[Progress-Summary] 15/09

[Lotemn102]

Done

(1) RMSE rate for using Kabsch algoritm without any improvments - 22.0343

• Tried to improve vicon projection points accuracy with Alon's methods:
  • (2) Remove annotated points where i am not 100% sure about their location - 16 points were chosen, RMSE was 21.9403.
  • Annotated less points on the rgb image - (2) had 16 points. I tried using different subsets of these points, best error rates were for 10 points - 20.3458. The subset were calculated brute forcely, i.e, for subset of 15 points, i've found all subsets of the 16 points of size 15 and calculated their Kabsch transformation. for subset of 14 points, i've found all subsets of the 16 points of size 14 and calculated their Kabsch transformation, and so on.
  • Run Kabsch algorithm, calculate for each point the error of the projection using the matrix Kabsch returned, remove points with high error rate, and re-run Kabsch - RMSE 22.1695
  • Use average depth value over several neighbor pixels (in the same frame) - RMSE 23.0545, sometimes even worse due to noise in the depth maps. The noise is due to the pixel alignment.

TODOs

• Try using different sessions (i've only tested on Sub007)
• Try taking different frames (i've only used 1-3 first frames for each Kabsch running)
• Try averaging the vicon points over several near frames