Update README.md

README.md

@@ -1,3 +1,92 @@
 ---
 license: cc-by-4.0
+language:
+- en
+task_categories:
+- object-detection
+- visual-question-answering
+tags:
+- heico
+- orena
+- surgical-data-science
 ---
+
+# OReNA FOCUS FRAME — HEICO Foreign-Object Annotations
+
+Hugging Face mirror of the **v14-heico** annotation release, derived from the source snapshot dated 14 September 2026.
+
+**Canonical release and citation:** [10.5281/zenodo.22751012](https://doi.org/10.5281/zenodo.22751012).
+
+## Authors
+
+1. [Tristan Kirscher](https://orcid.org/0009-0004-6646-6548) — ICube Laboratory, CNRS UMR-7357, University of Strasbourg, Strasbourg, France; CLCC Institut Strauss, Strasbourg, France.
+2. [Lalith Sharan](https://orcid.org/0000-0003-0835-042X) — CAMMA, ICube, UMR7357, University of Strasbourg, CNRS, INSERM, Strasbourg, France; Institute of Image-Guided Surgery, IHU Strasbourg, Strasbourg, France.
+3. Andrea Naclerio — CAMMA, ICube, UMR7357, University of Strasbourg, CNRS, INSERM, Strasbourg, France; Institute of Image-Guided Surgery, IHU Strasbourg, Strasbourg, France.
+4. [Luc Vedrenne](https://orcid.org/0009-0002-0577-4307) — CAMMA, ICube, UMR7357, University of Strasbourg, CNRS, INSERM, Strasbourg, France; Institute of Image-Guided Surgery, IHU Strasbourg, Strasbourg, France.
+5. [Nicolas Padoy](https://orcid.org/0000-0002-5010-4137) — CAMMA, ICube, UMR7357, University of Strasbourg, CNRS, INSERM, Strasbourg, France; Institute of Image-Guided Surgery, IHU Strasbourg, Strasbourg, France.
+
+## Dataset contents
+
+The release contains annotations for **1,274 frames from 20 HEICO training videos**, with **2,505 bounding-box instances**. Seven of the ten taxonomy categories are represented.
+
+| Category | Bounding boxes |
+| --- | ---: |
+| Sponge | 291 |
+| Clip | 896 |
+| Specimen Bag | 74 |
+| Silicone Loop | 410 |
+| External Drain | 524 |
+| Needle | 256 |
+| Gallstone | 0 |
+| Specimen | 54 |
+| Mesh | 0 |
+| Absorbable Hemostatic Agent (provisional) | 0 |
+
+Counts refer to instances across frames, not distinct physical objects tracked across videos. Zero examples do not establish clinical absence.
+
+## Download
+
+- [Complete annotation archive](orena_frame_heico_annotations_v14_zenodo.zip)
+- [Technical note (PDF)](HEICO_annotation_data_note.pdf)
+- [SHA-256 checksums](SHA256SUMS.txt)
+
+Download and extract the ZIP to access the individual files. The package includes structured annotations in JSONL and COCO formats, frame facts, DINO crop specifications, quality flags, derived question–answer pairs, a source-image mapping manifest, author metadata, and a citation file. `PACKAGE_MANIFEST.json` records file sizes and SHA-256 hashes.
+
+**No source images, videos, model weights or LapChole data are included.** DINO records describe crops but do not contain crop pixels.
+
+## Using the annotations
+
+`image_file_name`, `image_path`, and COCO `file_name` are unique export aliases, not original dataset paths. Use `source_image_manifest.jsonl` to map the alias or task ID to `source_relative_path`, relative to an authorized HEICO root containing `frames/`. Verify `image_sha256` before use. Original filenames can repeat across videos; do not flatten paths to basenames.
+
+Source-image hashes were preserved and checked for consistency against source metadata; they were not recomputed from original image bytes during release preparation. Obtain source images and their access terms from the dataset owners.
+
+There are **4,586 canonical QA rows**. The existing frame flag filter removes 86 frames, retaining **1,188 frames and 4,228 canonical QA rows**. The optional text-QA union contains **7,013 rows**. Use `augmented_vqa_clean.jsonl` OR its four constituent family files, never both. Names containing `balanced` describe upstream generation; the HEICO subsets are not exactly balanced.
+
+The full structured annotation, COCO and DINO files retain flagged frames. To apply the same filter to detection training, exclude flagged task IDs and their dependent records. Keep all frames, crops and derived QA from a video in the same downstream split.
+
+## Annotation process and limitations
+
+Source records designate 834 frames as label-assisted and 440 as image-only. The source handoff describes human acceptance, correction or drawing of detector/SAM 2 proposals. Workflow completion and exhaustiveness fields are not independent measurements of annotation correctness.
+
+This is selected training material, not an unbiased test set, prevalence sample or clinical reference standard. Structural consistency and package integrity were checked; independent image-level clinical adjudication was not performed during release preparation.
+
+Quality flags identify recorded discrepancies and ambiguity, not all possible annotation errors. A total of 1,012 boxes have uncertain visibility, including 1,010 after frame filtering. Historical Clip undercounts were observed in an earlier subset; residual undercounting in this release has not been quantified. The detailed technical note explains these limitations and the interpretation of default or uncertain attributes.
+
+Annotator identities, free-text notes, annotation timestamps and source question IDs were removed. Structured labels, visibility fields and quality flags were retained. Scientific information loss from removed notes has not been independently assessed. The exact use of this snapshot by the submitted model remains unconfirmed.
+
+The term “foreign object” follows the challenge taxonomy and does not imply that an object was unintentionally retained. Empty-frame and absence answers describe the visible image field, not the entire operation.
+
+## Source attribution
+
+Maier-Hein, L., Wagner, M., Ross, T. et al. *Heidelberg colorectal data set for surgical data science in the sensor operating room*. Scientific Data **8**, 101 (2021). [10.1038/s41597-021-00882-2](https://doi.org/10.1038/s41597-021-00882-2).
+
+- [OReNA challenge](https://or-arena.org/)
+- [HEICO FOCUS VQA source](https://huggingface.co/datasets/orena-dkfz/heico-focus-vqa)
+
+## License and citation
+
+The additional annotations and accompanying documentation are released under [Creative Commons Attribution 4.0 International (CC BY 4.0)](https://creativecommons.org/licenses/by/4.0/). Source images and videos retain their own access and reuse terms.
+
+Copyright 2026 Tristan Kirscher, Lalith Sharan, Andrea Naclerio, Luc Vedrenne, Nicolas Padoy.
+
+Please cite the [Zenodo record](https://doi.org/10.5281/zenodo.22751012) when using these annotations. The archive contains `CITATION.cff`. Cite the original HEICO publication as appropriate when using the source dataset.