Q1. Classification Model (40 points)

• the test accuracy of my best model is 95.4%

• visualize a few random test point clouds and mention the predicted classes for each. Also visualize at least 1 failure prediction for each class (chair, vase and lamp), and provide interpretation in a few sentences.

The left is Prediction class, and the right is GT class.

Chair:

Vase:

Lamp:

Failure Case:

The shape of the lamp is very similar to the vast

Q2. Segmentation Model (40 points)

• the test accuracy of my best model is 80.4%.

• visualize segmentation results of at least 5 objects (including 2 bad predictions) with corresponding ground truth, report the prediction accuracy for each object, and provide interpretation in a few sentences.

Auc: 94.4%

Auc: 88.3%

Auc: 74.9%

Auc: 24.3%

Auc: 27.3%

When the object is composed by multiple components, the segmentation accuracy is usually low.

Q3. Robustness Analysis (20 points)

1. I use pytorch 3D to rotate point clouds around x-axis for 20 degrees.

The accuracy of classification without and with rotation are 95.4% and 88.7%.

The accuracy of segmentation without and with rotation are 80.4% and 70.5%.

2. I tune the number of input point cloud.

The accuracy of classification for 1000, 5000, 10000 points are 95.2%, 95.3%, 95.4%

The accuracy of segmentation for 1000, 5000, 10000 points are 80.1%, 80.2%, 80.4%