**Assignment 5: Point Cloud Classification and Segmentation**
Student name: Kangle Deng
(##) Classification Model
The test accuracy of my best model is 97.69%. I visualize predicted chairs as white, vases as purple, and lamps as blue.
Correct samples:
Failure samples:
All chairs are predicted correctly. Below I show one wrong example for vases and lamps. Some vases and lamps are very similar in shape, so it is relatively diffcult to tell the difference without the texture.
(##) Segmentation Model
The test accuracy of my best model is 89.99%. I visualize the prediction on the left, and the groundtruth on the right.
Good Predictions:
Bad Predictions:
Predictions can be inaccurate when the sample is out-of-distribution, for example, has accessories or has a abnormal shape.
(##) Robustness Analysis
(###) Rotate
I rotate the pointclouds about y-axis by 180 degrees. Namely, we'are turning around the objects so that the back of them is towards front.
I get 32.21% for the Classification model, compared to 97.69% in Q1.
I get 49.07% for the Segmentaiton model, compared to 89.99% in Q2.
This indicates our learned model can only work reasonabley in the canonical space of the objects.
(###) Number of pointclouds
I decrease the number of points from 10000 to 1000.
I get 97.27% for the Classification model, compared to 97.69% in Q1.
I get 89.17% for the Segmentaiton model, compared to 89.99% in Q2.
This indicates our learned model is robust to the number of the points.
Failure samples:
All chairs are predicted correctly. Below I show one wrong example for vases and lamps. Some vases and lamps are very similar in shape, so it is relatively diffcult to tell the difference without the texture.
(##) Segmentation Model
The test accuracy of my best model is 89.99%. I visualize the prediction on the left, and the groundtruth on the right.
Good Predictions:
Bad Predictions:
Predictions can be inaccurate when the sample is out-of-distribution, for example, has accessories or has a abnormal shape.
(##) Robustness Analysis
(###) Rotate
I rotate the pointclouds about y-axis by 180 degrees. Namely, we'are turning around the objects so that the back of them is towards front.
I get 32.21% for the Classification model, compared to 97.69% in Q1.
I get 49.07% for the Segmentaiton model, compared to 89.99% in Q2.
This indicates our learned model can only work reasonabley in the canonical space of the objects.
(###) Number of pointclouds
I decrease the number of points from 10000 to 1000.
I get 97.27% for the Classification model, compared to 97.69% in Q1.
I get 89.17% for the Segmentaiton model, compared to 89.99% in Q2.
This indicates our learned model is robust to the number of the points.