rongliangzi / dense-scale-network-for-crowd-counting

An unofficial implement of paper "Dense Scale Network for Crowd Counting", link: https://arxiv.org/abs/1906.09707

Dense-Scale-Network-for-Crowd-Counting

An unofficial implement of paper "Dense Scale Network for Crowd Counting".

Dataset setup

Download the shanghaitech dataset from here, UCF-QNRF dataset from here.

Data preparation

In make_sh_gt.py, modify variable root in line 18 to your dataset path and set the min_size in line 16 for image. Then run the .py file. It will save images and .h5 file in root/{dataset}_preprocessed/train/ and root/{dataset}_preprocessed/test/.

Train

In main.py, set train_path to root/{dataset}_preprocessed/train/ and test_path to root/{dataset}_preprocessed/test/ in line 81 and 82. Also specify the save_path. When training shanghaitech PartA dataset, the model shows faster convergence if learning rate is set as 1e-4 compared to 1e-5 which is claimed by the paper.

Test

Test on one image

python test_one_image.py --gpu 0 --model_path pretrained_model_path --test_img_path your_image_path

Test on a dataset

python test_dataset.py --gpu 0 --model_path pretrained_model_path --test_img_dir your_image_directory

Result

Dataset	MAE	MSE
sha	69.35	104.4
shb	8.58	14.87
qrnf	tbd	tbd

Anyone interested in implementing crowd counting models is welcomed to contact me.

dense-scale-network-for-crowd-counting's Issues

Network architecture issues

Many thanks for your great work to attempt to reproduce the DSNet paper!
After reading your code, in particular the dsnet.py file, I notice 4 potential issues :

1. Vgg-18 ten first layers should be no trainable (not clear in the paper)
2. There is no bias for the dilated convolution layer inside DDCB block. According to the paper, it is similar to DenseASPP layer implementation example
3. At the end of DDCB block, there is no RELU layer according to the paper
4. Inside DDCB block again, the last concat should include x1_raw, ie. x3 = torch.cat([x, x1_raw, x2_raw, x3_raw], 1)
   I think there is a typo in the figure 2 of the paper, the DDCB paragraph overlines a full connection of dilation layers.

I am training a network with these changes on ShangaiTech B dataset, let see if I retrieve paper results....

Multi-scale density level consistency loss

Thanks again for your work!
Regarding the Multi-scale density level consistency loss, I think the L1 loss has to be reduced by sum instead of mean (default pytorch paramater).
Indeed, even if it it not specified clearly in the paper, as the norm is divided by the square of the average pooling output size, it is more logical to sum every component of the average pooling operation than average them.

What do you think?

Why don't you use 'batchnorm' in your model?

得到人的数量

网络输出一张热力图，最后是dmp.sum()得到一个count，但是这个count是一个小数，怎么获得这张热力图的实际人数呢？

How can I test the model and create.h5 file?

Hi,

I am a newbie who is studying crowd counting.
According to the survey paper, your work performs excellent.
I have traced your code and trained the model, it seems need a h5 to test your model? How can I use a photo from the internet to test your model? And how can I create a h5 file of it?
Thanks

cv2 interpolation to resize the target

It would be more accurate to use the INTER_AREA interpolation than the INTER_CUBIC one for the resize of the target.
Indeed, this downsampling is more natural here because it computes the weighted average of pixels.

Moreover, the rescale should be normalized with int(target.shape[0]) * int(target.shape[1])/ (int(target.shape[1]/ratio) * int(target.shape[0]/ratio)) rather than ratio**2.

It is not that minor, the current resize produces 1% of error.

I will make a PR for that!

Thanks

提示object 'density' doesn't exsit

您好，对于您的程序有几个问题：
1、程序中使用的是python2 还是python3？
2、Shanghaitech的数据集中groundtruth为.mat格式，程序中使用h5py读取，这样读出来之后为什么为空呢？
3、读取groundtruth之后，索引['density']键时，会提示该键不存在，可能是什么原因呢？