You need to agree to share your contact information to access this dataset
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
This is a mirror of the Autodesk Fusion 360 Gallery Dataset (Segmentation subset), redistributed with permission from the rights holder. The upstream licence is Autodesk's custom licence, NOT an open/CC licence: the data may be used ONLY FOR NON-COMMERCIAL RESEARCH PURPOSES, and you must pass that same restriction, together with attribution to the Fusion 360 Gallery Dataset, on to anyone you share the data or derivatives with. You must not attempt to re-associate the models with their creators or any identifiable individual. If you are employed by a for-profit entity, your employer is also bound by the licence. Full terms: https://github.com/AutodeskAILab/Fusion360GalleryDataset/blob/master/LICENSE.md . State your name, affiliation and intended use; access is granted manually.
Log in or Sign Up to review the conditions and access this dataset content.
Fusion 360 Gallery — Segmentation (逐面建模操作分割 · 原始 B-rep)
Autodesk Fusion 360 Gallery Dataset 的 Segmentation 子集,已转为团队统一 7 字段 SFT schema。 真人提交的 Fusion 360 设计,标签直接从参数化时间线自动导出:每张 B-rep 面标注 **"它是被哪个 CAD 建模操作造出来的"**。官方基线 BRepNet(CVPR 2021)。
本库是原始 B-rep,
image为 null;渲染由团队后续统一处理(与mfcad_original/mfinstseg同口径)。
⚠ 许可:仅限非商业研究(这条会传递给你)
上游是 Autodesk 自定义的非商用研究许可,不是 CC 也不是 MIT。本镜像经权利方许可 (permission obtained) 再分发。下载即表示你接受:
- 仅可用于非商业研究目的(LICENSE.md 第 1 条)。任何商业用途都不在授权范围内。
- 你若再向他人提供本数据(或其衍生),必须把"仅非商业研究"这条限制一并传递下去, 并要求对方继续传递,同时保留全部署名与 notice(第 3.2 / 3.3 条)。
- 衍生集不得明示或暗示自己就是原数据集,须显著标注为 "portion or modification"(第 3.1 条)。
- 不得尝试把数据集里的模型与其创作者或任何可识别个人重新关联(第 4 条)。
- 若研究中用到,请按学术惯例署名 "Fusion 360 Gallery Dataset" 并引用下方论文(第 5 条)。
- 受雇于营利性实体者注意第 11 条:你的使用会把你的雇主一并纳入本许可约束。
- Autodesk 保留随时终止许可的权利(第 9 条)。
许可原文:https://github.com/AutodeskAILab/Fusion360GalleryDataset/blob/master/LICENSE.md(Updated 11/2021;包内根目录另带同文 docx)。 上游仓库:https://github.com/AutodeskAILab/Fusion360GalleryDataset。官方基线代码 BRepNet 另有一份 license(CC BY-NC-SA 4.0,同样 NC)。
两个 config
| config | 源包 | 模型数 | 官方 train / test | 内容 |
|---|---|---|---|---|
main(默认) |
s2.0.1.zip(3,326,258,457 B) |
35,680 | 30,314 / 5,366 | 主集 |
extended |
s2.0.1_extended_step.zip(506,333,119 B) |
42,912 | 36,458 / 6,454 | 主集 + 额外 7,232 个(metadata.is_additional) |
extended 是 main 的超集(35,680 + 7,232 = 42,912),
多出来的那批是"做不出边数接近 2500 的三角网、因此没进 MeshCNN 基线"的复杂模型
(docs/segmentation.md:125 "help fatten the long tail of more complex models")。
只做 B-rep、不做跨表示对照时,用 extended。两个 config 各自从各自的 zip 构建,
不假设重叠部分逐字节相同(官方说是 drop-in replacement,但我们没做逐文件 CRC 比对)。
本次转换推送的 config:main。
⚠ **本次只推送了 config
main**,另一个 config 这次没构建,可能尚未存在于本仓库。
8 个类
ExtrudeSide / ExtrudeEnd / CutSide / CutEnd / Fillet / Chamfer / RevolveSide / RevolveEnd
即 {extrude, cut(=减材 extrude), revolve} × {side, end} + fillet + chamfer
(docs/segmentation.md:14)。类名读自包内 segment_names.json(143 bytes),脚本里没有硬编码。
注意 revolve/extrude 的 Start 与 End 在 timeline_info 里是分开的,落到 8 类时合并成一个 *End。
本次转换实测的面标签分布(config main,面级样本 526,069 个):
| idx | 类别 | 面数 |
|---|---|---|
| 0 | ExtrudeSide |
266,947 |
| 1 | ExtrudeEnd |
82,937 |
| 2 | CutSide |
67,554 |
| 3 | CutEnd |
12,315 |
| 4 | Fillet |
52,749 |
| 5 | Chamfer |
15,478 |
| 6 | RevolveSide |
27,654 |
| 7 | RevolveEnd |
435 |
论文口径的比例(paper §A.1 / §A.4 正文):ExtrudeSide 50.60%、ExtrudeEnd 15.72%、Fillet 10.22%、
Chamfer 2.94%、CutEnd 2.35%、**RevolveEnd 0.08%**。而 BRepNet 的 IoU 是 8 类等权 macro
(brepnet.py:766 mean_iou /= num_classes),0.08% 的类和 50.6% 的类权重一样 —— 少数类数字很不稳。
⚠ 坑 1:STEP 坐标是 ×10 的
docs/segmentation.md:112 原文:"Due to a unit conversion in the data exchange process, the STEP files
are scaled by a factor of 10. i.e. the longest side of the bounding box for the STEP files is [-10, 10]."
