Experiment setup
Models were trained for 250 epochs on a Tesla T4 using Adam with learning rate 0.001. Each object contains 10,000 XYZ points. Classification uses 6,784 training and 953 test objects across chair, vase, and lamp categories. Segmentation uses 4,489 training and 617 test chairs with six point labels. Batch sizes were 16 for classification and 8 for segmentation.
Following the starter training loop, the best checkpoint is selected using test accuracy each epoch. Results are from one run per model. Classification accuracy counts objects and segmentation accuracy counts points. All plots display Y vertically while inference uses the original coordinates.
Q1. Classification
A shared MLP maps each point through 3 → 64 → 128 → 256 channels with ReLU. Global max pooling yields an object feature, followed by a 256 → 128 → 3 classifier with dropout 0.3. Cross-entropy consumes raw logits.
Test accuracy 97.06% · 925 / 953 objects
Random test examples




One failure per class



Q2. Chair part segmentation
A shared 3 → 64 → 64 MLP produces point features. A 64 → 128 → 256 branch and global max pooling produce chair context. Concatenating this context with each 64-dimensional point feature feeds a 320 → 128 → 64 → 6 segmentation head. The model predicts six logits per point.
Test point accuracy 89.75%
The two lowest-scoring chairs and examples at the 25th, 50th, and 75th percentiles illustrate the range of predictions. Colors identify consistent label IDs across ground truth and prediction. Part descriptions below refer to visible geometry.





Q3. Robustness
Both experiments evaluate the original Q1 and Q2 checkpoints without retraining. The baseline uses all 10,000 points.
Experiment 1 — Rotation
Rotate every test cloud about the dataset Y axis by 30°, 60°, or 90°, keeping all points and their semantic labels. The same angle is applied to every object.
| Rotation | Classification | Segmentation |
|---|---|---|
| 0° | 97.06% | 89.75% |
| 30° | 86.04% | 71.78% |
| 60° | 58.55% | 58.13% |
| 90° | 63.59% | 52.55% |
Both models are sensitive to rotation. Classification falls by 38.51 percentage points at 60°, while segmentation falls by 37.20 points at 90°. Point-order invariance from max pooling does not imply rotation invariance, since the shared MLP consumes XYZ coordinates directly. Classification partially recovers at 90°, so its response is not monotonic across these angles.




Experiment 2 — Point subsampling
A single permutation of the 10,000 point indices is generated with NumPy default_rng(42). Its first 2,048, 1,024, and 256 indices form nested subsets shared across objects. Segmentation applies the same indices to ground-truth labels and scores only retained points. These measurements use one sampling seed.
| Points | Classification | Segmentation |
|---|---|---|
| 10,000 | 97.06% | 89.75% |
| 2,048 | 96.85% | 89.48% |
| 1,024 | 96.64% | 88.75% |
| 256 | 95.80% | 84.64% |
At 256 points, classification falls by 1.26 percentage points and segmentation by 5.11 points. Sparse samples often retain enough global shape to identify a category, while part labeling is more affected by the changed global feature and reduced geometric coverage. The displayed segmentation examples were selected for large drops and are more severe than the average.




Q4. Locality with voxel pooling
Each object is partitioned into a 12 × 12 × 12 cubic grid. The grid starts at the coordinate-wise minimum and its side length is the largest bounding-box extent. A shared 3 → 32 → 32 MLP encodes points. Cell-wise max pooling is gathered back to each point and concatenated with XYZ. The resulting 35-dimensional input feeds the original classification or segmentation network. This supplies local cell context alongside global PointNet pooling.
| Task | PointNet | With voxel pooling | Change |
|---|---|---|---|
| Classification | 97.06% | 97.69% | +0.63 pp |
| Segmentation | 89.75% | 90.01% | +0.26 pp |
Both models use all 10,000 points and the same training schedule as their baselines. The aggregate improvements are modest in this single-run comparison. Local pooling corrects some difficult examples and introduces some new errors. Changes in parameter count and training randomness also contribute to this comparison.





