GeoText-1652 Leaderboard

The primary ranking metric is Mean, computed as the average of Text R@1/R@5/R@10 and Image R@1/R@5/R@10 means. All scores are percentages.

Use the official evaluation command from the README and report the metrics printed by Method/re_bbox.py.

Full Test (951 buildings)

Official GeoText-1652 test_951_version.json evaluation. This is the primary leaderboard.

Rank	Method	Submitter	Training Data	Backbone	Text R@1	Text R@5	Text R@10	Image R@1	Image R@5	Image R@10	Mean	Links	Verified
1	NGCG-MLLM (InternVL3.5-1B)	Chen et al.	GeoText-1652 train	InternVL3.5-1B + LoRA	25.8	41.0	49.4	34.4	63.3	75.1	48.2	paper code	pending
2	SAA-DGL (X2-VLM)	Yuan et al.	GeoText-1652 train	X2-VLM	16.7	29.2	36.2	29.4	57.2	69.4	39.7	paper	pending
3	SAA-DGL (X-VLM)	Yuan et al.	GeoText-1652 train	X-VLM	14.9	28.4	35.8	28.2	56.5	69.0	38.8	paper	pending
4	HCCM	Ruan et al.	GeoText-1652 train	X-VLM	14.7	26.0	32.5	28.8	57.3	69.9	38.2	paper code	pending
5	GeoText-1652 official checkpoint	GeoText-1652 Team	GeoText-1652 train	Swin-B + BERT	13.6	24.6	31.2	26.3	53.7	66.9	36.1	paper code ckpt	yes

24GB Test

GPU-friendly test_24G_version.json evaluation for quick comparison on smaller hardware.

Rank	Method	Submitter	Training Data	Backbone	Text R@1	Text R@5	Text R@10	Image R@1	Image R@5	Image R@10	Mean	Links	Verified
1	NGCG-MLLM (InternVL3.5-1B)	Chen et al.	GeoText-1652 train	InternVL3.5-1B + LoRA	45.7	62.9	70.3	58.3	85.9	92.9	69.3	paper code	pending
2	SAA-DGL (X2-VLM)	Yuan et al.	GeoText-1652 train	X2-VLM	31.9	49.8	57.9	53.3	83.9	91.7	61.4	paper	pending
3	SAA-DGL (X-VLM)	Yuan et al.	GeoText-1652 train	X-VLM	30.7	49.2	57.6	52.1	83.3	91.4	60.7	paper	pending
4	GeoText-1652 official checkpoint	GeoText-1652 Team	GeoText-1652 train	Swin-B + BERT	29.9	46.3	54.1	50.1	81.2	90.3	58.7	paper code ckpt	yes

Submit Results

Open a pull request that edits leaderboard/data.json, then run:

python3 scripts/render_leaderboard.py

Required fields for each entry:

• id: stable lowercase identifier, unique within the split
• method: model or system name
• submitter: person, team, or organization
• training_data: training data used, including any external data
• backbone: visual/text backbone or main model family
• metrics: txt_r1, txt_r5, txt_r10, img_r1, img_r5, img_r10

Please include paper/code/checkpoint links when public. Mark verified as false unless the maintainers have reproduced or audited the result.