Baseline for MCS2022.Car models verification competition

This is a repository with a baseline solution for the MCS2022. Cars verification competition. In this competition, participants need to train a model to verify car models (models are the same, not the same car).

The idea of the basic solution is to train a classifier of car models, remove the classification layer and use embeddings to measure the proximity between two images.

Steps for working with baseline

0. Download CompCars dataset

To train the model in this offline the CompCars dataset is used. You can download it here.

If you have problems getting box labels in datasets, you can use a duplicate of the labels that we posted here.

1. Prepare data for classification

Launch prepare_data.py to crop images on bboxes and generate lists for training and validation phases.

python prepare_data.py --data_path ../CompCars/data/ --annotation_path ../CompCars/annotation/

2. Run model training

CUDA_VISIBLE_DEVICES=0 python main.py --cfg config/baseline_mcs.yml

3. Create a submission file

CUDA_VISIBLE_DEVICES=0 python create_submission.py --exp_cfg config/baseline_mcs.yml \
                                                   --checkpoint_path experiments/baseline_mcs/model_0077.pth \
                                                   --inference_cfg config/inference_config.yml