/BiAM

[ICCV 2021] Official Pytorch implementation for Discriminative Region-based Multi-Label Zero-Shot Learning SOTA results on NUS-WIDE and OpenImages

Primary LanguagePythonApache License 2.0Apache-2.0

PWC PWC

Discriminative Region-based Multi-Label Zero-Shot Learning (ICCV 2021)

[Paper][Video][Project Page]

Sanath Narayan*, Akshita Gupta*, Salman Khan, Fahad Shahbaz Khan, Ling Shao, Mubarak Shah

(:star2: denotes equal contribution)

Installation

The codebase is built on PyTorch 1.1.0 and tested on Ubuntu 16.04 environment (Python3.6, CUDA9.0, cuDNN7.5).

For installing, follow these intructions

conda create -n mlzsl python=3.6
conda activate mlzsl
conda install pytorch=1.1 torchvision=0.3 cudatoolkit=9.0 -c pytorch
pip install matplotlib scikit-image scikit-learn opencv-python yacs joblib natsort h5py tqdm pandas

Install warmup scheduler

cd pytorch-gradual-warmup-lr; python setup.py install; cd ..

Attention Visualization

Results

Our approach on NUS-WIDE Dataset.

Our approach on OpenImages Dataset.

Training and Evaluation

NUS-WIDE

Step 1: Data preparation

  1. Download pre-computed features from here and store them at features folder inside BiAM/datasets/NUS-WIDE directory.
  2. [Optional] You can extract the features on your own by using the original NUS-WIDE dataset from here and run the below script:
python feature_extraction/extract_nus_wide.py

Step 2: Training from scratch

To train and evaluate multi-label zero-shot learning model on full NUS-WIDE dataset, please run:

sh scripts/train_nus.sh

Step 3: Evaluation using pretrained weights

To evaluate the multi-label zero-shot model on NUS-WIDE. You can download the pretrained weights from here and store them at NUS-WIDE folder inside pretrained_weights directory.

sh scripts/evaluate_nus.sh

OPEN-IMAGES

Step 1: Data preparation

  1. Please download the annotations for training, validation, and testing into this folder.

  2. Store the annotations inside BiAM/datasets/OpenImages.

  3. To extract the features for OpenImages-v4 dataset run the below scripts for crawling the images and extracting features of them:

## Crawl the images from web
python ./datasets/OpenImages/download_imgs.py  #`data_set` == `train`: download images into `./image_data/train/`
python ./datasets/OpenImages/download_imgs.py  #`data_set` == `validation`: download images into `./image_data/validation/`
python ./datasets/OpenImages/download_imgs.py  #`data_set` == `test`: download images into `./image_data/test/`

## Run feature extraction codes for all the 3 splits
python feature_extraction/extract_openimages_train.py
python feature_extraction/extract_openimages_test.py
python feature_extraction/extract_openimages_val.py

Step 2: Training from scratch

To train and evaluate multi-label zero-shot learning model on full OpenImages-v4 dataset, please run:

sh scripts/train_openimages.sh
sh scripts/evaluate_openimages.sh

Step 3: Evaluation using pretrained weights

To evaluate the multi-label zero-shot model on OpenImages. You can download the pretrained weights from here and store them at OPENIMAGES folder inside pretrained_weights directory.

sh scripts/evaluate_openimages.sh

License

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

Citation

If you find this repository useful, please consider giving a star ⭐ and citation 🎊:

@article{narayan2021discriminative,
title={Discriminative Region-based Multi-Label Zero-Shot Learning},
author={Narayan, Sanath and Gupta, Akshita and Khan, Salman and  Khan, Fahad Shahbaz and Shao, Ling and Shah, Mubarak},
journal={Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
publisher = {IEEE},
year={2021}
}

Contact

Should you have any question, please contact 📧 akshita.gupta@inceptioniai.org