File size: 3,240 Bytes
4a7b926 8b37795 4a7b926 8b37795 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 | ---
license: cc-by-4.0
tags:
- mesh-segmentation
- part-segmentation
- point-cloud
- vehicle
- onnx
- 3d
- qtmesheditor
library_name: onnx
pipeline_tag: other
---
# QtMeshEditor — Vehicle Part Segmentation
A point-cloud part-segmentation network (PointNet++-style) that labels each
point of a **vehicle** mesh (car / truck / plane / helicopter) as
`vehicle_body`, `wheel`, `window`, `wing`, or `rotor` (propeller), exported
to **ONNX** for local inference via ONNX Runtime.
One of the category-specialised segmentation models built for
**[QtMeshEditor](https://github.com/fernandotonon/QtMeshEditor)** (epic #818,
Track B2) — a free, open-source 3D mesh & animation editor. The app
auto-detects the mesh category with a companion
[point-cloud classifier](https://huggingface.co/fernandotonon/QtMeshEditor-mesh-segmentation-category)
and dispatches to this model for vehicles; siblings:
[body](https://huggingface.co/fernandotonon/QtMeshEditor-mesh-segmentation),
[vegetation](https://huggingface.co/fernandotonon/QtMeshEditor-mesh-segmentation-vegetation),
[building](https://huggingface.co/fernandotonon/QtMeshEditor-mesh-segmentation-building).
Aggregate download source used by the app:
[QtMeshEditor-models](https://huggingface.co/fernandotonon/QtMeshEditor-models)
(`segment/meshseg_vehicle.onnx`).
## Model
- **Input:** a sampled point cloud `float32 [1, N, 3]` (normalised to a centred
unit box; +Y up, vehicle nose facing +Z).
- **Output:** per-point class logits over 6 channels
(`unknown, vehicle_body, wheel, window, wing, rotor`); argmax → label,
scattered back to mesh vertices/faces by nearest sampled point.
- **Architecture:** shared per-point MLP + two kNN local-aggregation blocks
(in-graph `cdist`+`topk`, ONNX-exportable) + a global max-pooled feature;
~0.78 MB. Trained at the app's inference sample size (4096 points).
## Training data & license
Trained **from scratch, 100% on procedurally generated synthetic vehicles we
own** (no third-party data at all): parametric cars/trucks (body + cabin +
proud window panes + 4–6 wheels), planes (fuselage, main/tail wings, vertical
fin, optional nose prop + landing gear + canopy), and helicopters (body +
tail boom, main/tail rotors, skids, canopy) — labels are exact by
construction. Weights released under **CC-BY-4.0**; please credit
*QtMeshEditor*.
## Evaluation
- Held-out synthetic validation accuracy: **92.8%** (per-point, unknown
masked; v1.1's harder detached-part-augmented data — v1.0 scored 93.5% on
the easier all-attached set). Real-world CC0 vehicle packs are the planned
next data slice (mined via submesh/material-name labels — "Wheel_FL",
"glass", …).
## Reproducing
`scripts/export-meshseg-onnx.py --category vehicle` in the QtMeshEditor repo
(one-time, offline; the app never runs Python). Strategy + roadmap:
`docs/MESH_SEGMENTATION_STRATEGY.md`.
## Versions
- **v1.1.0** (current) — detached-part robustness: wheel/part clusters are
randomly offset during training (real exports often ship wheels as
separate nodes below the hull — the verified real-world failure case), so
detached wheels still label as `wheel`.
- **v1.0.0** — initial synthetic-only release (#818 Track B2).
|