CrossSectAI: Cross-Section Parameter Extraction using Artificial Intelligence

Bridges are vital components of transportation infrastructure, but many, particularly in countries like Germany, urgently require maintenance due to aging. Digital methods such as Building Information Modeling (BIM) and Digital Twinning support efficient maintenance planning. However, these methods rely heavily on accurate digital models, which are often unavailable for existing bridges. Consequently, digital reconstruction from construction drawings becomes necessary. Automating this reconstruction significantly accelerates the adoption of BIM and Digital Twinning for bridge maintenance.

This repository provides an implementation of the end-to-end pipeline introduced in our paper. The pipeline automatically extracts bridge cross-section geometries from existing construction drawings. The process uses YOLOv8 to detect cross-sections and the Segment Anything Model (SAM) to generate precise pixel-wise masks. These masks feed into a global optimization algorithm that matches parametric templates to the masks by minimizing a custom loss function, thus deriving accurate geometric parameters.

---

Usage

Setup

See below for quickstart installation and usage examples.

Install

The project uses pipenv to manage dependencies. To set up the environment, run:

pip install pipenv
pipenv install

The pre-trained YOLOv8 weights are already included in the directory of this repository. To use SAM, manually download the checkpoint from the official repository.

Ensure that the correct paths to the model weights are set in the configuration file (default.yaml):

Usage

You can use CrossSectAI directly from the Command Line Interface:

pipenv run python main.py -i <input_path> -o <output_dir> [OPTIONS]

The command supports various modes, accepting arguments:

• -i, --input (required) Path to a PNG image or a folder containing PNG images.

• -o, --output (required) Path to the output directory.

• -c, --config Path to the YAML configuration file. Defaults to default.yaml.

• --draw-results If set, draws and saves intermediate visualizations such as bounding boxes, masks, and polygons.

• --save-coco If set, saves detection and segmentation results in COCO format.

• --template-type Selects the cross-section template: 0 = Slab Girder, 1 = T-Girder, 2 = Tapered T-Girder (default).

Example

To run the included example, execute the following command from the project root:

pipenv run python main.py -i examples/img.png -o ./ --draw-results --save-coco

Allplan Bridge

To use the provided TCL scripts in Allplan Bridge:

1. Copy all required files into a single directory. This includes:

   • variables.tcl
   • The relevant cross-section script(s)
   • The main bridge script

2. In the main bridge script, set the base_dir variable to the path of this directory.

Allplan Bridge will sequentially load the referenced files and generate the bridge superstructure accordingly.

Note: After importing the scripts in Allplan Bridge, it may be necessary to trigger a recalculation to finalize the geometry.

License

This software is licensed under MIT License.