Research map / Detection and segmentation / Open-set and few-shot detection
MOD-UV: Learning Mobile Object Detectors from Unlabeled Videos
Yihong Sun, Bharath Hariharan
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 |
|---|---|---|---|
| add Moving2Mobile self-training stage (ablation) | MOD-UV‡ · Waymo Open val | Box AR / AP: AR 12.6 -> 17.5; AP 7.0 -> 10.9 | Improved |
| add Large2Small small-object detector (ablation) | MOD-UV‡ · Waymo Open val | Box AR_S / AP_S: AR_S reduces nearly 4x and AP_S 3x when removed; AR 12.5 -> 17.5 | Improved |
| add final self-training round on aggregated proposals (ablation) | MOD-UV‡ · Waymo Open val | Box AR: +8.1 vs large-only, +4.1 vs small-only detector | Improved |
| cluster pseudo-3D points from monocular depth instead of 2D contours for pseudo-labels (ablation) | motion-segmentation pseudo-labels · Waymo train | Moving Mask AR_L: +9.4 Moving AR_L, +2.5 Moving AR | Improved |
| train Moving2Mobile beyond 3 epochs (ablation) | MOD-UV‡ · Waymo train pseudo-labels | Mask AR (all): 5.3 (3 epochs) -> 4.3 (5 epochs) -> 4.1 (20 epochs) | Hurt |
| Moving2Mobile alone for small objects (ablation) | MOD-UV‡ · Waymo train pseudo-labels | AR_S: 0.0 -> 0.0 | No clear effect |
| add Large2Small to CutLER (no motion cues) | CutLER · Waymo Open val | Box AR: 11.7 -> 15.0 | Improved |
| replace NMS with proposed mask aggregation of large and small detector outputs (ablation) | MOD-UV · Waymo Open val | AR0.5 / AP50: +1.8 AR0.5, +1.5 AP50 | Improved |
| use motion-based pseudo-labels instead of image-feature-only (CutLER^L2S) | MOD-UV · Waymo Open val | Box AP: 4.3 -> 12.6 | Improved |
| train HASSOD only on Waymo instead of COCO | HASSOD · Waymo Open val | Box AR: 12.7 -> 8.3 | Hurt |
| ImageNet-initialized backbone instead of Waymo-initialized | MOD-UV · nuScenes zero-shot | Box AR: 17.3 -> 21.9 | Improved |
| apply Waymo-trained MOD-UV to COCO (domain shift) | MOD-UV · COCO val2017 ground vehicles | Box AR: 26.6 vs HASSOD 27.9 | Mixed |
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.
- HASSOD — HASSOD: Hierarchical Adaptive Self-Supervised Object Detection (NeurIPS 2023) · better on 5 of 6 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
- Learning Features by Watching Objects Move (CVPR 2017 · similarity 0.88)
- Discovering Objects That Can Move (CVPR 2022 · similarity 0.88)
- Track, Check, Repeat: An EM Approach to Unsupervised Tracking (CVPR 2021 · similarity 0.87)
- MOVE: Unsupervised Movable Object Segmentation and Detection (NeurIPS 2022 · similarity 0.87)
- Exploiting Spatial Invariance for Scalable Unsupervised Object Tracking (AAAI 2020 · similarity 0.87)
- Unidentified Video Objects: A Benchmark for Dense, Open-World Segmentation (ICCV 2021 · similarity 0.87)
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.
