search

microsoft / AirSim

Open source simulator for autonomous vehicles built on Unreal Engine / Unity, from Microsoft AI & Research
https://microsoft.github.io/AirSim/
Other
16.07k stars 4.49k forks source link

Point cloud stiching #1405

Open xiahaa opened 5 years ago

xiahaa commented 5 years ago

Version: 1.2.0 Binary Coding: Python

Problem: I tried to do point cloud stiching using several depth images (depth planner). However, the result showed there may be some synchronization issues for scene image and depth image. result for 1080P 1080p result for QVGA qvga Clearly, you can see some dragging effect. Scene and depth images were recorded with "Recording" function. I did tranformation like $$$ P_ground = RP_camera + t $$$ where R is constructed using quaternion.

Can anyone give some ideas about the problem? Thanks

xiahaa commented 5 years ago

confirm again using sfm software. 2 tests with same setting, same image, the only difference are:

  1. let pose being camera pose logged with Airsim recording. Since those poses should be ground truth, I made uncertainty of those poses to be very low, 1mm.
  2. don't use any prior pose, just let SFM do automatic feature matching to determine the camera poses. Result: 2 is better than 1, more points, less artefacts.

My guess right now would be: At least for the recording, it seems that pose and image are not exactly aligned. And this problem becomes severe if the resolution of the image increases, let's say from VGA to 4K, the delay will increase.

Could anyone provide some information about this alignment issue?

Many thanks.