This project implements Face Occlusion Detection using DeepLabV3+ to generate ground truth images for occluded face images. The output consists of binary masks with two classes:
- Black (0) - Background
- White (255) - Occlusion
The dataset used is WebFace OCC, but the full dataset is available only upon request from [email protected].
- Uses DeepLabV3+ with ResNet-101 backbone for segmentation.
- Generates ground truth occlusion masks for face images.
- Processes nested folder structures for batch mask generation.
- Outputs grayscale binary masks (background vs occlusion).
The project is trained using WebFace OCC. However, due to restrictions, the complete dataset is not included. If you require the full dataset, you must request access from:
To run this project, install the required dependencies:
pip install torch torchvision numpy pillow matplotlibModify dataset paths in face_occlusion_ground_truth_generation.py and run:
python face_occlusion_ground_truth_generation.pyNote that initially 14 ground truth images are created manually using Adobe Photoshop in order to train the model. I've given access to the same in the assets/GroundTruth.
To generate masks for a folder of images, update train_image_dir, train_mask_dir, input_folder and output_folder in the script and run:
python face_occlusion_ground_truth_generation.pyThe project uses DeepLabV3+ with a ResNet-101 backbone, pre-trained on COCO, and fine-tuned for occlusion segmentation with two classes:
- Occlusion (White - 255)
- Background (Black - 0)
Feel free to contribute by improving the segmentation model or optimizing the dataset usage. Submit a pull request if you'd like to collaborate!
This project is for research purposes only. Please cite the dataset owners if you use WebFace OCC.
For any queries, feel free to reach out!