Fuzzy image output with camera moverment

[Aozhi]

Excuse me. I have a question of the robustness of the algorithm. If I keep the camera center fixed, there's no ghost in the output image. But if I move the camera, then ghost will probabaly appear. Do you have solution on this problem?

[ppwwyyxx]

With random camera movement, the stitching problem is not a mathematically well-defined problem at all. There is no elegant way to do this.

[Aozhi]

Thank you for the explanation.

[Aozhi]

Do you have suggested parameters in config.cfg if I stitch photos using a smart phone? Also, you mention OpenMP in your code to speed up the running. Can I have it compiled just like compiling the Eigen or libjpeg?

[ppwwyyxx]

No suggestions. It depends on how the stitching goes. OpenMP is a feature of the compiler. If you use gcc or a sufficiently new clang it will be there.

[Aozhi]

Hi. Have you ever tried with other feature detectors? I saw there's a brief detector in your code. Have you tested its performance?

[ppwwyyxx]

That's for fun only . It's never really used and not very useful.

[Aozhi]

Thank you for your patient reply. This is a very good program of panorama stitching from scratch. Have you ever programmed the same thing with OpenCV? How do you think of the performance of stitching API of OpenCV?

[ppwwyyxx]

I think the stitching quality of OpenCV isn't very good. And it's slower.

[Aozhi]

Thx. I also think OpenCV is slow.

I find one more problem that: if I would like to stitch the images even though the overlap situation is bad, how should I change the code? Since I found a sentence that: Found a tree size 11!=12 when the overlap is not good.

[ppwwyyxx]

You'll need more features or more matches. A number of configs can have this effect, but it's hard to say which to use without looking at the data.

[Aozhi]

Hello, thank you so much for the explations. I noticed that you mention RootSIFT in your blog. Is this change already included in your code?

[ppwwyyxx]

yes

[Aozhi]

struct Edge { int v1, v2; float weight; Edge(int a, int b, float v):v1(a), v2(b), weight(v) {} bool operator < (const Edge& r) const { return weight < r.weight; } };

Excuse me, I notice that you define a graph and use it in the file camera_estimate.cc. What are the nodes and edges representing? What is the purpose of the traverse here? Thank you.

[ppwwyyxx]

Node is each image. Edge is a matching pair. weight is how confident the matching is. traverse gives an order to do bundle adjustment one image by one image.

[Aozhi]

Thank you again.

[Aozhi]

One more question. In some cases the program will give the error "the tree size is != (a number)", "image n and image m are no connected well". In which case this error will occur?

[ppwwyyxx]

When not all images are not in the same connected components in the graph, which means certain pairs of images are not matched very well. And there could be a lot of reasons for that..