Kaichen Zhou

Kaichen Zhou, Lanqing Hong, Changhao Chen et al.

Self-supervised depth learning from monocular images normally relies on the 2D pixel-wise photometric relation between temporally adjacent image frames. However, they neither fully exploit the 3D point-wise geometric correspondences, nor effectively tackle the ambiguities in the photometric warping caused by occlusions or illumination inconsistency. To address these problems, this work proposes Density Volume Construction Network (DevNet), a novel self-supervised monocular depth learning framework, that can consider 3D spatial information, and exploit stronger geometric constraints among adjacent camera frustums. Instead of directly regressing the pixel value from a single image, our DevNet divides the camera frustum into multiple parallel planes and predicts the pointwise occlusion probability density on each plane. The final depth map is generated by integrating the density along corresponding rays. During the training process, novel regularization strategies and loss functions are introduced to mitigate photometric ambiguities and overfitting. Without obviously enlarging model parameters size or running time, DevNet outperforms several representative baselines on both the KITTI-2015 outdoor dataset and NYU-V2 indoor dataset. In particular, the root-mean-square-deviation is reduced by around 4% with DevNet on both KITTI-2015 and NYU-V2 in the task of depth estimation. Code is available at https://github.com/gitkaichenzhou/DevNet.

Jia-Xing Zhong, Kaichen Zhou, Qingyong Hu et al.

Scene flow is a powerful tool for capturing the motion field of 3D point clouds. However, it is difficult to directly apply flow-based models to dynamic point cloud classification since the unstructured points make it hard or even impossible to efficiently and effectively trace point-wise correspondences. To capture 3D motions without explicitly tracking correspondences, we propose a kinematics-inspired neural network (Kinet) by generalizing the kinematic concept of ST-surfaces to the feature space. By unrolling the normal solver of ST-surfaces in the feature space, Kinet implicitly encodes feature-level dynamics and gains advantages from the use of mature backbones for static point cloud processing. With only minor changes in network structures and low computing overhead, it is painless to jointly train and deploy our framework with a given static model. Experiments on NvGesture, SHREC'17, MSRAction-3D, and NTU-RGBD demonstrate its efficacy in performance, efficiency in both the number of parameters and computational complexity, as well as its versatility to various static backbones. Noticeably, Kinet achieves the accuracy of 93.27% on MSRAction-3D with only 3.20M parameters and 10.35G FLOPS.

Kaichen Zhou, Jia-Xing Zhong, Sangyun Shin et al.

The introduction of neural radiance fields has greatly improved the effectiveness of view synthesis for monocular videos. However, existing algorithms face difficulties when dealing with uncontrolled or lengthy scenarios, and require extensive training time specific to each new scenario. To tackle these limitations, we propose DynPoint, an algorithm designed to facilitate the rapid synthesis of novel views for unconstrained monocular videos. Rather than encoding the entirety of the scenario information into a latent representation, DynPoint concentrates on predicting the explicit 3D correspondence between neighboring frames to realize information aggregation. Specifically, this correspondence prediction is achieved through the estimation of consistent depth and scene flow information across frames. Subsequently, the acquired correspondence is utilized to aggregate information from multiple reference frames to a target frame, by constructing hierarchical neural point clouds. The resulting framework enables swift and accurate view synthesis for desired views of target frames. The experimental results obtained demonstrate the considerable acceleration of training time achieved - typically an order of magnitude - by our proposed method while yielding comparable outcomes compared to prior approaches. Furthermore, our method exhibits strong robustness in handling long-duration videos without learning a canonical representation of video content.

Ao Qu, Han Zheng, Zijian Zhou et al.

Large language model (LLM)-based evolution is a promising approach for open-ended discovery, where progress requires sustained search and knowledge accumulation. Existing methods still rely heavily on fixed heuristics and hard-coded exploration rules, which limit the autonomy of LLM agents. We present CORAL, the first framework for autonomous multi-agent evolution on open-ended problems. CORAL replaces rigid control with long-running agents that explore, reflect, and collaborate through shared persistent memory, asynchronous multi-agent execution, and heartbeat-based interventions. It also provides practical safeguards, including isolated workspaces, evaluator separation, resource management, and agent session and health management. Evaluated on diverse mathematical, algorithmic, and systems optimization tasks, CORAL sets new state-of-the-art results on 10 tasks, achieving 3-10 times higher improvement rates with far fewer evaluations than fixed evolutionary search baselines across tasks. On Anthropic's kernel engineering task, four co-evolving agents improve the best known score from 1363 to 1103 cycles. Mechanistic analyses further show how these gains arise from knowledge reuse and multi-agent exploration and communication. Together, these results suggest that greater agent autonomy and multi-agent evolution can substantially improve open-ended discovery. Code is available at https://github.com/Human-Agent-Society/CORAL.

Jia-Xing Zhong, Ta-Ying Cheng, Yuhang He et al.

A truly generalizable approach to rigid segmentation and motion estimation is fundamental to 3D understanding of articulated objects and moving scenes. In view of the closely intertwined relationship between segmentation and motion estimates, we present an SE(3) equivariant architecture and a training strategy to tackle this task in an unsupervised manner. Our architecture is composed of two interconnected, lightweight heads. These heads predict segmentation masks using point-level invariant features and estimate motion from SE(3) equivariant features, all without the need for category information. Our training strategy is unified and can be implemented online, which jointly optimizes the predicted segmentation and motion by leveraging the interrelationships among scene flow, segmentation mask, and rigid transformations. We conduct experiments on four datasets to demonstrate the superiority of our method. The results show that our method excels in both model performance and computational efficiency, with only 0.25M parameters and 0.92G FLOPs. To the best of our knowledge, this is the first work designed for category-agnostic part-level SE(3) equivariance in dynamic point clouds.

