layumi/University1652-Baseline

What it solves

This project provides a benchmark and baseline for drone-based geo-localization. It addresses the challenge of matching images from different perspectives—specifically drone-view, satellite-view, and street-view—to identify the location of a target building.

How it works

The system uses deep learning models (such as ResNet or VGG-16) to extract features from images. It supports two primary tasks:

1. Drone-view target localization: Matching a drone-view image or video to the most similar satellite-view image to locate a building.
2. Drone navigation: Using a satellite-view image to find relevant drone-view images from a flight history to navigate back to a target location.

Who it’s for

This is primarily for researchers and developers working on spatial intelligence, UAV (Unmanned Aerial Vehicle) navigation, and cross-view image retrieval.

Highlights

• Multi-source Dataset: Includes images of 1,652 buildings from 72 universities worldwide across drone, satellite, and street views.
• Training Features: Supports native PyTorch Float16 and BFloat16, linear warm-up, and random erasing.
• Evaluation Tools: Includes capabilities for re-ranking, multiple query evaluation, and visualization of ranking results.
• Extensibility: Provides a baseline for triplet loss and integrates with other datasets like CVUSA and CVACT.