源包里的 smt / meshes / point_clouds 都已归一化到 [-1, 1],只有 STEP 是例外。
渲染前、或与别的子集拼数据前,先除 10。
⚠ 坑 2:面序契约,而且跟本组另两个 STEP 库都不一样
annot 的键、metadata.face_labels 的下标、timeline 里 faces 数组的下标,是同一个面序:
Fusion 360 API 枚举面的顺序(docs/segmentation.md:44),.seg 文件的行序就是它。
不要默认它等于 OpenCascade 读 STEP 后的面遍历序。 Autodesk 自己都不敢默认 ——
BRepNet/pipeline/face_index_validator.py:11-23 的三步校验是:先比面数 → 再比颜色暗示的面顺序 →
不一致就退而比三角形包围盒 → 差太多直接丢弃该文件("we reject the file as the labels are not
guaranteed to correspond to the correct face")。s1.0.0 → s2.0.0 掉的那 178 个就是这么掉的。
跨库对照(复用对齐代码必错):
| 库 | 面序约定 |
|---|---|
mfcad_original |
ADVANCED_FACE 的 name 整数 |
mfinstseg |
pythonocc TopologyExplorer 遍历序 |
| 本库 | Fusion 360 API 面序(需用颜色/几何校验确认 OCC 读出来是同一个序) |
坑 3:STEP 带颜色,但不泄题(已实测)
.stp 的 header 里 originating_system 是 ColorSegmentExtractor,面上确实挂着
STYLED_ITEM / COLOUR_RGB。这一度让人怀疑「面是按 segment 上色的 = 答案写在几何里」。
2026-09-23 实测否掉了这个怀疑。 方法:用 HTTP Range 从 S3 上的 zip 直接抠成员(未下载全包),
取 4 个面数 10、标签种类 3–5 的模型,解析 STYLED_ITEM -> ADVANCED_FACE 与
STYLED_ITEM -> ... -> COLOUR_RGB 建「面 → 颜色」,再与 .seg 的「面 → 标签」对照:
| 模型 | 面数 | 标签种类 | 颜色种类 |
|---|---|---|---|
26403_1f4a2618_10 |
10 | 3 | 10 |
22537_c6fcdc16_12 |
10 | 5 | 10 |
143157_747397db_3 |
10 | 3 | 10 |
132529_37236ada_1 |
10 | 3 | 10 |
4/4 都是一面一色,同标签的面颜色也各不相同 —— 颜色是面的身份标识,不是类别标识。
这与官方 pipeline/face_index_validator.py 拿颜色校验「面的顺序对不对」的用途完全吻合。
结论:STEP 可以原样使用,不需要剥色。 唯一要注意的是颜色能反推面序 —— 如果你的实验要求模型自己推断面的顺序,需要自行处理这一点。
⚠ 坑 4:官方划分是 per-body 切的,泄漏很重 —— 所以它是列不是 HF split
split 列取自包内 train_test.json(只有 train / test 两个 key,没有 validation)。
主集实测 30,314 / 5,366。
⚠⚠
docs/segmentation.md:122写的 30,459 / 5,399 是过期的 s1.0.0 数字 (两者之和 35,858 = s1.0.0 的规模,也是论文 §4/§A.1 的口径)。以包内文件为准。
test body 里 4,733/5,366 = 88.2%(本次转换从包内 train_test.json 实测) 其来源设计文档同时出现在 train 里
(extended:5,718/6,454 = 88.6%(本次未构建该 config,此数来自上游核验记录里未经复核的一条,仅供参考))。
35,680 个 body 只来自 8,832 个设计文档,平均一个设计 4 个 body,
而去重只删了"面数边数相同且面积体积差 1% 内"的(paper §A.1.1 明说
"Models with small topological differences were not considered to be duplicates")。
本库故意不把它做成 HF split,就是为了不让人 ds["test"] 一把梭然后报一个虚的泛化数字。
要测真泛化,请按 metadata.design_doc(= <designID>_<hash>)重切。
另:论文 70/15/15 里的 validation 不在 train_test.json 里,是 BRepNet 建 dataset.json 时用
固定种子 random_state=567 从 train 里一次性切好落盘的(pipeline/build_dataset_file.py:143,
默认 --test_split 0.15 / --validation_split 0.18),不是每次跑随机切。
⚠ 坑 5:形状被官方管线改过