Sangyun Shin, Stuart Golodetz, Madhu Vankadari et al.

Deep learning has led to great progress in the detection of mobile (i.e. movement-capable) objects in urban driving scenes in recent years. Supervised approaches typically require the annotation of large training sets; there has thus been great interest in leveraging weakly, semi- or self-supervised methods to avoid this, with much success. Whilst weakly and semi-supervised methods require some annotation, self-supervised methods have used cues such as motion to relieve the need for annotation altogether. However, a complete absence of annotation typically degrades their performance, and ambiguities that arise during motion grouping can inhibit their ability to find accurate object boundaries. In this paper, we propose a new self-supervised mobile object detection approach called SCT. This uses both motion cues and expected object sizes to improve detection performance, and predicts a dense grid of 3D oriented bounding boxes to improve object discovery. We significantly outperform the state-of-the-art self-supervised mobile object detection method TCR on the KITTI tracking benchmark, and achieve performance that is within 30% of the fully supervised PV-RCNN++ method for IoUs <= 0.5.

Zhuoyang Liu, Jiaming Liu, Jiadong Xu et al.

Vision-language-action models (VLAs) have shown generalization capabilities in robotic manipulation tasks by inheriting from vision-language models (VLMs) and learning action generation. Most VLA models focus on interpreting vision and language to generate actions, whereas robots must perceive and interact within the spatial-physical world. This gap highlights the need for a comprehensive understanding of robotic-specific multisensory information, which is crucial for achieving complex and contact-rich control. To this end, we introduce a multisensory language-action (MLA) model that collaboratively perceives heterogeneous sensory modalities and predicts future multisensory objectives to facilitate physical world modeling. Specifically, to enhance perceptual representations, we propose an encoder-free multimodal alignment scheme that innovatively repurposes the large language model itself as a perception module, directly interpreting multimodal cues by aligning 2D images, 3D point clouds, and tactile tokens through positional correspondence. To further enhance MLA's understanding of physical dynamics, we design a future multisensory generation post-training strategy that enables MLA to reason about semantic, geometric, and interaction information, providing more robust conditions for action generation. For evaluation, the MLA model outperforms the previous state-of-the-art 2D and 3D VLA methods by 12% and 24% in complex, contact-rich real-world tasks, respectively, while also demonstrating improved generalization to unseen configurations.

Kaichen Zhou, Yuzhen Chen, Fangneng Zhan et al.

Video world models can generate realistic futures from a single instruction, but they often fail to preserve consistent point-level motion over time. As a result, the generated videos appear plausible, yet lack the physical grounding required for reliable action execution, such as robot manipulation. We present GEM-4D, a geometry-grounded video world model that resolves this limitation by injecting dense 4D correspondence supervision, distilled from a pretrained geometry foundation model, into the video generative backbone during training. This supervision enables the model to jointly capture appearance and geometric structure while retaining a single-stream architecture with no additional inference cost. We further introduce an inverse dynamics module that converts correspondence-consistent video rollouts into executable robot trajectories, enabling direct deployment in both real-world and simulated manipulation. GEM-4D achieves state-of-the-art performance on both video prediction and geometric consistency across simulation and realistic scenarios and improves real-world manipulation success from 61% to 81%. Additional results are available at the project page: https://anonymous-submission-20.github.io/gem.github.io/.

Kaichen Zhou, Zeyang Bai, Xinhai Chang et al.

View-conditioned 3D generators such as SAM 3D, TRELLIS and Hunyuan3D produce high-quality object reconstructions from a single view, but real-world visual observation often arrives as long monocular streams. Naively applying these generators to each streaming frame independently leads to severe temporal inconsistency in the generated results. To address this problem, we propose Stream3D, the first training-free streaming mechanism that turns a frozen view-conditioned 3D generator into a streaming generator with constant cross-chunk memory. Stream3D achieves this by maintaining a compact evidential memory, which selectively caches the most informative historical frames based on a proposed evidence score mechanism. As the stream progresses, the memory dynamically updates to retain a fixed number of informative frames, preventing the memory footprint from growing linearly with sequence length. This also prevents degradation over long sequences and keeps the underlying generator completely unchanged without retraining, architectural modifications, or auxiliary losses. Evaluated on both realistic and synthetic streaming benchmarks, Stream3D outperforms latent-transport baselines, including KV-cache reuse and flow-based feature editing, across both photometric and geometric metrics. More details can be found at: https://anonymous-submission-20.github.io/streaming3D.github.io/.

Jialu Wang, Kaichen Zhou, Andrew Markham et al.

