This repository is not going to be updated anymore. The new detection model will be published here: TARTDetection

Code accompanying the paper "Expecting the Unexpected: Training Detectors for Unusual Pedestrians with Adversarial Imposters(CVPR2017)". As for the generator for synthetic data, please take this repo for reference.

• ubuntu or Mac OS
• tensorflow==1.1+
• pip install image
• pip install sklearn
• pip install scipy
• image_pylib(This repository should be put under the same folder with RPNplus.)

Run Demo:

• Download model files(RPN_model & VGG16_model) first, and put them in the ./models/ folder.
• The number 0 is your GPU index, and you can change to any available GPU index.
• This demo will test the images in the ./images/ folder and output the results to ./results/ folder.

python demo.py 0

Train:

• The number 0 is your GPU index, and you can change to any available GPU index.
• Open train.py and set imageLoadDir and anoLoadDir to proper values(imageLoadDir means where you store your images and anoLoadDir means where you store your annotation files).

python train.py 0

• Precarious Dataset
• Synthetic Dataset

• MIKKI Dataset

Please cite our paper if you use this code or our datasets in your own work:

@InProceedings{Huang_2017_CVPR,
author = {Huang, Shiyu and Ramanan, Deva},
title = {Expecting the Unexpected: Training Detectors for Unusual Pedestrians With Adversarial Imposters},
booktitle = {The IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
month = {July},
year = {2017}
}

• Our code is based on Yinpeng Dong's code and this repo: https://github.com/machrisaa/tensorflow-vgg

Shiyu Huang([email protected])