Home

Awesome

Illuminating Pedestrians via Simultaneous Detection & Segmentation

Garrick Brazil, Xi Yin, Xiaoming Liu

Introduction

Pedestrian detection framework using simultaneous detection and segmentation as detailed in arXiv report, accepted to ICCV 2017.

Our SDS-RCNN framework is derivative work of Faster R-CNN and RPN+BF. Tested with Ubuntu 14.04, CUDA 7.5, Matlab 2016a, Titan X GPU, and a modified version of Caffe v1.0 as provided. Unless otherwise stated the below scripts and instructions assume cwd in MATLAB is the project root of SDS-RCNN.

@inproceedings{brazil2017illuminating,
    title={Illuminating Pedestrians via Simultaneous Detection \& Segmentation},
    author={Brazil, Garrick and Yin, Xi and Liu, Xiaoming},
    booktitle={Proceedings of the IEEE International Conference on Computer Vision},
    address={Venice, Italy},
    year={2017}
}

Setup

Training

Training both stages takes about 18 hours on a single Titan X.

rpn_config  = 'caltech_VGG16_weak_seg';
rcnn_config	= 'caltech_VGG16_weak_seg';
gpu_id = 1;

% train both stages
train_all(rpn_config, rcnn_config, gpu_id);

Testing

We provide the collective SDS-RCNN trained models for RPN and BCN (7.36% MR), as well as the RPN only file with cost-sensitive off (9.63% MR). There are associated artifact files of anchors, bbox_stds, bbox_means, and basic configurations which should be loaded into memory at test time as depicted below. All files are packed into SDS-RCNN-Release.zip.

load('rpn_conf.mat');
load('rcnn_conf.mat');
load('anchors.mat');
load('bbox_means.mat');
load('bbox_stds.mat');
gpu_id = 1;

% test RPN only
test_rpn(test_prototxt_path, weights_path, rpn_conf, anchors, bbox_means, bbox_stds, gpu_id)

% test RPN and BCN (full SDS-RCNN)
test_rcnn(test_prototxt_path, weights_path, rpn_conf, anchors, bbox_means, bbox_stds, ...
    rcnn_prototxt, rcnn_weights, rcnn_conf, gpu_id)