Research map / Detection and segmentation / CNN features
Video Object Detection via Object-level Temporal Aggregation
Chun-Han Yao, Chen Fang, Xiaohui Shen, Yangyue Wan, Ming-Hsuan Yang
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 confidence reassignment (temporal max) (ablation) | YOLOv3-SPP + KCF · ImageNet VID val (YOLOv3-SPP + KCF, Dbase=1) | mAP: 58.6 -> 62.1 | Improved |
| add class probability aggregation (gamma=1.0) to confidence reassignment (ablation) | YOLOv3-SPP + KCF · ImageNet VID val (YOLOv3-SPP + KCF, Dbase=1) | mAP: 62.1 -> 64.4 | Improved |
| raise gamma from 0.6 to 1.0 in class probability aggregation (ablation) | YOLOv3-SPP + KCF · ImageNet VID val (YOLOv3-SPP + KCF, Dbase=1) | mAP: 63.8 -> 64.4 | Improved |
| add box coordinate refinement (ablation) | YOLOv3-SPP + KCF · ImageNet VID val (YOLOv3-SPP + KCF, Dbase=1) | mAP: 64.4 -> 64.8 | Improved |
| replace heuristic aggregation with LSTM-based aggregation (ablation) | YOLOv3-SPP + KCF · ImageNet VID val (YOLOv3-SPP + KCF, Dbase=1) | mAP: 64.8 -> 65.8 | Improved |
| fine-tune detector on VID+DET+COCO instead of COCO only (ablation) | YOLOv3-SPP + KCF + LSTM · ImageNet VID val (14 COCO classes, Dbase=1) | mAP: 62.6 -> 66.6 | Improved |
| train aggregation without fine-tuning detector on VID (ablation) | CenterNet + KCF + LSTM · ImageNet VID val (14 COCO classes, Dbase=1) | mAP gain: +3.7 -> +4.3 | No clear effect |
| use RL keyframe scheduler instead of fixed interval (ablation) | YOLOv3-SPP + KCF · ImageNet VID val (LSTM aggregation, Dbase=7) | mAP: 60.6 -> 63.2 | Improved |
| use heuristic keyframe scheduler instead of fixed interval (ablation) | YOLOv3-SPP + KCF · ImageNet VID val (LSTM aggregation, Dbase=7) | mAP: 60.6 -> 61.2 | Improved |
| use RL scheduler instead of fixed interval (cost in speed) (ablation) | YOLOv3-SPP + KCF · ImageNet VID val (LSTM aggregation, Dbase=7) | speed (fps): 37.1 -> 31.4 | Hurt |
| replace KCF with SiamFC tracker (ablation) | YOLOv3-SPP + RL · ImageNet VID val (LSTM aggregation, Dbase=7) | mAP: 63.2 -> 63.7 | Improved |
| replace KCF with SiamFC tracker (ablation) | YOLOv3-SPP + RL · ImageNet VID val (LSTM aggregation, Dbase=7) | speed (fps): 31.4 -> 20.2 | Hurt |
Closest papers by meaning
- Practical Video Object Detection via Feature Selection and Aggregation (IJCV 2026 · similarity 0.90)
- Temporal Context Enhanced Feature Aggregation for Video Object Detection (AAAI 2020 · similarity 0.90)
- Flow-Guided Feature Aggregation for Video Object Detection (ICCV 2017 · similarity 0.90)
- Leveraging Long-Range Temporal Relationships Between Proposals for Video Object Detection (ICCV 2019 · similarity 0.89)
- Feature Aggregated Queries for Transformer-Based Video Object Detectors (CVPR 2023 · similarity 0.88)
- YOLOV: Making Still Image Object Detectors Great at Video Object Detection (AAAI 2023 · similarity 0.88)
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.
