OpenDriveLab / ViDAR

[CVPR 2024 Highlight] Visual Point Cloud Forecasting
https://arxiv.org/abs/2312.17655
Apache License 2.0
277 stars 17 forks source link

About custom datasets and num_pred_height in config #38

Open qiuqc1 opened 3 months ago

qiuqc1 commented 3 months ago

When I try to use my own dataset to connect to vidar debugging, point_cloud_range = [-70, -150, -10.0, 70, 150, 10.0], num_pred_height = 16, but there are always some idx data whose loss in the first frame and the second frame is 0. The reason is that when generating cur_origin_points and cur_gt_points, the height of z exceeds num_pred_height, resulting in cur_valid_mask being all false when calculating loss, and subsequent dense.loss is all 0 image image image

I would like to know if the range of num_pred_height and point_cloud_range is related to this? I would like to ask if the author has ever encountered such a situation?

tomztyang commented 3 months ago

Hi,

how about adjusting the num_pred_height or some parameters related to voxel_size accordingly? I haven't tested the code with other ranges in the z-axis (e.g., [-10, 10] in your case).

Best, Zetong

counterattack-hub commented 3 months ago

Isn't 'cur_valid_mask' all about the dimension of x?