Understanding 2D computer-aided design (CAD) drawings plays a crucial role for creating 3D prototypes in architecture, engineering and construction (AEC) industries. The task of automated panoptic symbol spotting, i.e., to spot and parse both countable object instances (windows, doors, tables, etc.) and uncountable stuff (wall, railing, etc.) from CAD drawings, has recently drawn interests from the computer vision community. Unfortunately, the highly irregular ordering and orientations set major roadblocks for this task. Existing methods, based on convolutional neural networks (CNNs) and/or graph neural networks (GNNs), regress instance bounding boxes in the pixel domain and then convert the predictions into symbols. In this paper, we present a novel framework named CAD Transformer, that can painlessly modify existing vision transformer (ViT) backbones to tackle the above limitations for the panoptic symbol spotting task.
| Authors | Fan, Zhiwen and Chen, Tianlong and Wang, Peihao and Wang, Zhangyang |
|---|---|
| Year | 2022 |
| Venue | — |
| DOI | 10.1109/cvpr52688.2022.01071 |
| Source Database | EBSCO (IEEE Xplore Digital Library) |
| Bridge-to-GNN | Category Borderline |
| GNN Architecture | GNN (heuristic — verify) |
| Graph Encoding | 2D Drawing Graph (inferred from title) |
| AEC Task | Not yet extracted from PDF |
| Cohort | Peak Adoption (2022) |
| Key Finding | novel framework named CAD Transformer, that can painlessly modify existing vision transformer (ViT) backbones to tackle the above limitations for the panoptic symbol spotting task (heuristic — verify) |
|---|---|
| Quality Assessment | 1/6 · rigor: Low · Code/Data: ✓ (heuristic — verify) |
| GNN architectures detected | GNN (general) |
|---|---|
| Code repositories found in text | https://github.com/VITA-Group/ |
| Quality (heuristic, 6-flag) | 1/6 · rigor: Low Code/Data: ✓Benchmark: —Baseline: —CV/Split: —Ablation: —Reproducible: — |
Part of the GML/GNN in AEC Systematic Review (PRISMA 2020) — 112 papers, 2020–2026