cadena-bench
收藏资源简介:
CADENA-Bench是一个用于CAD逆向工程基准测试的数据集,专注于真实机械零件。数据来源于MCB、TMCAD和CADNet三个公开数据集,经过去除平凡和重复几何后,根据零件功能划分为六个家族:轴与衬套(762个)、齿轮与轴承(749个)、壳体与框架(359个)、杠杆与型材(463个)、弹簧与紧固件(1046个)、工装与量具(17个),总计3396个零件。每个样本包含零件ID、家族标签、网格文件、网格二进制数据、八个视角(六个轴对齐和两个等轴测)拼接而成的图像,以及用于逐步推理的提示文本。该数据集用于评估CAD逆向工程方法,特别是针对旋转和阵列几何体的重建能力。评估指标采用GMS(几何匹配分数),要求预测点与目标点距离足够近且法线方向一致,无效预测不计入平均分。数据集已与强化学习训练集去重,确保无重叠。
CADENA-Bench is a benchmark dataset for CAD reverse engineering, focusing on real mechanical parts. The data is sourced from three public datasets: MCB, TMCAD, and CADNet. After removing trivial and duplicate geometries, the parts are categorized into six families based on their functions: Shafts and Bushings (762), Gears and Bearings (749), Housings and Frames (359), Levers and Profiles (463), Springs and Fasteners (1046), and Tools and Gauges (17), totaling 3,396 parts. Each sample includes a part ID, family label, mesh file, mesh binary data, an image composed of eight views (six axis-aligned and two isometric), and a prompt text for step-by-step reasoning. The dataset is used to evaluate CAD reverse engineering methods, particularly their ability to reconstruct rotational and array geometries. The evaluation metric is GMS (Geometry Matching Score), which requires predicted points to be sufficiently close to target points with consistent normal directions; invalid predictions are excluded from the average score. The dataset has been deduplicated with the reinforcement learning training set to ensure no overlap.
CADENA-Bench 数据集概述
基本信息
- 数据集名称:CADENA-Bench
- 任务类型:图像到文本(image-to-text)
- 许可协议:MIT
- 数据规模:1K < n < 10K,共 3396 个样本
- 数据集大小:约 13.05 GB(下载大小约 6.4 GB)
数据集简介
CADENA-Bench 是一个用于 CAD 逆向工程的基准测试数据集,专注于真实机械零件。数据来源于 MCB、TMCAD 和 CADNet,剔除了平凡和重复的几何形状,并按六个零件族进行分组,以便更清晰地暴露方法的弱点。数据集中所有零件均来自真实机械零件,合计 3396 个零件。
数据划分
- 训练集:3396 个样本,总字节数约 13.05 GB
零件族分布
| 零件族 | 数量 |
|---|---|
| shafts_and_bushings | 762 |
| gears_and_bearings | 749 |
| housings_and_frames | 359 |
| levers_and_profiles | 463 |
| springs_and_fasteners | 1046 |
| tooling_and_gauges | 17 |
| 总计 | 3396 |
注:
tooling_and_gauges样本量较小,其对应的评分不应作为排名参考。
数据字段
| 字段名 | 内容说明 |
|---|---|
part_id |
零件标识符 |
family |
所属零件族 |
mesh |
目标网格的二进制数据 |
mesh_file |
目标网格文件名 |
image |
八张目标视图拼接图(六个轴对齐视图 + 两个等距视图) |
prompt |
用于逐步推断的提示词 |
数据纯净性
原始评估集包含 3504 个零件,其中 108 个网格与强化学习训练语料库中的网格字节完全相同,这些重复网格已从该基准中移除。因此,本基准与 RL 训练集无交集。移除过程可通过哈希对比所有网格并剔除精确内容匹配来复现。
评估方法
论文采用 GMS(几何匹配分数) 对预测结果进行评分。GMS 要求预测点既要在距离上足够接近目标点,又要法线方向一致(满足 ‖p − q‖ ≤ τ 且 n_p · n_q ≥ cos α),从而反映表面是否被正确重建,而非仅关注体积是否填充。预测失败或非水密的模型结果视为无效,无效结果计入无效率并从均值统计中排除。
使用示例
python from datasets import load_dataset
ds = load_dataset("kulibinai/cadena-bench", split="train") gears = ds.filter(lambda r: r["family"] == "gears_and_bearings")
相关资源
- 论文:CADENA: Stepwise CAD Reverse Engineering(arXiv:2608.00799)
- 代码:https://github.com/zhemdi/cadena
- 模型:https://huggingface.co/kulibinai/cadena
引用格式
bibtex @article{cadena2026, title = {CADENA: Stepwise CAD Reverse Engineering}, author = {Kabisov, Soslan and Savrasov, Gennadiy and Elistratov, Maksim and Rodriguez, Antonio and Ignatiev, Daniil and Gavrilov, Nikita and Uzdenov, Rustam and Boyko, Alexey I. and Pasechnik, Igor and Konushin, Anton and Kuznetsov, Andrey and Zhemchuzhnikov, Dmitrii}, year = {2026}, eprint = {2608.00799}, archivePrefix = {arXiv}, primaryClass = {cs.CV}, url = {https://arxiv.org/abs/2608.00799} }




