Face occlusion classification

git clone https://github.com/LamKser/face-occlusion-classification.git
cd face-occlusion-classification

💻 Hardware & Environment

• All the train and test processes are done on google colab with GPU Tesla T4

  conda env create -f environment.yml
  conda activate face-occlusion
  

📚 Dataset

• Crawl 9,749 images from the internet, crop the face by using FaceMaskDetection and divide into 2 classes:
  • 0 - Non-occluded face
  • 1 - Occluded face

Figure 1: Non-occluded face example

Figure 2: Occluded face example

• Then split the dataset into 3 sets (7 - 2 - 1):

  • Train set : 6,826 images
  • Val set : 1,945 images
  • Test set : 978 images

• Data structure:

  face_occlusion
  ├───Train
  │   ├───1
  │   │   ├─face_0.jpg
  │   │   ├─face_1.jpg
  │   │   └...
  │   └───0
  ├───Val
  │   ├───1
  │   └───0
  └───Test
      ├───1
      └───0
  

• 🔗 Data link: face occlusion dataset

📐 Config

• To use other model or change hyperparameters, you can edit train.yml and test.yml in configs folder
• Available models: densenet169, resnet18, resnet50

🏗️ Train model

• Train

  python train.py --opt configs/train.yml
  

• Show the training and validation progress

  tensorboard --logdir logger
  

• If using wandb to log training process:

  wandb:
      project: <Type your project>
      name: <Type experiment name>
  

📈 Test model

• Test the model

  python test_model.py --opt configs/test.yml
  

• Test single image

  python test_single_image.py --model <model_name> --weight <weight_path> --image <image_path>
  

• ONNX model

  • Convert pytorch model to onnx

    python onnx/convert_2_onnx.py --model <model name> \\
                                  --weight <weight and checkpoint file> \\
                                  --save <path/to/save/onnx/*.onnx> \\
                                  --opset_version <version> (optional)
    

  • Run onnx model

    python onnx/run_onnx.py --onnx <onnx file> --img <your image>
    

📊 Results (Train/Val/Test)

• All the trained model: trained model
• The pretrained models are trained with 30 epochs

Last model

Model	Params (M)	Infer (ms)	Accuracy	Precision	Recall	F1	Weights
VGG16	134.2	7.76	0.9805	0.981	0.9789	0.9799	link
VGG19	139.5	9.36	0.9836	0.9831	0.9831	0.9831	link
VGG16-BN	134.2	8.3	0.9734	0.9746	0.9705	0.9725	link
VGG19-BN	139.5	10.01	0.9713	0.9765	0.9642	0.9703	link
DenseNet169	12.4	25.46	0.9795	0.9729	0.9852	0.979	link
DenseNet201	18	31.06	0.9744	0.9787	0.9684	0.9735	link
ResNet18	11.1	3.69	0.9703	0.9665	0.9726	0.9695	link
ResNet50	23.5	7.15	0.9754	0.9787	0.9705	0.9746	link
ResNet152	58.1	19.31	0.9805	0.983	0.9768	0.9799	link
ConvNeXt-Base	87.5	13.26	0.9867	0.9894	0.9831	0.9862	link
ConvNeXt-Small	49.4	11.54	0.9887	0.9853	0.9915	0.9884	link
ConvNeXt-Tiny	27.8	7.24	0.9867	0.9832	0.9894	0.9863	link