orena-dkfz/heico-focus-vqa

HeiCo-FOCUS (Beta release)

A clinically grounded dataset for long-context video understanding in minimally invasive surgery.

📄 Paper  •  🤗 Dataset  •  💻 Code  •  🏆 Challenge  •  ⚖️ CC BY-NC-SA 4.0

Until 31 July 2026, we will employ a restricted post-release review period. During this time, we kindly ask the community to provide feedback to help us increase quality control and make continuous improvements.

---

Overview

Recent advances in Vision-Language Models (VLMs) have enabled impressive performance on video understanding benchmarks. However, existing evaluations largely focus on short-term reasoning and fail to assess a critical capability: maintaining cumulative temporal consistency over extended time horizons.

HeiCo-FOCUS closes this evaluation gap by providing a clinically grounded dataset for long-context video understanding through the task of Foreign Object Contextual Understanding in Surgery. Built on a dataset of Heidelberg Colorectal surgeries, the dataset requires models to continuously track multiple objects as they are inserted, manipulated, occluded, and removed over procedures lasting up to several hours.

From a long-video benchmarking perspective, foreign object tracking is an attractive evaluation target because it inherently demands persistent memory, temporal consistency, and aggregation over extended time horizons. Unlike short-context tasks, accurate counting requires that models integrate information across the full duration of a procedure — errors accumulate over time and cannot be corrected from a single frame or short clip. The task is objectively verifiable and clinically meaningful, linking technical performance directly to real-world patient safety.

Figure 1: Overview of the HeiCo-FOCUS dataset. a) Clinical motivation. Ensuring the retrieval of foreign objects at the end of a surgery is critical, as retained objects can lead to severe complications. b) Dataset overview. The dataset comprises 96 hours of surgical video annotated with 15,000 VQA pairs spanning five core capabilities.

---

Dataset at a Glance

Property	Value
Videos	30 full-length laparoscopic procedures
Procedure types	Rectum resections, proctectomies, sigma resections
VQA pairs	15,000
Capability groups	5
Leaf capabilities	15
Evaluation tracks	3 (FRAME, SEGMENT, PROCEDURE)
Annotators	39 domain experts + crowd workers
Video duration	Up to several hours

---

Clinical Background

Despite rigorous manual protocols, retained surgical items (RSIs) remain a critical patient safety problem. Surgical sponges, needles, clips, and other instruments inserted into the body cavity during minimally invasive surgery must be fully accounted for before wound closure. Unintentional retention is rare but can cause serious complications.

HeiCo-FOCUS addresses this problem by providing the first expert-validated dataset designed to assess whether VLMs can provide reliable, temporally consistent reasoning about foreign objects in endoscopic video — an essential prerequisite for deployment in high-stakes clinical settings.

---

Evaluation Tracks

The dataset uses a multi-track framework that systematically increases temporal and contextual demands:

Track	Config name	Visual input	Description
FRAME	frame	Single frame	Tests short-context perception. No temporal modelling required.
SEGMENT	segment	<= 5min clip	Tests understanding of motion and event context within a short window.
PROCEDURE	procedure	Up to full video	Tests long-horizon reasoning over complete procedures lasting up to hours.

Models may be evaluated on any subset of tracks. Comparing performance across tracks reveals whether capability degradation stems from temporal scale or from the task itself. The all_tracks config combines all three tracks into a single split for convenience.

---

Capability Taxonomy

VQA pairs are structured around five capability groups spanning from basic perception to high-level clinical inference.

Figure 2: Hierarchical taxonomy of VQA capabilities with example questions.

