[ECCV2022] OmniBenchmark

Yuanhan Zhang¹  Zhenfei Yin²  Jing Shao²  Ziwe Liu¹

¹S-Lab, Nanyang Technological University  ²SenseTime Research 

---

[Paper] • [Project Page] • [Leaderboard]
[Challenge:ImageNet1k-Pretrain Track] • [Challenge:Open-Pretrain Track]

Updates

[12/2022] For users in mainland China, you can also download OmniBenchmark v2 at [HERE]. Thanks for the OpenDataLab.

[11/2022] We upload the tar file of the OmiBenchmark V2 at [HERE]. md5sum

[08/2022] We release OmniBenchmark V2 [Statistics].

[07/2022] OmniBenchmark Challenge ECCV@2022 will start together with ECCV 2022 SenseHuman Workshop.

[07/2022] Dataset with hidden test has been released.

[07/2022] Code for ReCo has been released.

[07/2022] arXiv paper has been released.

About OmniBenchmark

OmniBenchamrk V2

Clean Strategy

We use Bamboo_ViT-B16 to clean up the OmniBenchmark following two solutions, producing the OmniBenchmark V2 (meta_url_4_challenge_v2).

• Delete images whose inference result lies outside its belonging realm. e.g. delete the image from the "bird" realm if its inference class is "tiger."
• Clustering images by K-means and deleting clusters whose images are less than 2. Images from the such cluster are mostly noise.

Trainid and its name

The trainid of OmniBenchamrk V2 is different from V1, we release the mapping file trainid2name_v2.json

Evaluation Protocol

• You should train on train.txt, hyper-parameter search on val.txt and finally evaluate on test.txt.
• The V2 results of recent methods will be released soon.

Download data and annotations

IMPORTANT:

• You can download the data and annotation of OmniBenchamrk V2 at [HERE]. Afer you have downloaded 9 omnibenchmark_v2_onedrive.tar.* files, you can untar them using

cat omnibenchmark_v2_onedrive.tar.gz.* | tar -xvf

For the downloading of OmniBenchamrk V1, you should follow the following step.

cd download_tool
#it may cost 2 hours
pythoon download_image.py

After downlaoding you should see the following folder structure, i.e., a separate folder of images per realm:

<meta>
...
|--- activity
|   |--- activity.train
|   |   |---images/ #data
|   |   |    |---*.jpg
|   |   |---record.txt #annotation
|   |--- activity.val
|   |   |images/ #data
|   |   |    |---*.jpg
|   |   |--- record.txt #annotation
|   |--- activity.test
|   |   |images/ #data
|   |   |    |---*.jpg
|   |   |--- record.txt #image_path + pseudo_class
...

Please refer to download_tool/README.txt for the detail information of your downloaded files.

Find the class name

In downloaded meta files (e.g. car.val), each line of the file is a data record, including the local image path and the corresponding label, separated by a space.

#path trainid
XXXXXX 0
XXXXXX 1
XXXXXX 2
...

You can find the name of trainid through trainid2name.json (trainid2name_v2.json).

Evaluating a model on the OmniBenchmark

Step1: Model preparation

Public models

Inspired by ImageNet-CoG, we use ResNet50 as a reference model, and evaluate 22 models that are divided into three groups. You can download these models at HERE. You can check the reference papers of these model in the paper.

After you download models, you should update their path in their config files in the linear_probe/model_cfg/.

e.g. if you download beit_b16 model in the ./weights/beit_base_patch16_224_pt22k_ft22kto1k.pth

• vim linear_probe/model_cfg/beit_b16.yaml
• Change /mnt/lustre/zhangyuanhan/architech/beit_base_patch16_224_pt22k_ft22kto1k.pth to ./weights/beit_base_patch16_224_pt22k_ft22kto1k.pth.

Customer models

• Upload your model files in linear_probe/models/ABC.config, ABC is your model name.
• Upload the corresponding config files in linear_probe/configs/model_cfg/.

Step2: Data preparation

Updating the path of your downloaded data and annotation in linear_probe/configs/100p/.

e.g. add the information of activity dataset.

• vim linear_probe/100p/config_activity.yaml
• Update the root in line 13/19 and meta in line 14/20

Step3: Linear probing

• vim linear_probe/multi_run_100p.sh
• Change models=(beit_b16 effenetb4) to models=(beit_b16 effenetb4 ABC). Separating each model name in space.
• Change datasets=(activity aircraft) to datasets=(activity aircraft DEF GHI). DEF and GHI is the dataset name you want to evaluate, refered to linear_probe/configs/100p/config_DEF.yaml.
• sh linear_probe/multi_run_100p.sh

About relational contrastive (ReCo) learning

Similarity information in ImageNet1k

./ReCo/ImageNet1K.visual.3_hump.relation.depth_version.json provides the similarity information of classes in ImageNet1k (Equation 4 in the paper).

ReCo loss

We can use ReCo loss ./ReCo/losses.py in any supervised contrastive learning framework. Here we use Parametric-Contrastive-Learning (PaCo) in our experiments.

#Run ReCo
sh ./sh/train_resnet50_reco_imagenet1k.sh

Citation

If you use this code in your research, please kindly cite this work.

@inproceedings{zhang2022benchmarking,
  title={Benchmarking omni-vision representation through the lens of visual realms},
  author={Zhang, Yuanhan and Yin, Zhenfei and Shao, Jing and Liu, Ziwei},
  booktitle={European Conference on Computer Vision},
  pages={594--611},
  year={2022},
  organization={Springer}
}

Acknowledgement

Thanks to Siyu Chen (https://github.com/Siyu-C) for implementing the linear_probe.
Thanks to Qinghong Sun for coordinating the data collection.
Part of the ReCo code is borrowed from Parametric-Contrastive-Learning.

License

This dataset is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License.