danielemateria/hd_epic_interaction_anticipation_bench_som_gaze

Dataset Card for HD-EPIC Gaze Interaction Anticipation: SoM & Gaze-Augmented Clips

Dataset Summary

This dataset provides visually-augmented versions of the Gaze Interaction Anticipation subset of the HD-EPIC VQA benchmark (Perrett et al., CVPR 2025), built for the paper:

Leveraging Gaze and Set-of-Mark in VLLMs for Human-Object Interaction Anticipation from Egocentric Videos Daniele Materia, Francesco Ragusa, Giovanni Maria Farinella — ICPR 2026

The original benchmark consists of 1,000 multiple-choice VQA items, each paired with a 10-second egocentric video clip from HD-EPIC, trimmed to end 0.3s after the user's gaze primes the next object interaction. This release does not re-record or re-label those items — it adds a processing pipeline that renders each clip into several visual variants used as VLLM input:

• Set-of-Mark (SoM) masks — semantic segmentation masks generated with Semantic-SAM (α = 0.05 transparency), overlaid on the last frame only of each clip. No alphanumeric tags are used, only the masks.
• Gaze-trajectory overlays — for each frame, the W = 15 most recent ground-truth 2D gaze fixations are rendered as connected circles that fade from red (most recent) to blue (oldest), visualizing the user's recent attention.
• Combinations of the two (SoM + Gaze), plus the unmodified RGB clips as a baseline.

The goal is to let a Vision-Language Model be probed under different visual-cue conditions (raw / gaze / SoM / SoM+gaze) without re-running the annotation pipeline.

Supported Tasks

• Human-object interaction anticipation, formulated as multiple-choice VQA: "What object will the person interact with next, ignoring ongoing interactions?" — 1 correct object name + 4 distractor names sampled from other objects moved within the surrounding ~2-minute segment.
• Suitable as a VLLM probing benchmark for visual-grounding and gaze-conditioning ablations (this was its original use in the paper: zero-shot evaluation of LLaVA-OneVision-7B and Gemini 2.0 Flash).

Languages

English (question text and object-name candidates, inherited from the HD-EPIC annotations).

Dataset Structure

Directory layout

Following the processing pipeline in the official repo, clips are organized by modality:

Folder	Content	Corresponding --mode
video_segments/	Raw RGB 10s clips, extracted from HD-EPIC long videos	standard
Gaze_video_segments/	RGB clips with gaze-trajectory overlay	gaze
SoM_last_video_segments/	RGB clips with SoM masks on the last frame	som / som_last
SoM_last_Gaze_video_segments/	SoM masks (last frame) + gaze trajectory combined	som_gaze / som_last_gaze

Each folder mirrors the same 1,000 clip identifiers, so the same video_id can be looked up across modalities.

Data Instances

A single instance corresponds to one HD-EPIC "gaze interaction anticipation" question. In its augmented form it consists of:

video_id: <HD-EPIC clip identifier>
clips:
  standard: video_segments/<video_id>.mp4
  gaze:     Gaze_video_segments/<video_id>.mp4
  som:      SoM_last_video_segments/<video_id>.mp4
  som_gaze: SoM_last_Gaze_video_segments/<video_id>.mp4
question: "What object will the person interact with next, ignoring ongoing interactions?"
candidates: [<correct object>, <distractor_1>, ..., <distractor_4>]
answer: <correct object>

The question / candidates / answer fields come directly from HD-EPIC's gaze_interaction_anticipation.json (see Source Data below) — this release does not redistribute that file, only the derived video clips keyed to the same video_ids.

Data Fields

• video_id (string) — unique clip identifier, shared across all four modality folders.
• start_time, end_time (string) — clip boundaries in the source HD-EPIC long video.
• question (string) — fixed VQA prompt template.
• candidates (list[string], length 5) — 1 correct object name + 4 distractors.
• answer (string) — ground-truth object name.

Data Splits

There is a single split of 1,000 items, used exclusively for zero-shot evaluation in the paper (no train/val split — no fine-tuning was performed on this subset).

Dataset Creation

Source Data

• Base dataset: HD-EPIC (Perrett et al., CVPR 2025) — 41 hours of unscripted, multi-day egocentric kitchen recordings across 9 kitchens, captured with Meta Project Aria glasses, with dense 3D-grounded annotations and eye-gaze data.
• Base benchmark subset: the Gaze Interaction Anticipation question type from HD-EPIC's 26K-question VQA benchmark (1,000 of the 26K questions).
• Distractor construction (inherited from HD-EPIC, not modified here): 4 negative candidates per question, randomly sampled from other objects moved within the surrounding ~2-minute segment.

Annotation / Augmentation Process

This is not a manual re-annotation — it is an automated visual-augmentation pipeline (3 steps, see the repo for exact scripts):

1. Clip extraction — 10-second segments are cut from raw HD-EPIC long videos using timestamps aligned to the original benchmark annotations.
2. SoM masking — Semantic-SAM generates semantic segmentation masks for the final frame of each clip (α = 0.05 transparency, no alphanumeric tags).
3. Gaze rendering — ground-truth 2D gaze fixations (projected from HD-EPIC's raw gaze + Aria VRS data) are drawn as a fading red→blue trajectory of the last 15 fixations, optionally on top of the SoM-masked frame.

Licensing Information

• The processing code in the official repository is released under the MIT License.
• The underlying HD-EPIC video/annotation data is distributed by its authors under a Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0) license.

Citation Information

If you use this dataset, please cite both the paper that produced this augmented version and the original HD-EPIC paper.

@inproceedings{materia2026leveraging,
  title     = {Leveraging Gaze and Set-of-Mark in VLLMs for Human-Object Interaction Anticipation from Egocentric Videos},
  author    = {Materia, Daniele and Ragusa, Francesco and Farinella, Giovanni Maria},
  booktitle = {International Conference on Pattern Recognition (ICPR)},
  year      = {2026}
}

@InProceedings{perrett2025hdepic,
  author    = {Perrett, Toby and Darkhalil, Ahmad and Sinha, Saptarshi and Emara, Omar and Pollard, Sam and Parida, Kranti and Liu, Kaiting and Gatti, Prajwal and Bansal, Siddhant and Flanagan, Kevin and Chalk, Jacob and Zhu, Zhifan and Guerrier, Rhodri and Abdelazim, Fahd and Zhu, Bin and Moltisanti, Davide and Wray, Michael and Doughty, Hazel and Damen, Dima},
  title     = {HD-EPIC: A Highly-Detailed Egocentric Video Dataset},
  booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  year      = {2025},
  month     = {June}
}

Additional Information

• Paper (arXiv): https://arxiv.org/abs/2604.03667
• Code: https://github.com/fpv-iplab/leveraging_gaze_som_vllms_human_obj_anticipation
• Base dataset website: https://hd-epic.github.io
• Corresponding author: Daniele Materia — daniele.materia@studium.unict.it