#	Group	Leaf capabilities
1	Object Recognition	Identification, Instance Matching, Attributes, Spatial (camera), Spatial (situs)
2	Temporal Grounding	Temporal Localization, Duration Estimation
3	Aggregation	Object Aggregation, Event Aggregation
4	Event & Procedural Understanding	FO Interaction Recognition, FO Usage Purpose, Temporal Ordering
5	Complex Reasoning	Functional Reasoning, Causal & Consequence Reasoning, Multi-step Reasoning

---

Dataset Structure

QA annotations are stored as parquet files, one per track and split:

data/
  frame/
    train.parquet
    test.parquet
  segment/
    train.parquet
    test.parquet
  procedure/
    train.parquet
    test.parquet

The all_tracks config merges all six files, making it easy to load the full dataset in one call. Video files are hosted separately in the videos/ folder of the repository.

---

Dataset Schema

Each row in the dataset follows this schema:

Field	Type	Description
id	string	Unique question identifier
video	string	Video filename (e.g. 0015 - Heico - Rektum - 6.avi)
timestamp_start	string	Start of the relevant time window (HH:MM:SS)
timestamp_end	string	End of the relevant time window (HH:MM:SS)
procedure_type	string	Surgical procedure name
question	string	The clinically grounded question
answer	string	Expert-validated ground-truth answer
answer_format	string	Expected answer format (e.g. number, binary, time, fo_class)
primary_capability	string	Primary capability (e.g. object_identification)
secondary_capabilities	list[string]	Additional capabilities required to answer correctly
clinical_relevance	bool	Whether the question is directly clinically relevant
ood	bool	Whether the question is considered out-of-distribution

---

Usage

Loading annotations

QA annotations are fetched directly from HuggingFace — no local setup required:

from datasets import load_dataset

ds = load_dataset("orena-dkfz/heico-focus-vqa", "segment", split="test")
print(ds[0]["question"])   # "How many sponges are visible?"
print(ds[0]["answer"])     # "2"

Using the orena-focus library

The orena-focus library provides dataset loaders, answer-format parsing, and a full evaluation framework:

pip install orena-focus

from focus import FocusDataset, DatasetSplit, Track

ds = FocusDataset("heico", DatasetSplit.TEST, Track.SEGMENT)
request, reference = ds[0]
print(request.question)        # "How many sponges are visible?"
print(reference.answer)        # "2"
print(reference.format.type)   # "number"

To run inference with video input, download the video files first:

from focus import FocusConfig, set_config, download

set_config(FocusConfig(root_dir="/data/focus"))
download("heico")  # downloads video files into /data/focus/heico/videos/

See the library repository for end-to-end inference and evaluation examples.

---

Citation

If you use HeiCo-FOCUS-VQA in your research, please cite the dataset paper and the original HeiCo dataset:

@article{luttnermayer2026heicofocus,
  author  = {[TODO]},
  title   = {HeiCo-FOCUS: A Clinically Grounded Benchmark for Long-Context Video Understanding in Minimally Invasive Surgery},
  journal = {arXiv preprint},
  year    = {2026},
  doi     = {[TODO]},
}

@article{maier2021heidelberg,
  author    = {Maier-Hein, Lena and Wagner, Martin and Ross, Tobias and Reinke, Annika and Bodenstedt, Sebastian and Full, Peter M and Hempe, Hellena and Mindroc-Filimon, Diana and Scholz, Patrick and Tran, Thuy Nuong and others},
  title     = {Heidelberg colorectal data set for surgical data science in the sensor operating room},
  journal   = {Scientific Data},
  volume    = {8},
  number    = {1},
  pages     = {101},
  year      = {2021},
  publisher = {Nature Publishing Group UK London},
  doi       = {10.1038/s41597-021-00882-2},
}

---

Acknowledgements

HeiCo-FOCUS-VQA was developed at the Division of Intelligent Medical Systems (IMSY), German Cancer Research Center (DKFZ), Heidelberg. The dataset is the basis of the ORena SAVE FOCUS challenge at MICCAI 2026. The project was partially funded through the SAVE program.

The underlying video data is from the HeiCo dataset (Maier-Hein et al., 2021). We gratefully acknowledge all annotation contributors, crowd workers, medical students, and other domain experts whose effort made this dataset possible.

---

License

The dataset is released under CC BY-NC-SA 4.0. Non-commercial use only, with attribution and share-alike conditions.