Recent advances in foundational 3D reconstruction models, such as DUSt3R and MASt3R, have shown great potential in 2D and 3D correspondence in static scenes. In this paper, we propose to adapt them for the task of point tracking through 3D grounded correspondence. We first demonstrate that these models are competitive point trackers when focusing on static points, present in current point tracking benchmarks (+33.5% on EgoPoints vs. CoTracker2). We propose to combine the reconstruction loss with training for dynamic correspondence along with a visibility head, and fine-tuning MASt3R for point tracking using a relatively small amount of synthetic data. Importantly, we only train and evaluate on pairs of frames where one contains the query point, effectively removing any temporal context.Using a mix of dynamic and static point correspondences, we achieve competitive or superior point tracking results on four datasets (e.g. competitive on TAP-Vid-DAVIS 73.8 δavg/ 85.8% occlusion acc. for PointSt3R compared to 75.7 / 88.3% for CoTracker2; and significantly outperform CoTracker3 on EgoPoints 61.3 vs 54.2 and RGB-S 87.0 vs 82.8). We also present results on 3D point tracking along with several ablations on training datasets and percentage of dynamic correspondences.

We introduce 4D Motion Scaffolds (MoSca), a modern 4D reconstruction system designed to reconstruct and synthesize novel views of dynamic scenes from monocular videos captured casually in the wild. To address such a challenging and ill-posed inverse problem, we leverage prior knowledge from foundational vision models and lift the video data to a novel Motion Scaffold (MoSca) representation, which compactly and smoothly encodes the underlying motions/deformations. The scene geometry and appearance are then disentangled from the deformation field and are encoded by globally fusing the Gaussians anchored onto the MoSca and optimized via Gaussian Splatting. Additionally, camera focal length and poses can be solved using bundle adjustment without the need of any other pose estimation tools. Experiments demonstrate state-of-the-art performance on dynamic rendering benchmarks and its effectiveness on real videos.

We introduce AllTracker: a model that estimates long-range point tracks by way of estimating the flow field between a query frame and every other frame of a video. Unlike existing point tracking methods, our approach delivers high-resolution and dense (all-pixel) correspondence fields, which can be visualized as flow maps. Unlike existing optical flow methods, our approach corresponds one frame to hundreds of subsequent frames, rather than just the next frame. We develop a new architecture for this task, blending techniques from existing work in optical flow and point tracking: the model performs iterative inference on low-resolution grids of correspondence estimates, propagating information spatially via 2D convolution layers, and propagating information temporally via pixel-aligned attention layers. The model is fast and parameter-efficient (16 million parameters), and delivers state-of-the-art point tracking accuracy at high resolution (i.e., tracking 768x1024 pixels, on a 40G GPU). A benefit of our design is that we can train jointly on optical flow datasets and point tracking datasets, and we find that doing so is crucial for top performance. We provide an extensive ablation study on our architecture details and training recipe, making it clear which details matter most. Our code and model weights are available at https://alltracker.github.io

The ability to predict collision-free future trajectories from egocentric observations is crucial in applications such as humanoid robotics, VR / AR, and assistive navigation. In this work, we introduce the challenging problem of predicting a sequence of future 6D head poses from an egocentric video. In particular, we predict both head translations and rotations to learn the active information-gathering behavior expressed through head-turning events. To solve this task, we propose a framework that reasons over temporally aggregated 3D latent features, which models the geometric and semantic constraints for both the static and dynamic parts of the environment. Motivated by the lack of training data in this space, we further contribute a data collection pipeline using the Project Aria glasses, and present a dataset collected through this approach. Our dataset, dubbed Aria Navigation Dataset (AND), consists of 4 hours of recording of users navigating in real-world scenarios. It includes diverse situations and navigation behaviors, providing a valuable resource for learning real-world egocentric navigation policies. Extensive experiments show that our model learns human-like navigation behaviors such as waiting / slowing down, rerouting, and looking around for traffic while generalizing to unseen environments. Check out our project webpage at https://sites.google.com/stanford.edu/lookout.

