ECCV2022-OOD-CV-Challenge-detection-Track-3rd-Place-Program

Competition Open Source Solutions

1. Environment setting

1.0. Package

• Several important packages
  • torch == 1.8.1+cu111
  • trochvision == 0.9.1+cu111

1.1. Dataset

In the classification track, we use only the OOD detection data and labels:

• ECCV-OOD

1.2. OS

• Windows10
• Ubuntu20.04
• macOS (CPU only)

2. Train

• Single GPU Training
• DataParallel (single machine multi-gpus)
• DistributedDataParallel

(more information: https://pytorch.org/tutorials/intermediate/ddp_tutorial.html)

2.1. data

train data and test data structure:

├── data/
│   ├── train
|   |    ├── Images
│   |    └── train.json
│   ├── test
|   |    ├── Images
│   |    └── phase2-det.json
│   ├── val
|   |    ├── Images
│   |    └── iid_test.json
│   └── occlusion

The structure flow of the generated file is as follows：

2.1.1 move picture

Put the training set image "ROBINv1.1-det" in

./data/train/Images/ 

Put the test set picture "phase2-det" in

./data/test/Images/

2.1.2 check picture

Since the given dataset contains gif images, we need to convert gif images into jpg images. We take the first frame of gif images as jpg images, and the generated jpg images automatically replace the original gif images.

python ./tools/check_gif.py  ./data/test/Images

2.2. run.

CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 ./tools/dist_train.sh  \
./config/cascade_rcnn_r50_fpn_1x_coco_backbone_convnextLarge_OnlyAdamW_cos_colorjitter_softmax_corrupt.py 8 \
--seed 0 \ 
--deterministic \ 
--work-dir ./work_dirs/  

3. Evaluation

CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 ./tools/dist_test_final.sh \
./config/cascade_rcnn_r50_fpn_1x_coco_backbone_convnextLarge_OnlyAdamW_cos_colorjitter_softmax_corrupt.py \ 
./work_dirs/epoch_15.pth 8 \
--format-only \ 
--options "jsonfile_prefix=./work_dirs/out" > ./work_dirs/out-test.out & 

4. Challenge's final checkpoints

It can be downloaded from Baidu Cloud Disk: https://pan.baidu.com/s/1scW9Z-PjZbrqi3VL7SNJ5w Extraction code：tc77

It can be directly used for model reasoning and get final result.

Acknowledgment

• Thanks to timm for Pytorch implementation.