Bad feature matching

asked 2016-01-07 17:09:49 -0500

laxn_pander gravatar image

Hello there people,

I am currently working with OpenCV 3.1 trying to make some realtime stitching of aerial images possible. However I am just starting and do have my troubles with feature matching. It's just not good enough and I don't really know what may left to do.

Workflow: - Loading images - Detecting features with ORB - Extracting features with ORB - Matching with BruteForce (Hamming-(2)) - Ratio test - Symmetry test - Calculating Homography using RANSAC - Warping

But still, my matches (without RANSAC) look like that:

Matched Features

edit retag flag offensive close merge delete

Comments

1

Your problem is the texture of the parts you want to stitch. Your ORB features on trees and sand are not unique and can be easily reproduced by other parts of the image. Thats the downside of having not enough unique features by buildings, roads, ... not sure if you can do much about it. Can you post the stitching result?

StevenPuttemans gravatar imageStevenPuttemans ( 2016-01-08 03:20:20 -0500 )edit

I was afraid that is the answer to it. But yesterday I tested a little bit with the amount of image features in relation to image size, realizing this was my biggest issue. My images were full sized with way too few features. As a result I got the warped image below pre resizing. After resizing my results still don't match perfectly, but I guess without bundle adjustment that's the best shot I can get.

Warped image pre resizing

Warped image post resizing

laxn_pander gravatar imagelaxn_pander ( 2016-01-09 06:53:37 -0500 )edit

I think the second result is quite good for this kind of textureless images... Some research of a previous collegue I worked with might be interesting for you. Here you can find his poster of his publication on textureless mapping, related to 3D dense matching, but might be interesting here too.

StevenPuttemans gravatar imageStevenPuttemans ( 2016-01-11 02:41:00 -0500 )edit