Assignment 5 — Point Cloud Classification & Segmentation

Q1. Classification Model (40 points)

Objective: Classify point clouds of chairs, vases, and lamps using a PointNet-based architecture.

Test Accuracy: ≈ 97.6%

Visualization of Random Test Samples

Failure Cases

Interpretation: Misclassifications mainly occur between lamps and vases due to their similar elongated structures in point space.

Q2. Segmentation Model (40 points)

Objective: Predict per-point segmentation labels for chair point clouds (6 semantic parts).

Test Accuracy: ≈ 90.2%

Segmentation Results on 5 Objects

The following GIFs visualize the ground truth and model predictions for five chair objects. Two of them represent bad predictions (accuracy < 0.6).

Interpretation: The model performs well on objects 0–2, accurately segmenting key parts such as the seat, legs, and backrest. However, on objects 26 and 61, the accuracy drops significantly — likely due to partial occlusion or irregular leg geometry, which causes confusion among spatially similar parts. Overall, the segmentation network captures strong local geometric cues but remains sensitive to point sparsity and unusual structural patterns.

Q3. Robustness Analysis (20 points)

Objective: Evaluate how model performance changes under rotations, varying point counts, and Gaussian noise. Both classification and segmentation models are tested for stability and invariance.

(A) Classification Robustness

1️⃣ Rotation Robustness — Classification

Interpretation: Model accuracy drops sharply beyond 30°, indicating limited rotational invariance. Around 180°, some symmetry causes partial recovery in accuracy.

2️⃣ Point Count Robustness — Classification

Interpretation: The model remains highly robust to downsampling; performance is stable above 2k points.

3️⃣ Noise Robustness — Classification

Interpretation: The classifier is robust to mild Gaussian noise, losing only ~0.5% accuracy at σ=0.02.

(B) Segmentation Robustness

1️⃣ Rotation Robustness — Segmentation

Interpretation: The segmentation model shows stronger degradation under rotation than the classifier, confirming limited orientation invariance.

2️⃣ Point Count Robustness — Segmentation

Interpretation: The segmentation model remains relatively stable above 2k points, showing high spatial consistency even under sparse sampling.

3️⃣ Noise Robustness — Segmentation

Interpretation: The segmentation network shows strong resilience to Gaussian perturbations — small noise has negligible effect, though accuracy drops gradually at σ ≥ 0.01.

Overall Summary:

• Both models demonstrate robustness to sampling and noise perturbations.
• Rotation remains the dominant weakness, suggesting future improvement via data augmentation or rotation-invariant point embeddings.