorous experiments and\ncomparisons with existing state-of-the-art methods, demonstrating the efficacy\nand scalability of the proposed approaches.\n", "date": "2024-04-01", "categories": []}
{"id": "2304.09704", "title": "Learnable Earth Parser: Discovering 3D Prototypes in Aerial Scans", "abstract": "  We propose an unsupervised method for parsing large 3D scans of real-world\nscenes with easily-interpretable shapes. This work aims to provide a practical\ntool for analyzing 3D scenes in the context of aerial surveying and mapping,\nwithout the need for user annotations. Our approach is based on a probabilistic\nreconstruction model that decomposes an input 3D point cloud into a small set\nof learned prototypical 3D shapes. The resulting reconstruction is visually\ninterpretable and can be used to perform unsupervised instance and low-shot\nsemantic segmentation of complex scenes. We demonstrate the usefulness of our\nmodel on a novel dataset of seven large aerial LiDAR scans from diverse\nreal-world scenarios. Our approach outperforms state-of-the-art unsupervised\nmethods in terms of decomposition accuracy while remaining visually\ninterpretable. Our code and dataset are available at\nhttps://romainloiseau.fr/learnable-earth-parser/\n", "date": "2023-04-19", "categories": []}
{"id": "2403.19474", "title": "SG-PGM: Partial Graph Matching Network with Semantic Geometric Fusion\n  for 3D Scene Graph Alignment and Its Downstream Tasks", "abstract": "  Scene graphs have been recently introduced into 3D spatial understanding as a\ncomprehensive representation of the scene. The alignment between 3D scene\ngraphs is the first step of many downstream tasks such as scene graph aided\npoint cloud registration, mosaicking, overlap checking, and robot navigation.\nIn this work, we treat 3D scene graph alignment as a partial graph-matching\nproblem and propose to solve it with a graph neural network. We reuse the\ngeometric features learned by a point cloud registration method and associate\nthe clustered point-level geometric features with the node-level semantic\nfeature via our designed feature fusion module. Partial matching is enabled by\nusing a learnable method to select the top-k similar node pairs. Subsequent\ndownstream tasks such as point cloud registration are achieved by running a\npre-trained registration network within the matched regions. We further propose\na point-matching rescoring method, that uses the node-wise alignment of the 3D\nscene graph to reweight the matching candidates from a pre-trained point cloud\nregistration method. It reduces the false point correspondences estimated\nespecially in low-overlapping cases. Experiments show that our method improves\nthe alignment accuracy by 10~20% in low-overlap and random transformation\nscenarios and outperforms the existing work in multiple downstream tasks.\n", "date": "2024-03-28", "categories": []}
{"id": "2312.09138", "title": "Living Scenes: Multi-object Relocalization and Reconstruction in\n  Changing 3D Environments", "abstract": "  Research into dynamic 3D scene understanding has primarily focused on\nshort-term change tracking from dense observations, while little attention has\nbeen paid to long-term changes with sparse observations. We address this gap\nwith MoRE, a novel approach for multi-object relocalization and reconstruction\nin evolving environments. We view these environments as \"living scenes\" and\nconsider the problem of transforming scans taken at different points in time\ninto a 3D reconstruction of the object instances, whose accuracy and\ncompleteness increase over time. At the core of our method lies an\nSE(3)-equivariant representation in a single encoder-decoder network, trained\non synthetic data. This representation enables us to seamlessly tackle instance\nmatching, registration, and reconstruction. We also introduce a joint\noptimization algorithm that facilitates the accumulation of point clouds\noriginating from the same instance across multiple scans taken at different\npoints in time. We validate our method on synthetic and real-world data and\ndemonstrate state-of-the-art performance in both end-to-end performance and\nindividual subtasks.\n", "date": "2023-12-14", "categories": []}
{"id": "2311.15383", "title": "Visual Programming for Zero-shot Open-Vocabulary 3D Visual Grounding", "abstract": "  3D Visual Grounding (3DVG) aims at localizing 3D object based on textual\ndescriptions. Conventional supervised methods for 3DVG often necessitate\nextensive annotations and a predefined vocabulary, which can be restrictive. To\naddress this issue, we propose a novel visual programming approach for\nzero-shot open-vocabulary 3DVG, leveraging the capabilities of large language\nmodels (LLMs). Our approach begins with a unique dialog-based method, engaging\nwith LLMs to establish a foundational understanding of zero-shot 3DVG. Building\non this, we design a visual program that consists of three types of modules,\ni.e., view-independent, view-dependent, and functional modules. These modules,\nspecifically tailored for 3D scenarios, work collaboratively to perform complex\nreasoning and inference. Furthermore, we develop an innovative language-object\ncorrelation module to extend the scope of existing 3D object detectors into\nopen-vocabulary scenarios. Extensive experiments demonstrate that our zero-shot\napproach can outperform some supervised baselines, marking a significant stride\ntowards effective 3DVG.\n", "date": "2023-11-26", "categories": []}
{"id": "2403.11401", "title": "Scene-LLM: Extending Language Model for 3D Visual Understanding and\n  Reasoning", "abstract": "  This paper introduces Scene-LLM, a 3D-visual-language model that enhances\nembodied agents' abilities in interactive 3D indoor environments by integrating\nthe reasoning strengths of Large Language Models (LLMs). Scene-LLM adopts a\nhybrid 3D visual feature representation, that incorporates dense spatial\ninformation and supports scene state updates. The model employs a projection\nlayer to efficiently project these features in the pre-trained textual\nembedding space, enabling effective interpretation of 3D visual information.\nUnique to our approach is the integration of both scene-level and ego-centric\n3D information. This combination is pivotal for interactive planning, where\nscene-level data supports global planning and ego-centric data is important for\nlocalization. Notably, we use ego-centric 3D frame features for feature\nalignment, an efficient technique that enhances the model's ability to align\nfeatures of small objects within the scene. Our experiments with Scene-LLM\ndemonstrate its strong capabilities in dense captioning, question answering,\nand interactive planning. We believe Scene-LLM advances the field of 3D visual\nunderstanding and reasoning, offering new possibilities for sophisticated agent\ninteractions in indoor settings.\n", "date": "2024-03-18", "categories": []}
{"id": "2403.15019", "title": "BSNet: Box-Supervised Simulation-assisted Mean Teacher for 3D Instance\n  Segmentation", "abstract": "  3D instance segmentation (3DIS) is a crucial task, but point-level\nannotations are tedious in fully supervised settings. Thus, using bounding\nboxes (bboxes) as annotations has shown great potential. The current mainstream\napproach is a two-step process, involving the generation of pseudo-labels from\nbox annotations and the training of a 3DIS network with the pseudo-labels.\nHowever, due to the presence of intersections among bboxes, not every point has\na determined instance label, especially in overlapping areas. To generate\nhigher quality pseudo-labels and achieve more precise weakly supervised 3DIS\nresults, we propose the Box-Supervised Simulation-assisted Mean Teacher for 3D\nInstance Segmentation (BSNet), which devises a novel pseudo-labeler called\nSimulation-assisted Transformer. The labeler consists of two main components.\nThe first is Simulation-assisted Mean Teacher, which introduces Mean Teacher\nfor the first time in this task and constructs simulated samples to assist the\nlabeler in acquiring prior knowledge about overlapping areas. To better model\nlocal-global structure, we also propose Local-Global Aware Attention as the\ndecoder for teacher and student labelers. Extensive experiments conducted on\nthe ScanNetV2 and S3DIS datasets verify the superiority of our designs. Code is\navailable at\n\\href{https://github.com/peoplelu/BSNet}{https://github.com/peoplelu/BSNet}.\n", "date": "2024-03-22", "categories": []}
{"id": "2403.16431", "title": "DOCTR: Disentangled Object-Centric Transformer for Point Scene\n  Understanding", "abstract": "  Point scene understanding is a challenging task to process real-world scene\npoint cloud, which aims at segmenting each object, estimating its pose, and\nreconstructing its mesh simultaneously. Recent state-of-the-art method first\nsegments each object and then processes them independently with multiple stages\nfor the different sub-tasks. This leads to a complex pipeline to optimize and\nmakes it hard to leverage the relationship constraints between multiple\nobjects. In this work, we propose a novel Disentangled Object-Centric\nTRansformer (DOCTR) that explores object-centric representation to facilitate\nlearning with multiple objects for the multiple sub-tasks in a unified manner.\nEach object is represented as a query, and a Transformer decoder is adapted to\niteratively optimize all the queries involving their relationship. In\nparticular, we introduce a semantic-geometry disentangled query (SGDQ) design\nthat enables the query features to attend separately to semantic information\nand geometric information relevant to the corresponding sub-tasks. A hybrid\nbipartite matching module is employed to well use the supervisions from all the\nsub-tasks during training. Qualitative and quantitative experimental results\ndemonstrate that our method achieves state-of-the-art performance on the\nchallenging ScanNet dataset. Code is available at\nhttps://github.com/SAITPublic/DOCTR.\n", "date": "2024-03-25", "categories": []}
{"id": "2403.14366", "title": "SurroundSDF: Implicit 3D Scene Understanding Based on Signed Distance\n  Field", "abstract": "  Vision-centric 3D environment understanding is both vital and challenging for\nautonomous driving systems. Recently, object-free methods have attracted\nconsiderable attention. Such methods perceive the world by predicting the\nsemantics of discrete voxel grids but fail to construct continuous and accurate\nobstacle surfaces. To this end, in this paper, we propose SurroundSDF to\nimplicitly predict the signed distance field (SDF) and semantic field for the\ncontinuous perception from surround images. Specifically, we introduce a\nquery-based approach and utilize SDF constrained by the Eikonal formulation to\naccurately describe the surfaces of obstacles. Furthermore, considering the\nabsence of precise SDF ground truth, we propose a novel weakly supervised\nparadigm for SDF, referred to as the Sandwich Eikonal formulation, which\nemphasizes applying correct and dense constraints on both sides of the surface,\nthereby enhancing the perceptual accuracy of the surface. Experiments suggest\nthat our method achieves SOTA for both occupancy prediction and 3D scene\nreconstruction tasks on the nuScenes dataset.\n", "date": "2024-03-21", "categories": []}
{"id": "2403.14619", "title": "ClusteringSDF: Self-Organized Neural Implicit Surfaces for 3D\n  Decomposition", "abstract": "  3D decomposition/segmentation still remains a challenge as large-scale 3D\nannotated data is not readily available. Contemporary approaches typically\nleverage 2D machine-generated segments, integrating them for 3D consistency.\nWhile the majority of these methods are based on NeRFs, they face a potential\nweakness that the instance/semantic embedding features derive from independent\nMLPs, thus preventing the segmentation network from learning the geometric\ndetails of the objects directly through radiance and density. In this paper, we\npropose ClusteringSDF, a novel approach to achieve both segmentation and\nreconstruction in 3D via the neural implicit surface representation,\nspecifically Signal Distance Function (SDF), where the segmentation rendering\nis directly integrated with the volume rendering of neural implicit surfaces.\nAlthough based on ObjectSDF++, ClusteringSDF no longer requires the\nground-truth segments for supervision while maintaining the capability of\nreconstructing individual object surfaces, but purely with the noisy and\ninconsistent labels from pre-trained models.As the core of ClusteringSDF, we\nintroduce a high-efficient clustering mechanism for lifting the 2D labels to 3D\nand the experimental results on the challenging scenes from ScanNet and Replica\ndatasets show that ClusteringSDF can achieve competitive performance compared\nagainst the state-of-the-art with significantly reduced training time.\n", "date": "2024-03-21", "categories": []}
{"id": "2312.03774", "title": "OctreeOcc: Efficient and Multi-Granularity Occupancy Prediction Using\n  Octree Queries", "abstract": "  Occupancy prediction has increasingly garnered attention in recent years for\nits fine-grained understanding of 3D scenes. Traditional approaches typically\nrely on dense, regular grid representations, which often leads to excessive\ncomputational demands and a loss of spatial details for small objects. This\npaper introduces OctreeOcc, an innovative 3D occupancy prediction framework\nthat leverages the octree representation to adaptively capture valuable\ninformation in 3D, offering variable granularity to accommodate object shapes\nand semantic regions of varying sizes and complexities. In particular, we\nincorporate image semantic information to improve the accuracy of initial\noctree structures and design an effective rectification mechanism to refine the\noctree structure iteratively. Our extensive evaluations show that OctreeOcc not\nonly surpasses state-of-the-art methods in occupancy prediction, but also\nachieves a 15%-24% reduction in computational overhead compared to\ndense-grid-based methods.\n", "date": "2023-12-06", "categories": []}
{"id": "2403.11697", "title": "Urban Scene Diffusion through Semantic Occupancy Map", "abstract": "  Generating unbounded 3D scenes is crucial for large-scale scene understanding\nand simulation. Urban scenes, unlike natural landscapes, consist of various\ncomplex man-made objects and structures such as roads, traffic signs, vehicles,\nand buildings. To create a realistic and detailed urban scene, it is crucial to\naccurately represent the geometry and semantics of the underlying objects,\ngoing beyond their visual appearance. In this work, we propose UrbanDiffusion,\na 3D diffusion model that is conditioned on a Bird's-Eye View (BEV) map and\ngenerates an urban scene with geometry and semantics in the form of semantic\noccupancy map. Our model introduces a novel paradigm that learns the data\ndistribution of scene-level structures within a latent space and further\nenables the expansion of the synthesized scene into an arbitrary scale. After\ntraining on real-world driving datasets, our model can generate a wide range of\ndiverse urban scenes given the BEV maps from the held-out set and also\ngeneralize to the synthesized maps from a driving simulator. We further\ndemonstrate its application to scene image synthesis with a pretrained image\ngenerator as a prior.\n", "date": "2024-03-18", "categories": []}
{"id": "2403.12301", "title": "R3DS: Reality-linked 3D Scenes for Panoramic Scene Understanding", "abstract": "  We introduce the Reality-linked 3D Scenes (R3DS) dataset of synthetic 3D\nscenes mirroring the real-world scene arrangements from Matterport3D panoramas.\nCompared to prior work, R3DS has more complete and densely populated scenes\nwith objects linked to real-world observations in panoramas. R3DS also provides\nan object support hierarchy, and matching object sets (e.g., same chairs around\na dining table) for each scene. Overall, R3DS contains 19K objects represented\nby 3,784 distinct CAD models from over 100 object categories. We demonstrate\nthe effectiveness of R3DS on the Panoramic Scene Understanding task. We find\nthat: 1) training on R3DS enables better generalization; 2) support relation\nprediction trained with R3DS improves performance compared to heuristically\ncalculated support; and 3) R3DS offers a challenging benchmark for future work\non panoramic scene understanding.\n", "date": "2024-03-18", "categories": []}
{"id": "2403.12722", "title": "HUGS: Holistic Urban 3D Scene Understanding via Gaussian Splatting", "abstract": "  Holistic understanding of urban scenes based on RGB images is a challenging\nyet important problem. It encompasses understanding both the geometry and\nappearance to enable novel view synthesis, parsing semantic labels, and\ntracking moving objects. Despite considerable progress, existing approaches\noften focus on specific aspects of this task and require additional inputs such\nas LiDAR scans or manually annotated 3D bounding boxes. In this paper, we\nintroduce a novel pipeline that utilizes 3D Gaussian Splatting for holistic\nurban scene understanding. Our main idea involves the joint optimization of\ngeometry, appearance, semantics, and motion using a combination of static and\ndynamic 3D Gaussians, where moving object poses are regularized via physical\nconstraints. Our approach offers the ability to render new viewpoints in\nreal-time, yielding 2D and 3D semantic information with high accuracy, and\nreconstruct dynamic scenes, even in scenarios where 3D bounding box detection\nare highly noisy. Experimental results on KITTI, KITTI-360, and Virtual KITTI 2\ndemonstrate the effectiveness of our approach.\n", "date": "2024-03-19", "categories": []}
{"id": "2402.15321", "title": "OpenSUN3D: 1st Workshop Challenge on Open-Vocabulary 3D Scene\n  Understanding", "abstract": "  This report provides an overview of the challenge hosted at the OpenSUN3D\nWorkshop on Open-Vocabulary 3D Scene Understanding held in conjunction with\nICCV 2023. The goal of this workshop series is to provide a platform for\nexploration and discussion of open-vocabulary 3D scene understanding tasks,\nincluding but not limited to segmentation, detection and mapping. We provide an\noverview of the challenge hosted at the workshop, present the challenge\ndataset, the evaluation methodology, and brief descriptions of the winning\nmethods. For additional details, please see\nhttps://opensun3d.github.io/index_iccv23.html.\n", "date": "2024-02-23", "categories": []}
{"id": "2403.11835", "title": "Agent3D-Zero: An Agent for Zero-shot 3D Understanding", "abstract": "  The ability to understand and reason the 3D real world is a crucial milestone\ntowards artificial general intelligence. The current common practice is to\nfinetune Large Language Models (LLMs) with 3D data and texts to enable 3D\nunderstanding. Despite their effectiveness, these approaches are inherently\nlimited by the scale and diversity of the available 3D data. Alternatively, in\nthis work, we introduce Agent3D-Zero, an innovative 3D-aware agent framework\naddressing the 3D scene understanding in a zero-shot manner. The essence of our\napproach centers on reconceptualizing the challenge of 3D scene perception as a\nprocess of understanding and synthesizing insights from multiple images,\ninspired by how our human beings attempt to understand 3D scenes. By\nconsolidating this idea, we propose a novel way to make use of a Large Visual\nLanguage Model (VLM) via actively selecting and analyzing a series of\nviewpoints for 3D understanding. Specifically, given an input 3D scene,\nAgent3D-Zero first processes a bird's-eye view image with custom-designed\nvisual prompts, then iteratively chooses the next viewpoints to observe and\nsummarize the underlying knowledge. A distinctive advantage of Agent3D-Zero is\nthe introduction of novel visual prompts, which significantly unleash the VLMs'\nability to identify the most informative viewpoints and thus facilitate\nobserving 3D scenes. Extensive experiments demonstrate the effectiveness of the\nproposed framework in understanding diverse and previously unseen 3D\nenvironments.\n", "date": "2024-03-18", "categories": []}
{"id": "2403.10452", "title": "Robust Shape Fitting for 3D Scene Abstraction", "abstract": "  Humans perceive and construct the world as an arrangement of simple\nparametric models. In particular, we can often describe man-made environments\nusing volumetric primitives such as cuboids or cylinders. Inferring these\nprimitives is important for attaining high-level, abstract scene descriptions.\nPrevious approaches for primitive-based abstraction estimate shape parameters\ndirectly and are only able to reproduce simple objects. In contrast, we propose\na robust estimator for primitive fitting, which meaningfully abstracts complex\nreal-world environments using cuboids. A RANSAC estimator guided by a neural\nnetwork fits these primitives to a depth map. We condition the network on\npreviously detected parts of the scene, parsing it one-by-one. To obtain\ncuboids from single RGB images, we additionally optimise a depth estimation CNN\nend-to-end. Naively minimising point-to-primitive distances leads to large or\nspurious cuboids occluding parts of the scene. We thus propose an improved\nocclusion-aware distance metric correctly handling opaque scenes. Furthermore,\nwe present a neural network based cuboid solver which provides more\nparsimonious scene abstractions while also reducing inference time. The\nproposed algorithm does not require labour-intensive labels, such as cuboid\nannotations, for training. Results on the NYU Depth v2 dataset demonstrate that\nthe proposed algorithm successfully abstracts cluttered real-world 3D scene\nlayouts.\n", "date": "2024-03-15", "categories": []}
{"id": "2403.07560", "title": "Unleashing Network Potentials for Semantic Scene Completion", "abstract": "  Semantic scene completion (SSC) aims to predict complete 3D voxel occupancy\nand semantics from a single-view RGB-D image, and recent SSC methods commonly\nadopt multi-modal inputs. However, our investigation reveals two limitations:\nineffective feature learning from single modalities and overfitting to limited\ndatasets. To address these issues, this paper proposes a novel SSC framework -\nAdversarial Modality Modulation Network (AMMNet) - with a fresh perspective of\noptimizing gradient updates. The proposed AMMNet introduces two core modules: a\ncross-modal modulation enabling the interdependence of gradient flows between\nmodalities, and a customized adversarial training scheme leveraging dynamic\ngradient competition. Specifically, the cross-modal modulation adaptively\nre-calibrates the features to better excite representation potentials from each\nsingle modality. The adversarial training employs a minimax game of evolving\ngradients, with customized guidance to strengthen the generator's perception of\nvisual fidelity from both geometric completeness and semantic correctness.\nExtensive experimental results demonstrate that AMMNet outperforms\nstate-of-the-art SSC methods by a large margin, providing a promising direction\nfor improving the effectiveness and generalization of SSC methods.\n", "date": "2024-03-12", "categories": []}
{"id": "2403.08885", "title": "SLCF-Net: Sequential LiDAR-Camera Fusion for Semantic Scene Completion\n  using a 3D Recurrent U-Net", "abstract": "  We introduce SLCF-Net, a novel approach for the Semantic Scene Completion\n(SSC) task that sequentially fuses LiDAR and camera data. It jointly estimates\nmissing geometry and semantics in a scene from sequences of RGB images and\nsparse LiDAR measurements. The images are semantically segmented by a\npre-trained 2D U-Net and a dense depth prior is estimated from a\ndepth-conditioned pipeline fueled by Depth Anything. To associate the 2D image\nfeatures with the 3D scene volume, we introduce Gaussian-decay Depth-prior\nProjection (GDP). This module projects the 2D features into the 3D volume along\nthe line of sight with a Gaussian-decay function, centered around the depth\nprior. Volumetric semantics is computed by a 3D U-Net. We propagate the hidden\n3D U-Net state using the sensor motion and design a novel loss to ensure\ntemporal consistency. We evaluate our approach on the SemanticKITTI dataset and\ncompare it with leading SSC approaches. The SLCF-Net excels in all SSC metrics\nand shows great temporal consistency.\n", "date": "2024-03-13", "categories": []}
{"id": "2403.08766", "title": "MonoOcc: Digging into Monocular Semantic Occupancy Prediction", "abstract": "  Monocular Semantic Occupancy Prediction aims to infer the complete 3D\ngeometry and semantic information of scenes from only 2D images. It has\ngarnered significant attention, particularly due to its potential to enhance\nthe 3D perception of autonomous vehicles. However, existing methods rely on a\ncomplex cascaded framework with relatively limited information to restore 3D\nscenes, including a dependency on supervision solely on the whole network's\noutput, single-frame input, and the utilization of a small backbone. These\nchallenges, in turn, hinder the optimization of the framework and yield\ninferior prediction results, particularly concerning smaller and long-tailed\nobjects. To address these issues, we propose MonoOcc. In particular, we (i)\nimprove the monocular occupancy prediction framework by proposing an auxiliary\nsemantic loss as supervision to the shallow layers of the framework and an\nimage-conditioned cross-attention module to refine voxel features with visual\nclues, and (ii) employ a distillation module that transfers temporal\ninformation and richer knowledge from a larger image backbone to the monocular\nsemantic occupancy prediction framework with low cost of hardware. With these\nadvantages, our method yields state-of-the-art performance on the camera-based\nSemanticKITTI Scene Completion benchmark. Codes and models can be accessed at\nhttps://github.com/ucaszyp/MonoOcc\n", "date": "2024-03-13", "categories": []}
{"id": "2403.07469", "title": "A Comprehensive Survey of 3D Dense Captioning: Localizing and Describing\n  Objects in 3D Scenes", "abstract": "  Three-Dimensional (3D) dense captioning is an emerging vision-language\nbridging task that aims to generate multiple detailed and accurate descriptions\nfor 3D scenes. It presents significant potential and challenges due to its\ncloser representation of the real world compared to 2D visual captioning, as\nwell as complexities in data collection and processing of 3D point cloud\nsources. Despite the popularity and success of existing methods, there is a\nlack of comprehensive surveys summarizing the advancements in this field, which\nhinders its progress. In this paper, we provide a comprehensive review of 3D\ndense captioning, covering task definition, architecture classification,\ndataset analysis, evaluation metrics, and in-depth prosperity discussions.\nBased on a synthesis of previous literature, we refine a standard pipeline that\nserves as a common paradigm for existing methods. We also introduce a clear\ntaxonomy of existing models, summarize technologies involved in different\nmodules, and conduct detailed experiment analysis. Instead of a chronological\norder introduction, we categorize the methods into different classes to\nfacilitate exploration and analysis of the differences and connections among\nexisting techniques. We also provide a reading guideline to assist readers with\ndifferent backgrounds and purposes in reading efficiently. Furthermore, we\npropose a series of promising future directions for 3D dense captioning by\nidentifying challenges and aligning them with the development of related tasks,\noffering valuable insights and inspiring future research in this field. Our aim\nis to provide a comprehensive understanding of 3D dense captioning, foster\nfurther investigations, and contribute to the development of novel applications\nin multimedia and related domains.\n", "date": "2024-03-12", "categories": []}
{"id": "2403.06974", "title": "Memory-based Adapters for Online 3D Scene Perception", "abstract": "  In this paper, we propose a new framework for online 3D scene perception.\nConventional 3D scene perception methods are offline, i.e., take an already\nreconstructed 3D scene geometry as input, which is not applicable in robotic\napplications where the input data is streaming RGB-D videos rather than a\ncomplete 3D scene reconstructed from pre-collected RGB-D videos. To deal with\nonline 3D scene perception tasks where data collection and perception should be\nperformed simultaneously, the model should be able to process 3D scenes frame\nby frame and make use of the temporal information. To this end, we propose an\nadapter-based plug-and-play module for the backbone of 3D scene perception\nmodel, which constructs memory to cache and aggregate the extracted RGB-D\nfeatures to empower offline models with temporal learning ability.\nSpecifically, we propose a queued memory mechanism to cache the supporting\npoint cloud and image features. Then we devise aggregation modules which\ndirectly perform on the memory and pass temporal information to current frame.\nWe further propose 3D-to-2D adapter to enhance image features with strong\nglobal context. Our adapters can be easily inserted into mainstream offline\narchitectures of different tasks and significantly boost their performance on\nonline tasks. Extensive experiments on ScanNet and SceneNN datasets demonstrate\nour approach achieves leading performance on three 3D scene perception tasks\ncompared with state-of-the-art online methods by simply finetuning existing\noffline models, without any model and task-specific designs.\n\\href{https://xuxw98.github.io/Online3D/}{Project page}.\n", "date": "2024-03-11", "categories": []}
{"id": "2309.12708", "title": "PointSSC: A Cooperative Vehicle-Infrastructure Point Cloud Benchmark for\n  Semantic Scene Completion", "abstract": "  Semantic Scene Completion (SSC) aims to jointly generate space occupancies\nand semantic labels for complex 3D scenes. Most existing SSC models focus on\nvolumetric representations, which are memory-inefficient for large outdoor\nspaces. Point clouds provide a lightweight alternative but existing benchmarks\nlack outdoor point cloud scenes with semantic labels. To address this, we\nintroduce PointSSC, the first cooperative vehicle-infrastructure point cloud\nbenchmark for semantic scene completion. These scenes exhibit long-range\nperception and minimal occlusion. We develop an automated annotation pipeline\nleveraging Semantic Segment Anything to efficiently assign semantics. To\nbenchmark progress, we propose a LiDAR-based model with a Spatial-Aware\nTransformer for global and local feature extraction and a Completion and\nSegmentation Cooperative Module for joint completion and segmentation. PointSSC\nprovides a challenging testbed to drive advances in semantic point cloud\ncompletion for real-world navigation. The code and datasets are available at\nhttps://github.com/yyxssm/PointSSC.\n", "date": "2023-09-22", "categories": []}
{"id": "2402.16936", "title": "Disentangled 3D Scene Generation with Layout Learning", "abstract": "  We introduce a method to generate 3D scenes that are disentangled into their\ncomponent objects. This disentanglement is unsupervised, relying only on the\nknowledge of a large pretrained text-to-image model. Our key insight is that\nobjects can be discovered by finding parts of a 3D scene that, when rearranged\nspatially, still produce valid configurations of the same scene. Concretely,\nour method jointly optimizes multiple NeRFs from scratch - each representing\nits own object - along with a set of layouts that composite these objects into\nscenes. We then encourage these composited scenes to be in-distribution\naccording to the image generator. We show that despite its simplicity, our\napproach successfully generates 3D scenes decomposed into individual objects,\nenabling new capabilities in text-to-3D content creation. For results and an\ninteractive demo, see our project page at https://dave.ml/layoutlearning/\n", "date": "2024-02-26", "categories": []}
{"id": "2402.15933", "title": "Bridging the Gap between 2D and 3D Visual Question Answering: A Fusion\n  Approach for 3D VQA", "abstract": "  In 3D Visual Question Answering (3D VQA), the scarcity of fully annotated\ndata and limited visual content diversity hampers the generalization to novel\nscenes and 3D concepts (e.g., only around 800 scenes are utilized in ScanQA and\nSQA dataset). Current approaches resort supplement 3D reasoning with 2D\ninformation. However, these methods face challenges: either they use top-down\n2D views that introduce overly complex and sometimes question-irrelevant visual\nclues, or they rely on globally aggregated scene/image-level representations\nfrom 2D VLMs, losing the fine-grained vision-language correlations. To overcome\nthese limitations, our approach utilizes question-conditional 2D view selection\nprocedure, pinpointing semantically relevant 2D inputs for crucial visual\nclues. We then integrate this 2D knowledge into the 3D-VQA system via a\ntwo-branch Transformer structure. This structure, featuring a Twin-Transformer\ndesign, compactly combines 2D and 3D modalities and captures fine-grained\ncorrelations between modalities, allowing them mutually augmenting each other.\nIntegrating proposed mechanisms above, we present BridgeQA, that offers a fresh\nperspective on multi-modal transformer-based architectures for 3D-VQA.\nExperiments validate that BridgeQA achieves state-of-the-art on 3D-VQA datasets\nand significantly outperforms existing solutions. Code is available at\n$\\href{https://github.com/matthewdm0816/BridgeQA}{\\text{this URL}}$.\n", "date": "2024-02-24", "categories": []}
{"id": "2402.14215", "title": "Swin3D++: Effective Multi-Source Pretraining for 3D Indoor Scene\n  Understanding", "abstract": "  Data diversity and abundance are essential for improving the performance and\ngeneralization of models in natural language processing and 2D vision. However,\n3D vision domain suffers from the lack of 3D data, and simply combining\nmultiple 3D datasets for pretraining a 3D backbone does not yield significant\nimprovement, due to the domain discrepancies among different 3D datasets that\nimpede effective feature learning. In this work, we identify the main sources\nof the domain discrepancies between 3D indoor scene datasets, and propose\nSwin3D++, an enhanced architecture based on Swin3D for efficient pretraining on\nmulti-source 3D point clouds. Swin3D++ introduces domain-specific mechanisms to\nSwin3D's modules to address domain discrepancies and enhance the network\ncapability on multi-source pretraining. Moreover, we devise a simple\nsource-augmentation strategy to increase the pretraining data scale and\nfacilitate supervised pretraining. We validate the effectiveness of our design,\nand demonstrate that Swin3D++ surpasses the state-of-the-art 3D pretraining\nmethods on typical indoor scene understanding tasks. Our code and models will\nbe released at https://github.com/microsoft/Swin3D\n", "date": "2024-02-22", "categories": []}
{"id": "2402.12792", "title": "OccFlowNet: Towards Self-supervised Occupancy Estimation via\n  Differentiable Rendering and Occupancy Flow", "abstract": "  Semantic occupancy has recently gained significant traction as a prominent 3D\nscene representation. However, most existing methods rely on large and costly\ndatasets with fine-grained 3D voxel labels for training, which limits their\npracticality and scalability, increasing the need for self-monitored learning\nin this domain. In this work, we present a novel approach to occupancy\nestimation inspired by neural radiance field (NeRF) using only 2D labels, which\nare considerably easier to acquire. In particular, we employ differentiable\nvolumetric rendering to predict depth and semantic maps and train a 3D network\nbased on 2D supervision only. To enhance geometric accuracy and increase the\nsupervisory signal, we introduce temporal rendering of adjacent time steps.\nAdditionally, we introduce occupancy flow as a mechanism to handle dynamic\nobjects in the scene and ensure their temporal consistency. Through extensive\nexperimentation we demonstrate that 2D supervision only is sufficient to\nachieve state-of-the-art performance compared to methods using 3D labels, while\noutperforming concurrent 2D approaches. When combining 2D supervision with 3D\nlabels, temporal rendering and occupancy flow we outperform all previous\noccupancy estimation models significantly. We conclude that the proposed\nrendering supervision and occupancy flow advances occupancy estimation and\nfurther bridges the gap towards self-supervised learning in this domain.\n", "date": "2024-02-20", "categories": []}
{"id": "2211.14310", "title": "Efficient 3D Reconstruction, Streaming and Visualization of Static and\n  Dynamic Scene Parts for Multi-client Live-telepresence in Large-scale\n  Environments", "abstract": "  Despite the impressive progress of telepresence systems for room-scale scenes\nwith static and dynamic scene entities, expanding their capabilities to\nscenarios with larger dynamic environments beyond a fixed size of a few\nsquare-meters remains challenging.\n  In this paper, we aim at sharing 3D live-telepresence experiences in\nlarge-scale environments beyond room scale with both static and dynamic scene\nentities at practical bandwidth requirements only based on light-weight scene\ncapture with a single moving consumer-grade RGB-D camera. To this end, we\npresent a system which is built upon a novel hybrid volumetric scene\nrepresentation in terms of the combination of a voxel-based scene\nrepresentation for the static contents, that not only stores the reconstructed\nsurface geometry but also contains information about the object semantics as\nwell as their accumulated dynamic movement over time, and a point-cloud-based\nrepresentation for dynamic scene parts, where the respective separation from\nstatic parts is achieved based on semantic and instance information extracted\nfor the input frames. With an independent yet simultaneous streaming of both\nstatic and dynamic content, where we seamlessly integrate potentially moving\nbut currently static scene entities in the static model until they are becoming\ndynamic again, as well as the fusion of static and dynamic data at the remote\nclient, our system is able to achieve VR-based live-telepresence at close to\nreal-time rates. Our evaluation demonstrates the potential of our novel\napproach in terms of visual quality, performance, and ablation studies\nregarding involved design choices.\n", "date": "2022-11-25", "categories": []}
{"id": "2402.06826", "title": "Neural Rendering based Urban Scene Reconstruction for Autonomous Driving", "abstract": "  Dense 3D reconstruction has many applications in automated driving including\nautomated annotation validation, multimodal data augmentation, providing ground\ntruth annotations for systems lacking LiDAR, as well as enhancing auto-labeling\naccuracy. LiDAR provides highly accurate but sparse depth, whereas camera\nimages enable estimation of dense depth but noisy particularly at long ranges.\nIn this paper, we harness the strengths of both sensors and propose a\nmultimodal 3D scene reconstruction using a framework combining neural implicit\nsurfaces and radiance fields. In particular, our method estimates dense and\naccurate 3D structures and creates an implicit map representation based on\nsigned distance fields, which can be further rendered into RGB images, and\ndepth maps. A mesh can be extracted from the learned signed distance field and\nculled based on occlusion. Dynamic objects are efficiently filtered on the fly\nduring sampling using 3D object detection models. We demonstrate qualitative\nand quantitative results on challenging automotive scenes.\n", "date": "2024-02-09", "categories": []}
