Image presorting

[lole-elol]

Trying to presort images based on similarity in order to increase point cloud quality. Algorithms:

• PCA
• PCA on image histograms
• PCA on image fft2

[lole-elol]

DELIVERY PCA

PCA HIST

PCA FFT2

GARDEN PCA

PCA HIST

PCA FFT2

KICKER PCA

PCA HIST

PCA FFT2

PIPES PCA

PCA HIST

PCA FFT2

RELIEF PCA

PCA HIST

PCA FFT2

STATUE PCA

PCA HIST

PCA FFT2

TERRAINS PCA

PCA HIST

PCA FFT2

[danielsanjosepro]

Sieht gut aus. 1D projection ist leider nicht gut gemug, wir müssen da was besseres machen.

[lufixSch]

I think the results are pretty good. Especially using Histogram or FFT2 before running the PCA.

I too think more dimensional projections/features could help. We could try using more dimensions from the PCA or combining FFT2 and HIST features into one vector. Maybe the second one would help as they both (HIST, FFT2) perform better in different cases but I don't know how this would affect computation time.

Besides Traveling Salesman we could also try nearest Neighbour to find a path in n dimensional space. I know it's not as exact as traveling salesman but much faster and easier to implement.

[lufixSch]

I added resorting to the 3D Reconstruction Pipeline in order to benchmark the results.

For this to be easier I combined all resort algorithms in one function. You can now choose between different algorithms with a function argument.

[lufixSch]

I will run some rough benchmarks using only one of the datasets for now.

@lole-elol tomorrow we can take a closer look at the results and/or run some more benchmarks and tune the algorithms.

[lufixSch]

Results of the rough benchmarks:

HIST and FFT2 produce okay results but only if the nearest neighbor sorting with 1D results from the PCA is used

Any other settings produce complete garbage results.

[danielsanjosepro]

@lole-elol @lufixSch is this ready for review?

[lufixSch]

@lole-elol @lufixSch is this ready for review?

I would like to wait until the meeting tomorrow.

Edit: I read "merge"... Would be nice If you could take a look at it. But I think we will talk about the Algorithms tomorrow.

[lufixSch]

As far as I'm concerned this can be merged, if you are okay with the structure of the function and it's position in the 3d reconstruction pipeline.