Datasets:
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 dataset is redistributed with permission from the rights holder, under Autodesk's custom Fusion 360 Gallery Dataset License: it may be used ONLY FOR NON-COMMERCIAL RESEARCH PURPOSES. That restriction travels with the data and applies to you. If you are employed by a for-profit entity, Item 11 of the upstream license also binds your employer. This is a format-converted modification of the original dataset, not Autodesk's original release. 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 — Assembly-Joint (j1.0.0) · 原始 B-Rep 整包镜像
真实 Fusion 360 用户设计里的装配关节。给两个孤立的 B-Rep 零件,预测设计师会选哪一对
B-Rep 实体(每边一个 face 或 edge)来定义 joint axis。真值不是事后人工标注,而是从 .f3d 里
直接读出来的设计师选择(论文 A.1.4)。官方基线 JoinABLe(CVPR 2022)。
再分发经权利方许可(permission obtained)。上游许可不变:仅限非商用研究。 见下方 License 一节 —— 这条限制随数据一起传递给你,不是我们这层可以放宽的。
🔴 样本单位:19,156,不是 32,148
官方把话说死了(docs/assembly_joint.md:12 同一行的两句):
19,156 joint sets / 32,148 joints / 23,029 parts,以及
"We consider a joint set to be a single data sample."
| 数字 | 是什么 | 不是什么 |
|---|---|---|
| 19,156 | 样本数(joint set) | — |
| 32,148 | joint 总数 | ❌ 不是样本数(当样本数会高估 1.68 倍) |
| 23,029 | 去重零件数 | ❌ 不是样本数 |
README 的下载表印的是「32,148 joints / 23,029 parts」,很多登记表就是在这里错的。
本库的 default config 19,156 行 = joint set 数。转换脚本在两处对账、
任一处对不上就直接崩、不产出短数据集:开跑前数归档头里的 joint_set_*.json 个数,
收尾前数实际写出的行数与源文件里的 joint 总数(必须分别等于 19,156 和 32,148)。
约 70% 的 joint set 只含 1 个 joint、其余 30% 含多个(论文 A.1.4)——
一个样本可能有多个正确答案,本次实测:单 joint 13,381 / 多 joint 5,775,
joint 合计 32,148。
两个 config:为什么零件要单独一张表
23,029 个零件被 38,312 个 joint-set 槽位复用。把 body 文件按样本各复制一份等于 1.66 倍的浪费,
所以拆成两张表,用 body_id 连接:
| config | 行数 | 一行是什么 |
|---|---|---|
default |
19,156 | 一个 joint set(7 字段 SFT + 标注原文 + 装配态缩略图 + 两个 body 外键) |
parts |
23,029 | 一个去重零件(.smt / .step / .obj / graph JSON + 该零件的 holes 标注) |
from datasets import load_dataset
js = load_dataset("AI4Manufacturing/fusion360_joint") # joint sets(官方四个 split)
parts = load_dataset("AI4Manufacturing/fusion360_joint", "parts") # 去重零件载荷,按 body_id 连接
Gated — request access; granted manually.
