Understanding multimodal signals in egocentric vision, such as RGB video, depth, camera poses, and gaze, is essential for applications in augmented reality, robotics, and human-computer interaction, enabling systems to better interpret the camera wearer's actions, intentions, and surrounding environment. However, building large-scale egocentric multimodal and multitask models presents unique challenges. Egocentric data are inherently heterogeneous, with large variations in modality coverage across devices and settings. Generating pseudo-labels for missing modalities, such as gaze or head-mounted camera trajectories, is often infeasible, making standard supervised learning approaches difficult to scale. Furthermore, dynamic camera motion and the complex temporal and spatial structure of first-person video pose additional challenges for the direct application of existing multimodal foundation models. To address these challenges, we introduce a set of efficient temporal tokenizers and propose EgoM2P, a masked modeling framework that learns from temporally-aware multimodal tokens to train a large, general-purpose model for egocentric 4D understanding. This unified design supports multitasking across diverse egocentric perception and synthesis tasks, including gaze prediction, egocentric camera tracking, and monocular depth estimation from egocentric video, and also serves as a generative model for conditional egocentric video synthesis. Across these tasks, EgoM2P matches or outperforms specialist models while being an order of magnitude faster. We will fully open-source EgoM2P to support the community and advance egocentric vision research. Project page: https://egom2p.github.io/.
Obstacle avoidance is essential for ensuring the safety of autonomous vehicles. Accurate perception and motion planning are crucial to enabling vehicles to navigate complex environments while avoiding collisions. In this paper, we propose an efficient obstacle avoidance pipeline that leverages a camera-only perception module and a Frenet-Pure Pursuit-based planning strategy. By integrating advancements in computer vision, the system utilizes YOLOv11 for object detection and state-of-the-art monocular depth estimation models, such as Depth Anything V2, to estimate object distances. A comparative analysis of these models provides valuable insights into their accuracy, efficiency, and robustness in real-world conditions. The system is evaluated in diverse scenarios on a university campus, demonstrating its effectiveness in handling various obstacles and enhancing autonomous navigation. The video presenting the results of the obstacle avoidance experiments is available at: https://www.youtube.com/watch?v=FoXiO5S_tA8
We introduce $\pi^3$, a feed-forward neural network that offers a novel approach to visual geometry reconstruction, breaking the reliance on a conventional fixed reference view. Previous methods often anchor their reconstructions to a designated viewpoint, an inductive bias that can lead to instability and failures if the reference is suboptimal. In contrast, $\pi^3$ employs a fully permutation-equivariant architecture to predict affine-invariant camera poses and scale-invariant local point maps without any reference frames. This design makes our model inherently robust to input ordering and highly scalable. These advantages enable our simple and bias-free approach to achieve state-of-the-art performance on a wide range of tasks, including camera pose estimation, monocular/video depth estimation, and dense point map reconstruction. Code and models are publicly available.
The pursuit of a generalizable stereo matching model, capable of performing across varying resolutions and disparity ranges without dataset-specific fine-tuning, has revealed a fundamental trade-off. Iterative local search methods achieve high scores on constrained benchmarks, but their core mechanism inherently limits the global consistency required for true generalization. On the other hand, global matching architectures, while theoretically more robust, have been historically rendered infeasible by prohibitive computational and memory costs. We resolve this dilemma with $S^2M^2$: a global matching architecture that achieves both state-of-the-art accuracy and high efficiency without relying on cost volume filtering or deep refinement stacks. Our design integrates a multi-resolution transformer for robust long-range correspondence, trained with a novel loss function that concentrates probability on feasible matches. This approach enables a more robust joint estimation of disparity, occlusion, and confidence. $S^2M^2$ establishes a new state of the art on the Middlebury v3 and ETH3D benchmarks, significantly outperforming prior methods across most metrics while reconstructing high-quality details with competitive efficiency.
