cschoeller/constant_velocity_pedestrian_motion

Code for "What the Constant Velocity Model Can Teach Us About Pedestrian Motion Prediction", Schöller et al., Robotics and Automation Letters (RA-L), 2020

constant_velocity_pedestrian_motion's Issues

Question about the ADE in metrics.py

Hi, thanks for sharing your codes. You have done a great job. But I have a question about the ADE in the codes:

constant_velocity_pedestrian_motion/metrics.py

Lines 4 to 15 in 7fe0716

	def avg_disp(y_pred, y_true):
	""" Average displacement error. """
	y_true, masks = y_true
	seq_lengths = masks.sum(1)
	batch_size = len(seq_lengths)
	squared_dist = (y_true - y_pred)**2
	l2_dist = masks * torch.sqrt(squared_dist.sum(2))
	avg_l2_dist = (1./batch_size) * ((1./seq_lengths) * l2_dist.sum(1)).sum()
	return avg_l2_dist.item()

avg_l2_dist = (1./batch_size) * ((1./seq_lengths) * l2_dist.sum(1)).sum()
Why do you still multiply 1./batch_size while you have already multiplied 1./seq_lengths. I can't figure out the reason. Let's assume that all the positions in the predicted trajectory are valid, which means seq_legths = 12 and batch_size = 12. So in your code, you divide the l2_dist.sum(1) by 12*12. I think the right way is to divide it by 12. Could you tell me the reason why you do that?

Figure generation code

Hi,

My name is Abbie Lee, and I'm a senior at MIT who is working on intent inferencing for autonomous systems as part of a final project for a robotics class. My group is very interested in using CVM as a baseline for pedestrian motion prediction for another algorithm we are working with, and I was wondering if you would be able to provide the code for generating the figures that appear in your paper.

Thanks!
Abbie

Could you provide the full code for <FloMo: Tractable Motion Prediction with Normalizing Flows>？

Dear Christoph,

I am very sorry to put forward the issue on this repository since I cannot see the issue in the corresponding repository.

Thank you for your work. Your work is excellent and inspires me. However, due to the inability to handle all details, I would like to know if you can provide the full code for <FloMo: Tractable Motion Prediction with Normalizing Flows> for better understanding of your excellent work.

Best regards,
Honghui

Results table1

Hi,

first of all thanks for your work. In table 1, the result of social gan is much better than the improved version reported in,
https://github.com/agrimgupta92/sgan/blob/master/MODEL_ZOO.md.
Are you sure that you evaluation is the same?

Dataset pre-processing

Hi Christoph,

First of all, thanks for sharing your code. I was wondering, how did you generate the .json files in the folder ../data/eth_univ/data/. For instance, if we look inside the file "3.json" we get the next entry: {"timestamp": 80, "object_list": [{"position": [10.67, 3.99], "id": 1}, {"position": [13.64, 5.8], "id": 2}], "frame": 80, "size": 2} but the original dataset (obsmat.txt file) gives the next entry the first time pedestrian with id:2 appears:

8.0400000e+02 2.0000000e+00 1.3017548e+01 0.0000000e+00 5.7825914e+00 -2.3244442e+00 0.0000000e+00 -7.6605938e-02

Where 1.3017548e+01 and 5.7825914e+00 are the x, y coordinates. So, my question is where is 13.64 (x of pedestrian 2) coming from? The same thing applies to many other positions. Thanks in advance.