---
title: Fusion 360 Gallery Segmentation
url: https://www.emergentmind.com/topics/fusion-360-gallery-segmentation-dataset
type: topic
---

# Fusion 360 Gallery Segmentation

The Fusion 360 Gallery Segmentation Dataset is a large-scale, publicly available benchmark comprising boundary representation (B-rep) models specifically annotated for face-level operation segmentation. Consisting of over 35,000 bodies sampled from the Autodesk Fusion 360 Online Gallery, this dataset provides high-fidelity CAD data structured for robust evaluation of machine learning and geometric reasoning techniques targeting Computer-Aided Design (CAD) model understanding. Each B-rep model is comprehensively labeled with the modeling operation that generated each face, providing a granular testbed for segmentation networks such as BRepNet and for advancing the reverse engineering of CAD workflows [2104.00706].

## 1. Data Scope, Structure, and Representations

The dataset contains 35,858 B-rep bodies, totaling approximately 500,000 faces (with individual models ranging from 1 to 421 faces and a median of 9). Each B-rep model is described at multiple levels:

- **B-rep (Boundary Representation):** Provided in Autodesk Fusion 360 (.f3d) and/or STEP format, encoding full parametric geometry (faces, edges, coedges) with explicit sparse adjacency and incidence matrices (`N`, `P`, `M`, `F`, `E`).
- **Triangle Mesh:** High-quality manifold meshes encoded in .obj files with ~3,000 triangles per body.
- **Point Cloud:** Uniformly sampled sets of 2,048 surface points per body (.ply or .npz).

In addition to raw geometry, Python-readable files (HDF5, pickle) are provided with matrices of per-face, per-edge, and per-coedge features (`X^{f} \in \mathbb{R}^{F \times p}`, `X^{e} \in \mathbb{R}^{E \times q}`, `X^{c} \in \mathbb{R}^{C \times r}`), adjacency matrices, and per-face label vectors $y \in \{1, \ldots, 8\}^F$. Meshes and point clouds inherit face-level labels through index mapping.

Each B-rep’s topology is structured as follows:

| Entity    | Symbol | Description                      |
|-----------|--------|----------------------------------|
| Face      | $F$    | Distinct B-rep patch             |
| Edge      | $E$    | Curve boundary shared by faces   |
| Coedge    | $C$    | Directed edge (face–edge pair)   |

Topology is encoded with five sparse matrices: ‘next’, ‘previous’, and ‘mate’ relations among coedges ($N, P, M \in \{0,1\}^{C \times C}$), and coedge-to-face/edge mappings ($F \in \{0,1\}^{C \times F}$, $E \in \{0,1\}^{C \times E}$).

## 2. Label Taxonomy and Annotation Schema

The fusion segmentation scheme designates eight exhaustive, mutually exclusive face-level classes reflecting high-level CAD modeling operations:

- **ExtrudeSide** (side walls of additive extrusion)
- **ExtrudeEnd** (cap of additive extrusion)
- **CutSide** (side walls of subtractive extrusion)
- **CutEnd** (cap of subtractive extrusion)
- **RevolveSide** (swept surface of non-360° revolve)
- **RevolveEnd** (end cap of non-360° revolve)
- **Fillet**
- **Chamfer**

Each face is labeled with *exactly one* of these classes. Additionally, for reverse engineering, operation IDs per face enable grouping faces originating from the same modeling action. Extrusions and revolves may be flagged as {Start,Side,End} regarding their generative stages (with `RevolveEnd` only for sweeps < $360^\circ$). All annotation data is stored in Python/HDF5 files with direct array mapping for ML integration. Mesh and point cloud representations inherit per-face labels via lookup tables between face, triangle, and point indices.

## 3. Train/Validation/Test Splitting Methodology

The dataset adheres to a 70%/15%/15% split for training, validation, and testing:

| Split      | Bodies (rounded) |
|------------|------------------|
| Training   | 25,080           |
| Validation | 5,379            |
| Test       | 5,399            |

Splitting is performed post-deduplication to avoid near-duplicate leakage (identical $F$, $E$ plus $<1\%$ difference in surface area and volume) and all splits disallow any duplicate or near-duplicate overlap.

