Assignment 5: Point Cloud Classification and Segmentation

PointNet classification and segmentation results, robustness experiments, and a Point Transformer bonus model.


Q1. Classification Model

Test accuracy

97.59% (930 / 953 test objects)

Random test point clouds and predicted classes

One failure prediction for each class

Chair failure: The chair has a tall, narrow structure, so its legs and back produce a silhouette resembling a lamp stand. PointNet's global max-pooled feature appears to emphasize this overall vertical shape more than the spatial relationship between the seat and back.

Vase failure: The vase's narrow neck and wider body resemble the stem-and-shade proportions of a lamp. This suggests that the classifier relies strongly on coarse proportions when the local surface structure is ambiguous.

Lamp failure: The lamp has a rounded body or base connected to a narrow section, which resembles a vase profile. Sparse or thin lamp components contribute fewer prominent features, allowing the smoother vase-like regions to dominate the global representation.


Q2. Segmentation Model

Test accuracy

89.89% (5,546,210 / 6,170,000 test points)

Five segmentation results, including two bad predictions

Object 351 — 48.01% accuracy (bad prediction)

Object 096 — 51.27% accuracy (bad prediction)

Object 135 — 74.24% accuracy

Object 520 — 86.06% accuracy

Object 013 — 90.21% accuracy

Interpretation: The model performs well when chair parts have clear spatial separation. The worst examples contain ambiguous boundaries and structures, causing large connected regions to receive the wrong part label.


Q3. Robustness Analysis

Experiment 1: Number of input points

Procedure: Evaluate both trained models with 10,000, 5,000, 1,000, and 100 sampled points per object. Smaller point sets are subsets of the larger sets.

PointsClassification accuracySegmentation accuracy
10,00097.59%89.89%
5,00097.38%89.91%
1,00096.96%89.02%
10094.44%79.79%
Accuracy versus number of points
Accuracy versus number of input points.

Classification examples: 10,000 versus 100 points

Segmentation examples: 10,000 versus 100 points

Interpretation: Both models remain stable through 1,000 points. At 100 points, classification falls by 3.15 percentage points and segmentation by 10.10 points. Dense segmentation depends more strongly on sufficient spatial coverage.

Experiment 2: Point-cloud rotation

Procedure: Rotate each input around the y-axis by 0°, 45°, 90°, 135°, and 180° without retraining. Segmentation labels remain attached to their points.

RotationClassification accuracySegmentation accuracy
0°97.59%89.89%
45°77.54%71.50%
90°77.33%51.34%
135°27.39%32.11%
180°32.84%39.75%
Accuracy versus rotation angle
Accuracy versus rotation around the y-axis.

Classification examples: 0° versus 90°

Segmentation examples: 0° versus 90°

Interpretation: Rotation causes a much larger drop than point removal. The models learn orientation-dependent coordinate patterns rather than rotation invariance. Segmentation is especially sensitive because rotation changes the coordinate patterns used to identify individual parts.


Q4. Bonus: Point Transformer

Implemented model

Our compact Point Transformer selects 1,024 anchors with farthest-point sampling and applies vector attention over each point's K = 16 nearest neighbors. For segmentation, anchor features are interpolated back to all 10,000 points.

Accuracy comparison

TaskPointNetPoint TransformerDifference
Classification97.59%96.96%−0.63
Segmentation89.89%91.13%+1.24

Segmentation comparison

The Point Transformer improves segmentation by 1.24 percentage points and labels 76,564 additional test points correctly. Classification is 0.63 points below PointNet.

Point Transformer robustness: number of input points

PointsClassification accuracySegmentation accuracy
10,00096.96%91.13%
5,00097.17%91.16%
1,00096.54%90.71%
10041.76%83.47%
Point Transformer accuracy versus number of input points
Point Transformer accuracy versus number of input points.
Point Transformer classification example with 100 points
Point Transformer classification example with 100 input points.

Interpretation: Transformer segmentation remains comparatively robust at 100 points, dropping 7.66 percentage points. Classification falls by 55.20 points. The classifier was trained using up to 1,024 attention anchors and does not generalize well to an extremely sparse 100-point input. In comparison, PointNet classification still reaches 94.44% with 100 points.

Point Transformer robustness: point-cloud rotation

RotationClassification accuracySegmentation accuracy
0°96.96%91.13%
45°80.27%82.77%
90°77.02%63.83%
135°30.01%55.79%
180°35.05%62.88%
Point Transformer accuracy versus rotation angle
Point Transformer accuracy versus rotation around the y-axis.
Point Transformer segmentation at 90-degree rotation
Point Transformer segmentation example at 90°.

Interpretation: Point Transformer classification has approximately the same rotation sensitivity as PointNet. Transformer segmentation is more rotation robust: at 90° it reaches 63.83% versus PointNet's 51.34%, and at 180° it reaches 62.88% versus 39.75%.