PVG icon indicating copy to clipboard operation
PVG copied to clipboard

The problem of bigger scene reconstruction

Open huangqiusheng opened this issue 10 months ago • 4 comments

Admirable work! When I try to reconstruct a 198-frame waymo dynamic scene, the reconstruction effect is very poor. It seems that it can only succeed under 50 frames. Fine-tuning some hyper-parameters does not solve this problem. Please give me some advice, thanks!

huangqiusheng avatar Apr 02 '24 11:04 huangqiusheng

Admirable work! When I try to reconstruct a 198-frame waymo dynamic scene, the reconstruction effect is very poor. It seems that it can only succeed under 50 frames. Fine-tuning some hyper-parameters does not solve this problem. Please give me some advice, thanks!

Can you share the config and scripts you used on waymo scene? I want to use PVG on my waymo dataset. Thanks a lot.

cxzhou35 avatar Apr 22 '24 13:04 cxzhou35

Admirable work! When I try to reconstruct a 198-frame waymo dynamic scene, the reconstruction effect is very poor. It seems that it can only succeed under 50 frames. Fine-tuning some hyper-parameters does not solve this problem. Please give me some advice, thanks!

Can you share the config and scripts you used on waymo scene? I want to use PVG on my waymo dataset. Thanks a lot.

config from author

huangqiusheng avatar Apr 23 '24 07:04 huangqiusheng

Thanks!

How do you convert the raw .tfrecord waymo data to kitti format? I just use the sciprts provided by Waymo_kitti_converter, but the results can't be used directly.

cxzhou35 avatar Apr 24 '24 05:04 cxzhou35

@huangqiusheng hi, you could try to add "fix_radius: 10.0" to the waymo config to alleviate this problem. This parameter will fix the threshold controlling whether a point is going to split or clone (otherwise this threshold will grow as the scene becomes bigger and lead to a "under-split" case).

Fumore avatar Apr 24 '24 12:04 Fumore