NOTE

We are currently updating this repository. We are preparing training and testing launchers, will be pushed soon. Moreover, we will very soon upload some pre-trained models for different benchmarks.

Pedestron

Pedestron is an MMetection based repository that focuses on the advancement of reserach on pedestrian detection. We provide processed annotations and scripts to process the annotation of different pedestrian detection benchmarks.

Installation

We refer to the installation and list of dependencies of MMdetection to the official installation file. Clone this repo and follow installation of mmdetection.

List of detectors

*Currently we provide configurations for Cascade Mask-RCNN

Following datasets are currently supported

• Caltech
• CityPersons
• EuroCity Persons
• CrowdHuman
• WiderPedestrian Challenge

Preparation

1. Download the datasets from the official sites. Fill in the copyright forms where applicable.
2. Place them in ./datasets folder in the follwoing heararchy, for example CityPersons should be (./datsets/CityPersons/images/) and (./datsets/CityPersons/annotations/) for images and annotations respectively.
3. Use our pre-processing script to convert the annotations into Pedestron acceptable format.

Pre-Trained models

1. CityPersons.
2. Caltech.
3. EuroCity Persons

Please cite the following work

ArXiv version

@article{irtiza20elephant,
  title={Pedestrian Detection: The Elephant In The Room},
  author={Hasan, Irtiza and Liao, Shengcai and Li, Jinpeng and Akram, Saad Ullah and Ling, Shao},
  journal={arXiv preprint},
  year={2020}}