{"id": "2305.13876", "title": "Cross3DVG: Cross-Dataset 3D Visual Grounding on Different RGB-D Scans", "abstract": "  We present a novel task for cross-dataset visual grounding in 3D scenes\n(Cross3DVG), which overcomes limitations of existing 3D visual grounding\nmodels, specifically their restricted 3D resources and consequent tendencies of\noverfitting a specific 3D dataset. We created RIORefer, a large-scale 3D visual\ngrounding dataset, to facilitate Cross3DVG. It includes more than 63k diverse\ndescriptions of 3D objects within 1,380 indoor RGB-D scans from 3RScan, with\nhuman annotations. After training the Cross3DVG model using the source 3D\nvisual grounding dataset, we evaluate it without target labels using the target\ndataset with, e.g., different sensors, 3D reconstruction methods, and language\nannotators. Comprehensive experiments are conducted using established visual\ngrounding models and with CLIP-based multi-view 2D and 3D integration designed\nto bridge gaps among 3D datasets. For Cross3DVG tasks, (i) cross-dataset 3D\nvisual grounding exhibits significantly worse performance than learning and\nevaluation with a single dataset because of the 3D data and language variants\nacross datasets. Moreover, (ii) better object detector and localization modules\nand fusing 3D data and multi-view CLIP-based image features can alleviate this\nlower performance. Our Cross3DVG task can provide a benchmark for developing\nrobust 3D visual grounding models to handle diverse 3D scenes while leveraging\ndeep language understanding.\n", "date": "2023-05-23", "categories": []}
{"id": "2402.04671", "title": "V2VSSC: A 3D Semantic Scene Completion Benchmark for Perception with\n  Vehicle to Vehicle Communication", "abstract": "  Semantic scene completion (SSC) has recently gained popularity because it can\nprovide both semantic and geometric information that can be used directly for\nautonomous vehicle navigation. However, there are still challenges to overcome.\nSSC is often hampered by occlusion and short-range perception due to sensor\nlimitations, which can pose safety risks. This paper proposes a fundamental\nsolution to this problem by leveraging vehicle-to-vehicle (V2V) communication.\nWe propose the first generalized collaborative SSC framework that allows\nautonomous vehicles to share sensing information from different sensor views to\njointly perform SSC tasks. To validate the proposed framework, we further build\nV2VSSC, the first V2V SSC benchmark, on top of the large-scale V2V perception\ndataset OPV2V. Extensive experiments demonstrate that by leveraging V2V\ncommunication, the SSC performance can be increased by 8.3% on geometric metric\nIoU and 6.0% mIOU.\n", "date": "2024-02-07", "categories": []}
{"id": "2402.03327", "title": "Uni3D-LLM: Unifying Point Cloud Perception, Generation and Editing with\n  Large Language Models", "abstract": "  In this paper, we introduce Uni3D-LLM, a unified framework that leverages a\nLarge Language Model (LLM) to integrate tasks of 3D perception, generation, and\nediting within point cloud scenes. This framework empowers users to\neffortlessly generate and modify objects at specified locations within a scene,\nguided by the versatility of natural language descriptions. Uni3D-LLM harnesses\nthe expressive power of natural language to allow for precise command over the\ngeneration and editing of 3D objects, thereby significantly enhancing\noperational flexibility and controllability. By mapping point cloud into the\nunified representation space, Uni3D-LLM achieves cross-application\nfunctionality, enabling the seamless execution of a wide array of tasks,\nranging from the accurate instantiation of 3D objects to the diverse\nrequirements of interactive design. Through a comprehensive suite of rigorous\nexperiments, the efficacy of Uni3D-LLM in the comprehension, generation, and\nediting of point cloud has been validated. Additionally, we have assessed the\nimpact of integrating a point cloud perception module on the generation and\nediting processes, confirming the substantial potential of our approach for\npractical applications.\n", "date": "2024-01-09", "categories": []}
{"id": "2210.01595", "title": "FreDSNet: Joint Monocular Depth and Semantic Segmentation with Fast\n  Fourier Convolutions", "abstract": "  In this work we present FreDSNet, a deep learning solution which obtains\nsemantic 3D understanding of indoor environments from single panoramas.\nOmnidirectional images reveal task-specific advantages when addressing scene\nunderstanding problems due to the 360-degree contextual information about the\nentire environment they provide. However, the inherent characteristics of the\nomnidirectional images add additional problems to obtain an accurate detection\nand segmentation of objects or a good depth estimation. To overcome these\nproblems, we exploit convolutions in the frequential domain obtaining a wider\nreceptive field in each convolutional layer. These convolutions allow to\nleverage the whole context information from omnidirectional images. FreDSNet is\nthe first network that jointly provides monocular depth estimation and semantic\nsegmentation from a single panoramic image exploiting fast Fourier\nconvolutions. Our experiments show that FreDSNet has similar performance as\nspecific state of the art methods for semantic segmentation and depth\nestimation. FreDSNet code is publicly available in\nhttps://github.com/Sbrunoberenguel/FreDSNet\n", "date": "2022-10-04", "categories": []}
{"id": "2402.01466", "title": "Scaled 360 layouts: Revisiting non-central panoramas", "abstract": "  From a non-central panorama, 3D lines can be recovered by geometric\nreasoning. However, their sensitivity to noise and the complex geometric\nmodeling required has led these panoramas being very little investigated. In\nthis work we present a novel approach for 3D layout recovery of indoor\nenvironments using single non-central panoramas. We obtain the boundaries of\nthe structural lines of the room from a non-central panorama using deep\nlearning and exploit the properties of non-central projection systems in a new\ngeometrical processing to recover the scaled layout. We solve the problem for\nManhattan environments, handling occlusions, and also for Atlanta environments\nin an unified method. The experiments performed improve the state-of-the-art\nmethods for 3D layout recovery from a single panorama. Our approach is the\nfirst work using deep learning with non-central panoramas and recovering the\nscale of single panorama layouts.\n", "date": "2024-02-02", "categories": []}
{"id": "2401.17075", "title": "Non-central panorama indoor dataset", "abstract": "  Omnidirectional images are one of the main sources of information for\nlearning based scene understanding algorithms. However, annotated datasets of\nomnidirectional images cannot keep the pace of these learning based algorithms\ndevelopment. Among the different panoramas and in contrast to standard central\nones, non-central panoramas provide geometrical information in the distortion\nof the image from which we can retrieve 3D information of the environment [2].\nHowever, due to the lack of commercial non-central devices, up until now there\nwas no dataset of these kinds of panoramas. In this data paper, we present the\nfirst dataset of non-central panoramas for indoor scene understanding. The\ndataset is composed by {\\bf 2574} RGB non-central panoramas taken in around 650\ndifferent rooms. Each panorama has associated a depth map and annotations to\nobtain the layout of the room from the image as a structural edge map, list of\ncorners in the image, the 3D corners of the room and the camera pose. The\nimages are taken from photorealistic virtual environments and pixel-wise\nautomatically annotated.\n", "date": "2024-01-30", "categories": []}
{"id": "2306.12681", "title": "One at a Time: Progressive Multi-step Volumetric Probability Learning\n  for Reliable 3D Scene Perception", "abstract": "  Numerous studies have investigated the pivotal role of reliable 3D volume\nrepresentation in scene perception tasks, such as multi-view stereo (MVS) and\nsemantic scene completion (SSC). They typically construct 3D probability\nvolumes directly with geometric correspondence, attempting to fully address the\nscene perception tasks in a single forward pass. However, such a single-step\nsolution makes it hard to learn accurate and convincing volumetric probability,\nespecially in challenging regions like unexpected occlusions and complicated\nlight reflections. Therefore, this paper proposes to decompose the complicated\n3D volume representation learning into a sequence of generative steps to\nfacilitate fine and reliable scene perception. Considering the recent advances\nachieved by strong generative diffusion models, we introduce a multi-step\nlearning framework, dubbed as VPD, dedicated to progressively refining the\nVolumetric Probability in a Diffusion process. Extensive experiments are\nconducted on scene perception tasks including multi-view stereo (MVS) and\nsemantic scene completion (SSC), to validate the efficacy of our method in\nlearning reliable volumetric representations. Notably, for the SSC task, our\nwork stands out as the first to surpass LiDAR-based methods on the\nSemanticKITTI dataset.\n", "date": "2023-06-22", "categories": []}
{"id": "2305.00393", "title": "DynaVol: Unsupervised Learning for Dynamic Scenes through Object-Centric\n  Voxelization", "abstract": "  Unsupervised learning of object-centric representations in dynamic visual\nscenes is challenging. Unlike most previous approaches that learn to decompose\n2D images, we present DynaVol, a 3D scene generative model that unifies\ngeometric structures and object-centric learning in a differentiable volume\nrendering framework. The key idea is to perform object-centric voxelization to\ncapture the 3D nature of the scene, which infers the probability distribution\nover objects at individual spatial locations. These voxel features evolve over\ntime through a canonical-space deformation function, forming the basis for\nglobal representation learning via slot attention. The voxel features and\nglobal features are complementary and are both leveraged by a compositional\nNeRF decoder for volume rendering. DynaVol remarkably outperforms existing\napproaches for unsupervised dynamic scene decomposition. Once trained, the\nexplicitly meaningful voxel features enable additional capabilities that 2D\nscene decomposition methods cannot achieve: it is possible to freely edit the\ngeometric shapes or manipulate the motion trajectories of the objects.\n", "date": "2023-04-30", "categories": []}
{"id": "2308.02982", "title": "Beyond First Impressions: Integrating Joint Multi-modal Cues for\n  Comprehensive 3D Representation", "abstract": "  In recent years, 3D understanding has turned to 2D vision-language\npre-trained models to overcome data scarcity challenges. However, existing\nmethods simply transfer 2D alignment strategies, aligning 3D representations\nwith single-view 2D images and coarse-grained parent category text. These\napproaches introduce information degradation and insufficient synergy issues,\nleading to performance loss. Information degradation arises from overlooking\nthe fact that a 3D representation should be equivalent to a series of\nmulti-view images and more fine-grained subcategory text. Insufficient synergy\nneglects the idea that a robust 3D representation should align with the joint\nvision-language space, rather than independently aligning with each modality.\nIn this paper, we propose a multi-view joint modality modeling approach, termed\nJM3D, to obtain a unified representation for point cloud, text, and image.\nSpecifically, a novel Structured Multimodal Organizer (SMO) is proposed to\naddress the information degradation issue, which introduces contiguous\nmulti-view images and hierarchical text to enrich the representation of vision\nand language modalities. A Joint Multi-modal Alignment (JMA) is designed to\ntackle the insufficient synergy problem, which models the joint modality by\nincorporating language knowledge into the visual modality. Extensive\nexperiments on ModelNet40 and ScanObjectNN demonstrate the effectiveness of our\nproposed method, JM3D, which achieves state-of-the-art performance in zero-shot\n3D classification. JM3D outperforms ULIP by approximately 4.3% on PointMLP and\nachieves an improvement of up to 6.5% accuracy on PointNet++ in top-1 accuracy\nfor zero-shot 3D classification on ModelNet40. The source code and trained\nmodels for all our experiments are publicly available at\nhttps://github.com/Mr-Neko/JM3D.\n", "date": "2023-08-06", "categories": []}
{"id": "2401.09419", "title": "GARField: Group Anything with Radiance Fields", "abstract": "  Grouping is inherently ambiguous due to the multiple levels of granularity in\nwhich one can decompose a scene -- should the wheels of an excavator be\nconsidered separate or part of the whole? We present Group Anything with\nRadiance Fields (GARField), an approach for decomposing 3D scenes into a\nhierarchy of semantically meaningful groups from posed image inputs. To do this\nwe embrace group ambiguity through physical scale: by optimizing a\nscale-conditioned 3D affinity feature field, a point in the world can belong to\ndifferent groups of different sizes. We optimize this field from a set of 2D\nmasks provided by Segment Anything (SAM) in a way that respects coarse-to-fine\nhierarchy, using scale to consistently fuse conflicting masks from different\nviewpoints. From this field we can derive a hierarchy of possible groupings via\nautomatic tree construction or user interaction. We evaluate GARField on a\nvariety of in-the-wild scenes and find it effectively extracts groups at many\nlevels: clusters of objects, objects, and various subparts. GARField inherently\nrepresents multi-view consistent groupings and produces higher fidelity groups\nthan the input SAM masks. GARField's hierarchical grouping could have exciting\ndownstream applications such as 3D asset extraction or dynamic scene\nunderstanding. See the project website at https://www.garfield.studio/\n", "date": "2024-01-17", "categories": []}
{"id": "2211.15759", "title": "PIDS: Joint Point Interaction-Dimension Search for 3D Point Cloud", "abstract": "  The interaction and dimension of points are two important axes in designing\npoint operators to serve hierarchical 3D models. Yet, these two axes are\nheterogeneous and challenging to fully explore. Existing works craft point\noperator under a single axis and reuse the crafted operator in all parts of 3D\nmodels. This overlooks the opportunity to better combine point interactions and\ndimensions by exploiting varying geometry/density of 3D point clouds. In this\nwork, we establish PIDS, a novel paradigm to jointly explore point interactions\nand point dimensions to serve semantic segmentation on point cloud data. We\nestablish a large search space to jointly consider versatile point interactions\nand point dimensions. This supports point operators with various\ngeometry/density considerations. The enlarged search space with heterogeneous\nsearch components calls for a better ranking of candidate models. To achieve\nthis, we improve the search space exploration by leveraging predictor-based\nNeural Architecture Search (NAS), and enhance the quality of prediction by\nassigning unique encoding to heterogeneous search components based on their\npriors. We thoroughly evaluate the networks crafted by PIDS on two semantic\nsegmentation benchmarks, showing ~1% mIOU improvement on SemanticKITTI and\nS3DIS over state-of-the-art 3D models.\n", "date": "2022-11-28", "categories": []}
{"id": "2306.00965", "title": "BUOL: A Bottom-Up Framework with Occupancy-aware Lifting for Panoptic 3D\n  Scene Reconstruction From A Single Image", "abstract": "  Understanding and modeling the 3D scene from a single image is a practical\nproblem. A recent advance proposes a panoptic 3D scene reconstruction task that\nperforms both 3D reconstruction and 3D panoptic segmentation from a single\nimage. Although having made substantial progress, recent works only focus on\ntop-down approaches that fill 2D instances into 3D voxels according to\nestimated depth, which hinders their performance by two ambiguities. (1)\ninstance-channel ambiguity: The variable ids of instances in each scene lead to\nambiguity during filling voxel channels with 2D information, confusing the\nfollowing 3D refinement. (2) voxel-reconstruction ambiguity: 2D-to-3D lifting\nwith estimated single view depth only propagates 2D information onto the\nsurface of 3D regions, leading to ambiguity during the reconstruction of\nregions behind the frontal view surface. In this paper, we propose BUOL, a\nBottom-Up framework with Occupancy-aware Lifting to address the two issues for\npanoptic 3D scene reconstruction from a single image. For instance-channel\nambiguity, a bottom-up framework lifts 2D information to 3D voxels based on\ndeterministic semantic assignments rather than arbitrary instance id\nassignments. The 3D voxels are then refined and grouped into 3D instances\naccording to the predicted 2D instance centers. For voxel-reconstruction\nambiguity, the estimated multi-plane occupancy is leveraged together with depth\nto fill the whole regions of things and stuff. Our method shows a tremendous\nperformance advantage over state-of-the-art methods on synthetic dataset\n3D-Front and real-world dataset Matterport3D. Code and models are available in\nhttps://github.com/chtsy/buol.\n", "date": "2023-06-01", "categories": []}
{"id": "2401.03201", "title": "3DMIT: 3D Multi-modal Instruction Tuning for Scene Understanding", "abstract": "  The remarkable potential of multi-modal large language models (MLLMs) in\ncomprehending both vision and language information has been widely\nacknowledged. However, the scarcity of 3D scenes-language pairs in comparison\nto their 2D counterparts, coupled with the inadequacy of existing approaches in\nunderstanding of 3D scenes by LLMs, poses a significant challenge. In response,\nwe collect and construct an extensive dataset comprising 75K\ninstruction-response pairs tailored for 3D scenes. This dataset addresses tasks\nrelated to 3D VQA, 3D grounding, and 3D conversation. To further enhance the\nintegration of 3D spatial information into LLMs, we introduce a novel and\nefficient prompt tuning paradigm, 3DMIT. This paradigm eliminates the alignment\nstage between 3D scenes and language and extends the instruction prompt with\nthe 3D modality information including the entire scene and segmented objects.\nWe evaluate the effectiveness of our method across diverse tasks in the 3D\nscene domain and find that our approach serves as a strategic means to enrich\nLLMs' comprehension of the 3D world. Our code is available at\nhttps://github.com/staymylove/3DMIT.\n", "date": "2024-01-06", "categories": []}
{"id": "2401.02241", "title": "Slot-guided Volumetric Object Radiance Fields", "abstract": "  We present a novel framework for 3D object-centric representation learning.\nOur approach effectively decomposes complex scenes into individual objects from\na single image in an unsupervised fashion. This method, called slot-guided\nVolumetric Object Radiance Fields (sVORF), composes volumetric object radiance\nfields with object slots as a guidance to implement unsupervised 3D scene\ndecomposition. Specifically, sVORF obtains object slots from a single image via\na transformer module, maps these slots to volumetric object radiance fields\nwith a hypernetwork and composes object radiance fields with the guidance of\nobject slots at a 3D location. Moreover, sVORF significantly reduces memory\nrequirement due to small-sized pixel rendering during training. We demonstrate\nthe effectiveness of our approach by showing top results in scene decomposition\nand generation tasks of complex synthetic datasets (e.g., Room-Diverse).\nFurthermore, we also confirm the potential of sVORF to segment objects in\nreal-world scenes (e.g., the LLFF dataset). We hope our approach can provide\npreliminary understanding of the physical world and help ease future research\nin 3D object-centric representation learning.\n", "date": "2024-01-04", "categories": []}
{"id": "2312.11451", "title": "Language-Assisted 3D Scene Understanding", "abstract": "  The scale and quality of point cloud datasets constrain the advancement of\npoint cloud learning. Recently, with the development of multi-modal learning,\nthe incorporation of domain-agnostic prior knowledge from other modalities,\nsuch as images and text, to assist in point cloud feature learning has been\nconsidered a promising avenue. Existing methods have demonstrated the\neffectiveness of multi-modal contrastive training and feature distillation on\npoint clouds. However, challenges remain, including the requirement for paired\ntriplet data, redundancy and ambiguity in supervised features, and the\ndisruption of the original priors. In this paper, we propose a\nlanguage-assisted approach to point cloud feature learning (LAST-PCL),\nenriching semantic concepts through LLMs-based text enrichment. We achieve\nde-redundancy and feature dimensionality reduction without compromising textual\npriors by statistical-based and training-free significant feature selection.\nFurthermore, we also delve into an in-depth analysis of the impact of text\ncontrastive training on the point cloud. Extensive experiments validate that\nthe proposed method learns semantically meaningful point cloud features and\nachieves state-of-the-art or comparable performance in 3D semantic\nsegmentation, 3D object detection, and 3D scene classification tasks.\n", "date": "2023-12-18", "categories": []}
{"id": "2209.08263", "title": "Scalable SoftGroup for 3D Instance Segmentation on Point Clouds", "abstract": "  This paper considers a network referred to as SoftGroup for accurate and\nscalable 3D instance segmentation. Existing state-of-the-art methods produce\nhard semantic predictions followed by grouping instance segmentation results.\nUnfortunately, errors stemming from hard decisions propagate into the grouping,\nresulting in poor overlap between predicted instances and ground truth and\nsubstantial false positives. To address the abovementioned problems, SoftGroup\nallows each point to be associated with multiple classes to mitigate the\nuncertainty stemming from semantic prediction. It also suppresses false\npositive instances by learning to categorize them as background. Regarding\nscalability, the existing fast methods require computational time on the order\nof tens of seconds on large-scale scenes, which is unsatisfactory and far from\napplicable for real-time. Our finding is that the $k$-Nearest Neighbor ($k$-NN)\nmodule, which serves as the prerequisite of grouping, introduces a\ncomputational bottleneck. SoftGroup is extended to resolve this computational\nbottleneck, referred to as SoftGroup++. The proposed SoftGroup++ reduces time\ncomplexity with octree $k$-NN and reduces search space with class-aware pyramid\nscaling and late devoxelization. Experimental results on various indoor and\noutdoor datasets demonstrate the efficacy and generality of the proposed\nSoftGroup and SoftGroup++. Their performances surpass the best-performing\nbaseline by a large margin (6\\% $\\sim$ 16\\%) in terms of AP$_{50}$. On datasets\nwith large-scale scenes, SoftGroup++ achieves a 6$\\times$ speed boost on\naverage compared to SoftGroup. Furthermore, SoftGroup can be extended to\nperform object detection and panoptic segmentation with nontrivial improvements\nover existing methods. The source code and trained models are available at\n\\url{https://github.com/thangvubk/SoftGroup}.\n", "date": "2022-09-17", "categories": []}
{"id": "2306.08659", "title": "Explore In-Context Learning for 3D Point Cloud Understanding", "abstract": "  With the rise of large-scale models trained on broad data, in-context\nlearning has become a new learning paradigm that has demonstrated significant\npotential in natural language processing and computer vision tasks. Meanwhile,\nin-context learning is still largely unexplored in the 3D point cloud domain.\nAlthough masked modeling has been successfully applied for in-context learning\nin 2D vision, directly extending it to 3D point clouds remains a formidable\nchallenge. In the case of point clouds, the tokens themselves are the point\ncloud positions (coordinates) that are masked during inference. Moreover,\nposition embedding in previous works may inadvertently introduce information\nleakage. To address these challenges, we introduce a novel framework, named\nPoint-In-Context, designed especially for in-context learning in 3D point\nclouds, where both inputs and outputs are modeled as coordinates for each task.\nAdditionally, we propose the Joint Sampling module, carefully designed to work\nin tandem with the general point sampling operator, effectively resolving the\naforementioned technical issues. We conduct extensive experiments to validate\nthe versatility and adaptability of our proposed methods in handling a wide\nrange of tasks.\n", "date": "2023-06-14", "categories": []}
{"id": "2312.17232", "title": "Segment3D: Learning Fine-Grained Class-Agnostic 3D Segmentation without\n  Manual Labels", "abstract": "  Current 3D scene segmentation methods are heavily dependent on manually\nannotated 3D training datasets. Such manual annotations are labor-intensive,\nand often lack fine-grained details. Importantly, models trained on this data\ntypically struggle to recognize object classes beyond the annotated classes,\ni.e., they do not generalize well to unseen domains and require additional\ndomain-specific annotations. In contrast, 2D foundation models demonstrate\nstrong generalization and impressive zero-shot abilities, inspiring us to\nincorporate these characteristics from 2D models into 3D models. Therefore, we\nexplore the use of image segmentation foundation models to automatically\ngenerate training labels for 3D segmentation. We propose Segment3D, a method\nfor class-agnostic 3D scene segmentation that produces high-quality 3D\nsegmentation masks. It improves over existing 3D segmentation models\n(especially on fine-grained masks), and enables easily adding new training data\nto further boost the segmentation performance -- all without the need for\nmanual training labels.\n", "date": "2023-12-28", "categories": []}
{"id": "2312.16170", "title": "EmbodiedScan: A Holistic Multi-Modal 3D Perception Suite Towards\n  Embodied AI", "abstract": "  In the realm of computer vision and robotics, embodied agents are expected to\nexplore their environment and carry out human instructions. This necessitates\nthe ability to fully understand 3D scenes given their first-person observations\nand contextualize them into language for interaction. However, traditional\nresearch focuses more on scene-level input and output setups from a global\nview. To address the gap, we introduce EmbodiedScan, a multi-modal, ego-centric\n3D perception dataset and benchmark for holistic 3D scene understanding. It\nencompasses over 5k scans encapsulating 1M ego-centric RGB-D views, 1M language\nprompts, 160k 3D-oriented boxes spanning over 760 categories, some of which\npartially align with LVIS, and dense semantic occupancy with 80 common\ncategories. Building upon this database, we introduce a baseline framework\nnamed Embodied Perceptron. It is capable of processing an arbitrary number of\nmulti-modal inputs and demonstrates remarkable 3D perception capabilities, both\nwithin the two series of benchmarks we set up, i.e., fundamental 3D perception\ntasks and language-grounded tasks, and in the wild. Codes, datasets, and\nbenchmarks will be available at https://github.com/OpenRobotLab/EmbodiedScan.\n", "date": "2023-12-26", "categories": []}
{"id": "2312.14465", "title": "FM-OV3D: Foundation Model-based Cross-modal Knowledge Blending for\n  Open-Vocabulary 3D Detection", "abstract": "  The superior performances of pre-trained foundation models in various visual\ntasks underscore their potential to enhance the 2D models' open-vocabulary\nability. Existing methods explore analogous applications in the 3D space.\nHowever, most of them only center around knowledge extraction from singular\nfoundation models, which limits the open-vocabulary ability of 3D models. We\nhypothesize that leveraging complementary pre-trained knowledge from various\nfoundation models can improve knowledge transfer from 2D pre-trained visual\nlanguage models to the 3D space. In this work, we propose FM-OV3D, a method of\nFoundation Model-based Cross-modal Knowledge Blending for Open-Vocabulary 3D\nDetection, which improves the open-vocabulary localization and recognition\nabilities of 3D model by blending knowledge from multiple pre-trained\nfoundation models, achieving true open-vocabulary without facing constraints\nfrom original 3D datasets. Specifically, to learn the open-vocabulary 3D\nlocalization ability, we adopt the open-vocabulary localization knowledge of\nthe Grounded-Segment-Anything model. For open-vocabulary 3D recognition\nability, We leverage the knowledge of generative foundation models, including\nGPT-3 and Stable Diffusion models, and cross-modal discriminative models like\nCLIP. The experimental results on two popular benchmarks for open-vocabulary 3D\nobject detection show that our model efficiently learns knowledge from multiple\nfoundation models to enhance the open-vocabulary ability of the 3D model and\nsuccessfully achieves state-of-the-art performance in open-vocabulary 3D object\ndetection tasks. Code is released at\nhttps://github.com/dmzhang0425/FM-OV3D.git.\n", "date": "2023-12-22", "categories": []}
{"id": "2306.09077", "title": "Estimating Generic 3D Room Structures from 2D Annotations", "abstract": "  Indoor rooms are among the most common use cases in 3D scene understanding.\nCurrent state-of-the-art methods for this task are driven by large annotated\ndatasets. Room layouts are especially important, consisting of structural\nelements in 3D, such as wall, floor, and ceiling. However, they are difficult\nto annotate, especially on pure RGB video. We propose a novel method to produce\ngeneric 3D room layouts just from 2D segmentation masks, which are easy to\nannotate for humans. Based on these 2D annotations, we automatically\nreconstruct 3D plane equations for the structural elements and their spatial\nextent in the scene, and connect adjacent elements at the appropriate contact\nedges. We annotate and publicly release 2246 3D room layouts on the\nRealEstate10k dataset, containing YouTube videos. We demonstrate the high\nquality of these 3D layouts annotations with extensive experiments.\n", "date": "2023-06-15", "categories": []}
{"id": "2303.11048", "title": "SGFormer: Semantic Graph Transformer for Point Cloud-based 3D Scene\n  Graph Generation", "abstract": "  In this paper, we propose a novel model called SGFormer, Semantic Graph\nTransFormer for point cloud-based 3D scene graph generation. The task aims to\nparse a point cloud-based scene into a semantic structural graph, with the core\nchallenge of modeling the complex global structure. Existing methods based on\ngraph convolutional networks (GCNs) suffer from the over-smoothing dilemma and\ncan only propagate information from limited neighboring nodes. In contrast,\nSGFormer uses Transformer layers as the base building block to allow global\ninformation passing, with two types of newly-designed layers tailored for the\n3D scene graph generation task. Specifically, we introduce the graph embedding\nlayer to best utilize the global information in graph edges while maintaining\ncomparable computation costs. Furthermore, we propose the semantic injection\nlayer to leverage linguistic knowledge from large-scale language model (i.e.,\nChatGPT), to enhance objects' visual features. We benchmark our SGFormer on the\nestablished 3DSSG dataset and achieve a 40.94% absolute improvement in\nrelationship prediction's R@50 and an 88.36% boost on the subset with complex\nscenes over the state-of-the-art. Our analyses further show SGFormer's\nsuperiority in the long-tail and zero-shot scenarios. Our source code is\navailable at https://github.com/Andy20178/SGFormer.\n", "date": "2023-03-20", "categories": []}
{"id": "2312.12418", "title": "LASA: Instance Reconstruction from Real Scans using A Large-scale\n  Aligned Shape Annotation Dataset", "abstract": "  Instance shape reconstruction from a 3D scene involves recovering the full\ngeometries of multiple objects at the semantic instance level. Many methods\nleverage data-driven learning due to the intricacies of scene complexity and\nsignificant indoor occlusions. Training these methods often requires a\nlarge-scale, high-quality dataset with aligned and paired shape annotations\nwith real-world scans. Existing datasets are either synthetic or misaligned,\nrestricting the performance of data-driven methods on real data. To this end,\nwe introduce LASA, a Large-scale Aligned Shape Annotation Dataset comprising\n10,412 high-quality CAD annotations aligned with 920 real-world scene scans\nfrom ArkitScenes, created manually by professional artists. On this top, we\npropose a novel Diffusion-based Cross-Modal Shape Reconstruction (DisCo)\nmethod. It is empowered by a hybrid feature aggregation design to fuse\nmulti-modal inputs and recover high-fidelity object geometries. Besides, we\npresent an Occupancy-Guided 3D Object Detection (OccGOD) method and demonstrate\nthat our shape annotations provide scene occupancy clues that can further\nimprove 3D object detection. Supported by LASA, extensive experiments show that\nour methods achieve state-of-the-art performance in both instance-level scene\nreconstruction and 3D object detection tasks.\n", "date": "2023-12-19", "categories": []}
{"id": "2312.08054", "title": "Semantic Complete Scene Forecasting from a 4D Dynamic Point Cloud\n  Sequence", "abstract": "  We study a new problem of semantic complete scene forecasting (SCSF) in this\nwork. Given a 4D dynamic point cloud sequence, our goal is to forecast the\ncomplete scene corresponding to the future next frame along with its semantic\nlabels. To tackle this challenging problem, we properly model the synergetic\nrelationship between future forecasting and semantic scene completion through a\nnovel network named SCSFNet. SCSFNet leverages a hybrid geometric\nrepresentation for high-resolution complete scene forecasting. To leverage\nmulti-frame observation as well as the understanding of scene dynamics to ease\nthe completion task, SCSFNet introduces an attention-based skip connection\nscheme. To ease the need to model occlusion variations and to better focus on\nthe occluded part, SCSFNet utilizes auxiliary visibility grids to guide the\nforecasting task. To evaluate the effectiveness of SCSFNet, we conduct\nexperiments on various benchmarks including two large-scale indoor benchmarks\nwe contributed and the outdoor SemanticKITTI benchmark. Extensive experiments\nshow SCSFNet outperforms baseline methods on multiple metrics by a large\nmargin, and also prove the synergy between future forecasting and semantic\nscene completion.\n", "date": "2023-12-13", "categories": []}
{"id": "2312.07378", "title": "X4D-SceneFormer: Enhanced Scene Understanding on 4D Point Cloud Videos\n  through Cross-modal Knowledge Transfer", "abstract": "  The field of 4D point cloud understanding is rapidly developing with the goal\nof analyzing dynamic 3D point cloud sequences. However, it remains a\nchallenging task due to the sparsity and lack of texture in point clouds.\nMoreover, the irregularity of point cloud poses a difficulty in aligning\ntemporal information within video sequences. To address these issues, we\npropose a novel cross-modal knowledge transfer framework, called\nX4D-SceneFormer. This framework enhances 4D-Scene understanding by transferring\ntexture priors from RGB sequences using a Transformer architecture with\ntemporal relationship mining. Specifically, the framework is designed with a\ndual-branch architecture, consisting of an 4D point cloud transformer and a\nGradient-aware Image Transformer (GIT). During training, we employ multiple\nknowledge transfer techniques, including temporal consistency losses and masked\nself-attention, to strengthen the knowledge transfer between modalities. This\nleads to enhanced performance during inference using single-modal 4D point\ncloud inputs. Extensive experiments demonstrate the superior performance of our\nframework on various 4D point cloud video understanding tasks, including action\nrecognition, action segmentation and semantic segmentation. The results achieve\n1st places, i.e., 85.3% (+7.9%) accuracy and 47.3% (+5.0%) mIoU for 4D action\nsegmentation and semantic segmentation, on the HOI4D\nchallenge\\footnote{\\url{http://www.hoi4d.top/}.}, outperforming previous\nstate-of-the-art by a large margin. We release the code at\nhttps://github.com/jinglinglingling/X4D\n", "date": "2023-12-12", "categories": []}
{"id": "2312.03015", "title": "PartSLIP++: Enhancing Low-Shot 3D Part Segmentation via Multi-View\n  Instance Segmentation and Maximum Likelihood Estimation", "abstract": "  Open-world 3D part segmentation is pivotal in diverse applications such as\nrobotics and AR/VR. Traditional supervised methods often grapple with limited\n3D data availability and struggle to generalize to unseen object categories.\nPartSLIP, a recent advancement, has made significant strides in zero- and\nfew-shot 3D part segmentation. This is achieved by harnessing the capabilities\nof the 2D open-vocabulary detection module, GLIP, and introducing a heuristic\nmethod for converting and lifting multi-view 2D bounding box predictions into\n3D segmentation masks. In this paper, we introduce PartSLIP++, an enhanced\nversion designed to overcome the limitations of its predecessor. Our approach\nincorporates two major improvements. First, we utilize a pre-trained 2D\nsegmentation model, SAM, to produce pixel-wise 2D segmentations, yielding more\nprecise and accurate annotations than the 2D bounding boxes used in PartSLIP.\nSecond, PartSLIP++ replaces the heuristic 3D conversion process with an\ninnovative modified Expectation-Maximization algorithm. This algorithm\nconceptualizes 3D instance segmentation as unobserved latent variables, and\nthen iteratively refines them through an alternating process of 2D-3D matching\nand optimization with gradient descent. Through extensive evaluations, we show\nthat PartSLIP++ demonstrates better performance over PartSLIP in both low-shot\n3D semantic and instance-based object part segmentation tasks. Code released at\nhttps://github.com/zyc00/PartSLIP2.\n", "date": "2023-12-05", "categories": []}
{"id": "2312.03026", "title": "Uni3DL: Unified Model for 3D and Language Understanding", "abstract": "  In this work, we present Uni3DL, a unified model for 3D and Language\nunderstanding. Distinct from existing unified vision-language models in 3D\nwhich are limited in task variety and predominantly dependent on projected\nmulti-view images, Uni3DL operates directly on point clouds. This approach\nsignificantly expands the range of supported tasks in 3D, encompassing both\nvision and vision-language tasks in 3D. At the core of Uni3DL, a query\ntransformer is designed to learn task-agnostic semantic and mask outputs by\nattending to 3D visual features, and a task router is employed to selectively\ngenerate task-specific outputs required for diverse tasks. With a unified\narchitecture, our Uni3DL model enjoys seamless task decomposition and\nsubstantial parameter sharing across tasks. Uni3DL has been rigorously\nevaluated across diverse 3D vision-language understanding tasks, including\nsemantic segmentation, object detection, instance segmentation, visual\ngrounding, 3D captioning, and text-3D cross-modal retrieval. It demonstrates\nperformance on par with or surpassing state-of-the-art (SOTA) task-specific\nmodels. We hope our benchmark and Uni3DL model will serve as a solid step to\nease future research in unified models in the realm of 3D and language\nunderstanding. Project page: https://uni3dl.github.io.\n", "date": "2023-12-05", "categories": []}
{"id": "2302.02858", "title": "TR3D: Towards Real-Time Indoor 3D Object Detection", "abstract": "  Recently, sparse 3D convolutions have changed 3D object detection. Performing\non par with the voting-based approaches, 3D CNNs are memory-efficient and scale\nto large scenes better. However, there is still room for improvement. With a\nconscious, practice-oriented approach to problem-solving, we analyze the\nperformance of such methods and localize the weaknesses. Applying modifications\nthat resolve the found issues one by one, we end up with TR3D: a fast\nfully-convolutional 3D object detection model trained end-to-end, that achieves\nstate-of-the-art results on the standard benchmarks, ScanNet v2, SUN RGB-D, and\nS3DIS. Moreover, to take advantage of both point cloud and RGB inputs, we\nintroduce an early fusion of 2D and 3D features. We employ our fusion module to\nmake conventional 3D object detection methods multimodal and demonstrate an\nimpressive boost in performance. Our model with early feature fusion, which we\nrefer to as TR3D+FF, outperforms existing 3D object detection approaches on the\nSUN RGB-D dataset. Overall, besides being accurate, both TR3D and TR3D+FF\nmodels are lightweight, memory-efficient, and fast, thereby marking another\nmilestone on the way toward real-time 3D object detection. Code is available at\nhttps://github.com/SamsungLabs/tr3d .\n", "date": "2023-02-06", "categories": []}
