Pedestrian detector for Lua/Torch7

Train/test/benchmark a pedestrian detector using lua/torch7. These detector uses a modified Fast R-CNN network + a pedestrian oriented roi proposal generator for detection.

The available networks for feature extraction for use are the following:

• AlexNet
• ZeilerNet
• VGG16/19
• Googlenet (inception v3)
• ResNet (18, 32, 50, 101, 152, 200)

Note: Some of these networks require GPUs with 10Gb+ ram to be trained.

For Region-of-Interest (ROI) proposal generation, this code has three methods for roi proposal generation:

• ACF (paper, code)
• LDCF (paper, code)
• EdgeBoxes (paper, code)

Installation

Requirements

To run the code in this repository you'll need the following resources:

• Torch7
• Fast R-CNN module
• dbcollection
• Python (>=2.7 or >=3.5)
• Matlab >= 2012a (for benchmark and roi proposal generation)

Packages/dependencies installation

To use this example code, some packages are required for it to work: fastrcnn and dbcollection.

fastrcnn

To install the Fast R-CNN package do the following:

• install all the necessary dependencies.

luarocks install tds
luarocks install cudnn
luarocks install inn
luarocks install matio
luarocks install torchnet

• download and install the package.

git clone https://github.com/farrajota/fast-rcnn-torch
cd fast-rcnn-torch && luarocks make rocks/*

For more information about the fastrcnn package see here.

dbcollection

To install the dbcollection package do the following:

• install the Python module (Python>=2.7 and >=3.5).

  pip install dbcollection==0.1.7
  

  or

  conda install -c farrajota dbcollection==0.1.7
  

• install the Lua/Torch7 dbcollection wrapper:

  1. download the Lua/Torch7 git repo to disk.

  git clone https://github.com/dbcollection/dbcollection-torch7
  

  2. install the package.

  cd dbcollection-torch7 && luarocks make
  

For more information about the dbcollection package see here.

Getting started

Usage

To start using the code, clone this repo to your home directory:

git clone --recursive https://github.com/farrajota/pedestrian_detector_torch

If you clone the repo into a different directory, please make sure you modify projectdir.lua and point to the new path before using the code.

Data setup

The necessary data is available for download by calling the following command in the terminal:

th download/download_all.lua

This will download the following data:

• pre-trained models on Imagenet
• pre-processed roi proposals on the caltech pedestrian dataset
• algorithm scores of several methods for benchmark

Note: this data can also be downloaded manually by following the next steps.

Download pre-trained models and benchmark algorithm results

Several pre-trained models are available to be used in this code. To download all networks, simply run the following command in the terminal:

th download/download_pretrained_models.lua

For benchmark, other evaluation algorithms need to be downloaded before proceeding to evaluate the trained network. To download all available algorithms simply do the following:

th download/download_extract_algorithms.lua

RoI Proposals

To generate region proposals, run th generate_rois.lua on the terminal. This will automatically generate the RoI proposals needed for training and testing the pedestrian detector.

Warning: processing the region proposals takes around a week to process. For this reason, pre-processed proposals are available for download:

th download/download_proposals.lua

Dataset

To run the code, first it is necessary to setup the dataset's data. This code uses the dbcollection package for data setup/management.

To setup a dataset do the following:

dbc = require 'dbcollection'
caltech = dbc.load{name='caltech_pedestrian',
                   data_dir='save/files/to/dir'}

This will download and pre-process the dataset's data and store the all files to the selected path. If data_dir is not defined or left empty, the data files will be stored in the dbcollection/ folder in your home directory in a folder with the dataset's name.

In case you already have the necessary data files, you can manually set the directory path of the files by doing the following commands:

dbc = require 'dbcollection'
caltech = dbc.load{name='caltech_pedestrian',
                   data_dir='path/to/dataset/files'}

Note: All datasets used here can be downloaded via dbcollection.

To download and extract the relevant data, please run the following scripts: th download_extract_dataset.lua -save_dir <store_dir_path> and download_extract_algorithms.lua -save_dir <store_dir_path>. The -save_dir <store_dir_path> allows for the user to save the downloaded data into another directory than the root dir of the code.

Available datasets

The following datasets are available for training/testing a pedestrian detector using this repo:

• Caltech Pedestrian Dataset

Train and test a model using the example code

This repository contains scripts for training and testing a pedestrian detector network using a pre-trained network on ImageNet for feature extraction such as the alexnet or vgg.

Note: several options are available for configuring the training/testing parameters (see options.lua for a complete set of available parameters).

Training a network

To train a model run th train.lua. To change the default settings, use the input arguments of your choice. To see all available option's parameters do th train.lua --help or check options.lua.

• You can select one of the following imagenet pre-trained networks for feature extraction: AlexNet, ZeilerNet, VGG (16, 19), ResNet (19, 34, 50, 101, 152, 200), and GoogleNet v3.

• Snapshots of the network's training procedure are stored to disk other information such as the configuration file, loss logs and model parameters of the training procedure (default path is ./data/exp). You can change this directory by passing the -expDirnew path to save the experiment option.

Testing a network (mAP accuracy)

To test a network's accuracy, run th test.lua and define the -expID, -dataset and -expDir input options (if changed) to compute the mean average-precision.

Note: The test script only uses a single GPU.

Demo: detecting persons

To run the basic demo code you'll first need to train a network model. After this is done, just simply run the demo on the terminal:

qlua demo.lua -expID <exp_name>

This selects random images from the testing set and it should give you a glimpse of how the net performs.

Benchmark evaluation

This repo contains scripts for benchmarking the results of your network against other top-performing methods on the Caltech dataset. This requires that evaluation results of other algorithms have been downloaded in order to run.

After this is setup, simply run th benchmark.lua -expID '<exp_name> and this will generate a series of plots which are stored in the experiment's folder in ./data/.

The default name for the method is OURS. To change to any other name just set a different name using -eval_plot_name input option.

License

MIT license (see the LICENSE file)

Acknowledgements

The evaluation/benchmarking code is an adaptation of @pdollar toolbox/evaluation code.