Research map / Detection and segmentation / CNN features
Leveraging Long-Range Temporal Relationships Between Proposals for Video Object Detection
Mykhailo Shvets, Wei Liu, Alexander C. Berg
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 the proposal-level temporal relation block (causal, K=2) to the single-frame detector | FPN detector · ImageNet VID val (ResNet-50) | mAP: 72.4 -> 78.4 | Improved |
| add zero-centered feature normalization (vs uncentered, 10-dim) (ablation) | relation block · ImageNet VID val (ResNet-50) | mAP: 77.7 -> 78.4 | Improved |
| add supervised graph loss on the relation matrix (ablation) | relation block · ImageNet VID val (ResNet-50) | mAP: 77.4 -> 78.4 | Improved |
| reduce graph embedding dimension from 1024 to 10 (uncentered) (ablation) | relation block · ImageNet VID val (ResNet-50) | mAP: 77.5 -> 77.7 | No clear effect |
| use two relation blocks (fc6->{fc6,fc7}) instead of one at fc7->fc7 (ablation) | prediction head · ImageNet VID val (ResNet-50) | mAP: 77.5 -> 78.4 | Improved |
| add second relation block on top of single fc7 block (ablation) | causal model · ImageNet VID val (ResNet-101) | mAP: 80.6 -> 80.7 | Improved |
| add multi-head aggregation in relation block (ablation) | relation block · ImageNet VID val | mAP | No clear effect |
| add feature normalization, graph loss and larger embedding to Non-Local style block (ablation) | relation block · ImageNet VID val (ResNet-101) | mAP: 79.0 -> 80.7 | Improved |
| aggregate over multiple temporal strides at inference (ablation) | relation block · ImageNet VID val | mAP | Improved |
| use symmetric mode (past and future frames) instead of causal (ablation) | relation block · ImageNet VID val (ResNet-101) | mAP: 80.7 -> 81.0 | Improved |
| use frozen support model instead of single model (ablation) | relation block · ImageNet VID val (ResNet-101) | mAP: 80.7 -> 80.6 | No clear effect |
| pretrain the full system on COCO (ablation) | causal ResNet-50 model · ImageNet VID val | mAP: 78.4 -> 78.9 | Improved |
Closest papers by meaning
- Mining Inter-Video Proposal Relations for Video Object Detection (ECCV 2020 · similarity 0.89)
- Video Object Detection via Object-level Temporal Aggregation (ECCV 2020 · similarity 0.89)
- Relation Distillation Networks for Video Object Detection (ICCV 2019 · similarity 0.88)
- Objects Do Not Disappear: Video Object Detection by Single-Frame Object Location Anticipation (ICCV 2023 · similarity 0.88)
- Temporal Object-Aware Vision Transformer for Few-Shot Video Object Detection (AAAI 2026 · similarity 0.87)
- Context Enhanced Transformer for Single Image Object Detection in Video Data (AAAI 2024 · 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.