We introduce TAPIP3D, a novel approach for long-term 3D point tracking in monocular RGB and RGB-D videos. TAPIP3D represents videos as camera-stabilized spatio-temporal feature clouds, leveraging depth and camera motion information to lift 2D video features into a 3D world space where camera movement is effectively canceled out. Within this stabilized 3D representation, TAPIP3D iteratively refines multi-frame motion estimates, enabling robust point tracking over long time horizons. To handle the irregular structure of 3D point distributions, we propose a 3D Neighborhood-to-Neighborhood (N2N) attention mechanism—a 3D-aware contextualization strategy that builds informative, spatially coherent feature neighborhoods to support precise trajectory estimation. Our 3D-centric formulation significantly improves performance over existing 3D point tracking methods and even surpasses state-of-the-art 2D pixel trackers in accuracy when reliable depth is available. The model supports inference in both camera-centric (unstabilized) and world-centric (stabilized) coordinates, with experiments showing that compensating for camera motion leads to substantial gains in tracking robustness. By replacing the conventional 2D square correlation windows used in prior 2D and 3D trackers with a spatially grounded 3D attention mechanism, TAPIP3D achieves strong and consistent results across multiple 3D point tracking benchmarks. Our code and trained checkpoints will be public.

We introduce EgoPoints, a benchmark for point tracking in egocentric videos. We annotate 4.7K challenging tracks in egocentric sequences. Compared to the popular TAP-Vid-DAVIS [7] evaluation benchmark, we include 9x more points that go out-of-view and 59x more points that require re-identification (ReID) after returning to view. To measure the performance of models on these challenging points, we introduce evaluation metrics that specifically monitor tracking performance on points in-view, out-of-view, and points that require re-identification. We then propose a pipeline to create semi-real sequences, with automatic ground truth. We generate 11 K such sequences by combining dynamic Kubric [17] objects with scene points from EPIC Fields [38]. When fine-tuning point tracking methods on these sequences and evaluating on our annotated EgoPoints sequences, we improve Co-Tracker [22] across all metrics, including the tracking accuracy$\delta^{\ast}_{avg}$by 2.7 percentage points and accuracy on ReID sequences ($ReID\delta_{avg}$) by 2.4 points. We also improve$\delta^{\ast}_{avg}$and$ReID\delta_{avg}$of PIPs++ [42] by 0.3 and 2.8 respectively.

State-of-the-art models on contemporary 3D segmentation benchmarks like ScanNet consume and label dataset- provided 3D point clouds, obtained through post processing of sensed multiview RGB-D images. They are typically trained in-domain, forego large-scale 2D pre-training and outperform alternatives that featurize the posed RGB- D multiview images instead. The gap in performance between methods that consume posed images versus post- processed 3D point clouds has fueled the belief that 2D and 3D perception require distinct model architectures. In this paper, we challenge this view and propose ODIN (Omni-Dimensional INstance segmentation), a model that can segment and label both 2D RGB images and 3D point clouds, using a transformer architecture that alternates between 2D within-view and 3D cross-view information fusion. Our model differentiates 2D and 3D feature oper-ations through the positional encodings of the tokens in-volved, which capture pixel coordinates for 2D patch tokens and 3D coordinates for 3D feature tokens. ODIN achieves state-of-the-art performance on ScanNet200, Matterport3D and AI2THOR 3D instance segmentation benchmarks, and competitive performance on ScanNet, S3DIS and COCO. It outperforms all previous works by a wide margin when the sensed 3D point cloud is used in place of the point cloud sampled from 3D mesh. When used as the 3D perception engine in an instructable embodied agent architecture, it sets a new state-of-the-art on the TEACh action-from-dialogue benchmark. Our code and checkpoints can be found at the project website

Object tracking is central to robot perception and scene understanding, allowing robots to parse a video stream in terms of moving objects with names. Tracking-by-detection has long been a dominant paradigm for object tracking of specific object categories [1], [2]. Recently, large-scale pre-trained models have shown promising advances in detecting and segmenting objects and parts in 2D static images in the wild. This raises the question: can we re-purpose these large-scale pre-trained static image models for open-vocabulary video tracking? In this paper, we combine an open-vocabulary detector [3], segmenter [4], and dense optical flow estimator [5], into a model that tracks and segments any object in 2D videos. Given a monocular video input, our method predicts object and part mask tracks with associated language descriptions, rebuilding the pipeline of Tractor [6] with modern large pre-trained models for static image detection and segmentation: we detect open-vocabulary object instances and propagate their boxes from frame to frame using a flow-based motion model, refine the propagated boxes with the box regression module of the visual detector, and prompt an open-world segmenter with the refined box to segment the objects. We decide the termination of an object track based on the objectness score of the propagated boxes as well as forward-backward optical flow consistency. We re-identify objects across occlusions using deep feature matching. We show that our model achieves strong performance on multiple established benchmarks [7], [8], [9], [10], and can produce reasonable tracks in manipulation data [11]. In particular, our model outperforms previous state-of-the-art in UVO and BURST, benchmarks for open-world object tracking and segmentation, despite never being explicitly trained for tracking. We hope that our approach can serve as a simple and extensible framework for future research and enable imitation learning from videos with unconventional objects.