Location information is pivotal for the automation and intelligence of terminal devices and edge-cloud IoT systems, such as autonomous vehicles and augmented reality. However, achieving reliable positioning across diverse IoT applications remains challenging due to significant training costs and the necessity of densely collected data. To tackle these issues, we have innovatively applied the selective state space (SSM) model to visual localization, introducing a new model named MambaLoc. The proposed model demonstrates exceptional training efficiency by capitalizing on the SSM model's strengths in efficient feature extraction, rapid computation, and memory optimization, and it further ensures robustness in sparse data environments due to its parameter sparsity. Additionally, we propose the Global Information Selector (GIS), which leverages selective SSM to implicitly achieve the efficient global feature extraction capabilities of Non-local Neural Networks. This design leverages the computational efficiency of the SSM model alongside the Non-local Neural Networks' capacity to capture long-range dependencies with minimal layers. Consequently, the GIS enables effective global information capture while significantly accelerating convergence. Our extensive experimental validation using public indoor and outdoor datasets first demonstrates our model's effectiveness, followed by evidence of its versatility with various existing localization models. Our code and models are publicly available to support further research and development in this area.

Shu Pu, Boya Zeng, Kaichen Zhou et al.

Generative models are increasingly used in 3D vision to synthesize novel shapes, yet it remains unclear whether their generation relies on memorizing training shapes. Understanding their memorization could help prevent training data leakage and improve the diversity of generated results. In this paper, we design an evaluation framework to quantify memorization in 3D generative models and study the influence of different data and modeling designs on memorization. We first apply our framework to quantify memorization in existing methods. Next, through controlled experiments with a latent vector-set (Vecset) diffusion model, we find that, on the data side, memorization depends on data modality, and increases with data diversity and finer-grained conditioning; on the modeling side, it peaks at a moderate guidance scale and can be mitigated by longer Vecsets and simple rotation augmentation. Together, our framework and analysis provide an empirical understanding of memorization in 3D generative models and suggest simple yet effective strategies to reduce it without degrading generation quality. Our code is available at https://github.com/zlab-princeton/3d_mem.

Guibiao Liao, Kaichen Zhou, Zhenyu Bao et al.

The development of Neural Radiance Fields (NeRFs) has provided a potent representation for encapsulating the geometric and appearance characteristics of 3D scenes. Enhancing the capabilities of NeRFs in open-vocabulary 3D semantic perception tasks has been a recent focus. However, current methods that extract semantics directly from Contrastive Language-Image Pretraining (CLIP) for semantic field learning encounter difficulties due to noisy and view-inconsistent semantics provided by CLIP. To tackle these limitations, we propose OV-NeRF, which exploits the potential of pre-trained vision and language foundation models to enhance semantic field learning through proposed single-view and cross-view strategies. First, from the single-view perspective, we introduce Region Semantic Ranking (RSR) regularization by leveraging 2D mask proposals derived from Segment Anything (SAM) to rectify the noisy semantics of each training view, facilitating accurate semantic field learning. Second, from the cross-view perspective, we propose a Cross-view Self-enhancement (CSE) strategy to address the challenge raised by view-inconsistent semantics. Rather than invariably utilizing the 2D inconsistent semantics from CLIP, CSE leverages the 3D consistent semantics generated from the well-trained semantic field itself for semantic field training, aiming to reduce ambiguity and enhance overall semantic consistency across different views. Extensive experiments validate our OV-NeRF outperforms current state-of-the-art methods, achieving a significant improvement of 20.31% and 18.42% in mIoU metric on Replica and ScanNet, respectively. Furthermore, our approach exhibits consistent superior results across various CLIP configurations, further verifying its robustness. Project page: https://github.com/pcl3dv/OV-NeRF.

Sangyun Shin, Kaichen Zhou, Madhu Vankadari et al.

Coarse-to-fine 3D instance segmentation methods show weak performances compared to recent Grouping-based, Kernel-based and Transformer-based methods. We argue that this is due to two limitations: 1) Instance size overestimation by axis-aligned bounding box(AABB) 2) False negative error accumulation from inaccurate box to the refinement phase. In this work, we introduce Spherical Mask, a novel coarse-to-fine approach based on spherical representation, overcoming those two limitations with several benefits. Specifically, our coarse detection estimates each instance with a 3D polygon using a center and radial distance predictions, which avoids excessive size estimation of AABB. To cut the error propagation in the existing coarse-to-fine approaches, we virtually migrate points based on the polygon, allowing all foreground points, including false negatives, to be refined. During inference, the proposal and point migration modules run in parallel and are assembled to form binary masks of instances. We also introduce two margin-based losses for the point migration to enforce corrections for the false positives/negatives and cohesion of foreground points, significantly improving the performance. Experimental results from three datasets, such as ScanNetV2, S3DIS, and STPLS3D, show that our proposed method outperforms existing works, demonstrating the effectiveness of the new instance representation with spherical coordinates. The code is available at: https://github.com/yunshin/SphericalMask

Zhenyu Bao, Guibiao Liao, Kaichen Zhou et al.

Despite the photorealistic novel view synthesis (NVS) performance achieved by the original 3D Gaussian splatting (3DGS), its rendering quality significantly degrades with sparse input views. This performance drop is mainly caused by the limited number of initial points generated from the sparse input, insufficient supervision during the training process, and inadequate regularization of the oversized Gaussian ellipsoids. To handle these issues, we propose the LoopSparseGS, a loop-based 3DGS framework for the sparse novel view synthesis task. In specific, we propose a loop-based Progressive Gaussian Initialization (PGI) strategy that could iteratively densify the initialized point cloud using the rendered pseudo images during the training process. Then, the sparse and reliable depth from the Structure from Motion, and the window-based dense monocular depth are leveraged to provide precise geometric supervision via the proposed Depth-alignment Regularization (DAR). Additionally, we introduce a novel Sparse-friendly Sampling (SFS) strategy to handle oversized Gaussian ellipsoids leading to large pixel errors. Comprehensive experiments on four datasets demonstrate that LoopSparseGS outperforms existing state-of-the-art methods for sparse-input novel view synthesis, across indoor, outdoor, and object-level scenes with various image resolutions.

