search

agrimgupta92 / sgan

Code for "Social GAN: Socially Acceptable Trajectories with Generative Adversarial Networks", Gupta et al, CVPR 2018
MIT License
813 stars 261 forks source link

Why I could not get the same result with the paper? #41

Closed mingbocui closed 5 years ago

mingbocui commented 5 years ago

I just run evaluate_model.py with the trained model provided by the author, but the error of ADE and FDE are both much bigger than the results displayed in the paper. Is there anything else I have to take care when running the evaluate_model.py?

HorryChen commented 5 years ago

Hello, I just started to contact the related research on pedestrian trajectory prediction. This is the first time that I contacted this project such as SGAN. I met some difficulties and doubts in my study. Would you like to give me your email address? Thank you. And my email address is wisionpeng@163.com. I'm looking forward to your reply. Sincerely Wision Peng

amiryanj commented 5 years ago

We have also explored a new trajectory prediction architecture, called Social-Ways. We will present this work at CVPR workshop. The paper will be released soon (on arxiv and CVF), and I would be interested to see your feedbacks. javad.amirian@inria.fr

amiryanj commented 5 years ago

The mentioned paper is now available on arxiv: https://arxiv.org/abs/1904.09507

We will start to publish the source codes, ASAP.

oyjGithub commented 5 years ago

Can you reproduce the experimental results in this paper? I have problems with the results of my operation, and the gap is very large.@amiryanj

amiryanj commented 5 years ago

I just used the pre-trained models and the results are quite the same as in the paper. @onlylove321

mingbocui commented 5 years ago

@amiryanj have you tried just used the hyperparameters provided by the scripts and train the model by yourself? I have got different results, hope for your insights.

amiryanj commented 5 years ago

Hi @mingbocui I'm afraid that I've never get that prediction accuracies by training, always thinking I'm doing something wrong. I just know that the l2_loss_weight is set to zero while it should be something, otherwise you may let such GAN run for few days to get converged.

ybpaopao commented 5 years ago

Hi, @mingbocui, have you found any method that can solve the problem you mentioned? I also trained SGAN with different hyperparameters but the results are far away from the paper.