Research map / Detection and segmentation / CNN features
Crafting GBD-Net for Object Detection
Xingyu Zeng, Wanli Ouyang, Junjie Yan, Hongsheng Li, Tong Xiao, Kun Wang, Yu Liu, Yucong Zhou, Bin Yang, Zhe Wang, Hui Zhou, Xiaogang Wang
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 GBD message passing to BN-net on VOC07 | BN+FRCN · PASCAL VOC2007 test | mAP: 73.1 -> 77.2 | Improved |
| add GBD structure to BN+FRCN on COCO | BN+FRCN · MS COCO closed-test | AP: 24.4 -> 27.0 | Improved |
| add gate functions to message passing layers (random init) (ablation) | GBD-Net · ImageNet val2 | mAP: 42.3 -> 48.9 | Improved |
| place roi-pooling on raw pixels instead of after 6th inception module (ablation) | GBD-Net-v1 · ImageNet val2 | mAP: 48.9 -> 51.4 | Improved |
| enlarge single support region from padding 0.2 to 1.7 (ablation) | BN-net · ImageNet val2 | mAP: 46.3 -> 45.2 | Hurt |
| add more resolutions/contexts to GBD-Net (one to four regions) (ablation) | GBD-Net-v1 · ImageNet val2 | mAP: 47.4 -> 48.0 -> 48.9 | Improved |
| combine GBD-Net with multi-region approach by score averaging | GBD-Net BN-net · ImageNet val2 | mAP: 48.9 -> 51.2 | Improved |
| use GBD-v2 instead of GBD-v1 | BN-net · ImageNet val2 | mAP: +2.6 -> +3.7 | Improved |
| scale messages by beta=0.1 instead of 1 (ablation) | GBD BN-net · ImageNet val2 | mAP: 52 -> 53.6 | Improved |
| add new GBD to backbones | ResNet-152 / ResNet-269 · ImageNet val2 | mAP: 54 -> 56.5; 56.6 -> 58.8 | Improved |
| object-centric pretraining without roi-pooling vs image-centric (ablation) | BN-net · ImageNet val2 | mAP: 49.1 -> 48.4 | Hurt |
| add roi-pooling to object-centric pretraining (ablation) | BN-net · ImageNet val2 | mAP: 48.4 -> 49.9 | 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 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
- Gated Bi-directional CNN for Object Detection (ECCV 2016 · similarity 0.87)
- G-CNN: An Iterative Grid Based Object Detector (CVPR 2016 · similarity 0.86)
- R-FCN: Object Detection via Region-based Fully Convolutional Networks (NeurIPS 2016 · similarity 0.86)
- Feature Pyramid Networks for Object Detection (CVPR 2017 · similarity 0.86)
- Parallel Feature Pyramid Network for Object Detection (ECCV 2018 · similarity 0.86)
- Object Detection Networks on Convolutional Feature Maps (TPAMI 2016 · 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.
