/Track4_aicity_2022

:fire::fire::fire: Winner of AI City Challenge 2022 - Track 4 :fire::fire::fire:

Primary LanguagePythonGNU General Public License v3.0GPL-3.0

Code release for AI City Challenge 2022 - Track 4 - Team 9

💥💥💥 Paper: Improving Domain Generalization by Learning without Forgetting: Application in Retail Checkout

💥💥💥 Our solution wins the 1st place Track 4. https://www.aicitychallenge.org/2022-challenge-winners/

Table of contents

  1. Operating Systems and Hardware Specs
  • Introduce the OS and hardware we use.
  1. Dataset preparation and environment setup
  • Give instructions to prepare datasets for training detector/classifier and set up environments.
  1. Train model
  • Provide scripts to train detector and classifier models.
  1. Run detector and ensemble detection results
  • Provide inference script for detector and give instructions on how to run ensemble.
  1. Create submission file
  • Provide script to create submission file to submit to test server.

Operating Systems and Hardware Specs

  • Ubuntu 18.04
  • Intel(R) Xeon(R) Silver 4210R CPU @ 2.40GHz
  • 4 GPUs Quadro RTX 8000 48GB
  • Driver: Version: 495.29.05, CUDA: 11.5
  • RAM: 256GB

Dataset preparation and environment setup

  • If you use our trained models, you can download here and ignore the dataset preparation step. Unzip the .zip file in the code directory. You should see the following structure:
    work_dirs
    ├── lwof_frcnn_swin_shapenet_gan
    │   └── epoch_1.pth
    ├── repvgg_a0_4xb64_10e_aicity22t4
    │   └── epoch_10.pth
    ├── res2net50_4xb64_10e_aicity22t4
    │   └── epoch_10.pth
    └── swins_4xb256_10e_aicity22t4
        └── epoch_10.pth
  • We split the dataset Auto-retail-syndata-release (provided by the challenge) into train and val sets. Please download the annotation files via the link. Then, unzip it into ./data/Auto-retail-syndata-release/anns as the following:
./data/Auto-retail-syndata-release/anns
├── train.txt
└── val.txt

Dataset preparation

  • We created datasets for training from random backgrounds and GANs pre-trained model.

  • Please run the following command to create datasets:

cd tools/data_process
bash download.sh
cd  ../..

Dataset will be created in data folder with following structure:

./data
├── annotations
├── biggan_imagenet
├── CelebA_128x128_N2M2S64
├── LSUN_256x256_N2M2S128
├── random_bg
├── Auto-retail-syndata-release
└── ShapeNetRendering
Folder description 
- "annotation" contains coco format json file for 'random_bg' images, 'random_bg' contain random backgrounds which were created to train our model.
- "CelebA_128x128_N2M2S64" and "LSUN_256x256_N2M2S128" are folders containing random backgrounds which are created by GAN model at [COCO-GAN](https://hubert0527.github.io/COCO-GAN/).
- "biggan_imagenet" contains random backgrounds which are created by [biggan](https://github.com/huggingface/pytorch-pretrained-BigGAN).
- ShapeNetRendering contained ShapeNet images
- "Auto-retail-syndata-release" contains AI city challenge dataset, please download and extract data by following organizers guide.

Environment setup

Create environment and install dependencies 
# create conda env
conda create -n aicity22_track4 python=3.7 -y
conda activate aicity22_track4
# Install others
conda install pytorch=1.10.0 torchvision cudatoolkit=10.2 -c pytorch -y
conda install cython -y
pip install click==7.1.2 openmim future tensorboard sklearn timm==0.4.12 tqdm torch_optimizer shapely scikit-learn scikit-image albumentations lap cython_bbox numba
pip install git+https://github.com/thuyngch/cvut
mim install mmcv-full==1.4.6 -f https://download.openmmlab.com/mmcv/dist/cu102/torch1.10.0/index.html -y
mim install mmcls==0.21.0 -y
mim install mmsegmentation==0.22.0 -y
mim install mmdet==2.22.0 -y
python setup.py develop
pip install Pillow==6.1 # Ignore the error

Train model

  • If you use our trained models, you can download here and ignore the Train model step. Unzip the .zip file and you should see the folder structure like in Dataset preparation section.

Train detection model

  • The detection model must be trained using 4 GPUs to be able to pre-produce the result.