Transformers are increasingly prevalent for multi-view computer vision tasks, where geometric relationships between viewpoints are critical for 3D perception. To leverage these relationships, multi-view transformers must use camera geometry to ground visual tokens in 3D space. In this work, we compare techniques for conditioning transformers on cameras: token-level raymap encodings, attention-level relative pose encodings, and a new relative encoding we propose -- Projective Positional Encoding (PRoPE) -- that captures complete camera frustums, both intrinsics and extrinsics, as a relative positional encoding. Our experiments begin by showing how relative camera conditioning improves performance in feedforward novel view synthesis, with further gains from PRoPE. This holds across settings: scenes with both shared and varying intrinsics, when combining token- and attention-level conditioning, and for generalization to inputs with out-of-distribution sequence lengths and camera intrinsics. We then verify that these benefits persist for different tasks, stereo depth estimation and discriminative spatial cognition, as well as larger model sizes.
Depth estimation is a fundamental task in 3D computer vision, crucial for applications such as 3D reconstruction, free-viewpoint rendering, robotics, autonomous driving, and AR/VR technologies. Traditional methods relying on hardware sensors like LiDAR are often limited by high costs, low resolution, and environmental sensitivity, limiting their applicability in real-world scenarios. Recent advances in vision-based methods offer a promising alternative, yet they face challenges in generalization and stability due to either the low-capacity model architectures or the reliance on domain-specific and small-scale datasets. The emergence of scaling laws and foundation models in other domains has inspired the development of "depth foundation models": deep neural networks trained on large datasets with strong zero-shot generalization capabilities. This paper surveys the evolution of deep learning architectures and paradigms for depth estimation across the monocular, stereo, multi-view, and monocular video settings. We explore the potential of these models to address existing challenges and provide a comprehensive overview of large-scale datasets that can facilitate their development. By identifying key architectures and training strategies, we aim to highlight the path towards robust depth foundation models, offering insights into their future research and applications.
Learning-based Multi-View Stereo (MVS) methods aim to predict depth maps for a sequence of calibrated images to recover dense point clouds. However, existing MVS methods often struggle with challenging regions, such as textureless regions and reflective surfaces, where feature matching fails. In contrast, monocular depth estimation inherently does not require feature matching, allowing it to achieve robust relative depth estimation in these regions. To bridge this gap, we propose MonoMVSNet, a novel monocular feature and depth guided MVS network that integrates powerful priors from a monocular foundation model into multi-view geometry. Firstly, the monocular feature of the reference view is integrated into source view features by the attention mechanism with a newly designed cross-view position encoding. Then, the monocular depth of the reference view is aligned to dynamically update the depth candidates for edge regions during the sampling procedure. Finally, a relative consistency loss is further designed based on the monocular depth to supervise the depth prediction. Extensive experiments demonstrate that MonoMVSNet achieves state-of-the-art performance on the DTU and Tanks-and-Temples datasets, ranking first on the Tanks-and-Temples Intermediate and Advanced benchmarks. The source code is available at https://github.com/JianfeiJ/MonoMVSNet.
Calisthenics skill classification is the computer vision task of inferring the skill performed by an athlete from images, enabling automatic performance assessment and personalized analytics. Traditional methods for calisthenics skill recognition are based on pose estimation methods to determine the position of skeletal data from images, which is later fed to a classification algorithm to infer the performed skill. Despite the progress in human pose estimation algorithms, they still involve high computational costs, long inference times, and complex setups, which limit the applicability of such approaches in real-time applications or mobile devices. This work proposes a direct approach to calisthenics skill recognition, which leverages depth estimation and athlete patch retrieval to avoid the computationally expensive human pose estimation module. Using Depth Anything V2 for depth estimation and YOLOv10 for athlete localization, we segment the subject from the background rather than relying on traditional pose estimation techniques. This strategy increases efficiency, reduces inference time, and improves classification accuracy. Our approach significantly outperforms skeleton-based methods, achieving 38.3x faster inference with RGB image patches and improved classification accuracy with depth patches (0.837 vs. 0.815). Beyond these performance gains, the modular design of our pipeline allows for flexible replacement of components, enabling future enhancements and adaptation to real-world applications.
