生产运营/行业运营>> multi-stream-crowd-counting-extended>> 返回
项目作者： RQuispeC

项目描述 ：
Code for "Multi-Stream Networks and Ground-Truth Generation for Crowd Counting" presented in IJECES
高级语言： Python
项目地址: git://github.com/RQuispeC/multi-stream-crowd-counting-extended.git
创建时间： 2019-07-15T13:53:35Z
项目社区：https://github.com/RQuispeC/multi-stream-crowd-counting-extended
开源协议：MIT License
下载


Multi-Stream Networks and Ground-Truth Generation for Crowd Counting

Code of “Multi-Stream Networks and Ground-Truth Generation for Crowd Counting” presented in The International Journal of Electrical and Computer Engineering Systems 2020.


1. Environment

We used the following enviroment:

  • Python 3
  • PyTorch
  • OpenCV
  • Numpy
  • MatPlotLib
  • Ubuntu 16.04

You can also run the code using the docker image of ufoym/deepo.

To use ground truths generated by our face method, you have to install MXNET and other dependencias based on this repo.

2. Preparing data

We make available UCF-CC-50 and Shanghai Tech datasets here, download and unzip it into the root of the repo. Directories should have the following hierarchy:

  1. ROOT_OF_REPO
  2.     data
  3.         ucf_cc_50
  4.             UCF_CC_50
  5.                 images
  6.                 labels
  7.         ShanghaiTech
  8.             part_A
  9.                 train_data
  10.                     images
  11.                     ground-truth
  12.                 test_data
  13.                     images
  14.                     ground-truth
  15.             part_B
  16.                 train_data
  17.                     images
  18.                     ground-truth
  19.                 test_data
  20.                     images
  21.                     ground-truth

The code was developed such that data augmentation is computed before every other step and the results are stored in the hard drive. Thus, the first time you run the code it will take quite a long time. Augmented data is stored with the following hierarchy:

  1. ROOT_OF_REPO
  2.     data
  3.         ucf_cc_50
  4.             people_thr_0_gt_mode_same
  5.         ShanghaiTech
  6.             part_A
  7.                 people_thr_0_gt_mode_same
  8.             part_B
  9.                 people_thr_0_gt_mode_same

To run face method augmentation UCF-CC-50 dataset:

  1. python3 tiny_detection_mxnet.py -d ucf-cc-50 --save-plots --root-dir YOUR-DATA-DIR
  2. python3 main.py -d ucf-cc-50 --augment-only --force-augment --force-den-maps --gt-mode face

3. Training

To train using UCF-CC-50 (with all folds) and save the results log in log/multi-stream you can run:

  1. python3 main.py -d ucf-cc-50 --train-batch 32 --save-dir log/multi-stream

In case you want to run a specific fold or part you can use flag --units, check the Makefile for more examples.

The training log is stored in log_train.txt inside the corresponding log/fold/part directory.

4. Testing

After training you can re-load the trained weights (using flag --resume) and use them for testing:

  1. python3 main.py -d ucf-cc-50 --save-dir log/multi-stream --resume log/multi-stream/ucf-cc-50_people_thr_0_gt_mode_same --evaluate-only

The testing log is stored in log_test.txt inside the corresponding log/fold/part directory. You can also generate the plots of the predictions using flag --save-plots, results are stored in the directory plot-results-test inside the corresponding log/fold/part directory.

5. Final Notes

If you have any question regarding the code or method don’t hesitate to contact me. Further examples for running this code are available in the Makefile