{"id": "2312.02208", "title": "A Data-efficient Framework for Robotics Large-scale LiDAR Scene Parsing", "abstract": "  Existing state-of-the-art 3D point clouds understanding methods only perform\nwell in a fully supervised manner. To the best of our knowledge, there exists\nno unified framework which simultaneously solves the downstream high-level\nunderstanding tasks, especially when labels are extremely limited. This work\npresents a general and simple framework to tackle point clouds understanding\nwhen labels are limited. We propose a novel unsupervised region expansion based\nclustering method for generating clusters. More importantly, we innovatively\npropose to learn to merge the over-divided clusters based on the local\nlow-level geometric property similarities and the learned high-level feature\nsimilarities supervised by weak labels. Hence, the true weak labels guide\npseudo labels merging taking both geometric and semantic feature correlations\ninto consideration. Finally, the self-supervised reconstruction and data\naugmentation optimization modules are proposed to guide the propagation of\nlabels among semantically similar points within a scene. Experimental Results\ndemonstrate that our framework has the best performance among the three most\nimportant weakly supervised point clouds understanding tasks including semantic\nsegmentation, instance segmentation, and object detection even when limited\npoints are labeled, under the data-efficient settings for the large-scale 3D\nsemantic scene parsing. The developed techniques have postentials to be applied\nto downstream tasks for better representations in robotic manipulation and\nrobotic autonomous navigation. Codes and models are publicly available at:\nhttps://github.com/KangchengLiu.\n", "date": "2023-12-03", "categories": []}
{"id": "2312.01148", "title": "Has Anything Changed? 3D Change Detection by 2D Segmentation Masks", "abstract": "  As capturing devices become common, 3D scans of interior spaces are acquired\non a daily basis. Through scene comparison over time, information about objects\nin the scene and their changes is inferred. This information is important for\nrobots and AR and VR devices, in order to operate in an immersive virtual\nexperience. We thus propose an unsupervised object discovery method that\nidentifies added, moved, or removed objects without any prior knowledge of what\nobjects exist in the scene. We model this problem as a combination of a 3D\nchange detection and a 2D segmentation task. Our algorithm leverages generic 2D\nsegmentation masks to refine an initial but incomplete set of 3D change\ndetections. The initial changes, acquired through render-and-compare likely\ncorrespond to movable objects. The incomplete detections are refined through\ngraph optimization, distilling the information of the 2D segmentation masks in\nthe 3D space. Experiments on the 3Rscan dataset prove that our method\noutperforms competitive baselines, with SoTA results.\n", "date": "2023-12-02", "categories": []}
{"id": "2312.01262", "title": "A Review and A Robust Framework of Data-Efficient 3D Scene Parsing with\n  Traditional/Learned 3D Descriptors", "abstract": "  Existing state-of-the-art 3D point cloud understanding methods merely perform\nwell in a fully supervised manner. To the best of our knowledge, there exists\nno unified framework that simultaneously solves the downstream high-level\nunderstanding tasks including both segmentation and detection, especially when\nlabels are extremely limited. This work presents a general and simple framework\nto tackle point cloud understanding when labels are limited. The first\ncontribution is that we have done extensive methodology comparisons of\ntraditional and learned 3D descriptors for the task of weakly supervised 3D\nscene understanding, and validated that our adapted traditional PFH-based 3D\ndescriptors show excellent generalization ability across different domains. The\nsecond contribution is that we proposed a learning-based region merging\nstrategy based on the affinity provided by both the traditional/learned 3D\ndescriptors and learned semantics. The merging process takes both low-level\ngeometric and high-level semantic feature correlations into consideration.\nExperimental results demonstrate that our framework has the best performance\namong the three most important weakly supervised point clouds understanding\ntasks including semantic segmentation, instance segmentation, and object\ndetection even when very limited number of points are labeled. Our method,\ntermed Region Merging 3D (RM3D), has superior performance on ScanNet\ndata-efficient learning online benchmarks and other four large-scale 3D\nunderstanding benchmarks under various experimental settings, outperforming\ncurrent arts by a margin for various 3D understanding tasks without complicated\nlearning strategies such as active learning.\n", "date": "2023-12-03", "categories": []}
{"id": "2311.18482", "title": "Language Embedded 3D Gaussians for Open-Vocabulary Scene Understanding", "abstract": "  Open-vocabulary querying in 3D space is challenging but essential for scene\nunderstanding tasks such as object localization and segmentation.\nLanguage-embedded scene representations have made progress by incorporating\nlanguage features into 3D spaces. However, their efficacy heavily depends on\nneural networks that are resource-intensive in training and rendering. Although\nrecent 3D Gaussians offer efficient and high-quality novel view synthesis,\ndirectly embedding language features in them leads to prohibitive memory usage\nand decreased performance. In this work, we introduce Language Embedded 3D\nGaussians, a novel scene representation for open-vocabulary query tasks.\nInstead of embedding high-dimensional raw semantic features on 3D Gaussians, we\npropose a dedicated quantization scheme that drastically alleviates the memory\nrequirement, and a novel embedding procedure that achieves smoother yet high\naccuracy query, countering the multi-view feature inconsistencies and the\nhigh-frequency inductive bias in point-based representations. Our comprehensive\nexperiments show that our representation achieves the best visual quality and\nlanguage querying accuracy across current language-embedded representations,\nwhile maintaining real-time rendering frame rates on a single desktop GPU.\n", "date": "2023-11-30", "categories": []}
{"id": "2306.15670", "title": "Symphonize 3D Semantic Scene Completion with Contextual Instance Queries", "abstract": "  `3D Semantic Scene Completion (SSC) has emerged as a nascent and pivotal\nundertaking in autonomous driving, aiming to predict voxel occupancy within\nvolumetric scenes. However, prevailing methodologies primarily focus on\nvoxel-wise feature aggregation, while neglecting instance semantics and scene\ncontext. In this paper, we present a novel paradigm termed Symphonies\n(Scene-from-Insts), that delves into the integration of instance queries to\norchestrate 2D-to-3D reconstruction and 3D scene modeling. Leveraging our\nproposed Serial Instance-Propagated Attentions, Symphonies dynamically encodes\ninstance-centric semantics, facilitating intricate interactions between\nimage-based and volumetric domains. Simultaneously, Symphonies enables holistic\nscene comprehension by capturing context through the efficient fusion of\ninstance queries, alleviating geometric ambiguity such as occlusion and\nperspective errors through contextual scene reasoning. Experimental results\ndemonstrate that Symphonies achieves state-of-the-art performance on\nchallenging benchmarks SemanticKITTI and SSCBench-KITTI-360, yielding\nremarkable mIoU scores of 15.04 and 18.58, respectively. These results showcase\nthe paradigm's promising advancements. The code is available at\nhttps://github.com/hustvl/Symphonies.\n", "date": "2023-06-27", "categories": []}
{"id": "2311.12174", "title": "LABELMAKER: Automatic Semantic Label Generation from RGB-D Trajectories", "abstract": "  Semantic annotations are indispensable to train or evaluate perception\nmodels, yet very costly to acquire. This work introduces a fully automated\n2D/3D labeling framework that, without any human intervention, can generate\nlabels for RGB-D scans at equal (or better) level of accuracy than comparable\nmanually annotated datasets such as ScanNet. Our approach is based on an\nensemble of state-of-the-art segmentation models and 3D lifting through neural\nrendering. We demonstrate the effectiveness of our LabelMaker pipeline by\ngenerating significantly better labels for the ScanNet datasets and\nautomatically labelling the previously unlabeled ARKitScenes dataset. Code and\nmodels are available at https://labelmaker.org\n", "date": "2023-11-20", "categories": []}
{"id": "2302.06494", "title": "Explicit3D: Graph Network with Spatial Inference for Single Image 3D\n  Object Detection", "abstract": "  Indoor 3D object detection is an essential task in single image scene\nunderstanding, impacting spatial cognition fundamentally in visual reasoning.\nExisting works on 3D object detection from a single image either pursue this\ngoal through independent predictions of each object or implicitly reason over\nall possible objects, failing to harness relational geometric information\nbetween objects. To address this problem, we propose a dynamic sparse graph\npipeline named Explicit3D based on object geometry and semantics features.\nTaking the efficiency into consideration, we further define a relatedness score\nand design a novel dynamic pruning algorithm followed by a cluster sampling\nmethod for sparse scene graph generation and updating. Furthermore, our\nExplicit3D introduces homogeneous matrices and defines new relative loss and\ncorner loss to model the spatial difference between target pairs explicitly.\nInstead of using ground-truth labels as direct supervision, our relative and\ncorner loss are derived from the homogeneous transformation, which renders the\nmodel to learn the geometric consistency between objects. The experimental\nresults on the SUN RGB-D dataset demonstrate that our Explicit3D achieves\nbetter performance balance than the-state-of-the-art.\n", "date": "2023-02-13", "categories": []}
{"id": "2311.11666", "title": "OmniSeg3D: Omniversal 3D Segmentation via Hierarchical Contrastive\n  Learning", "abstract": "  Towards holistic understanding of 3D scenes, a general 3D segmentation method\nis needed that can segment diverse objects without restrictions on object\nquantity or categories, while also reflecting the inherent hierarchical\nstructure. To achieve this, we propose OmniSeg3D, an omniversal segmentation\nmethod aims for segmenting anything in 3D all at once. The key insight is to\nlift multi-view inconsistent 2D segmentations into a consistent 3D feature\nfield through a hierarchical contrastive learning framework, which is\naccomplished by two steps. Firstly, we design a novel hierarchical\nrepresentation based on category-agnostic 2D segmentations to model the\nmulti-level relationship among pixels. Secondly, image features rendered from\nthe 3D feature field are clustered at different levels, which can be further\ndrawn closer or pushed apart according to the hierarchical relationship between\ndifferent levels. In tackling the challenges posed by inconsistent 2D\nsegmentations, this framework yields a global consistent 3D feature field,\nwhich further enables hierarchical segmentation, multi-object selection, and\nglobal discretization. Extensive experiments demonstrate the effectiveness of\nour method on high-quality 3D segmentation and accurate hierarchical structure\nunderstanding. A graphical user interface further facilitates flexible\ninteraction for omniversal 3D segmentation.\n", "date": "2023-11-20", "categories": []}
{"id": "2311.10018", "title": "On the Overconfidence Problem in Semantic 3D Mapping", "abstract": "  Semantic 3D mapping, the process of fusing depth and image segmentation\ninformation between multiple views to build 3D maps annotated with object\nclasses in real-time, is a recent topic of interest. This paper highlights the\nfusion overconfidence problem, in which conventional mapping methods assign\nhigh confidence to the entire map even when they are incorrect, leading to\nmiscalibrated outputs. Several methods to improve uncertainty calibration at\ndifferent stages in the fusion pipeline are presented and compared on the\nScanNet dataset. We show that the most widely used Bayesian fusion strategy is\namong the worst calibrated, and propose a learned pipeline that combines fusion\nand calibration, GLFS, which achieves simultaneously higher accuracy and 3D map\ncalibration while retaining real-time capability. We further illustrate the\nimportance of map calibration on a downstream task by showing that\nincorporating proper semantic fusion on a modular ObjectNav agent improves its\nsuccess rates. Our code will be provided on Github for reproducibility upon\nacceptance.\n", "date": "2023-11-16", "categories": []}
{"id": "2301.08387", "title": "Occlusion Reasoning for Skeleton Extraction of Self-Occluded Tree\n  Canopies", "abstract": "  In this work, we present a method to extract the skeleton of a self-occluded\ntree canopy by estimating the unobserved structures of the tree. A tree\nskeleton compactly describes the topological structure and contains useful\ninformation such as branch geometry, positions and hierarchy. This can be\ncritical to planning contact interactions for agricultural manipulation, yet is\ndifficult to gain due to occlusion by leaves, fruits and other branches. Our\nmethod uses an instance segmentation network to detect visible trunk, branches,\nand twigs. Then, based on the observed tree structures, we build a custom 3D\nlikelihood map in the form of an occupancy grid to hypothesize on the presence\nof occluded skeletons through a series of minimum cost path searches. We show\nthat our method outperforms baseline methods in highly occluded scenes,\ndemonstrated through a set of experiments on a synthetic tree dataset.\nQualitative results are also presented on a real tree dataset collected from\nthe field.\n", "date": "2023-01-20", "categories": []}
{"id": "2209.05629", "title": "Leveraging Large (Visual) Language Models for Robot 3D Scene\n  Understanding", "abstract": "  Abstract semantic 3D scene understanding is a problem of critical importance\nin robotics. As robots still lack the common-sense knowledge about household\nobjects and locations of an average human, we investigate the use of\npre-trained language models to impart common sense for scene understanding. We\nintroduce and compare a wide range of scene classification paradigms that\nleverage language only (zero-shot, embedding-based, and structured-language) or\nvision and language (zero-shot and fine-tuned). We find that the best\napproaches in both categories yield $\\sim 70\\%$ room classification accuracy,\nexceeding the performance of pure-vision and graph classifiers. We also find\nsuch methods demonstrate notable generalization and transfer capabilities\nstemming from their use of language.\n", "date": "2022-09-12", "categories": []}
{"id": "2304.13219", "title": "ZRG: A Dataset for Multimodal 3D Residential Rooftop Understanding", "abstract": "  A crucial part of any home is the roof over our heads to protect us from the\nelements. In this paper we present the Zeitview Rooftop Geometry (ZRG) dataset\nfor residential rooftop understanding. ZRG is a large-scale residential rooftop\ndataset of over 20k properties collected through roof inspections from across\nthe U.S. and contains multiple modalities including high resolution aerial\northomosaics, digital surface models (DSM), colored point clouds, and 3D roof\nwireframe annotations. We provide an in-depth analysis and perform several\nexperimental baselines including roof outline extraction, monocular height\nestimation, and planar roof structure extraction, to illustrate a few of the\nnumerous potential applications unlocked by this dataset.\n", "date": "2023-04-26", "categories": []}
{"id": "2309.15702", "title": "SGRec3D: Self-Supervised 3D Scene Graph Learning via Object-Level Scene\n  Reconstruction", "abstract": "  In the field of 3D scene understanding, 3D scene graphs have emerged as a new\nscene representation that combines geometric and semantic information about\nobjects and their relationships. However, learning semantic 3D scene graphs in\na fully supervised manner is inherently difficult as it requires not only\nobject-level annotations but also relationship labels. While pre-training\napproaches have helped to boost the performance of many methods in various\nfields, pre-training for 3D scene graph prediction has received little\nattention. Furthermore, we find in this paper that classical contrastive point\ncloud-based pre-training approaches are ineffective for 3D scene graph\nlearning. To this end, we present SGRec3D, a novel self-supervised pre-training\nmethod for 3D scene graph prediction. We propose to reconstruct the 3D input\nscene from a graph bottleneck as a pretext task. Pre-training SGRec3D does not\nrequire object relationship labels, making it possible to exploit large-scale\n3D scene understanding datasets, which were off-limits for 3D scene graph\nlearning before. Our experiments demonstrate that in contrast to recent point\ncloud-based pre-training approaches, our proposed pre-training improves the 3D\nscene graph prediction considerably, which results in SOTA performance,\noutperforming other 3D scene graph models by +10% on object prediction and +4%\non relationship prediction. Additionally, we show that only using a small\nsubset of 10% labeled data during fine-tuning is sufficient to outperform the\nsame model without pre-training.\n", "date": "2023-09-27", "categories": []}
{"id": "2311.01989", "title": "Leveraging Large-Scale Pretrained Vision Foundation Models for\n  Label-Efficient 3D Point Cloud Segmentation", "abstract": "  Recently, large-scale pre-trained models such as Segment-Anything Model (SAM)\nand Contrastive Language-Image Pre-training (CLIP) have demonstrated remarkable\nsuccess and revolutionized the field of computer vision. These foundation\nvision models effectively capture knowledge from a large-scale broad data with\ntheir vast model parameters, enabling them to perform zero-shot segmentation on\npreviously unseen data without additional training. While they showcase\ncompetence in 2D tasks, their potential for enhancing 3D scene understanding\nremains relatively unexplored. To this end, we present a novel framework that\nadapts various foundational models for the 3D point cloud segmentation task.\nOur approach involves making initial predictions of 2D semantic masks using\ndifferent large vision models. We then project these mask predictions from\nvarious frames of RGB-D video sequences into 3D space. To generate robust 3D\nsemantic pseudo labels, we introduce a semantic label fusion strategy that\neffectively combines all the results via voting. We examine diverse scenarios,\nlike zero-shot learning and limited guidance from sparse 2D point labels, to\nassess the pros and cons of different vision foundation models. Our approach is\nexperimented on ScanNet dataset for 3D indoor scenes, and the results\ndemonstrate the effectiveness of adopting general 2D foundation models on\nsolving 3D point cloud segmentation tasks.\n", "date": "2023-11-03", "categories": []}
{"id": "2311.02873", "title": "OVIR-3D: Open-Vocabulary 3D Instance Retrieval Without Training on 3D\n  Data", "abstract": "  This work presents OVIR-3D, a straightforward yet effective method for\nopen-vocabulary 3D object instance retrieval without using any 3D data for\ntraining. Given a language query, the proposed method is able to return a\nranked set of 3D object instance segments based on the feature similarity of\nthe instance and the text query. This is achieved by a multi-view fusion of\ntext-aligned 2D region proposals into 3D space, where the 2D region proposal\nnetwork could leverage 2D datasets, which are more accessible and typically\nlarger than 3D datasets. The proposed fusion process is efficient as it can be\nperformed in real-time for most indoor 3D scenes and does not require\nadditional training in 3D space. Experiments on public datasets and a real\nrobot show the effectiveness of the method and its potential for applications\nin robot navigation and manipulation.\n", "date": "2023-11-06", "categories": []}
{"id": "2305.08776", "title": "Bridging the Domain Gap: Self-Supervised 3D Scene Understanding with\n  Foundation Models", "abstract": "  Foundation models have achieved remarkable results in 2D and language tasks\nlike image segmentation, object detection, and visual-language understanding.\nHowever, their potential to enrich 3D scene representation learning is largely\nuntapped due to the existence of the domain gap. In this work, we propose an\ninnovative methodology called Bridge3D to address this gap by pre-training 3D\nmodels using features, semantic masks, and captions sourced from foundation\nmodels. Specifically, our method employs semantic masks from foundation models\nto guide the masking and reconstruction process for the masked autoencoder,\nenabling more focused attention on foreground representations. Moreover, we\nbridge the 3D-text gap at the scene level using image captioning foundation\nmodels, thereby facilitating scene-level knowledge distillation. We further\nextend this bridging effort by introducing an innovative object-level knowledge\ndistillation method that harnesses highly accurate object-level masks and\nsemantic text data from foundation models. Our methodology significantly\nsurpasses the performance of existing state-of-the-art methods in 3D object\ndetection and semantic segmentation tasks. For instance, on the ScanNet\ndataset, Bridge3D improves the baseline by a notable margin of 6.3%. Code will\nbe available at: https://github.com/Zhimin-C/Bridge3D\n", "date": "2023-05-15", "categories": []}
{"id": "2311.00457", "title": "Single-view 3D Scene Reconstruction with High-fidelity Shape and Texture", "abstract": "  Reconstructing detailed 3D scenes from single-view images remains a\nchallenging task due to limitations in existing approaches, which primarily\nfocus on geometric shape recovery, overlooking object appearances and fine\nshape details. To address these challenges, we propose a novel framework for\nsimultaneous high-fidelity recovery of object shapes and textures from\nsingle-view images. Our approach utilizes the proposed Single-view neural\nimplicit Shape and Radiance field (SSR) representations to leverage both\nexplicit 3D shape supervision and volume rendering of color, depth, and surface\nnormal images. To overcome shape-appearance ambiguity under partial\nobservations, we introduce a two-stage learning curriculum incorporating both\n3D and 2D supervisions. A distinctive feature of our framework is its ability\nto generate fine-grained textured meshes while seamlessly integrating rendering\ncapabilities into the single-view 3D reconstruction model. This integration\nenables not only improved textured 3D object reconstruction by 27.7% and 11.6%\non the 3D-FRONT and Pix3D datasets, respectively, but also supports the\nrendering of images from novel viewpoints. Beyond individual objects, our\napproach facilitates composing object-level representations into flexible scene\nrepresentations, thereby enabling applications such as holistic scene\nunderstanding and 3D scene editing. We conduct extensive experiments to\ndemonstrate the effectiveness of our method.\n", "date": "2023-11-01", "categories": []}
{"id": "2206.01203", "title": "Box2Mask: Weakly Supervised 3D Semantic Instance Segmentation Using\n  Bounding Boxes", "abstract": "  Current 3D segmentation methods heavily rely on large-scale point-cloud\ndatasets, which are notoriously laborious to annotate. Few attempts have been\nmade to circumvent the need for dense per-point annotations. In this work, we\nlook at weakly-supervised 3D semantic instance segmentation. The key idea is to\nleverage 3D bounding box labels which are easier and faster to annotate.\nIndeed, we show that it is possible to train dense segmentation models using\nonly bounding box labels. At the core of our method, \\name{}, lies a deep\nmodel, inspired by classical Hough voting, that directly votes for bounding box\nparameters, and a clustering method specifically tailored to bounding box\nvotes. This goes beyond commonly used center votes, which would not fully\nexploit the bounding box annotations. On ScanNet test, our weakly supervised\nmodel attains leading performance among other weakly supervised approaches (+18\nmAP@50). Remarkably, it also achieves 97% of the mAP@50 score of current fully\nsupervised models. To further illustrate the practicality of our work, we train\nBox2Mask on the recently released ARKitScenes dataset which is annotated with\n3D bounding boxes only, and show, for the first time, compelling 3D instance\nsegmentation masks.\n", "date": "2022-06-02", "categories": []}
{"id": "2310.18773", "title": "CityRefer: Geography-aware 3D Visual Grounding Dataset on City-scale\n  Point Cloud Data", "abstract": "  City-scale 3D point cloud is a promising way to express detailed and\ncomplicated outdoor structures. It encompasses both the appearance and geometry\nfeatures of segmented city components, including cars, streets, and buildings,\nthat can be utilized for attractive applications such as user-interactive\nnavigation of autonomous vehicles and drones. However, compared to the\nextensive text annotations available for images and indoor scenes, the scarcity\nof text annotations for outdoor scenes poses a significant challenge for\nachieving these applications. To tackle this problem, we introduce the\nCityRefer dataset for city-level visual grounding. The dataset consists of 35k\nnatural language descriptions of 3D objects appearing in SensatUrban city\nscenes and 5k landmarks labels synchronizing with OpenStreetMap. To ensure the\nquality and accuracy of the dataset, all descriptions and labels in the\nCityRefer dataset are manually verified. We also have developed a baseline\nsystem that can learn encoded language descriptions, 3D object instances, and\ngeographical information about the city's landmarks to perform visual grounding\non the CityRefer dataset. To the best of our knowledge, the CityRefer dataset\nis the largest city-level visual grounding dataset for localizing specific 3D\nobjects.\n", "date": "2023-10-28", "categories": []}
{"id": "2310.19516", "title": "Generating Context-Aware Natural Answers for Questions in 3D Scenes", "abstract": "  3D question answering is a young field in 3D vision-language that is yet to\nbe explored. Previous methods are limited to a pre-defined answer space and\ncannot generate answers naturally. In this work, we pivot the question\nanswering task to a sequence generation task to generate free-form natural\nanswers for questions in 3D scenes (Gen3DQA). To this end, we optimize our\nmodel directly on the language rewards to secure the global sentence semantics.\nHere, we also adapt a pragmatic language understanding reward to further\nimprove the sentence quality. Our method sets a new SOTA on the ScanQA\nbenchmark (CIDEr score 72.22/66.57 on the test sets).\n", "date": "2023-10-30", "categories": []}
{"id": "2310.15712", "title": "GNeSF: Generalizable Neural Semantic Fields", "abstract": "  3D scene segmentation based on neural implicit representation has emerged\nrecently with the advantage of training only on 2D supervision. However,\nexisting approaches still requires expensive per-scene optimization that\nprohibits generalization to novel scenes during inference. To circumvent this\nproblem, we introduce a generalizable 3D segmentation framework based on\nimplicit representation. Specifically, our framework takes in multi-view image\nfeatures and semantic maps as the inputs instead of only spatial information to\navoid overfitting to scene-specific geometric and semantic information. We\npropose a novel soft voting mechanism to aggregate the 2D semantic information\nfrom different views for each 3D point. In addition to the image features, view\ndifference information is also encoded in our framework to predict the voting\nscores. Intuitively, this allows the semantic information from nearby views to\ncontribute more compared to distant ones. Furthermore, a visibility module is\nalso designed to detect and filter out detrimental information from occluded\nviews. Due to the generalizability of our proposed method, we can synthesize\nsemantic maps or conduct 3D semantic segmentation for novel scenes with solely\n2D semantic supervision. Experimental results show that our approach achieves\ncomparable performance with scene-specific approaches. More importantly, our\napproach can even outperform existing strong supervision-based approaches with\nonly 2D annotations. Our source code is available at:\nhttps://github.com/HLinChen/GNeSF.\n", "date": "2023-10-24", "categories": []}
{"id": "2310.16494", "title": "Lang3DSG: Language-based contrastive pre-training for 3D Scene Graph\n  prediction", "abstract": "  D scene graphs are an emerging 3D scene representation, that models both the\nobjects present in the scene as well as their relationships. However, learning\n3D scene graphs is a challenging task because it requires not only object\nlabels but also relationship annotations, which are very scarce in datasets.\nWhile it is widely accepted that pre-training is an effective approach to\nimprove model performance in low data regimes, in this paper, we find that\nexisting pre-training methods are ill-suited for 3D scene graphs. To solve this\nissue, we present the first language-based pre-training approach for 3D scene\ngraphs, whereby we exploit the strong relationship between scene graphs and\nlanguage. To this end, we leverage the language encoder of CLIP, a popular\nvision-language model, to distill its knowledge into our graph-based network.\nWe formulate a contrastive pre-training, which aligns text embeddings of\nrelationships (subject-predicate-object triplets) and predicted 3D graph\nfeatures. Our method achieves state-of-the-art results on the main semantic 3D\nscene graph benchmark by showing improved effectiveness over pre-training\nbaselines and outperforming all the existing fully supervised scene graph\nprediction methods by a significant margin. Furthermore, since our scene graph\nfeatures are language-aligned, it allows us to query the language space of the\nfeatures in a zero-shot manner. In this paper, we show an example of utilizing\nthis property of the features to predict the room type of a scene without\nfurther training.\n", "date": "2023-10-25", "categories": []}
{"id": "2302.07241", "title": "ConceptFusion: Open-set Multimodal 3D Mapping", "abstract": "  Building 3D maps of the environment is central to robot navigation, planning,\nand interaction with objects in a scene. Most existing approaches that\nintegrate semantic concepts with 3D maps largely remain confined to the\nclosed-set setting: they can only reason about a finite set of concepts,\npre-defined at training time. Further, these maps can only be queried using\nclass labels, or in recent work, using text prompts.\n  We address both these issues with ConceptFusion, a scene representation that\nis (1) fundamentally open-set, enabling reasoning beyond a closed set of\nconcepts and (ii) inherently multimodal, enabling a diverse range of possible\nqueries to the 3D map, from language, to images, to audio, to 3D geometry, all\nworking in concert. ConceptFusion leverages the open-set capabilities of\ntoday's foundation models pre-trained on internet-scale data to reason about\nconcepts across modalities such as natural language, images, and audio. We\ndemonstrate that pixel-aligned open-set features can be fused into 3D maps via\ntraditional SLAM and multi-view fusion approaches. This enables effective\nzero-shot spatial reasoning, not needing any additional training or finetuning,\nand retains long-tailed concepts better than supervised approaches,\noutperforming them by more than 40% margin on 3D IoU. We extensively evaluate\nConceptFusion on a number of real-world datasets, simulated home environments,\na real-world tabletop manipulation task, and an autonomous driving platform. We\nshowcase new avenues for blending foundation models with 3D open-set multimodal\nmapping.\n  For more information, visit our project page https://concept-fusion.github.io\nor watch our 5-minute explainer video\nhttps://www.youtube.com/watch?v=rkXgws8fiDs\n", "date": "2023-02-14", "categories": []}
{"id": "2210.05633", "title": "Habitat-Matterport 3D Semantics Dataset", "abstract": "  We present the Habitat-Matterport 3D Semantics (HM3DSEM) dataset. HM3DSEM is\nthe largest dataset of 3D real-world spaces with densely annotated semantics\nthat is currently available to the academic community. It consists of 142,646\nobject instance annotations across 216 3D spaces and 3,100 rooms within those\nspaces. The scale, quality, and diversity of object annotations far exceed\nthose of prior datasets. A key difference setting apart HM3DSEM from other\ndatasets is the use of texture information to annotate pixel-accurate object\nboundaries. We demonstrate the effectiveness of HM3DSEM dataset for the Object\nGoal Navigation task using different methods. Policies trained using HM3DSEM\nperform outperform those trained on prior datasets. Introduction of HM3DSEM in\nthe Habitat ObjectNav Challenge lead to an increase in participation from 400\nsubmissions in 2021 to 1022 submissions in 2022.\n", "date": "2022-10-11", "categories": []}
{"id": "2309.14616", "title": "NDC-Scene: Boost Monocular 3D Semantic Scene Completion in Normalized\n  Device Coordinates Space", "abstract": "  Monocular 3D Semantic Scene Completion (SSC) has garnered significant\nattention in recent years due to its potential to predict complex semantics and\ngeometry shapes from a single image, requiring no 3D inputs. In this paper, we\nidentify several critical issues in current state-of-the-art methods, including\nthe Feature Ambiguity of projected 2D features in the ray to the 3D space, the\nPose Ambiguity of the 3D convolution, and the Computation Imbalance in the 3D\nconvolution across different depth levels. To address these problems, we devise\na novel Normalized Device Coordinates scene completion network (NDC-Scene) that\ndirectly extends the 2D feature map to a Normalized Device Coordinates (NDC)\nspace, rather than to the world space directly, through progressive restoration\nof the dimension of depth with deconvolution operations. Experiment results\ndemonstrate that transferring the majority of computation from the target 3D\nspace to the proposed normalized device coordinates space benefits monocular\nSSC tasks. Additionally, we design a Depth-Adaptive Dual Decoder to\nsimultaneously upsample and fuse the 2D and 3D feature maps, further improving\noverall performance. Our extensive experiments confirm that the proposed method\nconsistently outperforms state-of-the-art methods on both outdoor SemanticKITTI\nand indoor NYUv2 datasets. Our code are available at\nhttps://github.com/Jiawei-Yao0812/NDCScene.\n", "date": "2023-09-26", "categories": []}
{"id": "2310.07522", "title": "S4C: Self-Supervised Semantic Scene Completion with Neural Fields", "abstract": "  3D semantic scene understanding is a fundamental challenge in computer\nvision. It enables mobile agents to autonomously plan and navigate arbitrary\nenvironments. SSC formalizes this challenge as jointly estimating dense\ngeometry and semantic information from sparse observations of a scene. Current\nmethods for SSC are generally trained on 3D ground truth based on aggregated\nLiDAR scans. This process relies on special sensors and annotation by hand\nwhich are costly and do not scale well. To overcome this issue, our work\npresents the first self-supervised approach to SSC called S4C that does not\nrely on 3D ground truth data. Our proposed method can reconstruct a scene from\na single image and only relies on videos and pseudo segmentation ground truth\ngenerated from off-the-shelf image segmentation network during training. Unlike\nexisting methods, which use discrete voxel grids, we represent scenes as\nimplicit semantic fields. This formulation allows querying any point within the\ncamera frustum for occupancy and semantic class. Our architecture is trained\nthrough rendering-based self-supervised losses. Nonetheless, our method\nachieves performance close to fully supervised state-of-the-art methods.\nAdditionally, our method demonstrates strong generalization capabilities and\ncan synthesize accurate segmentation maps for far away viewpoints.\n", "date": "2023-10-11", "categories": []}
{"id": "2310.08326", "title": "NSM4D: Neural Scene Model Based Online 4D Point Cloud Sequence\n  Understanding", "abstract": "  Understanding 4D point cloud sequences online is of significant practical\nvalue in various scenarios such as VR/AR, robotics, and autonomous driving. The\nkey goal is to continuously analyze the geometry and dynamics of a 3D scene as\nunstructured and redundant point cloud sequences arrive. And the main challenge\nis to effectively model the long-term history while keeping computational costs\nmanageable. To tackle these challenges, we introduce a generic online 4D\nperception paradigm called NSM4D. NSM4D serves as a plug-and-play strategy that\ncan be adapted to existing 4D backbones, significantly enhancing their online\nperception capabilities for both indoor and outdoor scenarios. To efficiently\ncapture the redundant 4D history, we propose a neural scene model that\nfactorizes geometry and motion information by constructing geometry tokens\nseparately storing geometry and motion features. Exploiting the history becomes\nas straightforward as querying the neural scene model. As the sequence\nprogresses, the neural scene model dynamically deforms to align with new\nobservations, effectively providing the historical context and updating itself\nwith the new observations. By employing token representation, NSM4D also\nexhibits robustness to low-level sensor noise and maintains a compact size\nthrough a geometric sampling scheme. We integrate NSM4D with state-of-the-art\n4D perception backbones, demonstrating significant improvements on various\nonline perception benchmarks in indoor and outdoor settings. Notably, we\nachieve a 9.6% accuracy improvement for HOI4D online action segmentation and a\n3.4% mIoU improvement for SemanticKITTI online semantic segmentation.\nFurthermore, we show that NSM4D inherently offers excellent scalability to\nlonger sequences beyond the training set, which is crucial for real-world\napplications.\n", "date": "2023-10-12", "categories": []}
{"id": "2209.14941", "title": "EDA: Explicit Text-Decoupling and Dense Alignment for 3D Visual\n  Grounding", "abstract": "  3D visual grounding aims to find the object within point clouds mentioned by\nfree-form natural language descriptions with rich semantic cues. However,\nexisting methods either extract the sentence-level features coupling all words\nor focus more on object names, which would lose the word-level information or\nneglect other attributes. To alleviate these issues, we present EDA that\nExplicitly Decouples the textual attributes in a sentence and conducts Dense\nAlignment between such fine-grained language and point cloud objects.\nSpecifically, we first propose a text decoupling module to produce textual\nfeatures for every semantic component. Then, we design two losses to supervise\nthe dense matching between two modalities: position alignment loss and semantic\nalignment loss. On top of that, we further introduce a new visual grounding\ntask, locating objects without object names, which can thoroughly evaluate the\nmodel's dense alignment capacity. Through experiments, we achieve\nstate-of-the-art performance on two widely-adopted 3D visual grounding\ndatasets, ScanRefer and SR3D/NR3D, and obtain absolute leadership on our\nnewly-proposed task. The source code is available at\nhttps://github.com/yanmin-wu/EDA.\n", "date": "2022-09-29", "categories": []}
