mfinstseg / README.md
Kriscamp's picture
Upload README.md with huggingface_hub
0b405d6 verified
|
Raw
History Blame Contribute Delete
7.28 kB
---
dataset_info:
features:
- name: query
dtype: string
- name: step
dtype: string
- name: annot
dtype: string
- name: inst
dtype: string
- name: bottom
dtype: string
- name: inst_partition
dtype: string
- name: split
dtype: string
- name: reasoning
dtype: string
- name: image
dtype: string
- name: cate
dtype: string
- name: task
dtype: string
- name: metadata
dtype: string
splits:
- name: train
num_bytes: 6216409233
num_examples: 62495
download_size: 1903755580
dataset_size: 6216409233
configs:
- config_name: default
data_files:
- split: train
path: data/train-*
extra_gated_prompt: Released for research use; access granted manually. State your
name, affiliation, and intended use.
license: cc0-1.0
pretty_name: MFInstSeg
tags:
- smart-manufacturing
- cad
- brep
- machining-features
- instance-segmentation
extra_gated_fields:
Name: text
Affiliation: text
Intended use: text
---
# MFInstSeg (加工特征实例分割 · 原始 B-rep)
MFInstSeg:**62495** 行, 每行一个 STEP(B-rep) 实体。唯一模型数 **62,477**;常被引用的
62,495 是上游划分文件的行数, 被 18 条重复撑出来的(train.txt 内重复 12、val.txt 内重复 1,
另有 5 个 id 同时出现在两个 split), 别再拿它当模型数。
任务是**三合一**(出处 AAGNet, Wu et al., RCIM 86:102661, 2024), 三套标签都在:
1. **语义分割** `annot`:每个面一个 0..24 的特征类别;
2. **实例分组** `inst` / `inst_partition`:哪些面属于同一个特征实例;
3. **底面识别** `bottom`:每个面是否为特征底面(0/1)。
> **按威威裁决上传原始 dataset 本身**("direct VLM 如果需要 render 的那种可以先传 dataset 本身,我来看看有没有办法弄一下 rendering")。
> 本库是**原始 B-rep**,尚未渲染成图像;渲染由团队后续统一处理。`image` 为 null。
## 面索引怎么对齐(最重要的一段)
`annot`/`bottom` 的键 k 和 `inst` 矩阵的行列号 k, 指的都是**用 pythonocc
`TopologyExplorer(shape).faces()`(`ignore_orientation` 保持默认 True)枚举出来的第 k 个面**,
等价于 occwl `Solid(shape).faces()` 的顺序。
- STEP 文件里所有 ADVANCED_FACE 的 name 都是**空串**(上游导出用了 save_face_label=False), 没有名字可对;
- **禁止**按 STEP 文件里 ADVANCED_FACE 的出现顺序对齐, 也别 set 去重、别按 hash 排序——都会把整套标签静默错位;
- 和我们已发布的 [`mfcad_D01_TB4`](https://huggingface.co/datasets/AI4Manufacturing/mfcad_D01_TB4) 不同:
MFCAD 是"名字索引"、MFInstSeg 是"位置索引", 两个库**不能共用**面对齐逻辑;
- 诚实声明: 这个约定来自审读上游源码链(get_segmentaion_label → TopologyExplorer; occwl entity_mapper),
跨 OCC 版本的保序性上游从未验证过; 抽样重算几何做自洽检查(标 through_hole 的面应是圆柱面之类)
还没做, 是遗留校验项。
## aag/ 为什么重要(虽然没入库)
源包另含 `aag/`(约 20GB 预算好的几何属性邻接图)。它**不是**"与 VLM 无关的杂物":
`aag/graphs.json` 的**节点顺序就是标签假定的面序**, 是唯一不依赖本地 OCC 版本就能安全还原
面序的参照物。本库因体积未收录; 需要严格核对面序时, 请回源包取 graphs.json, 把它的节点序
当对齐参照。(只抽 face_order 摘要入库还是仅在此指路, 团队待定。)
