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

Question about the datasets used for training #61

Open cuihenggang opened 5 years ago

cuihenggang commented 5 years ago

I find the training instruction says

We support five datasets ETH, ZARA1, ZARA2, HOTEL and UNIV. We use leave-one-out approach, train on 4 sets and test on the remaining set.

I am a little bit confused by this. Does it mean for each model (e.g., ETH model), we will actually use data that are outside the corresponding dataset to train this model? This looks very unnatural. I thought for each model (e.g., ETH model), we should train it using only the data in that dataset (e.g., ETH dataset), right?

Can someone explain?

amiryanj commented 5 years ago

@cuihenggang That's true. They dont use the data from test environment in training. Because the goal is to learn the 'social behavior' which should be independent of the env. Well, in my point of view, this is a nice approach to have a generalized social model, but at the same time it would ignore all env dependencies during prediction which is not a good idea.

cuihenggang commented 5 years ago

Are the evaluation results on the Social-GAN paper also evaluated in this way? I am just curious whether in the paper they train each models with only one dataset or multiple datasets.

simmonssong commented 3 years ago

@amiryanj Hi, Did you check the distributions of train-set, validation-set, and test-set? Are the proportions between different kinds of trajectories the same? For instance, turning left, turning right, or going straight forward.

amiryanj commented 3 years ago

Hey @simmonssong, In OpenTraj we studied the Angular Deviation for different datasets: we didn't find a significant "deviation from linear motion" in ETH and UCY datasets. Though, we didn't compute the values for trajnet data.