OskarK1/MMLandmarks_baseline

CVPR 2026 - MMLandmarks: a Cross-View Instance-Level Benchmark for Geo-Spatial Understanding

🤖 Pretrained model:

We provide a pretrained model (125849_GSTC_textfirst_losscomplete_bs512_lr0.0001_ep20) trained with the following setup:

• modalities: GSTC
• outdoor_only: false
• text_sampling: first
• last_sat_only: false
• loss: complete

Results:

	medR	mAP@1k	R@1	R@5	R@10
G2S	57	24.88	19.40	43.77	56.45
S2G	29	16.86	27.30	48.40	57.70
T2G	77	16.31	28.10	44.60	52.40
T2S	712	17.6	13.80	31.70	40.80

	Dist @ 1km	Dist @ 25km	Dist @ 200km	Dist @ 750km	Dist @ 2500km
G2C	16.6	36.0	50.77	74.6	92.3
S2C	26.4	50.3	71.3	91.9	98.9
T2C(no_cues)	6.1	17.8	28.7	51.9	87.7

Description

Explore the dataset [here].

MMLandmarks is a multi-modal, instance-level and continental scale dataset of landmarks collect from the United States of America. We address the shortcomings of datasets and specialized models used for geospatial tasks, such as for geolocalization, retrieval and cross-view localization, and instead propose a dataset that enables training and benchmarking models on all these tasks. In fact, for each landmark we collect multiple ground view images, temporal satellite images, as well as the landmarks' GPS coordinates and textual information. MMLandmarks provides data for $18{,}557$ unique landmarks, where each landmark contains all four modalities. With this dataset, we hope to motivate further multimodal research for geospatial tasks, and multimodal learning in general.

Github repo:

Visit the Github repository to get access to the official training/evaluation code.

Dataset

Visit the MMLandmarks codebase to download the full dataset. The page also extensively covers the contents of the dataset, as well as its structure.