16-889 Assignment 5

Luyuan Wang (luyuanw@andrew.cmu.edu)

1. Classification

Test Accuracy: 97.48%

Visualization:

PointCloud
Prediction	Chair	Vase	Lamp

Failure case:

The GT label is 1 (Vase), but the predicted label is 2 (lamp). It makes some sense, as the flower somehow looks simliar to a lampshade, and the stem looks like the neck of a lamp.

The network is designed following by PointNet (without Transform blocks). The shared MLPs are implemented using Conv1D, as CNNs work as scanning MLPs. There are BatchNorm layers and ReLU layers between convolutional layers. I didn't use dropout layers though.

2. Segmentation

Test Accuracy: 87.95%

Visualization:

GT	Predicted	Acc
		93.06%
		93.06%
		98.43

Failure Cases:

GT	Predicted	Acc
		51.98%
		52.16%

The results are pretty impressive, it's hard to tell any differences for the success cases. For the failure cases, the first one classify part of the 'back' (light blue) as 'base' (dark blue), and part of the seat (red) as 'back' (light blue). While the error of the second case is more obvious, part of the 'base' (red and purple) are classified as 'back' (light blue). It's understandable, as these parts are close and pretty similar.

The network structure is almost the same with the classification network. The global feature vector are copied and concatenated to each row of the previous tensor.

3. Robustness Analysis

Adding rotation to point clouds

A 20 degree rotation is added to x-axis, y-axis, and z-axis, repectively. The rotation matrix is defined as follow:

A visualization of the rotated point cloud:

Classification Test Accuracy:

An example of the rotated point cloud for the segmentation task:

GT	Predicted

The network is not very robust to rotation. By adding Transform blocks (T-Net) may be helpful.

Changing number of points

The network is pretty robust to the number of points per object.