Research map / Detection and segmentation / CNN features
Supervised Transformer Network for Efficient Face Detection
Dong Chen, Gang Hua, Fang Wen, Jian Sun
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 multi-task RPN with facial landmarks (instead of direct rectangle regression) (ablation) | Supervised Transformer Network · FDDB | recall rate at 10 false alarms: 87.1 -> 89.6 | Improved |
| add the Supervised Transformer to RPN + feature combination (ablation) | Supervised Transformer Network · FDDB | recall rate at 10 false alarms: 88.8 -> 89.6 | Improved |
| add the Supervised Transformer to the multi-task RPN (ablation) | Supervised Transformer Network · FDDB | recall rate at 10 false alarms: 87.1 -> 88.3 | Improved |
| add multi-granularity feature combination (ablation) | Supervised Transformer Network · FDDB | recall rate at 10 false alarms: 88.3 -> 89.6 | Improved |
| add feature combination to the multi-task RPN (ablation) | Supervised Transformer Network · FDDB | recall rate at 10 false alarms: 87.1 -> 88.8 | Improved |
| add the multi-task RPN to feature combination alone (ablation) | Supervised Transformer Network · FDDB | recall rate at 10 false alarms: 88.0 -> 88.8 | Improved |
| replace NMS with non-top K suppression (K=3) (ablation) | Supervised Transformer Network · FDDB | recall rate: 88.7 -> 89.3 | Improved |
| increase K in non-top K suppression from 3 to 5 (ablation) | Supervised Transformer Network · FDDB | recall rate: 89.3 -> 89.4 | Improved |
| add ROI convolution with pre-filter threshold 0 | Supervised Transformer Network · FDDB (640x480) | total time and recall: 107.2 -> 55.3 ms; 89.3% -> 89.2% recall | Improved |
| add ROI convolution with pre-filter threshold 3 | Supervised Transformer Network · FDDB (640x480) | total time and recall: 107.2 -> 29.3 ms; 89.3% -> 88.1% recall | Mixed |
| raise pre-filter threshold from 2 to 3 | Supervised Transformer Network · FDDB (640x480) | total time and recall: 32.1 -> 29.3 ms; 88.7% -> 88.1% recall | Mixed |
| initialize canonical positions too large, too small or offset | Supervised Transformer Network · learned canonical positions | canonical point convergence | No clear effect |
Closest papers by meaning
- Detecting Faces Using Inside Cascaded Contextual CNN (ICCV 2017 · similarity 0.89)
- Joint Training of Cascaded CNN for Face Detection (CVPR 2016 · similarity 0.88)
- Mnemonic Descent Method: A Recurrent Process Applied for End-To-End Face Alignment (CVPR 2016 · similarity 0.88)
- Recursive Spatial Transformer (ReST) for Alignment-Free Face Recognition (ICCV 2017 · similarity 0.88)
- HyperFace: A Deep Multi-Task Learning Framework for Face Detection, Landmark Localization, Pose Estimation, and Gender Recognition (TPAMI 2017 · similarity 0.87)
- Recurrent Convolutional Shape Regression (TPAMI 2018 · similarity 0.87)
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.
