Research map / Video / Video segmentation and tracking
TrackVerse: A Large-Scale Object-Centric Video Dataset for Image-Level Representation Learning
Yibing Wei, Samuel Church, Victor Suciu, Jinhong Lin, Cheng-En Wu, Pedro Morgado
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 |
|---|---|---|---|
| pretrain on TrackVerse tracks with VA-MoCo instead of ImageNet-1K with MoCo (equal budget) | ViT-B/16 MoCo · TrackVerse FSL | accuracy (%): 47.7 -> 71.1 | Improved |
| pretrain on TrackVerse tracks instead of ImageNet-1K | ViT-B/16 MoCo · SSv2 Ft | top-1 accuracy (%): 56.8 -> 61.8 | Improved |
| use object-centric tracks instead of full scene clips (ablation) | VA-MoCo · LVIS-IN NN | accuracy (%): 35.3 -> 50.1 | Improved |
| use TrackVerse instead of TAO tracking dataset (ablation) | VA-MoCo · LVIS-IN NN | accuracy (%): 14.8 -> 50.1 | Improved |
| scale TrackVerse subset from 0.4M to 3.8M tracks (ablation) | VA-MoCo · TrackVerse FSL | accuracy (%): 55.3 -> 71.1 | Improved |
| add variance-aware predictor conditioning on augmentation parameters (ablation) | MoCo on TrackVerse (0.4M-CB1000) · TrackVerse NN | accuracy (%): 33.7 -> 39.9 | Improved |
| train longer (100 to 600 epochs) on dynamic tracks versus static frames (ablation) | MoCo on TrackVerse (0.4M-CB1000) · TrackVerse NN | accuracy (%): 35.96 -> 39.94 -> 40.43 | Improved |
| scale TrackVerse subset from 0.4M to 3.8M tracks (ablation) | VA-MoCo · DAVIS-2017 | J&Fm: 65.4 -> 67.1 | Improved |
| scale TrackVerse subset from 2.4M to 3.8M tracks (ablation) | VA-MoCo · MIT-States Pair Ft. | accuracy (%): 50.0 -> 50.0 | No clear effect |
| use ImageNet-1K MoCo versus early small TrackVerse subset (0.4M) on MIT-States Pair Ft. | ViT-B/16 MoCo · MIT-States Pair Ft. | accuracy (%): 46.8 -> 40.8 | Hurt |
Closest papers by meaning
- Benchmarking Unsupervised Object Representations for Video Sequences (JMLR 2021 · similarity 0.88)
- Tracktention: Leveraging Point Tracking to Attend Videos Faster and Better (CVPR 2025 · similarity 0.87)
- Video Diffusion Models Excel at Tracking Similar-Looking Objects Without Supervision (NeurIPS 2025 · similarity 0.86)
- DisMo: Disentangled Motion Representations for Open-World Motion Transfer (NeurIPS 2025 · similarity 0.86)
- Temporally Consistent Object-Centric Learning by Contrasting Slots (CVPR 2025 · similarity 0.86)
- TrackCraft3R: Repurposing Video Diffusion Transformers for Dense 3D Tracking (NeurIPS 2026 · similarity 0.85)
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.