Given the difficulty of manually annotating motion in video, the current best motion estimation methods are trained with synthetic data, and therefore struggle somewhat due to a train/test gap. Self-supervised methods hold the promise of training directly on real video, but typically perform worse. These include methods trained with warp error (i.e., color constancy) combined with smoothness terms, and methods that encourage cycle-consistency in the estimates (i.e., tracking backwards should yield the opposite trajectory as tracking forwards). In this work, we take on the challenge of improving state-of-the-art supervised models with self-supervised training. We find that when the initialization is supervised weights, most existing self-supervision techniques actually make performance worse instead of better, which suggests that the benefit of seeing the new data is overshadowed by the noise in the training signal. Focusing on obtaining a "clean" training signal from real-world unlabelled video, we propose to separate label-making and training into two distinct stages. In the first stage, we use the pre-trained model to estimate motion in a video, and then select the subset of motion estimates which we can verify with cycle-consistency. This produces a sparse but accurate pseudo-labelling of the video. In the second stage, we fine-tune the model to reproduce these outputs, while also applying augmentations on the input. We complement this boot-strapping method with simple techniques that densify and re-balance the pseudo-labels, ensuring that we do not merely train on "easy" tracks. We show that our method yields reliable gains over fully-supervised methods in real videos, for both short-term (flow-based) and long-range (multiframe) pixel tracking. Our code can be found here: https: //github.com/AlexSunNik/refining-motion-code.

Gaussian splatting has become a popular representation for novel-view synthesis, exhibiting clear strengths in efficiency, photometric quality, and compositional edibility. Following its success, many works have extended Gaussians to 4D, showing that dynamic Gaussians maintain these benefits while also tracking scene geometry far better than alternative representations. Yet, these methods assume dense multi-view videos as supervision. In this work, we are interested in extending the capability of Gaussian scene representations to casually captured monocular videos. We show that existing 4D Gaussian methods dramatically fail in this setup because the monocular setting is underconstrained. Building off this finding, we propose a method we call Dynamic Gaussian Marbles, which consist of three core modifications that target the difficulties of the monocular setting. First, we use isotropic Gaussian "marbles'', reducing the degrees of freedom of each Gaussian. Second, we employ a hierarchical divide and-conquer learning strategy to efficiently guide the optimization towards solutions with globally coherent motion. Finally, we add image-level and geometry-level priors into the optimization, including a tracking loss that takes advantage of recent progress in point tracking. By constraining the optimization, Dynamic Gaussian Marbles learns Gaussian trajectories that enable novel-view rendering and accurately capture the 3D motion of the scene elements. We evaluate on the Nvidia Dynamic Scenes dataset and the DyCheck iPhone dataset, and show that Gaussian Marbles significantly outperforms other Gaussian baselines in quality, and is on-par with non-Gaussian representations, all while maintaining the efficiency, compositionality, editability, and tracking benefits of Gaussians. Our project page can be found here https://geometry.stanford.edu/projects/dynamic-gaussian-marbles.github.io/.

We introduce PointOdyssey, a large-scale synthetic dataset, and data generation framework, for the training and evaluation of long-term fine-grained tracking algorithms. Our goal is to advance the state-of-the-art by placing emphasis on long videos with naturalistic motion. Toward the goal of naturalism, we animate deformable characters using real-world motion capture data, we build 3D scenes to match the motion capture environments, and we render camera viewpoints using trajectories mined via structure-from-motion on real videos. We create combinatorial diversity by randomizing character appearance, motion profiles, materials, lighting, 3D assets, and atmospheric effects. Our dataset currently includes 104 videos, averaging 2,000 frames long, with orders of magnitude more correspondence annotations than prior work. We show that existing methods can be trained from scratch in our dataset and outperform the published variants. Finally, we introduce modifications to the PIPs point tracking method, greatly widening its temporal receptive field, which improves its performance on PointOdyssey as well as on two real-world benchmarks. Our data and code are publicly available at: https://pointodyssey.com

