In this work, we introduce a new learning target named local counting map, and show its feasibility and advantages in local counting regression. Meanwhile, we propose an adaptive mixture regression framework in a coarse-to-fine manner. It reports marked improvements in counting accuracy and the stability of the training phase, and achieves the start-of-the-art performances on several author- itative datasets. For more details, please refer to our arXiv paper.

• Prerequisites

  • Python >= 3.5
  • Pytorch >= 1.0.1
  • other libs in requirements.txt, run pip install -r requirements.txt.

• Data Preparation

  • Download ShanghaiTech, UCF-QNRF, UCF_CC_50 datasets from the official websites and unzip them in ./ProcessedData.
  • Run cd ./datasets/XXX/ and python prepare_XXX_mod64.py to resize images and generate training labels.
  • UCF-QNRF: [home link], ShanghaiTech: [home link], UCF-CC-50: [home link]

• Pretrained Model (Only for Training)

  • Some Counting Networks (such as VGG, CSRNet and so on) adopt the pre-trained models on ImageNet. Download vgg16-397923af.pth from torchvision.models.
  • Place the pre-trained model to ./models/Pretrain_model/.
  • vgg16-397923af.pth : download link

• Folder Tree

  +-- source_code
  |   +-- datasets
      |   +-- SHHA
      |   +-- ......
  |   +-- misc     
  |   +-- models
      |   +-- Prerain_Model
      |   +-- SCC_Model
      |   +-- ......
  |   +-- ProcessedData
      |   +-- shanghaitech_part_A
      |   +-- ......
  

QNRF-model (MAE/MSE: 86.6/152.1):

Google Drive: download link, Baidu Yun: download link (key: pe2r)

• put test images in ./demo_image.
• run python demo.py.
• results are saved at ./demo_image/result.

• set the parameters (such as MODEL_PATH) in test_config.py.
• run python test.py.

https://github.com/gjy3035/C-3-Framework