diptodip / counting Goto Github PK

First I want to thank you for your paper and this repository, which really helped me a lot. Now I am working on a counting task but not for counting people, also each image could have several hundreds objects. This may different to human crowd counting.

The annotation point, should it be the center of an object or other part, because these objects not like humans always stand on the ground and up head. For instance I want to count how many animals in an image, animals could have a lot more complicated poses and occlusions.

I am new to computer vision area, some detailed information within your paper and reference papers currently not very clear to me, so I want to know what directions should I go? What hyper parameters should I adjust for this task?

I have already trained a AMDCN model about animal counting, it performs good, but due to lack of image types, the generalization ability of the model is bad. Anyway, I tried to change some hyper parameters, then I found out that none of these models performed better than your original default hyper parameters. Why did this happen? the task is very different. So I'm here to get some inspiration from you.

Thank you!

Hello, @diptodip

Thank you for sharing such beautiful work.

When I run train_models.py on UCF-50 datesets, it occurs
tensorflow.python.framework.errors_impl.InvalidArgumentError: Incompatible shapes: [1,380,380,1] vs. [1,256,256,1]

I cannot solve it. Can you give me some advices?

hi @diptodip ,
do you have a plan to release AMDCN? if have, when do you release? thanks.

Hi there,

I downloaded the UCF dataset, there are 50 .jpg and 50 .mat files, but where can I find or generate label .png files, quote from your README file: "The label images should be named as ndots.png where n is the original image number. Once the label images are in data/UCF/B/, run the included blur.py in data/UCSD/B/ which will generate the ground truth density maps. "

Thanks!

I want to predict number of persons in a crowd image. But I haven't found any option in your code for this. In test_model.py file

X = f["data"]
y = f["label"]
"y_predict = model.predict_generator(gen.flow(X,y, 0, 0, None), X.shape[0])"

This means that I have to annotate my jpg image to generate label. Then using this label, I have to predict count! This is not feasible for me. Because I have to deploy an IP camera on crowd scene and feed the image to your model. Then I hope your model can give the crowd count. Can it be done using your model?

I am a new student for the research about crowd counting. Now I'm in trouble in the WorldExpo'10 datasets' 'roi.mat'. I don't not how to use it to make it's mask. Can you help me? Thanks a lot!

Hi,

Can you tell me the detials about training WorldExpo'10? Only train one model for 5 scenes testing data?
Looking forward to your reply.

Thanks a lot!