vt-le / astnet Goto Github PK

Hi, is it possible to train the model? I could not find any code to train model.

I am wondering how to create the pre-trained model (UCSD Ped2, CUHK Avenue, ShanghaiTech Campus). Is this repository inference only?

Thank you.

Thanks for your sharing!I have a question about your models.In your model folder，there have two models named wresnet1024_cattn_tsm.py and wresnet2048_multiscale_cattn_tsmplus_layer6.py.What‘s the differences and connections between them?

How should I solve this situation when testing with the ped2 dataset？

I want to know how to make anomaly score graph in animation image, such as gif...
Can you tell be about that??

In the paper the "anomaly score" is defined as the normalized PSNR, denoted $S(t)$ in the paper, and calculated as such in the code.
However, higher PSNR means more "normal", so shouldn't the anomaly score be $1-S(t)$, which I believe is what is shown in the example animation and labeled as "anomaly score"?

Hello, Thank you for your work! I have a question. What is the purpose of the pretrained wider_resnet38.pth.tar file, and where can I find this file in the following link (https://github.com/mapillary/inplace_abn/)? Looking forward to your reply!

Let me start out by saying that this repo is very refreshing in being able to get up and running with a pretrained model, so thank you very much for that. I'm testing out the pretrained shanghaitech model, and I get 73.6% AUC which is correct according to the paper. When I qualitatively look at the anomaly score however, I see an odd periodic spiking which doesn't seem to appear in a similar visualization in the README. I'm wondering if this is different than expected? I've made an animated gif below, similar to the one in the readme (note that I'm plotting $1-S(t)$ which I believe should be the anomaly score).

Hi, execuse me, Where to obtain the ped2.mat file

Hello, I want to write the train model myself, but the AUC of the training model I wrote is only 79% on the ped2 dataset, you may help me

i used your pretrained shangaitech model for testing and got this assertion:

assert scores.shape == labels.shape, f'Ground truth has {labels.shape[0]} frames, BUT got {scores.shape[0]} detected frames!'
AssertionError: Ground truth has 40363 frames, BUT got 40389 detected frames!

Can you provide drive link for avenue dataset. Can you also please mention the frame per second (FPS) specification of all the datasets used for your models.

Hi what are the weights that you guys have used in calculating the loss function. Could you please help
mail: [email protected]

I am not able to download ped2 dataset from the provided link: http://www.svcl.ucsd.edu/projects/anomaly/dataset.html
could you provide with the drive link for the dataset if possible?

Hi tried building a training code for this, but got only 76% AUROC.
Could you help me with the debug where it's going wrong.
repo: https://github.com/Devansh-Sehrawat/astnet/blob/main/trainMod.py
mail: [email protected]

What part of the code is Residual channel attention block？

Please tell me， thank you

Hello, how many rounds do you get the best model for each dataset?

Hi, there are two Pillow versions in the requirements.txt. When you run pip install -r requirements.txt there is an error

ERROR: Cannot install Pillow==6.2.2 and Pillow==9.5.0 because these package versions have conflicting dependencies.

Which version should we use in the requirements.txt?

Hi，Where to obtain the avenue.mat file

Hello, Thank you for the code!
I have tried training with the ped2 dataset, but I got this error with the default pred2_wresnet.yaml.

python train.py \
    --cfg config/ped2_wresmet.yaml

RuntimeError: Given groups=1, weight of size [2048, 16384, 1, 1], expected input[2, 8192, 28, 36] to have 16384 channels, but got 8192 channels instead

Could you provide me with ped2_wresnet.yaml that works for the training of UCSped2? (http://www.svcl.ucsd.edu/projects/anomaly/dataset.html)

Thanks for the code , how can one detect anomaly on test video

Hi, I was evaluating a pre-trained model UCSDped2 using the command python test.py --cfg config/ped2_wresnet.yaml --model-file ped2.pth. After the 12 ephocs, I had this error AssertionError: Ground truth has 2012 frames, BUT got 1962 detected frames!. I have downloaded the ped2.pth and ped2.mat files correctly. Do you know how to solve this? Thank you.

Hello, I am trying to train and inference with your code.

Could you tell me if license is defined or not? (Apache License, Version 2.0, MIT, etc.)

Thank you.

Hello, thank you for sharing. I want to know how a picture like this is drawn? What data and code are needed?

When I used the shanghaitech dataset for training, I found that the original training dataset was a video file, so I performed a frame extraction operation, but I still couldn't reach the 274514 training frames mentioned in your paper. In addition, when I used the shanghaitech.pth file you provided for testing, the following error occurred: AssertionError: Ground truth has 0 videos, BuT got 107 detected videos! I suspect it's an issue with the dataset, as the shanghaitech dataset does not have files similar to ped2. mat like ped2. I hope you can answer my questions. Thank you.

Hi, thanks for your work, i am wondering could you share the avenue test dataset which you used for this project please, email([email protected]) or google driver. Thank you.

Hello dear,
First of all thank you for sharing this with us
Can I apply this model on my project ? If so can you provide me with steps ? I want to use it to detect the crime activity in video surveillance so what are the is pre processing steps?

Thanks in advance

Hi,

Thank you for providing the repo. May I know where can I download the wider_resnet38.pth.tar file that is supposed to be in ../../datasets/seg_weights folder mentioned in the ped2_wresnet.yaml file.

Thank you