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 sisi321/ldtr_pytorch:version0

1. Run LDTR docker image

nvidia-docker run -t -i -v {local_dir}:{docker_dir} -p 8888:8888  sisi321/ldtr_pytorch:version00 /bin/bash

Training Data Generation

Generate positive data containing target lines

python generate_positive_training_data.py \
   --input_dir <path_to_input_dir> \
   --obj_name <line_object_name> \
   --output_dir <path_to_generated_training_data> \
   --start_index <starting_index_for_output_files> \
   --stride <stride_for_cropping_patches>

The above command runs the generate_positive_training_data.py script with the following parameters:

• --input_dir <path_to_input_dir>: Specifies the directory containing the input data.
• --obj_name <line_object_name>: Defines the name of the object to be processed.
• --output_dir <path_to_generated_training_data>: Indicates the directory where the generated training data will be saved.
• --start_index <starting_index_for_output_files>: Sets the starting index for the output files.
• --stride <stride_for_cropping_patches>: Determines the stride value for cropping patches from the input data.

Generate negative data containing target lines

python generate_negative_training_data.py \
   --input_dir <path_to_input_dir> \
   --obj_name <line_object_name> \
   --output_dir <path_to_generated_training_data> \
   --start_index <starting_index_for_output_files> \
   --stride <stride_for_cropping_patches>

The above command runs the generate_negative_training_data.py script with the following parameters:

• --input_dir <path_to_input_dir>: Specifies the directory containing the input data.
• --obj_name <line_object_name>: Defines the name of the object to be processed.
• --output_dir <path_to_generated_training_data>: Indicates the directory where the generated training data will be saved.
• --start_index <starting_index_for_output_files>: Sets the starting index for the output files.
• --stride <stride_for_cropping_patches>: Determines the stride value for cropping patches from the input data.

The expected input_dir should be

/path_to_input_dir/
├── tif_map_name_1.tif
├── tif_map_name_1_line_name.png
└── tif_map_name_1_line_name.p
├── tif_map_name_2.tif
├── tif_map_name_2_line_name.png
└── tif_map_name_2_line_name.p
└── ...

• <map_name>.tif is the map image.
• <map_name>_<line_name>.png is the raster ground truth.
• <map_name>_<line_name>.p is the vector ground truth.

** The output_dir structure is

/path_to_generated_training_data/
├── raw/
│   ├── sample_000001_data.png
│   ├── sample_000002_data.png
│   └── ...
└── seg/
|  ├── sample_000000_seg.png
|  ├── sample_000001_seg.png
|  └── ...
└── vtp/
|  ├── sample_000000_graph.vtp
|  ├── sample_000001_graph.vtp
|  └── ...

• raw/ contains map image patches.
• seg/ contains raster linear object ground truth patches.
• vtp/ contains vector linear object ground truth patches.

Here is a link to an example training dataset:

• Training Data for thrust fault lines: Download Here

Setup Instructions

1. Download and unzip the training data
2. Update the configuration file
   • Replace ./config/usgs_railroads.yaml with the one for thrust fault line in the Model weights table below
   • Locate the DATA_PATH field in the config file and set it to the unzipped training data directory.

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'

Model weights

Object Name	Config file	Model weights	Model card
Railroads	Download	Download	Download
Waterlines	Download	Download	Download
Scarp lines	Download	Download	Download
Thrust fault lines	Download	Download	Download

Use LDTR to detect desired linear objects

Generate Testing Data

Testing data is generated by cropping images from a map. The image file names should follow the format <row>_<col>.png, where <row> and <col> represent the row and column indices of the top-left corner of the cropped patches in the original map.

The expected input_dir structure for testing dataset is

/path_to_output_dir/
├── 0_0.png
├── 0_1.png
├── 1_0.png
└── 1_1.png
└── ...

• Each file represents a cropped patch from the input map.

Here is a link to an example testing dataset:

• Testing Data: Download Here
• The original input map and raster ground truth can be downloaded here. The file names are CO_Alamosa.png and CO_Alamosa_thrust_fault_line.png.

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 to the original map path.

Inference Using LDTR

python inference_on_map.py --cuda_visible_device 0 \
         --config <path_to_yaml_config_file> \
         --checkpoint <path_to_pretrained_LDTR> \
         --predict_raster \
         --predict_vector \
         --line_feature_name <target_line_name> \
         --prediction_dir <the_folder_to_save_the_detection> \
         --map_name <input_map_file_name>

The above command runs the inference_on_map.py script with the following parameters:

• --cuda_visible_device 0: Specifies the GPU device to be used.
• --config <path_to_yaml_config_file>: Path to the YAML configuration file.
• --checkpoint <path_to_pretrained_LDTR>: Path to the pretrained LDTR model checkpoint.
• --predict_raster: Flag to enable raster prediction.
• --predict_vector: Flag to enable vector prediction.
• --map_name <input_map_file_name>: Name of the input map file.
• --line_feature_name <target_line_name>: Name of the target line feature to detect. The detection output will be named as <map_name>_<line_feature_name>
• --prediction_dir <the_folder_to_save_the_detection>: Directory to save the detection results.

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

Inference Using LDTR on Map Patches

For debugging purposes, we provide a script to generate direct detection results at the patch level using LDTR.

python inference_on_patch.py --cuda_visible_device 0 \
         --config <path_to_yaml_config_file> \
         --checkpoint <path_to_pretrained_LDTR> \
         --output_dir <dir_to_save_detection_patches>

The above command runs the inference_on_patch.py script with the following parameters:

• --cuda_visible_device 0: Specifies the GPU device to be used.
• --config <path_to_yaml_config_file>: Path to the YAML configuration file.
• --checkpoint <path_to_pretrained_LDTR>: Path to the pretrained LDTR model checkpoint.
• --output_dir <dir_to_save_detection_patches>: Directory to save the detection results for each patch.

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