Research map / Detection and segmentation / CNN features
Objects Do Not Disappear: Video Object Detection by Single-Frame Object Location Anticipation
Xin Liu, Fatemeh Karimi Nejadasl, Jan C. van Gemert, Olaf Booij, Silvia L. Pintea
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 motion anticipation supervision with simulated smooth motion (ablation) | static detector on MovingDigits · MovingDigits keyframe | keyframe mAP (%): 73.51 -> 76.57 | Improved |
| supervise with randomized (non-smooth) box positions between keyframes (ablation) | static detector on MovingDigits · MovingDigits keyframe | keyframe mAP (%): 73.51 -> 62.68 | Hurt |
| use annotated ground-truth motion instead of simulated motion (ablation) | trajectory anticipation · MovingDigits keyframe | keyframe mAP (%): 76.57 -> 79.31 | Improved |
| use trajectory loss L_traj instead of bag loss L_bag (ablation) | trajectory network · ImageNet VID subset, T=8 | mAP (%): 83.97 -> 86.15 | Improved |
| predict offsets between neighboring boxes instead of box coordinates (ablation) | trajectory network · ImageNet VID subset | mAP (%): +0.81% (T=4), +0.89% (T=8) | Improved |
| use only the L_bag(delta) term without L_sum (ablation) | L_traj · ImageNet VID subset, T=8 | mAP (%): 86.15 -> 79.73 | Hurt |
| use only L_sum without L_bag(delta) (ablation) | L_traj · ImageNet VID subset, T=4 | mAP (%): 89.20 -> 87.95 | Hurt |
| sparse annotation loss with interpolated trajectories instead of full supervision (ablation) | L_traj · ImageNet VID subset keyframes, T=4 | keyframe mAP (%): 91.03 -> 90.76 | No clear effect |
| use only the offset term in the sparse annotation loss (ablation) | L_traj^(sa) · ImageNet VID subset keyframes, T=4 | keyframe mAP (%): 90.76 -> 74.61 | Hurt |
| predict longer trajectories (sample fewer keyframes) (ablation) | full method · ImageNet VID subset | mAP vs FPS: mAP 89.2% at 39.6 FPS; noticeable drop beyond 10 frames | Mixed |
| predict trajectories from keyframes instead of per-frame detection on ImageNet VID | Faster RCNN R101 · ImageNet VID val | mAP (%): 73.6 -> 87.2 | Improved |
| add sparse-annotation trajectory loss with interpolated pseudo trajectories | Faster RCNN · Waymo Open AP/L1 | AP/L1 (%): 55.66 -> 64.53 | 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.
- Faster-RCNN — Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks (TPAMI 2016) · better on 5 of 5 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
- Minimum Delay Moving Object Detection (CVPR 2017 · similarity 0.89)
- Learning Motion Patterns in Videos (CVPR 2017 · similarity 0.88)
- Leveraging Long-Range Temporal Relationships Between Proposals for Video Object Detection (ICCV 2019 · similarity 0.88)
- Video Object Detection via Object-level Temporal Aggregation (ECCV 2020 · similarity 0.88)
- Beyond Local Search: Tracking Objects Everywhere With Instance-Specific Proposals (CVPR 2016 · similarity 0.88)
- TrackIME: Enhanced Video Point Tracking via Instance Motion Estimation (NeurIPS 2024 · 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.