凡不属于 extrude / revolve / fillet / chamfer 的 feature(孔、阵列、扫掠、放样、抽壳、镜像、布尔……) 全被 suppress。docs/segmentation.md:115 "we 'suppress' some CAD modeling features, slightly modifying the design from its original state";paper §A.1.1 "While this procedure modified the shapes of some models, it greatly increased the number of B-rep bodies available for the dataset." 换来的是 8 类闭集干净(没有 other/unknown),代价是形状分布不再是真实 CAD 件的分布—— 不能宣称这是"未经修改的真实工业几何"。
⚠ 坑 6:标签有不可约歧义
paper §5.4 自承:同一个圆柱,设计者可以用"挤出一个圆"做,也可以用"旋转一个矩形"做。
CutEnd 与 ExtrudeEnd 都是平面,几何上看不出差别,只能靠邻域凹凸判断。
准确率天花板不是 100%。
timeline 列:白送的实例级标注
timeline 是 timeline_info/<stem>.json 原文:每张面挂 feature(UUID)+ location_in_feature
(StartFace / EndFace / SideFace),另有 features 字典给每个 UUID 的 name / type / operation /
timeline_index(type ∈ ExtrudeFeature / ChamferFeature / FilletFeature / RevolveFeature;
operation ∈ NewBody / Join / Cut / IntersectFeatureOperation)。
8 类 .seg 是把它压扁的结果 —— 同类型的不同次操作在 .seg 里分不开,在这里分得开。
本库另把它派生成 metadata.feature_instances(按 feature UUID 分组的面索引 + 该操作的 type /
operation / timeline_index),做实例分割、操作序列预测、逆向工程直接可用。
官方基线 BRepNet 一行都没用它(整仓 .py/.md/.sh 零处出现 timeline_info),
所以这条线没有官方基线数字可比。
⚠ face_count 与 faces 数组嵌在 "body" 下面,不在顶层。
没有收进来的东西(以及为什么)
源包一个 stem 对应 10 个文件,本库只收 stp + breps/seg + timeline_info 三件套:
| 没收 | 理由 |
|---|---|
meshes/*.obj、point_clouds/*.xyz |
每个 10^5 字节量级,×35,680 各是 GB 级;渲染由团队统一处理,image=null |
meshes/*.{seg,fidx}、point_clouds/*.{seg,fidx} |
只有配上对应的 .obj/.xyz 才有意义(.fidx 是 triangle→face / point→face 的唯一桥梁),单收是死数据 |
breps/smt/*.smt |
Autodesk 专有格式,只有 Fusion 360 / Autodesk 产品能读,开源栈用不了 |
需要 mesh / 点云 / 无损几何请回源包取。 注意 STEP 转换是有损的:圆柱/球/锥会沿 seam line
被切开(拓扑变了)、部分环面退化成 rational NURBS
(BRepNet/docs/differences_in_open_cascade_pipeline.md:3,:9),.smt 才是未经转换的原始几何。
Schema
query(任务提示,含 8 个类名与面序说明) · step(STEP 原文,抽样 5/5 为 AP214、未全量扫,⚠ 坐标 ×10) ·
timeline(timeline_info JSON 原文) · split(train/test,⚠ 见坑 4) ·
image(null —— 需渲染) · annot({面序号: 类名} JSON) · reasoning(null) ·
cate=B / task=T-B4(与 mfcad_original / mfinstseg / deepcad 同族) ·
metadata(model_id、design_id、**design_doc**、body_index、split_official、n_faces、
face_labels(原始 0..7 下标,保真)、face_label_names、segment_names、classes_present、
feature_instances / n_feature_instances、is_additional(仅 extended),
以及 face_index_semantics / step_scale_note / step_color_note / split_note / shape_note /
taxonomy_note / timeline_note / license_note 八条随行说明)
⚠ step / timeline / split 是本库在 7 字段之外额外加的三列,理由见上。
这不是加工特征识别
标签语义是建模操作归属(设计意图视角),不是加工特征(制造工艺视角)。
与 mfcad_original /
mfcadpp / mfinstseg
标签体系不可互换。撞名的情况要分库看:
| 库 | 与本库 8 类撞名的 |
|---|---|
