vastai algorithm: 1) git clone GCP 2) install datasets 3) change config for dataset paths (for dataset and model separately). Batch size too 4) change train/launch_train scripts as needed

GPU Configuration Total Batch Size Estimated Training Time 2x RTX 4090 6 ~1.8 days (44 hours) 2x RTX 5090 10 ~1 day (25 hours)

Building Instance Segmentation Project

Project Overview

Automatic detection and generation of building outlines on images.

Author

KazGisa - Sagi Abdashim

Technical Specification

Task

Develop a model for automatic recognition and drawing of building and structure outlines based on Google Maps overlays (or Google satellite images), or on aerial photography (AFS).

Requirements

Functionality

• Automatic detection and generation of building outlines on images.
• If a new building outline appears on the overlay/AFS, the system should automatically add it to the database and render it.
• Rescan the area and update data upon detecting changes (optional).

Accuracy

• The geometric accuracy and shape of the outlines must match the example provided in the reference image.
• Deviation in coordinates should not exceed X meters/pixels (to be specified based on the example).

Data Source

• Satellite imagery / Google Maps overlays (or Google Satellite)
• Available aerial photographs with 5cm accuracy.

Output Data

• Vector outlines of buildings (formats: GeoJSON, Shapefile).
• Accompanying attribute information:
  • Date of appearance
  • Building status (residential, non-residential, etc.) # Optional
  • Area
  • Coordinates

Features

• [x] Instance segmentation support via MMDetection
• [x] Command-line interface for training, evaluation, and visualization
• [x] Integration with Weights & Biases for experiment tracking
• [ ] TorchScript and ONNX export (planned)
• [ ] Post-processing (planned)

Getting Started

Instructions for setup and usage will be provided in a future update.

Model Architecture and Approach

Initially, we used P2PFormer (paper), but later adopted the GCP model (paper) with a ResNet50 backbone for improved performance. Training GCP involves two stages (GitHub repo): 1. Train Mask2Former for mask extraction 2. Train GCP for polygionization of masks by adding a polygonizer head, freezing all other layers.

The multi-class instance segmentation approach was abandoned due to the complexity of distinguishing residential from non-residential buildings.

Alternative Experiments

• Multi-class instance segmentation (abandoned)
• Training only on AFS (aerial) images (in progress)
• Training only on Google Satellite images (planned)
• Combining both datasets (planned)