官方划分(本次实测,不是按比例反推)
官方只说「approximate 70/10/10/10%」(docs/assembly_joint.md:234),精确条数上游从未公布。
本次转换实测:
| split | 行数 |
|---|---|
test |
1,926 |
train |
13,405 |
uniform_test |
1,909 |
validation |
1,916 |
划分文件 train_test.json 的原始键与条目数:{"train": 13405, "validation": 1916, "test": 1926, "uniform_test": 1909}。
validation / test 是人为清洗过的(剔掉了几何相似但没被标注的 sibling 实体),
mix_test 保留原始分布 —— 官方自承它不能可靠判分(论文 A.2.6)。
跨集泄漏(同一个 joint set 出现在多个 split 里)本次实测 0 个,按首现认领、未剔除。
⚠ 零件跨 split 复用:split 划在 joint set 上,而零件是共享的。本次实测
4,967 个零件出现在不止一个 split 的 joint set 里
(每行的 metadata.splits_referencing 列了具体是哪些)。拿 parts 做预训练时这就是泄漏面。
🔴 索引口径(四个 STEP 库各不相同,别跨库复用对齐代码)
每个零件的图节点顺序是先全部 face、再全部 edge(docs/assembly_joint.md:26);
JoinABLe 把 edge 的图节点号算成 n_faces + edge_index(offset_joint_index, joint_graph_dataset.py:860-870)。
本库 annot 里的 index 是 JSON 原文里按类型分开的原始索引,不是图节点号。
要和 graph 对齐,请自己用 parts config 的 metadata.n_faces 加偏移。
对照:mfcad_original 按 ADVANCED_FACE 的 name 整数对齐、mfcadpp 按 STEP 文本出现序、
mfinstseg 按 TopologyExplorer 遍历序。四套各用各的。
🔴 UV-grid:face 10×10×7,edge 10×6
graph 列里带 UV-Net 风格的 UV-grid。face 是 10×10×7(3 point + 3 normal + 1 trimming mask,
joint_graph_dataset.py:83-90 / 93-97);edge 是 1D 的 10×6 = 60 维(3 point + 3 tangent,
:106-109 用的是 grid_size*3 而不是 grid_len*3),edge 没有 trimming mask。
把 10×10×7 套到 edge 上会高估特征体量。(官方默认训练配置其实不喂 UV-grid,也不喂任何 XYZ 坐标:
face 9 维 / edge 10 维,只有 entity type + length + reversed 位。)
🔴 标签是 PU 的,评测口径必须带上
- 99.9% 的标签为负,未标注的正样本约为已标注的 3 倍(论文 A.2.4)。标成 0 的实体对里 相当一部分其实能装,只是设计师当时没用。不能当干净的二分类真值。
- 测试时正标签是
"Joint,JointEquivalent"—— 共轴等价实体也算命中 (args_train.py:73-80→joint_graph_dataset.py:296把 scheme 内标签压成 1 →metrics.py:263-264判max_label==1); **训练时只用"Joint"**。 - 所以官方 top-1 79.53%(clean test)是等价口径下的数;换到保留原分布的
mix_test上 Ours 掉到 62.22、人类从 80.00 掉到 69.00。人类在 No Hole 子集上更是掉到 58.33% / 7.69%(论文 A.2.5 正文;Table 1/6 里印的是 "-")。引用任何数字都要带口径。
⚠ 缩略图会泄题
thumbnail_png 是 joint_set_XXXXX.png 的原字节 —— 那是装配态渲染(第一个 joint 装好之后),
等于把答案画出来了。所以统一口径的 image 一律为 null;要拿缩略图当输入,先确认你不是在泄题。
(渲染由团队后续统一处理。)
收了什么、没收什么
- 只有 rigid joint,且只有每边单选一个实体的 joint;多实体选择的那 9% 不在本子集 (留在 assembly 子集,论文 A.1.4)。
- JSON 里躺着 7 种 joint type(Rigid/Revolute/Slider/Cylindrical/PinSlot/Planar/Ball)和完整的 motion limits,
但官方基线一个都不预测(
get_label_matrix从不读joint_motion)。想做 joint type 分类得自己定义任务与指标。 contacts(装配态下 0.1mm 容差内贴合的面对)与holes都在joint_set原文里,保真收录。holes是机器跑的:Autodesk Shape Manager 特征识别工具自动标注,不是人工(docs/assembly_joint.md:184)。 本库把它按 body 归集去重后放进partsconfig 的annot,当真值用之前请抽检。 本次实测:不同 joint set 给出不一致孔集的零件 4,623 个(保留首见并计数); 没拿到 holes 条目的零件 10,433 个。- 单位:长度 cm,角度弧度(