In the aftermath of earthquakes, social media images have become a crucial resource for disaster reconnaissance, providing immediate insights into the extent of damage. Traditional approaches to damage severity assessment in post-earthquake social media images often rely on classification methods, which are inherently subjective and incapable of accounting for the varying extents of damage within an image. Addressing these limitations, this study proposes a novel approach by framing damage severity assessment as a semantic segmentation problem, aiming for a more objective analysis of damage in earthquake-affected areas. The methodology involves the construction of a segmented damage severity dataset, categorizing damage into three degrees: undamaged structures, damaged structures, and debris. Utilizing this dataset, the study fine-tunes a SegFormer model to generate damage severity segmentations for post-earthquake social media images. Furthermore, a new damage severity scoring system is introduced, quantifying damage by considering the varying degrees of damage across different areas within images, adjusted for depth estimation. The application of this approach allows for the quantification of damage severity in social media images in a more objective and comprehensive manner. By providing a nuanced understanding of damage, this study enhances the ability to offer precise guidance to disaster reconnaissance teams, facilitating more effective and targeted response efforts in the aftermath of earthquakes.
We introduce ByDeWay, a training-free framework designed to enhance the performance of Multimodal Large Language Models (MLLMs). ByDeWay uses a novel prompting strategy called Layered-Depth-Based Prompting (LDP), which improves spatial reasoning and grounding without modifying any model parameters. It segments the scene into closest, mid-range, and farthest layers using monocular depth estimation, then generates region-specific captions with a grounded vision-language model. These structured, depth-aware captions are appended to the image-question prompt, enriching it with spatial context. This guides MLLMs to produce more grounded and less hallucinated responses. Our method is lightweight, modular, and compatible with black-box MLLMs. Experiments on hallucination-sensitive (POPE) and reasoning-intensive (GQA) benchmarks show consistent improvements across multiple MLLMs, validating the effectiveness of depth-aware prompting in a zero-training setting.
Recent advances in 3D Gaussian Splatting (3DGS) have enabled real-time novel view synthesis (NVS) with impressive quality in indoor scenes. However, achieving high-fidelity rendering requires meticulously captured images covering the entire scene, limiting accessibility for general users. We aim to develop a practical 3DGS-based NVS framework using simple panorama-style motion with a handheld camera (e.g., mobile device). While convenient, this rotation-dominant motion and narrow baseline make accurate camera pose and 3D point estimation challenging, especially in textureless indoor scenes. To address these challenges, we propose LighthouseGS, a novel framework inspired by the lighthouse-like sweeping motion of panoramic views. LighthouseGS leverages rough geometric priors, such as mobile device camera poses and monocular depth estimation, and utilizes the planar structures often found in indoor environments. We present a new initialization method called plane scaffold assembly to generate consistent 3D points on these structures, followed by a stable pruning strategy to enhance geometry and optimization stability. Additionally, we introduce geometric and photometric corrections to resolve inconsistencies from motion drift and auto-exposure in mobile devices. Tested on collected real and synthetic indoor scenes, LighthouseGS delivers photorealistic rendering, surpassing state-of-the-art methods and demonstrating the potential for panoramic view synthesis and object placement.
Monocular depth estimation has recently advanced to provide not only relative but also metric depth predictions. However, its reliability in underwater environments remains limited due to light attenuation and scattering, color distortion, turbidity, and the lack of high-quality metric ground-truth data. In this paper, we present a comprehensive benchmark of zero-shot and fine-tuned monocular metric depth estimation models on real-world underwater datasets with metric depth annotations, such as FLSea and SQUID. We evaluate a diverse set of state-of-the-art models across a range of underwater conditions with different ranges. Our results show that large-scale models trained on terrestrial (real or synthetic) data, while effective in in-air settings, perform poorly underwater due to significant domain shifts. To address this, we fine-tune Depth Anything V2 with a ViT-S backbone encoder on a synthetic underwater variant of the Hypersim dataset, which we generated using a physically based underwater image formation model. We demonstrate our fine-tuned model consistently improves performance across all benchmarks and outperforms baselines trained only on the clean in-air Hypersim dataset. Our study provides a detailed evaluation and visualization for monocular metric depth estimation in underwater scenes, highlighting the importance of domain adaptation and scale-aware supervision for achieving robust and generalizable metric depth predictions in challenging underwater environments for future research.
