herbwood / CrowdDet

Detection in Crowded Scenes: One Proposal, Multiple Predictions

This is the pytorch implementation of our paper "Detection in Crowded Scenes: One Proposal, Multiple Predictions", https://arxiv.org/abs/2003.09163, published in CVPR 2020.

Our method aiming at detecting highly-overlapped instances in crowded scenes.

The key of our approach is to let each proposal predict a set of instances that might be highly overlapped rather than a single one in previous proposal-based frameworks. With this scheme, the predictions of nearby proposals are expected to infer the same set of instances, rather than distinguishing individuals, which is much easy to be learned. Equipped with new techniques such as EMD Loss and Set NMS, our detector can effectively handle the difficulty of detecting highly overlapped objects.

The network structure and results are shown here:

Citation

If you use the code in your research, please cite:

@InProceedings{Chu_2020_CVPR,
author = {Chu, Xuangeng and Zheng, Anlin and Zhang, Xiangyu and Sun, Jian},
title = {Detection in Crowded Scenes: One Proposal, Multiple Predictions},
booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
month = {June},
year = {2020}
}

Run

1. Requirements:

   • python 3.6.8, pytorch 1.5.0, torchvision 0.6.0, cuda 10.1

2. CrowdHuman data:

   • CrowdHuman is a benchmark dataset to better evaluate detectors in crowd scenarios. The dataset can be downloaded from http://www.crowdhuman.org/. The path of the dataset is set in config.py.

3. Steps to run:

   • Step1: training. More training and testing settings can be set in config.py.

   cd tools
   python3 train.py -md rcnn_fpn_baseline
   

   • Step2: testing. If you have four GPUs, you can use -d 0-3 to use all of your GPUs. The result json file will be evaluated automatically.

   cd tools
   python3 test.py -md rcnn_fpn_baseline -r 40
   

   • Step3: evaluating json, inference one picture and visulization json file. -r means resume epoch, -n means number of visulization pictures.

   cd tools
   python3 eval_json.py -f your_json_path.json
   python3 inference.py -md rcnn_fpn_baseline -r 40 -i your_image_path.png 
   python3 visulize_json.py -f your_json_path.json -n 3
   

Models

We use MegEngine in the research (https://github.com/megvii-model/CrowdDetection), this proiect is a re-implementation based on Pytorch.

We use pre-trained model from MegEngine Model Hub and convert this model to pytorch. You can get this model from here. These models can also be downloaded from Baidu Netdisk(code:yx46).

Model	Top1 acc	Top5 acc
ResNet50	76.254	93.056

All models are based on ResNet-50 FPN.

	AP	MR	JI	Model
RCNN FPN Baseline (convert from MegEngine)	0.8718	0.4239	0.7949	rcnn_fpn_baseline_mge.pth
RCNN EMD Simple (convert from MegEngine)	0.9052	0.4196	0.8209	rcnn_emd_simple_mge.pth
RCNN EMD with RM (convert from MegEngine)	0.9097	0.4102	0.8271	rcnn_emd_refine_mge.pth
RCNN FPN Baseline (trained with PyTorch)	0.8665	0.4243	0.7949	rcnn_fpn_baseline.pth
RCNN EMD Simple (trained with PyTorch)	0.8997	0.4167	0.8225	rcnn_emd_simple.pth
RCNN EMD with RM (trained with PyTorch)	0.9030	0.4128	0.8263	rcnn_emd_refine.pth
RetinaNet FPN Baseline	0.8188	0.5644	0.7316	retina_fpn_baseline.pth
RetinaNet EMD Simple	0.8292	0.5481	0.7393	retina_emd_simple.pth

Contact

If you have any questions, please do not hesitate to contact Xuangeng Chu ([e-mail]).