Renjie Gu, Kaichen Zhou, Yan Luo et al.

Modern Vision-Language Models (VLMs) achieve strong semantic recognition, yet remain brittle on elementary spatial relations such as left of, on, behind, and between. One cause of this failure arises before language reasoning begins: the visual pathway may compress or discard critical 3D structural cues during feature extraction, so the language model receives image representations that are already insufficient for reliable spatial judgment. We introduce GeoWorld-VLM, a VLM-side distillation framework that transfers geometric structure from frozen camera-conditioned video world models into VLMs. GeoWorld-VLM fine-tunes only the image encoder and multimodal projector, aligning post-projector image features with intermediate world-model representations while leaving the main backbone frozen. Given images, a prompt, and a sampled camera trajectory, the world-model teacher converts static visual input into a synthetic multi-view spatial signal. Training combines spatial answer supervision, teacher-student feature alignment, and a preservation anchor to the original VLM. Since the language model remains frozen, GeoWorld-VLM preserves the original model's linguistic capabilities while attributing spatial improvements to the enhanced visual pathway. To evaluate the effectiveness and generality of the proposed method, we apply GeoWorld-VLM to two distinct VLM architectures and observe consistent improvements across both backbones. GeoWorld-VLM improves performance by approximately 4 percent on both the What'sUp and VSR benchmarks, suggesting that world-model-guided visual alignment generalizes across model structures and spatial reasoning datasets.

Yan Xia, Ran Ding, Ziyuan Qin et al.

Recent advances in predicting 6D grasp poses from a single depth image have led to promising performance in robotic grasping. However, previous grasping models face challenges in cluttered environments where nearby objects impact the target object's grasp. In this paper, we first establish a new benchmark dataset for TARget-driven Grasping under Occlusions, named TARGO. We make the following contributions: 1) We are the first to study the occlusion level of grasping. 2) We set up an evaluation benchmark consisting of large-scale synthetic data and part of real-world data, and we evaluated five grasp models and found that even the current SOTA model suffers when the occlusion level increases, leaving grasping under occlusion still a challenge. 3) We also generate a large-scale training dataset via a scalable pipeline, which can be used to boost the performance of grasping under occlusion and generalized to the real world. 4) We further propose a transformer-based grasping model involving a shape completion module, termed TARGO-Net, which performs most robustly as occlusion increases. Our benchmark dataset can be found at https://TARGO-benchmark.github.io/.

Yuxuan Wan, Kaichen Zhou, jinhong Chen et al.

Autonomous assembly in robotics and 3D vision presents significant challenges, particularly in ensuring assembly correctness. Presently, predominant methods such as MEPNet focus on assembling components based on manually provided images. However, these approaches often fall short in achieving satisfactory results for tasks requiring long-term planning. Concurrently, we observe that integrating a self-correction module can partially alleviate such issues. Motivated by this concern, we introduce the Single-Step Assembly Error Correction Task, which involves identifying and rectifying misassembled components. To support research in this area, we present the LEGO Error Correction Assembly Dataset (LEGO-ECA), comprising manual images for assembly steps and instances of assembly failures. Additionally, we propose the Self-Correct Assembly Network (SCANet), a novel method to address this task. SCANet treats assembled components as queries, determining their correctness in manual images and providing corrections when necessary. Finally, we utilize SCANet to correct the assembly results of MEPNet. Experimental results demonstrate that SCANet can identify and correct MEPNet's misassembled results, significantly improving the correctness of assembly. Our code and dataset could be found at https://scanet-iros2024.github.io/.

Kaichen Zhou, Lanqing Hong, Xinhai Chang et al.

A key challenge in fine-grained 3D-based interactive editing is the absence of an efficient representation that balances diverse modifications with high-quality view synthesis under a given memory constraint. While 3D meshes provide robustness for various modifications, they often yield lower-quality view synthesis compared to 3D Gaussian Splatting, which, in turn, suffers from instability during extensive editing. A straightforward combination of these two representations results in suboptimal performance and fails to meet memory constraints. In this paper, we introduce SplatMesh, a novel fine-grained interactive 3D segmentation and editing algorithm that integrates 3D Gaussian Splat with a precomputed mesh and could adjust the memory request based on the requirement. Specifically, given a mesh, \method simplifies it while considering both color and shape, ensuring it meets memory constraints. Then, SplatMesh aligns Gaussian splats with the simplified mesh by treating each triangle as a new reference point. By segmenting and editing the simplified mesh, we can effectively edit the Gaussian splats as well, which will lead to extensive experiments on real and synthetic datasets, coupled with illustrative visual examples, highlighting the superiority of our approach in terms of representation quality and editing performance. Code of our paper can be found here: https://github.com/kaichen-z/SplatMesh.

Yifan Liu, Fangneng Zhan, Kaichen Zhou et al.

