Multi-modal Aerial View Object Classification Challenge (MAVOC 2022) Solution

This repo contains the solution for MAVOC challenge track 2 (SAR+EO), which utilized data augmentation, focal loss, semi-supervised learning, label calibration etc. techniques to tackle with the given task. Detailed information can be found in method description. This solution ranked 6th in the final leaderboard in test phase.

Dataset structures

The datasets should be organized as follows. The /path/to/dataset in codes and configs refers to the root of this structure.

dataset
    - train_images
        - 0
        - 1
        ...
        - 9
    - test_images
        - test_eo
        - test_sar
    - valid_images
        - valid_eo
        - valid_sar

Training and Inference

Firstly, use the scripts in ./preprocess_scripts to generate csv file for dataloaders, then pip install required libraries listed in requirements.txt. Change the dataroot and csv_file path to your own customized paths, then use the following line for training:

sh bash_run_train_mavoc.sh --opt configs/004_final_sareo_light_semisuper_simpledual_focalloss_aug.yml

After the training process is finished, inference the test images using the following commands:

python test_mavoc.py --opt configs/004_final_sareo_light_semisuper_simpledual_focalloss_aug.yml