// the find
atulapra/Emotion-detection
Real-time Facial Emotion Detection using deep learning
A real-time webcam emotion classifier using a 4-layer CNN trained on FER-2013, with Haar cascade for face detection. Classifies faces into 7 emotions at 63.2% test accuracy. This is a learning project or demo baseline — not something you'd ship.
- Minimal stack: OpenCV + TensorFlow/Keras, no exotic dependencies, easy to run locally
- Pre-trained model available for download so you can see it working without the training setup
- The pipeline is straightforward to follow — face detection → crop → CNN → argmax — good for understanding the basics
- FER-2013 is the standard benchmark here, so the 63.2% accuracy number is comparable to other published baselines
- 63.2% accuracy on FER-2013 is near the floor for CNNs on this dataset — modern approaches hit 70%+ and FER+ re-labeling pushes further; this architecture is circa 2018
- Haar cascade face detection fails badly on non-frontal faces, partial occlusion, or poor lighting — anyone building on this will hit this wall immediately
- Last meaningful commit is years old; TFLearn is in the topics but TensorFlow 2.x is used — the project has some identity confusion and no active maintenance
- Single-file implementation with no tests, no packaging, no configuration surface — fine for a demo but you're on your own if you want to integrate this into anything real