Learning-based stereo matching models struggle in adverse weather conditions due to the scarcity of corresponding training data and the challenges in extracting discriminative features from degraded images. These limitations significantly hinder zero-shot generalization to out-of-distribution weather conditions. In this paper, we propose \textbf{RobuSTereo}, a novel framework that enhances the zero-shot generalization of stereo matching models under adverse weather by addressing both data scarcity and feature extraction challenges. First, we introduce a diffusion-based simulation pipeline with a stereo consistency module, which generates high-quality stereo data tailored for adverse conditions. By training stereo matching models on our synthetic datasets, we reduce the domain gap between clean and degraded images, significantly improving the models' robustness to unseen weather conditions. The stereo consistency module ensures structural alignment across synthesized image pairs, preserving geometric integrity and enhancing depth estimation accuracy. Second, we design a robust feature encoder that combines a specialized ConvNet with a denoising transformer to extract stable and reliable features from degraded images. The ConvNet captures fine-grained local structures, while the denoising transformer refines global representations, effectively mitigating the impact of noise, low visibility, and weather-induced distortions. This enables more accurate disparity estimation even under challenging visual conditions. Extensive experiments demonstrate that \textbf{RobuSTereo} significantly improves the robustness and generalization of stereo matching models across diverse adverse weather scenarios.
World models have become indispensable tools for embodied intelligence, serving as powerful simulators capable of generating realistic robotic videos while addressing critical data scarcity challenges. However, current embodied world models exhibit limited physical awareness, particularly in modeling 3D geometry and motion dynamics, resulting in unrealistic video generation for contact-rich robotic scenarios. In this paper, we present RoboScape, a unified physics-informed world model that jointly learns RGB video generation and physics knowledge within an integrated framework. We introduce two key physics-informed joint training tasks: temporal depth prediction that enhances 3D geometric consistency in video rendering, and keypoint dynamics learning that implicitly encodes physical properties (e.g., object shape and material characteristics) while improving complex motion modeling. Extensive experiments demonstrate that RoboScape generates videos with superior visual fidelity and physical plausibility across diverse robotic scenarios. We further validate its practical utility through downstream applications including robotic policy training with generated data and policy evaluation. Our work provides new insights for building efficient physics-informed world models to advance embodied intelligence research. The code is available at: https://github.com/tsinghua-fib-lab/RoboScape.
Video Instance Segmentation (VIS) fundamentally struggles with pervasive challenges including object occlusions, motion blur, and appearance variations during temporal association. To overcome these limitations, this work introduces geometric awareness to enhance VIS robustness by strategically leveraging monocular depth estimation. We systematically investigate three distinct integration paradigms. Expanding Depth Channel (EDC) method concatenates the depth map as input channel to segmentation networks; Sharing ViT (SV) designs a uniform ViT backbone, shared between depth estimation and segmentation branches; Depth Supervision (DS) makes use of depth prediction as an auxiliary training guide for feature learning. Though DS exhibits limited effectiveness, benchmark evaluations demonstrate that EDC and SV significantly enhance the robustness of VIS. When with Swin-L backbone, our EDC method gets 56.2 AP, which sets a new state-of-the-art result on OVIS benchmark. This work conclusively establishes depth cues as critical enablers for robust video understanding.
Recent large-scale Vision Language Action (VLA) models have shown superior performance in robotic manipulation tasks guided by natural language. However, their generalization remains limited when applied to novel objects or unfamiliar environments that lie outside the training distribution. To address this, many existing approaches integrate additional components such as depth estimation, segmentation, or even diffusion to improve generalization, at the cost of adding significant computation overhead, resulting in low efficiency. This motivates the exploration of efficient action prediction methods, which are independent of additional high-level visual representations or diffusion techniques. In this work, we propose VOTE, an efficient and general framework for the optimization and acceleration of VLA models. In details, we propose a novel tokenizer-free fine-tuning approach for parallel accurate action prediction, which reduces computational overhead and accelerates inference speed. Additionally, we adopt an ensemble voting strategy for the action sampling, which significantly improves model performance and enhances generalization. Experimental results show that our method achieves state-of-the-art performance with 35$\times$ faster inference and 145 Hz throughput. All the details and codes will be open-sourced.
