Related Experiment Video
Updated: Nov 25, 2025
04:48
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
3.2K
DASGIL: Domain Adaptation for Semantic and Geometric-Aware Image-Based Localization
Summary
This study introduces a new method for visual localization, fusing geometric and semantic data for robust place recognition in changing environments. The approach enhances autonomous systems
Area of Science:
- Computer Vision
- Robotics
- Machine Learning
Background:
- Long-term visual localization is critical for autonomous systems but challenged by environmental changes (season, illumination).
- Image retrieval offers an efficient solution for visual place recognition.
Purpose of the Study:
- To propose a novel multi-task architecture for fusing geometric and semantic information into multi-scale latent embeddings.
- To enable domain adaptation from synthetic to real-world datasets using adversarial training without human annotation.
Main Methods:
- A novel multi-task architecture integrating geometric and semantic features.
- Multi-scale latent embedding representation for visual place recognition.
- Adversarial training with a multi-scale feature discriminator for domain adaptation (synthetic-to-real KITTI).
Main Results:
- The proposed approach demonstrates superior performance on challenging datasets (Extended CMU-Seasons, Oxford RobotCar).
- Outperforms state-of-the-art baselines in retrieval-based localization and large-scale place recognition.
- Effective domain adaptation from synthetic to real-world data achieved.
Conclusions:
- The novel architecture effectively fuses multi-modal information for robust visual place recognition.
- Adversarial domain adaptation significantly improves performance in real-world, changing environments.
- The method advances long-term visual localization capabilities for autonomous driving and robotics.