{"id": "2310.03923", "title": "Open-Fusion: Real-time Open-Vocabulary 3D Mapping and Queryable Scene\n  Representation", "abstract": "  Precise 3D environmental mapping is pivotal in robotics. Existing methods\noften rely on predefined concepts during training or are time-intensive when\ngenerating semantic maps. This paper presents Open-Fusion, a groundbreaking\napproach for real-time open-vocabulary 3D mapping and queryable scene\nrepresentation using RGB-D data. Open-Fusion harnesses the power of a\npre-trained vision-language foundation model (VLFM) for open-set semantic\ncomprehension and employs the Truncated Signed Distance Function (TSDF) for\nswift 3D scene reconstruction. By leveraging the VLFM, we extract region-based\nembeddings and their associated confidence maps. These are then integrated with\n3D knowledge from TSDF using an enhanced Hungarian-based feature-matching\nmechanism. Notably, Open-Fusion delivers outstanding annotation-free 3D\nsegmentation for open-vocabulary without necessitating additional 3D training.\nBenchmark tests on the ScanNet dataset against leading zero-shot methods\nhighlight Open-Fusion's superiority. Furthermore, it seamlessly combines the\nstrengths of region-based VLFM and TSDF, facilitating real-time 3D scene\ncomprehension that includes object concepts and open-world semantics. We\nencourage the readers to view the demos on our project page:\nhttps://uark-aicv.github.io/OpenFusion\n", "date": "2023-10-05", "categories": []}
{"id": "2307.11932", "title": "RIC: Rotate-Inpaint-Complete for Generalizable Scene Reconstruction", "abstract": "  General scene reconstruction refers to the task of estimating the full 3D\ngeometry and texture of a scene containing previously unseen objects. In many\npractical applications such as AR/VR, autonomous navigation, and robotics, only\na single view of the scene may be available, making the scene reconstruction\ntask challenging. In this paper, we present a method for scene reconstruction\nby structurally breaking the problem into two steps: rendering novel views via\ninpainting and 2D to 3D scene lifting. Specifically, we leverage the\ngeneralization capability of large visual language models (Dalle-2) to inpaint\nthe missing areas of scene color images rendered from different views. Next, we\nlift these inpainted images to 3D by predicting normals of the inpainted image\nand solving for the missing depth values. By predicting for normals instead of\ndepth directly, our method allows for robustness to changes in depth\ndistributions and scale. With rigorous quantitative evaluation, we show that\nour method outperforms multiple baselines while providing generalization to\nnovel objects and scenes.\n", "date": "2023-07-21", "categories": []}
{"id": "2309.16956", "title": "Model2Scene: Learning 3D Scene Representation via Contrastive\n  Language-CAD Models Pre-training", "abstract": "  Current successful methods of 3D scene perception rely on the large-scale\nannotated point cloud, which is tedious and expensive to acquire. In this\npaper, we propose Model2Scene, a novel paradigm that learns free 3D scene\nrepresentation from Computer-Aided Design (CAD) models and languages. The main\nchallenges are the domain gaps between the CAD models and the real scene's\nobjects, including model-to-scene (from a single model to the scene) and\nsynthetic-to-real (from synthetic model to real scene's object). To handle the\nabove challenges, Model2Scene first simulates a crowded scene by mixing\ndata-augmented CAD models. Next, we propose a novel feature regularization\noperation, termed Deep Convex-hull Regularization (DCR), to project point\nfeatures into a unified convex hull space, reducing the domain gap. Ultimately,\nwe impose contrastive loss on language embedding and the point features of CAD\nmodels to pre-train the 3D network. Extensive experiments verify the learned 3D\nscene representation is beneficial for various downstream tasks, including\nlabel-free 3D object salient detection, label-efficient 3D scene perception and\nzero-shot 3D semantic segmentation. Notably, Model2Scene yields impressive\nlabel-free 3D object salient detection with an average mAP of 46.08\\% and\n55.49\\% on the ScanNet and S3DIS datasets, respectively. The code will be\npublicly available.\n", "date": "2023-09-29", "categories": []}
{"id": "2309.15940", "title": "Context-Aware Entity Grounding with Open-Vocabulary 3D Scene Graphs", "abstract": "  We present an Open-Vocabulary 3D Scene Graph (OVSG), a formal framework for\ngrounding a variety of entities, such as object instances, agents, and regions,\nwith free-form text-based queries. Unlike conventional semantic-based object\nlocalization approaches, our system facilitates context-aware entity\nlocalization, allowing for queries such as ``pick up a cup on a kitchen table\"\nor ``navigate to a sofa on which someone is sitting\". In contrast to existing\nresearch on 3D scene graphs, OVSG supports free-form text input and\nopen-vocabulary querying. Through a series of comparative experiments using the\nScanNet dataset and a self-collected dataset, we demonstrate that our proposed\napproach significantly surpasses the performance of previous semantic-based\nlocalization techniques. Moreover, we highlight the practical application of\nOVSG in real-world robot navigation and manipulation experiments.\n", "date": "2023-09-27", "categories": []}
{"id": "2309.16650", "title": "ConceptGraphs: Open-Vocabulary 3D Scene Graphs for Perception and\n  Planning", "abstract": "  For robots to perform a wide variety of tasks, they require a 3D\nrepresentation of the world that is semantically rich, yet compact and\nefficient for task-driven perception and planning. Recent approaches have\nattempted to leverage features from large vision-language models to encode\nsemantics in 3D representations. However, these approaches tend to produce maps\nwith per-point feature vectors, which do not scale well in larger environments,\nnor do they contain semantic spatial relationships between entities in the\nenvironment, which are useful for downstream planning. In this work, we propose\nConceptGraphs, an open-vocabulary graph-structured representation for 3D\nscenes. ConceptGraphs is built by leveraging 2D foundation models and fusing\ntheir output to 3D by multi-view association. The resulting representations\ngeneralize to novel semantic classes, without the need to collect large 3D\ndatasets or finetune models. We demonstrate the utility of this representation\nthrough a number of downstream planning tasks that are specified through\nabstract (language) prompts and require complex reasoning over spatial and\nsemantic concepts. (Project page: https://concept-graphs.github.io/ Explainer\nvideo: https://youtu.be/mRhNkQwRYnc )\n", "date": "2023-09-28", "categories": []}
{"id": "2304.14880", "title": "SGAligner : 3D Scene Alignment with Scene Graphs", "abstract": "  Building 3D scene graphs has recently emerged as a topic in scene\nrepresentation for several embodied AI applications to represent the world in a\nstructured and rich manner. With their increased use in solving downstream\ntasks (eg, navigation and room rearrangement), can we leverage and recycle them\nfor creating 3D maps of environments, a pivotal step in agent operation? We\nfocus on the fundamental problem of aligning pairs of 3D scene graphs whose\noverlap can range from zero to partial and can contain arbitrary changes. We\npropose SGAligner, the first method for aligning pairs of 3D scene graphs that\nis robust to in-the-wild scenarios (ie, unknown overlap -- if any -- and\nchanges in the environment). We get inspired by multi-modality knowledge graphs\nand use contrastive learning to learn a joint, multi-modal embedding space. We\nevaluate on the 3RScan dataset and further showcase that our method can be used\nfor estimating the transformation between pairs of 3D scenes. Since benchmarks\nfor these tasks are missing, we create them on this dataset. The code,\nbenchmark, and trained models are available on the project website.\n", "date": "2023-04-28", "categories": []}
{"id": "2309.09514", "title": "PanoMixSwap Panorama Mixing via Structural Swapping for Indoor Scene\n  Understanding", "abstract": "  The volume and diversity of training data are critical for modern deep\nlearningbased methods. Compared to the massive amount of labeled perspective\nimages, 360 panoramic images fall short in both volume and diversity. In this\npaper, we propose PanoMixSwap, a novel data augmentation technique specifically\ndesigned for indoor panoramic images. PanoMixSwap explicitly mixes various\nbackground styles, foreground furniture, and room layouts from the existing\nindoor panorama datasets and generates a diverse set of new panoramic images to\nenrich the datasets. We first decompose each panoramic image into its\nconstituent parts: background style, foreground furniture, and room layout.\nThen, we generate an augmented image by mixing these three parts from three\ndifferent images, such as the foreground furniture from one image, the\nbackground style from another image, and the room structure from the third\nimage. Our method yields high diversity since there is a cubical increase in\nimage combinations. We also evaluate the effectiveness of PanoMixSwap on two\nindoor scene understanding tasks: semantic segmentation and layout estimation.\nOur experiments demonstrate that state-of-the-art methods trained with\nPanoMixSwap outperform their original setting on both tasks consistently.\n", "date": "2023-09-18", "categories": []}
{"id": "2207.04526", "title": "Efficient Multi-Task RGB-D Scene Analysis for Indoor Environments", "abstract": "  Semantic scene understanding is essential for mobile agents acting in various\nenvironments. Although semantic segmentation already provides a lot of\ninformation, details about individual objects as well as the general scene are\nmissing but required for many real-world applications. However, solving\nmultiple tasks separately is expensive and cannot be accomplished in real time\ngiven limited computing and battery capabilities on a mobile platform. In this\npaper, we propose an efficient multi-task approach for RGB-D scene\nanalysis~(EMSANet) that simultaneously performs semantic and instance\nsegmentation~(panoptic segmentation), instance orientation estimation, and\nscene classification. We show that all tasks can be accomplished using a single\nneural network in real time on a mobile platform without diminishing\nperformance - by contrast, the individual tasks are able to benefit from each\nother. In order to evaluate our multi-task approach, we extend the annotations\nof the common RGB-D indoor datasets NYUv2 and SUNRGB-D for instance\nsegmentation and orientation estimation. To the best of our knowledge, we are\nthe first to provide results in such a comprehensive multi-task setting for\nindoor scene analysis on NYUv2 and SUNRGB-D.\n", "date": "2022-07-10", "categories": []}
{"id": "2309.14010", "title": "Variational Inference for Scalable 3D Object-centric Learning", "abstract": "  We tackle the task of scalable unsupervised object-centric representation\nlearning on 3D scenes. Existing approaches to object-centric representation\nlearning show limitations in generalizing to larger scenes as their learning\nprocesses rely on a fixed global coordinate system. In contrast, we propose to\nlearn view-invariant 3D object representations in localized object coordinate\nsystems. To this end, we estimate the object pose and appearance representation\nseparately and explicitly map object representations across views while\nmaintaining object identities. We adopt an amortized variational inference\npipeline that can process sequential input and scalably update object latent\ndistributions online. To handle large-scale scenes with a varying number of\nobjects, we further introduce a Cognitive Map that allows the registration and\nquery of objects on a per-scene global map to achieve scalable representation\nlearning. We explore the object-centric neural radiance field (NeRF) as our 3D\nscene representation, which is jointly modeled within our unsupervised\nobject-centric learning framework. Experimental results on synthetic and real\ndatasets show that our proposed method can infer and maintain object-centric\nrepresentations of 3D scenes and outperforms previous models.\n", "date": "2023-09-25", "categories": []}
{"id": "2309.11133", "title": "Shape Anchor Guided Holistic Indoor Scene Understanding", "abstract": "  This paper proposes a shape anchor guided learning strategy (AncLearn) for\nrobust holistic indoor scene understanding. We observe that the search space\nconstructed by current methods for proposal feature grouping and instance point\nsampling often introduces massive noise to instance detection and mesh\nreconstruction. Accordingly, we develop AncLearn to generate anchors that\ndynamically fit instance surfaces to (i) unmix noise and target-related\nfeatures for offering reliable proposals at the detection stage, and (ii)\nreduce outliers in object point sampling for directly providing well-structured\ngeometry priors without segmentation during reconstruction. We embed AncLearn\ninto a reconstruction-from-detection learning system (AncRec) to generate\nhigh-quality semantic scene models in a purely instance-oriented manner.\nExperiments conducted on the challenging ScanNetv2 dataset demonstrate that our\nshape anchor-based method consistently achieves state-of-the-art performance in\nterms of 3D object detection, layout estimation, and shape reconstruction. The\ncode will be available at https://github.com/Geo-Tell/AncRec.\n", "date": "2023-09-20", "categories": []}
{"id": "2309.09456", "title": "Object2Scene: Putting Objects in Context for Open-Vocabulary 3D\n  Detection", "abstract": "  Point cloud-based open-vocabulary 3D object detection aims to detect 3D\ncategories that do not have ground-truth annotations in the training set. It is\nextremely challenging because of the limited data and annotations (bounding\nboxes with class labels or text descriptions) of 3D scenes. Previous approaches\nleverage large-scale richly-annotated image datasets as a bridge between 3D and\ncategory semantics but require an extra alignment process between 2D images and\n3D points, limiting the open-vocabulary ability of 3D detectors. Instead of\nleveraging 2D images, we propose Object2Scene, the first approach that\nleverages large-scale large-vocabulary 3D object datasets to augment existing\n3D scene datasets for open-vocabulary 3D object detection. Object2Scene inserts\nobjects from different sources into 3D scenes to enrich the vocabulary of 3D\nscene datasets and generates text descriptions for the newly inserted objects.\nWe further introduce a framework that unifies 3D detection and visual\ngrounding, named L3Det, and propose a cross-domain category-level contrastive\nlearning approach to mitigate the domain gap between 3D objects from different\ndatasets. Extensive experiments on existing open-vocabulary 3D object detection\nbenchmarks show that Object2Scene obtains superior performance over existing\nmethods. We further verify the effectiveness of Object2Scene on a new benchmark\nOV-ScanNet-200, by holding out all rare categories as novel categories not seen\nduring training.\n", "date": "2023-09-18", "categories": []}
{"id": "2302.02871", "title": "Top-Down Beats Bottom-Up in 3D Instance Segmentation", "abstract": "  Most 3D instance segmentation methods exploit a bottom-up strategy, typically\nincluding resource-exhaustive post-processing. For point grouping, bottom-up\nmethods rely on prior assumptions about the objects in the form of\nhyperparameters, which are domain-specific and need to be carefully tuned. On\nthe contrary, we address 3D instance segmentation with a TD3D: the pioneering\ncluster-free, fully-convolutional and entirely data-driven approach trained in\nan end-to-end manner. This is the first top-down method outperforming bottom-up\napproaches in 3D domain. With its straightforward pipeline, it demonstrates\noutstanding accuracy and generalization ability on the standard indoor\nbenchmarks: ScanNet v2, its extension ScanNet200, and S3DIS, as well as on the\naerial STPLS3D dataset. Besides, our method is much faster on inference than\nthe current state-of-the-art grouping-based approaches: our flagship\nmodification is 1.9x faster than the most accurate bottom-up method, while\nbeing more accurate, and our faster modification shows state-of-the-art\naccuracy running at 2.6x speed. Code is available at\nhttps://github.com/SamsungLabs/td3d .\n", "date": "2023-02-06", "categories": []}
{"id": "2303.14727", "title": "You Only Need One Thing One Click: Self-Training for Weakly Supervised\n  3D Scene Understanding", "abstract": "  3D scene understanding, e.g., point cloud semantic and instance segmentation,\noften requires large-scale annotated training data, but clearly, point-wise\nlabels are too tedious to prepare. While some recent methods propose to train a\n3D network with small percentages of point labels, we take the approach to an\nextreme and propose ``One Thing One Click,'' meaning that the annotator only\nneeds to label one point per object. To leverage these extremely sparse labels\nin network training, we design a novel self-training approach, in which we\niteratively conduct the training and label propagation, facilitated by a graph\npropagation module. Also, we adopt a relation network to generate the\nper-category prototype to enhance the pseudo label quality and guide the\niterative training. Besides, our model can be compatible to 3D instance\nsegmentation equipped with a point-clustering strategy. Experimental results on\nboth ScanNet-v2 and S3DIS show that our self-training approach, with\nextremely-sparse annotations, outperforms all existing weakly supervised\nmethods for 3D semantic and instance segmentation by a large margin, and our\nresults are also comparable to those of the fully supervised counterparts.\nCodes and models are available at\nhttps://github.com/liuzhengzhe/One-Thing-One-Click.\n", "date": "2023-03-26", "categories": []}
{"id": "2303.09248", "title": "Cross-Dimensional Refined Learning for Real-Time 3D Visual Perception\n  from Monocular Video", "abstract": "  We present a novel real-time capable learning method that jointly perceives a\n3D scene's geometry structure and semantic labels. Recent approaches to\nreal-time 3D scene reconstruction mostly adopt a volumetric scheme, where a\nTruncated Signed Distance Function (TSDF) is directly regressed. However, these\nvolumetric approaches tend to focus on the global coherence of their\nreconstructions, which leads to a lack of local geometric detail. To overcome\nthis issue, we propose to leverage the latent geometric prior knowledge in 2D\nimage features by explicit depth prediction and anchored feature generation, to\nrefine the occupancy learning in TSDF volume. Besides, we find that this\ncross-dimensional feature refinement methodology can also be adopted for the\nsemantic segmentation task by utilizing semantic priors. Hence, we proposed an\nend-to-end cross-dimensional refinement neural network (CDRNet) to extract both\n3D mesh and 3D semantic labeling in real time. The experiment results show that\nthis method achieves a state-of-the-art 3D perception efficiency on multiple\ndatasets, which indicates the great potential of our method for industrial\napplications.\n", "date": "2023-03-16", "categories": []}
{"id": "2309.05339", "title": "PAg-NeRF: Towards fast and efficient end-to-end panoptic 3D\n  representations for agricultural robotics", "abstract": "  Precise scene understanding is key for most robot monitoring and intervention\ntasks in agriculture. In this work we present PAg-NeRF which is a novel\nNeRF-based system that enables 3D panoptic scene understanding. Our\nrepresentation is trained using an image sequence with noisy robot odometry\nposes and automatic panoptic predictions with inconsistent IDs between frames.\nDespite this noisy input, our system is able to output scene geometry,\nphoto-realistic renders and 3D consistent panoptic representations with\nconsistent instance IDs. We evaluate this novel system in a very challenging\nhorticultural scenario and in doing so demonstrate an end-to-end trainable\nsystem that can make use of noisy robot poses rather than precise poses that\nhave to be pre-calculated. Compared to a baseline approach the peak signal to\nnoise ratio is improved from 21.34dB to 23.37dB while the panoptic quality\nimproves from 56.65% to 70.08%. Furthermore, our approach is faster and can be\ntuned to improve inference time by more than a factor of 2 while being memory\nefficient with approximately 12 times fewer parameters.\n", "date": "2023-09-11", "categories": []}
{"id": "2303.11910", "title": "360BEV: Panoramic Semantic Mapping for Indoor Bird's-Eye View", "abstract": "  Seeing only a tiny part of the whole is not knowing the full circumstance.\nBird's-eye-view (BEV) perception, a process of obtaining allocentric maps from\negocentric views, is restricted when using a narrow Field of View (FoV) alone.\nIn this work, mapping from 360{\\deg} panoramas to BEV semantics, the 360BEV\ntask, is established for the first time to achieve holistic representations of\nindoor scenes in a top-down view. Instead of relying on narrow-FoV image\nsequences, a panoramic image with depth information is sufficient to generate a\nholistic BEV semantic map. To benchmark 360BEV, we present two indoor datasets,\n360BEV-Matterport and 360BEV-Stanford, both of which include egocentric\npanoramic images and semantic segmentation labels, as well as allocentric\nsemantic maps. Besides delving deep into different mapping paradigms, we\npropose a dedicated solution for panoramic semantic mapping, namely 360Mapper.\nThrough extensive experiments, our methods achieve 44.32% and 45.78% in mIoU on\nboth datasets respectively, surpassing previous counterparts with gains of\n+7.60% and +9.70% in mIoU. Code and datasets are available at the project page:\nhttps://jamycheung.github.io/360BEV.html.\n", "date": "2023-03-21", "categories": []}
{"id": "2304.03763", "title": "Clutter Detection and Removal in 3D Scenes with View-Consistent\n  Inpainting", "abstract": "  Removing clutter from scenes is essential in many applications, ranging from\nprivacy-concerned content filtering to data augmentation. In this work, we\npresent an automatic system that removes clutter from 3D scenes and inpaints\nwith coherent geometry and texture. We propose techniques for its two key\ncomponents: 3D segmentation from shared properties and 3D inpainting, both of\nwhich are important problems. The definition of 3D scene clutter\n(frequently-moving objects) is not well captured by commonly-studied object\ncategories in computer vision. To tackle the lack of well-defined clutter\nannotations, we group noisy fine-grained labels, leverage virtual rendering,\nand impose an instance-level area-sensitive loss. Once clutter is removed, we\ninpaint geometry and texture in the resulting holes by merging inpainted RGB-D\nimages. This requires novel voting and pruning strategies that guarantee\nmulti-view consistency across individually inpainted images for mesh\nreconstruction. Experiments on ScanNet and Matterport dataset show that our\nmethod outperforms baselines for clutter segmentation and 3D inpainting, both\nvisually and quantitatively.\n", "date": "2023-04-07", "categories": []}
{"id": "2309.00615", "title": "Point-Bind & Point-LLM: Aligning Point Cloud with Multi-modality for 3D\n  Understanding, Generation, and Instruction Following", "abstract": "  We introduce Point-Bind, a 3D multi-modality model aligning point clouds with\n2D image, language, audio, and video. Guided by ImageBind, we construct a joint\nembedding space between 3D and multi-modalities, enabling many promising\napplications, e.g., any-to-3D generation, 3D embedding arithmetic, and 3D\nopen-world understanding. On top of this, we further present Point-LLM, the\nfirst 3D large language model (LLM) following 3D multi-modal instructions. By\nparameter-efficient fine-tuning techniques, Point-LLM injects the semantics of\nPoint-Bind into pre-trained LLMs, e.g., LLaMA, which requires no 3D instruction\ndata, but exhibits superior 3D and multi-modal question-answering capacity. We\nhope our work may cast a light on the community for extending 3D point clouds\nto multi-modality applications. Code is available at\nhttps://github.com/ZiyuGuo99/Point-Bind_Point-LLM.\n", "date": "2023-09-01", "categories": []}
{"id": "2308.15050", "title": "iBARLE: imBalance-Aware Room Layout Estimation", "abstract": "  Room layout estimation predicts layouts from a single panorama. It requires\ndatasets with large-scale and diverse room shapes to train the models. However,\nthere are significant imbalances in real-world datasets including the\ndimensions of layout complexity, camera locations, and variation in scene\nappearance. These issues considerably influence the model training performance.\nIn this work, we propose the imBalance-Aware Room Layout Estimation (iBARLE)\nframework to address these issues. iBARLE consists of (1) Appearance Variation\nGeneration (AVG) module, which promotes visual appearance domain\ngeneralization, (2) Complex Structure Mix-up (CSMix) module, which enhances\ngeneralizability w.r.t. room structure, and (3) a gradient-based layout\nobjective function, which allows more effective accounting for occlusions in\ncomplex layouts. All modules are jointly trained and help each other to achieve\nthe best performance. Experiments and ablation studies based on\nZInD~\\cite{cruz2021zillow} dataset illustrate that iBARLE has state-of-the-art\nperformance compared with other layout estimation baselines.\n", "date": "2023-08-29", "categories": []}
{"id": "2302.10425", "title": "Instance-incremental Scene Graph Generation from Real-world Point Clouds\n  via Normalizing Flows", "abstract": "  This work introduces a new task of instance-incremental scene graph\ngeneration: Given a scene of the point cloud, representing it as a graph and\nautomatically increasing novel instances. A graph denoting the object layout of\nthe scene is finally generated. It is an important task since it helps to guide\nthe insertion of novel 3D objects into a real-world scene in vision-based\napplications like augmented reality. It is also challenging because the\ncomplexity of the real-world point cloud brings difficulties in learning object\nlayout experiences from the observation data (non-empty rooms with labeled\nsemantics). We model this task as a conditional generation problem and propose\na 3D autoregressive framework based on normalizing flows (3D-ANF) to address\nit. First, we represent the point cloud as a graph by extracting the label\nsemantics and contextual relationships. Next, a model based on normalizing\nflows is introduced to map the conditional generation of graphic elements into\nthe Gaussian process. The mapping is invertible. Thus, the real-world\nexperiences represented in the observation data can be modeled in the training\nphase, and novel instances can be autoregressively generated based on the\nGaussian process in the testing phase. To evaluate the performance of our\nmethod sufficiently, we implement this new task on the indoor benchmark dataset\n3DSSG-O27R16 and our newly proposed graphical dataset of outdoor scenes GPL3D.\nExperiments show that our method generates reliable novel graphs from the\nreal-world point cloud and achieves state-of-the-art performance on the\ndatasets.\n", "date": "2023-02-21", "categories": []}
{"id": "2212.02501", "title": "SceneRF: Self-Supervised Monocular 3D Scene Reconstruction with Radiance\n  Fields", "abstract": "  3D reconstruction from a single 2D image was extensively covered in the\nliterature but relies on depth supervision at training time, which limits its\napplicability. To relax the dependence to depth we propose SceneRF, a\nself-supervised monocular scene reconstruction method using only posed image\nsequences for training. Fueled by the recent progress in neural radiance fields\n(NeRF) we optimize a radiance field though with explicit depth optimization and\na novel probabilistic sampling strategy to efficiently handle large scenes. At\ninference, a single input image suffices to hallucinate novel depth views which\nare fused together to obtain 3D scene reconstruction. Thorough experiments\ndemonstrate that we outperform all baselines for novel depth views synthesis\nand scene reconstruction, on indoor BundleFusion and outdoor SemanticKITTI.\nCode is available at https://astra-vision.github.io/SceneRF .\n", "date": "2022-12-05", "categories": []}
{"id": "2308.11417", "title": "ScanNet++: A High-Fidelity Dataset of 3D Indoor Scenes", "abstract": "  We present ScanNet++, a large-scale dataset that couples together capture of\nhigh-quality and commodity-level geometry and color of indoor scenes. Each\nscene is captured with a high-end laser scanner at sub-millimeter resolution,\nalong with registered 33-megapixel images from a DSLR camera, and RGB-D streams\nfrom an iPhone. Scene reconstructions are further annotated with an open\nvocabulary of semantics, with label-ambiguous scenarios explicitly annotated\nfor comprehensive semantic understanding. ScanNet++ enables a new real-world\nbenchmark for novel view synthesis, both from high-quality RGB capture, and\nimportantly also from commodity-level images, in addition to a new benchmark\nfor 3D semantic scene understanding that comprehensively encapsulates diverse\nand ambiguous semantic labeling scenarios. Currently, ScanNet++ contains 460\nscenes, 280,000 captured DSLR images, and over 3.7M iPhone RGBD frames.\n", "date": "2023-08-22", "categories": []}
{"id": "2304.13980", "title": "A Review of Panoptic Segmentation for Mobile Mapping Point Clouds", "abstract": "  3D point cloud panoptic segmentation is the combined task to (i) assign each\npoint to a semantic class and (ii) separate the points in each class into\nobject instances. Recently there has been an increased interest in such\ncomprehensive 3D scene understanding, building on the rapid advances of\nsemantic segmentation due to the advent of deep 3D neural networks. Yet, to\ndate there is very little work about panoptic segmentation of outdoor\nmobile-mapping data, and no systematic comparisons. The present paper tries to\nclose that gap. It reviews the building blocks needed to assemble a panoptic\nsegmentation pipeline and the related literature. Moreover, a modular pipeline\nis set up to perform comprehensive, systematic experiments to assess the state\nof panoptic segmentation in the context of street mapping. As a byproduct, we\nalso provide the first public dataset for that task, by extending the NPM3D\ndataset to include instance labels. That dataset and our source code are\npublicly available. We discuss which adaptations are need to adapt current\npanoptic segmentation methods to outdoor scenes and large objects. Our study\nfinds that for mobile mapping data, KPConv performs best but is slower, while\nPointNet++ is fastest but performs significantly worse. Sparse CNNs are in\nbetween. Regardless of the backbone, Instance segmentation by clustering\nembedding features is better than using shifted coordinates.\n", "date": "2023-04-27", "categories": []}
{"id": "2308.08769", "title": "Chat-3D: Data-efficiently Tuning Large Language Model for Universal\n  Dialogue of 3D Scenes", "abstract": "  3D scene understanding has gained significant attention due to its wide range\nof applications. However, existing methods for 3D scene understanding are\nlimited to specific downstream tasks, which hinders their practicality in\nreal-world applications. This paper presents Chat-3D, which combines the 3D\nvisual perceptual ability of pre-trained 3D representations and the impressive\nreasoning and conversation capabilities of advanced LLMs to achieve the first\nuniversal dialogue systems for 3D scenes. Specifically, we align 3D\nrepresentations into the feature space of LLMs, thus enabling LLMs to perceive\nthe 3D world. Given the scarcity of 3D scene-text data, we propose a\nthree-stage training strategy to efficiently utilize the available data for\nbetter alignment. To enhance the reasoning ability and develop a user-friendly\ninteraction scheme, we further construct a high-quality object-centric 3D\ninstruction dataset and design an associated object-centric prompt. Our\nexperiments show that Chat-3D achieves an impressive ability to comprehend\ndiverse instructions for 3D scenes, engage in intricate spatial reasoning, and\nincorporate external knowledge into its responses. Chat-3D achieves a 75.6%\nrelative score compared with GPT-4 on the constructed instruction dataset.\n", "date": "2023-08-17", "categories": []}
{"id": "2304.06906", "title": "Swin3D: A Pretrained Transformer Backbone for 3D Indoor Scene\n  Understanding", "abstract": "  The use of pretrained backbones with fine-tuning has been successful for 2D\nvision and natural language processing tasks, showing advantages over\ntask-specific networks. In this work, we introduce a pretrained 3D backbone,\ncalled {\\SST}, for 3D indoor scene understanding. We design a 3D Swin\ntransformer as our backbone network, which enables efficient self-attention on\nsparse voxels with linear memory complexity, making the backbone scalable to\nlarge models and datasets. We also introduce a generalized contextual relative\npositional embedding scheme to capture various irregularities of point signals\nfor improved network performance. We pretrained a large {\\SST} model on a\nsynthetic Structured3D dataset, which is an order of magnitude larger than the\nScanNet dataset. Our model pretrained on the synthetic dataset not only\ngeneralizes well to downstream segmentation and detection on real 3D point\ndatasets, but also outperforms state-of-the-art methods on downstream tasks\nwith +2.3 mIoU and +2.2 mIoU on S3DIS Area5 and 6-fold semantic segmentation,\n+1.8 mIoU on ScanNet segmentation (val), +1.9 mAP@0.5 on ScanNet detection, and\n+8.1 mAP@0.5 on S3DIS detection. A series of extensive ablation studies further\nvalidate the scalability, generality, and superior performance enabled by our\napproach. The code and models are available at\nhttps://github.com/microsoft/Swin3D .\n", "date": "2023-04-14", "categories": []}
{"id": "2307.04246", "title": "Convex Decomposition of Indoor Scenes", "abstract": "  We describe a method to parse a complex, cluttered indoor scene into\nprimitives which offer a parsimonious abstraction of scene structure. Our\nprimitives are simple convexes. Our method uses a learned regression procedure\nto parse a scene into a fixed number of convexes from RGBD input, and can\noptionally accept segmentations to improve the decomposition. The result is\nthen polished with a descent method which adjusts the convexes to produce a\nvery good fit, and greedily removes superfluous primitives. Because the entire\nscene is parsed, we can evaluate using traditional depth, normal, and\nsegmentation error metrics. Our evaluation procedure demonstrates that the\nerror from our primitive representation is comparable to that of predicting\ndepth from a single image.\n", "date": "2023-07-09", "categories": []}
{"id": "2308.06719", "title": "3D Scene Graph Prediction on Point Clouds Using Knowledge Graphs", "abstract": "  3D scene graph prediction is a task that aims to concurrently predict object\nclasses and their relationships within a 3D environment. As these environments\nare primarily designed by and for humans, incorporating commonsense knowledge\nregarding objects and their relationships can significantly constrain and\nenhance the prediction of the scene graph. In this paper, we investigate the\napplication of commonsense knowledge graphs for 3D scene graph prediction on\npoint clouds of indoor scenes. Through experiments conducted on a real-world\nindoor dataset, we demonstrate that integrating external commonsense knowledge\nvia the message-passing method leads to a 15.0 % improvement in scene graph\nprediction accuracy with external knowledge and $7.96\\%$ with internal\nknowledge when compared to state-of-the-art algorithms. We also tested in the\nreal world with 10 frames per second for scene graph generation to show the\nusage of the model in a more realistic robotics setting.\n", "date": "2023-08-13", "categories": []}
{"id": "2308.04352", "title": "3D-VisTA: Pre-trained Transformer for 3D Vision and Text Alignment", "abstract": "  3D vision-language grounding (3D-VL) is an emerging field that aims to\nconnect the 3D physical world with natural language, which is crucial for\nachieving embodied intelligence. Current 3D-VL models rely heavily on\nsophisticated modules, auxiliary losses, and optimization tricks, which calls\nfor a simple and unified model. In this paper, we propose 3D-VisTA, a\npre-trained Transformer for 3D Vision and Text Alignment that can be easily\nadapted to various downstream tasks. 3D-VisTA simply utilizes self-attention\nlayers for both single-modal modeling and multi-modal fusion without any\nsophisticated task-specific design. To further enhance its performance on 3D-VL\ntasks, we construct ScanScribe, the first large-scale 3D scene-text pairs\ndataset for 3D-VL pre-training. ScanScribe contains 2,995 RGB-D scans for 1,185\nunique indoor scenes originating from ScanNet and 3R-Scan datasets, along with\npaired 278K scene descriptions generated from existing 3D-VL tasks, templates,\nand GPT-3. 3D-VisTA is pre-trained on ScanScribe via masked language/object\nmodeling and scene-text matching. It achieves state-of-the-art results on\nvarious 3D-VL tasks, ranging from visual grounding and dense captioning to\nquestion answering and situated reasoning. Moreover, 3D-VisTA demonstrates\nsuperior data efficiency, obtaining strong performance even with limited\nannotations during downstream task fine-tuning.\n", "date": "2023-08-08", "categories": []}
{"id": "2308.02883", "title": "Cross-modal & Cross-domain Learning for Unsupervised LiDAR Semantic\n  Segmentation", "abstract": "  In recent years, cross-modal domain adaptation has been studied on the paired\n2D image and 3D LiDAR data to ease the labeling costs for 3D LiDAR semantic\nsegmentation (3DLSS) in the target domain. However, in such a setting the\npaired 2D and 3D data in the source domain are still collected with additional\neffort. Since the 2D-3D projections can enable the 3D model to learn semantic\ninformation from the 2D counterpart, we ask whether we could further remove the\nneed of source 3D data and only rely on the source 2D images. To answer it,\nthis paper studies a new 3DLSS setting where a 2D dataset (source) with\nsemantic annotations and a paired but unannotated 2D image and 3D LiDAR data\n(target) are available. To achieve 3DLSS in this scenario, we propose\nCross-Modal and Cross-Domain Learning (CoMoDaL). Specifically, our CoMoDaL aims\nat modeling 1) inter-modal cross-domain distillation between the unpaired\nsource 2D image and target 3D LiDAR data, and 2) the intra-domain cross-modal\nguidance between the target 2D image and 3D LiDAR data pair. In CoMoDaL, we\npropose to apply several constraints, such as point-to-pixel and\nprototype-to-pixel alignments, to associate the semantics in different\nmodalities and domains by constructing mixed samples in two modalities. The\nexperimental results on several datasets show that in the proposed setting, the\ndeveloped CoMoDaL can achieve segmentation without the supervision of labeled\nLiDAR data. Ablations are also conducted to provide more analysis. Code will be\navailable publicly.\n", "date": "2023-08-05", "categories": []}
