Why can only point clouds of objects within four meters be output?

[xihaofei]

Hello, I use the quantized int8 model (lightstereo-m-sceneflow-general) on rk3588 for stereo matching, and finally convert the obtained parallax map (the obtained parallax map is normal) into point cloud data. Currently, point clouds can be output normally, but only for objects within 4 meters.
My camera parameters are as follows:
Resolution: 640*480@60FPS
Focal length: 3.0mm
Field of view: 80° (no distortion)
Baseline: 12cm
What are the factors that affect the point cloud distance, and how to improve the distance of the output point cloud? Looking forward to your reply.

[zz990099]

Could you provide the depth2cloud code？

[xihaofei]

Hello, using the image as input, the output disparity map is as follows:

Thank you for your reply. At this moment, I can only output point clouds at a distance of 4 meters. This morning, when testing with a 4.2mm camera, I found that point clouds for objects over 5 meters can be output. May I ask how far the point cloud information for this model (Lightstereo-M-SceneFlow-General) can be output? How can this distance be estimated in practice? I look forward to your reply.

[zz990099]

This issue may be caused by insufficient model accuracy. According to the official repository, when testing the Lightstereo-M-SceneFlow-General model, input images with lower resolution are used to verify its inference speed. However, at lower resolutions, the EPE accuracy is actually quite poor. At higher resolutions, the accuracy can reach the metrics described in the paper.