GeoLocate Benchmark Final

This repository package contains the final publishable benchmark data for GeoLocate remote-sensing grounding and detection experiments.

Datasets

• HRSID/: SAR ship benchmark package with HRSID image chips and pure background images.
• DOTA/: optical multi-class aerial-object benchmark package with DOTA train/val source images staged into the final split protocol.

Each dataset directory contains:

• annotations/images.jsonl: image-level metadata.
• annotations/objects.jsonl: object-level ground truth.
• annotations/tasks_train.jsonl, tasks_val.jsonl, tasks_test.jsonl: benchmark tasks and answers.
• reports/stats.json: dataset statistics.
• manifest.json: build metadata and hashes for release metadata files.
• README.md: dataset-specific protocol notes.

Coordinate Contract

All canonical benchmark coordinates are original-image pixel coordinates (coordinate_system=original_image_pixel). This applies to object geometries, task answers, and future prediction/submission files used for evaluation.

Tile-local coordinates, normalized token coordinates, view-local coordinates, and model-output strings such as <box>...</box> are derived formats only. Predictions from any derived view must be mapped back to original-image pixel coordinates before evaluation.

Current Package Summary

• HRSID: 6,004 images, 16,969 objects, 20,038 tasks.
• DOTA: 1,869 images, 127,843 objects, 20,746 tasks.
• Total staged size: about 15 GB.

Run the local release audits before publishing or after modifying the package:

/nfs/dataset-ofs-perception/detection/chenzhisheng_i/envs/.venv-geolocate/bin/python scripts/audit_hrsid_final.py
/nfs/dataset-ofs-perception/detection/chenzhisheng_i/envs/.venv-geolocate/bin/python scripts/audit_dota_final.py