{"id": "2308.00353", "title": "Lowis3D: Language-Driven Open-World Instance-Level 3D Scene\n  Understanding", "abstract": "  Open-world instance-level scene understanding aims to locate and recognize\nunseen object categories that are not present in the annotated dataset. This\ntask is challenging because the model needs to both localize novel 3D objects\nand infer their semantic categories. A key factor for the recent progress in 2D\nopen-world perception is the availability of large-scale image-text pairs from\nthe Internet, which cover a wide range of vocabulary concepts. However, this\nsuccess is hard to replicate in 3D scenarios due to the scarcity of 3D-text\npairs. To address this challenge, we propose to harness pre-trained\nvision-language (VL) foundation models that encode extensive knowledge from\nimage-text pairs to generate captions for multi-view images of 3D scenes. This\nallows us to establish explicit associations between 3D shapes and\nsemantic-rich captions. Moreover, to enhance the fine-grained visual-semantic\nrepresentation learning from captions for object-level categorization, we\ndesign hierarchical point-caption association methods to learn semantic-aware\nembeddings that exploit the 3D geometry between 3D points and multi-view\nimages. In addition, to tackle the localization challenge for novel classes in\nthe open-world setting, we develop debiased instance localization, which\ninvolves training object grouping modules on unlabeled data using\ninstance-level pseudo supervision. This significantly improves the\ngeneralization capabilities of instance grouping and thus the ability to\naccurately locate novel objects. We conduct extensive experiments on 3D\nsemantic, instance, and panoptic segmentation tasks, covering indoor and\noutdoor scenes across three datasets. Our method outperforms baseline methods\nby a significant margin in semantic segmentation (e.g. 34.5%$\\sim$65.3%),\ninstance segmentation (e.g. 21.8%$\\sim$54.0%) and panoptic segmentation (e.g.\n14.7%$\\sim$43.3%). Code will be available.\n", "date": "2023-08-01", "categories": []}
{"id": "2307.14392", "title": "Human-centric Scene Understanding for 3D Large-scale Scenarios", "abstract": "  Human-centric scene understanding is significant for real-world applications,\nbut it is extremely challenging due to the existence of diverse human poses and\nactions, complex human-environment interactions, severe occlusions in crowds,\netc. In this paper, we present a large-scale multi-modal dataset for\nhuman-centric scene understanding, dubbed HuCenLife, which is collected in\ndiverse daily-life scenarios with rich and fine-grained annotations. Our\nHuCenLife can benefit many 3D perception tasks, such as segmentation,\ndetection, action recognition, etc., and we also provide benchmarks for these\ntasks to facilitate related research. In addition, we design novel modules for\nLiDAR-based segmentation and action recognition, which are more applicable for\nlarge-scale human-centric scenarios and achieve state-of-the-art performance.\n", "date": "2023-07-26", "categories": []}
{"id": "2307.14570", "title": "Physically Plausible 3D Human-Scene Reconstruction from Monocular RGB\n  Image using an Adversarial Learning Approach", "abstract": "  Holistic 3D human-scene reconstruction is a crucial and emerging research\narea in robot perception. A key challenge in holistic 3D human-scene\nreconstruction is to generate a physically plausible 3D scene from a single\nmonocular RGB image. The existing research mainly proposes optimization-based\napproaches for reconstructing the scene from a sequence of RGB frames with\nexplicitly defined physical laws and constraints between different scene\nelements (humans and objects). However, it is hard to explicitly define and\nmodel every physical law in every scenario. This paper proposes using an\nimplicit feature representation of the scene elements to distinguish a\nphysically plausible alignment of humans and objects from an implausible one.\nWe propose using a graph-based holistic representation with an encoded physical\nrepresentation of the scene to analyze the human-object and object-object\ninteractions within the scene. Using this graphical representation, we\nadversarially train our model to learn the feasible alignments of the scene\nelements from the training data itself without explicitly defining the laws and\nconstraints between them. Unlike the existing inference-time optimization-based\napproaches, we use this adversarially trained model to produce a per-frame 3D\nreconstruction of the scene that abides by the physical laws and constraints.\nOur learning-based method achieves comparable 3D reconstruction quality to\nexisting optimization-based holistic human-scene reconstruction methods and\ndoes not need inference time optimization. This makes it better suited when\ncompared to existing methods, for potential use in robotic applications, such\nas robot navigation, etc.\n", "date": "2023-07-27", "categories": []}
{"id": "2307.07938", "title": "CVSformer: Cross-View Synthesis Transformer for Semantic Scene\n  Completion", "abstract": "  Semantic scene completion (SSC) requires an accurate understanding of the\ngeometric and semantic relationships between the objects in the 3D scene for\nreasoning the occluded objects. The popular SSC methods voxelize the 3D\nobjects, allowing the deep 3D convolutional network (3D CNN) to learn the\nobject relationships from the complex scenes. However, the current networks\nlack the controllable kernels to model the object relationship across multiple\nviews, where appropriate views provide the relevant information for suggesting\nthe existence of the occluded objects. In this paper, we propose Cross-View\nSynthesis Transformer (CVSformer), which consists of Multi-View Feature\nSynthesis and Cross-View Transformer for learning cross-view object\nrelationships. In the multi-view feature synthesis, we use a set of 3D\nconvolutional kernels rotated differently to compute the multi-view features\nfor each voxel. In the cross-view transformer, we employ the cross-view fusion\nto comprehensively learn the cross-view relationships, which form useful\ninformation for enhancing the features of individual views. We use the enhanced\nfeatures to predict the geometric occupancies and semantic labels of all\nvoxels. We evaluate CVSformer on public datasets, where CVSformer yields\nstate-of-the-art results.\n", "date": "2023-07-16", "categories": []}
{"id": "2209.07896", "title": "3D VSG: Long-term Semantic Scene Change Prediction through 3D Variable\n  Scene Graphs", "abstract": "  Numerous applications require robots to operate in environments shared with\nother agents, such as humans or other robots. However, such shared scenes are\ntypically subject to different kinds of long-term semantic scene changes. The\nability to model and predict such changes is thus crucial for robot autonomy.\nIn this work, we formalize the task of semantic scene variability estimation\nand identify three main varieties of semantic scene change: changes in the\nposition of an object, its semantic state, or the composition of a scene as a\nwhole. To represent this variability, we propose the Variable Scene Graph\n(VSG), which augments existing 3D Scene Graph (SG) representations with the\nvariability attribute, representing the likelihood of discrete long-term change\nevents. We present a novel method, DeltaVSG, to estimate the variability of\nVSGs in a supervised fashion. We evaluate our method on the 3RScan long-term\ndataset, showing notable improvements in this novel task over existing\napproaches. Our method DeltaVSG achieves an accuracy of 77.1% and a recall of\n72.3%, often mimicking human intuition about how indoor scenes change over\ntime. We further show the utility of VSG prediction in the task of active\nrobotic change detection, speeding up task completion by 66.0% compared to a\nscene-change-unaware planner. We make our code available as open-source.\n", "date": "2022-09-16", "categories": []}
{"id": "2307.05873", "title": "OG: Equip vision occupancy with instance segmentation and visual\n  grounding", "abstract": "  Occupancy prediction tasks focus on the inference of both geometry and\nsemantic labels for each voxel, which is an important perception mission.\nHowever, it is still a semantic segmentation task without distinguishing\nvarious instances. Further, although some existing works, such as\nOpen-Vocabulary Occupancy (OVO), have already solved the problem of open\nvocabulary detection, visual grounding in occupancy has not been solved to the\nbest of our knowledge. To tackle the above two limitations, this paper proposes\nOccupancy Grounding (OG), a novel method that equips vanilla occupancy instance\nsegmentation ability and could operate visual grounding in a voxel manner with\nthe help of grounded-SAM. Keys to our approach are (1) affinity field\nprediction for instance clustering and (2) association strategy for aligning 2D\ninstance masks and 3D occupancy instances. Extensive experiments have been\nconducted whose visualization results and analysis are shown below. Our code\nwill be publicly released soon.\n", "date": "2023-07-12", "categories": []}
{"id": "2307.01492", "title": "FB-OCC: 3D Occupancy Prediction based on Forward-Backward View\n  Transformation", "abstract": "  This technical report summarizes the winning solution for the 3D Occupancy\nPrediction Challenge, which is held in conjunction with the CVPR 2023 Workshop\non End-to-End Autonomous Driving and CVPR 23 Workshop on Vision-Centric\nAutonomous Driving Workshop. Our proposed solution FB-OCC builds upon FB-BEV, a\ncutting-edge camera-based bird's-eye view perception design using\nforward-backward projection. On top of FB-BEV, we further study novel designs\nand optimization tailored to the 3D occupancy prediction task, including joint\ndepth-semantic pre-training, joint voxel-BEV representation, model scaling up,\nand effective post-processing strategies. These designs and optimization result\nin a state-of-the-art mIoU score of 54.19% on the nuScenes dataset, ranking the\n1st place in the challenge track. Code and models will be released at:\nhttps://github.com/NVlabs/FB-BEV.\n", "date": "2023-07-04", "categories": []}
{"id": "2306.15349", "title": "SSC-RS: Elevate LiDAR Semantic Scene Completion with Representation\n  Separation and BEV Fusion", "abstract": "  Semantic scene completion (SSC) jointly predicts the semantics and geometry\nof the entire 3D scene, which plays an essential role in 3D scene understanding\nfor autonomous driving systems. SSC has achieved rapid progress with the help\nof semantic context in segmentation. However, how to effectively exploit the\nrelationships between the semantic context in semantic segmentation and\ngeometric structure in scene completion remains under exploration. In this\npaper, we propose to solve outdoor SSC from the perspective of representation\nseparation and BEV fusion. Specifically, we present the network, named SSC-RS,\nwhich uses separate branches with deep supervision to explicitly disentangle\nthe learning procedure of the semantic and geometric representations. And a BEV\nfusion network equipped with the proposed Adaptive Representation Fusion (ARF)\nmodule is presented to aggregate the multi-scale features effectively and\nefficiently. Due to the low computational burden and powerful representation\nability, our model has good generality while running in real-time. Extensive\nexperiments on SemanticKITTI demonstrate our SSC-RS achieves state-of-the-art\nperformance.\n", "date": "2023-06-27", "categories": []}
{"id": "2306.02851", "title": "Scene as Occupancy", "abstract": "  Human driver can easily describe the complex traffic scene by visual system.\nSuch an ability of precise perception is essential for driver's planning. To\nachieve this, a geometry-aware representation that quantizes the physical 3D\nscene into structured grid map with semantic labels per cell, termed as 3D\nOccupancy, would be desirable. Compared to the form of bounding box, a key\ninsight behind occupancy is that it could capture the fine-grained details of\ncritical obstacles in the scene, and thereby facilitate subsequent tasks. Prior\nor concurrent literature mainly concentrate on a single scene completion task,\nwhere we might argue that the potential of this occupancy representation might\nobsess broader impact. In this paper, we propose OccNet, a multi-view\nvision-centric pipeline with a cascade and temporal voxel decoder to\nreconstruct 3D occupancy. At the core of OccNet is a general occupancy\nembedding to represent 3D physical world. Such a descriptor could be applied\ntowards a wide span of driving tasks, including detection, segmentation and\nplanning. To validate the effectiveness of this new representation and our\nproposed algorithm, we propose OpenOcc, the first dense high-quality 3D\noccupancy benchmark built on top of nuScenes. Empirical experiments show that\nthere are evident performance gain across multiple tasks, e.g., motion planning\ncould witness a collision rate reduction by 15%-58%, demonstrating the\nsuperiority of our method.\n", "date": "2023-06-05", "categories": []}
{"id": "2304.10950", "title": "Factored Neural Representation for Scene Understanding", "abstract": "  A long-standing goal in scene understanding is to obtain interpretable and\neditable representations that can be directly constructed from a raw monocular\nRGB-D video, without requiring specialized hardware setup or priors. The\nproblem is significantly more challenging in the presence of multiple moving\nand/or deforming objects. Traditional methods have approached the setup with a\nmix of simplifications, scene priors, pretrained templates, or known\ndeformation models. The advent of neural representations, especially neural\nimplicit representations and radiance fields, opens the possibility of\nend-to-end optimization to collectively capture geometry, appearance, and\nobject motion. However, current approaches produce global scene encoding,\nassume multiview capture with limited or no motion in the scenes, and do not\nfacilitate easy manipulation beyond novel view synthesis. In this work, we\nintroduce a factored neural scene representation that can directly be learned\nfrom a monocular RGB-D video to produce object-level neural presentations with\nan explicit encoding of object movement (e.g., rigid trajectory) and/or\ndeformations (e.g., nonrigid movement). We evaluate ours against a set of\nneural approaches on both synthetic and real data to demonstrate that the\nrepresentation is efficient, interpretable, and editable (e.g., change object\ntrajectory). Code and data are available at\nhttp://geometry.cs.ucl.ac.uk/projects/2023/factorednerf .\n", "date": "2023-04-21", "categories": []}
{"id": "2306.11414", "title": "Multi-Scale Occ: 4th Place Solution for CVPR 2023 3D Occupancy\n  Prediction Challenge", "abstract": "  In this report, we present the 4th place solution for CVPR 2023 3D occupancy\nprediction challenge. We propose a simple method called Multi-Scale Occ for\noccupancy prediction based on lift-splat-shoot framework, which introduces\nmulti-scale image features for generating better multi-scale 3D voxel features\nwith temporal fusion of multiple past frames. Post-processing including model\nensemble, test-time augmentation, and class-wise thresh are adopted to further\nboost the final performance. As shown on the leaderboard, our proposed\noccupancy prediction method ranks the 4th place with 49.36 mIoU.\n", "date": "2023-06-20", "categories": []}
{"id": "2305.16133", "title": "OVO: Open-Vocabulary Occupancy", "abstract": "  Semantic occupancy prediction aims to infer dense geometry and semantics of\nsurroundings for an autonomous agent to operate safely in the 3D environment.\nExisting occupancy prediction methods are almost entirely trained on\nhuman-annotated volumetric data. Although of high quality, the generation of\nsuch 3D annotations is laborious and costly, restricting them to a few specific\nobject categories in the training dataset. To address this limitation, this\npaper proposes Open Vocabulary Occupancy (OVO), a novel approach that allows\nsemantic occupancy prediction of arbitrary classes but without the need for 3D\nannotations during training. Keys to our approach are (1) knowledge\ndistillation from a pre-trained 2D open-vocabulary segmentation model to the 3D\noccupancy network, and (2) pixel-voxel filtering for high-quality training data\ngeneration. The resulting framework is simple, compact, and compatible with\nmost state-of-the-art semantic occupancy prediction models. On NYUv2 and\nSemanticKITTI datasets, OVO achieves competitive performance compared to\nsupervised semantic occupancy prediction approaches. Furthermore, we conduct\nextensive analyses and ablation studies to offer insights into the design of\nthe proposed framework. Our code is publicly available at\nhttps://github.com/dzcgaara/OVO.\n", "date": "2023-05-25", "categories": []}
{"id": "2306.09117", "title": "UniOcc: Unifying Vision-Centric 3D Occupancy Prediction with Geometric\n  and Semantic Rendering", "abstract": "  In this technical report, we present our solution, named UniOCC, for the\nVision-Centric 3D occupancy prediction track in the nuScenes Open Dataset\nChallenge at CVPR 2023. Existing methods for occupancy prediction primarily\nfocus on optimizing projected features on 3D volume space using 3D occupancy\nlabels. However, the generation process of these labels is complex and\nexpensive (relying on 3D semantic annotations), and limited by voxel\nresolution, they cannot provide fine-grained spatial semantics. To address this\nlimitation, we propose a novel Unifying Occupancy (UniOcc) prediction method,\nexplicitly imposing spatial geometry constraint and complementing fine-grained\nsemantic supervision through volume ray rendering. Our method significantly\nenhances model performance and demonstrates promising potential in reducing\nhuman annotation costs. Given the laborious nature of annotating 3D occupancy,\nwe further introduce a Depth-aware Teacher Student (DTS) framework to enhance\nprediction accuracy using unlabeled data. Our solution achieves 51.27\\% mIoU on\nthe official leaderboard with single model, placing 3rd in this challenge.\n", "date": "2023-06-15", "categories": []}
{"id": "2306.07809", "title": "Low-Resource White-Box Semantic Segmentation of Supporting Towers on 3D\n  Point Clouds via Signature Shape Identification", "abstract": "  Research in 3D semantic segmentation has been increasing performance metrics,\nlike the IoU, by scaling model complexity and computational resources, leaving\nbehind researchers and practitioners that (1) cannot access the necessary\nresources and (2) do need transparency on the model decision mechanisms. In\nthis paper, we propose SCENE-Net, a low-resource white-box model for 3D point\ncloud semantic segmentation. SCENE-Net identifies signature shapes on the point\ncloud via group equivariant non-expansive operators (GENEOs), providing\nintrinsic geometric interpretability. Our training time on a laptop is 85~min,\nand our inference time is 20~ms. SCENE-Net has 11 trainable geometrical\nparameters and requires fewer data than black-box models. SCENE--Net offers\nrobustness to noisy labeling and data imbalance and has comparable IoU to\nstate-of-the-art methods. With this paper, we release a 40~000 Km labeled\ndataset of rural terrain point clouds and our code implementation.\n", "date": "2023-06-13", "categories": []}
{"id": "2206.03591", "title": "ObPose: Leveraging Pose for Object-Centric Scene Inference and\n  Generation in 3D", "abstract": "  We present ObPose, an unsupervised object-centric inference and generation\nmodel which learns 3D-structured latent representations from RGB-D scenes.\nInspired by prior art in 2D representation learning, ObPose considers a\nfactorised latent space, separately encoding object location (where) and\nappearance (what). ObPose further leverages an object's pose (i.e. location and\norientation), defined via a minimum volume principle, as a novel inductive bias\nfor learning the where component. To achieve this, we propose an efficient,\nvoxelised approximation approach to recover the object shape directly from a\nneural radiance field (NeRF). As a consequence, ObPose models each scene as a\ncomposition of NeRFs, richly representing individual objects. To evaluate the\nquality of the learned representations, ObPose is evaluated quantitatively on\nthe YCB, MultiShapeNet, and CLEVR datatasets for unsupervised scene\nsegmentation, outperforming the current state-of-the-art in 3D scene inference\n(ObSuRF) by a significant margin. Generative results provide qualitative\ndemonstration that the same ObPose model can both generate novel scenes and\nflexibly edit the objects in them. These capacities again reflect the quality\nof the learned latents and the benefits of disentangling the where and what\ncomponents of a scene. Key design choices made in the ObPose encoder are\nvalidated with ablations.\n", "date": "2022-06-07", "categories": []}
{"id": "2212.09948", "title": "MM-3DScene: 3D Scene Understanding by Customizing Masked Modeling with\n  Informative-Preserved Reconstruction and Self-Distilled Consistency", "abstract": "  Masked Modeling (MM) has demonstrated widespread success in various vision\nchallenges, by reconstructing masked visual patches. Yet, applying MM for\nlarge-scale 3D scenes remains an open problem due to the data sparsity and\nscene complexity. The conventional random masking paradigm used in 2D images\noften causes a high risk of ambiguity when recovering the masked region of 3D\nscenes. To this end, we propose a novel informative-preserved reconstruction,\nwhich explores local statistics to discover and preserve the representative\nstructured points, effectively enhancing the pretext masking task for 3D scene\nunderstanding. Integrated with a progressive reconstruction manner, our method\ncan concentrate on modeling regional geometry and enjoy less ambiguity for\nmasked reconstruction. Besides, such scenes with progressive masking ratios can\nalso serve to self-distill their intrinsic spatial consistency, requiring to\nlearn the consistent representations from unmasked areas. By elegantly\ncombining informative-preserved reconstruction on masked areas and consistency\nself-distillation from unmasked areas, a unified framework called MM-3DScene is\nyielded. We conduct comprehensive experiments on a host of downstream tasks.\nThe consistent improvement (e.g., +6.1 mAP@0.5 on object detection and +2.2%\nmIoU on semantic segmentation) demonstrates the superiority of our approach.\n", "date": "2022-12-20", "categories": []}
{"id": "2306.02329", "title": "Multi-CLIP: Contrastive Vision-Language Pre-training for Question\n  Answering tasks in 3D Scenes", "abstract": "  Training models to apply common-sense linguistic knowledge and visual\nconcepts from 2D images to 3D scene understanding is a promising direction that\nresearchers have only recently started to explore. However, it still remains\nunderstudied whether 2D distilled knowledge can provide useful representations\nfor downstream 3D vision-language tasks such as 3D question answering. In this\npaper, we propose a novel 3D pre-training Vision-Language method, namely\nMulti-CLIP, that enables a model to learn language-grounded and transferable 3D\nscene point cloud representations. We leverage the representational power of\nthe CLIP model by maximizing the agreement between the encoded 3D scene\nfeatures and the corresponding 2D multi-view image and text embeddings in the\nCLIP space via a contrastive objective. To validate our approach, we consider\nthe challenging downstream tasks of 3D Visual Question Answering (3D-VQA) and\n3D Situated Question Answering (3D-SQA). To this end, we develop novel\nmulti-modal transformer-based architectures and we demonstrate how our\npre-training method can benefit their performance. Quantitative and qualitative\nexperimental results show that Multi-CLIP outperforms state-of-the-art works\nacross the downstream tasks of 3D-VQA and 3D-SQA and leads to a well-structured\n3D scene feature space.\n", "date": "2023-06-04", "categories": []}
{"id": "2306.04184", "title": "StructuredMesh: 3D Structured Optimization of Fa\\c{c}ade Components on\n  Photogrammetric Mesh Models using Binary Integer Programming", "abstract": "  The lack of fa\\c{c}ade structures in photogrammetric mesh models renders them\ninadequate for meeting the demands of intricate applications. Moreover, these\nmesh models exhibit irregular surfaces with considerable geometric noise and\ntexture quality imperfections, making the restoration of structures\nchallenging. To address these shortcomings, we present StructuredMesh, a novel\napproach for reconstructing fa\\c{c}ade structures conforming to the regularity\nof buildings within photogrammetric mesh models. Our method involves capturing\nmulti-view color and depth images of the building model using a virtual camera\nand employing a deep learning object detection pipeline to semi-automatically\nextract the bounding boxes of fa\\c{c}ade components such as windows, doors, and\nbalconies from the color image. We then utilize the depth image to remap these\nboxes into 3D space, generating an initial fa\\c{c}ade layout. Leveraging\narchitectural knowledge, we apply binary integer programming (BIP) to optimize\nthe 3D layout's structure, encompassing the positions, orientations, and sizes\nof all components. The refined layout subsequently informs fa\\c{c}ade modeling\nthrough instance replacement. We conducted experiments utilizing building mesh\nmodels from three distinct datasets, demonstrating the adaptability,\nrobustness, and noise resistance of our proposed methodology. Furthermore, our\n3D layout evaluation metrics reveal that the optimized layout enhances\nprecision, recall, and F-score by 6.5%, 4.5%, and 5.5%, respectively, in\ncomparison to the initial layout.\n", "date": "2023-06-07", "categories": []}
{"id": "2306.03908", "title": "SAM3D: Segment Anything in 3D Scenes", "abstract": "  In this work, we propose SAM3D, a novel framework that is able to predict\nmasks in 3D point clouds by leveraging the Segment-Anything Model (SAM) in RGB\nimages without further training or finetuning. For a point cloud of a 3D scene\nwith posed RGB images, we first predict segmentation masks of RGB images with\nSAM, and then project the 2D masks into the 3D points. Later, we merge the 3D\nmasks iteratively with a bottom-up merging approach. At each step, we merge the\npoint cloud masks of two adjacent frames with the bidirectional merging\napproach. In this way, the 3D masks predicted from different frames are\ngradually merged into the 3D masks of the whole 3D scene. Finally, we can\noptionally ensemble the result from our SAM3D with the over-segmentation\nresults based on the geometric information of the 3D scenes. Our approach is\nexperimented with ScanNet dataset and qualitative results demonstrate that our\nSAM3D achieves reasonable and fine-grained 3D segmentation results without any\ntraining or finetuning of SAM.\n", "date": "2023-06-06", "categories": []}
{"id": "2305.12497", "title": "PanoContext-Former: Panoramic Total Scene Understanding with a\n  Transformer", "abstract": "  Panoramic image enables deeper understanding and more holistic perception of\n$360^\\circ$ surrounding environment, which can naturally encode enriched scene\ncontext information compared to standard perspective image. Previous work has\nmade lots of effort to solve the scene understanding task in a bottom-up form,\nthus each sub-task is processed separately and few correlations are explored in\nthis procedure. In this paper, we propose a novel method using depth prior for\nholistic indoor scene understanding which recovers the objects' shapes,\noriented bounding boxes and the 3D room layout simultaneously from a single\npanorama. In order to fully utilize the rich context information, we design a\ntransformer-based context module to predict the representation and relationship\namong each component of the scene. In addition, we introduce a real-world\ndataset for scene understanding, including photo-realistic panoramas,\nhigh-fidelity depth images, accurately annotated room layouts, and oriented\nobject bounding boxes and shapes. Experiments on the synthetic and real-world\ndatasets demonstrate that our method outperforms previous panoramic scene\nunderstanding methods in terms of both layout estimation and 3D object\ndetection.\n", "date": "2023-05-21", "categories": []}
{"id": "2209.10663", "title": "Convolutional Bayesian Kernel Inference for 3D Semantic Mapping", "abstract": "  Robotic perception is currently at a cross-roads between modern methods,\nwhich operate in an efficient latent space, and classical methods, which are\nmathematically founded and provide interpretable, trustworthy results. In this\npaper, we introduce a Convolutional Bayesian Kernel Inference (ConvBKI) layer\nwhich learns to perform explicit Bayesian inference within a depthwise\nseparable convolution layer to maximize efficency while maintaining reliability\nsimultaneously. We apply our layer to the task of real-time 3D semantic\nmapping, where we learn semantic-geometric probability distributions for LiDAR\nsensor information and incorporate semantic predictions into a global map. We\nevaluate our network against state-of-the-art semantic mapping algorithms on\nthe KITTI data set, demonstrating improved latency with comparable semantic\nlabel inference results.\n", "date": "2022-09-21", "categories": []}
{"id": "2304.14382", "title": "Analogy-Forming Transformers for Few-Shot 3D Parsing", "abstract": "  We present Analogical Networks, a model that encodes domain knowledge\nexplicitly, in a collection of structured labelled 3D scenes, in addition to\nimplicitly, as model parameters, and segments 3D object scenes with analogical\nreasoning: instead of mapping a scene to part segments directly, our model\nfirst retrieves related scenes from memory and their corresponding part\nstructures, and then predicts analogous part structures for the input scene,\nvia an end-to-end learnable modulation mechanism. By conditioning on more than\none retrieved memories, compositions of structures are predicted, that mix and\nmatch parts across the retrieved memories. One-shot, few-shot or many-shot\nlearning are treated uniformly in Analogical Networks, by conditioning on the\nappropriate set of memories, whether taken from a single, few or many memory\nexemplars, and inferring analogous parses. We show Analogical Networks are\ncompetitive with state-of-the-art 3D segmentation transformers in many-shot\nsettings, and outperform them, as well as existing paradigms of meta-learning\nand few-shot learning, in few-shot settings. Analogical Networks successfully\nsegment instances of novel object categories simply by expanding their memory,\nwithout any weight updates. Our code and models are publicly available in the\nproject webpage: http://analogicalnets.github.io/.\n", "date": "2023-04-27", "categories": []}
{"id": "2305.12427", "title": "VL-Fields: Towards Language-Grounded Neural Implicit Spatial\n  Representations", "abstract": "  We present Visual-Language Fields (VL-Fields), a neural implicit spatial\nrepresentation that enables open-vocabulary semantic queries. Our model encodes\nand fuses the geometry of a scene with vision-language trained latent features\nby distilling information from a language-driven segmentation model. VL-Fields\nis trained without requiring any prior knowledge of the scene object classes,\nwhich makes it a promising representation for the field of robotics. Our model\noutperformed the similar CLIP-Fields model in the task of semantic segmentation\nby almost 10%.\n", "date": "2023-05-21", "categories": []}
{"id": "2305.10714", "title": "Vision-Language Pre-training with Object Contrastive Learning for 3D\n  Scene Understanding", "abstract": "  In recent years, vision language pre-training frameworks have made\nsignificant progress in natural language processing and computer vision,\nachieving remarkable performance improvement on various downstream tasks.\nHowever, when extended to point cloud data, existing works mainly focus on\nbuilding task-specific models, and fail to extract universal 3D vision-language\nembedding that generalize well. We carefully investigate three common tasks in\nsemantic 3D scene understanding, and derive key insights into the development\nof a pre-training model. Motivated by these observations, we propose a\nvision-language pre-training framework 3DVLP (3D vision-language pre-training\nwith object contrastive learning), which transfers flexibly on 3D\nvision-language downstream tasks. 3DVLP takes visual grounding as the proxy\ntask and introduces Object-level IoU-guided Detection (OID) loss to obtain\nhigh-quality proposals in the scene. Moreover, we design Object-level\nCross-Contrastive alignment (OCC) task and Object-level Self-Contrastive\nlearning (OSC) task to align the objects with descriptions and distinguish\ndifferent objects in the scene, respectively. Extensive experiments verify the\nexcellent performance of 3DVLP on three 3D vision-language tasks, reflecting\nits superiority in semantic 3D scene understanding.\n", "date": "2023-05-18", "categories": []}
{"id": "2305.09078", "title": "PanelNet: Understanding 360 Indoor Environment via Panel Representation", "abstract": "  Indoor 360 panoramas have two essential properties. (1) The panoramas are\ncontinuous and seamless in the horizontal direction. (2) Gravity plays an\nimportant role in indoor environment design. By leveraging these properties, we\npresent PanelNet, a framework that understands indoor environments using a\nnovel panel representation of 360 images. We represent an equirectangular\nprojection (ERP) as consecutive vertical panels with corresponding 3D panel\ngeometry. To reduce the negative impact of panoramic distortion, we incorporate\na panel geometry embedding network that encodes both the local and global\ngeometric features of a panel. To capture the geometric context in room design,\nwe introduce Local2Global Transformer, which aggregates local information\nwithin a panel and panel-wise global context. It greatly improves the model\nperformance with low training overhead. Our method outperforms existing methods\non indoor 360 depth estimation and shows competitive results against\nstate-of-the-art approaches on the task of indoor layout estimation and\nsemantic segmentation.\n", "date": "2023-05-16", "categories": []}
{"id": "2305.02743", "title": "Incremental 3D Semantic Scene Graph Prediction from RGB Sequences", "abstract": "  3D semantic scene graphs are a powerful holistic representation as they\ndescribe the individual objects and depict the relation between them. They are\ncompact high-level graphs that enable many tasks requiring scene reasoning. In\nreal-world settings, existing 3D estimation methods produce robust predictions\nthat mostly rely on dense inputs. In this work, we propose a real-time\nframework that incrementally builds a consistent 3D semantic scene graph of a\nscene given an RGB image sequence. Our method consists of a novel incremental\nentity estimation pipeline and a scene graph prediction network. The proposed\npipeline simultaneously reconstructs a sparse point map and fuses entity\nestimation from the input images. The proposed network estimates 3D semantic\nscene graphs with iterative message passing using multi-view and geometric\nfeatures extracted from the scene entities. Extensive experiments on the 3RScan\ndataset show the effectiveness of the proposed method in this challenging task,\noutperforming state-of-the-art approaches.\n", "date": "2023-05-04", "categories": []}
{"id": "2305.02627", "title": "UrbanBIS: a Large-scale Benchmark for Fine-grained Urban Building\n  Instance Segmentation", "abstract": "  We present the UrbanBIS benchmark for large-scale 3D urban understanding,\nsupporting practical urban-level semantic and building-level instance\nsegmentation. UrbanBIS comprises six real urban scenes, with 2.5 billion\npoints, covering a vast area of 10.78 square kilometers and 3,370 buildings,\ncaptured by 113,346 views of aerial photogrammetry. Particularly, UrbanBIS\nprovides not only semantic-level annotations on a rich set of urban objects,\nincluding buildings, vehicles, vegetation, roads, and bridges, but also\ninstance-level annotations on the buildings. Further, UrbanBIS is the first 3D\ndataset that introduces fine-grained building sub-categories, considering a\nwide variety of shapes for different building types. Besides, we propose B-Seg,\na building instance segmentation method to establish UrbanBIS. B-Seg adopts an\nend-to-end framework with a simple yet effective strategy for handling\nlarge-scale point clouds. Compared with mainstream methods, B-Seg achieves\nbetter accuracy with faster inference speed on UrbanBIS. In addition to the\ncarefully-annotated point clouds, UrbanBIS provides high-resolution\naerial-acquisition photos and high-quality large-scale 3D reconstruction\nmodels, which shall facilitate a wide range of studies such as multi-view\nstereo, urban LOD generation, aerial path planning, autonomous navigation, road\nnetwork extraction, and so on, thus serving as an important platform for many\nintelligent city applications.\n", "date": "2023-05-04", "categories": []}
{"id": "2301.04101", "title": "Neural Radiance Field Codebooks", "abstract": "  Compositional representations of the world are a promising step towards\nenabling high-level scene understanding and efficient transfer to downstream\ntasks. Learning such representations for complex scenes and tasks remains an\nopen challenge. Towards this goal, we introduce Neural Radiance Field Codebooks\n(NRC), a scalable method for learning object-centric representations through\nnovel view reconstruction. NRC learns to reconstruct scenes from novel views\nusing a dictionary of object codes which are decoded through a volumetric\nrenderer. This enables the discovery of reoccurring visual and geometric\npatterns across scenes which are transferable to downstream tasks. We show that\nNRC representations transfer well to object navigation in THOR, outperforming\n2D and 3D representation learning methods by 3.1% success rate. We demonstrate\nthat our approach is able to perform unsupervised segmentation for more complex\nsynthetic (THOR) and real scenes (NYU Depth) better than prior methods (29%\nrelative improvement). Finally, we show that NRC improves on the task of depth\nordering by 5.5% accuracy in THOR.\n", "date": "2023-01-10", "categories": []}
{"id": "2304.12624", "title": "Shape-Net: Room Layout Estimation from Panoramic Images Robust to\n  Occlusion using Knowledge Distillation with 3D Shapes as Additional Inputs", "abstract": "  Estimating the layout of a room from a single-shot panoramic image is\nimportant in virtual/augmented reality and furniture layout simulation. This\ninvolves identifying three-dimensional (3D) geometry, such as the location of\ncorners and boundaries, and performing 3D reconstruction. However, occlusion is\na common issue that can negatively impact room layout estimation, and this has\nnot been thoroughly studied to date. It is possible to obtain 3D shape\ninformation of rooms as drawings of buildings and coordinates of corners from\nimage datasets, thus we propose providing both 2D panoramic and 3D information\nto a model to effectively deal with occlusion. However, simply feeding 3D\ninformation to a model is not sufficient to utilize the shape information for\nan occluded area. Therefore, we improve the model by introducing 3D\nIntersection over Union (IoU) loss to effectively use 3D information. In some\ncases, drawings are not available or the construction deviates from a drawing.\nConsidering such practical cases, we propose a method for distilling knowledge\nfrom a model trained with both images and 3D information to a model that takes\nonly images as input. The proposed model, which is called Shape-Net, achieves\nstate-of-the-art (SOTA) performance on benchmark datasets. We also confirmed\nits effectiveness in dealing with occlusion through significantly improved\naccuracy on images with occlusion compared with existing models.\n", "date": "2023-04-25", "categories": []}
{"id": "2304.08492", "title": "STRAP: Structured Object Affordance Segmentation with Point Supervision", "abstract": "  With significant annotation savings, point supervision has been proven\neffective for numerous 2D and 3D scene understanding problems. This success is\nprimarily attributed to the structured output space; i.e., samples with high\nspatial affinity tend to share the same labels. Sharing this spirit, we study\naffordance segmentation with point supervision, wherein the setting inherits an\nunexplored dual affinity-spatial affinity and label affinity. By label\naffinity, we refer to affordance segmentation as a multi-label prediction\nproblem: A plate can be both holdable and containable. By spatial affinity, we\nrefer to a universal prior that nearby pixels with similar visual features\nshould share the same point annotation. To tackle label affinity, we devise a\ndense prediction network that enhances label relations by effectively\ndensifying labels in a new domain (i.e., label co-occurrence). To address\nspatial affinity, we exploit a Transformer backbone for global patch\ninteraction and a regularization loss. In experiments, we benchmark our method\non the challenging CAD120 dataset, showing significant performance gains over\nprior methods.\n", "date": "2023-04-17", "categories": []}
