Research map / Detection and segmentation / Tracking
Cannot See the Forest for the Trees: Aggregating Multiple Viewpoints To Better Classify Objects in Videos
Sukjun Hwang, Miran Heo, Seoung Wug Oh, Seon Joo Kim
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 |
|---|---|---|---|
| Add tracklet-level set classification to QDTrack | QDTrack · TAO validation | TrackAP50: 15.8 -> 19.9 | Improved |
| Add tracklet-level set classification to QDTrack | QDTrack · TAO test | TrackAP50: 12.4 -> 15.7 | Improved |
| Add tracklet-level set classification to QDTrack | QDTrack · YouTube-VIS 2019 | AP: 34.4 -> 37.7 | Improved |
| Allow tracklets to mix multiple identities while still restricting them to one class (ablation) | Set classifier · TAO validation | TrackAP50: 17.7 -> 17.2 | Hurt |
| Allow multiple classes in multi-identity tracklets (ablation) | Set classifier · TAO validation | TrackAP50: 17.2 -> 19.9 | Improved |
| Increase generated tracklets per batch from 32 to 256 (ablation) | Set classifier · TAO validation | TrackAP50: 17.7 -> 19.9 | Improved |
| Add video-derived tracklet training to image-derived tracklet training (ablation) | Set classifier · TAO validation | TrackAP50: 17.1 -> 19.9 | Improved |
| Train on videos only instead of images and videos (ablation) | Set classifier · TAO validation | TrackAP50: 19.9 -> 6.5 | Hurt |
| Use tracklets of length [16, 32) rather than [8, 16) or [32, 64) (ablation) | Set classifier · TAO validation | TrackAP50: 19.0, 18.3 -> 19.9 | Improved |
| Use tail-favoring RoI sampling with p=0.5 rather than uniform sampling (ablation) | Set classifier · TAO validation | TrackAP50: 15.6 -> 19.9 | Improved |
| Add instance-wise loss (ablation) | Set classifier · TAO validation | TrackAP50: 18.7 -> 18.9 | Improved |
| Add clustering loss alongside instance-wise loss (ablation) | Set classifier · TAO validation | TrackAP50: 18.9 -> 19.9 | Improved |
Closest papers by meaning
- TAO: A Large-Scale Benchmark for Tracking Any Object (ECCV 2020 · similarity 0.90)
- TrackingNet: A Large-Scale Dataset and Benchmark for Object Tracking in the Wild (ECCV 2018 · similarity 0.89)
- Bridging Images and Videos: A Simple Learning Framework for Large Vocabulary Video Object Detection (ECCV 2022 · similarity 0.88)
- VastTrack: Vast Category Visual Object Tracking (NeurIPS 2024 · similarity 0.88)
- OVT-B: A New Large-Scale Benchmark for Open-Vocabulary Multi-Object Tracking (NeurIPS 2024 · similarity 0.86)
- Track-On: Transformer-based Online Point Tracking with Memory (ICLR 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.
