Hi, I am kinda confused about pseudo lidar points from depth map ?

[nile649]

can I use depth map from any architecture to create pseudo lidar points? GT lidar points are required to train model to predict pseudo lidar points from generated depth map?

[mileyan]

Yes, you can I use depth map from any architecture to create pseudo lidar points. But if you fine tune your model on the KITTI training set LIDAR points, you will get a better result. But the difference should be less than 5%.

[nile649]

My question is how can I create pseudo lidar points, if I don't have velodyne ground truth for the training dataset. I am planning to test on custom dataset. Thanks

[mileyan]

You need the calib, image_2(left image) and image_3(right image) folders. And you need to generate the calibration files as the same format as KITTI object detection dataset.

[ghost]

@mileyan so when using a custom dataset, what transformation matrix should I use for Tr_velo_to_cam in the absence of any LiDAR device at all?

EDIT:
Sorry, wasn't thinking clearly last night! It's just a rotation and translation matrix. So would need to figure out where my camera would be relative road surface and camera2 on KITTI, then translate it to the KITTI LiDAR position. Correct?