Building 3D perception systems for autonomous vehicles that do not rely on high-density LiDAR is a critical research problem because of the expense of LiDAR systems compared to cameras and other sensors. Recent research has developed a variety of camera-only methods, where features are differentiably “lifted” from the multi-camera images onto the 2D ground plane, yielding a “bird's eye view” (BEV) feature representation of the 3D space around the vehicle. This line of work has produced a variety of novel “lifting” methods, but we observe that other details in the training setups have shifted at the same time, making it unclear what really matters in top-performing methods. We also observe that using cameras alone is not a real-world constraint, considering that additional sensors like radar have been integrated into real vehicles for years already. In this paper, we first of all attempt to elucidate the high-impact factors in the design and training protocol of BEV perception models. We find that batch size and input resolution greatly affect performance, while lifting strategies have a more modest effect-even a simple parameter-free lifter works well. Second, we demonstrate that radar data can provide a substantial boost to performance, helping to close the gap between camera-only and LiDAR-enabled systems. We analyze the radar usage details that lead to good performance, and invite the community to re-consider this commonly-neglected part of the sensor platform.

We propose an unsupervised method for detecting and tracking moving objects in 3D, in unlabelled RGB-D videos. The method begins with classic handcrafted techniques for segmenting objects using motion cues: we estimate optical flow and camera motion, and conservatively segment regions that appear to be moving independently of the background. Treating these initial segments as pseudo-labels, we learn an ensemble of appearance-based 2D and 3D detectors, under heavy data augmentation. We use this ensemble to detect new instances of the "moving" type, even if they are not moving, and add these as new pseudo-labels. Our method is an expectation-maximization algorithm, where in the expectation step we fire all modules and look for agreement among them, and in the maximization step we re-train the modules to improve this agreement. The constraint of ensemble agreement helps combat contamination of the generated pseudo-labels (during the E step), and data augmentation helps the modules generalize to yetunlabelled data (during the M step). We compare against existing unsupervised object discovery and tracking methods, using challenging videos from CATER and KITTI, and show strong improvements over the state-of-the-art.

This paper explores self-supervised learning of amodal 3D feature representations from RGB and RGB-D posed images and videos, agnostic to object and scene semantic content, and evaluates the resulting scene representations in the downstream tasks of visual correspondence, object tracking, and object detection. The model infers a latent 3D representation of the scene in the form of 3D feature points, where each continuous world 3D point is mapped to its corresponding feature vector. The model is trained for contrastive view prediction by rendering 3D feature clouds in queried viewpoints and matching against the 3D feature point cloud predicted from the query view. Notably, the representation can be queried for any 3D location, even if it is not visible from the input view. Our model brings together three powerful ideas of recent exciting research work: 3D feature grids as a neural bottleneck for view prediction, implicit functions for handling resolution limitations of 3D grids, and contrastive learning for unsupervised training of feature representations. We show the resulting 3D visual feature representations effectively scale across objects and scenes, imagine information occluded or missing from the input viewpoints, track objects over time, align semantically related objects in 3D, and improve 3D object detection. We outperform many existing state-of-the-art methods for 3D feature learning and view prediction, which are either limited by 3D grid spatial resolution, do not attempt to build amodal 3D representations, or do not handle combinatorial scene variability due to their non-convolutional bottlenecks.

We propose associating language utterances to 3D visual abstractions of the scene they describe. The 3D visual abstractions are encoded as 3-dimensional visual feature maps. We infer these 3D visual scene feature maps from RGB images of the scene via view prediction: when the generated 3D scene feature map is neurally projected from a camera viewpoint, it should match the corresponding RGB image. We present generative models that condition on the dependency tree of an utterance and generate a corresponding visual 3D feature map as well as reason about its plausibility, and detector models that condition on both the dependency tree of an utterance and a related image and localize the object referents in the 3D feature map inferred from the image. Our model outperforms models of language and vision that associate language with 2D CNN activations or 2D images by a large margin in a variety of tasks, such as, classifying plausibility of utterances, detecting referential expressions, and supplying rewards for trajectory optimization of object placement policies from language instructions. We perform numerous ablations and show the improved performance of our detectors is due to its better generalization across camera viewpoints and lack of object interferences in the inferred 3D feature space, and the improved performance of our generators is due to their ability to spatially reason about objects and their configurations in 3D when mapping from language to scenes.

