What, Where, and How: Probing Spatiotemporal Representations in Video Foundation Models

• URL: http://arxiv.org/abs/2609.01551v1
• Date: Tue, 01 Sep 2026 17:17:04 GMT
• Title: What, Where, and How: Probing Spatiotemporal Representations in Video Foundation Models
• Abstract summary: We leverage lightweight probes trained to discover three temporally grounded properties: (i) camera motion, understanding (ii) intuitive physics, and (iii) anomaly detection.<n>We find that temporal features from individual videos form smooth low-temporal trajectories in representation space, suggesting that camera motion is not only linearly decodable but also geometrically organized.
• Score: 15.808524963857119
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Self-supervised video foundation models learn rich spatiotemporal representations, yet it remains unclear what visual concepts these representations encode, where they emerge across transformer layers, and how they are geometrically organized. In this work, we tackle these three questions through a systematic layer-wise analysis of V-JEPA 2 and VideoMAE-v2. We leverage lightweight probes trained to discover three temporally grounded properties: (i) camera motion understanding, (ii) intuitive physics, and (iii) anomaly detection. Both models encode camera motion, with best results ($>90$ ROC AUC) emerging at 60-70% of network depth, and achieve moderate anomaly detection performance ($>60$ ROC AUC), but remain near chance on intuitive-physics tasks, suggesting a limited encoding of deeper physical reasoning. Beyond classification, we find that temporal features from individual videos form smooth low-dimensional trajectories in representation space, suggesting that camera motion is not only linearly decodable but also geometrically organized. Based on these results, we apply geometry-aware spline-based steering in the model's latent representations to interpolate camera motion, yielding steered videos with smoother trajectories and more coherent temporal progression than linear interpolation.

Related papers

• Neural Voxel Dynamics: Learning Implicit 3D Physics via Volumetric Feature Advection [0.14323566945483496]
  We present a self-dimensional framework for learning implicit 3D physical dynamics directly from video-derived supervisory signals.<n>We address this by shifting the predictive bottleneck from 2D image space to a lifted' 3D Volumetric Latent Space.<n>Unlike state-of-the-art hybrid models that rely on explicit classical simulators for training and/or inference, our architecture tracks material states implicitly within high-supervised V-JEPA features.
  arXiv  Detail & Related papers  (2026-06-24T22:06:35Z)
• RAYNOVA: Scale-Temporal Autoregressive World Modeling in Ray Space [51.441415833480505]
  RAYNOVA is a multiview world model for driving scenarios that employs a dual-causal autoregressive framework.<n>It constructs an isotropic-temporal representation across views, frames, and scales based on relative Plcker-ray positional encoding.
  arXiv  Detail & Related papers  (2026-02-24T08:41:40Z)
• Seeing World Dynamics in a Nutshell [132.79736435144403]
  NutWorld is a framework that transforms monocular videos into dynamic 3D representations in a single forward pass.<n>We demonstrate that NutWorld achieves high-fidelity video reconstruction quality while enabling downstream applications in real-time.
  arXiv  Detail & Related papers  (2025-02-05T18:59:52Z)
• GaussianVideo: Efficient Video Representation via Hierarchical Gaussian Splatting [28.981174430968643]
  We introduce a novel neural representation that combines 3D Gaussian splatting with continuous camera motion modeling.<n> Experimental results show that our hierarchical learning, combined with robust camera motion modeling, captures complex dynamic scenes with strong temporal consistency.<n>This memory-efficient approach achieves high-quality rendering at impressive speeds.
  arXiv  Detail & Related papers  (2025-01-08T19:01:12Z)
• Physical Informed Driving World Model [47.04423342994622]
  DrivePhysica is an innovative model designed to generate realistic driving videos that adhere to essential physical principles.<n>We achieve state-of-the-art performance in driving video generation quality (3.96 FID and 38.06 FVD on the Nuscenes dataset) and downstream perception tasks.
  arXiv  Detail & Related papers  (2024-12-11T14:29:35Z)
• MonST3R: A Simple Approach for Estimating Geometry in the Presence of Motion [118.74385965694694]
  We present Motion DUSt3R (MonST3R), a novel geometry-first approach that directly estimates per-timestep geometry from dynamic scenes.<n>By simply estimating a pointmap for each timestep, we can effectively adapt DUST3R's representation, previously only used for static scenes, to dynamic scenes.<n>We show that by posing the problem as a fine-tuning task, identifying several suitable datasets, and strategically training the model on this limited data, we can surprisingly enable the model to handle dynamics.
  arXiv  Detail & Related papers  (2024-10-04T18:00:07Z)
• Vision-based Discovery of Nonlinear Dynamics for 3D Moving Target [11.102585080028945]
  We propose a vision-based approach to automatically uncover governing equations of nonlinear dynamics for 3D moving targets via raw videos recorded by a set of cameras. This framework is capable of effectively handling the challenges associated with measurement data, e.g., noise in the video, imprecise tracking of the target that causes data missing, etc.
  arXiv  Detail & Related papers  (2024-04-27T11:13:55Z)
• Dynamic 3D Point Cloud Sequences as 2D Videos [81.46246338686478]
  3D point cloud sequences serve as one of the most common and practical representation modalities of real-world environments. We propose a novel generic representation called textitStructured Point Cloud Videos (SPCVs) SPCVs re-organizes a point cloud sequence as a 2D video with spatial smoothness and temporal consistency, where the pixel values correspond to the 3D coordinates of points.
  arXiv  Detail & Related papers  (2024-03-02T08:18:57Z)
• NeuPhysics: Editable Neural Geometry and Physics from Monocular Videos [82.74918564737591]
  We present a method for learning 3D geometry and physics parameters of a dynamic scene from only a monocular RGB video input. Experiments show that our method achieves superior mesh and video reconstruction of dynamic scenes compared to competing Neural Field approaches.
  arXiv  Detail & Related papers  (2022-10-22T04:57:55Z)
• Exploring Optical-Flow-Guided Motion and Detection-Based Appearance for Temporal Sentence Grounding [61.57847727651068]
  Temporal sentence grounding aims to localize a target segment in an untrimmed video semantically according to a given sentence query. Most previous works focus on learning frame-level features of each whole frame in the entire video, and directly match them with the textual information. We propose a novel Motion- and Appearance-guided 3D Semantic Reasoning Network (MA3SRN), which incorporates optical-flow-guided motion-aware, detection-based appearance-aware, and 3D-aware object-level features.
  arXiv  Detail & Related papers  (2022-03-06T13:57:09Z)
• Learning to Segment Rigid Motions from Two Frames [72.14906744113125]
  We propose a modular network, motivated by a geometric analysis of what independent object motions can be recovered from an egomotion field. It takes two consecutive frames as input and predicts segmentation masks for the background and multiple rigidly moving objects, which are then parameterized by 3D rigid transformations. Our method achieves state-of-the-art performance for rigid motion segmentation on KITTI and Sintel.
  arXiv  Detail & Related papers  (2021-01-11T04:20:30Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.