points error

[howthep]

Thanks for your works. I'm testing on KITTI-360, I found the synthesized point stored as .npy. So I use this to convert it to pcd.

import numpy as np
import open3d as op

path = "./log/kitti360_lidar4d_f4950_release/results/test_lidar4d_ep0639_0002_depth_lidar.npy"
points = np.load(path)
print(points.shape,points[:10])
pcd = op.geometry.PointCloud()
pcd.points = op.utility.Vector3dVector(points)
op.io.write_point_cloud("2.pcd",pcd)

But it looks like the depths are wrong. Below image is captured from pcl_viewer

Below image is corresponding depth map

This is the output of evaluation

Could you please help me find where is wrong?

[dyfcalid]

Hi, there.
The results of the evaluation are wrong, but the depth map looks normal. It seems that the depth scale is not correct. You can check if the data is preprocessed in the right way. We have provided an example kitti360_4950.txt in the configs folder, and the scale is about 0.01.
If the problem is still not solved, you can follow the instructions to double-check each step, or offer me more details.

[howthep]

this is the kitti360_4950.txt

dataloader = kitti360
path = data/kitti360
sequence_id = 4950
num_frames = 51
fov_lidar = [2.0, 26.9]
scale = 0.010504329815187737
offset = [1307.2250542680918, 3604.6116677769596, 110.30188369430098]

this is the command I run, I have trained it for 639 epoch

main_lidar4d.py \
--config configs/kitti360_4950.txt \
--workspace log/kitti360_lidar4d_f4950_release \
--lr 1e-2 \
--num_rays_lidar 1024 \
--iters 30000 \
--alpha_d 1 \
--alpha_i 0.1 \
--alpha_r 0.01 \
--test_eval 

before the code converting depth to lidar, I find the pred_depth's values span a narrow range from 0.199 to 0.201.

pred_lidar = pano_to_lidar(
                    pred_depth / self.opt.scale, loader._data.intrinsics_lidar
                )

I do not know how to solve this
Can you provide some instructions to double-check , or what details should I offer?

[dyfcalid]

It's weird, and I've never encountered it before. Could you provide information about your machine and environment, as well as the training and evaluation log (first and last several epochs)?

[howthep]

Yes, that's so weird. I remove all checkpoints and train again. It looks normal. Thans for your help.

last question

Could you provide your code for visualization, or which library do you use to visualize?
Like this, that's more pretty.

[dyfcalid]

OK.👌 You can try open3d for visualization and customize the color.