Facial expression recognition with residual masking

What it does Implements the Residual Masking Network architecture from an ICPR 2020 paper for facial expression recognition. The repository bundles training code for FER2013 and ImageNet, a suite of baseline models (VGG19, ResNet variants, CBAM), and a PyPI-installable library that wraps inference into a few lines of Python using OpenCV.

The interesting bit Despite being pitched for faces, the authors benchmark the architecture on ImageNet (74.16% top-1), suggesting the masking approach generalizes beyond emotion datasets. The published rmn package lowers the barrier to entry: instantiate RMN() and run detect_emotion_for_single_frame() or video_demo() on a webcam feed.

Key highlights

• Single model hits 74.14% on FER2013; an ensemble of seven models pushes that to 76.82%
• Ships as a PyPI package (rmn) with a minimalist API for image and webcam inference
• Includes pre-trained weights, Colab notebooks, and training scripts for both FER2013 and ImageNet
• Bundles extensive baselines (VGG19, ResNet variants, CBAM, DenseNet) for comparison
• ICPR 2020 paper with dissertation and presentation materials included

Caveats

• The ensemble workflow is manual: download weights, place them in ./saved/checkpoints/, edit gen_results, and run scripts for each model before fusing
• Core documentation (the dissertation) is in Vietnamese; English speakers get the paper and slides only
• The training script defaults to AlexNet unless you edit the config, which is an odd choice for a residual masking repo

Verdict Worth a look if you need off-the-shelf emotion detection in Python or want to study the masking architecture on small image benchmarks. Skip it if you need a production-grade face-analysis pipeline—this is research code with a convenient wrapper.