DEGNet: Decoupled Edge Guidance Network for Automatic Check-out

This repository implements DEGNet on the base of Pytorch. The implementation is heavily influenced by the project deep-learning-for-image-processing, and the implementation code of our Cascade RCNN framework references MMDetection.

Installation

Check requirements.txt for installation instructions. Similar torch versions should also work normally.

Data Preparation

The RPC dataset can be download by these two urls:
https://www.kaggle.com/diyer22/retail-product-checkout-dataset
https://pan.baidu.com/s/1vrrLaSpJe5JxT3zhYfOaog

Besides, you can use the method of DPSNet to generate synthesized images.

Necessary modifications

Before training, you should modify the contents of the paths in the following files.

1. train_utils/rpc/paths_catalog.py
   In these configuration files, you need modify the following parameters.

   rpc_train_syn: '{your synthesized data root path and annotation path}'
   rpc_instance: '{your rpc_validation_dataset root path and annotation path}'
   rpc_2019_test: '{your rpc_test_dataset root path and annotation path}'  
   

2. configs/xxx.py The configuration files we provide run on a 3090 GPU by default. You can adjust the global batch size and learning rate according to your graphics card configuration.

   batch_size: '{your batch_size}'
   lr: '{your learning rate}' 
   

Train

We train our model in one 3090 card, and the training command is:

python tools/train.py --config {config file} 

and an example is:

python tools/train.py --config configs/ffm_syn_att_da_pds.yaml 

Delete a file

You can delete the current file by clicking the Remove button in the file explorer. The file will be moved into the Trash folder and automatically deleted after 7 days of inactivity.

Test

If you only want to evaluate the model, you can execute:

 python tools/test.py --config {config file} --cpt {checkpoint_file_path}  

and an example is:

python tools/test.py --config configs/ffm_syn_att_da_pds.yaml  \
       --cpt result/ffm_syn_att_da_pds/latest.pt

Citation

Please consider citing our repo if it helps your research.