ECCV'22 ChaLearn Seasons in Drift Challenge (track 1: Day level)

1. Enviroment

• Ubuntu 18.04
• CUDA 11.1
• 8*3090 GPU
• Python3.7
• pytorch-1.7, torchvision-0.8.0
• mmdetection-2.22.0

2. Install

(1). Create a conda environment and activate it.

conda create --name openmmlab python=3.8 -y
conda activate openmmlab

(2). Install PyTorch 1.7 or highre

conda install pytorch==1.7.0 torchvision==0.8.0 torchaudio==0.7.0 cudatoolkit=11.0 -c pytorch

(3). Install MMdetection-2.22.0 and mmcv

pip install -r requirements.txt
pip install mmdet==2.22.0
pip install mmcv-full==1.3.17
pip install tqdm

3. data pross

（1）Place the training set in the data/train folder and unzip it.

（2）Place the testing set in the data/test folder and unzip it.

（3）Place the labels file in the data/labels folder and unzip it.

（4) Run dataprocess: sh data_process.sh

4. training models

Notice: We use 8*3090 GPU to train our models. If GPU numbers not equal 8, we should simply change the learning rate. For examole: If 4 GPUs are used, change the learning rate 0.02/2 in config_day/day.py .

Before training, pre-trained weights on coco dataset should download unzip, and put it in weights folder

wget https://github.com/CBNetwork/storage/releases/download/v1.0.0/cascade_rcnn_cbv2d1_r2_101_mdconv_fpn_20e_fp16_ms400-1400_giou_4conv1f_coco_swa.pth.zip

mv cascade_rcnn_cbv2d1_r2_101_mdconv_fpn_20e_fp16_ms400-1400_giou_4conv1f_coco_swa.pth.zip weights/

unzip cascade_rcnn_cbv2d1_r2_101_mdconv_fpn_20e_fp16_ms400-1400_giou_4conv1f_coco_swa.pth.zip

sh dist_train.sh

5. test and generate result file

sh test_submit.sh

Now we can find a zip file in submit_last folder.

If you have any problems, pleanse mail me directly.

Thanks for your efforts.