The performance of stereo vision of nitendo 3ds is so excellent. The merged image from the two camera is much better than any other images I have seen, also it runs in a real-time mode. So what's the exact stereo match algorithm for its stereo cameras?
Dear Yu Luo,
Look the links, may be useful.
http://nefeli.lib.teicrete.gr/browse/stef/epp/2015/ChoustoulakisEmmanouil/attached-document-1446199791-653661-11173/ChoustoulakisEmmanouil2015.pdf
http://www.cs.cmu.edu/~clz/ZK00Stereo.pdf
Regards, Shafagat
Dear Professor.Shafagat Mahmudova,
Thanks a lot for your reply! Your links are very useful!
I'm wondering that why we just use the simple color data for the stereo match, not more advanced data, like SIFT features. Based on these precise data, the final result might be more accurate.
Regards,
Yu
Under commercial pressures, it is likely that it is a performance issue decision, the more complex and accurate the algorithm the the more computation it is going to be required. This is on top of all the processing that is usually done for gaming. besides sometimes good is good enough.
Hi Arturo Geigel,
Thanks for your reply!
I have read many papers about the stereo match problem and implement some in my program. And I reconstructed the 2D image in the computer's 3D space with the disparity data computed by these algorithms. The 3D image is never as continuous and smooth as the Nitendo's.
And I tried to capture a toothbrush about 30cm away. I could even feel the depth difference of different parts of the toothbrush. That's much better than my work. And it's well implemented on the FPGA.
If try to achieve its accuracy, I can try some methods used by 3D scanner, but it's very slow. And I think Nitendo has a perfect solution, but I just don't know it.
Regards
Yu
What kind of pre processing are you doing? Also, are you processing in color, B&W or both? How good is the contour image that you can obtain without the stereo matching? Can you isolate the target(toothbrush) enough in a non stereo environment?
The answer to these questions will help in the assessment of the basics before going into stereo matching and reconstruction(assuming that is your target).
Hope this helps
Hi Arturo Geigel,
For now, I just use the original SSD(sum of suqared difference) stereo match with a little information about the gradient of center pixel. This method performced better than simple SSD match, but still many miss-match disparity.
The contour image you mentioned, I think it will be a good suggestion. The match based on the contour image will reduce the time cost and increase the accuracy. The contour image may isolate all the target in the image by colors.
For me, I'm considering some algorithm like SIFT features. Not just use the color of the matched pixels, but more scale-invariant information. With the epipolar constraint, the matched disparities would be more accurate.
Thanks for your advice!
Yu
Yu,
Let me know your progress and if you need anything post it here or you can directly write to me.
Regards
- Badges
- Science topic
- Similar topics
- Computer Science and Engineering
- Algorithms