imenebak / uav-crowd-counting

SOFT-CSRNET : Counting people in drone video footage

VisDrone_CC

This is the Soft CSRNET version repo for ECCV2020(Challenge-CrowdCounting), which delivered an optimization of the parameters of the deep CNN "CSRNET" which made the density estimation in real time.

Datasets

CC visdrone Dataset: web_site

Prerequisites

We used Google Colab as a perncipal environment to train and test our models.

Ground Truth

Please follow the DensityVisDrone.ipynb to generate the ground truth. You may use Visdrone2019_dotAnnotation.ipynb to change the type of the annotation in Visdrone VID 2019 dataset or else.

Training Process

Please follow the first part of TrainVal.ipynb to start training process.

Validation

Follow the second part TrainVal.ipynb to try the validation. You can try to modify the notebook and see the output of each image.

Test

Follow the Test_model.ipynb to test your model on 2020 ECCV CC. You can try to modify the notebook and see the output of each image.

References

Code

On this repos we based on the keras implementation of CSRNet by : https://github.com/Neerajj9/CSRNet-keras
We're finalyzing also the Pytorch implementation : https://github.com/imenebak/CSRNet-pytorch

SOFT-CSRNet paper

@INPROCEEDINGS{9378749,
  author={Bakour, Imene and Bouchali, Hadia Nesma and Allali, Sarah and Lacheheb, Hadjer},
  booktitle={2020 2nd International Workshop on Human-Centric Smart Environments for Health and Well-being (IHSH)}, 
  title={Soft-CSRNet: Real-time Dilated Convolutional Neural Networks for Crowd Counting with Drones}, 
  year={2021},
  volume={},
  number={},
  pages={28-33},
  doi={10.1109/IHSH51661.2021.9378749}}
  

CSRNet paper.

@inproceedings{li2018csrnet,
  title={CSRNet: Dilated convolutional neural networks for understanding the highly congested scenes},
  author={Li, Yuhong and Zhang, Xiaofan and Chen, Deming},
  booktitle={Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition},
  pages={1091--1100},
  year={2018}
}

ECCV2020 Challenge DroneCrowd.

@article{zhu2018vision,

title={Vision meets drones: A challenge},
author={Zhu, Pengfei and Wen, Longyin and Bian, Xiao and Ling, Haibin and Hu, Qinghua},
journal={arXiv preprint arXiv:1804.07437},
year={2018} }

@article{zhu2020vision,
title={Vision Meets Drones: Past, Present and Future},
author={Zhu, Pengfei and Wen, Longyin and Du, Dawei and Bian, Xiao and Hu, Qinghua and Ling, Haibin},
journal={arXiv preprint arXiv:2001.06303},
year={2020} }

uav-crowd-counting's Issues

Can you pls make a guide about dataset and order of use of notebooks

In the DensityVisDrone.ipynb VisdroneCrowdCounting/Visdrone.zip is utilized, I am not able to find this in https://github.com/VisDrone/VisDrone-Dataset
It would be incredibly helpful if you can tell us the exact dataset used here/ possible share the google drive link to your VIsdroneCrowdCounting

ValueError: Unknown loss function: accuracy

Hello!

First of all, thank you for this detailed walkthrough of your work. It's quite enjoyable to follow in the Colab notebooks :)
I have some doubts in some lines of code in TrainVal notebook and I would really appreciate if you can help me understand what is going on.

In the definition of the CrowdNet function, there is a line almost at the end stating:
model.compile(optimizer=Adam(1e-5), loss='accuracy', metrics=[mae, "mse"])

However, in the next blocks before fitting the model, it appears another compilation stating:
model.compile(optimizer=Adam(1e-5), loss='mae', metrics=[mae, "mse", 'accuracy'])

Why are they different?
When I check the code for CSRNet, both lines have the same structure:
model.compile(optimizer=sgd, loss=euclidean_distance_loss, metrics=['mse'])

When I try to make both lines equal in Soft-CSRNet, I can't use "accuracy" as loss because it throws the "Unknown loss function" error. I get no errors with "mae" as loss, but then I cannot manage to replicate the same range of values in the Y-axis of the "Training Loss on Dataset" plot.

Thanks in advance for your support.

Installation

How do we install and run this in the cloud?

Can u give me ur weights

hi author, i want ur weight for my research. Can u give me ur weight. Tks.

能放下你们训练出来的权重模型嘛？？

你们的代码有些问题
1、Trainval.ipynb中的create_img()函数上面那段代码无法运行.
2、划分的训练集、验证集、测试集是哪些.
3、Trainval.ipynb没有train_sum这个变量，没有定义logging这个变量
4、Trainval.ipynb中img = cv2.resize(target,(int(target.shape[1]//8),int(target.shape[0]//8)),interpolation = cv2.INTER_CUBIC)*64为什么用//