NOH-NMS: Improving Pedestrian Detection by Nearby Objects Hallucination

The offical implementation for the “NOH-NMS: Improving Pedestrian Detection by Nearby Objects Hallucination” which is published in ACM MM 2020.

We propose Nearby Objects Hallucinator (NOH), which pinpoints the objects nearby each proposal with a Gaussian distribution, together with NOH-NMS, which dynamically eases the suppression for the space that might contain other objects with a high likelihood.

This work has won the first place at the CrowdHuman Challenge, 2020.

This repo is implemented based on detectron2.

Performance

Model	Backbone	AP	Recall	MR	Weights
Faster RCNN	ResNet-50	85.0	87.5	44.5	faster_rcnn_model_final.pth
NOH-NMS	ResNet-50	88.8	92.6	43.7	noh_nms_model_final.pth

Prepare Datasets

Download the CrowdHuman Datasets from http://www.crowdhuman.org/, and then move them under the directory like:

./data/crowdhuman
├── annotations
│   └── annotation_train.odgt
│   └── annotation_val.odgt
├── images
│   └── train
│   └── val

Installation

  cd detectron2
  pip install -e . 
  #or rebuild
  sh build.sh

Quick Start

See GETTING_STARTED.md in detectron2

Acknowledgement

detectron2

Citation

if you find this project useful for your research, please cite:

@inproceedings{zhou2020noh,
  title={NOH-NMS: Improving Pedestrian Detection by Nearby Objects Hallucination},
  author={Zhou, Penghao and Zhou, Chong and Peng, Pai and Du, Junlong and Sun, Xing and Guo, Xiaowei and Huang, Feiyue},
  booktitle={Proceedings of the 28th ACM International Conference on Multimedia},
  pages={1967--1975},
  year={2020}
}