Vision-language models (VLMs) struggle with 3D-related tasks such as spatial cognition and physical understanding, which are crucial for real-world applications like robotics and embodied agents. We attribute this to a modality gap between the 3D tasks and the 2D training of VLM, which led to inefficient retrieval of 3D information from 2D input. To bridge this gap, we introduce SandboxVLM, a simple yet effective framework that leverages abstract bounding boxes to encode geometric structure and physical kinematics for VLM. Specifically, we design a 3D Sandbox reconstruction and perception pipeline comprising four stages: generating multi-view priors with abstract control, proxy elevation, multi-view voting and clustering, and 3D-aware reasoning. Evaluated in zero-shot settings across multiple benchmarks and VLM backbones, our approach consistently improves spatial intelligence, achieving an 8.3\% gain on SAT Real compared with baseline methods for instance. These results demonstrate that equipping VLMs with a 3D abstraction substantially enhances their 3D reasoning ability without additional training, suggesting new possibilities for general-purpose embodied intelligence.

Jialu Wang, Kaichen Zhou, Andrew Markham et al.

Despite the advancements in deep learning for camera relocalization tasks, obtaining ground truth pose labels required for the training process remains a costly endeavor. While current weakly supervised methods excel in lightweight label generation, their performance notably declines in scenarios with sparse views. In response to this challenge, we introduce WSCLoc, a system capable of being customized to various deep learning-based relocalization models to enhance their performance under weakly-supervised and sparse view conditions. This is realized with two stages. In the initial stage, WSCLoc employs a multilayer perceptron-based structure called WFT-NeRF to co-optimize image reconstruction quality and initial pose information. To ensure a stable learning process, we incorporate temporal information as input. Furthermore, instead of optimizing SE(3), we opt for $\mathfrak{sim}(3)$ optimization to explicitly enforce a scale constraint. In the second stage, we co-optimize the pre-trained WFT-NeRF and WFT-Pose. This optimization is enhanced by Time-Encoding based Random View Synthesis and supervised by inter-frame geometric constraints that consider pose, depth, and RGB information. We validate our approaches on two publicly available datasets, one outdoor and one indoor. Our experimental results demonstrate that our weakly-supervised relocalization solutions achieve superior pose estimation accuracy in sparse-view scenarios, comparable to state-of-the-art camera relocalization methods. We will make our code publicly available.

Yuxin Ray Song, Jinzhou Li, Rao Fu et al.

The human hand is our primary interface to the physical world, yet egocentric perception rarely knows when, where, or how forcefully it makes contact. Robust wearable tactile sensors are scarce, and no existing in-the-wild datasets align first-person video with full-hand touch. To bridge the gap between visual perception and physical interaction, we present OpenTouch, the first in-the-wild egocentric full-hand tactile dataset, containing 5.1 hours of synchronized video-touch-pose data and 2,900 curated clips with detailed text annotations. Using OpenTouch, we introduce retrieval and classification benchmarks that probe how touch grounds perception and action. We show that tactile signals provide a compact yet powerful cue for grasp understanding, strengthen cross-modal alignment, and can be reliably retrieved from in-the-wild video queries. By releasing this annotated vision-touch-pose dataset and benchmark, we aim to advance multimodal egocentric perception, embodied learning, and contact-rich robotic manipulation.

Kaichen Zhou, Jia-Wang Bian, Jian-Qing Zheng et al.

Despite advancements in self-supervised monocular depth estimation, challenges persist in dynamic scenarios due to the dependence on assumptions about a static world. In this paper, we present Manydepth2, to achieve precise depth estimation for both dynamic objects and static backgrounds, all while maintaining computational efficiency. To tackle the challenges posed by dynamic content, we incorporate optical flow and coarse monocular depth to create a pseudo-static reference frame. This frame is then utilized to build a motion-aware cost volume in collaboration with the vanilla target frame. Furthermore, to improve the accuracy and robustness of the network architecture, we propose an attention-based depth network that effectively integrates information from feature maps at different resolutions by incorporating both channel and non-local attention mechanisms. Compared to methods with similar computational costs, Manydepth2 achieves a significant reduction of approximately five percent in root-mean-square error for self-supervised monocular depth estimation on the KITTI-2015 dataset. The code could be found at https://github.com/kaichen-z/Manydepth2.

Kaichen Zhou, Changhao Chen, Bing Wang et al.

Recent learning-based approaches have achieved impressive results in the field of single-shot camera localization. However, how best to fuse multiple modalities (e.g., image and depth) and to deal with degraded or missing input are less well studied. In particular, we note that previous approaches towards deep fusion do not perform significantly better than models employing a single modality. We conjecture that this is because of the naive approaches to feature space fusion through summation or concatenation which do not take into account the different strengths of each modality. To address this, we propose an end-to-end framework, termed VMLoc, to fuse different sensor inputs into a common latent space through a variational Product-of-Experts (PoE) followed by attention-based fusion. Unlike previous multimodal variational works directly adapting the objective function of vanilla variational auto-encoder, we show how camera localization can be accurately estimated through an unbiased objective function based on importance weighting. Our model is extensively evaluated on RGB-D datasets and the results prove the efficacy of our model. The source code is available at https://github.com/kaichen-z/VMLoc.

Yuqin Dai, Guoqing Wang, Yuan Wang et al.

