LDTR: Linear Object Detection Transformer

Accurate Graph Generation by Learning N-hop Connectivity Information

Goal

LDTR aims to detect linear objects in topographic or geological maps.

---

Docker Image for Training & Testing LDTR

Run LDTR Docker Image

1. Pull LDTR Docker Image from Docker Hub

nvidia-docker pull weiweiduan/ldtr_pytorch:versions

1. Run LDTR docker image

nvidia-docker run -t -i -v {local_dir}:{docker_dir} -p 8888:8888 weiweiduan/ldtr_pytorch:version0 /bin/bash

Training Data Generation

python generate_noprune_data.py

Train LDTR

python train.py --config configs/usgs_railroads.yaml --cuda_visible_device 0

To update the parameters for data, model architecture, and training process, please update './configs/usgs_railroads.yaml'

Use LDTR to detect desired linear objects

python test_darpa_map_conflation_shapely.py --cuda_visible_device 0 --config ./configs/usgs_railroads.yaml --checkpoint ./trained_weights/runsrailroads_1965_g256_10/models/checkpoint_epoch=200.pt

Please update './configs/usgs_railroads.yaml' for the path to the testing images

Training and Testing Data

Training Data

Here is a link to an example training dataset:

• Training Data: Download Here

Setup Instructions

1. Download and unzip the training data
2. Update the configuration file (usgs_railroads.yaml):
   • Locate the DATA_PATH field.
   • Set it to the unzipped training data directory.

---

Testing Data

Here is a link to an example testing dataset:

• Testing Data: Download Here

Contents of the Testing Dataset

The zip file contains:

1. OR_Camas_g256_s100.zip → Direct input data for LDTR inference.
2. OR_Camas.tif → The original map used for testing.
3. OR_Camas_fault_line.tif → Ground truth raster image representing the fault line (linear object).

Setup Instructions

To test LDTR using this dataset:

1. Unzip the testing data
2. Update the configuration file (usgs_railroads.yaml):
   • Set TEST_DATA_PATH to the unzipped testing data directory.
   • Set TEST_MAP_PATH and TEST_TIF_PATH to the original map path.

Note: The image dimension in both training and testing examples is 256*256 pixels