Research map / Video / Video segmentation and tracking
Tracking by Predicting 3-D Gaussians Over Time
Tanish Baranwal, Himanshu Gaurav Singh, Jathushan Rajasegaran, Jitendra Malik
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| integrate mean (motion) deltas during pretraining (ablation) | Video-GMAE large · TAP-Vid Davis (large, finetuned) | AJ: 39.1 -> 44.4 | Improved |
| add RGB deltas on top of mean deltas (ablation) | Video-GMAE large · TAP-Vid Davis (large, finetuned) | AJ: 44.4 -> 44.7 | No clear effect |
| integrate RGB deltas only (ablation) | Video-GMAE large · TAP-Vid Davis (large, finetuned) | AJ: 39.1 -> 42.5 | Improved |
| pretrain on 24 frames instead of 2 (stride = frames) (ablation) | Video-GMAE base · TAP-Vid Davis (base, finetuned), stride = frames | AJ: 55.4 -> 24.7 | Hurt |
| pretrain on 4 or 8 frames vs 2, evaluated at stride 2 (ablation) | Video-GMAE base · TAP-Vid Davis (base, finetuned), stride 2 | AJ: 55.4 -> 64.2 (4), 63.5 (8); 57.5 (16), 52.4 (24) | Mixed |
| correspondence-aware Gaussian decoder instead of patch reconstruction | MAE-ST/VideoMAE · TAP-Vid Kinetics (frozen encoder) | AJ: 46.9 -> 65.1 (vs VideoMAE) | Improved |
| finetune encoder instead of freezing (ablation) | Video-GMAE large · TAP-Vid Kinetics | AJ: 65.1 -> 75.1 | Improved |
| scale base to large (frozen encoder) (ablation) | Video-GMAE · TAP-Vid Davis | AJ: 45.2 -> 46.7 | Improved |
| fine-tuned Video-GMAE vs supervised trackers | Video-GMAE large finetune · TAP-Vid Davis | AJ: 57.9 vs CoTracker3 63.8 | Hurt |
| zeroshot Gaussian tracking vs GMRW-C | Video-GMAE zeroshot · TAP-Vid Kinetics | AJ: 31.9 -> 60.1 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- VideoMAE — VideoMAE: Masked Autoencoders are Data-Efficient Learners for Self-Supervised Video Pre-Training (NeurIPS 2022) · better on 3 of 3 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Splatter a Video: Video Gaussian Representation for Versatile Processing (NeurIPS 2024 · similarity 0.88)
- TrackMAE: Video Representation Learning via Track Mask and Predict (CVPR 2026 · similarity 0.88)
- GaussianVideo: Efficient Video Representation via Hierarchical Gaussian Splatting (ICCV 2025 · similarity 0.88)
- Video Diffusion Models Excel at Tracking Similar-Looking Objects Without Supervision (NeurIPS 2025 · similarity 0.87)
- Autoregressive Sequential Pretraining for Visual Tracking (CVPR 2025 · similarity 0.86)
- Emergent Temporal Correspondences from Video Diffusion Transformers (NeurIPS 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
