Source code for paper "Localization Guided Learning for Pedestrian Attribute Recognition".
Accepted by British Machine Vision Conference (BMVC), 2018.
Pengze Liu, Xihui Liu, Junjie Yan, Jing Shao
Originally implemented on a customized Caffe with several additional layers:
- CalOverlap: Takes the IoU between two boxes as the affinity between proposals and attribute boxes. Ref: This Paper
- CamBox: Extract the "Class-Activation-Box" from the CAMs. Check CAM (B. Zhou, 2016) for more details.
- SigmoidCrossEntropyWeightLoss: Add a weight factor by the positive/negative ratio of each attribute. Check DeepMar for more details.
- DilatedConvolution: Implementation of Dilated Convolution with backpropagation. Check Multi-scale context aggregation with dilated convolutions for more details.
@article{liu2018localization,
title={Localization guided learning for pedestrian attribute recognition},
author={Liu, Pengze and Liu, Xihui and Yan, Junjie and Shao, Jing},
journal={arXiv preprint arXiv:1808.09102},
year={2018}
}