Monocular depth estimation methods traditionally train deep models to infer depth directly from RGB pixels. This implicit learning often overlooks explicit monocular cues that the human visual system relies on, such as occlusion boundaries, shading, and perspective. Rather than expecting a network to discover these cues unaided, we present ThirdEye, a cue-aware pipeline that deliberately supplies each cue through specialised, pre-trained, and frozen networks. These cues are fused in a three-stage cortical hierarchy (V1->V2->V3) equipped with a key-value working-memory module that weights them by reliability. An adaptive-bins transformer head then produces a high-resolution disparity map. Because the cue experts are frozen, ThirdEye inherits large amounts of external supervision while requiring only modest fine-tuning. This extended version provides additional architectural detail, neuroscientific motivation, and an expanded experimental protocol; quantitative results will appear in a future revision.
Recent video depth estimation methods achieve great performance by following the paradigm of image depth estimation, i.e., typically fine-tuning pre-trained video diffusion models with massive data. However, we argue that video depth estimation is not a naive extension of image depth estimation. The temporal consistency requirements for dynamic and static regions in videos are fundamentally different. Consistent video depth in static regions, typically backgrounds, can be more effectively achieved via stereo matching across all frames, which provides much stronger global 3D cues. While the consistency for dynamic regions still should be learned from large-scale video depth data to ensure smooth transitions, due to the violation of triangulation constraints. Based on these insights, we introduce StereoDiff, a two-stage video depth estimator that synergizes stereo matching for mainly the static areas with video depth diffusion for maintaining consistent depth transitions in dynamic areas. We mathematically demonstrate how stereo matching and video depth diffusion offer complementary strengths through frequency domain analysis, highlighting the effectiveness of their synergy in capturing the advantages of both. Experimental results on zero-shot, real-world, dynamic video depth benchmarks, both indoor and outdoor, demonstrate StereoDiff's SoTA performance, showcasing its superior consistency and accuracy in video depth estimation.
Monocular depth estimation and ego-motion estimation are significant tasks for scene perception and navigation in stable, accurate and efficient robot-assisted endoscopy. To tackle lighting variations and sparse textures in endoscopic scenes, multiple techniques including optical flow, appearance flow and intrinsic image decomposition have been introduced into the existing methods. However, the effective training strategy for multiple modules are still critical to deal with both illumination issues and information interference for self-supervised depth estimation in endoscopy. Therefore, a novel framework with multistep efficient finetuning is proposed in this work. In each epoch of end-to-end training, the process is divided into three steps, including optical flow registration, multiscale image decomposition and multiple transformation alignments. At each step, only the related networks are trained without interference of irrelevant information. Based on parameter-efficient finetuning on the foundation model, the proposed method achieves state-of-the-art performance on self-supervised depth estimation on SCARED dataset and zero-shot depth estimation on Hamlyn dataset, with 4\%$\sim$10\% lower error. The evaluation code of this work has been published on https://github.com/BaymaxShao/EndoMUST.
Dense metric depth estimation using millimeter-wave radar typically requires dense LiDAR supervision, generated via multi-frame projection and interpolation, to guide the learning of accurate depth from sparse radar measurements and RGB images. However, this paradigm is both costly and data-intensive. To address this, we propose RaCalNet, a novel framework that eliminates the need for dense supervision by using sparse LiDAR to supervise the learning of refined radar measurements, resulting in a supervision density of merely around 1% compared to dense-supervised methods. Unlike previous approaches that associate radar points with broad image regions and rely heavily on dense labels, RaCalNet first recalibrates and refines sparse radar points to construct accurate depth priors. These priors then serve as reliable anchors to guide monocular depth prediction, enabling metric-scale estimation without resorting to dense supervision. This design improves structural consistency and preserves fine details. Despite relying solely on sparse supervision, RaCalNet surpasses state-of-the-art dense-supervised methods, producing depth maps with clear object contours and fine-grained textures. Extensive experiments on the ZJU-4DRadarCam dataset and real-world deployment scenarios demonstrate its effectiveness, reducing RMSE by 35.30% and 34.89%, respectively.