Research map / Detection and segmentation / CNN features
CenterNet Heatmap Propagation for Real-time Video Object Detection
Zhujun Xu, Emir Hrustic, Damien Vivet
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 heatmap propagation (HP) | CenterNet ResNet-101 · ImageNet VID val | mAP50 (%): 73.6 -> 76.7 | Improved |
| add heatmap propagation (HP) with Seq-NMS post-processing | CenterNet ResNet-101 · ImageNet VID val | mAP50 (%): 75.9 -> 78.4 | Improved |
| add heatmap propagation (HP) | CenterNet DLA-34 · ImageNet VID val | mAP50 (%): 69.1 -> 71.3 | Improved |
| add simple interpolation between consecutive-frame outputs | CenterNet · ImageNet VID val | mAP50 (%): 73.6 -> 73.6 | No clear effect |
| add HP (cost: extra runtime) | CenterNet · ImageNet VID val | FPS: 47 -> 37 | Hurt |
| add propagation dilation, P from 0 to 2 (ablation) | CenterNet+HP · ImageNet VID val | mAP50 (%): 75 -> 76.7 | Improved |
| increase dilation size P from 2 to 5 (ablation) | CenterNet+HP · ImageNet VID val | mAP50 (%): 76.7 -> 75.5 | Hurt |
| use larger P for fast-motion objects (P=5 vs P=2) (ablation) | CenterNet+HP · ImageNet VID fast motion | AP (%): 67.6 -> 68.6 | Improved |
| add balance with network heatmap (alpha 0.98 instead of 1) (ablation) | CenterNet+HP · ImageNet VID val | mAP50 (%): 75.3 -> 76.7 | Improved |
| increase alpha from 0.5 to 0.98 (ablation) | CenterNet+HP · ImageNet VID val | mAP50 (%): 74.6 -> 76.7 | Improved |
| add tracklet-length weighting with beta=2 instead of beta=1000 (no length weighting limit) (ablation) | CenterNet+HP · ImageNet VID val | mAP50 (%): 75 -> 76.7 | Improved |
| increase beta from 2 to 100 (ablation) | CenterNet+HP · ImageNet VID val | mAP50 (%): 76.7 -> 75.1 | Hurt |
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.
- R-FCN — R-FCN: Object Detection via Region-based Fully Convolutional Networks (NeurIPS 2016) · better on 1 of 1 shared result
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
- CenterNet++ for Object Detection (TPAMI 2023 · similarity 0.87)
- Efficient One-Stage Video Object Detection by Exploiting Temporal Consistency (ECCV 2022 · similarity 0.86)
- Optimizing Video Object Detection via a Scale-Time Lattice (CVPR 2018 · similarity 0.86)
- CenterNet: Keypoint Triplets for Object Detection (ICCV 2019 · similarity 0.86)
- Tracking Objects as Points (ECCV 2020 · similarity 0.85)
- Object Detection using Event Camera: A MoE Heat Conduction based Detector and A New Benchmark Dataset (CVPR 2025 · similarity 0.84)
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.
