Set up environment

Our model and code are based on CBNetV2 of UniverseNet

git clone https://github.com/haiminh2001/Train4Ever.git

pip install -U openmim
mim install mmcv-full

cd Train4Ever
pip install -v -e .

Change permission to train

chmod 777 ./tools/dist_train.sh

Preprocessing and Gen label format COCO

Preprocessing images to PNG

python preprocessing.py --input_path <path_to_input_data> --output_path ./data/TrainImagesPNG

Gen label fomart COCO

python genLabelCocoFormat.py --input_labels_path <path_to_input_label_images> --output_folder_path ./data

Our preprocessed data, generated annotation file and pretrain model CBNetV2 on dataset COCO

To save time you can download our preprocessed data and generated annotation file here

Pretrain model CBNetV2 on dataset COCO here, put file weight pretrain at folder pretrains

After the download is complete you set up folder as picture below and put the corresponding dataset

Training

tools/dist_train.sh configs/cbnet/mask_rcnn_cbv2_swin_tiny_patch4_window7_mstrain_780-1100_adamw_3x_coco.py 1

You can change file config here

Inference

checkpoint our model here

python Infer.py --input_path <path_to_input_data> --config_path <path_to_input_config> --ckpt_path <path_to_input_ckpt> --output_path <path_to_output>