BLT: Balancing Long-Tailed Datasets with Adversarially-Perturbed Images

ACCV 2020 Paper, Supplemental material and Video

Original PyTorch implementation of the ACCV 2020 paper: BLT: Balancing Long-Tailed Datasets with Adversarially-Perturbed Images

Data

To recreate the experiments from the paper please download the original datasets from:

• ImageNet
• Places-365
• iNaturalist 2018

In 'main.py' modify the path to the datasets:

data_root = {'ImageNet': '/putYourPathHere/ImageNet_LT',
             'Places': '/putYourPathHere/Places/',
             'iNaturalist_2018': '/putYourPathHere/iNaturalist_2018'}

Exact train/val/test splits for all three datasets:

• Can be found here

Please extract them into the 'data' directory:

data
  |--ImageNet_LT
    |--ImageNet_LT_train.txt
    |--ImageNet_LT_test.txt
    |--ImageNet_LT_val.txt
  |--Places_LT
    |--Places_LT_train.txt
    |--Places_LT_test.txt
    |--Places_LT_val.txt
  |--iNaturalist_2018
    |--iNaturalist_2018_train.txt
    |--iNaturalist_2018_test.txt
    |--iNaturalist_2018_val.txt

Train

ImageNet-LT

To train both Stage 1 and Stage 2 run:

sh train.sh

To train just Stage 1:

python3 main.py --config ./config/ImageNet_LT/stage_1.py --no_sampler --no_hallucinations --gpu 0

And then to train Stage 2:

python3 main.py --config ./config/ImageNet_LT/stage_2.py --gpu 0

Places-LT

To train just Stage 1:

python3 main.py --config ./config/Places_LT/stage_1.py --no_sampler --no_hallucinations --gpu 0

And then to train Stage 2:

python3 main.py --config ./config/Places_LT/stage_2.py --gpu 0

iNaturalist-2018

To train just Stage 1:

python3 main.py --config ./config/iNaturalist_2018/stage_1.py --no_sampler --no_hallucinations --gpu 0

And then to train Stage 2:

python3 main.py --config ./config/iNaturalist_2018/stage_2.py --gpu 0

General information

Authors:

• Jedrzej Kozerawski (jkozerawski@ucsb.edu)
• Victor Fragoso (victor.fragoso@microsoft.com)
• Nikolaos Karianakis (nikolaos.karianakis@microsoft.com)
• Gaurav Mittal (gaurav.mittal@microsoft.com)
• Matthew Turk (mturk@ttic.edu)
• Mei Chen (mei.chen@microsoft.com)

Citation

If you use this code for your research, please cite the following paper:

[bibtex]
@InProceedings{Kozerawski_2020_ACCV,
    author    = {Kozerawski, Jedrzej and Fragoso, Victor and Karianakis, Nikolaos and Mittal, Gaurav and Turk, Matthew and Chen, Mei},
    title     = {BLT: Balancing Long-Tailed Datasets with Adversarially-Perturbed Images},
    booktitle = {Proceedings of the Asian Conference on Computer Vision (ACCV)},
    month     = {November},
    year      = {2020}
}

Video

YouTube: https://youtu.be/stMSOwrJToU