mfcad_original(16 类) |
只有 chamfer;它的类表里没有 fillet/round |
mfcadpp / mfinstseg(都是 25 类) |
Chamfer 和 Round —— 它们的 Round 就是本库的 Fillet(圆角) |
而且撞名处含义也不同:那边是"这张面属于一个倒角/圆角加工特征",
这边是"这张面是被 ChamferFeature / FilletFeature 这个 CAD 操作造出来的"。
形状来源也完全不同(那边是合成 stock 块上挖特征、多为平面;这边是真人设计,含圆柱/锥/球/环面/NURBS)。
官方基线数字(供参照,我们没复现)
paper Table 1(10 seed 均值 ± std,口径是 s1.0.0 的 35,858,不是本库的 35,680):
| 方法 | per-face acc | per-face IoU | 参数量 |
|---|---|---|---|
| BRepNet(winged edge) | 92.52 ± 0.15% | 77.10 ± 0.54% | 359,100 |
| ECC | 90.36% | 72.08% | 359,558 |
| PointNet++ | 74.00 ± 0.84% | 34.78 ± 2.23% | 1,403,784 |
| MeshCNN | 62.99 ± 0.37% | 20.59 ± 0.41% | 2,279,720 |
| ASM(规则法) | 64.57% | 49.53% | — |
⚠ 引用这些数字时必须带三个限定:(a) 它们是在 s1.0.0(35,858) 上、用官方那个泄漏的 per-body 划分
跑出来的;(b) "BRepNet 不吃坐标"只对论文配置成立 —— 仓库默认 --use_face_grids=1 /
--use_coedge_grids=1(brepnet.py:477,479)会走 UV-Net 的 xyz+法向网格,那就是坐标;
(c) 从 .stp 复现的数字与论文(基于 ASM 原生几何)不会完全一样,上游只说 "you will obtain very
similar numbers"(BRepNet/README.md:142),没给差多少。
Load
from datasets import load_dataset
ds = load_dataset("AI4Manufacturing/fusion360_segmentation") # config `main`,35,680 行
dse = load_dataset("AI4Manufacturing/fusion360_segmentation", "extended") # config `extended`,42,912 行
# 官方划分是**列**不是 split(理由见坑 4)
train = ds["train"].filter(lambda r: r["split"] == "train")
Gated — 需申请,人工审批。仅限非商业研究。
复现
原始数据不入库,只入 URL 与期望字节数:
| 文件 | URL | bytes |
|---|---|---|
s2.0.1.zip |
https://fusion-360-gallery-dataset.s3.us-west-2.amazonaws.com/segmentation/s2.0.1/s2.0.1.zip |
3,326,258,457 |
s2.0.1_extended_step.zip |
https://fusion-360-gallery-dataset.s3.us-west-2.amazonaws.com/segmentation/s2.0.1/s2.0.1_extended_step.zip |
506,333,119 |
set HF_TOKEN=...
set FORGE_DATASETS_SUB=<放 zip 的目录>
python fusion360_seg/convert_fusion360_seg.py --fetch
转换器逐成员从 zip 内部读、读完即弃,绝不整包解压(主包 3.3 GB 里的 .smt / .obj / .xyz
一个字节都不会被解压)。
Provenance
- 数据集:Fusion 360 Gallery Dataset — Segmentation 子集,https://github.com/AutodeskAILab/Fusion360GalleryDataset
(
docs/segmentation.md;主集取 s2.0.1,与 s2.0.0 内容相同但带 2021-11 更新版 license 文档)。 - 论文:Lambourne, J.G., Willis, K.D.D., Jayaraman, P.K., Sanghi, A., Meltzer, P., Shayani, H. BRepNet: A Topological Message Passing System for Solid Models. CVPR 2021, pp. 12773–12782. arXiv:2104.00706.
- 形状来源:Autodesk Online Gallery 上用户提交的 Fusion 360 设计(约 20,000 个设计中派生)。
- 许可:Autodesk 自定义非商用研究许可(见上),再分发经权利方许可。请署名 "Fusion 360 Gallery Dataset" 并引用上述论文。
- Downloads last month
- -