crowd-hat's Introduction

Boosting Detection in Crowd Analysis via Underutilized Output Features

Project Page | Paper

This repository contains the official PyTorch implementation of our paper Boosting Detection in Crowd Analysis via Underutilized Output Features. We provide a plug-and-play module leveraging the detection outputs as features to boost the performance in crowd analysis.

Boosting Detection in Crowd Analysis via Underutilized Output Features
Shaokai Wu*, Fengyu Yang*
Jilin University and University of Michigan
In CVPR 2023

This module works on top of the Head of detection methods and it can be tailored for each detection Head, thus we call it Crowd Hat. In this repository, Hat_LSC-CNN means Crowd Hat implemented on LSC-CNN, and likewise the others.

Implementation

Crowd Hat can be easily added on detection-based methods. Just follow the steps below:

Select a detection-based method and get the pipeline model. You can either train the model on crowd dataset or get the model from corresponding repositories (if any).
Fix the weights of the detection pipeline, do inference across the training dataset, and save the data to the disk.
Use the data above to train the Crowd Hat network, including the count decoder and region-adaptive NMS decoder.

This repository currently provide two pytorch implementation of Hat_LSC-CNN and Hat_TopoCount.

TODO

Crowd Hat + PSDNN and Crowd Hat + SDNet will be coming soon. SDNet is officially implemented with Tensorflow.

Citation

If you use this code for your research, please cite our paper.

@InProceedings{Wu_2023_boosting,
      author    = {Wu, Shaokai and Yang, Fengyu},
      title     = {Boosting Detection in Crowd Analysis via Underutilized Output Features},
      booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
      month     = {June},
      year      = {2023},
      pages     = {15609-15618}
}

Recommend Projects