Pedestron

Pedestron is a MMetection based repository that focuses on the advancement of research on pedestrian detection. We provide a list of detectors, both general purpose and pedestrian specific to train and test. Moreover, we provide pre-trained models and benchmarking of several detectors on different pedestrian detection datasets. Additionally, we provide processed annotations and scripts to process the annotation of different pedestrian detection benchmarks. If you use Pedestron, please consider citing respective sources.

Updates

• [NEW] Added configurations and pre-trained model for Hybrid Task Cascade (HTC) and MobileNet
• [NEW] Evaluation code for the Caltech dataset, added to the repository
• [NEW] Ported testing of Mask-Guided Attention Network for Occluded Pedestrian Detection (MGAN). (ICCV'19), into Pedestron along with pretrained model
• [NEW] Added a pre-trained model for generic person detection

YouTube

• YouTube link for qualitative results on Caltech. Pre-Trained model available.

Installation

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

List of detectors

Currently we provide configurations for with different backbones

• Cascade Mask-R-CNN
• Faster R-CNN
• RetinaNet
• Hybrid Task Cascade (HTC)
• MGAN

Following datasets are currently supported

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

Datasets Preparation

• We refer to Datasets preparation file for detailed instructions

Benchmarking of Pre-Trained models

Detector	Dataset	Backbone	Reasonable	Heavy
Cascade Mask R-CNN	CityPersons	HRNet	7.5	28.0
Cascade Mask R-CNN	CityPersons	MobileNet	10.2	37.3
Faster R-CNN	CityPersons	HRNet	10.2	36.2
RetinaNet	CityPersons	ResNeXt	14.6	39.5
Hybrid Task Cascade (HTC)	CityPersons	ResNeXt	9.5	35.8
MGAN	CityPersons	VGG	11.2	52.5
Cascade Mask R-CNN	Caltech	HRNet	1.7	25.7
Cascade Mask R-CNN	EuroCity Persons	HRNet	4.4	21.3

Pre-Trained models

Cascade Mask R-CNN

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

Faster R-CNN

1. CityPersons

RetinaNet

1. CityPersons

Hybrid Task Cascade (HTC)

1. CityPersons

MGAN

1. CityPersons

Getting Started

Running a demo using pre-trained model on few images

1. Pre-trained model can be evaluated on sample images in the following way

python tools/demo.py config checkpoint input_dir output_dir

Download one of our provided pre-trained model and place it in models_pretrained folder. Demo can be run using the following command

python tools/demo.py configs/elephant/cityperson/cascade_hrnet.py ./models_pretrained/epoch_5.pth.stu demo/ result_demo/ 

Training

Train with single GPU

python tools/train.py ${CONFIG_FILE}

Train with multiple GPUs

./tools/dist_train.sh ${CONFIG_FILE} ${GPU_NUM} [optional arguments]

For instance training on CityPersons using single GPU

python tools/train.py configs/elephant/cityperson/cascade_hrnet.py

Training on CityPersons using multiple(7 in this case) GPUs

./tools/dist_train.sh configs/elephant/cityperson/cascade_hrnet.py 7  

Testing

Test can be run using the following command.

python ./tools/TEST_SCRIPT_TO_RUN.py PATH_TO_CONFIG_FILE ./models_pretrained/epoch_ start end\
 --out Output_filename --mean_teacher 

For example for CityPersons inference can be done the following way

1. Download the pretrained CityPersons model and place it in the folder "models_pretrained/".
2. Run the following command:

python ./tools/test_city_person.py configs/elephant/cityperson/cascade_hrnet.py ./models_pretrained/epoch_ 5 6\
 --out result_citypersons.json --mean_teacher 

or without mean_teacher flag for MGAN

python ./tools/test_city_person.py configs/elephant/cityperson/mgan_vgg.py ./models_pretrained/epoch_ 1 2\
 --out result_citypersons.json  

• Similarly change respective paths for EuroCity Persons
• For Caltech refer to Datasets preparation file

Please cite the following work

ArXiv version

@article{hasan2020pedestrian,
  title={Pedestrian Detection: The Elephant In The Room},
  author={Hasan, Irtiza and Liao, Shengcai and Li, Jinpeng and Akram, Saad Ullah and Shao, Ling},
  journal={arXiv preprint arXiv:2003.08799},
  year={2020}
}