  • Use the following script to train:

    Train detector 
    #!/usr/bin/env bash
set -e

## make sure GPUs are not hang
export NCCL_P2P_DISABLE=1

#set which GPUs will be used to train
export CUDA_VISIBLE_DEVICES=0,1,2,3

export CUDA_LAUNCH_BLOCKING=1
export MKL_NUM_THREADS=1
export PYTHONPATH="$(dirname $0)/..":$PYTHONPATH

#path to your model config file
CFG="configs/ccdet/track4/lwof_frcnn_swin_shapenet_gan.py"

#Set number of GPUs will be used to train
GPUS_NUM=4

#Command to train our detection model
mim train mmdet $CFG --gpus $GPUS_NUM --launcher pytorch --no-validate

Note: The detector's checkpoint will be saved at './work_dirs/lwof_frcnn_swin_shapenet_gan' folder

Train classification model

  • Following scripts will train 3 classification models namely RepVGG-A0, SwinS, and Res2Net50.
  • Must use 4 GPUs to reproduce the results.
  • The checkpoints will be saved under "./work_dirs"
Train classifier 
#!/usr/bin/env bash
set -e

export GPUS=4
export NCCL_P2P_DISABLE=1
export CUDA_VISIBLE_DEVICES=0,1,2,3

CFG="configs/cccls/repvgg_a0_4xb64_10e_aicity22t4.py"
WORKDIR="./work_dirs/repvgg_a0_4xb64_10e_aicity22t4"
mim train mmcls $CFG --work-dir $WORKDIR --launcher pytorch --gpus $GPUS \
    --seed 0 --deterministic

CFG="configs/cccls/swins_4xb256_10e_aicity22t4.py"
WORKDIR="./work_dirs/swins_4xb256_10e_aicity22t4"
mim train mmcls $CFG --work-dir $WORKDIR --launcher pytorch --gpus $GPUS \
    --seed 0 --deterministic

CFG="configs/cccls/res2net50_4xb64_10e_aicity22t4.py"
WORKDIR="./work_dirs/res2net50_4xb64_10e_aicity22t4"
mim train mmcls $CFG --work-dir $WORKDIR --launcher pytorch --gpus $GPUS \
    --seed 0 --deterministic

Run detector and ensemble detection results

Run detector

  • Following script will run 5 models located in tools/tta_run/models.

  • <path/to/videos> is the path to test videos which are used to submit to the test server. For example, all videos are located under "./data/testB_videos", then <path/to/videos> will be "./data/testB_videos".

  • Run:

    mkdir data/full_video_pkl
bash ./tools/tta_run/run.sh <path/to/videos> "./data/full_video_pkl"
rm data/full_video_pkl/*.py

  • The detection result will be saved under "./data/full_video_pkl" with the following structure:

  ./data/full_video_pkl
  ├── detectors_htc_r101_20e_coco
  │   ├── testA_1.pkl # This is output detection result for video name TestA_1.mp4
  ├── lwof_frcnn_swin_shapenet_gan
  │   ├── testA_1.pkl
  ...

Ensemble detection results

  • In the following scripts, <path/to/videos> is the path containing videos, it must be similar to <path/to/videos> in Run detector.
  • The ensemble results of videos will be saved at: ./cache/det_ensemble
rm data/full_video_pkl/*.py
python tools/ensemble/0_0_organize_pkl_results.py
./tools/ensemble/0_1_test_time_ensemble_results_in_folder.sh <path/to/videos>

Create submission file

  • Following script will run classifier, tracker, and create file for submission.
  • <path/to/videos> is same as <path/to/videos> in Run detector.
  • Videos for visualization are saved under: ./cache/cls_ensemble
  • Submission file is saved at: ./cache/submission.txt, use this file to submit to test server.
./tools/create_submission_file.sh <path/to/videos>

ABOUT US

We are Cybercore Co., Ltd. We proudly provide our award winning academic-level, state-of-the-art custom-made technologies to our customers in the field of image processing, image recognition and artificial intelligence (AI)

■Head Office
Malios 10th floor, 2-9-1 Morioka-Eki Nishi-Dori, Morioka, Iwate, Japan 020-0045 Tel : +81-19-681-8776

■Tokyo Office
Toranomon-Hills Business Tower 15F, Minato-ku, Tokyo, Japan 105-6490

■Vietnam (Affiliated company)
12th TNR Tower, 180-192 Nguyen Cong Tru St., District 1, HCMC Tel : (+84)028 710 77 710