ACCV 2020 Paper, Supplemental material and Video
Original PyTorch implementation of the ACCV 2020 paper: BLT: Balancing Long-Tailed Datasets with Adversarially-Perturbed Images
To recreate the experiments from the paper please download the original datasets from:
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
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
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
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
- 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)
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}
}
YouTube: https://youtu.be/stMSOwrJToU