xiaojinu/ae-wtn

Scaling Object Detection by Transferring Classification Weights

Autoencoder Weight Transfer Network (AE-WTN)

This is the code to replicate the AE-WTN experiments in the Scaling Object Detection by Transferring Classification Weights paper accepted as an oral paper at ICCV 2019.

Please consider citing this paper in your publications if it helps your research.

@inproceedings{kuen2019scaling,
   title = {Scaling Object Detection by Transferring Classification Weights},
   author = {Kuen, Jason and Perazzi, Federico and Lin, Zhe and Zhang, Jianming and Tan, Yap-Peng},
   booktitle = {ICCV},
   year = {2019}
}

Installation

• Prerequisites: PyTorch v1.2.0, pandas
• For others, please refer to INSTALL.md of the official maskrcnn-benchmark repo.
• python setup.py build develop

Dataset Preparation

cd AE-WTN/datasets/openimages

# download the Open Images training annotations 
wget https://storage.googleapis.com/openimages/challenge_2018/train/challenge-2018-train-annotations-bbox.csv

## create symlinks (in datasets/openimages) to image directories of training and evaluation datasets

# Open Images (challenge/V4/V5) training images directory (about 1.58M images with all download parts combined)
# https://www.figure-eight.com/dataset/open-images-annotated-with-bounding-boxes/ (train_00.zip, train_01.zip, ...)
ln -s train /path_to_openimages_images/train

# Open Images (V4/V5) validation images (41,620 images)
# https://datasets.figure-eight.com/figure_eight_datasets/open-images/zip_files_copy/validation.zip
ln -s val_600 /path_to_openimages_images/validation

# Visual Genome (Version 1.2) images (108,079 images with part 1 & 2 combined)
# https://cs.stanford.edu/people/rak248/VG_100K_2/images.zip
# https://cs.stanford.edu/people/rak248/VG_100K_2/images2.zip
ln -s VG_100K /path_to_visualgenome_images

cd ../..

Training & Evaluation

By default, 4 GPU cards are utilized for training and evaluation.

Training checkpoints are stored in the same directory. Evaluation results are stored in the inference subdirectory.

cd experiment

# training
sh train.sh

# evaluate on the 3 evaluation datasets
sh test.sh

Pretrained model: download link (place it in the same directory before running evaluation)

License

AE-WTN is released under the MIT license. See LICENSE for additional details.