Ask Your Question

Increasing depth accuracy with more than two cameras

asked 2020-03-09 11:40:22 -0600

myoung gravatar image

I'm looking for advice to achieve a high resolution depth map from multiple cameras with OpenCV. My subject is a flower, which has both smooth petal surfaces as well as grass-like hairs protruding from the surface towards the camera. Because these hairs do not have a large surface area and rapidly change their depth, they are difficult to match across the images.

My thought is to use multiple cameras in a grid to both minimize holes in the reconstruction and improve accuracy. For example, with four cameras placed on the vertices of a square, you'd have 6 unique image pairs each with a known baseline.

My questions:

1) Is my subject (a flower) a good fit for stereo matching, or should I use other technology like laser scanners? I'm looking for millimeter accuracy.

2) Is using multiple cameras a good way to improve quality, or should I focus on higher-resolution cameras and more CPU horsepower?

3) Is there anything in OpenCV that would help me stitch together the image pairs? I've seen techniques for doing this with three cameras, but it assumes they are colinear which seems less optimal.

4) Does OpenCV endorse any consulting firms that would help me build this out?


edit retag flag offensive close merge delete

1 answer

Sort by ยป oldest newest most voted

answered 2020-03-20 13:01:14 -0600

I have a partial answer:

1) You may have two different applications of stereo matching here. You may need to filter the input images to: enhance the lower frequency smooth surfaces and solve for smooth, wide changes (petals), and then enhance the images for high-frequency narrow items pointing toward the camera, and solve that (stamens/pistils), then merge the results intelligently, rejecting the parts that are not of interest.

The stereo algorithms are not ideal for detecting depth of narrow things pointed right at the camera, but they can resolve surfaces that are more parallel to the camera fairly well. To image the narrow stems, you may need to have depth cameras imaging the scene more obliquely.

2) Multiple cameras are an excellent way to fill in parts of the solution hidden by tall objects pointing toward the camera. I've used this technique to good results. One needs to consider point cloud vs. depth map vs. modeled shape merging and filtering though, to solve this problem, and there are a lot of choices there.

I won't comment on stitching since that is not my area of expertise.

edit flag offensive delete link more

Question Tools

1 follower


Asked: 2020-03-09 11:40:22 -0600

Seen: 711 times

Last updated: Mar 20 '20