This week I mainly helped with the presentation and running a lot of tests to get good graphs for the slides. The slides can be found in the group slide deck, and I was mainly testing pixel skipping, laser threshold, and ICP number of point cloud inputs, and also did some testing on varying number of frames.
One key factor in development was time, as it took a lot of time to run any of these tests. From Saturday to Sunday when creating the slides I was basically rendering various number of frames such as 400, 700, and 1000, and also running the verification code on the side once one set of renders finished. Rendering is definitely the main bottle cap for the project in terms of how much time it takes to make changes, if the changes require re-rendering or rendering with new parameters.
I also fixed some bugs for ICP, with the main one being max iterations. Previously, our ICP was running max 30 iterations or until it converges, which I overlooked due to not printing the open3d inner debugging verbose output. However, from closer inspection, I realize that was causing a lot of the ICP outputs to not be nearly as perfectly aligned, and I set max iterations to 2000. For aligning point clouds, it usually takes between 40 to 150 iterations until convergence, and for our testing code to align meshes, it takes around 300 to 400 usually, so 2000 is a good upper cap which won’t restrict number of iterations a large majority of the time. Previously, ICP would have some random results and if we got lucky that the initial alignment was favorable then it would have better results, but now it actually converges on the local minima.
Next week I will be focusing on the demo videos and the final report.