Assignment 5

Jianchun Chen jianchuc@andrew.cmu.edu

---

Q1. Classification Model (40 points)

Best testing accuracy: 97.90

The left hand side images are samples that is mistakenly classified and right hand side are correctly classified.

The chair is classified as a lamp.

The vase is classified as a lamp.

The lamp is classified as a vase.

Basically the failure cases are because the input shape are unordinary, which looks similar to objects from other category.

Q2. Segmentation Model (40 points)

Best testing accuracy: 89.64

Left hand side images are for predictions and right hand side images are for ground truth. The first two are hard example and the failure is from misunderstanding of major component. Other objects has a high part-seg accuracy. Only points in the boundary are mistakenly segmented.

Accuracy: 0.4796

Accuracy: 0.4607

Accuracy: 0.9368

Accuracy: 0.9245

Accuracy: 0.9871

Q3. Robustness Analysis (20 points)

number of sample points:

num_points	100	1000	10000
cls	93.8	97.8	97.9
seg	81.7	88.6	89.6

Random rotation:

Rotated?	No	Yes
cls	97.9	30.33
seg	89.6	27.52

PointNet is robust to random point sampling, but not robust to rotation (new research such as VectorNeuron can learn rotation invariant feature).