## 字段
- `query`(任务提示) · `step`(STEP 原文) · `annot`(面→类别 JSON, 键=面索引) ·
`inst`(N×N 0/1 实例矩阵 JSON) · `bottom`(面→0/1 底面 JSON) · `inst_partition`(见下) ·
`split`(train/val/test/unassigned) · `image`(null, 未渲染) · `reasoning`(null) ·
`cate`/`task` = `B`/`T-B4` · `metadata`(model_id, n_faces, n_instances, inst_matrix_rows)。
- `inst` 的约定(照搬上游): 特征面对角线 inst[i][i]=1; **不属于任何特征的 stock 面整行全 0**。
别把对角线清零、也别把全零行当缺失丢——stock 的语义就存在这。
- `inst_partition` 是我们为 SFT 加的派生列: 把 N×N 矩阵按连通分量还原成
`[{"faces": [...], "class_id": c}]`,
类别取成员面 seg 标签多数票。生成式模型吃划分远比吃矩阵合理, 上游特征级指标本来也吃这个。
- `cate/task` 沿用已发布的 mfcad_D01_TB4 / fcad_D03_TB4 的 `T-B4`(同一类任务, 非臆造)。
## 词表(25 类)
类别索引 0..24。已核实锚点: **0=chamfer、23=round、24=stock**。权威表在 AAGNet
`dataset/feature_creation.py` L42-47 的 feat_names; **注意该 repo 的 dataloader/mfinstseg.py 和
main.py 的注释把 chamfer/round 写反了**, 以 feature_creation.py 为准。本次转换未载入 feat_names.json, inst_partition 只有 class_id 数字。
## Split 与已知泄漏
`split` 列来自上游 `MFInstseg_partition/{train,val,test}.txt`(官方唯一 id 43,733/9,374/9,375,
约 70/15/15, 随机 i.i.d. 打散, 不是跨域泛化划分)。本次转换实测: 划分文件唯一 id
{"train": 43733, "val": 9374, "test": 9375}, 文件内重复行 {"train": 12, "val": 1}, 入库行 split 分布
{"test": 9373, "train": 43733, "val": 9371, "unassigned": 18}。
**已知 5 个 id 同时出现在两个 split(跨集泄漏)**: 20221225_205134_10 / _32 / _40 / _44 / _52
(train∩val=3, train∩test=1, val∩test=1)。本库按首次出现(train→val→test 顺序)归属;
本次转换实测检出的跨集 id: 20221225_205134_10, 20221225_205134_32, 20221225_205134_40, 20221225_205134_44, 20221225_205134_52。是否剔除这 5 个, 下游自己定。
## 数据质量(本次转换实测)
- 解析失败跳过 0 条; seg/inst/bottom 长度或键不齐跳过 0 条;
inst 不对称跳过 0 条。解析失败**不会**变成空行入库。
- 上游已知坑, 引用其指标时留意: 特征级 localization 判 TP 要求面集**完全相等**(IoU=1.0, 不是
COCO 的 0.5); 面级实例 F1 是在零填充 batch 上算的、没有 mask, 不同实现间不可直接比。
## License 与使用口径
- **CC0(公共领域)**是**百度 AI Studio 源站页面的标注**; AAGNet GitHub repo 自身的 license
我们未核到, 两边如不一致以哪个为准待确认。我们按 researcher 非商用使用, 并署名上游。
## Load
```python
from datasets import load_dataset
ds = load_dataset("AI4Manufacturing/mfinstseg")
```
_Gated — request access; granted manually._
## 复现
原始数据不入库。百度 AI Studio 需登录、无匿名直链: 手动下载 data2.zip(约 1.7GB, 解出
steps/ 与 labels/), 连同上游 MFInstseg_partition/ 一起放到 FORGE_DATASETS_SUB, 然后:
`FORGE_DATASETS_SUB=<dir> python mfinstseg/convert_mfinstseg.py`。
## Provenance
- 源: 百度 AI Studio `aistudio.baidu.com/datasetdetail/211864`(data2.zip 1.7GB; 另有 aag/ 约 20GB, 未收录, 见上)。
- 出处论文: AAGNet(Wu et al., RCIM 86:102661, 2024)。相关工作: BRepFormer / BRepMAE /
FeatureFox / MFInstSeg++(AAGATNet)等。请引用上游。