Retrieval-Augmented Generation (RAG) has advanced open-domain question answering by incorporating external information into model reasoning. However, effectively leveraging external information to enhance reasoning presents the following challenges: (1) low signal-to-noise ratio, where answer-supportive external information is diluted by irrelevant material, and (2) error accumulation, which arises in multi-hop reasoning when incomplete or misleading information is incorporated. To address these challenges, we introduce EviNote-RAG, a framework that follows a retrieve-note-answer workflow. Instead of reasoning directly over raw external information, the model first produces Supportive-Evidence Notes (SENs), which concisely preserve answer-critical information and explicitly mark key and uncertainty information to improve accuracy. We further design an entailment-based Evidence Quality Reward (EQR) to ensure that SENs are logically sufficient to derive the final answer, thereby enhancing SENs' quality. Experiments on both in-domain and out-of-domain QA benchmarks show that EviNote-RAG achieves state-of-the-art performance, improving answer accuracy, training stability, robustness, and efficiency. In particular, it yields relative F1 gains of 20% on HotpotQA (+0.093), 40% on Bamboogle (+0.151), and 91% on 2Wiki (+0.256), benefiting from improvements in the reasoning process.

Jiahui Zhang, Yuelei Li, Anpei Chen et al. · eth-zurich

3D reconstruction and view synthesis are foundational problems in computer vision, graphics, and immersive technologies such as augmented reality (AR), virtual reality (VR), and digital twins. Traditional methods rely on computationally intensive iterative optimization in a complex chain, limiting their applicability in real-world scenarios. Recent advances in feed-forward approaches, driven by deep learning, have revolutionized this field by enabling fast and generalizable 3D reconstruction and view synthesis. This survey offers a comprehensive review of feed-forward techniques for 3D reconstruction and view synthesis, with a taxonomy according to the underlying representation architectures including point cloud, 3D Gaussian Splatting (3DGS), Neural Radiance Fields (NeRF), etc. We examine key tasks such as pose-free reconstruction, dynamic 3D reconstruction, and 3D-aware image and video synthesis, highlighting their applications in digital humans, SLAM, robotics, and beyond. In addition, we review commonly used datasets with detailed statistics, along with evaluation protocols for various downstream tasks. We conclude by discussing open research challenges and promising directions for future work, emphasizing the potential of feed-forward approaches to advance the state of the art in 3D vision.

Haoxiao Wang, Kaichen Zhou, Binrui Gu et al.

Manipulating transparent objects presents significant challenges due to the complexities introduced by their reflection and refraction properties, which considerably hinder the accurate estimation of their 3D shapes. To address these challenges, we propose a single-view RGB-D-based depth completion framework, TransDiff, that leverages the Denoising Diffusion Probabilistic Models(DDPM) to achieve material-agnostic object grasping in desktop. Specifically, we leverage features extracted from RGB images, including semantic segmentation, edge maps, and normal maps, to condition the depth map generation process. Our method learns an iterative denoising process that transforms a random depth distribution into a depth map, guided by initially refined depth information, ensuring more accurate depth estimation in scenarios involving transparent objects. Additionally, we propose a novel training method to better align the noisy depth and RGB image features, which are used as conditions to refine depth estimation step by step. Finally, we utilized an improved inference process to accelerate the denoising procedure. Through comprehensive experimental validation, we demonstrate that our method significantly outperforms the baselines in both synthetic and real-world benchmarks with acceptable inference time. The demo of our method can be found on https://wang-haoxiao.github.io/TransDiff/

Yingji Zhong, Kaichen Zhou, Zhihao Li et al.

Neural Radiance Fields (NeRF) have shown remarkable capabilities for photorealistic novel view synthesis. One major deficiency of NeRF is that dense inputs are typically required, and the rendering quality will drop drastically given sparse inputs. In this paper, we highlight the effectiveness of rendered semantics from dense novel views, and show that rendered semantics can be treated as a more robust form of augmented data than rendered RGB. Our method enhances NeRF's performance by incorporating guidance derived from the rendered semantics. The rendered semantic guidance encompasses two levels: the supervision level and the feature level. The supervision-level guidance incorporates a bi-directional verification module that decides the validity of each rendered semantic label, while the feature-level guidance integrates a learnable codebook that encodes semantic-aware information, which is queried by each point via the attention mechanism to obtain semantic-relevant predictions. The overall semantic guidance is embedded into a self-improved pipeline. We also introduce a more challenging sparse-input indoor benchmark, where the number of inputs is limited to as few as 6. Experiments demonstrate the effectiveness of our method and it exhibits superior performance compared to existing approaches.

Kaichen Zhou, Yuhan Wang, Grace Chen et al.

Recent 3D feed-forward models, such as the Visual Geometry Grounded Transformer (VGGT), have shown strong capability in inferring 3D attributes of static scenes. However, since they are typically trained on static datasets, these models often struggle in real-world scenarios involving complex dynamic elements, such as moving humans or deformable objects like umbrellas. To address this limitation, we introduce PAGE-4D, a feedforward model that extends VGGT to dynamic scenes, enabling camera pose estimation, depth prediction, and point cloud reconstruction -- all without post-processing. A central challenge in multi-task 4D reconstruction is the inherent conflict between tasks: accurate camera pose estimation requires suppressing dynamic regions, while geometry reconstruction requires modeling them. To resolve this tension, we propose a dynamics-aware aggregator that disentangles static and dynamic information by predicting a dynamics-aware mask -- suppressing motion cues for pose estimation while amplifying them for geometry reconstruction. Extensive experiments show that PAGE-4D consistently outperforms the original VGGT in dynamic scenarios, achieving superior results in camera pose estimation, monocular and video depth estimation, and dense point map reconstruction.

Bo Liang, Chen Gong, Haobo Wang et al.