{"id": "2210.07474", "title": "SQA3D: Situated Question Answering in 3D Scenes", "abstract": "  We propose a new task to benchmark scene understanding of embodied agents:\nSituated Question Answering in 3D Scenes (SQA3D). Given a scene context (e.g.,\n3D scan), SQA3D requires the tested agent to first understand its situation\n(position, orientation, etc.) in the 3D scene as described by text, then reason\nabout its surrounding environment and answer a question under that situation.\nBased upon 650 scenes from ScanNet, we provide a dataset centered around 6.8k\nunique situations, along with 20.4k descriptions and 33.4k diverse reasoning\nquestions for these situations. These questions examine a wide spectrum of\nreasoning capabilities for an intelligent agent, ranging from spatial relation\ncomprehension to commonsense understanding, navigation, and multi-hop\nreasoning. SQA3D imposes a significant challenge to current multi-modal\nespecially 3D reasoning models. We evaluate various state-of-the-art approaches\nand find that the best one only achieves an overall score of 47.20%, while\namateur human participants can reach 90.06%. We believe SQA3D could facilitate\nfuture embodied AI research with stronger situation understanding and reasoning\ncapability.\n", "date": "2022-10-14", "categories": []}
{"id": "2304.06061", "title": "CLIP-Guided Vision-Language Pre-training for Question Answering in 3D\n  Scenes", "abstract": "  Training models to apply linguistic knowledge and visual concepts from 2D\nimages to 3D world understanding is a promising direction that researchers have\nonly recently started to explore. In this work, we design a novel 3D\npre-training Vision-Language method that helps a model learn semantically\nmeaningful and transferable 3D scene point cloud representations. We inject the\nrepresentational power of the popular CLIP model into our 3D encoder by\naligning the encoded 3D scene features with the corresponding 2D image and text\nembeddings produced by CLIP. To assess our model's 3D world reasoning\ncapability, we evaluate it on the downstream task of 3D Visual Question\nAnswering. Experimental quantitative and qualitative results show that our\npre-training method outperforms state-of-the-art works in this task and leads\nto an interpretable representation of 3D scene features.\n", "date": "2023-04-12", "categories": []}
{"id": "2304.05316", "title": "OccFormer: Dual-path Transformer for Vision-based 3D Semantic Occupancy\n  Prediction", "abstract": "  The vision-based perception for autonomous driving has undergone a\ntransformation from the bird-eye-view (BEV) representations to the 3D semantic\noccupancy. Compared with the BEV planes, the 3D semantic occupancy further\nprovides structural information along the vertical direction. This paper\npresents OccFormer, a dual-path transformer network to effectively process the\n3D volume for semantic occupancy prediction. OccFormer achieves a long-range,\ndynamic, and efficient encoding of the camera-generated 3D voxel features. It\nis obtained by decomposing the heavy 3D processing into the local and global\ntransformer pathways along the horizontal plane. For the occupancy decoder, we\nadapt the vanilla Mask2Former for 3D semantic occupancy by proposing\npreserve-pooling and class-guided sampling, which notably mitigate the sparsity\nand class imbalance. Experimental results demonstrate that OccFormer\nsignificantly outperforms existing methods for semantic scene completion on\nSemanticKITTI dataset and for LiDAR semantic segmentation on nuScenes dataset.\nCode is available at \\url{https://github.com/zhangyp15/OccFormer}.\n", "date": "2023-04-11", "categories": []}
{"id": "2207.11232", "title": "Neural Groundplans: Persistent Neural Scene Representations from a\n  Single Image", "abstract": "  We present a method to map 2D image observations of a scene to a persistent\n3D scene representation, enabling novel view synthesis and disentangled\nrepresentation of the movable and immovable components of the scene. Motivated\nby the bird's-eye-view (BEV) representation commonly used in vision and\nrobotics, we propose conditional neural groundplans, ground-aligned 2D feature\ngrids, as persistent and memory-efficient scene representations. Our method is\ntrained self-supervised from unlabeled multi-view observations using\ndifferentiable rendering, and learns to complete geometry and appearance of\noccluded regions. In addition, we show that we can leverage multi-view videos\nat training time to learn to separately reconstruct static and movable\ncomponents of the scene from a single image at test time. The ability to\nseparately reconstruct movable objects enables a variety of downstream tasks\nusing simple heuristics, such as extraction of object-centric 3D\nrepresentations, novel view synthesis, instance-level segmentation, 3D bounding\nbox prediction, and scene editing. This highlights the value of neural\ngroundplans as a backbone for efficient 3D scene understanding models.\n", "date": "2022-07-22", "categories": []}
{"id": "2211.15654", "title": "OpenScene: 3D Scene Understanding with Open Vocabularies", "abstract": "  Traditional 3D scene understanding approaches rely on labeled 3D datasets to\ntrain a model for a single task with supervision. We propose OpenScene, an\nalternative approach where a model predicts dense features for 3D scene points\nthat are co-embedded with text and image pixels in CLIP feature space. This\nzero-shot approach enables task-agnostic training and open-vocabulary queries.\nFor example, to perform SOTA zero-shot 3D semantic segmentation it first infers\nCLIP features for every 3D point and later classifies them based on\nsimilarities to embeddings of arbitrary class labels. More interestingly, it\nenables a suite of open-vocabulary scene understanding applications that have\nnever been done before. For example, it allows a user to enter an arbitrary\ntext query and then see a heat map indicating which parts of a scene match. Our\napproach is effective at identifying objects, materials, affordances,\nactivities, and room types in complex 3D scenes, all using a single model\ntrained without any labeled 3D data.\n", "date": "2022-11-28", "categories": []}
{"id": "2301.04926", "title": "CLIP2Scene: Towards Label-efficient 3D Scene Understanding by CLIP", "abstract": "  Contrastive Language-Image Pre-training (CLIP) achieves promising results in\n2D zero-shot and few-shot learning. Despite the impressive performance in 2D,\napplying CLIP to help the learning in 3D scene understanding has yet to be\nexplored. In this paper, we make the first attempt to investigate how CLIP\nknowledge benefits 3D scene understanding. We propose CLIP2Scene, a simple yet\neffective framework that transfers CLIP knowledge from 2D image-text\npre-trained models to a 3D point cloud network. We show that the pre-trained 3D\nnetwork yields impressive performance on various downstream tasks, i.e.,\nannotation-free and fine-tuning with labelled data for semantic segmentation.\nSpecifically, built upon CLIP, we design a Semantic-driven Cross-modal\nContrastive Learning framework that pre-trains a 3D network via semantic and\nspatial-temporal consistency regularization. For the former, we first leverage\nCLIP's text semantics to select the positive and negative point samples and\nthen employ the contrastive loss to train the 3D network. In terms of the\nlatter, we force the consistency between the temporally coherent point cloud\nfeatures and their corresponding image features. We conduct experiments on\nSemanticKITTI, nuScenes, and ScanNet. For the first time, our pre-trained\nnetwork achieves annotation-free 3D semantic segmentation with 20.8% and 25.08%\nmIoU on nuScenes and ScanNet, respectively. When fine-tuned with 1% or 100%\nlabelled data, our method significantly outperforms other self-supervised\nmethods, with improvements of 8% and 1% mIoU, respectively. Furthermore, we\ndemonstrate the generalizability for handling cross-domain datasets. Code is\npublicly available https://github.com/runnanchen/CLIP2Scene.\n", "date": "2023-01-12", "categories": []}
{"id": "2212.06250", "title": "ScanEnts3D: Exploiting Phrase-to-3D-Object Correspondences for Improved\n  Visio-Linguistic Models in 3D Scenes", "abstract": "  The two popular datasets ScanRefer [16] and ReferIt3D [3] connect natural\nlanguage to real-world 3D data. In this paper, we curate a large-scale and\ncomplementary dataset extending both the aforementioned ones by associating all\nobjects mentioned in a referential sentence to their underlying instances\ninside a 3D scene. Specifically, our Scan Entities in 3D (ScanEnts3D) dataset\nprovides explicit correspondences between 369k objects across 84k natural\nreferential sentences, covering 705 real-world scenes. Crucially, we show that\nby incorporating intuitive losses that enable learning from this novel dataset,\nwe can significantly improve the performance of several recently introduced\nneural listening architectures, including improving the SoTA in both the Nr3D\nand ScanRefer benchmarks by 4.3% and 5.0%, respectively. Moreover, we\nexperiment with competitive baselines and recent methods for the task of\nlanguage generation and show that, as with neural listeners, 3D neural speakers\ncan also noticeably benefit by training with ScanEnts3D, including improving\nthe SoTA by 13.2 CIDEr points on the Nr3D benchmark. Overall, our carefully\nconducted experimental studies strongly support the conclusion that, by\nlearning on ScanEnts3D, commonly used visio-linguistic 3D architectures can\nbecome more efficient and interpretable in their generalization without needing\nto provide these newly collected annotations at test time. The project's\nwebpage is https://scanents3d.github.io/ .\n", "date": "2022-12-12", "categories": []}
{"id": "2304.00690", "title": "3D Semantic Segmentation in the Wild: Learning Generalized Models for\n  Adverse-Condition Point Clouds", "abstract": "  Robust point cloud parsing under all-weather conditions is crucial to level-5\nautonomy in autonomous driving. However, how to learn a universal 3D semantic\nsegmentation (3DSS) model is largely neglected as most existing benchmarks are\ndominated by point clouds captured under normal weather. We introduce\nSemanticSTF, an adverse-weather point cloud dataset that provides dense\npoint-level annotations and allows to study 3DSS under various adverse weather\nconditions. We study all-weather 3DSS modeling under two setups: 1) domain\nadaptive 3DSS that adapts from normal-weather data to adverse-weather data; 2)\ndomain generalizable 3DSS that learns all-weather 3DSS models from\nnormal-weather data. Our studies reveal the challenge while existing 3DSS\nmethods encounter adverse-weather data, showing the great value of SemanticSTF\nin steering the future endeavor along this very meaningful research direction.\nIn addition, we design a domain randomization technique that alternatively\nrandomizes the geometry styles of point clouds and aggregates their embeddings,\nultimately leading to a generalizable model that can improve 3DSS under various\nadverse weather effectively. The SemanticSTF and related codes are available at\n\\url{https://github.com/xiaoaoran/SemanticSTF}.\n", "date": "2023-04-03", "categories": []}
{"id": "2211.15658", "title": "Connecting the Dots: Floorplan Reconstruction Using Two-Level Queries", "abstract": "  We address 2D floorplan reconstruction from 3D scans. Existing approaches\ntypically employ heuristically designed multi-stage pipelines. Instead, we\nformulate floorplan reconstruction as a single-stage structured prediction\ntask: find a variable-size set of polygons, which in turn are variable-length\nsequences of ordered vertices. To solve it we develop a novel Transformer\narchitecture that generates polygons of multiple rooms in parallel, in a\nholistic manner without hand-crafted intermediate stages. The model features\ntwo-level queries for polygons and corners, and includes polygon matching to\nmake the network end-to-end trainable. Our method achieves a new\nstate-of-the-art for two challenging datasets, Structured3D and SceneCAD, along\nwith significantly faster inference than previous methods. Moreover, it can\nreadily be extended to predict additional information, i.e., semantic room\ntypes and architectural elements like doors and windows. Our code and models\nare available at: https://github.com/ywyue/RoomFormer.\n", "date": "2022-11-28", "categories": []}
{"id": "2303.16094", "title": "LinK: Linear Kernel for LiDAR-based 3D Perception", "abstract": "  Extending the success of 2D Large Kernel to 3D perception is challenging due\nto: 1. the cubically-increasing overhead in processing 3D data; 2. the\noptimization difficulties from data scarcity and sparsity. Previous work has\ntaken the first step to scale up the kernel size from 3x3x3 to 7x7x7 by\nintroducing block-shared weights. However, to reduce the feature variations\nwithin a block, it only employs modest block size and fails to achieve larger\nkernels like the 21x21x21. To address this issue, we propose a new method,\ncalled LinK, to achieve a wider-range perception receptive field in a\nconvolution-like manner with two core designs. The first is to replace the\nstatic kernel matrix with a linear kernel generator, which adaptively provides\nweights only for non-empty voxels. The second is to reuse the pre-computed\naggregation results in the overlapped blocks to reduce computation complexity.\nThe proposed method successfully enables each voxel to perceive context within\na range of 21x21x21. Extensive experiments on two basic perception tasks, 3D\nobject detection and 3D semantic segmentation, demonstrate the effectiveness of\nour method. Notably, we rank 1st on the public leaderboard of the 3D detection\nbenchmark of nuScenes (LiDAR track), by simply incorporating a LinK-based\nbackbone into the basic detector, CenterPoint. We also boost the strong\nsegmentation baseline's mIoU with 2.7% in the SemanticKITTI test set. Code is\navailable at https://github.com/MCG-NJU/LinK.\n", "date": "2023-03-28", "categories": []}
{"id": "2212.13545", "title": "Interactive Segmentation of Radiance Fields", "abstract": "  Radiance Fields (RF) are popular to represent casually-captured scenes for\nnew view synthesis and several applications beyond it. Mixed reality on\npersonal spaces needs understanding and manipulating scenes represented as RFs,\nwith semantic segmentation of objects as an important step. Prior segmentation\nefforts show promise but don't scale to complex objects with diverse\nappearance. We present the ISRF method to interactively segment objects with\nfine structure and appearance. Nearest neighbor feature matching using\ndistilled semantic features identifies high-confidence seed regions. Bilateral\nsearch in a joint spatio-semantic space grows the region to recover accurate\nsegmentation. We show state-of-the-art results of segmenting objects from RFs\nand compositing them to another scene, changing appearance, etc., and an\ninteractive segmentation tool that others can use.\n  Project Page: https://rahul-goel.github.io/isrf/\n", "date": "2022-12-27", "categories": []}
{"id": "2302.12251", "title": "VoxFormer: Sparse Voxel Transformer for Camera-based 3D Semantic Scene\n  Completion", "abstract": "  Humans can easily imagine the complete 3D geometry of occluded objects and\nscenes. This appealing ability is vital for recognition and understanding. To\nenable such capability in AI systems, we propose VoxFormer, a Transformer-based\nsemantic scene completion framework that can output complete 3D volumetric\nsemantics from only 2D images. Our framework adopts a two-stage design where we\nstart from a sparse set of visible and occupied voxel queries from depth\nestimation, followed by a densification stage that generates dense 3D voxels\nfrom the sparse ones. A key idea of this design is that the visual features on\n2D images correspond only to the visible scene structures rather than the\noccluded or empty spaces. Therefore, starting with the featurization and\nprediction of the visible structures is more reliable. Once we obtain the set\nof sparse queries, we apply a masked autoencoder design to propagate the\ninformation to all the voxels by self-attention. Experiments on SemanticKITTI\nshow that VoxFormer outperforms the state of the art with a relative\nimprovement of 20.0% in geometry and 18.1% in semantics and reduces GPU memory\nduring training to less than 16GB. Our code is available on\nhttps://github.com/NVlabs/VoxFormer.\n", "date": "2023-02-23", "categories": []}
{"id": "2303.10967", "title": "Real-time 3D Semantic Scene Completion Via Feature Aggregation and\n  Conditioned Prediction", "abstract": "  Semantic Scene Completion (SSC) aims to simultaneously predict the volumetric\noccupancy and semantic category of a 3D scene. In this paper, we propose a\nreal-time semantic scene completion method with a feature aggregation strategy\nand conditioned prediction module. Feature aggregation fuses feature with\ndifferent receptive fields and gathers context to improve scene completion\nperformance. And the conditioned prediction module adopts a two-step prediction\nscheme that takes volumetric occupancy as a condition to enhance semantic\ncompletion prediction. We conduct experiments on three recognized benchmarks\nNYU, NYUCAD, and SUNCG. Our method achieves competitive performance at a speed\nof 110 FPS on one GTX 1080 Ti GPU.\n", "date": "2023-03-20", "categories": []}
{"id": "2303.12417", "title": "CLIP$^2$: Contrastive Language-Image-Point Pretraining from Real-World\n  Point Cloud Data", "abstract": "  Contrastive Language-Image Pre-training, benefiting from large-scale\nunlabeled text-image pairs, has demonstrated great performance in open-world\nvision understanding tasks. However, due to the limited Text-3D data pairs,\nadapting the success of 2D Vision-Language Models (VLM) to the 3D space remains\nan open problem. Existing works that leverage VLM for 3D understanding\ngenerally resort to constructing intermediate 2D representations for the 3D\ndata, but at the cost of losing 3D geometry information. To take a step toward\nopen-world 3D vision understanding, we propose Contrastive Language-Image-Point\nCloud Pretraining (CLIP$^2$) to directly learn the transferable 3D point cloud\nrepresentation in realistic scenarios with a novel proxy alignment mechanism.\nSpecifically, we exploit naturally-existed correspondences in 2D and 3D\nscenarios, and build well-aligned and instance-based text-image-point proxies\nfrom those complex scenarios. On top of that, we propose a cross-modal\ncontrastive objective to learn semantic and instance-level aligned point cloud\nrepresentation. Experimental results on both indoor and outdoor scenarios show\nthat our learned 3D representation has great transfer ability in downstream\ntasks, including zero-shot and few-shot 3D recognition, which boosts the\nstate-of-the-art methods by large margins. Furthermore, we provide analyses of\nthe capability of different representations in real scenarios and present the\noptional ensemble scheme.\n", "date": "2023-03-22", "categories": []}
{"id": "2303.14408", "title": "VL-SAT: Visual-Linguistic Semantics Assisted Training for 3D Semantic\n  Scene Graph Prediction in Point Cloud", "abstract": "  The task of 3D semantic scene graph (3DSSG) prediction in the point cloud is\nchallenging since (1) the 3D point cloud only captures geometric structures\nwith limited semantics compared to 2D images, and (2) long-tailed relation\ndistribution inherently hinders the learning of unbiased prediction. Since 2D\nimages provide rich semantics and scene graphs are in nature coped with\nlanguages, in this study, we propose Visual-Linguistic Semantics Assisted\nTraining (VL-SAT) scheme that can significantly empower 3DSSG prediction models\nwith discrimination about long-tailed and ambiguous semantic relations. The key\nidea is to train a powerful multi-modal oracle model to assist the 3D model.\nThis oracle learns reliable structural representations based on semantics from\nvision, language, and 3D geometry, and its benefits can be heterogeneously\npassed to the 3D model during the training stage. By effectively utilizing\nvisual-linguistic semantics in training, our VL-SAT can significantly boost\ncommon 3DSSG prediction models, such as SGFN and SGGpoint, only with 3D inputs\nin the inference stage, especially when dealing with tail relation triplets.\nComprehensive evaluations and ablation studies on the 3DSSG dataset have\nvalidated the effectiveness of the proposed scheme. Code is available at\nhttps://github.com/wz7in/CVPR2023-VLSAT.\n", "date": "2023-03-25", "categories": []}
{"id": "2211.14306", "title": "RUST: Latent Neural Scene Representations from Unposed Imagery", "abstract": "  Inferring the structure of 3D scenes from 2D observations is a fundamental\nchallenge in computer vision. Recently popularized approaches based on neural\nscene representations have achieved tremendous impact and have been applied\nacross a variety of applications. One of the major remaining challenges in this\nspace is training a single model which can provide latent representations which\neffectively generalize beyond a single scene. Scene Representation Transformer\n(SRT) has shown promise in this direction, but scaling it to a larger set of\ndiverse scenes is challenging and necessitates accurately posed ground truth\ndata. To address this problem, we propose RUST (Really Unposed Scene\nrepresentation Transformer), a pose-free approach to novel view synthesis\ntrained on RGB images alone. Our main insight is that one can train a Pose\nEncoder that peeks at the target image and learns a latent pose embedding which\nis used by the decoder for view synthesis. We perform an empirical\ninvestigation into the learned latent pose structure and show that it allows\nmeaningful test-time camera transformations and accurate explicit pose\nreadouts. Perhaps surprisingly, RUST achieves similar quality as methods which\nhave access to perfect camera pose, thereby unlocking the potential for\nlarge-scale training of amortized neural scene representations.\n", "date": "2022-11-25", "categories": []}
{"id": "2211.16312", "title": "PLA: Language-Driven Open-Vocabulary 3D Scene Understanding", "abstract": "  Open-vocabulary scene understanding aims to localize and recognize unseen\ncategories beyond the annotated label space. The recent breakthrough of 2D\nopen-vocabulary perception is largely driven by Internet-scale paired\nimage-text data with rich vocabulary concepts. However, this success cannot be\ndirectly transferred to 3D scenarios due to the inaccessibility of large-scale\n3D-text pairs. To this end, we propose to distill knowledge encoded in\npre-trained vision-language (VL) foundation models through captioning\nmulti-view images from 3D, which allows explicitly associating 3D and\nsemantic-rich captions. Further, to foster coarse-to-fine visual-semantic\nrepresentation learning from captions, we design hierarchical 3D-caption pairs,\nleveraging geometric constraints between 3D scenes and multi-view images.\nFinally, by employing contrastive learning, the model learns language-aware\nembeddings that connect 3D and text for open-vocabulary tasks. Our method not\nonly remarkably outperforms baseline methods by 25.8% $\\sim$ 44.7% hIoU and\n14.5% $\\sim$ 50.4% hAP$_{50}$ in open-vocabulary semantic and instance\nsegmentation, but also shows robust transferability on challenging zero-shot\ndomain transfer tasks. See the project website at\nhttps://dingry.github.io/projects/PLA.\n", "date": "2022-11-29", "categories": []}
{"id": "2303.10585", "title": "Label Name is Mantra: Unifying Point Cloud Segmentation across\n  Heterogeneous Datasets", "abstract": "  Point cloud segmentation is a fundamental task in 3D vision that serves a\nwide range of applications. Although great progresses have been made these\nyears, its practical usability is still limited by the availability of training\ndata. Existing approaches cannot make full use of multiple datasets on hand due\nto the label mismatch among different datasets. In this paper, we propose a\nprincipled approach that supports learning from heterogeneous datasets with\ndifferent label sets. Our idea is to utilize a pre-trained language model to\nembed discrete labels to a continuous latent space with the help of their label\nnames. This unifies all labels of different datasets, so that joint training is\ndoable. Meanwhile, classifying points in the continuous 3D space by their\nvocabulary tokens significantly increase the generalization ability of the\nmodel in comparison with existing approaches that have fixed decoder\narchitecture. Besides, we also integrate prompt learning in our framework to\nalleviate data shifts among different data sources. Extensive experiments\ndemonstrate that our model outperforms the state-of-the-art by a large margin.\n", "date": "2023-03-19", "categories": []}
{"id": "2303.10962", "title": "Neural Implicit Vision-Language Feature Fields", "abstract": "  Recently, groundbreaking results have been presented on open-vocabulary\nsemantic image segmentation. Such methods segment each pixel in an image into\narbitrary categories provided at run-time in the form of text prompts, as\nopposed to a fixed set of classes defined at training time. In this work, we\npresent a zero-shot volumetric open-vocabulary semantic scene segmentation\nmethod. Our method builds on the insight that we can fuse image features from a\nvision-language model into a neural implicit representation. We show that the\nresulting feature field can be segmented into different classes by assigning\npoints to natural language text prompts. The implicit volumetric representation\nenables us to segment the scene both in 3D and 2D by rendering feature maps\nfrom any given viewpoint of the scene. We show that our method works on noisy\nreal-world data and can run in real-time on live sensor data dynamically\nadjusting to text prompts. We also present quantitative comparisons on the\nScanNet dataset.\n", "date": "2023-03-20", "categories": []}
{"id": "2303.09977", "title": "Semantic Scene Completion with Cleaner Self", "abstract": "  Semantic Scene Completion (SSC) transforms an image of single-view depth\nand/or RGB 2D pixels into 3D voxels, each of whose semantic labels are\npredicted. SSC is a well-known ill-posed problem as the prediction model has to\n\"imagine\" what is behind the visible surface, which is usually represented by\nTruncated Signed Distance Function (TSDF). Due to the sensory imperfection of\nthe depth camera, most existing methods based on the noisy TSDF estimated from\ndepth values suffer from 1) incomplete volumetric predictions and 2) confused\nsemantic labels. To this end, we use the ground-truth 3D voxels to generate a\nperfect visible surface, called TSDF-CAD, and then train a \"cleaner\" SSC model.\nAs the model is noise-free, it is expected to focus more on the \"imagination\"\nof unseen voxels. Then, we propose to distill the intermediate \"cleaner\"\nknowledge into another model with noisy TSDF input. In particular, we use the\n3D occupancy feature and the semantic relations of the \"cleaner self\" to\nsupervise the counterparts of the \"noisy self\" to respectively address the\nabove two incorrect predictions. Experimental results validate that our method\nimproves the noisy counterparts with 3.1% IoU and 2.2% mIoU for measuring scene\ncompletion and SSC, and also achieves new state-of-the-art accuracy on the\npopular NYU dataset.\n", "date": "2023-03-17", "categories": []}
{"id": "2303.04748", "title": "CLIP-FO3D: Learning Free Open-world 3D Scene Representations from 2D\n  Dense CLIP", "abstract": "  Training a 3D scene understanding model requires complicated human\nannotations, which are laborious to collect and result in a model only encoding\nclose-set object semantics. In contrast, vision-language pre-training models\n(e.g., CLIP) have shown remarkable open-world reasoning properties. To this\nend, we propose directly transferring CLIP's feature space to 3D scene\nunderstanding model without any form of supervision. We first modify CLIP's\ninput and forwarding process so that it can be adapted to extract dense pixel\nfeatures for 3D scene contents. We then project multi-view image features to\nthe point cloud and train a 3D scene understanding model with feature\ndistillation. Without any annotations or additional training, our model\nachieves promising annotation-free semantic segmentation results on\nopen-vocabulary semantics and long-tailed concepts. Besides, serving as a\ncross-modal pre-training framework, our method can be used to improve data\nefficiency during fine-tuning. Our model outperforms previous SOTA methods in\nvarious zero-shot and data-efficient learning benchmarks. Most importantly, our\nmodel successfully inherits CLIP's rich-structured knowledge, allowing 3D scene\nunderstanding models to recognize not only object concepts but also open-world\nsemantics.\n", "date": "2023-03-08", "categories": []}
{"id": "2303.09553", "title": "LERF: Language Embedded Radiance Fields", "abstract": "  Humans describe the physical world using natural language to refer to\nspecific 3D locations based on a vast range of properties: visual appearance,\nsemantics, abstract associations, or actionable affordances. In this work we\npropose Language Embedded Radiance Fields (LERFs), a method for grounding\nlanguage embeddings from off-the-shelf models like CLIP into NeRF, which enable\nthese types of open-ended language queries in 3D. LERF learns a dense,\nmulti-scale language field inside NeRF by volume rendering CLIP embeddings\nalong training rays, supervising these embeddings across training views to\nprovide multi-view consistency and smooth the underlying language field. After\noptimization, LERF can extract 3D relevancy maps for a broad range of language\nprompts interactively in real-time, which has potential use cases in robotics,\nunderstanding vision-language models, and interacting with 3D scenes. LERF\nenables pixel-aligned, zero-shot queries on the distilled 3D CLIP embeddings\nwithout relying on region proposals or masks, supporting long-tail\nopen-vocabulary queries hierarchically across the volume. The project website\ncan be found at https://lerf.io .\n", "date": "2023-03-16", "categories": []}
{"id": "2303.06884", "title": "SCPNet: Semantic Scene Completion on Point Cloud", "abstract": "  Training deep models for semantic scene completion (SSC) is challenging due\nto the sparse and incomplete input, a large quantity of objects of diverse\nscales as well as the inherent label noise for moving objects. To address the\nabove-mentioned problems, we propose the following three solutions: 1)\nRedesigning the completion sub-network. We design a novel completion\nsub-network, which consists of several Multi-Path Blocks (MPBs) to aggregate\nmulti-scale features and is free from the lossy downsampling operations. 2)\nDistilling rich knowledge from the multi-frame model. We design a novel\nknowledge distillation objective, dubbed Dense-to-Sparse Knowledge Distillation\n(DSKD). It transfers the dense, relation-based semantic knowledge from the\nmulti-frame teacher to the single-frame student, significantly improving the\nrepresentation learning of the single-frame model. 3) Completion label\nrectification. We propose a simple yet effective label rectification strategy,\nwhich uses off-the-shelf panoptic segmentation labels to remove the traces of\ndynamic objects in completion labels, greatly improving the performance of deep\nmodels especially for those moving objects. Extensive experiments are conducted\nin two public SSC benchmarks, i.e., SemanticKITTI and SemanticPOSS. Our SCPNet\nranks 1st on SemanticKITTI semantic scene completion challenge and surpasses\nthe competitive S3CNet by 7.2 mIoU. SCPNet also outperforms previous completion\nalgorithms on the SemanticPOSS dataset. Besides, our method also achieves\ncompetitive results on SemanticKITTI semantic segmentation tasks, showing that\nknowledge learned in the scene completion is beneficial to the segmentation\ntask.\n", "date": "2023-03-13", "categories": []}
{"id": "2208.07227", "title": "DM-NeRF: 3D Scene Geometry Decomposition and Manipulation from 2D Images", "abstract": "  In this paper, we study the problem of 3D scene geometry decomposition and\nmanipulation from 2D views. By leveraging the recent implicit neural\nrepresentation techniques, particularly the appealing neural radiance fields,\nwe introduce an object field component to learn unique codes for all individual\nobjects in 3D space only from 2D supervision. The key to this component is a\nseries of carefully designed loss functions to enable every 3D point,\nespecially in non-occupied space, to be effectively optimized even without 3D\nlabels. In addition, we introduce an inverse query algorithm to freely\nmanipulate any specified 3D object shape in the learned scene representation.\nNotably, our manipulation algorithm can explicitly tackle key issues such as\nobject collisions and visual occlusions. Our method, called DM-NeRF, is among\nthe first to simultaneously reconstruct, decompose, manipulate and render\ncomplex 3D scenes in a single pipeline. Extensive experiments on three datasets\nclearly show that our method can accurately decompose all 3D objects from 2D\nviews, allowing any interested object to be freely manipulated in 3D space such\nas translation, rotation, size adjustment, and deformation.\n", "date": "2022-08-15", "categories": []}
{"id": "2303.03361", "title": "Nerflets: Local Radiance Fields for Efficient Structure-Aware 3D Scene\n  Representation from 2D Supervision", "abstract": "  We address efficient and structure-aware 3D scene representation from images.\nNerflets are our key contribution -- a set of local neural radiance fields that\ntogether represent a scene. Each nerflet maintains its own spatial position,\norientation, and extent, within which it contributes to panoptic, density, and\nradiance reconstructions. By leveraging only photometric and inferred panoptic\nimage supervision, we can directly and jointly optimize the parameters of a set\nof nerflets so as to form a decomposed representation of the scene, where each\nobject instance is represented by a group of nerflets. During experiments with\nindoor and outdoor environments, we find that nerflets: (1) fit and approximate\nthe scene more efficiently than traditional global NeRFs, (2) allow the\nextraction of panoptic and photometric renderings from arbitrary views, and (3)\nenable tasks rare for NeRFs, such as 3D panoptic segmentation and interactive\nediting.\n", "date": "2023-03-06", "categories": []}
{"id": "2301.13510", "title": "3D Former: Monocular Scene Reconstruction with 3D SDF Transformers", "abstract": "  Monocular scene reconstruction from posed images is challenging due to the\ncomplexity of a large environment. Recent volumetric methods learn to directly\npredict the TSDF volume and have demonstrated promising results in this task.\nHowever, most methods focus on how to extract and fuse the 2D features to a 3D\nfeature volume, but none of them improve the way how the 3D volume is\naggregated. In this work, we propose an SDF transformer network, which replaces\nthe role of 3D CNN for better 3D feature aggregation. To reduce the explosive\ncomputation complexity of the 3D multi-head attention, we propose a sparse\nwindow attention module, where the attention is only calculated between the\nnon-empty voxels within a local window. Then a top-down-bottom-up 3D attention\nnetwork is built for 3D feature aggregation, where a dilate-attention structure\nis proposed to prevent geometry degeneration, and two global modules are\nemployed to equip with global receptive fields. The experiments on multiple\ndatasets show that this 3D transformer network generates a more accurate and\ncomplete reconstruction, which outperforms previous methods by a large margin.\nRemarkably, the mesh accuracy is improved by 41.8%, and the mesh completeness\nis improved by 25.3% on the ScanNet dataset. Project page:\nhttps://weihaosky.github.io/sdfformer.\n", "date": "2023-01-31", "categories": []}
{"id": "2302.14363", "title": "Efficient Implicit Neural Reconstruction Using LiDAR", "abstract": "  Modeling scene geometry using implicit neural representation has revealed its\nadvantages in accuracy, flexibility, and low memory usage. Previous approaches\nhave demonstrated impressive results using color or depth images but still have\ndifficulty handling poor light conditions and large-scale scenes. Methods\ntaking global point cloud as input require accurate registration and ground\ntruth coordinate labels, which limits their application scenarios. In this\npaper, we propose a new method that uses sparse LiDAR point clouds and rough\nodometry to reconstruct fine-grained implicit occupancy field efficiently\nwithin a few minutes. We introduce a new loss function that supervises directly\nin 3D space without 2D rendering, avoiding information loss. We also manage to\nrefine poses of input frames in an end-to-end manner, creating consistent\ngeometry without global point cloud registration. As far as we know, our method\nis the first to reconstruct implicit scene representation from LiDAR-only\ninput. Experiments on synthetic and real-world datasets, including indoor and\noutdoor scenes, prove that our method is effective, efficient, and accurate,\nobtaining comparable results with existing methods using dense input.\n", "date": "2023-02-28", "categories": []}
{"id": "2302.13540", "title": "OccDepth: A Depth-Aware Method for 3D Semantic Scene Completion", "abstract": "  3D Semantic Scene Completion (SSC) can provide dense geometric and semantic\nscene representations, which can be applied in the field of autonomous driving\nand robotic systems. It is challenging to estimate the complete geometry and\nsemantics of a scene solely from visual images, and accurate depth information\nis crucial for restoring 3D geometry. In this paper, we propose the first\nstereo SSC method named OccDepth, which fully exploits implicit depth\ninformation from stereo images (or RGBD images) to help the recovery of 3D\ngeometric structures. The Stereo Soft Feature Assignment (Stereo-SFA) module is\nproposed to better fuse 3D depth-aware features by implicitly learning the\ncorrelation between stereo images. In particular, when the input are RGBD\nimage, a virtual stereo images can be generated through original RGB image and\ndepth map. Besides, the Occupancy Aware Depth (OAD) module is used to obtain\ngeometry-aware 3D features by knowledge distillation using pre-trained depth\nmodels. In addition, a reformed TartanAir benchmark, named SemanticTartanAir,\nis provided in this paper for further testing our OccDepth method on SSC task.\nCompared with the state-of-the-art RGB-inferred SSC method, extensive\nexperiments on SemanticKITTI show that our OccDepth method achieves superior\nperformance with improving +4.82% mIoU, of which +2.49% mIoU comes from stereo\nimages and +2.33% mIoU comes from our proposed depth-aware method. Our code and\ntrained models are available at https://github.com/megvii-research/OccDepth.\n", "date": "2023-02-27", "categories": []}
{"id": "2302.14052", "title": "LODE: Locally Conditioned Eikonal Implicit Scene Completion from Sparse\n  LiDAR", "abstract": "  Scene completion refers to obtaining dense scene representation from an\nincomplete perception of complex 3D scenes. This helps robots detect\nmulti-scale obstacles and analyse object occlusions in scenarios such as\nautonomous driving. Recent advances show that implicit representation learning\ncan be leveraged for continuous scene completion and achieved through physical\nconstraints like Eikonal equations. However, former Eikonal completion methods\nonly demonstrate results on watertight meshes at a scale of tens of meshes.\nNone of them are successfully done for non-watertight LiDAR point clouds of\nopen large scenes at a scale of thousands of scenes. In this paper, we propose\na novel Eikonal formulation that conditions the implicit representation on\nlocalized shape priors which function as dense boundary value constraints, and\ndemonstrate it works on SemanticKITTI and SemanticPOSS. It can also be extended\nto semantic Eikonal scene completion with only small modifications to the\nnetwork architecture. With extensive quantitative and qualitative results, we\ndemonstrate the benefits and drawbacks of existing Eikonal methods, which\nnaturally leads to the new locally conditioned formulation. Notably, we improve\nIoU from 31.7% to 51.2% on SemanticKITTI and from 40.5% to 48.7% on\nSemanticPOSS. We extensively ablate our methods and demonstrate that the\nproposed formulation is robust to a wide spectrum of implementation\nhyper-parameters. Codes and models are publicly available at\nhttps://github.com/AIR-DISCOVER/LODE.\n", "date": "2023-02-27", "categories": []}
