inhwanbae/npsn

Official Code for "Non-Probability Sampling Network for Stochastic Human Trajectory Prediction (CVPR 2022)"

npsn's Issues

some issues about GCS dataset

Thanks for your great work!
I am a little confused about the data preprocessing for GCS Dataset. In the paper, you used the first 3.2s (8 frames) as the observed trajectory and the next 4.8s (12 frames) as GT. However, I checked the paper [59], and found that the FPS is 25 and the annotated fps is 1.25. Thus, the time interval between two neighboring annotated frames is 1/1.25=0.8s. This means 8 steps in the past trajectory spend 8*0.8=6.4s, which is different from your paper. I am wondering if my understanding is correct.
Also, could you please release the preprocessing code or provide more details about the data preprocessing for GCS dataset (train/test split)?

Question about the dimension of the Box-muller transform

Hi,

I have a question about the dimension of the Box-muller transform. In my understanding, the Box-muller transform should work on the number of samples (n) dimension. However, in the code, it seems that you use the BM tran. on the number of feature (s) dimension. See:

NPSN/train_npsn.py

Lines 71 to 78 in 97f56cf

	elif args.baseline == 'pecnet':
	obs_traj, pred_traj, mask, x, y, initial_pos, _ = model_forward_pre_hook(batch, data_sampler=data_sampler)
	# NPSN
	loc = model_npsn(obs_traj.unsqueeze(dim=0).transpose(-1, -2), mask=mask)
	loc = loc.squeeze(dim=0).permute(1, 0, 2)
	loc = box_muller_transform(loc)
	# PECNet
	all_guesses = model_forward(model, x, initial_pos, loc)

After line 75, loc has the shape of n, batch size, s.

And the bull_muller_transform() here

NPSN/npsn/utils.py

Lines 7 to 14 in 97f56cf

	def box_muller_transform(x: torch.FloatTensor):
	r"""Box-Muller transform"""
	shape = x.shape
	x = x.view(shape[:-1] + (-1, 2))
	z = torch.zeros_like(x, device=x.device)
	z[..., 0] = (-2 * x[..., 0].log()).sqrt() * (2 * np.pi * x[..., 1]).cos()
	z[..., 1] = (-2 * x[..., 0].log()).sqrt() * (2 * np.pi * x[..., 1]).sin()
	return z.view(shape)

seems to work on the last dimension of the tensor x, which is the s dimension.

Please let me know if there is anything wrong with my understanding. Many thanks!

Reproducing the results from the rest of the models and datasets

Could you please provide details on reproducing the results for the rest of the models and datasets?
Is there any particular reason you did not provide code for the other baselines and datasets?

Box-Muller

Thanks for your great work!
But I am confused about the Box-Muller transformation. I know how to do it, but i want to know why we should do it. I'll appreciate it if you could give me an answer.

PECNet+NPSN on SDD

Hi,
Thanks for the nice work, I have a question regarding reproducing the numbers in Table 1 on SDD with PECNet+NPSN, which is 8.56/11.85 on minADE/minFDE.

Do you have the pretrained models, and also the training pipeline?

Thanks in advance!