Code for our paper "Towards Automated Testing and Robustification by Semantic Adversarial Data Generation" (Paper)
- Python 3.6.7
- PyTorch 1.5
Download the coco dataset into data/coco directory Download the json file below which collects all the required meta data for COCO into the data/coco directory Download the pretrained models given below.
To run interpolations between randomly sampled objects and compare results run the script as below.
CUDA_VISIBLE_DEVICES=X python compare_interpolations.py -m modelA.pth.tar modelB.pth.tar -n savename_modelA savename_modelB
Code to train the synthesizer network and to run adversarial attack will be released soon.
- Pre-trained synthesizer model (our part-segmentation bottlneck) - Trained on COCO dataset for matching 18 pascal object categories.
- Alternate synthesizer model (Gaussian bottlneck) - Trained on COCO dataset for matching 18 pascal object categories.
- COCO dataset file - Single json file with metadata and annotations for the COCO dataset
If you find this code useful in your work, please cite the paper.
PaperBibtex
@inproceedings{shetty2020SemAdv,
title={Towards automated testing and robustification by semantic adversarial data generation},
author={Shetty, Rakshith and Fritz, Mario and Schiele, Bernt},
booktitle={European Conference on Computer Vision (ECCV)},
year={2020},
}