Dense points genereation from images that mounted on a car

[JyunPingJhan]

Hi

I am working on a project for pavement inspection, where two six-lens 360 spherical camera is mounted on a car for street view image collection. I use original unstitched images obtained from each lens, and import them as a multicamera system (the rig are calibrated in advance) for photo alignamet and pavement dense cloud generation.

Each project contains 10000 groups of images (total of 80000 images), and each group is equaly extracted in 1 m distance. It takes several hours for alignment, but I have encountered a problem in the processing of depth maps generation.

I understand that depth map generation is to use selected neighbor points (default is 100, and I modifed it to -1) that obtained from valid matches in overlapped area. 

However, since this is a close-range enviorment project, what I am interest is the pavement area and the overlap images are limit, i.e. neghbioring 5 groups, there is no necessary to process depth maps that from two groups of images at distance longer than 5 meters, even they are overlapped and have valid matches on buildings.

Therefore, I am wondering is there a way that I can select a group of images for depth map generation instead of using matches, as seaking all possible image pairs for depth generation is really time consuming. Though adjusting the tweaks may help, but is still not a good solution as the vaild matchs are also varied scene by scene, it is hard to chose a proper value.

The 100 default value is very quik, but resulting in missing points in the pavement and the confidence is too low, whiel -1 value gets complete and more confident dense points on pavement but the processing time is at least 10 times longer. 

As my project scene is stable, hope there is a option for me to process depth map in only a certain distance range.

Attatched are some of my results.

[Bzuco]

Hi,

I would try extract images every 2m or even every 2.5 - 3m, because 1m is too often if you are driving on more distant road lane. As one camera is always pointing perpendicular to the building, the second two neighbours cameras are really needed as a support for additional angles needed for better filtering unwanted noise and adding more details into pointcloud. Therefore, processing group of cameras only in 5m distance would be insufficient for depth maps/dense cloud.

With less photos and increasing neighbours tweak parameters you should be able to achieve what you need - processing speed and preserve dense cloud without holes on pavement.

[JyunPingJhan]

Hi Bzuco

Thanks for your comment, but 2.5-3 m is too long as neighbor images have only a small overlap on the pavement.
I need highly overlapped images to reduce the hole and increase the confidence of pavement dense clouds.

Processing 5 groups of images in 5 meter range (this value is ajdustable) is to reduce the unnecessary depth map generation.
For example, group 1 and group 10 has no pavement overlap but they have matches on surrounding buildings,  in such case, metashape will use this pair to generate depth map that I don't want if the tweak is set to -1.

[Kiesel]

Hi JyunPingJhan,

I think the camera you have choosen isn't the right one to get a good pavement inspection. You use only a very small part of your sensor for that (less than 50%?, most of the photo area is masked out) and additional the camera looks at a very low angle at the pavement, which isn't good for inspection too.
I would rather use a normal camera facing down to the street for that what you want to get. The camera needs to be mounted on a higher point for that. This way you don't have the problem with the unwanted pairs too.


Best regards,

Kiesel