Awesome

IIM - Crowd Localization

This repo is the official implementation of paper: Learning Independent Instance Maps for Crowd Localization. The code is developed based on C3F.

Progress

• Testing Code (2020.12.10)
• Training Code
  • NWPU (2020.12.14)
  • JHU (2021.01.05)
  • UCF-QNRF (2020.12.30)
  • ShanghaiTech Part A/B (2020.12.29)
  • FDST (2020.12.30)
• scale information for UCF-QNRF and ShanghaiTech Part A/B (2021.01.07)

Getting Started

Preparation

• Prerequisites

  • Python 3.7
  • Pytorch 1.6: http://pytorch.org .
  • other libs in requirements.txt, run pip install -r requirements.txt.

• Code

  • Clone this repo in the directory (Root/IIM):
  • Download the pre-trained HR models from this link. More details are availble at HRNet-Semantic-Segmentation and HRNet-Image-Classification.

• Datasets

  • Download NWPU-Crowd dataset from this link.

  • Unzip *zip files in turns and place images_part* into the same folder (Root/ProcessedData/NWPU/images).

  • Download the processing labels and val gt file from this link. Place them into Root/ProcessedData/NWPU/masks and Root/ProcessedData/NWPU, respectively.

  • If you want to reproduce the results on Shanghai Tech Part A/B , UCF-QNRF, and JHU datasets, you can follow the instructions in DATA.md to setup the datasets.

  • Finally, the folder tree is below:

   -- ProcessedData
   	|-- NWPU
   		|-- images
   		|   |-- 0001.jpg
   		|   |-- 0002.jpg
   		|   |-- ...
   		|   |-- 5109.jpg
   		|-- masks
   		|   |-- 0001.png
   		|   |-- 0002.png
   		|   |-- ...
   		|   |-- 3609.png
   		|-- train.txt
   		|-- val.txt
   		|-- test.txt
   		|-- val_gt_loc.txt
   -- PretrainedModels
     |-- hrnetv2_w48_imagenet_pretrained.pth
   -- IIM
     |-- datasets
     |-- misc
     |-- ...

Training

• run python train.py.
• run tensorboard --logdir=exp --port=6006.
• The validtion records are shown as follows:
• The sub images are the input image, GT, prediction map,localization result, and pixel-level threshold, respectively:

Computational Cost

• Devices: Two TITAN RTX 3090-24GB GPUs
• Training Time: ~50 hours for large-scale NWPU dataset. ~6 hours for some small-scale datasets, such as SHHB, SHHA.

Testing and Submitting

• Modify some key parameters in test.py:
  • netName.
  • model_path.
• Run python test.py. Then the output file (*_*_test.txt) will be generated, which can be directly submitted to CrowdBenchmark

Visualization on the val set

• Modify some key parameters in test.py:
  • test_list = 'val.txt'
  • netName.
  • model_path.
• Run python test.py. Then the output file (*_*_val.txt) will be generated.
• Modify some key parameters in vis4val.py:
  • pred_file.
• Run python vis4val.py.

Performance

The results (F1, Pre., Rec. under the sigma_l) and pre-trained models on NWPU val set, UCF-QNRF, SHT A, SHT B, and FDST:

References

1. Deep High-Resolution Representation Learning for Visual Recognition, T-PAMI, 2019.
2. Very Deep Convolutional Networks for Large-scale Image Recognition, arXiv, 2014.
3. Feature Pyramid Networks for Object Detection, CVPR, 2017.

About the leaderboard on the test set, please visit Crowd benchmark. Our submissions are the IIM(HRNet) and IIM (VGG16).

Video Demo

We test the pretrained HR Net model on the NWPU dataset in a real-world subway scene. Please visit bilibili or YouTube to watch the video demonstration.

Citation

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

@article{gao2020learning,
  title={Learning Independent Instance Maps for Crowd Localization},
  author={Gao, Junyu and Han, Tao and Yuan, Yuan and Wang, Qi},
  journal={arXiv preprint arXiv:2012.04164},
  year={2020}
}

Our code borrows a lot from the C^3 Framework, and you may cite:

@article{gao2019c,
  title={C$^3$ Framework: An Open-source PyTorch Code for Crowd Counting},
  author={Gao, Junyu and Lin, Wei and Zhao, Bin and Wang, Dong and Gao, Chenyu and Wen, Jun},
  journal={arXiv preprint arXiv:1907.02724},
  year={2019}
}

If you use pre-trained models in this repo (HR Net, VGG, and FPN), please cite them.