Assignment 5. Point Cloud Processing

Q1. Classification

Test Accuracy: 86.67%

Random test examples

Point cloud of a chair
GT: Chair
Predicted: Chair
Point cloud of an office chair
GT: Chair
Predicted: Chair
Point cloud of an ornate vase
GT: Vase
Predicted: Lamp
Point cloud of a floor lamp
GT: Lamp
Predicted: Lamp

Failures

Chair point cloud with a tall, narrow back
GT: Chair
Predicted: Lamp

The tall back dominates the silhouette. The short seat is easy to miss, so the overall shape resembles a narrow floor lamp.

Vase point cloud with a narrow middle and flared top
GT: Vase
Predicted: Lamp

The vase has a wide base and a narrow middle. The flared top could resemble a small lampshade above a stem.

Lamp point cloud with a round base and wide shade
GT: Lamp
Predicted: Vase

The rounded base runs into the shade with little visible separation. The outline resembles a single bulbous vessel.

Q2. Segmentation

Test Accuracy: 84.44%

Bad predictions

Chair 260. 35.24%

Chair 260 ground truth
GT: Part labels
Chair 260 prediction
Predicted: Part labels

The backrest is recognizable, but the model breaks up the chair’s arm and side into seat and leg labels. It also treats the separate piece in front as if it were another seat with legs.

Chair 389. 35.40%

Chair 389 ground truth
GT: Part labels
Chair 389 prediction
Predicted: Part labels

The matching footrest is separate from the armchair, but the model labels it as if it were part of another chair. On the armchair itself, the backrest label spills into the seat.

Other predictions

Chair 415. 78.71%

Chair 415 ground truth
GT: Part labels
Chair 415 prediction
Predicted: Part labels

The cutout backrest and legs are easy to pick out in both views. The prediction gets fuzzy where the seat meets the legs, with the leg label extending onto the seat.

Chair 581. 88.48%

Chair 581 ground truth
GT: Part labels
Chair 581 prediction
Predicted: Part labels

The stool’s backrest, seat, and long legs are recognizable. The seat label runs down the tops of the legs where they join the seat.

Chair 30. 93.79%

Chair 30 ground truth
GT: Part labels
Chair 30 prediction
Predicted: Part labels

The prediction preserves the major parts. The main slip is where the armrest reaches the backrest and some of it picks up the backrest label.

Q3. Robustness

Rotation

I rotated each test cloud about its centroid around the Z axis and evaluated the original Q1 and Q2 checkpoints with the same labels.

AngleClassificationSegmentation
0° (clean)86.67%84.44%
15°83.42%81.96%
45°65.48%65.32%
90°46.07%46.68%
180°43.55%31.78%

The larger rotations create an out-of-distribution test: the shapes stay the same, but their XYZ coordinates no longer follow the orientations the models saw during training. PointNet’s max pooling handles point order, while its encoder can still learn orientation cues from XYZ. For example, a tilted chair can move leg points into positions the model associates with the back or seat. By 90°, both accuracies fell below 47%.

Three classification examples at zero, 45, and 90 degrees
Classification
Three segmentation examples at zero, 45, and 90 degrees
Segmentation

Gaussian noise

I added independent zero-mean Gaussian noise to each XYZ coordinate. The table gives the standard deviation in the input coordinate units.

Standard deviation (σ)ClassificationSegmentation
0 (clean)86.67%84.44%
0.00586.57%84.37%
0.01586.36%84.04%
0.02585.94%83.45%
0.0585.94%81.03%

Both models were close to their clean accuracy at σ = 0.005. At σ = 0.05, segmentation lost 3.41 percentage points and classification lost 0.73 points. Jitter makes part boundaries less distinct while leaving the overall object shape recognizable.

Three classification examples with no noise and Gaussian noise at standard deviations 0.025 and 0.05
Classification
Three segmentation examples with no noise and Gaussian noise at standard deviations 0.025 and 0.05
Segmentation

Q4. Locality

EdgeConv is a message-passing layer for point clouds. For each point, I built a fixed graph using its eight nearest neighbors in XYZ. A shared MLP turns each neighbor’s offset and the center point’s coordinates into a message, and a max over those messages gives the point a feature that depends on its neighborhood. I used one EdgeConv layer in place of PointNet’s point encoder. The classifier pools the features over the whole cloud, and the segmentation head combines each point’s feature with the pooled cloud feature before labeling it. Wang et al., Dynamic Graph CNN for Learning on Point Clouds (2019).

ModelClassificationSegmentation
Original Q1/Q2 (10,000 points)86.67%84.44%
Matched PointNet (1,024 points)93.70%83.76%
EdgeConv (1,024 points)93.28%85.33%

The matched runs shared the same point samples, training and validation split, and 10-epoch budget. The figures evaluate all three checkpoints on the same 1,024 points per cloud.

Three classification examples comparing the original Q1, matched PointNet, and EdgeConv
Classification
Three segmentation examples comparing the original Q2, matched PointNet, and EdgeConv
Segmentation