A key promise of Virtual Reality (VR) is the possibility of remote social interaction that is more immersive than any prior telecommunication media. However, existing social VR experiences are mediated by inauthentic digital representations of the user (i.e., stylized avatars). These stylized representations have limited the adoption of social VR applications in precisely those cases where immersion is most necessary (e.g., professional interactions and intimate conversations). In this work, we present a bidirectional system that can animate avatar heads of both users' full likeness using consumer-friendly headset mounted cameras (HMC). There are two main challenges in doing this: unaccommodating camera views and the image-to-avatar domain gap. We address both challenges by leveraging constraints imposed by multiview geometry to establish precise image-to-avatar correspondence, which are then used to learn an end-to-end model for real-time tracking. We present designs for a training HMC, aimed at data-collection and model building, and a tracking HMC for use during interactions in VR. Correspondence between the avatar and the HMC-acquired images are automatically found through self-supervised multiview image translation, which does not require manual annotation or one-to-one correspondence between domains. We evaluate the system on a variety of users and demonstrate significant improvements over prior work.

Humans effortlessly "program" one another by communicating goals and desires in natural language. In contrast, humans program robotic behaviours by indicating desired object locations and poses to be achieved [5], by providing RGB images of goal configurations [19], or supplying a demonstration to be imitated [17]. None of these methods generalize across environment variations, and they convey the goal in awkward technical terms. This work proposes joint learning of natural language grounding and instructable behavioural policies reinforced by perceptual detectors of natural language expressions, grounded to the sensory inputs of the robotic agent. Our supervision is narrated visual demonstrations (NVD), which are visual demonstrations paired with verbal narration (as opposed to being silent). We introduce a dataset of NVD where teachers perform activities while describing them in detail. We map the teachers' descriptions to perceptual reward detectors, and use them to train corresponding behavioural policies in simulation. We empirically show that our instructable agents (i) learn visual reward detectors using a small number of examples by exploiting hard negative mined configurations from demonstration dynamics, (ii) develop pick-and-place policies using learned visual reward detectors, (iii) benefit from object-factorized state representations that mimic the syntactic structure of natural language goal expressions, and (iv) can execute behaviours that involve novel objects in novel locations at test time, instructed by natural language.

We introduce an approach to integrate segmentation information within a convolutional neural network (CNN). This counter-acts the tendency of CNNs to smooth information across regions and increases their spatial precision. To obtain segmentation information, we set up a CNN to provide an embedding space where region co-membership can be estimated based on Euclidean distance. We use these embeddings to compute a local attention mask relative to every neuron position. We incorporate such masks in CNNs and replace the convolution operation with a “segmentation-aware” variant that allows a neuron to selectively attend to inputs coming from its own region. We call the resulting network a segmentation-aware CNN because it adapts its filters at each image point according to local segmentation cues, while at the same time remaining fully-convolutional. We demonstrate the merit of our method on two widely different dense prediction tasks, that involve classification (semantic segmentation) and regression (optical flow). Our results show that in semantic segmentation we can replace DenseCRF inference with a cascade of segmentation-aware filters, and in optical flow we obtain clearly sharper responses than the ones obtained with comparable networks that do not use segmentation. In both cases segmentation-aware convolution yields systematic improvements over strong baselines.

Researchers have developed excellent feed-forward models that learn to map images to desired outputs, such as to the images’ latent factors, or to other images, using supervised learning. Learning such mappings from unlabelled data, or improving upon supervised models by exploiting unlabelled data, remains elusive. We argue that there are two important parts to learning without annotations: (i) matching the predictions to the input observations, and (ii) matching the predictions to known priors. We propose Adversarial Inverse Graphics networks (AIGNs): weakly supervised neural network models that combine feedback from rendering their predictions, with distribution matching between their predictions and a collection of ground-truth factors. We apply AIGNs to 3D human pose estimation and 3D structure and egomotion estimation, and outperform models supervised by only paired annotations. We further apply AIGNs to facial image transformation using super-resolution and inpainting renderers, while deliberately adding biases in the ground-truth datasets. Our model seamlessly incorporates such biases, rendering input faces towards young, old, feminine, masculine or Tom Cruiselike equivalents (depending on the chosen bias), or adding lip and nose augmentations while inpainting concealed lips and noses.

This paper presents a new state-of-the-art for document image classification and retrieval, using features learned by deep convolutional neural networks (CNNs). In object and scene analysis, deep neural nets are capable of learning a hierarchical chain of abstraction from pixel inputs to concise and descriptive representations. The current work explores this capacity in the realm of document analysis, and confirms that this representation strategy is superior to a variety of popular handcrafted alternatives. Extensive experiments show that (i) features extracted from CNNs are robust to compression, (ii) CNNs trained on non-document images transfer well to document analysis tasks, and (iii) enforcing region-specific feature-learning is unnecessary given sufficient training data. This work also makes available a new labelled subset of the IIT-CDIP collection, containing 400,000 document images across 16 categories.