apb-spatial-computer-vision 1.0.0

Python package to integrate image pyramids and raster/vector algorithms for computer vision at Port of Barcelona. Allows text, bounding box and point-based segmentation based on SAMGeo and GroundingDINO. Please refer to the homepage https://victorpachecoaznar.github.io/TFG/

TFG

This repository hosts the code for my Bachelors' thesis in Geoinfomation and Geomatics Engineering. The project has been tutored by Port de Barcelona and Universitat Politècnica de Catalunya.

Please refer to the homepage of the project:

🏡 https://victorpachecoaznar.github.io/TFG/

About this package

The goal of APB_spatial_computer_vision is to provide a framework to perform computer vision tasks on geospatial imagery. Its focus is semantic segmentation, mainly though SAM. The logic can be accessed in several points:

• Through text, to perform image to bbox tasks via Grounding DINO.
• Through bounding box, to perform SAM and achieve segmentation.
• Throgh a lower resolution vector dataset coming from another model, in an attempt to improve segmentation.

Tools

• 🦆 The program leverages Spatial-SQL DuckDB queries for precise vector operations
• ⚡ Image pyramids have been optimized via concurrency to GDAL-warping
• 📷 Integration of virtual layers through GDAL to SAMGeo
• 📄 MKDocs for automated docs+ jupyter visualization

Instructions

The entry points of the system are the following environment variables:

• TEXT_PROMPT: a string with the prompt to look for via Grounding DINO
• VECTOR_DATASET: a vector dataset
• DATA_DIR: location of the data
• BASE_DIR: where this package is located, acessible for deployments
• NAME_ORTOFOTO: basename for the image to be processed