Research map / Image generation / Autoregressive and latent generation
Weatherproofing Retrieval for Localization with Generative AI and Geometric Consistency
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Augments visual localization retrieval training with InstructPix2Pix-generated weather/season/time-of-day variants of SfM-120k images, filtered by geometric correspondence checks, and trains HOW-style models with a hybrid real/synthetic contrastive loss.
Problem
Image retrieval for visual localization degrades under large appearance changes such as weather, season, and day-night, which hurts downstream 6-DoF pose accuracy.
Approach
Use text-guided image editing (InstructPix2Pix) with 11 nameable domain-shift prompts to create synthetic variants of training images, keep only variants whose geometric correspondences in areas of interest survive (via dense matching or SfM co-observations), and train with a hybrid loss that swaps query/positive/negatives with valid synthetic pairs.
Key results
- Training with synthetic P11 data brings consistent gains across datasets and protocols, with the largest gains in the EWB protocol, especially top-1.
- For Global-SfM, gains exceed 5% on RobotCar Night under domain shift.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add P11 synthetic training images | HOW-based retrieval model · Aachen, ECMU, and RobotCar; EWB localization | successfully localized queries | Helped |
| Add P11 synthetic training images | HOW-based retrieval model · RobotCar Night; Global-SfM localization | successfully localized queries: more than 5% | Helped |
Limitations
- Prompts contain only one main alteration each, so more complex conditions are not covered.
- Generation is performed as a preprocessing step rather than on the fly, which is costly.
Experimental setup
Datasets: SfM-120k RobotCar Seasons Extended CMU Aachen day-night
Benchmarks: Aachen day-night RobotCar Seasons v1 Extended CMU
Metrics: percentage of queries localized at low (5m,10°), mid (0.5m,5
Baselines: HOW HOW+ASMK FIRe AugMix
Builds on: HOW ASMK InstructPix2Pix AugMix
Closest papers by meaning
- WeatherPrompt: Multi-modality Representation Learning for All-Weather Drone Visual Geo-Localization (NeurIPS 2025 · similarity 0.87)
- Scaling Image Geo-Localization to Continent Level (NeurIPS 2025 · similarity 0.86)
- Semantic Ambiguity Modeling and Propagation for Fine-Grained Visual Cross View Geo-Localization (AAAI 2025 · similarity 0.86)
- RELOCATE: A Simple Training-Free Baseline for Visual Query Localization Using Region-Based Representations (CVPR 2025 · similarity 0.86)
- Recognition through Reasoning: Reinforcing Image Geo-localization with Large Vision-Language Models (NeurIPS 2025 · similarity 0.85)
- GeoRanker: Distance-Aware Ranking for Worldwide Image Geolocalization (NeurIPS 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.