Millimeter-wave Human Pose Estimation (mmWave HPE) promises privacy but suffers from poor generalization under distribution shifts. We demonstrate that brute-force data scaling is ineffective for out-of-distribution (OOD) robustness; efficiency and coverage are the true bottlenecks. To address this, we introduce WiCompass, a coverage-aware data-collection framework. WiCompass leverages large-scale motion-capture corpora to build a universal pose space ``oracle'' that quantifies dataset redundancy and identifies underrepresented motions. Guided by this oracle, WiCompass employs a closed-loop policy to prioritize collecting informative missing samples. Experiments show that WiCompass consistently improves OOD accuracy at matched budgets and exhibits superior scaling behavior compared to conventional collection strategies. By shifting focus from brute-force scaling to coverage-aware data acquisition, this work offers a practical path toward robust mmWave sensing.

Kaichen Zhou, Laura Dodds, Sayed Saad Afzal et al.

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 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 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.

Mingxu Zhang, Xiaoqi Li, Jiahui Xu et al.

Recent advancements in 3D robotic manipulation have improved grasping of everyday objects, but transparent and specular materials remain challenging due to depth sensing limitations. While several 3D reconstruction and depth completion approaches address these challenges, they suffer from setup complexity or limited observation information utilization. To address this, leveraging the power of single view 3D object reconstruction approaches, we propose a training free framework SR3D that enables robotic grasping of transparent and specular objects from a single view observation. Specifically, given single view RGB and depth images, SR3D first uses the external visual models to generate 3D reconstructed object mesh based on RGB image. Then, the key idea is to determine the 3D object's pose and scale to accurately localize the reconstructed object back into its original depth corrupted 3D scene. Therefore, we propose view matching and keypoint matching mechanisms,which leverage both the 2D and 3D's inherent semantic and geometric information in the observation to determine the object's 3D state within the scene, thereby reconstructing an accurate 3D depth map for effective grasp detection. Experiments in both simulation and real world show the reconstruction effectiveness of SR3D.

Chuyan Xiong, Chengyu Shen, Xiaoqi Li et al.

The ability to reflect on and correct failures is crucial for robotic systems to interact stably with real-life objects.Observing the generalization and reasoning capabilities of Multimodal Large Language Models (MLLMs), previous approaches have aimed to utilize these models to enhance robotic systems accordingly.However, these methods typically focus on high-level planning corrections using an additional MLLM, with limited utilization of failed samples to correct low-level contact poses which is particularly prone to occur during articulated object manipulation.To address this gap, we propose an Autonomous Interactive Correction (AIC) MLLM, which makes use of previous low-level interaction experiences to correct SE(3) pose predictions for articulated object. Specifically, AIC MLLM is initially fine-tuned to acquire both pose prediction and feedback prompt comprehension abilities.We design two types of prompt instructions for interactions with objects: 1) visual masks to highlight unmovable parts for position correction, and 2) textual descriptions to indicate potential directions for rotation correction. During inference, a Feedback Information Extraction module is introduced to recognize the failure cause, allowing AIC MLLM to adaptively correct the pose prediction using the corresponding prompts.To further enhance manipulation stability, we devise a Test Time Adaptation strategy that enables AIC MLLM to better adapt to the current scene configuration.Finally, extensive experiments are conducted in both simulated and real-world environments to evaluate the proposed method. The results demonstrate that our AIC MLLM can efficiently correct failure samples by leveraging interaction experience prompts.Our project website is https://sites.google.com/view/aic-mllm.

Jiaming Liu, Mengzhen Liu, Zhenyu Wang et al.

A fundamental objective in robot manipulation is to enable models to comprehend visual scenes and execute actions. Although existing Vision-Language-Action (VLA) models for robots can handle a range of basic tasks, they still face challenges in two areas: (1) insufficient reasoning ability to tackle complex tasks, and (2) high computational costs for VLA model fine-tuning and inference. The recently proposed state space model (SSM) known as Mamba demonstrates promising capabilities in non-trivial sequence modeling with linear inference complexity. Inspired by this, we introduce RoboMamba, an end-to-end robotic VLA model that leverages Mamba to deliver both robotic reasoning and action capabilities, while maintaining efficient fine-tuning and inference. Specifically, we first integrate the vision encoder with Mamba, aligning visual tokens with language embedding through co-training, empowering our model with visual common sense and robotic-related reasoning. To further equip RoboMamba with SE(3) pose prediction abilities, we explore an efficient fine-tuning strategy with a simple policy head. We find that once RoboMamba possesses sufficient reasoning capability, it can acquire manipulation skills with minimal fine-tuning parameters (0.1\% of the model) and time. In experiments, RoboMamba demonstrates outstanding reasoning capabilities on general and robotic evaluation benchmarks. Meanwhile, our model showcases impressive pose prediction results in both simulation and real-world experiments, achieving inference speeds 3 times faster than existing VLA models. Our project web page: https://sites.google.com/view/robomamba-web

Xinhai Chang, Kaichen Zhou

Reconstructing accurate surfaces from sparse multi-view images remains challenging due to severe geometric ambiguity and occlusions. Existing generalizable neural surface reconstruction methods primarily rely on cost volumes that summarize multi-view features using simple statistics (e.g., mean and variance), which discard critical view-dependent geometric structure and often lead to over-smoothed reconstructions. We propose EpiS, a generalizable neural surface reconstruction framework that explicitly leverages epipolar geometry for sparse-view inputs. Instead of directly regressing geometry from cost-volume statistics, EpiS uses coarse cost-volume features to guide the aggregation of fine-grained epipolar features sampled along corresponding epipolar lines across source views. An epipolar transformer fuses multi-view information, followed by ray-wise aggregation to produce SDF-aware features for surface estimation. To further mitigate information loss under sparse views, we introduce a geometry regularization strategy that leverages a pretrained monocular depth model through scale-invariant global and local constraints. Extensive experiments on DTU and BlendedMVS demonstrate that EpiS significantly outperforms state-of-the-art generalizable surface reconstruction methods under sparse-view settings, while maintaining strong generalization without per-scene optimization.

