shiziqiang / checkout-data-generation Goto Github PK

This repo is to prepare synthetic data for training classifiers and detectors used in https://github.com/ShiZiqiang/aicity-23/. Details on how to generate the data can be found in our paper "CheckSORT: Refined Synthetic Data Combination and Optimized SORT for Automatic Retail Checkout".

Run the following command to generate the backgrounds

cd generate_backgrounds
bash download.sh
cd  ../

The generated images are placed in ./generate_backgrounds/temp/

Then resize the background image from 256X256 to 381X491:

mkdir -p ./data/backgrounds/LSUN_256x256_N2M2S128/
mkdir -p ./data/backgrounds/CelebA_128x128_N2M2S64/
mkdir -p ./data/backgrounds/biggan_imagenet/
mkdir -p ./data/backgrounds/random_bg/
python ./tools/batch_resize.py

Put the original 116500 training images provided by the organizer into ./data/AIC23_Track4_Automated_Checkout/train/.

Put segmentation labels of the original 116500 training images provided by the organizer into ./data/AIC23_Track4_Automated_Checkout/segmentation_labels.

Do the MSRCR enhancement of training data use retinex (borrowed and adapted from https://github.com/muggledy/retinex)

mkdir ./data/AIC23_Track4_Automated_Checkout/train_MSRCR
python ./retinex/batch_MSRCR.py

Convert the RGB into RGBA png image

mkdir ./data/AIC23_Track4_Automated_Checkout/train_RGBA
python ./tools/batch_to_rgba.py
mv ./data/AIC23_Track4_Automated_Checkout/train_RGBA/ ./data/objects/

First to install the required packages.

pip install -r requirements.txt

Then do the optimized synthesized training image generation (This program will generate images all the time without stopping, when it runs for 12 hours, it can be stopped manually).

python src/tools/generate_synthetic_data_train.py (Run 12 hours, then ctrl+c.)

Then combine the json files corresponding to all the images.

python src/tools/join_json.py

Generate part of the images for training classifiers (crop from the synthetic pictures just generated)

mkdir -p ./alladd2/train/
python src/tools/batch_mkdir.py
python src/tools/crop_117_class_objects.py
python src/tools/crop_classification_data.py

Comine with original 116500 training images for training classifiers

python batch_cp.py

Move the training data for detectors

mkdir coco_offline_MSRCR_GB_halfbackground_size100_no-ob_1
mv ./data/aicity23_train_scale0.4-0.8_iou0.5/train/ ./coco_offline_MSRCR_GB_halfbackground_size100_no-ob_1/

There is an instances_train.json in ./coco_offline_MSRCR_GB_halfbackground_size100_no-ob_1. Change the last row of instances_train.json from "name": "box" to "name": "1".

So far we have completed the training data preparation for the classifier and detector, which are in folders ./alladd2 and ./coco_offline_MSRCR_GB_halfbackground_size100_no-ob_1/ respectively.

If you have any questions, feel free to contact Ziqiang Shi ([email protected]).

@InProceedings{shi23AIC23,
	author = {Ziqiang Shi and Zhongling Liu and Liu Liu and Rujie Liu and Takuma Yamamoto and Xiaoyu Mi and Daisuke Uchida},
	title = {CheckSORT: Refined Synthetic Data Combination and Optimized SORT for Automatic Retail Checkout},
	booktitle = {The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Workshops},
	month = {June},
	year = {2023},
}