Celebrity Face Recognition

This repository contains the code developed for celebrity face recognition using a subset of the Celebrity Face Recognition Dataset. The following steps were undertaken to achieve accurate recognition:

Steps

1. Face Cropping with MTCNN:

   • Faces were cropped from images using MTCNN (Multi-Task Cascaded Convolutional Neural Network). To execute face cropping, run cropper_MCNN.py.

2. Model Fine-Tuning:

   • Several models were fine-tuned using the cropped dataset to optimize performance for face recognition.

3. Sample Reweighing:

   • Due to imbalance and noise in the dataset, sample reweighing techniques were applied, as described in this paper.

4. Ensembling Predictions:

   • Predictions from multiple fine-tuned models were ensembled using combined_predict.py to generate the final result.

Results

The approach resulted in an overall accuracy of 86.18%, which was the highest achieved for the competition.

Usage

To replicate the process:

• Execute cropper_MCNN.py to crop faces using MTCNN.
• Fine-tune models on the cropped dataset and train using 'train_using_<base_model>.py'.
• Use combined_predict.py to ensemble predictions for final accuracy assessment.