Shoukang Hu, Kaichen Zhou, Kaiyu Li et al.

Neural Radiance Fields (NeRF) has demonstrated remarkable 3D reconstruction capabilities with dense view images. However, its performance significantly deteriorates under sparse view settings. We observe that learning the 3D consistency of pixels among different views is crucial for improving reconstruction quality in such cases. In this paper, we propose ConsistentNeRF, a method that leverages depth information to regularize both multi-view and single-view 3D consistency among pixels. Specifically, ConsistentNeRF employs depth-derived geometry information and a depth-invariant loss to concentrate on pixels that exhibit 3D correspondence and maintain consistent depth relationships. Extensive experiments on recent representative works reveal that our approach can considerably enhance model performance in sparse view conditions, achieving improvements of up to 94% in PSNR, 76% in SSIM, and 31% in LPIPS compared to the vanilla baselines across various benchmarks, including DTU, NeRF Synthetic, and LLFF.

Kaichen Zhou, Lanqing Hong, Shoukang Hu et al.

Automated machine learning (AutoML) usually involves several crucial components, such as Data Augmentation (DA) policy, Hyper-Parameter Optimization (HPO), and Neural Architecture Search (NAS). Although many strategies have been developed for automating these components in separation, joint optimization of these components remains challenging due to the largely increased search dimension and the variant input types of each component. In parallel to this, the common practice of searching for the optimal architecture first and then retraining it before deployment in NAS often suffers from low performance correlation between the searching and retraining stages. An end-to-end solution that integrates the AutoML components and returns a ready-to-use model at the end of the search is desirable. In view of these, we propose DHA, which achieves joint optimization of Data augmentation policy, Hyper-parameter and Architecture. Specifically, end-to-end NAS is achieved in a differentiable manner by optimizing a compressed lower-dimensional feature space, while DA policy and HPO are regarded as dynamic schedulers, which adapt themselves to the update of network parameters and network architecture at the same time. Experiments show that DHA achieves state-of-the-art (SOTA) results on various datasets and search spaces. To the best of our knowledge, we are the first to efficiently and jointly optimize DA policy, NAS, and HPO in an end-to-end manner without retraining.

Chengliang Dai, Shuo Wang, Yuanhan Mo et al.

Machine learning has been widely adopted for medical image analysis in recent years given its promising performance in image segmentation and classification tasks. As a data-driven science, the success of machine learning, in particular supervised learning, largely depends on the availability of manually annotated datasets. For medical imaging applications, such annotated datasets are not easy to acquire. It takes a substantial amount of time and resource to curate an annotated medical image set. In this paper, we propose an efficient annotation framework for brain tumour images that is able to suggest informative sample images for human experts to annotate. Our experiments show that training a segmentation model with only 19% suggestively annotated patient scans from BraTS 2019 dataset can achieve a comparable performance to training a model on the full dataset for whole tumour segmentation task. It demonstrates a promising way to save manual annotation cost and improve data efficiency in medical imaging applications.

Kaichen Zhou, Shiji Song, Gao Huang et al.

In large-scale classification problems, the data set always be faced with frequent updates when a part of the data is added to or removed from the original data set. In this case, conventional incremental learning, which updates an existing classifier by explicitly modeling the data modification, is more efficient than retraining a new classifier from scratch. However, sometimes, we are more interested in determining whether we should update the classifier or performing some sensitivity analysis tasks. To deal with these such tasks, we propose an algorithm to make rational inferences about the updated linear classifier without exactly updating the classifier. Specifically, the proposed algorithm can be used to estimate the upper and lower bounds of the updated classifier's coefficient matrix with a low computational complexity related to the size of the updated dataset. Both theoretical analysis and experiment results show that the proposed approach is superior to existing methods in terms of tightness of coefficients' bounds and computational complexity.

Kaichen Zhou, Shiji Song, Anke Xue et al.

During recent decades, the automatic train operation (ATO) system has been gradually adopted in many subway systems for its low-cost and intelligence. This paper proposes two smart train operation algorithms by integrating the expert knowledge with reinforcement learning algorithms. Compared with previous works, the proposed algorithms can realize the control of continuous action for the subway system and optimize multiple critical objectives without using an offline speed profile. Firstly, through learning historical data of experienced subway drivers, we extract the expert knowledge rules and build inference methods to guarantee the riding comfort, the punctuality, and the safety of the subway system. Then we develop two algorithms for optimizing the energy efficiency of train operation. One is the smart train operation (STO) algorithm based on deep deterministic policy gradient named (STOD) and the other is the smart train operation algorithm based on normalized advantage function (STON). Finally, we verify the performance of proposed algorithms via some numerical simulations with the real field data from the Yizhuang Line of the Beijing Subway and illustrate that the developed smart train operation algorithm are better than expert manual driving and existing ATO algorithms in terms of energy efficiency. Moreover, STOD and STON can adapt to different trip times and different resistance conditions.