## 4. Data Statistics

**Label distribution:**

| Class        | Frequency |
|--------------|-----------|
| ExtrudeSide  | 50.6%     |
| ExtrudeEnd   | 15.7%     |
| CutSide      | 11.9%     |
| CutEnd       | 2.35%     |
| Fillet       | 10.2%     |
| Chamfer      | 2.94%     |
| RevolveSide  | 7.1%      |
| RevolveEnd   | 0.08%     |

**Model complexity:** Median faces per body = 9; 25% have $\leq5$ faces; maximum is 421. Most bodies (50%) have at least one operation, 31% have at least two, only 1% have more than ten operations (maximum observed: 59).

Sample visualizations in the source work illustrate per-face class mosaics and distribution histograms [2104.00706].

## 5. Segmentation Task Definition and Evaluation Protocol

The formal segmentation task is: given a B-rep body with $F$ faces, predict the correct operation-based class $y_j \in \{1, ..., 8\}$ for each face $f_j$.

**Evaluation metrics are:**

- **Per-class intersection-over-union:** $IoU_c = \frac{TP_c}{TP_c + FP_c + FN_c}$
- **Mean IoU:** $mIoU = \frac{1}{8} \sum_{c=1}^8 IoU_c$
- **Accuracy:** $Acc = \frac{\sum_c TP_c}{\sum_c (TP_c + FP_c)}$
- **Precision, recall, F1:** Computed per class $c$ as: $Precision_c = \frac{TP_c}{TP_c+FP_c}$, $Recall_c = \frac{TP_c}{TP_c+FN_c}$, $F1_c = \frac{2 \cdot Precision_c \cdot Recall_c}{Precision_c + Recall_c}$

All deep-learning baselines are trained to minimize per-face cross-entropy:

$$
\mathcal{L} = -\sum_{j=1}^F \sum_{c=1}^8 1_{[y_j=c]} \cdot \log p_{j,c}
$$

**Benchmark results:** (mean ± std over 10 runs)

| Method            | Params    | Acc (%)         | mIoU (%)       | Other metrics                |
|-------------------|-----------|-----------------|---------------|------------------------------|
| BRepNet (winged)  | 359,100   | 92.52 ± 0.15    | 77.10 ± 0.54  | —                            |
| ECC graph net     | 359,558   | 90.36 ± 0.23    | 72.08 ± 0.50  | —                            |
| ASM (rule-based)  | —         | 64.57           | 49.53          | —                            |
| PointNet++        | 1,403,784 | 74.00 ± 0.84    | 34.78 ± 2.23  | Per-point Acc 82.78; IoU 36.49|
| MeshCNN           | 2,279,720 | 62.99 ± 0.37    | 20.59 ± 0.41  | Per-edge Acc 73.81; IoU 24.07 |

This quantifies the advantage of working on B-rep structures directly, as BRepNet surpasses both mesh- and point-based deep learning methods [2104.00706].

## 6. Access, Licensing, and Preprocessing

The dataset is released at https://github.com/AutodeskAILab/Fusion360GalleryDataset under the Autodesk Open Source License / CC-BY 4.0 (as per repository). For setup, users are instructed to:

- Clone or download the repository; ensure Python 3 and standard ML libraries are available.
- Load B-rep data using provided scripts to access all feature arrays, matrices, and label vectors.
- For non-B-rep (mesh, point cloud) research, load `.obj` or `.ply` files and use bundled mappings from triangle/point indices to B-rep faces.
- All geometries are pre-normalized so that the longest bounding box dimension is 2 units and centered at the origin; no manual scaling or centering is required.

## 7. Research Applications and Significance

The Fusion 360 Gallery Segmentation Dataset is a foundational resource for learning about CAD structure via B-reps, enabling the development and evaluation of neural architectures (e.g., BRepNet) that leverage topological message passing in solid modeling. Its comprehensive annotation of generative operations facilitates both semantic segmentation and reverse engineering of CAD models. The benchmark has established new standards for accuracy and robustness in operation-based segmentation tasks, providing a reproducible foundation for ongoing research in geometric deep learning and CAD model analysis [2104.00706].

Source: https://www.emergentmind.com/topics/fusion-360-gallery-segmentation-dataset