Skip to content

Latest commit

 

History

History
87 lines (65 loc) · 2.64 KB

README.md

File metadata and controls

87 lines (65 loc) · 2.64 KB
Raw

COCO-O: A Benchmark for Object Detectors under Natural Distribution Shifts (ICCV'2023)

Benchmark Results This repository contains the data and guidelines of COCO-O dataset, which is proposed in our ICCV2023 paper "COCO-O: A Benchmark for Object Detectors under Natural Distribution Shifts".

Paper: https://arxiv.org/abs/2307.12730

COCO-O

COCO-O(ut-of-distribution) contains 6 domains (sketch, cartoon, painting, weather, handmake, tattoo) of COCO objects which are hard to be detected by most existing detectors. The dataset has a total of 6,782 images and 26,624 labelled bounding boxes.

If you want to use our COCO-O, first download dataset from here and then unzip the ood_coco.zip into /path/to/ood_coco

After that, you can follow the section of Benchmarking Detectors to evaluate on coco-o.

COCO-O

Benchmarking Detectors

COCO-O is organized with a similar file structure with COCO validation. So the standard COCO evaluation protocol works consistently by changing the directory /path/to/coco to /path/to/ood_coco/sketch

COCO Validation

coco
|-- val2017
    |-- xxx.jpg
    |-- ...
|-- annotations
    |-- instances_val2017.json

COCO-O

coco-o
|-- sketch
    |-- val2017
        |-- xxx.jpg
        |-- ...
    |-- annotations
        |-- instances_val2017.json
|-- cartoon
    |-- val2017
        |-- xxx.jpg
        |-- ...
    |-- annotations
        |-- instances_val2017.json
|-- painting
    ...

Following is the example for using COCO-O in MMdetection or Detectron2

MMDetection

git clone https://github.com/open-mmlab/mmdetection.git
cd mmdetection
mkdir data
ln -s /path/to/ood_coco/sketch data/coco

python tools/test.py ${CONFIG_FILE} ${CHECKPOINT_FILE} --eval bbox

Detectron2

git clone https://github.com/facebookresearch/detectron2.git
cd detectron2
mkdir datasets
ln -s /path/to/ood_coco/sketch datasets/coco

./tools/lazyconfig_train_net.py --config-file ${CONFIG_FILE} \
train.init_checkpoint=${CHECKPOINT_FILE} \
--eval-only

Citation

If you find this useful in your research, please consider citing:

@article{mao2023coco,
  title={COCO-O: A Benchmark for Object Detectors under Natural Distribution Shifts},
  author={Mao, Xiaofeng and Chen, Yuefeng and Zhu, Yao and Chen, Da and Su, Hang and Zhang, Rong and Xue, Hui},
  journal={Proceedings of the IEEE/CVF International Conference on Computer Vision},
  year={2023}
}