Test Accuracy: 97.80%
Random test point cloud samples with predictions:
Ground Truth: Chair
Predicted: Chair
Ground Truth: Chair
Predicted: Chair
Ground Truth: Chair
Predicted: Chair
Ground Truth: Chair
Predicted: Chair
Ground Truth: Chair
Predicted: Chair
Here are some examples from the rotation experiment where the model gets confused:
Ground Truth: Chair
Predicted: Vase
Interpretation: At 45° rotation, the model misclassifies this chair as a vase. The vertical structure of the rotated chair probably resembles a vase to the network since it mostly saw upright chairs during training.
Ground Truth: Chair
Predicted: Vase
Interpretation: Rotating a chair 90° makes it look cylindrical to the model - hence the vase prediction. This happens frequently at this angle.
Ground Truth: Chair
Predicted: Lamp
Interpretation: Flipping a chair upside down leads to lamp predictions. The network clearly learned to recognize specific orientations rather than actual 3D geometric properties.
Overall Test Accuracy: 90.22%
Visualization of 5+ objects with corresponding ground truth:
Ground Truth
Prediction
Ground Truth
Prediction
Ground Truth
Prediction
Ground Truth
Prediction
Ground Truth
Prediction
Ground Truth
Prediction
Ground Truth
Prediction
Interpretation: The segmentation model does a good job identifying different chair parts (back, seat, legs, armrests). It works well on simple chairs but struggles with more complex or unusual designs. The trickiest parts seem to be the boundaries where different components meet, like where the legs connect to the seat.
I rotated all the test point clouds around the z-axis by different angles (0°, 45°, 90°, and 180°) and checked how well the model still classified them. This tests whether the network actually learned rotation-invariant features or just memorized what upright objects look like.
| Rotation Angle | Test Accuracy | Accuracy Change |
|---|---|---|
| 0° (Original) | 98.11% | - |
| 45° | 30.33% | -67.78% |
| 90° | 23.08% | -75.03% |
| 180° | 41.66% | -56.45% |
0° Rotation
GT: Chair, Pred: Chair ✓
45° Rotation (Rare Success)
GT: Chair, Pred: Chair ✓
45° Rotation (Typical Failure)
GT: Chair, Pred: Vase ✗
90° Rotation
GT: Chair, Pred: Vase ✗
180° Rotation
GT: Chair, Pred: Lamp ✗
Ground Truth
Prediction
Result: Good segmentation accuracy
Ground Truth
Prediction
Result: Segmentation quality degrades
Ground Truth
Prediction
Result: Significant accuracy loss
Ground Truth
Prediction
Result: Poor segmentation performance
Interpretation: The model exhibits severe lack of rotation invariance. Despite PointNet's theoretical rotation-invariance through max-pooling, the network clearly memorized object orientations from the training data. At 90° rotation, accuracy drops to near-random performance (23.08%). Rotation augmentation during training would be necessary for real-world deployment.
I tested the model with different point cloud densities by randomly sampling 1000, 2500, 5000, 7000, and 10000 points from each test object. This simulates what would happen with different sensors or sampling methods in practice.
| Number of Points | Test Accuracy | Accuracy Change |
|---|---|---|
| 10000 (Original) | 97.80% | - |
| 7000 | 98.11% | +0.31% |
| 5000 | 98.01% | +0.21% |
| 2500 | 97.90% | +0.10% |
| 1000 | 97.59% | -0.21% |
10000 Points
Accuracy: 97.80%
7000 Points
Accuracy: 98.11%
5000 Points
Accuracy: 98.01%
2500 Points
Accuracy: 97.90%
1000 Points
Accuracy: 97.59%
Ground Truth
Prediction
Points: Sparse but maintains structure
Ground Truth
Prediction
Points: Good balance
Ground Truth
Prediction
Points: Dense representation
Ground Truth
Prediction
Points: Very dense
Interpretation: The model demonstrates strong robustness to varying point cloud densities - accuracy remains stable between 97.59% and 98.11% across a 10x variation in point count. Interestingly, performance slightly improves with fewer points (7000 appears optimal), possibly due to reduced noise. The max-pooling operation effectively captures important geometric features regardless of input density, making it suitable for applications with varying sensor quality.