Learning Hierarchical Graph for Occluded Pedestrian Detection

This is the official implementation of our paper "Learning Hierarchical Graph for Occluded Pedestrian Detection", https://dl.acm.org/doi/abs/10.1145/3394171.3413983, published in ACM MM 2020.

To address the occlusion issue in pedestrian detection, we propose a novel Hierarchical Graph Pedestrian Detector (HGPD), which integrates semantic and spatial relation information to construct two graphs named intra-proposal graph and inter-proposal graph. With the intra-proposal graph, we can model precise occlusion patterns and effectively suppress noisy features; with the inter-proposal graph, the weak visual cues of occluded persons can be enriched. The graph structure is shown:

Get Started

Please refer to INSTALL.md for installation and dataset preparation.

Please see GETTING_STARTED.md for the basic usage of MMDetection.

Prepare

CityPersons data:

• CityPersons is a popular benchmark for pedestrian detection. The dataset can be downloaded from here. The path of the dataset is set in configs/faster_rcnn_vgg16_1x.py. We also provide our annotation files in json format (code:kko8).

Train

CUDA_VISIBLE_DEVICES=gpu_id python tools/train.py configs/faster_rcnn_vgg16_citypersons.py --work_dir xxx

Some errors will happen when using distributed training, we will fix this bug later.

Test

CUDA_VISIBLE_DEVICES=gpu_id python tools/test.py configs/faster_rcnn_vgg16_citypersons.py path_to_your_model --eval box --out path_to_save_detection_results

Models

All models are trained and evaluated on CityPersons with the input scale of 1x. To show the occlusion handling of our method, we employ different subsets of training samples, which differ in the occlusion level.

	Reasonable	Heavy Occlusion	Model
visibility ≥ 0%	11.51	41.34	code:kko8
visibility ≥ 30%	12.24	42.65	code:kko8
visibility ≥ 50%	11.53	45.91	code:kko8

Citation

If you find HGPD useful in your research, please consider citing:

@inproceedings{li2020learning,
  title={Learning Hierarchical Graph for Occluded Pedestrian Detection},
  author={Li, Gang and Li, Jian and Zhang, Shanshan and Yang, Jian},
  booktitle={Proceedings of the 28th ACM International Conference on Multimedia},
  pages={1597--1605},
  year={2020}
}

Contacts

If you have any questions, please do not hesitate to contact gang li (李钢), gang.li@njust.edu.cn