This repository provides scoring scripts for:

• ✅ Task A: Gender Classification
• ✅ Task B: Face Recognition (Binary Verification)

Install required libraries:

pip install -r requirements.txt

Required config files:

Download Face Embeddings from Google Drive

Download Gender model weights from Google Drive

Download ArcFace model weights from Google Drive

We used the CLIP (Contrastive Language-Image Pretraining) ViT-B/32 model to extract image features then apply an MLP head to classify the gender of the person in the images.

We have provided our training jupyter notebook clip-genderclassification.ipynb which has the validation result and the training result is provided in the taska-scoregenerator.ipynb.

We have also made a separate score_task_a.py which can be used a tool to classify images. To classify images using the scorer file a sample Test folder structure is given below and also in the repo.

val_task_a/
├── male/
│   ├── img1.jpg
│   ├── img2.jpg
│   └── ...
├── female/
│   ├── img1.jpg
│   ├── img2.jpg
│   └── ...

After having the above folder structure download the model_weights.pt from gdrive and use the below command line prompt to run the scorer.

python score_task_a.py --val_path ./val_task_a --weights_path ./saved_model/model_weights.pt

We used a retriever model based on ArcFace Face Embeddings and vectorDB ChromaDB to make the Face Recognition model.

We have also provided with model development jupyter notebook task_B_dev.ipynb. We have used the ArcFace model through the DeepFace package's interface and stored all the face embeddings in ChromaDB with labels as metadata.

We have provided the faces_collection.json in gdrive which has the embeddings collection. score_task_b.py can be used to do face recognition based on FaceCOM dataset.

val_task_b/
├── Person_A/
│   ├── img.jpg
│   └── distortion/
│       └── distorted_1.jpg
├── Person_B/
│   ├── ...

python score_task_b.py --val_path ./val_task_b --collection faces_collection.json

Here, the recognized stands for the images that are from the FaceCOM dataset and not_recognized for the images not present in FaceCOM. Results where generated for 1000 images on the training and validation set respectively.

For a detailed System Description see : System Description