Research map / Detection and segmentation / CNN features
DeepProposals: Hunting Objects and Actions by Cascading Deep Convolutional Layers
Amir Ghodrati, Ali Diba, Marco Pedersoli, Tinne Tuytelaars, Luc Van Gool
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 a 2x2 spatial pyramid to average-pooled window features (ablation) | single-layer sliding window proposals · PASCAL VOC 2007 recall | recall: gain between 2 and 4 percent in recall | Improved |
| use 4x4 instead of 2x2 pyramid (ablation) | window representation · PASCAL VOC 2007 recall | recall vs cost: 5 times slower for 2x2 vs 21 times slower for 4x4 | Mixed |
| add stage 2 re-scoring on layer 3 with a 2x2 pyramid after layer 5 stage (ablation) | DeepProposals cascade · PASCAL VOC 2007 recall | recall: improves the recall of the cascade by 5% | Improved |
| add layer 4 to the cascade (ablation) | DeepProposals cascade · PASCAL VOC 2007 Recall@1000 | recall: 79% -> 79% | No clear effect |
| use layer 2 instead of layer 1 for contour refinement (ablation) | stage 3 contour detector · BSDS | AP: 0.61 -> 0.72 | Improved |
| add stage 3 local refinement on layer 2 contours (ablation) | DeepProposals cascade · PASCAL VOC 2007 high IoU | recall | Improved |
| use layer 3 features instead of layer 5 (ablation) | single-layer sliding window · PASCAL VOC 2007 IoU 0.5 recall@1000 | recall: 88% -> 92% | Improved |
| use layer 1 instead of layer 3 (ablation) | single-layer sliding window · PASCAL VOC 2007 IoU 0.5 recall@1000 | recall: 92% -> 73% | Hurt |
| increase number of window sizes (ablation) | sliding window proposals · PASCAL VOC 2007 | recall and cost | Mixed |
| increase windows per frame from 100 to 300 for action tubes (ablation) | action proposals · UCF-Sports | recall and Viterbi time: 1.3 and 12.1 seconds for N=100 and N=300 | Mixed |
| train proposal classifier on only 5 COCO classes instead of more classes (ablation) | DeepProposals · COCO recall at IoU 0.5, 1000 proposals | recall: 56% vs 59%-60% | Hurt |
| use DeepProposals instead of selective search proposals in detector with few proposals | Faster-RCNN detector · VOC 2007 test, 50 proposals | mAP: 28.1 -> superior to 50 | Improved |
Closest papers by meaning
- Proposal-Contrastive Pretraining for Object Detection from Fewer Data (ICLR 2023 · similarity 0.85)
- Rethinking the Faster R-CNN Architecture for Temporal Action Localization (CVPR 2018 · similarity 0.84)
- Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks (TPAMI 2016 · similarity 0.84)
- Feedback-Prop: Convolutional Neural Network Inference Under Partial Evidence (CVPR 2018 · similarity 0.84)
- Spatially Invariant Unsupervised Object Detection with Convolutional Neural Networks (AAAI 2019 · similarity 0.84)
- CRAFT Objects From Images (CVPR 2016 · 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.
