Assignment 5

Yu Han

Andrew ID: yuhan2

1

test accuracy of my best model: 0.979

successful predictions for chairs:

Failed predictions for chairs, predict as lamp:

successful predictions for vases:

Failed predictions for vases, predict as lamp:

successful predictions for lamps:

Failed predictions for lamps, predict as vase:

The geometry for these class diverses a lot. If the chair is folded like first one, the geometry is very different from the normal chairs. Also I can hardly tell which class it belongs to for these failure case.

2

best accuracy: 0.901

Good examples:

Prediction acc 0.961 &gt

Prediction acc 0.984 &gt

Prediction acc 0.885 &gt

bad examples (accuracy<0.6):

Prediction acc 0.5494 &gt

Prediction acc 0.5566 &gt

If the boundary for each semantic part is clear and sharp, then the prediction is good. However, if the boundary is unclear and soft, the prediction is bad.

3

Rotation:

For classification task:

Original, test accuracy: 0.979

Rotate pi/6 around z-axis, test accuracy: 0.770

Rotate pi/6=3 around z-axis, test accuracy: 0.377

The accuracy drops a lot. The model is not robust to rotation.

For segmentation task:

Original, test accuracy: 0.901

Rotate pi/6 around z-axis, test accuracy: 0.770

Rotate pi/6=3 around z-axis, test accuracy: 0.377

The accuracy drops a lot. The model is not robust to rotation.

Number of points:

For classification task:

Original 10000 points, test accuracy: 0.979

Decrease the number of points to 5000, test accuracy: 0.979

Decrease the number of points to 1250, test accuracy: 0.978

Decrease the number of points to 100, test accuracy: 0.922

The accuracy drops a little. The model is very robust to the number of points.

For segmentation task:

Original 10000 points, test accuracy: 0.901

Decrease the number of points to 5000, test accuracy: 0.901

Decrease the number of points to 1250, test accuracy: 0.895

Decrease the number of points to 100, test accuracy: 0.822

The accuracy drops a little. The model is very robust to the number of points.