/PAMS-FCN

Primary LanguageJupyter NotebookMIT LicenseMIT

A Part-Aware Multi-Scale Fully Convolutional Network for Pedestrian Detection

Introduction

We present a part-aware multi-scale FCN for pedestrian detection. In this method, the part-aware RoI pooling module can generate high detection confidence score for occluded pedestrians, while the multi-scale FCN is constructed to detect small-scale and large-scale pedestrians respectively on feature maps of different resolutions.

Requirements

  1. Software: Please use the Microsoft-version Caffe and follow the usual instructions.

  2. Hardware: NVIDIA GPU with 8GB or larger memory is required.

Installation

  1. Clone the PAMS-FCN repository into $PAMS_ROOT
git clone https://github.com/ypeiyu/PAMS-FCN.git
  1. Build the Cython modules
cd $PAMS_ROOT/lib
make
  1. Build Caffe and pycaffe
cd $PAMS_ROOT/caffe
make -j8 && make pycaffe

Train/Eval

  1. It should be prepare the training, val and test data as VOC format.
  2. Train/Test the PAMS-FCN detector. Outputs are svaed under $PAMS_ROOT/output/
cd $PAMS_ROOT
./experiments/scripts/pams_fcn_end2end/[_ohem].sh [GPU_ID] [NET] [DATASET] [--set ...]

Main Results

  1. We conduct an ablation study to validate the effectivenss of the proposed network.

Table 1. Result of the ablation study on the Heavy Occlusion subset of the Caltech Dataset and the Caltech-New dataset. Runtime is evaluated on a single GTX 1080Ti GPU per image.

ablation study

  1. The main detection results can be found here.