Skip to content

hustaia/Decoupled_Two-stage_Counting

Repository files navigation

Decoupled Two-Stage Counting

This repository includes the official implementation of D2CNet for crowd counting and localization, presented in our paper:

Decoupled Two-Stage Crowd Counting and Beyond

IEEE Transactions on Image Processing, 2021

Jian Cheng, Haipeng Xiong, Zhiguo Cao, Hao Lu

Huazhong University of Science and Technology

equal contribution

Overview

Visualization

Highlights

  • New representation of crowd: We introduce the concept of the probability map which describes the probability of a point being an object;
  • Simple and effective: With only a basic network architecture and simple training procedures, D2CNet obtains state-of-the-art counting and localization performance;
  • Multifunctional: D2C framework not only unifies object counting and object localization but also enables synthesized data augmentation and flexible counting error diagnosis.

Installation

The code has been tested on Python 3.7.6 and PyTorch 1.5.0. Please follow the official instructions to configure your environment. See other required packages in requirements.txt.

Data Preparation

  1. Download the official ShanghaiTech and UCF_QNRF dataset under ./data. Resize images in UCF_QNRF to ensure the longer side is not greater than 1920.
cd prepare_dataset
python normalize_QNRF.py
  1. Generate synthesized dataset by running generate_synthesized_dataset/main.m in Matlab.

  2. Generate probability maps, density maps and dot maps. It may take several days to prepare the UCF_QNRF dataset, so it's better to split it into several parts and prepare them in parallel.

python make_dataset.py
  1. Generate json files of training and testing data.
cd ..
python prepare_dataset/filepath2json.py ./data/shanghaitech/part_A_final/train_data/images/ part_A_train.json

Training

python train.py part_A_train.json part_A_test.json 0 0

Testing

Download the pretrained models to the main folder.

PartA (MAE: 55.673, MSE: 87.161) BaiduYun (code: 3qal)

PartB (MAE: 6.367, MSE: 10.716) BaiduYun (code: ybfp)

QNRF (MAE: 81.698, MSE: 137.925) BaiduYun (code: xrjq)

python train.py --pre partA.pth.tar part_A_train.json part_A_test.json 0 0 --mode 1

Mode 1 is for counting evaluation, and mode 2 is for localization evaluation.

Permission

The code are only for non-commercial purposes. Copyrights reserved.

Contact: Jian Cheng (jian_cheng@hust.edu.cn) and Hao Lu (hlu@hust.edu.cn)

About

No description, website, or topics provided.

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published