Research map / Label-efficient and robust learning / Recognition and long-tail
Occlusion Robust Face Recognition Based on Mask Learning With Pairwise Differential Siamese Network
Lingxue Song, Dihong Gong, Zhifeng Li, Changsong Liu, Wei Liu
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 differential input and pairwise loss to mask learning (ablation) | PDSN · AR sunglasses (Protocol 2) | rank-1 identification accuracy (%): 95.97 -> 98.19 | Improved |
| add differential input and pairwise loss to mask learning (ablation) | PDSN · AR scarf (Protocol 2) | rank-1 identification accuracy (%): 97.92 -> 98.33 | Improved |
| add differential input and pairwise loss to mask learning (ablation) | PDSN · occluded Facescrub MF1occ | accuracy (%): 54.80 -> 56.34 | Improved |
| binarize mask instead of using soft weights (ablation) | FDM · AR sunglasses (Protocol 2) | rank-1 identification accuracy (%): 96.67 -> 98.19 | Improved |
| binarize mask instead of using soft weights (ablation) | FDM · AR scarf (Protocol 2) | rank-1 identification accuracy (%): 97.22 -> 98.33 | Improved |
| zero low mask values but keep the rest unchanged (Soft+Binary) instead of fully binary (ablation) | FDM · AR scarf (Protocol 2) | rank-1 identification accuracy (%): 98.33 -> 99.03 | No clear effect |
| increase discarding threshold tau from 0 to 0.25 (ablation) | FDM · AR sunglasses+scarf | rank-1 identification accuracy (%): 95.84 -> 98.26 | Improved |
| increase discarding threshold tau from 0.25 to 0.45 (ablation) | FDM · AR sunglasses+scarf | rank-1 identification accuracy (%): 98.26 -> 97.92 | Hurt |
| fine-tune trunk CNN on synthetic occluded faces (Baseline) | Trunk CNN · LFW | verification accuracy (%): 99.20 -> 98.68 | Hurt |
| fine-tune trunk CNN on synthetic occluded faces (Baseline) | Trunk CNN · MegaFace Challenge 1 (small) | rank-1 identification accuracy (%): 74.40 -> 68.81 | Hurt |
| add PDSN feature discarding masks to trunk CNN | Trunk CNN · occluded Facescrub MF1occ | accuracy (%): 51.86 -> 56.34 | Improved |
| add PDSN feature discarding masks to trunk CNN | Trunk CNN · AR sunglasses (Protocol 2) | rank-1 identification accuracy (%): 95.14 -> 98.19 | Improved |
Closest papers by meaning
- End2End Occluded Face Recognition by Masking Corrupted Features (TPAMI 2021 · similarity 0.91)
- MSML: Enhancing Occlusion-Robustness by Multi-Scale Segmentation-Based Mask Learning for Face Recognition (AAAI 2022 · similarity 0.88)
- Robust Facial Landmark Detection via Occlusion-Adaptive Deep Networks (CVPR 2019 · similarity 0.88)
- Masked Face Recognition with Generative-to-Discriminative Representations (ICML 2024 · similarity 0.87)
- Occlusion-Free Face Alignment: Deep Regression Networks Coupled With De-Corrupt AutoEncoders (CVPR 2016 · similarity 0.86)
- Detecting Masked Faces in the Wild With LLE-CNNs (CVPR 2017 · 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.
