TriView2CAD

# TriView2CAD Dataset

**TriView2CAD** is the first large-scale, open-source benchmark specifically designed for orthographic projection reasoning in Computer-Aided Design (CAD). It supports tasks ranging from engineering drawing interpretation to 3D geometry reconstruction, and serves as a foundation for training and evaluating multi-modal models in industrial scenarios.

## 📦 Dataset Overview

TriView2CAD consists of:
- **200,000 synthetic samples** for training and in-domain evaluation;

Each sample is annotated with **precise dimension labels** and includes **six interoperable data modalities**:
1. **Parameter Tables** (JSON)
2. **Vector CAD Drawings** (DXF)
3. **Raster Images** (PNG; front, top, side views)
4. **Executable CAD Scripts**
5. **STEP Files** (3D models)
6. **B-Rep Files** (3D models)

This rich format enables seamless support for tasks across design, simulation, and manufacturing workflows.

## ⚙️ Dataset Construction

We focus on **prefabricated bridge piers**, a modular structure commonly found in civil engineering. A constraint-guided pipeline ensures the geometric and structural validity of each sample:

### Step 1: Constraint-Guided Parameter Sampling
- A 15-dimensional parameter space is defined.
- **Intra-view constraints** ensure physical validity and topology closure.
- **Inter-view constraints** guarantee consistency across orthographic views (height, width, depth).

### Step 2: 2D Drawing & Raster Image Generation
- The `ezdxf` library converts sampled parameters into vectorized DXF drawings.
- Geometric primitives (lines, circles, arcs) are organized into semantic layers.
- High-resolution raster images (front, top, side) are rendered in FreeCAD.

### Step 3: 3D Model Generation
- Using FreeCAD's scripting interface, 3D models are programmatically reconstructed.
- Both **STEP** and **B-Rep** files are generated to support CAD simulation and interoperability.

## 🧠 Supported Reasoning Tasks

TriView2CAD enables evaluation of complex reasoning capabilities via three key task categories:

1. **Dimension Recognition and Mapping**
Identify each annotated dimension and match it to its geometric feature.

2. **Primitive Counting**
Count the number of specified components (e.g., pier columns, pile bases).

3. **Composite Parameter Computation**
Derive engineering-critical values based on multiple parameters (e.g., spacing = width + gap).

A total of **15 parameters** are evaluated per sample, including:
- 6 recognition attributes
- 3 counting attributes
- 6 computed attributes

Evaluation accuracy is computed per parameter and averaged across the test set.

## 📊 Benchmarking & Generalization

TriView2CAD has been used to benchmark seven leading vision–language models. Our experiments demonstrate:
- Significant performance gaps between pretrained VLMs and reasoning-tuned models.
- Real-world samples introduce realistic noise and complexity (e.g., overlapping annotations, occlusions).
- Chain-of-Thought (CoT) and curriculum-based fine-tuning dramatically improve reasoning accuracy and OOD robustness.

# TriView2CAD 数据集

**TriView2CAD** 是首个专为计算机辅助设计（Computer-Aided Design，CAD）领域的正投影推理任务打造的大规模开源基准数据集。它支持从工程图纸解读到三维几何重建的各类任务，可作为工业场景下多模态模型训练与评估的基础平台。

## 📦 数据集概览

TriView2CAD 包含以下内容：
- **20万条合成样本**，用于训练与域内评估；

每条样本均标注有**精准的尺寸标签**，并包含六种可互操作的数据模态：
1. **参数表**（JSON格式）
2. **矢量CAD图纸**（DXF格式）
3. **光栅图像**（PNG格式，包含主视、俯视、侧视三视图）
4. **可执行CAD脚本**
5. **STEP文件**（三维模型）
6. **边界表示（Boundary Representation，B-Rep）文件**（三维模型）

该丰富的数据格式可无缝支持设计、仿真与制造全流程中的各类任务。

## ⚙️ 数据集构建流程

本次构建聚焦**预制桥墩**——这是土木工程中常见的模块化结构。我们采用约束引导的构建流程，确保每条样本的几何与结构合法性：

### 步骤1：约束引导的参数采样
- 定义了15维参数空间。
- **视图内约束**确保物理合理性与拓扑闭合性。
- **视图间约束**保证正投影三视图（高度、宽度、深度）的一致性。

### 步骤2：二维图纸与光栅图像生成
- 借助`ezdxf`库将采样得到的参数转换为矢量化DXF图纸。
- 将几何图元（直线、圆、圆弧）按语义分层组织。
- 在FreeCAD中渲染得到高分辨率的主视、俯视、侧视光栅图像。

### 步骤3：三维模型生成
- 通过FreeCAD的脚本接口，以编程方式重建三维模型。
- 生成**STEP**与**B-Rep**两类文件，以支持CAD仿真与跨平台互操作。

## 🧠 支持的推理任务

TriView2CAD 可通过三大核心任务类别，实现对复杂推理能力的评估：

1. **尺寸识别与映射**
识别每条标注尺寸，并将其与对应几何特征进行匹配。

2. **图元计数**
统计指定构件的数量（例如桥墩柱、桩基础）。

3. **复合参数计算**
基于多组参数推导工程关键数值（例如间距=宽度+间隙）。

每条样本共评估**15项参数**，涵盖：
- 6项识别属性
- 3项计数属性
- 6项计算属性

评估准确率按单参数计算，并在测试集上取平均。

## 📊 基准测试与泛化能力

目前TriView2CAD已用于7款主流视觉语言模型（Vision-Language Models，VLMs）的基准测试。我们的实验结果表明：
- 预训练视觉语言模型与经过推理微调的模型之间存在显著性能差距。
- 真实世界样本引入了符合实际的噪声与复杂度（例如标注重叠、遮挡问题）。
- 思维链（Chain-of-Thought，CoT）与基于课程的微调方法可显著提升推理准确率与分布外（Out-of-Distribution，OOD）鲁棒性。

搜集汇总

背景与挑战

背景概述

TriView2CAD是首个专为CAD正交投影推理设计的大规模开源基准，包含20万个合成样本，每个样本提供六种互操作数据模态（如参数表、矢量图纸和3D模型），专注于预制桥墩结构。它支持维度识别、原始计数和复合参数计算等任务，用于训练和评估工业场景中的多模态模型，并已基准测试多个领先模型以提升推理准确性。