{"id": "2302.03288", "title": "Object-Centric Scene Representations using Active Inference", "abstract": "  Representing a scene and its constituent objects from raw sensory data is a\ncore ability for enabling robots to interact with their environment. In this\npaper, we propose a novel approach for scene understanding, leveraging a\nhierarchical object-centric generative model that enables an agent to infer\nobject category and pose in an allocentric reference frame using active\ninference, a neuro-inspired framework for action and perception. For evaluating\nthe behavior of an active vision agent, we also propose a new benchmark where,\ngiven a target viewpoint of a particular object, the agent needs to find the\nbest matching viewpoint given a workspace with randomly positioned objects in\n3D. We demonstrate that our active inference agent is able to balance epistemic\nforaging and goal-driven behavior, and outperforms both supervised and\nreinforcement learning baselines by a large margin.\n", "date": "2023-02-07", "categories": []}
{"id": "2301.13865", "title": "From Semi-supervised to Omni-supervised Room Layout Estimation Using\n  Point Clouds", "abstract": "  Room layout estimation is a long-existing robotic vision task that benefits\nboth environment sensing and motion planning. However, layout estimation using\npoint clouds (PCs) still suffers from data scarcity due to annotation\ndifficulty. As such, we address the semi-supervised setting of this task based\nupon the idea of model exponential moving averaging. But adapting this scheme\nto the state-of-the-art (SOTA) solution for PC-based layout estimation is not\nstraightforward. To this end, we define a quad set matching strategy and\nseveral consistency losses based upon metrics tailored for layout quads.\nBesides, we propose a new online pseudo-label harvesting algorithm that\ndecomposes the distribution of a hybrid distance measure between quads and PC\ninto two components. This technique does not need manual threshold selection\nand intuitively encourages quads to align with reliable layout points.\nSurprisingly, this framework also works for the fully-supervised setting,\nachieving a new SOTA on the ScanNet benchmark. Last but not least, we also push\nthe semi-supervised setting to the realistic omni-supervised setting,\ndemonstrating significantly promoted performance on a newly annotated\nARKitScenes testing set. Our codes, data and models are released in this\nrepository.\n", "date": "2023-01-31", "categories": []}
{"id": "2301.06015", "title": "Diffusion-based Generation, Optimization, and Planning in 3D Scenes", "abstract": "  We introduce SceneDiffuser, a conditional generative model for 3D scene\nunderstanding. SceneDiffuser provides a unified model for solving\nscene-conditioned generation, optimization, and planning. In contrast to prior\nworks, SceneDiffuser is intrinsically scene-aware, physics-based, and\ngoal-oriented. With an iterative sampling strategy, SceneDiffuser jointly\nformulates the scene-aware generation, physics-based optimization, and\ngoal-oriented planning via a diffusion-based denoising process in a fully\ndifferentiable fashion. Such a design alleviates the discrepancies among\ndifferent modules and the posterior collapse of previous scene-conditioned\ngenerative models. We evaluate SceneDiffuser with various 3D scene\nunderstanding tasks, including human pose and motion generation, dexterous\ngrasp generation, path planning for 3D navigation, and motion planning for\nrobot arms. The results show significant improvements compared with previous\nmodels, demonstrating the tremendous potential of SceneDiffuser for the broad\ncommunity of 3D scene understanding.\n", "date": "2023-01-15", "categories": []}
{"id": "2301.06962", "title": "Long Range Pooling for 3D Large-Scale Scene Understanding", "abstract": "  Inspired by the success of recent vision transformers and large kernel design\nin convolutional neural networks (CNNs), in this paper, we analyze and explore\nessential reasons for their success. We claim two factors that are critical for\n3D large-scale scene understanding: a larger receptive field and operations\nwith greater non-linearity. The former is responsible for providing long range\ncontexts and the latter can enhance the capacity of the network. To achieve the\nabove properties, we propose a simple yet effective long range pooling (LRP)\nmodule using dilation max pooling, which provides a network with a large\nadaptive receptive field. LRP has few parameters, and can be readily added to\ncurrent CNNs. Also, based on LRP, we present an entire network architecture,\nLRPNet, for 3D understanding. Ablation studies are presented to support our\nclaims, and show that the LRP module achieves better results than large kernel\nconvolution yet with reduced computation, due to its nonlinearity. We also\ndemonstrate the superiority of LRPNet on various benchmarks: LRPNet performs\nthe best on ScanNet and surpasses other CNN-based methods on S3DIS and\nMatterport3D. Code will be made publicly available.\n", "date": "2023-01-17", "categories": []}
{"id": "2212.09802", "title": "Panoptic Lifting for 3D Scene Understanding with Neural Fields", "abstract": "  We propose Panoptic Lifting, a novel approach for learning panoptic 3D\nvolumetric representations from images of in-the-wild scenes. Once trained, our\nmodel can render color images together with 3D-consistent panoptic segmentation\nfrom novel viewpoints.\n  Unlike existing approaches which use 3D input directly or indirectly, our\nmethod requires only machine-generated 2D panoptic segmentation masks inferred\nfrom a pre-trained network. Our core contribution is a panoptic lifting scheme\nbased on a neural field representation that generates a unified and multi-view\nconsistent, 3D panoptic representation of the scene. To account for\ninconsistencies of 2D instance identifiers across views, we solve a linear\nassignment with a cost based on the model's current predictions and the\nmachine-generated segmentation masks, thus enabling us to lift 2D instances to\n3D in a consistent way. We further propose and ablate contributions that make\nour method more robust to noisy, machine-generated labels, including test-time\naugmentations for confidence estimates, segment consistency loss, bounded\nsegmentation fields, and gradient stopping.\n  Experimental results validate our approach on the challenging Hypersim,\nReplica, and ScanNet datasets, improving by 8.4, 13.8, and 10.6% in scene-level\nPQ over state of the art.\n", "date": "2022-12-19", "categories": []}
{"id": "2211.14091", "title": "Language-Assisted 3D Feature Learning for Semantic Scene Understanding", "abstract": "  Learning descriptive 3D features is crucial for understanding 3D scenes with\ndiverse objects and complex structures. However, it is usually unknown whether\nimportant geometric attributes and scene context obtain enough emphasis in an\nend-to-end trained 3D scene understanding network. To guide 3D feature learning\ntoward important geometric attributes and scene context, we explore the help of\ntextual scene descriptions. Given some free-form descriptions paired with 3D\nscenes, we extract the knowledge regarding the object relationships and object\nattributes. We then inject the knowledge to 3D feature learning through three\nclassification-based auxiliary tasks. This language-assisted training can be\ncombined with modern object detection and instance segmentation methods to\npromote 3D semantic scene understanding, especially in a label-deficient\nregime. Moreover, the 3D feature learned with language assistance is better\naligned with the language features, which can benefit various 3D-language\nmultimodal tasks. Experiments on several benchmarks of 3D-only and 3D-language\ntasks demonstrate the effectiveness of our language-assisted 3D feature\nlearning. Code is available at\nhttps://github.com/Asterisci/Language-Assisted-3D.\n", "date": "2022-11-25", "categories": []}
{"id": "2211.16882", "title": "MVRackLay: Monocular Multi-View Layout Estimation for Warehouse Racks\n  and Shelves", "abstract": "  In this paper, we propose and showcase, for the first time, monocular\nmulti-view layout estimation for warehouse racks and shelves. Unlike typical\nlayout estimation methods, MVRackLay estimates multi-layered layouts, wherein\neach layer corresponds to the layout of a shelf within a rack. Given a sequence\nof images of a warehouse scene, a dual-headed Convolutional-LSTM architecture\noutputs segmented racks, the front and the top view layout of each shelf within\na rack. With minimal effort, such an output is transformed into a 3D rendering\nof all racks, shelves and objects on the shelves, giving an accurate 3D\ndepiction of the entire warehouse scene in terms of racks, shelves and the\nnumber of objects on each shelf. MVRackLay generalizes to a diverse set of\nwarehouse scenes with varying number of objects on each shelf, number of\nshelves and in the presence of other such racks in the background. Further,\nMVRackLay shows superior performance vis-a-vis its single view counterpart,\nRackLay, in layout accuracy, quantized in terms of the mean IoU and mAP\nmetrics. We also showcase a multi-view stitching of the 3D layouts resulting in\na representation of the warehouse scene with respect to a global reference\nframe akin to a rendering of the scene from a SLAM pipeline. To the best of our\nknowledge, this is the first such work to portray a 3D rendering of a warehouse\nscene in terms of its semantic components - Racks, Shelves and Objects - all\nfrom a single monocular camera.\n", "date": "2022-11-30", "categories": []}
{"id": "2207.11514", "title": "Semantic Abstraction: Open-World 3D Scene Understanding from 2D\n  Vision-Language Models", "abstract": "  We study open-world 3D scene understanding, a family of tasks that require\nagents to reason about their 3D environment with an open-set vocabulary and\nout-of-domain visual inputs - a critical skill for robots to operate in the\nunstructured 3D world. Towards this end, we propose Semantic Abstraction\n(SemAbs), a framework that equips 2D Vision-Language Models (VLMs) with new 3D\nspatial capabilities, while maintaining their zero-shot robustness. We achieve\nthis abstraction using relevancy maps extracted from CLIP, and learn 3D spatial\nand geometric reasoning skills on top of those abstractions in a\nsemantic-agnostic manner. We demonstrate the usefulness of SemAbs on two\nopen-world 3D scene understanding tasks: 1) completing partially observed\nobjects and 2) localizing hidden objects from language descriptions.\nExperiments show that SemAbs can generalize to novel vocabulary,\nmaterials/lighting, classes, and domains (i.e., real-world scans) from training\non limited 3D synthetic data. Code and data is available at\nhttps://semantic-abstraction.cs.columbia.edu/\n", "date": "2022-07-23", "categories": []}
{"id": "2212.00836", "title": "UniT3D: A Unified Transformer for 3D Dense Captioning and Visual\n  Grounding", "abstract": "  Performing 3D dense captioning and visual grounding requires a common and\nshared understanding of the underlying multimodal relationships. However,\ndespite some previous attempts on connecting these two related tasks with\nhighly task-specific neural modules, it remains understudied how to explicitly\ndepict their shared nature to learn them simultaneously. In this work, we\npropose UniT3D, a simple yet effective fully unified transformer-based\narchitecture for jointly solving 3D visual grounding and dense captioning.\nUniT3D enables learning a strong multimodal representation across the two tasks\nthrough a supervised joint pre-training scheme with bidirectional and\nseq-to-seq objectives. With a generic architecture design, UniT3D allows\nexpanding the pre-training scope to more various training sources such as the\nsynthesized data from 2D prior knowledge to benefit 3D vision-language tasks.\nExtensive experiments and analysis demonstrate that UniT3D obtains significant\ngains for 3D dense captioning and visual grounding.\n", "date": "2022-12-01", "categories": []}
{"id": "2210.06853", "title": "NeuralRoom: Geometry-Constrained Neural Implicit Surfaces for Indoor\n  Scene Reconstruction", "abstract": "  We present a novel neural surface reconstruction method called NeuralRoom for\nreconstructing room-sized indoor scenes directly from a set of 2D images.\nRecently, implicit neural representations have become a promising way to\nreconstruct surfaces from multiview images due to their high-quality results\nand simplicity. However, implicit neural representations usually cannot\nreconstruct indoor scenes well because they suffer severe shape-radiance\nambiguity. We assume that the indoor scene consists of texture-rich and flat\ntexture-less regions. In texture-rich regions, the multiview stereo can obtain\naccurate results. In the flat area, normal estimation networks usually obtain a\ngood normal estimation. Based on the above observations, we reduce the possible\nspatial variation range of implicit neural surfaces by reliable geometric\npriors to alleviate shape-radiance ambiguity. Specifically, we use multiview\nstereo results to limit the NeuralRoom optimization space and then use reliable\ngeometric priors to guide NeuralRoom training. Then the NeuralRoom would\nproduce a neural scene representation that can render an image consistent with\nthe input training images. In addition, we propose a smoothing method called\nperturbation-residual restrictions to improve the accuracy and completeness of\nthe flat region, which assumes that the sampling points in a local surface\nshould have the same normal and similar distance to the observation center.\nExperiments on the ScanNet dataset show that our method can reconstruct the\ntexture-less area of indoor scenes while maintaining the accuracy of detail. We\nalso apply NeuralRoom to more advanced multiview reconstruction algorithms and\nsignificantly improve their reconstruction quality.\n", "date": "2022-10-13", "categories": []}
{"id": "2211.13702", "title": "CasFusionNet: A Cascaded Network for Point Cloud Semantic Scene\n  Completion by Dense Feature Fusion", "abstract": "  Semantic scene completion (SSC) aims to complete a partial 3D scene and\npredict its semantics simultaneously. Most existing works adopt the voxel\nrepresentations, thus suffering from the growth of memory and computation cost\nas the voxel resolution increases. Though a few works attempt to solve SSC from\nthe perspective of 3D point clouds, they have not fully exploited the\ncorrelation and complementarity between the two tasks of scene completion and\nsemantic segmentation. In our work, we present CasFusionNet, a novel cascaded\nnetwork for point cloud semantic scene completion by dense feature fusion.\nSpecifically, we design (i) a global completion module (GCM) to produce an\nupsampled and completed but coarse point set, (ii) a semantic segmentation\nmodule (SSM) to predict the per-point semantic labels of the completed points\ngenerated by GCM, and (iii) a local refinement module (LRM) to further refine\nthe coarse completed points and the associated labels from a local perspective.\nWe organize the above three modules via dense feature fusion in each level, and\ncascade a total of four levels, where we also employ feature fusion between\neach level for sufficient information usage. Both quantitative and qualitative\nresults on our compiled two point-based datasets validate the effectiveness and\nsuperiority of our CasFusionNet compared to state-of-the-art methods in terms\nof both scene completion and semantic segmentation. The codes and datasets are\navailable at: https://github.com/JinfengX/CasFusionNet.\n", "date": "2022-11-24", "categories": []}
{"id": "2211.14157", "title": "Learning 3D Scene Priors with 2D Supervision", "abstract": "  Holistic 3D scene understanding entails estimation of both layout\nconfiguration and object geometry in a 3D environment. Recent works have shown\nadvances in 3D scene estimation from various input modalities (e.g., images, 3D\nscans), by leveraging 3D supervision (e.g., 3D bounding boxes or CAD models),\nfor which collection at scale is expensive and often intractable. To address\nthis shortcoming, we propose a new method to learn 3D scene priors of layout\nand shape without requiring any 3D ground truth. Instead, we rely on 2D\nsupervision from multi-view RGB images. Our method represents a 3D scene as a\nlatent vector, from which we can progressively decode to a sequence of objects\ncharacterized by their class categories, 3D bounding boxes, and meshes. With\nour trained autoregressive decoder representing the scene prior, our method\nfacilitates many downstream applications, including scene synthesis,\ninterpolation, and single-view reconstruction. Experiments on 3D-FRONT and\nScanNet show that our method outperforms state of the art in single-view\nreconstruction, and achieves state-of-the-art results in scene synthesis\nagainst baselines which require for 3D supervision.\n", "date": "2022-11-25", "categories": []}
{"id": "2211.11354", "title": "Object-level 3D Semantic Mapping using a Network of Smart Edge Sensors", "abstract": "  Autonomous robots that interact with their environment require a detailed\nsemantic scene model. For this, volumetric semantic maps are frequently used.\nThe scene understanding can further be improved by including object-level\ninformation in the map. In this work, we extend a multi-view 3D semantic\nmapping system consisting of a network of distributed smart edge sensors with\nobject-level information, to enable downstream tasks that need object-level\ninput. Objects are represented in the map via their 3D mesh model or as an\nobject-centric volumetric sub-map that can model arbitrary object geometry when\nno detailed 3D model is available. We propose a keypoint-based approach to\nestimate object poses via PnP and refinement via ICP alignment of the 3D object\nmodel with the observed point cloud segments. Object instances are tracked to\nintegrate observations over time and to be robust against temporary occlusions.\nOur method is evaluated on the public Behave dataset where it shows pose\nestimation accuracy within a few centimeters and in real-world experiments with\nthe sensor network in a challenging lab environment where multiple chairs and a\ntable are tracked through the scene online, in real time even under high\nocclusions.\n", "date": "2022-11-21", "categories": []}
{"id": "2211.09375", "title": "3D-QueryIS: A Query-based Framework for 3D Instance Segmentation", "abstract": "  Previous top-performing methods for 3D instance segmentation often maintain\ninter-task dependencies and the tendency towards a lack of robustness. Besides,\ninevitable variations of different datasets make these methods become\nparticularly sensitive to hyper-parameter values and manifest poor\ngeneralization capability. In this paper, we address the aforementioned\nchallenges by proposing a novel query-based method, termed as 3D-QueryIS, which\nis detector-free, semantic segmentation-free, and cluster-free. Specifically,\nwe propose to generate representative points in an implicit manner, and use\nthem together with the initial queries to generate the informative instance\nqueries. Then, the class and binary instance mask predictions can be produced\nby simply applying MLP layers on top of the instance queries and the extracted\npoint cloud embeddings. Thus, our 3D-QueryIS is free from the accumulated\nerrors caused by the inter-task dependencies. Extensive experiments on multiple\nbenchmark datasets demonstrate the effectiveness and efficiency of our proposed\n3D-QueryIS method.\n", "date": "2022-11-17", "categories": []}
{"id": "2206.01256", "title": "PETRv2: A Unified Framework for 3D Perception from Multi-Camera Images", "abstract": "  In this paper, we propose PETRv2, a unified framework for 3D perception from\nmulti-view images. Based on PETR, PETRv2 explores the effectiveness of temporal\nmodeling, which utilizes the temporal information of previous frames to boost\n3D object detection. More specifically, we extend the 3D position embedding (3D\nPE) in PETR for temporal modeling. The 3D PE achieves the temporal alignment on\nobject position of different frames. A feature-guided position encoder is\nfurther introduced to improve the data adaptability of 3D PE. To support for\nmulti-task learning (e.g., BEV segmentation and 3D lane detection), PETRv2\nprovides a simple yet effective solution by introducing task-specific queries,\nwhich are initialized under different spaces. PETRv2 achieves state-of-the-art\nperformance on 3D object detection, BEV segmentation and 3D lane detection.\nDetailed robustness analysis is also conducted on PETR framework. We hope\nPETRv2 can serve as a strong baseline for 3D perception. Code is available at\n\\url{https://github.com/megvii-research/PETR}.\n", "date": "2022-06-02", "categories": []}
{"id": "2210.06240", "title": "Explore Contextual Information for 3D Scene Graph Generation", "abstract": "  3D scene graph generation (SGG) has been of high interest in computer vision.\nAlthough the accuracy of 3D SGG on coarse classification and single relation\nlabel has been gradually improved, the performance of existing works is still\nfar from being perfect for fine-grained and multi-label situations. In this\npaper, we propose a framework fully exploring contextual information for the 3D\nSGG task, which attempts to satisfy the requirements of fine-grained entity\nclass, multiple relation labels, and high accuracy simultaneously. Our proposed\napproach is composed of a Graph Feature Extraction module and a Graph\nContextual Reasoning module, achieving appropriate information-redundancy\nfeature extraction, structured organization, and hierarchical inferring. Our\napproach achieves superior or competitive performance over previous methods on\nthe 3DSSG dataset, especially on the relationship prediction sub-task.\n", "date": "2022-10-12", "categories": []}
{"id": "2211.00562", "title": "Leveraging commonsense for object localisation in partial scenes", "abstract": "  We propose an end-to-end solution to address the problem of object\nlocalisation in partial scenes, where we aim to estimate the position of an\nobject in an unknown area given only a partial 3D scan of the scene. We propose\na novel scene representation to facilitate the geometric reasoning, Directed\nSpatial Commonsense Graph (D-SCG), a spatial scene graph that is enriched with\nadditional concept nodes from a commonsense knowledge base. Specifically, the\nnodes of D-SCG represent the scene objects and the edges are their relative\npositions. Each object node is then connected via different commonsense\nrelationships to a set of concept nodes. With the proposed graph-based scene\nrepresentation, we estimate the unknown position of the target object using a\nGraph Neural Network that implements a novel attentional message passing\nmechanism. The network first predicts the relative positions between the target\nobject and each visible object by learning a rich representation of the objects\nvia aggregating both the object nodes and the concept nodes in D-SCG. These\nrelative positions then are merged to obtain the final position. We evaluate\nour method using Partial ScanNet, improving the state-of-the-art by 5.9% in\nterms of the localisation accuracy at a 8x faster training speed.\n", "date": "2022-11-01", "categories": []}
{"id": "2210.12935", "title": "360-MLC: Multi-view Layout Consistency for Self-training and\n  Hyper-parameter Tuning", "abstract": "  We present 360-MLC, a self-training method based on multi-view layout\nconsistency for finetuning monocular room-layout models using unlabeled\n360-images only. This can be valuable in practical scenarios where a\npre-trained model needs to be adapted to a new data domain without using any\nground truth annotations. Our simple yet effective assumption is that multiple\nlayout estimations in the same scene must define a consistent geometry\nregardless of their camera positions. Based on this idea, we leverage a\npre-trained model to project estimated layout boundaries from several camera\nviews into the 3D world coordinate. Then, we re-project them back to the\nspherical coordinate and build a probability function, from which we sample the\npseudo-labels for self-training. To handle unconfident pseudo-labels, we\nevaluate the variance in the re-projected boundaries as an uncertainty value to\nweight each pseudo-label in our loss function during training. In addition,\nsince ground truth annotations are not available during training nor in\ntesting, we leverage the entropy information in multiple layout estimations as\na quantitative metric to measure the geometry consistency of the scene,\nallowing us to evaluate any layout estimator for hyper-parameter tuning,\nincluding model selection without ground truth annotations. Experimental\nresults show that our solution achieves favorable performance against\nstate-of-the-art methods when self-training from three publicly available\nsource datasets to a unique, newly labeled dataset consisting of multi-view of\nthe same scenes.\n", "date": "2022-10-24", "categories": []}
{"id": "2210.11472", "title": "VIBUS: Data-efficient 3D Scene Parsing with VIewpoint Bottleneck and\n  Uncertainty-Spectrum Modeling", "abstract": "  Recently, 3D scenes parsing with deep learning approaches has been a heating\ntopic. However, current methods with fully-supervised models require manually\nannotated point-wise supervision which is extremely user-unfriendly and\ntime-consuming to obtain. As such, training 3D scene parsing models with sparse\nsupervision is an intriguing alternative. We term this task as data-efficient\n3D scene parsing and propose an effective two-stage framework named VIBUS to\nresolve it by exploiting the enormous unlabeled points. In the first stage, we\nperform self-supervised representation learning on unlabeled points with the\nproposed Viewpoint Bottleneck loss function. The loss function is derived from\nan information bottleneck objective imposed on scenes under different\nviewpoints, making the process of representation learning free of degradation\nand sampling. In the second stage, pseudo labels are harvested from the sparse\nlabels based on uncertainty-spectrum modeling. By combining data-driven\nuncertainty measures and 3D mesh spectrum measures (derived from normal\ndirections and geodesic distances), a robust local affinity metric is obtained.\nFinite gamma/beta mixture models are used to decompose category-wise\ndistributions of these measures, leading to automatic selection of thresholds.\nWe evaluate VIBUS on the public benchmark ScanNet and achieve state-of-the-art\nresults on both validation set and online test server. Ablation studies show\nthat both Viewpoint Bottleneck and uncertainty-spectrum modeling bring\nsignificant improvements. Codes and models are publicly available at\nhttps://github.com/AIR-DISCOVER/VIBUS.\n", "date": "2022-10-20", "categories": []}
{"id": "2207.06205", "title": "Trans4Map: Revisiting Holistic Bird's-Eye-View Mapping from Egocentric\n  Images to Allocentric Semantics with Vision Transformers", "abstract": "  Humans have an innate ability to sense their surroundings, as they can\nextract the spatial representation from the egocentric perception and form an\nallocentric semantic map via spatial transformation and memory updating.\nHowever, endowing mobile agents with such a spatial sensing ability is still a\nchallenge, due to two difficulties: (1) the previous convolutional models are\nlimited by the local receptive field, thus, struggling to capture holistic\nlong-range dependencies during observation; (2) the excessive computational\nbudgets required for success, often lead to a separation of the mapping\npipeline into stages, resulting the entire mapping process inefficient. To\naddress these issues, we propose an end-to-end one-stage Transformer-based\nframework for Mapping, termed Trans4Map. Our egocentric-to-allocentric mapping\nprocess includes three steps: (1) the efficient transformer extracts the\ncontextual features from a batch of egocentric images; (2) the proposed\nBidirectional Allocentric Memory (BAM) module projects egocentric features into\nthe allocentric memory; (3) the map decoder parses the accumulated memory and\npredicts the top-down semantic segmentation map. In contrast, Trans4Map\nachieves state-of-the-art results, reducing 67.2% parameters, yet gaining a\n+3.25% mIoU and a +4.09% mBF1 improvements on the Matterport3D dataset. Code\nat: https://github.com/jamycheung/Trans4Map.\n", "date": "2022-07-13", "categories": []}
{"id": "2210.07991", "title": "Novel 3D Scene Understanding Applications From Recurrence in a Single\n  Image", "abstract": "  We demonstrate the utility of recurring pattern discovery from a single image\nfor spatial understanding of a 3D scene in terms of (1) vanishing point\ndetection, (2) hypothesizing 3D translation symmetry and (3) counting the\nnumber of RP instances in the image.\n  Furthermore, we illustrate the feasibility of leveraging RP discovery output\nto form a more precise, quantitative text description of the scene. Our\nquantitative evaluations on a new 1K+ Recurring Pattern (RP) benchmark with\ndiverse variations show that visual perception of recurrence from one single\nview leads to scene understanding outcomes that are as good as or better than\nexisting supervised methods and/or unsupervised methods that use millions of\nimages.\n", "date": "2022-10-14", "categories": []}
{"id": "2206.06922", "title": "Object Scene Representation Transformer", "abstract": "  A compositional understanding of the world in terms of objects and their\ngeometry in 3D space is considered a cornerstone of human cognition.\nFacilitating the learning of such a representation in neural networks holds\npromise for substantially improving labeled data efficiency. As a key step in\nthis direction, we make progress on the problem of learning 3D-consistent\ndecompositions of complex scenes into individual objects in an unsupervised\nfashion. We introduce Object Scene Representation Transformer (OSRT), a\n3D-centric model in which individual object representations naturally emerge\nthrough novel view synthesis. OSRT scales to significantly more complex scenes\nwith larger diversity of objects and backgrounds than existing methods. At the\nsame time, it is multiple orders of magnitude faster at compositional rendering\nthanks to its light field parametrization and the novel Slot Mixer decoder. We\nbelieve this work will not only accelerate future architecture exploration and\nscaling efforts, but it will also serve as a useful tool for both\nobject-centric as well as neural scene representation learning communities.\n", "date": "2022-06-14", "categories": []}
{"id": "2209.08776", "title": "NeRF-SOS: Any-View Self-supervised Object Segmentation on Complex Scenes", "abstract": "  Neural volumetric representations have shown the potential that Multi-layer\nPerceptrons (MLPs) can be optimized with multi-view calibrated images to\nrepresent scene geometry and appearance, without explicit 3D supervision.\nObject segmentation can enrich many downstream applications based on the\nlearned radiance field. However, introducing hand-crafted segmentation to\ndefine regions of interest in a complex real-world scene is non-trivial and\nexpensive as it acquires per view annotation. This paper carries out the\nexploration of self-supervised learning for object segmentation using NeRF for\ncomplex real-world scenes. Our framework, called NeRF with Self-supervised\nObject Segmentation NeRF-SOS, couples object segmentation and neural radiance\nfield to segment objects in any view within a scene. By proposing a novel\ncollaborative contrastive loss in both appearance and geometry levels, NeRF-SOS\nencourages NeRF models to distill compact geometry-aware segmentation clusters\nfrom their density fields and the self-supervised pre-trained 2D visual\nfeatures. The self-supervised object segmentation framework can be applied to\nvarious NeRF models that both lead to photo-realistic rendering results and\nconvincing segmentation maps for both indoor and outdoor scenarios. Extensive\nresults on the LLFF, Tank & Temple, and BlendedMVS datasets validate the\neffectiveness of NeRF-SOS. It consistently surpasses other 2D-based\nself-supervised baselines and predicts finer semantics masks than existing\nsupervised counterparts. Please refer to the video on our project page for more\ndetails:https://zhiwenfan.github.io/NeRF-SOS.\n", "date": "2022-09-19", "categories": []}
{"id": "2210.05891", "title": "Point Cloud Scene Completion with Joint Color and Semantic Estimation\n  from Single RGB-D Image", "abstract": "  We present a deep reinforcement learning method of progressive view\ninpainting for colored semantic point cloud scene completion under volume\nguidance, achieving high-quality scene reconstruction from only a single RGB-D\nimage with severe occlusion. Our approach is end-to-end, consisting of three\nmodules: 3D scene volume reconstruction, 2D RGB-D and segmentation image\ninpainting, and multi-view selection for completion. Given a single RGB-D\nimage, our method first predicts its semantic segmentation map and goes through\nthe 3D volume branch to obtain a volumetric scene reconstruction as a guide to\nthe next view inpainting step, which attempts to make up the missing\ninformation; the third step involves projecting the volume under the same view\nof the input, concatenating them to complete the current view RGB-D and\nsegmentation map, and integrating all RGB-D and segmentation maps into the\npoint cloud. Since the occluded areas are unavailable, we resort to a A3C\nnetwork to glance around and pick the next best view for large hole completion\nprogressively until a scene is adequately reconstructed while guaranteeing\nvalidity. All steps are learned jointly to achieve robust and consistent\nresults. We perform qualitative and quantitative evaluations with extensive\nexperiments on the 3D-FUTURE data, obtaining better results than\nstate-of-the-arts.\n", "date": "2022-10-12", "categories": []}
{"id": "2210.03815", "title": "Scene-level Tracking and Reconstruction without Object Priors", "abstract": "  We present the first real-time system capable of tracking and reconstructing,\nindividually, every visible object in a given scene, without any form of prior\non the rigidness of the objects, texture existence, or object category. In\ncontrast with previous methods such as Co-Fusion and MaskFusion that first\nsegment the scene into individual objects and then process each object\nindependently, the proposed method dynamically segments the non-rigid scene as\npart of the tracking and reconstruction process. When new measurements indicate\ntopology change, reconstructed models are updated in real-time to reflect that\nchange. Our proposed system can provide the live geometry and deformation of\nall visible objects in a novel scene in real-time, which makes it possible to\nbe integrated seamlessly into numerous existing robotics applications that rely\non object models for grasping and manipulation. The capabilities of the\nproposed system are demonstrated in challenging scenes that contain multiple\nrigid and non-rigid objects.\n", "date": "2022-10-07", "categories": []}
{"id": "2209.14858", "title": "4D-StOP: Panoptic Segmentation of 4D LiDAR using Spatio-temporal Object\n  Proposal Generation and Aggregation", "abstract": "  In this work, we present a new paradigm, called 4D-StOP, to tackle the task\nof 4D Panoptic LiDAR Segmentation. 4D-StOP first generates spatio-temporal\nproposals using voting-based center predictions, where each point in the 4D\nvolume votes for a corresponding center. These tracklet proposals are further\naggregated using learned geometric features. The tracklet aggregation method\neffectively generates a video-level 4D scene representation over the entire\nspace-time volume. This is in contrast to existing end-to-end trainable\nstate-of-the-art approaches which use spatio-temporal embeddings that are\nrepresented by Gaussian probability distributions. Our voting-based tracklet\ngeneration method followed by geometric feature-based aggregation generates\nsignificantly improved panoptic LiDAR segmentation quality when compared to\nmodeling the entire 4D volume using Gaussian probability distributions. 4D-StOP\nachieves a new state-of-the-art when applied to the SemanticKITTI test dataset\nwith a score of 63.9 LSTQ, which is a large (+7%) improvement compared to\ncurrent best-performing end-to-end trainable methods. The code and pre-trained\nmodels are available at: https://github.com/LarsKreuzberg/4D-StOP.\n", "date": "2022-09-29", "categories": []}
{"id": "2210.03043", "title": "Feature-Realistic Neural Fusion for Real-Time, Open Set Scene\n  Understanding", "abstract": "  General scene understanding for robotics requires flexible semantic\nrepresentation, so that novel objects and structures which may not have been\nknown at training time can be identified, segmented and grouped. We present an\nalgorithm which fuses general learned features from a standard pre-trained\nnetwork into a highly efficient 3D geometric neural field representation during\nreal-time SLAM. The fused 3D feature maps inherit the coherence of the neural\nfield's geometry representation. This means that tiny amounts of human\nlabelling interacting at runtime enable objects or even parts of objects to be\nrobustly and accurately segmented in an open set manner.\n", "date": "2022-10-06", "categories": []}
{"id": "2209.13156", "title": "Towards Multimodal Multitask Scene Understanding Models for Indoor\n  Mobile Agents", "abstract": "  The perception system in personalized mobile agents requires developing\nindoor scene understanding models, which can understand 3D geometries, capture\nobjectiveness, analyze human behaviors, etc. Nonetheless, this direction has\nnot been well-explored in comparison with models for outdoor environments\n(e.g., the autonomous driving system that includes pedestrian prediction, car\ndetection, traffic sign recognition, etc.). In this paper, we first discuss the\nmain challenge: insufficient, or even no, labeled data for real-world indoor\nenvironments, and other challenges such as fusion between heterogeneous sources\nof information (e.g., RGB images and Lidar point clouds), modeling\nrelationships between a diverse set of outputs (e.g., 3D object locations,\ndepth estimation, and human poses), and computational efficiency. Then, we\ndescribe MMISM (Multi-modality input Multi-task output Indoor Scene\nunderstanding Model) to tackle the above challenges. MMISM considers RGB images\nas well as sparse Lidar points as inputs and 3D object detection, depth\ncompletion, human pose estimation, and semantic segmentation as output tasks.\nWe show that MMISM performs on par or even better than single-task models;\ne.g., we improve the baseline 3D object detection results by 11.7% on the\nbenchmark ARKitScenes dataset.\n", "date": "2022-09-27", "categories": []}
{"id": "2209.06641", "title": "CMR3D: Contextualized Multi-Stage Refinement for 3D Object Detection", "abstract": "  Existing deep learning-based 3D object detectors typically rely on the\nappearance of individual objects and do not explicitly pay attention to the\nrich contextual information of the scene. In this work, we propose\nContextualized Multi-Stage Refinement for 3D Object Detection (CMR3D)\nframework, which takes a 3D scene as input and strives to explicitly integrate\nuseful contextual information of the scene at multiple levels to predict a set\nof object bounding-boxes along with their corresponding semantic labels. To\nthis end, we propose to utilize a context enhancement network that captures the\ncontextual information at different levels of granularity followed by a\nmulti-stage refinement module to progressively refine the box positions and\nclass predictions. Extensive experiments on the large-scale ScanNetV2 benchmark\nreveal the benefits of our proposed method, leading to an absolute improvement\nof 2.0% over the baseline. In addition to 3D object detection, we investigate\nthe effectiveness of our CMR3D framework for the problem of 3D object counting.\nOur source code will be publicly released.\n", "date": "2022-09-13", "categories": []}
