All test accuracies use the full 10,000 points per object unless
stated otherwise. Class labels: 0 = chair, 1 = vase, 2 =
lamp. Segmentation colours follow viz_seg: white,
magenta, cyan, yellow, blue and red for parts 0–5.
PointNet without the transformation blocks: a shared per-point MLP (3 → 64 → 128 → 1024), then max-pooling over points, then an FC head (1024 → 512 → 3).
Test accuracy: 97.27%
| Object 39 | Object 256 | Object 293 |
|---|---|---|
![]() |
![]() |
![]() |
| GT chair, pred chair | GT chair, pred chair | GT chair, pred chair |
| Object 485 | Object 604 | Object 806 |
|---|---|---|
![]() |
![]() |
![]() |
| GT chair, pred chair | GT chair, pred chair | GT lamp, pred lamp |
| Chair → predicted lamp (406) | Vase → predicted lamp (618) | Lamp → predicted vase (726) |
|---|---|---|
![]() |
![]() |
![]() |
Errors by true class: 1 chair, 16 vases and 9 lamps are misclassified.
Vases and lamps cause almost all the errors. Both are tall shapes with rotational symmetry, so a single global max-pooled feature cannot easily tell them apart. Telling them apart needs local cues, such as a thin stem or a hollow opening, that the max-pool can wash out.
This uses the same per-point encoder. The 64-d local feature of each point is concatenated with the 1024-d global feature, then a per-point MLP (1088 → 512 → 256 → 128 → 6) predicts each point’s part.
Test accuracy: 89.95% (fraction of points labelled correctly)
| Object | Ground truth | Prediction | Accuracy |
|---|---|---|---|
| 138 | ![]() |
![]() |
97.1% |
| 428 | ![]() |
![]() |
96.5% |
| 552 | ![]() |
![]() |
98.7% |
| 255 (bad) | ![]() |
![]() |
43.8% |
| 351 (bad) | ![]() |
![]() |
47.2% |
Objects 138, 428 and 552 are random picks from the better half of the test set. Objects 255 and 351 are the two worst objects in the test set.
Interpretation. On typical chairs with a clear back, seat and legs, the model is above 96%. Its errors sit mostly at the boundaries between parts, for example where the seat meets the legs. The two failures are both unusual, boxy shapes:
In both cases the model uses where a point sits on the chair as a stand-in for part identity. That works for typical chairs and fails when the layout is unusual.
Both experiments use the Q1/Q2 checkpoints without retraining. The 0° and 10,000-point rows are the Q1/Q2 baselines. The visualisations use test object 0, a chair, for both tasks. Classification predictions are given as labels (0 = chair, 2 = lamp).
Procedure. I rotate every test point cloud about the
x-axis. The y-axis is “up”, so this tips objects forward. Angles: 0°,
15°, 30°, 45°, 60°, 90° and 180°. Then I re-evaluate with
python eval_cls.py --rot_deg D and
python eval_seg.py --rot_deg D.
| Rotation | Cls accuracy | Seg accuracy |
|---|---|---|
| 0° (Q1/Q2) | 97.3% | 90.0% |
| 15° | 91.1% | 83.6% |
| 30° | 68.8% | 70.6% |
| 45° | 36.8% | 59.9% |
| 60° | 27.1% | 44.9% |
| 90° | 25.4% | 29.0% |
| 180° | 65.1% | 30.6% |
| Rotation | Classification (object 0) | Seg ground truth | Seg prediction |
|---|---|---|---|
| 0° | ![]() pred chair ✓ |
![]() |
![]() |
| 30° | ![]() pred lamp ✗ |
![]() |
![]() |
| 90° | ![]() pred lamp ✗ |
![]() |
![]() |
| 180° | ![]() pred chair ✓ |
![]() |
![]() |
Interpretation.
Procedure. I evaluate with --num_points
set to 10000, 5000, 1000, 500, 100, 50 and 10. Each run draws a fixed
random subset of points for all objects (seed 0).
| Points | Cls accuracy | Seg accuracy |
|---|---|---|
| 10000 (Q1/Q2) | 97.3% | 90.0% |
| 5000 | 97.2% | 89.9% |
| 1000 | 97.0% | 88.9% |
| 500 | 96.5% | 87.0% |
| 100 | 94.6% | 78.1% |
| 50 | 91.1% | 73.6% |
| 10 | 72.9% | 64.2% |
| Points | Classification (object 0) | Seg ground truth | Seg prediction |
|---|---|---|---|
| 10000 | ![]() pred chair ✓ |
![]() |
![]() |
| 500 | ![]() pred chair ✓ |
![]() |
![]() |
| 50 | ![]() pred chair ✓ |
![]() |
![]() |
| 10 | ![]() pred chair ✓ |
![]() |
![]() |
Interpretation.
Model. I implemented PointNet++ with single-scale
grouping in plain PyTorch (models.py), for both tasks.
python train.py --task {cls,seg} --model pointnet2.| Task | PointNet (Q1/Q2) | PointNet++ |
|---|---|---|
| Classification | 97.27% | 98.64% |
| Segmentation | 89.95% | 91.54% |
In classification, PointNet++ fixes 18 of PointNet’s 26 test errors and introduces 5 new ones. All three Q1 failure cases from above are now correct:
| Object | PointNet | PointNet++ |
|---|---|---|
![]() 406 (chair) |
lamp ✗ | chair ✓ |
![]() 618 (vase) |
lamp ✗ | vase ✓ |
![]() 726 (lamp) |
vase ✗ | lamp ✓ |
Segmentation, on the same objects as Q2:
| Object | Ground truth | PointNet | PointNet++ |
|---|---|---|---|
| 138 | ![]() |
![]() 97.1% |
![]() 96.9% |
| 552 | ![]() |
![]() 98.7% |
![]() 98.1% |
| 255 | ![]() |
![]() 43.8% |
![]() 67.1% |
| 351 | ![]() |
![]() 47.2% |
![]() 55.3% |
Interpretation.