Benchmarking

[lufixSch]

Tuning Hyperparameters

[lufixSch]

Was able to tune hyper parameters for 3d reconstruction.

Was able to get some small improvements but no major changes.

• Delivery Area: good
• Kicker: recognizable but distorted
• Old Computer: Unrecognizable (For me. I don't know the dataset well)
• Pipes: Unrecognizable
• Relief: recognizable but distorted (a lot of points but camera drift ruins it)

Possible Fazit for Poster/Poster session: Results would improve a lot with better sorting/similarity matching You can really see structures from the real room in most point clouds but you can also observe how those structures are reconstructed at multiple locations because of drifting cameras.

[lufixSch]

What exactly does lazy presorting do?

It creates a weight matrix and multiplies it with the dissimilarity matrix to increase similarity for close images (close With regard to the original order

[lufixSch]

Results for Delivery Area and Kicker:

Cuboid fitting is still not ideal. I will try to improve that tomorrow.

[lole-elol]

What exactly does lazy presorting do?

It creates a weight matrix and multiplies it with the dissimilarity matrix to increase similarity for close images (close With regard to the original order

And the unlayzy way would be then to not multiply with weights, right ?

[MariusGhica]

LGTM. It is amazing that many hyperparameters from the 3D reconstruction are the default values of the Matlab functions.

[lufixSch]

LGTM. It is amazing that many hyperparameters from the 3D reconstruction are the default values of the Matlab functions.

I habe to admit I took a look at the default parameters and compared them against our defaults.

I only had time to compare 2-3 values for each parameter. With some more Metrics and not just looking at the pointcloud would probably make the process much better

[lufixSch]

What exactly does lazy presorting do?

It creates a weight matrix and multiplies it with the dissimilarity matrix to increase similarity for close images (close With regard to the original order

And the unlayzy way would be then to not multiply with weights, right ?

Exactly

[lufixSch]

Improved Model Detection. Results especially for Delivery Area are now much better.

Ps: I had some problems with inconsistency regarding the point cloud alignment when using the pipeline function. Usually everything works but sometimes the rotation is way of.

[lufixSch]

I think this is ready to merge

[lufixSch]

I am going to merge this now.