{"id": "2209.03494", "title": "Neural Feature Fusion Fields: 3D Distillation of Self-Supervised 2D\n  Image Representations", "abstract": "  We present Neural Feature Fusion Fields (N3F), a method that improves dense\n2D image feature extractors when the latter are applied to the analysis of\nmultiple images reconstructible as a 3D scene. Given an image feature\nextractor, for example pre-trained using self-supervision, N3F uses it as a\nteacher to learn a student network defined in 3D space. The 3D student network\nis similar to a neural radiance field that distills said features and can be\ntrained with the usual differentiable rendering machinery. As a consequence,\nN3F is readily applicable to most neural rendering formulations, including\nvanilla NeRF and its extensions to complex dynamic scenes. We show that our\nmethod not only enables semantic understanding in the context of scene-specific\nneural fields without the use of manual labels, but also consistently improves\nover the self-supervised 2D baselines. This is demonstrated by considering\nvarious tasks, such as 2D object retrieval, 3D segmentation, and scene editing,\nin diverse sequences, including long egocentric videos in the EPIC-KITCHENS\nbenchmark.\n", "date": "2022-09-07", "categories": []}
{"id": "2208.07870", "title": "Language-guided Semantic Style Transfer of 3D Indoor Scenes", "abstract": "  We address the new problem of language-guided semantic style transfer of 3D\nindoor scenes. The input is a 3D indoor scene mesh and several phrases that\ndescribe the target scene. Firstly, 3D vertex coordinates are mapped to RGB\nresidues by a multi-layer perceptron. Secondly, colored 3D meshes are\ndifferentiablly rendered into 2D images, via a viewpoint sampling strategy\ntailored for indoor scenes. Thirdly, rendered 2D images are compared to\nphrases, via pre-trained vision-language models. Lastly, errors are\nback-propagated to the multi-layer perceptron to update vertex colors\ncorresponding to certain semantic categories. We did large-scale qualitative\nanalyses and A/B user tests, with the public ScanNet and SceneNN datasets. We\ndemonstrate: (1) visually pleasing results that are potentially useful for\nmultimedia applications. (2) rendering 3D indoor scenes from viewpoints\nconsistent with human priors is important. (3) incorporating semantics\nsignificantly improve style transfer quality. (4) an HSV regularization term\nleads to results that are more consistent with inputs and generally rated\nbetter. Codes and user study toolbox are available at\nhttps://github.com/AIR-DISCOVER/LASST\n", "date": "2022-08-16", "categories": []}
{"id": "2207.09686", "title": "Object-Compositional Neural Implicit Surfaces", "abstract": "  The neural implicit representation has shown its effectiveness in novel view\nsynthesis and high-quality 3D reconstruction from multi-view images. However,\nmost approaches focus on holistic scene representation yet ignore individual\nobjects inside it, thus limiting potential downstream applications. In order to\nlearn object-compositional representation, a few works incorporate the 2D\nsemantic map as a cue in training to grasp the difference between objects. But\nthey neglect the strong connections between object geometry and instance\nsemantic information, which leads to inaccurate modeling of individual\ninstance. This paper proposes a novel framework, ObjectSDF, to build an\nobject-compositional neural implicit representation with high fidelity in 3D\nreconstruction and object representation. Observing the ambiguity of\nconventional volume rendering pipelines, we model the scene by combining the\nSigned Distance Functions (SDF) of individual object to exert explicit surface\nconstraint. The key in distinguishing different instances is to revisit the\nstrong association between an individual object's SDF and semantic label.\nParticularly, we convert the semantic information to a function of object SDF\nand develop a unified and compact representation for scene and objects.\nExperimental results show the superiority of ObjectSDF framework in\nrepresenting both the holistic object-compositional scene and the individual\ninstances. Code can be found at https://qianyiwu.github.io/objectsdf/\n", "date": "2022-07-20", "categories": []}
{"id": "2207.14268", "title": "MonteBoxFinder: Detecting and Filtering Primitives to Fit a Noisy Point\n  Cloud", "abstract": "  We present MonteBoxFinder, a method that, given a noisy input point cloud,\nfits cuboids to the input scene. Our primary contribution is a discrete\noptimization algorithm that, from a dense set of initially detected cuboids, is\nable to efficiently filter good boxes from the noisy ones. Inspired by recent\napplications of MCTS to scene understanding problems, we develop a stochastic\nalgorithm that is, by design, more efficient for our task. Indeed, the quality\nof a fit for a cuboid arrangement is invariant to the order in which the\ncuboids are added into the scene. We develop several search baselines for our\nproblem and demonstrate, on the ScanNet dataset, that our approach is more\nefficient and precise. Finally, we strongly believe that our core algorithm is\nvery general and that it could be extended to many other problems in 3D scene\nunderstanding.\n", "date": "2022-07-28", "categories": []}
{"id": "2207.14287", "title": "Depth Field Networks for Generalizable Multi-view Scene Representation", "abstract": "  Modern 3D computer vision leverages learning to boost geometric reasoning,\nmapping image data to classical structures such as cost volumes or epipolar\nconstraints to improve matching. These architectures are specialized according\nto the particular problem, and thus require significant task-specific tuning,\noften leading to poor domain generalization performance. Recently, generalist\nTransformer architectures have achieved impressive results in tasks such as\noptical flow and depth estimation by encoding geometric priors as inputs rather\nthan as enforced constraints. In this paper, we extend this idea and propose to\nlearn an implicit, multi-view consistent scene representation, introducing a\nseries of 3D data augmentation techniques as a geometric inductive prior to\nincrease view diversity. We also show that introducing view synthesis as an\nauxiliary task further improves depth estimation. Our Depth Field Networks\n(DeFiNe) achieve state-of-the-art results in stereo and video depth estimation\nwithout explicit geometric constraints, and improve on zero-shot domain\ngeneralization by a wide margin.\n", "date": "2022-07-28", "categories": []}
{"id": "2207.12394", "title": "Dynamic 3D Scene Analysis by Point Cloud Accumulation", "abstract": "  Multi-beam LiDAR sensors, as used on autonomous vehicles and mobile robots,\nacquire sequences of 3D range scans (\"frames\"). Each frame covers the scene\nsparsely, due to limited angular scanning resolution and occlusion. The\nsparsity restricts the performance of downstream processes like semantic\nsegmentation or surface reconstruction. Luckily, when the sensor moves, frames\nare captured from a sequence of different viewpoints. This provides\ncomplementary information and, when accumulated in a common scene coordinate\nframe, yields a denser sampling and a more complete coverage of the underlying\n3D scene. However, often the scanned scenes contain moving objects. Points on\nthose objects are not correctly aligned by just undoing the scanner's\nego-motion. In the present paper, we explore multi-frame point cloud\naccumulation as a mid-level representation of 3D scan sequences, and develop a\nmethod that exploits inductive biases of outdoor street scenes, including their\ngeometric layout and object-level rigidity. Compared to state-of-the-art scene\nflow estimators, our proposed approach aims to align all 3D points in a common\nreference frame correctly accumulating the points on the individual objects.\nOur approach greatly reduces the alignment errors on several benchmark\ndatasets. Moreover, the accumulated point clouds benefit high-level tasks like\nsurface reconstruction.\n", "date": "2022-07-25", "categories": []}
{"id": "2207.09291", "title": "3D Room Layout Estimation from a Cubemap of Panorama Image via Deep\n  Manhattan Hough Transform", "abstract": "  Significant geometric structures can be compactly described by global\nwireframes in the estimation of 3D room layout from a single panoramic image.\nBased on this observation, we present an alternative approach to estimate the\nwalls in 3D space by modeling long-range geometric patterns in a learnable\nHough Transform block. We transform the image feature from a cubemap tile to\nthe Hough space of a Manhattan world and directly map the feature to the\ngeometric output. The convolutional layers not only learn the local\ngradient-like line features, but also utilize the global information to\nsuccessfully predict occluded walls with a simple network structure. Unlike\nmost previous work, the predictions are performed individually on each cubemap\ntile, and then assembled to get the layout estimation. Experimental results\nshow that we achieve comparable results with recent state-of-the-art in\nprediction accuracy and performance. Code is available at\nhttps://github.com/Starrah/DMH-Net.\n", "date": "2022-07-19", "categories": []}
{"id": "2207.03204", "title": "MCTS with Refinement for Proposals Selection Games in Scene\n  Understanding", "abstract": "  We propose a novel method applicable in many scene understanding problems\nthat adapts the Monte Carlo Tree Search (MCTS) algorithm, originally designed\nto learn to play games of high-state complexity. From a generated pool of\nproposals, our method jointly selects and optimizes proposals that minimize the\nobjective term. In our first application for floor plan reconstruction from\npoint clouds, our method selects and refines the room proposals, modelled as 2D\npolygons, by optimizing on an objective function combining the fitness as\npredicted by a deep network and regularizing terms on the room shapes. We also\nintroduce a novel differentiable method for rendering the polygonal shapes of\nthese proposals. Our evaluations on the recent and challenging Structured3D and\nFloor-SP datasets show significant improvements over the state-of-the-art,\nwithout imposing hard constraints nor assumptions on the floor plan\nconfigurations. In our second application, we extend our approach to\nreconstruct general 3D room layouts from a color image and obtain accurate room\nlayouts. We also show that our differentiable renderer can easily be extended\nfor rendering 3D planar polygons and polygon embeddings. Our method shows high\nperformance on the Matterport3D-Layout dataset, without introducing hard\nconstraints on room layout configurations.\n", "date": "2022-07-07", "categories": []}
{"id": "2206.04585", "title": "Extracting Zero-shot Common Sense from Large Language Models for Robot\n  3D Scene Understanding", "abstract": "  Semantic 3D scene understanding is a problem of critical importance in\nrobotics. While significant advances have been made in simultaneous\nlocalization and mapping algorithms, robots are still far from having the\ncommon sense knowledge about household objects and their locations of an\naverage human. We introduce a novel method for leveraging common sense embedded\nwithin large language models for labelling rooms given the objects contained\nwithin. This algorithm has the added benefits of (i) requiring no task-specific\npre-training (operating entirely in the zero-shot regime) and (ii) generalizing\nto arbitrary room and object labels, including previously-unseen ones -- both\nof which are highly desirable traits in robotic scene understanding algorithms.\nThe proposed algorithm operates on 3D scene graphs produced by modern spatial\nperception systems, and we hope it will pave the way to more generalizable and\nscalable high-level 3D scene understanding for robotics.\n", "date": "2022-06-09", "categories": []}
{"id": "2206.07028", "title": "Learning 3D Object Shape and Layout without 3D Supervision", "abstract": "  A 3D scene consists of a set of objects, each with a shape and a layout\ngiving their position in space. Understanding 3D scenes from 2D images is an\nimportant goal, with applications in robotics and graphics. While there have\nbeen recent advances in predicting 3D shape and layout from a single image,\nmost approaches rely on 3D ground truth for training which is expensive to\ncollect at scale. We overcome these limitations and propose a method that\nlearns to predict 3D shape and layout for objects without any ground truth\nshape or layout information: instead we rely on multi-view images with 2D\nsupervision which can more easily be collected at scale. Through extensive\nexperiments on 3D Warehouse, Hypersim, and ScanNet we demonstrate that our\napproach scales to large datasets of realistic images, and compares favorably\nto methods relying on 3D ground truth. On Hypersim and ScanNet where reliable\n3D ground truth is not available, our approach outperforms supervised\napproaches trained on smaller and less diverse datasets.\n", "date": "2022-06-14", "categories": []}
{"id": "2206.02366", "title": "Scan2Part: Fine-grained and Hierarchical Part-level Understanding of\n  Real-World 3D Scans", "abstract": "  We propose Scan2Part, a method to segment individual parts of objects in\nreal-world, noisy indoor RGB-D scans. To this end, we vary the part hierarchies\nof objects in indoor scenes and explore their effect on scene understanding\nmodels. Specifically, we use a sparse U-Net-based architecture that captures\nthe fine-scale detail of the underlying 3D scan geometry by leveraging a\nmulti-scale feature hierarchy. In order to train our method, we introduce the\nScan2Part dataset, which is the first large-scale collection providing detailed\nsemantic labels at the part level in the real-world setting. In total, we\nprovide 242,081 correspondences between 53,618 PartNet parts of 2,477 ShapeNet\nobjects and 1,506 ScanNet scenes, at two spatial resolutions of 2 cm$^3$ and 5\ncm$^3$. As output, we are able to predict fine-grained per-object part labels,\neven when the geometry is coarse or partially missing.\n", "date": "2022-06-06", "categories": []}
{"id": "2606.21292", "title": "Lightweight 3D Feature Pretraining by Bayesian Inversion of 2D Foundation Models", "abstract": "We present Casper3D, a lightweight probabilistic framework for converting noisy multi-view 2D foundation-model embeddings into a latent 3D semantic representation. We model view-level semantic features as noisy observations of an underlying 3D semantic state and infer this state with a set-based variational model that incorporates relative pose during multi-view reasoning. Casper3D is trained by predicting held-out semantic observations from novel viewpoints, while remaining aligned with visual and text semantic spaces for open-vocabulary 3D understanding. The framework is backbone-agnostic and applies to both language-aligned and self-supervised embeddings. Experiments show that Casper3D produces more stable 3D semantics than simple multi-view pooling, especially in ambiguous and noisy settings.", "date": "2026-06-19", "categories": []}
{"id": "2606.21373", "title": "FLM-Occ: Feed-forward Likelihood Maximization for Efficient Indoor Occupancy Prediction", "abstract": "Recent indoor occupancy prediction methods adopt Gaussian primitives as a sparse 3D representation for computational efficiency. However, their training relies on voxel classification, which imposes only local constraints and lacks global supervision on the distribution of the primitives. Therefore, they inevitably predict spurious primitives in empty regions, undermining both representational and computational efficiency. To address this, we propose Feed-forward Likelihood Maximization (FLM), a novel framework that reformulates occupancy prediction as voxel distribution estimation. In FLM, a network is trained to predict a mixture model that maximizes the likelihood over ground-truth occupied voxels in a feed-forward manner. To enable end-to-end training of networks and voxelization of a standard mixture model, we define mixture weights as normalized primitive volumes to implicitly enforce simplex constraints and derive novel voxelization formulas. Based on FLM, our FLM-Occ, a novel method that is capable of relocating randomly initialized primitives over long distances to model a scene. On Occ-ScanNet, FLM-Occ achieves superior accuracy using only 32 superquadrics, 2.7% of the prior SoTA, while running 3.7 times faster.", "date": "2026-06-19", "categories": []}
{"id": "2606.21596", "title": "$\\phi$-Scene: Physically Grounded Image-to-3D Scene Reconstruction", "abstract": "Reconstructing compositional 3D scenes from a single image is a fundamental challenge in 3D world modeling. Recent methods can recover high-fidelity, complete 3D objects and predict plausible scene arrangements, but most still treat scene reconstruction primarily as a visual and geometric prediction problem. Their outputs may therefore contain floating objects, interpenetrations, or unstable-contact artifacts, limiting their physical validity and downstream usability in simulation, robotics, and interactive environments. We present $\\phi$-Scene, a physically grounded approach to open-vocabulary and compositional image-to-3D scene reconstruction. The key premise is that a reconstructed scene should not be treated merely as a set of objects with predicted poses, but as a stable physical system. Accordingly, $\\phi$-Scene formulates reconstruction as topology-driven physical assembly: it infers how objects support one another, orders them accordingly, and progressively settles each object against its already stabilized support context. For each object in topological order, SDF-based optimization first resolves penetrations against the pre-settled support context, and rigid-body simulation then settles the object into a stable contact configuration under real-world physical constraints. Experiments on 3D-Front show that $\\phi$-Scene achieves the strongest overall performance among out-of-domain methods and remains highly competitive with in-domain baselines. Human and VLM evaluations further show strong preference for $\\phi$-Scene in visual quality, reference alignment, and physical plausibility. Finally, dedicated physical plausibility metrics covering static contact quality and dynamic stability demonstrate that $\\phi$-Scene substantially reduces penetration artifacts while producing much lower post-simulation drift, indicating more stable and physically grounded 3D scenes.", "date": "2026-06-19", "categories": []}
{"id": "2606.22971", "title": "Humanoid-OmniOcc: Stereo-Based Full-View Occupancy Dataset for Embodied AI", "abstract": "Occupancy prediction at voxel-level granularity is essential for safe robotic navigation and interaction in complex environments. Existing occupancy datasets, however, are predominantly designed for autonomous driving with vehicle-centric biases -- forward-facing cameras, far-field geometry, and static road priors -- limiting their applicability to embodied humanoid perception. We present Humanoid-OmniOcc, a large-scale panoramic stereo-based occupancy dataset tailored for humanoid robots. The dataset encompasses 15 diverse simulated indoor scenes and 5 real-world environments, yielding over 155K samples with broad scene and style diversity. Importantly, the dataset is designed around a Real2Sim2Real closed-loop paradigm: real sensor specifications drive physically accurate simulation, simulation produces large-scale annotated training data, and models trained in simulation are directly evaluated on real-world captures -- enabling iterative refinement of the sim-to-real pipeline. We further propose \\textbf{H}umanoid \\textbf{S}urround \\textbf{S}tereo-guided \\textbf{Occ}upancy model (Humanoid-OmniOcc) that exploits robust depth priors for accurate 2D-to-3D lifting. Extensive experiments show that Humanoid-OmniOcc consistently outperforms monocular baselines and generalizes well to both unseen simulated test scenes and real-world environments, validating the effectiveness of the Real2Sim2Real design. Code and data will be available upon acceptance at https://d-robotics-ai-lab.github.io/humanoid-omniocc.", "date": "2026-06-22", "categories": []}
{"id": "2606.24180", "title": "Deep Learning Approaches for 3D Medical Scene Completion: From Geometric Modeling to Generative Paradigms", "abstract": "Three-dimensional scene completion has evolved as a major problem in computer vision and robotics, and its applications are diverse, including autonomous navigation and augmented reality. In this study, a systematic review has been conducted to compile the research contributions made in the last ten years, i.e., 2016 to 2026, which has revolutionized the field from the voxel semantic completion paradigm represented by SSCNet to the latest paradigm that combines generative diffusion priors with real-time rendering using a Gaussian splatting technique. The evolution in representation paradigms, such as voxel grids, point learning, implicit neural fields, transformer networks, diffusion networks, and the latest paradigm based on rendering-aware 3D Gaussian primitives, has been discussed in this study. A comprehensive analysis has been carried out on the contributions made in the last ten years, and a taxonomy has been developed to provide a clear idea about the contributions made in the field. The study has also discussed the research contributions made in the field, along with the challenges that still need to be addressed. Finally, the study has presented a research agenda that will provide a clear idea about the directions that can be followed in the development of the next-generation system", "date": "2026-06-23", "categories": []}
{"id": "2606.24649", "title": "Agentic Collaborative Cognition for Zero-Shot 3D Understanding", "abstract": "Recent advancements have explored agentic zero-shot 3D understanding by reformulating it as video keyframe understanding with Multimodal Large Language Models (MLLMs). However, existing methods face an intrinsic bottleneck due to the finite observation perspectives inherent in videos and the implicit perception of 3D scenes. In this paper, we propose a collaborative multi-agent framework that assigns a Planning Agent to handle high-level viewpoint planning and supplement novel perspectives, and a Perception Agent to explicitly summarize the 3D scene into a structured holistic cognitive map. Specifically, Planning Agent first analyzes this cognitive map to determine query-relevant viewpoints and supplements missing critical perspectives to ensure comprehensive observation. Subsequently, Perception Agent documents object-level attributes from these views by assigning consistent instance identifiers across viewpoints, thereby integrating fragmented observations into the holistic cognitive map. In parallel, it provides feedback to filter out mismatched candidate objects and guide subsequent viewpoint planning. Through this closed-loop iterative process, two agents collaboratively figure out candidates until Perception Agent determines that sufficient information has been captured to complete the task. Extensive experiments demonstrate that our method achieves state-of-the-art performance on 6 benchmarks, with improvements of 11.1\\% Acc@0.5 on ScanRefer, 14.6 BLEU-1 on 3D-assisted dialog, and 2.1 EM on SQA3D.", "date": "2026-06-23", "categories": []}
{"id": "2606.27412", "title": "Not All Relations Rotate Alike: Transformation-Aware Decoupling for Viewpoint-Robust 3D Scene Graph Generation", "abstract": "3D Scene Graph Generation (3DSGG) represents 3D scenes as structured object-relation-object graphs, providing a compact relational abstraction for spatial understanding. In embodied intelligence settings, the same 3D scene may be observed by agents from viewpoints that differ by yaw rotations. However, current 3DSGG models often fail to produce relation predictions that follow the expected transformation behavior under such viewpoint shifts. This behavior reveals an empirical mismatch related to predicate-level transformation heterogeneity: directional predicates such as left, front, right, and behind should transform with the observation frame, whereas most contact, support, and semantic predicates such as standing on and attached to should remain stable. To reduce this mismatch, we propose Transformation-Aware Decoupling (TAD), a viewpoint-robust 3DSGG framework that decouples relation reasoning according to predicate transformation behavior and is supported by viewpoint-stable object representations. TAD decomposes relation reasoning into two parts: one learns cues that should stay stable across viewpoints, while the other learns directional cues that should change with the observation frame. The two parts are merged for standard multi-label predicate prediction. Transformation-specific descriptors and group-aware auxiliary supervision encourage the two branches to capture complementary relation cues. Extensive experiments on 3DSSG show that TAD achieves state-of-the-art robustness under yaw viewpoint changes without training-time rotation augmentation, while maintaining competitive performance under the standard benchmark. The project page is available at https://tad-predicate.github.io/.", "date": "2026-06-25", "categories": []}
{"id": "2606.27575", "title": "Perceptual 3D Simulation With Physical World Modeling", "abstract": "Predicting how a scene will evolve after a desired 3D transformation from images is a central goal in vision, graphics, and robotics. Yet unlike ideal simulators with full access to 3D geometry and dynamics, real world systems must rely on perceptual inputs and local actions that are inherently partial and incomplete. In this work, we present P3Sim, a physical world modeling system that simulates future scene states under both partial observations and incomplete 3D transformation signals. P3Sim is composed of three interacting components: a learned physical world model, a geometric conditioning module, and a persistent scene memory. The world model interprets perception as probabilistic inference over multimodal scene variables, providing predictions of the distributions of any scene variable conditioned on any combination of others. The geometric conditioning module provides a partial 3D transform signal for conditioning the world model at inference time. The persistent scene memory integrates predictions over time, enabling online updates and consistency under uncertainty. By combining learned inference with explicit geometric structure, P3Sim balances data-driven flexibility with built-in inductive bias. This design yields a flexible perceptual simulator that generalizes across diverse 3D transformation tasks, such as novel view synthesis, object manipulation, and dynamic scene prediction, advancing toward general purpose 3D scene understanding and transformation.", "date": "2026-06-25", "categories": []}
{"id": "2606.27720", "title": "Scene and Human in One World: Reconstruction in a Feedforward Pass", "abstract": "Reconstructing humans in dynamic scenes from moving monocular cameras remains challenging due to scale ambiguity, human-scene misalignment, and occlusion interference. Rather than treating human mesh recovery and scene reconstruction as separate tasks, we believe that accurate human-scene reconstruction requires the two tasks to mutually inform each other: parametric human models offer semantic structure and metric-scale priors, while scene geometry provides spatial context for human localization and alignment. Built on this insight, we introduce SHOW, a mask-promptable human mesh recovery framework that couples feed-forward 3D scene reconstruction with Human Mesh Recovery in a unified metric space. SHOW injects human semantics and scale priors from parametric human models into normalized point-map prediction, enabling metric-scale scene reconstruction from inherently scale-ambiguous monocular input. In turn, the recovered scene geometry constrains human mesh estimation, encouraging spatially consistent human placement and improved human-scene alignment. To handle complex multi-person and cluttered scenes, SHOW further incorporates a promptable masking mechanism that enables flexible target-human selection while suppressing background distractions and occlusion interference. Through joint training, the model learns both human-aware geometric features and geometry-constrained human features, producing aligned metric-scale reconstructions from monocular human-centric videos. Extensive experiments demonstrate that SHOW improves metric-scale consistency, human-scene alignment, and reconstruction accuracy under challenging camera motion, occlusion, and cluttered backgrounds.", "date": "2026-06-26", "categories": []}
{"id": "2606.28060", "title": "ReScene: Structured Indoor Scene Reconstruction from Multi-View Captures", "abstract": "Constructing simulation-ready 3D scenes from multi-view captures is a key bottleneck for Embodied Artificial Intelligence, as downstream tasks require object-level structure, explicit inter-object relations, and physical plausibility. Existing approaches either rely on specialized capture hardware, suffer from single-view bias in object reconstruction, or yield layouts that are geometrically reasonable but physically inconsistent. We identify that the problem is not single-object reconstruction but cross-view relation fusion and physically plausible scene assembly. To address this challenge, we present ReScene, a framework that threads multi-view geometry throughout the pipeline as a unifying prior. Our method consists of two main components: HierView prioritizes reconstruction views based on semantic consistency and 3D coverage completeness, replacing the largest-mask heuristic that conflates image occupancy with object coverage; and Relation-Aware Assembly fuses multi-frame relation predictions from a vision-language model with geometric and room-shell priors into a confidence-weighted scene graph, enabling physically consistent scene assembly. ReScene sets a new state of the art across geometry, rendering, and perceptual quality on a set of ScanNet scenes, achieving a 17% reduction in Chamfer Distance and 26% in LPIPS over the strongest prior baseline, while running up to 10x faster than prior multi-view methods. Based on the reconstructed scenes, we also generate an embodied visual question answering dataset, on which fine-tuned Qwen-VL approaches the performance of strong closed-source models on several spatial reasoning tasks.", "date": "2026-06-26", "categories": []}
{"id": "2606.29376", "title": "SAD-GS: Learning Reliable 3D Semantic Gaussian Fields via Dynamic Geo-Semantic Anchoring", "abstract": "Open-vocabulary 3D semantic Gaussian field learning relies on multi-view 2D supervision, whose semantic targets and spatial assignments are often unreliable. Across varying viewpoints, view-dependent features cause semantic identity drift, while propagated tracker masks introduce boundary leakage and identity switches. Directly optimizing against these unreliable 2D targets forces the 3D representation to absorb multi-view contradictions, leading to severe error accumulation. To resolve this limitation, we propose SAD-GS, a framework for learning reliable 3D semantic Gaussian fields via dynamic geo-semantic anchoring. Specifically, Semantic Anchor Distillation (SAD) distills per-view visual embeddings into consensus text anchors to establish a viewpoint-invariant semantic identity. Concurrently, the Geo-Semantic Feedback Loop (GSFL) leverages the evolving 3D field to actively filter tracker anomalies and refine spatial mask assignments via a conservative three-gate update rule. Extensive evaluations on LERF-OVS, 3D-OVS, and Mip-NeRF360 show that SAD-GS consistently achieves the best overall performance in both open-vocabulary localization and semantic segmentation. These comprehensive improvements validate the effectiveness and robustness of dynamic geo-semantic anchoring for reliable 3D semantic Gaussian field learning.", "date": "2026-06-28", "categories": []}
{"id": "2606.29513", "title": "Scenes as Objects, Not Primitives: Instance-Structured 3D Tokenization from Unposed Views", "abstract": "A 3D scene is understood through its objects, not the primitives that compose them. Yet feed-forward reconstruction methods output dense, unstructured sets of points or Gaussians, leaving object-level structure to be recovered after the fact. We propose a feed-forward framework that decomposes a scene into instance-structured 3D token groups directly from unposed multi-view images -- compact object-centric units from which reconstruction, segmentation, and manipulation all follow. Each token group pairs an instance token capturing entity-level identity with anchor tokens that encode local geometry and appearance, which are decoded into a set of 3D Gaussians. This two-level factorization decouples object identity from local appearance, making object instances a native interface of the representation rather than a derived product. The token groups are learned through differentiable rendering with joint reconstruction and segmentation supervision, requiring no 3D annotations. Our feed-forward model surpasses per-scene optimization baselines in class-agnostic instance segmentation while remaining competitive in novel view synthesis. Beyond these metrics, the same token groups directly unlock instance-level scene editing -- removing, translating, or inserting objects by operating on their groups -- as well as efficient open-vocabulary 3D instance retrieval, where retrieval complexity scales with the number of instances rather than primitives.", "date": "2026-06-28", "categories": []}
{"id": "2606.29786", "title": "OP3DSG: Open-Vocabulary Part-Aware 3D Scene Graph Generation for Real-World Environments", "abstract": "3D scene graphs (3DSGs) provide a compact and structured abstraction of 3D environments. Although advances in foundation models have enabled open-vocabulary 3DSG generation, existing approaches remain object-centric and encode limited relational information -- restricting their applicability in real-world scenarios that require fine-grained understanding. We propose OP3DSG, an open-vocabulary part-aware 3DSG generation framework that constructs unified graphs that jointly model objects, interactive parts, spatial relations, functional relations, and affordances. OP3DSG integrates object-part knowledge-guided detection with part-aware 3D fusion to preserve small and interaction-relevant components, and employs a geometry-initialized prior graph with LLM-based refinement to reduce spurious relational predictions while enabling efficient graph construction. To systematically evaluate unified 3D scene graph construction, we introduce UniGraph3D, a benchmark designed for part-aware perception and multi-level relational reasoning. Experimental results show that OP3DSG achieves state-of-the-art performance and demonstrates its effectiveness as a perception backbone in diverse real-world robotics tasks.", "date": "2026-06-29", "categories": []}
{"id": "2606.30638", "title": "Open-Vocabulary and Referring Segmentation for 3D Gaussians Using 2D Detectors", "abstract": "3D Gaussian Splatting (3DGS) has emerged at the forefront of 3D scene reconstruction. Extending 3DGS with language-driven, open-vocabulary understanding has gained significant attention for real-world applications such as embodied AI. Recent methods achieve this by learning an instance feature attribute and assigning semantics by distilling high-dimensional Contrastive Language-Image Pretraining (CLIP) features directly into the scene representation. However, the instance grouping mechanisms of these methods either require a predefined number of instances or suffer from noise in their bottom-up grouping strategies. Furthermore, the reliance on CLIP restricts semantic understanding to simple noun phrases, preventing complex spatial reasoning and referential expression grounding. We present GaussDet, a method that circumvents the need for dense CLIP features by leveraging discrete, open-vocabulary 2D object detectors with referring expression capabilities. We learn instance features for individual Gaussians to decompose the scene into 3D instance groups. By rendering these groups and aggregating semantic votes from multi-view 2D detections, we generate a robust View-Aggregated Semantic Label Distribution (VASD) for each 3D instance. This view-aggregation strategy acts as a strong regularizer, attenuating spurious labels caused by low-quality instance grouping. Our approach enables a straightforward, zero-shot extension from simple language queries to complex referential grounding. Extensive evaluations across two key tasks -- open-vocabulary segmentation (LeRF-OVS, ScanNet) and referring expression grounding (Ref-LeRF) -- demonstrate that GaussDet achieves consistent improvements over existing methods. Most notably, we achieve a substantial 16.7% mIoU improvement in referential grounding within a strict zero-shot setting.", "date": "2026-06-29", "categories": []}
{"id": "2606.30754", "title": "Streaming Gaussian Encoding for 4D Panoptic Occupancy Tracking", "abstract": "Camera-based 4D panoptic occupancy tracking (4D-POT) is a promising paradigm for holistic scene understanding from multi-view imagery, enabling joint reasoning about geometry, semantics, and object identities across time. Recent mask-based pipelines achieve strong performance by propagating instance queries across frames. However, their underlying volumetric representations are typically recomputed at each timestep, limiting geometric temporal consistency, particularly under occlusion and for static scene elements. To address this limitation, we propose a streaming Gaussian encoder that maintains a persistent volumetric scene representation for 4D-POT. Our method models the scene as a fixed-size set of latent Gaussian queries that are propagated via ego-motion compensation and refreshed under a confidence-guided budget constraint. Crucially, we shape Gaussian opacities through depth-based supervision to serve as proxy for visibility, enabling confidence to accumulate as a temporally aggregated measure of persistent scene support. Together with a warmup-based multi-frame training strategy, this yields representation-level temporal coherence beyond decoder-only tracking. Extensive experiments on Occ3D-extended nuScenes and Waymo establish a new state-of-the-art for camera-based 4D-POT, improving tracking consistency with negligible computational overhead while remaining fully compatible with existing mask-based pipelines. We provide code and models at https://sge.cs.uni-freiburg.de.", "date": "2026-06-29", "categories": []}
{"id": "2606.31471", "title": "Think While You Map: Asynchronous Vision-Language Agents for Incremental 3D Scene Graphs", "abstract": "Open-vocabulary 3D scene graph methods typically operate in two stages: first reconstruct, then enrich with vision-language models, leaving the graph unqueryable during exploration. We argue that this sequential coupling is unnecessary and propose an asynchronous architecture in which lightweight online mapping runs concurrently with heavyweight semantic refinement. A probabilistic voxel-based backbone maintains stable object identities incrementally, while background VLM agents progressively enrich the graph. This framework resolves duplicate object tracks through semantic loop closure, attaches fine-grained visual attributes and derives spatial relations between objects. A multi-target frame scheduler amortizes VLM cost by selecting a small set of informative frames that jointly cover multiple targets. The resulting scene graph is queryable during exploration and grows in semantic richness over time. Our method matches or outperforms existing open-vocabulary 3D scene graph methods on semantic segmentation (ScanNet, Replica) and surpasses the prior state-of-the-art across three visual grounding benchmarks (Sr3D+, Nr3D, ScanRefer) by 15.3 to 18.8 A@0.25. Project page: https://denizbickici.github.io/thinkgraphs/", "date": "2026-06-30", "categories": []}