Q1. PointNet classification
Test accuracy: 96.12% (916/953 objects). The classifier got 615 of 617 chairs, 82 of 102 vases and 219 of 234 lamps right. Confusion between vases and lamps accounts for 35 of the 37 errors.
The same MLP maps each XYZ point to 128 features. Max pooling over the points produces one vector per object, independent of point order. A small classification head then scores chair, vase and lamp. I trained for 250 epochs with 512 randomly sampled points per object and chose the checkpoint using 678 held-out training objects. For the test results here, the model saw all 10,000 points per object.
Random test objects

Where the predictions went wrong

Chair #406, predicted lamp. The rounded top stands out, while the seat and back blend together. Compared with the two correctly classified chairs above, the horizontal seat is harder to pick out.
Vase #629, predicted lamp. A narrow stem connects a wide base to a flared top. That outline resembles a table lamp more than the round-bodied vase in the random examples.
Lamp #883, predicted vase. The cloud forms a squat globe with a small foot. Without a visible stem or separate shade, the geometry alone gives little indication that this is a lamp.
Q2. Chair part segmentation
Test point accuracy: 89.21% (5,504,116 of 6,170,000 points across 617 chairs). Each point gets one of six labels. PointNet combines its local features and XYZ coordinates with a max-pooled feature vector for the whole chair.
I trained for 60 epochs with 768 randomly sampled points per chair. I selected the checkpoint on 448 held-out training chairs (89.52% validation accuracy), then evaluated all 10,000 points per test chair.
Ground truth and prediction
Each pair shows ground truth on the left and prediction on the right, using the same colors and camera path. The first three chairs are near the 75th, 50th and 25th percentiles of per-chair accuracy; the last two are the lowest-scoring chairs.










Q3. Robustness without retraining
1. Tilting the objects
I rotated every test cloud about the X axis through the origin. The model still saw all 10,000 points, with their labels unchanged. The 0° row reproduces Q1 and Q2.
| Tilt | Classification | Segmentation |
|---|---|---|
| 0° | 96.12% | 89.21% |
| 30° | 76.18% | 71.24% |
| 60° | 29.38% | 38.42% |
| 90° | 26.55% | 32.65% |
Even a 30° tilt costs roughly 20 percentage points in classification and 18 in segmentation. Max pooling makes PointNet insensitive to point order, not to rotation: when a chair is on its side, its seat and back no longer occupy the height ranges the model learned.


2. How few points are enough?
With the objects upright, I used one seeded permutation of point indices and took its first N entries for every cloud. Smaller samples are subsets of larger ones. Segmentation labels were sampled at the same indices; the score does not include discarded points.
| Points per object | Classification | Segmentation |
|---|---|---|
| 10,000 | 96.12% | 89.21% |
| 2,048 | 96.22% | 89.15% |
| 512 | 96.33% | 88.83% |
| 128 | 95.91% | 84.91% |
| 32 | 87.83% | 75.25% |
At 512 points, the classifier gets only two more objects right than at 10,000. That small difference on one draw is not evidence that fewer points are better. Segmentation starts losing accuracy sooner: 128 points often preserve the overall outline but leave little evidence around thin parts and boundaries. A single chair’s score can rise when subsampling removes difficult points.


Q4. Adding local neighborhoods
In the original PointNet, two nearby points do not exchange information before the whole-object max pool. I added a 6 × 6 × 6 grid over each cloud’s bounding cube. Within each occupied cell, an MLP encodes each point’s offset from the cell center. I average those 16-dimensional features and give the average back to each point in the cell, along with its XYZ coordinates. The remaining layers are the same as in Q1 and Q2.
Both local models trained for 250 epochs with the same splits, batch sizes and point sampling as their respective baselines. For both tasks the best validation checkpoint was epoch 191. Test inputs contain all 10,000 points per object.
| Test accuracy | PointNet (Q1/Q2) | With neighborhoods |
|---|---|---|
| Classification | 96.12% | 96.85% (923/953) |
| Segmentation | 89.21% | 89.74% (5,536,951/6,170,000) |
The classifier gets seven more objects right overall. It now separates vase #629 from lamp #793, both of which the baseline confused. For segmentation, the largest change among these examples is chair #272: more of its solid base is labeled blue instead of red, raising accuracy from 49.4% to 75.9%. Chair #161 is essentially unchanged. Local pooling helps on these examples, but it does not fix every boundary.

