[ECCV 2022] VL-LTR: Learning Class-wise Visual-Linguistic Representation for Long-Tailed Visual Recognition

Usage

First, install PyTorch 1.7.1+, torchvision 0.8.2+ and other required packages as follows：

conda install -c pytorch pytorch torchvision
pip install timm==0.3.2
pip install ftfy regex tqdm
pip install git+https://github.com/openai/CLIP.git
pip install mmcv==1.3.14

Data preparation

ImageNet-LT

Download and extract ImageNet train and val images from here. The directory structure is the standard layout for the torchvision datasets.ImageFolder, and the training and validation data is expected to be in the train/ folder and val/ folder respectively.

Then download and extract the wiki text into the same directory, and the directory tree of data is expected to be like this:

./data/imagenet/
  train/
    class1/
      img1.jpeg
    class2/
      img2.jpeg
  val/
    class1/
      img3.jpeg
    class2/
      img4.jpeg
  wiki/
  	desc_1.txt
  ImageNet_LT_test.txt
  ImageNet_LT_train.txt
  ImageNet_LT_val.txt
  labels.txt

After that, download the CLIP's pretrained weight RN50.pt and ViT-B-16.pt into the pretrained directory from https://github.com/openai/CLIP.

Places-LT

Download the places365_standard data from here. (For researchers in Chinese Mainland, you can download the long-tailed version directly from Baidu Netdisk 6x8u.)

Then download and extract the wiki text into the same directory. The directory tree of data is expected to be like this (almost the same as ImageNet-LT):

./data/places/
  train/
    class1/
      img1.jpeg
    class2/
      img2.jpeg
  val/
    class1/
      img3.jpeg
    class2/
      img4.jpeg
  wiki/
  	desc_1.txt
  Places_LT_test.txt
  Places_LT_train.txt
  Places_LT_val.txt
  labels.txt

iNaturalist 2018

Download the iNaturalist 2018 data from here.

Then download and extract the wiki text into the same directory. The directory tree of data is expected to be like this:

./data/iNat/
  train_val2018/
  wiki/
  	desc_1.txt
  categories.json
  test2018.json
  train2018.json
  val2018.json

Evaluation

To evaluate VL-LTR with a single GPU run:

• Pre-training stage

bash eval.sh ${CONFIG_PATH} 1 --eval-pretrain

• Fine-tuning stage:

bash eval.sh ${CONFIG_PATH} 1

The ${CONFIG_PATH} is the relative path of the corresponding configuration file in the config directory.

Training

To train VL-LTR on a single node with 8 GPUs for:

• Pre-training stage, run:

bash dist_train_arun.sh ${CONFIG_PATH} 8

• Fine-tuning stage:

  • First, calculate the $\mathcal L_{\text{lin}}$ of each sentence for AnSS method by running this:

  bash eval.sh ${CONFIG_PATH} 1 --eval-pretrain --select

  • then, running this:

  bash dist_train_arun.sh ${CONFIG_PATH} 8

The ${CONFIG_PATH} is the relative path of the corresponding configuration file in the config directory.

Results

Below list our model's performance on ImageNet-LT, Places-LT, and iNaturalist 2018.

Dataset	Backbone	Top-1 Accuracy	Download
ImageNet-LT	ResNet-50	70.1	Weights
ImageNet-LT	ViT-Base-16	77.2	Weights
Places-LT	ResNet-50	48.0	Weights
Places-LT	ViT-Base-16	50.1	Weights
iNaturalist 2018	ResNet-50	74.6	Weights
iNaturalist 2018	ViT-Base-16	76.8	Weights

For more detailed information, please refer to our paper directly.

Citation

If you are interested in our work, please cite as follows:

@article{tian2021vl,
  title={VL-LTR: Learning Class-wise Visual-Linguistic Representation for Long-Tailed Visual Recognition},
  author={Tian, Changyao and Wang, Wenhai and Zhu, Xizhou and Wang, Xiaogang and Dai, Jifeng and Qiao, Yu},
  journal={arXiv preprint arXiv:2111.13579},
  year={2021}
}

License

This repository is released under the Apache 2.0 license as found in the LICENSE file.