Bamboo: 4 times larger than ImageNet; 2 time larger than Object365; Built by active learning.
Bamboo is a mega-scale and information-dense dataset for classification and detection pre-training. It is built upon integrating 24 public datasets (e.g. ImagenNet, Places365, Object365, OpenImages) and added new annotations through active learning. Bamboo has 69M image classification annotations (4 times larger than ImageNet) and 32M object bounding boxes (2 times larger than Object365).
π₯³! π₯³!
[11/2022] We release Bamboo-Det. [10/2022] We won the first place in Computer Vision in the Wild Challenge(ImageNet-1K in Pre-training track). π₯³! [06/2022] We split Bamboo-CLS into 30 datasets that represent different realms (e.g. car, mammals, food and etc.) in the natural worlds: HERE [06/2022] We release Bamboo-CLS with FC layer, it can classify 115,217 categories. [06/2022] We release our label system with many useful attributes!. [03/2022] Bamboo-CLS ResNet-50 and Bamboo-CLS ViT B/16 have been released. [03/2022] arXiv paper has been released.
NAME: XXX
ORGANIZATION: XXX (Bamboo is only for academic research and non-commercial use)
We provide the hierarchy for our label system at HERE. This JSON file includes the following attrubutes of each concept. We hope this information will be beneficial for your research.
We take concept/class dog
as an example.
#input
with open('PATH-TO-JSON-FILE.json') as f:
bamboo = json.load(f)
print(bamboo.keys())
#output
'father2child', 'child2father', 'id2name', 'id2desc', 'id2desc_zh', 'id2name_zh'
id (n02084071)
of the dog
on HERE.bamboo['child2father']['n02084071']
: domestic_animals, canine.bamboo['father2child']['n02084071']
: husky, griffon, shiba inu and etc.bamboo['id2desc']['n02084071']
: a member of the genus Canis (probably descended from the common wolf) that has been domesticated by man since prehistoric times; occurs in many breeds.bamboo['id2state']['n02084071']['academic']
: openimage, iWildCam2020, STL10, cifar10, iNat2021, ImageNet21K, coco, OpenImage, object365.Downloading the whole dataset might be unnecessary for most purposes. We provide meta files based on the following dimension.
gdown
pip install gdown
id
of the filesgdown https://drive.google.com/uc?id=1WEKQ_68Y9i9FzakvPYU6Yj5SOvkZCIEm
Model | Link | Data | cifar10 | cifar100 | food | pet | flower | sun | stanfordcar | dtd | caltech | fgvc-aircraft | AVG |
---|---|---|---|---|---|---|---|---|---|---|---|---|---|
ResNet-50 | Official | CLIP | 88.7 | 70.3 | 86.4 | 88.2 | 96.1 | 73.3 | 78.3 | 76.4 | 89.6 | 49.1 | 79.64 |
ViT B/16 | Official | CLIP | 96.2 | 83.1 | 92.8 | 93.1 | 98.1 | 78.4 | 86.7 | 79.2 | 94.7 | 59.5 | 86.18 |
ResNet-50 | link | Bamboo-CLS | 93.6 | 81.7 | 85.6 | 93.0 | 99.4 | 71.6 | 92.3 | 78.2 | 93.6 | 84.4 | 87.33 |
ViT B/16 | link link_with-FC | Bamboo-CLS | 98.5 | 91.0 | 93.3 | 95.3 | 99.7 | 79.5 | 93.9 | 81.9 | 94.8 | 88.8 | 91.65 |
Dataset | Model | Link | VOC (AP50) | CITY (MR) | COCO (mmAP) |
---|---|---|---|---|---|
OpenImages | ResNet-50 + FPN | Official | 82.4 | 16.8 | 37.4 |
Object365 | ResNet-50 + FPN | Official | 86.4 | 14.7 | 39.3 |
Bamboo-DET(Detectron2) | ResNet-50 + FPN | link | 87.5 | 12.6 | 43.9 |
# Create conda environment
conda create -n bamboo python=3.7
conda activate bamboo
# Install Pytorch
conda install pytorch==1.8.0 torchvision==0.9.0 cudatoolkit=10.2 -c pytorch
# Clone and install
git clone https://github.com/Davidzhangyuanhan/Bamboo.git
Downloading and organizing each downstream dataset as follows
data
βββ flowers
β βββ train/
β βββ test/
β βββ train_meta.list
β βββ test_meta.list
Changing root and meta in Bamboo-Benchmark/configs/100p/config_\*.yaml
Writing the path of the downloaded/your model config in Bamboo-Benchmark/configs/models_cfg/\*.yaml
Writing the name of the downloaded/your model in Bamboo-Benchmark/multi_run_100p.sh
sh Bamboo-Benchmark/multi_run_100p.sh
If you use this code in your research, please kindly cite the following papers.
@misc{zhang2022bamboo,
title={Bamboo: Building Mega-Scale Vision Dataset Continually with Human-Machine Synergy},
author={Yuanhan Zhang and Qinghong Sun and Yichun Zhou and Zexin He and Zhenfei Yin and Kun Wang and Lu Sheng and Yu Qiao and Jing Shao and Ziwei Liu},
year={2022},
eprint={2203.07845},
archivePrefix={arXiv},
primaryClass={cs.CV}
}
Thanks to Siyu Chen (https://github.com/Siyu-C) for implementing the Bamboo-Benchmark.