Research map / Detection and segmentation / CNN features
DSFD: Dual Shot Face Detector
Check your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 FEM to FSSD with VGG16 (ablation) | FSSD+VGG16 · WIDER FACE validation Easy, Medium, Hard | AP: 92.6% -> 93.0%; 90.2% -> 91.4%; 79.1% -> 84.6% | Helped |
| Add FEM to FSSD with ResNet50 (ablation) | FSSD+ResNet50 · WIDER FACE validation Easy, Medium, Hard | AP: 93.7% -> 95.0%; 92.2% -> 94.1%; 81.8% -> 88.0% | Helped |
| Add PAL to FSSD+FEM with ResNet50 (ablation) | FSSD+ResNet50+FEM · WIDER FACE validation Easy, Medium, Hard | AP: 95.0% -> 95.3%; 94.1% -> 94.4%; 88.0% -> 88.6% | Helped |
| Add IAM to FSSD+FEM with ResNet101 (ablation) | FSSD+ResNet101+FEM · WIDER FACE validation Easy, Medium, Hard | AP: 95.8% -> 96.1%; 95.1% -> 95.2%; 89.7% -> 90.0% | Helped |
| Add PAL after FEM and IAM with ResNet101 (ablation) | FSSD+ResNet101+FEM+IAM · WIDER FACE validation Easy, Medium, Hard | AP: 96.1% -> 96.3%; 95.2% -> 95.4%; 90.0% -> 90.1% | Helped |
| Replace ResNet101 with ResNet152 (ablation) | FSSD+FEM+IAM+PAL · WIDER FACE validation Easy, Medium, Hard | AP: 96.3% -> 96.6%; 95.4% -> 95.7%; 90.1% -> 90.4% | Helped |
| Increase training batch size (ablation) | DSFD with ResNet152 · WIDER FACE validation Hard | AP: 90.4% -> 91.2% | Helped |
| Replace FEM with RFB (ablation) | DSFD with ResNet101 · WIDER FACE Easy, Medium, Hard | AP: 96.3 -> 96.0; 95.4 -> 94.5; 90.1 -> 87.2 | Hurt |
| Use a more complex or higher ImageNet Top-1-accuracy backbone (ablation) | DSFD · WIDER FACE validation Easy, Medium, Hard | AP: ResNet101: 96.3%, 95.4%, 90.1%; DPN98: 96.3%, 95.5%, 90.4%; SE-ResNeXt101: 95.7%, 94.8%, 88.9% | Mixed |
| Use IAM anchor-based face sampling and matching (ablation) | DSFD anchor matching · WIDER FACE training matched anchors | average matched anchors per face: 6.4 to about 6.9 | Helped |
Closest papers by meaning
- Scale-Aware Face Detection (CVPR 2017 · similarity 0.84)
- DamoFD: Digging into Backbone Design on Face Detection (ICLR 2023 · similarity 0.84)
- Selective Refinement Network for High Performance Face Detection (AAAI 2019 · similarity 0.84)
- FA-RPN: Floating Region Proposals for Face Detection (CVPR 2019 · similarity 0.82)
- Seeing Small Faces From Robust Anchor's Perspective (CVPR 2018 · similarity 0.82)
- DFD: Distilling the Feature Disparity Differently for Detectors (ICML 2024 · similarity 0.80)
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.