Papers › PPL: Pairwise Prototype Learning for Masked Face Recognition
PPL: Pairwise Prototype Learning for Masked Face Recognition
Kim, Minsoo; Nam, Gi Pyo; Hong, Yu-Jin; Kim, Ig-Jae
The occlusion caused by a facemask emerged as a new challenge in face recognition as the pandemic of COVID-19 made wearing a facemask an everyday practice. The recognition performance of the previous approaches degraded in recognising a face with a facemask since the models were trained to extract features from the overall face and not many face samples with a facemask were available. Although the previously proposed method of using both face data without facemasks and those with synthesised facemasks for training improved the recognition performance, a certain drop in the recognition performance for faces without facemasks was observed. Thus, we propose an approach that can achieve robust recognition of faces with facemasks without compromising that of faces without facemasks. This study broke free from using a single prototype and designed Pairwise Prototype Learning (PPL) which separated the prototype depending on the facemask condition of the face data. Models trained with the proposed PPL method outperformed those trained with previously suggested methods in recognising both faces with and without facemasks. On top of presenting a new MFW+ dataset for masked face recognition benchmark, our study found a simple and intuitive way to improve recognition performance on all benchmarks, overcoming the limitation of using a single prototype in face recognition for faces with facemasks. All codes of PPL are available at https://github.com/kim1102/PPL-MFR.
Code
Repository list and official/mentioned flags are the archive's, frozen 2025-07-28. Reachability, where shown, is from one Syntology probe window (2026-09-16 to 2026-09-18); repositories not probed show nothing. GitHub stars are not tracked.
Code Syntology ran Syntology
Not run by Syntology. Nothing on this page verifies that the listed code works.
Tasks
Datasets
Introduced by this paper, per the archive.
Results from the paper archive 2025-07-28
| Task | Dataset | Model | Metric | Value | Rank at snapshot | Leaderboard | Report |
|---|---|---|---|---|---|---|---|
| Face Recognition | MFW+ (M-M) | ArcFace + PPL | TAR@FAR=0.0001 | 78.80 | #1 of 3 | Archive leaderboard | report |
| Face Recognition | MFW+ (M-M) | MaskInv-HG | TAR@FAR=0.0001 | 78.36 | #2 of 3 | Archive leaderboard | report |
| Face Recognition | MFW+ (M-M) | FocusFace | TAR@FAR=0.0001 | 77.77 | #3 of 3 | Archive leaderboard | report |
| Face Recognition | MFW+ (U-M) | AdaFace+PPL | TAR@FAR=0.0001 | 84.27 | #1 of 3 | Archive leaderboard | report |
| Face Recognition | MFW+ (U-M) | MaskInv-HG | TAR@FAR=0.0001 | 83.82 | #2 of 3 | Archive leaderboard | report |
| Face Recognition | MFW+ (U-M) | FocusFace | TAR@FAR=0.0001 | 83.40 | #3 of 3 | Archive leaderboard | report |
Ranks are positions in the archive's leaderboards as they stood at the 2025-07-28 snapshot. Results published since then are not among these rows, so a rank here is not a current standing.
Report a problem or propose a change · a person checks every report against the paper or source before anything changes; decisions are listed on /corrections