PSP: Automatic Check-out via Prototype-based Classifier Learning from Single-Product Exemplars

Environment

Python 3.7.11
Pytorch 1.7.1
torchvision 0.8.2
cudatoolkit 10.1.243
mmcv-full 1.3.9
mmdet 2.14.0

Installation

The installation process of our code can follow the MMDetection in the url: https://mmdetection.readthedocs.io/en/v2.14.0/get_started.html and overlay the original files by our code.
Or you can download our code and the file "README.md" of MMDetection, then run ''python setup.py develop'' .

Dataset

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

Necessary modifications

Before training or testing, you must modify the contents of the paths in the following files.

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

   data_root = '{your_dataset_root_path}'
   img_prefix = '{your_train_or_test_dataset_path}'
   ann_file = '{your_train_or_test_annotation_file}'  
   

2. rpc_eval.py
   The default value of --ann_file can be modified as {your_train_or_test_annotation_file}.

3. rpc_one_img_py
   In this file, you need modify the test annotation file in following:

   with open('{your_test_annotation_file}', 'r') as f:  
   img ='{test_image_file}'  
   out_file='{result_out_file}'  
   

Train

We train our model in one 2080Ti card, and the training command is:

python tools/train.py {config file}  

and an example is:

python tools/train.py configs/rpc/faster_rcnn_r50_fpn_3x_rpc_protoS_srr_mll.py  

Test

Our testing command is:

 python tools/test.py {config file} {checkpoint_file} [--out {result_file}] [--eval bbox]  

and an example is:

python tools/test.py configs/rpc/faster_rcnn_r50_fpn_3x_rpc_protoS_srr_mll.py \
       result/faster_rcnn_r50_fpn_3x_rpc_protoS_srr_mll/latest.pth --out \
       result/faster_rcnn_r50_fpn_3x_rpc_protoS_srr_mll/result.pkl --eval bbox  

Evaluate

Since 4 metrics other than mAP50 and mmAP are used, we perform the computation of these 4 metrics separately for the result file, i.e. rpc_eval.py.
This evaluation file can print 4 metric values for all clutter patterns. And the evaluating command is:

 python rpc_eval.py [--root {result_root_path}]  

and an example is:

python rpc_eval.py --root ./result/faster_rcnn_r50_fpn_3x_rpc_protoS_srr_mll/  

Note

If you want to train, test and evaluate the model sequentially at once, you can run the "rpc_run.py" file directly.

python rpc_run.py

If you want to use other configs, you only need modify the corresponding file names or paths in the "rpc_run.py" file.