docs/assembly_joint.md:45-46)。
Schema
default:query(任务提示,含索引口径) · image(null,见上) · annot(每个 joint 的
实体对 + equivalents + joint_type/offset/angle/is_flipped,JSON) · reasoning(null) ·
cate=B / task=T-B4 · metadata · **joint_set**(joint_set_XXXXX.json 原文,保真镜像) ·
**thumbnail_png**(PNG 原字节) · body_one / body_two(连 parts) · split。
**parts**:同样的 7 字段(annot = 该零件的 holes 标注;没有条目时为 null) ·
**body_id**(主键) · **smt**(Autodesk Shape Manager 原生 B-Rep,官方称之为 ground truth geometry) ·
**step**(中性格式备份) · obj(三角网格,g face N / g halfedge M edge K 分组,不保证 manifold) ·
**graph**(NetworkX node-link 邻接图 + UV-grid 特征)。
本次 graph 节点判据统计:{"by:surface_type": 23029}
(metadata.n_faces 为 null 表示这个零件的 graph 节点写法我们没认出来,没有猜)。
cate/task = B/T-B4,与 mfcad_original / mfinstseg / deepcad 同族。
📜 License —— 非商用研究,限制随数据传递给你
上游是 Autodesk 自定义许可,不是 CC / MIT:
- 数据只能用于非商用研究(LICENSE.md 第 6 条:
only for non-commercial research purposes)。 - 本次再分发经权利方许可(permission obtained);许可条款本身不变,我们按第 3 条的要求 把非商用限制连同第 7/8 条的免责声明一并传递给下游,并保留对上游的署名与原仓库链接。
- 本库是原始发布物的格式转换版(修改集),不是 Autodesk 的原始数据集; 要原件请从上游 Fusion360GalleryDataset 取。
- 第 4 条禁止把模型反向关联到创作者;第 9 条 Autodesk 保留随时终止授权的权利。
- ⚠ 第 11 条:若你受雇于营利实体,该许可同时约束你的雇主,且你声明自己有权代表雇主接受。 适用法为加州法(第 12 条)。
- 代码另算:JoinABLe 仓库的 LICENSE 是 MIT(Copyright (c) 2022 Autodesk, Inc)—— 基线代码可商用、数据不可商用,两者别混谈。
- 全文:https://github.com/AutodeskAILab/Fusion360GalleryDataset/blob/master/LICENSE.md(Updated 11/2021)
Provenance
- 数据:
https://fusion-360-gallery-dataset.s3.us-west-2.amazonaws.com/assembly/j1.0.0/j1.0.0.7z—— 2,975,602,672 字节,Last-Modified Tue, 30 Nov 2021 05:23:00 GMT (字节数与日期由 HTTP HEAD 独立复核两次)。解包后合计 35,423,840,413 字节(35.42 GB, 由归档头统计,未整包落地)。 另有预处理包j1.0.0_preprocessed.7z—— 538,969,936 字节,Last-Modified Thu, 21 Apr 2022 19:31:57 GMT, 那是 JoinABLe 自用的.pickle派生物,不是原始数据,本库未收录。 ⚠ 两个包日期不同,不要说「数据冻结在 2021-11-30」。 - 文档:
docs/assembly_joint.md(本子集唯一权威文档)、README.md、LICENSE.md。 - 基线:AutodeskAILab/JoinABLe(默认分支是
main,不是 master)。 - 论文(对外引用用 CVPR 版):Willis, Jayaraman, Chu, Tian, Li, Grandi, Sanghi, Tran, Lambourne, Solar-Lezama, Matusik. JoinABLe: Learning Bottom-up Assembly of Parametric CAD Joints. CVPR 2022. arXiv:2111.12772.
- 上游还有三个同源子集(Assembly / Reconstruction / Segmentation),都派生自同一批约 20,000 个
.f3d设计, 零件级形状很可能互相重叠 —— 与本组已有的 Fusion 360 Gallery 子集混用前请先做零件级比对(尚未核)。
- Downloads last month
- -