Sufficiency of low-resolution images for accurate visual localization

Determine whether downsampled images at approximately 560×560 resolution preserve sufficient information to enable accurate camera pose estimation in image-based visual localization using dense matching (e.g., RoMa), and rigorously characterize the contribution of high-frequency image details to localization performance across diverse datasets and conditions, with possible implications for other perception tasks.

Background

The paper proposes ImLoc, an image-based visual localization pipeline that stores posed RGB images and precomputed depth maps, leveraging dense matching for 2D–3D lifting. To reduce storage and computation, the authors use relatively low-resolution images (560×560) and observe empirically strong performance and efficiency.

Within the Dense Image Matching discussion, the authors explicitly state a conjecture that low-resolution images may retain most of the information needed for localization (and possibly other perception tasks), suggesting that high-frequency details could be less important. Validating this conjecture would impact design choices for map compression, storage budgets, and runtime efficiency across datasets and conditions.