search

agrimgupta92 / sgan

Code for "Social GAN: Socially Acceptable Trajectories with Generative Adversarial Networks", Gupta et al, CVPR 2018
MIT License
810 stars 259 forks source link

annotation sampling is different than original ETH #84

Open os1a opened 4 years ago

os1a commented 4 years ago

Hi,

Thanks for your nice work, I am wondering why the annotations are sampled differently than the original ETH?

For example, in ETH: your dataset contains annotations for frames (780, 790, 800.....) and the total number of annotations (num of lines in biwi_eth.txt) is 12380. While in the original ETH (https://icu.ee.ethz.ch/research/datsets.html), the annotations are provided for frames (780, 786, 792,....) and the total number of annotations is 8908.

Am I missing somthing?

zhangpur commented 4 years ago

Hi, The author said they sampled data by linear interpolation. When comparing with the original data in real word coordinates, the annotation of the same person at the same frame are identical. It seems that the sample rate of sgan was wrong in the author's pre-processing. But it's still confusing why there are more annotations.

JarlLemmens commented 2 years ago

The author mentions in the paper that the data is interpolated to obtain values at every 0.4 seconds, but then they use a sample rate of 10 frames (or frame IDs at least). But in the original ETH they also claim that the data is already annotated for 0.4 seconds with a sample rate of 6 frames. So it seems that the author's time steps are off by a factor 1.6. This is not correct right? Would it work to simply replace the ETH data file with a correct sample rate, or does it have to be steps of 10? Did anyone perhaps solve this already?

zhangpur commented 2 years ago

Hi JarlLemmens, I agree with you that the sample rate is incorrect. Absolutely it can work with a correct sample rate. You can refer to the paper 'Social-aware pedestrian trajectory prediction via states refinement LSTM' of the performance comparison on the corrected sample rate.