EatBench-2.7K: A Benchmark for Fine-Grained Eating Action Grounding in Videos

Dataset Summary

EatBench-2.7K is the first video benchmark for fine-grained eating-action grounding. It contains 525 video clips with 2,690 temporally annotated micro-action instances spanning 10 food categories and 3 eating stages. The benchmark is designed to evaluate models on precise temporal localization of sub-second eating micro-actions, posing distinct challenges that current vision-language models (VLMs) are not designed for.

Dataset Details

Eating Action Taxonomy

Each video is annotated with a three-stage eating-action taxonomy:

Label	Abbreviation	Description
Contacting Food	CF	Interaction between hand/utensil and food (reaching, grasping, scooping)
Food Approaching Mouth	FAM	Transporting food from hand/utensil up to the mouth
Food in Mouth	FIM	Ingestion phase after food reaches the mouth (chewing, swallowing)

Statistics

Split	Videos	Action Instances
Test (full)	525	2,690

EatBench-2.7K is a test-only benchmark. The entire dataset serves as the evaluation set for assessing pretrained and training-free methods without task-specific fine-tuning.

Per-Category Statistics

Category	Videos	Instances	Actions/Video	Avg Video Duration (s)	Avg Action Duration (s)
Burger	55	251	4.56	21.82	3.81
Cake	68	364	5.35	20.42	2.87
Carrots	49	237	4.84	21.40	3.28
Chips	58	297	5.12	20.74	3.15
Doughnuts	43	196	4.56	21.01	3.53
Hotdog	42	167	3.98	21.79	4.25
Ice Cream	64	453	7.08	19.15	1.84
Nachos	35	189	5.40	21.78	3.22
Spaghetti	72	319	4.43	20.25	3.59
Watermelon	39	217	5.56	21.72	3.09
All	525	2,690	5.12	20.86	3.11

Action Class Statistics

Class	Count	%	Mean (s)	Median (s)	Min (s)	Max (s)
CF	436	16.2%	3.341	2.312	0.083	21.500
FAM	1,036	38.5%	0.962	0.792	0.083	10.375
FIM	1,218	45.3%	4.860	3.667	0.042	25.208
All	2,690	—	3.113	1.583	0.042	25.208

Key properties:

• 600× temporal scale variation (0.042s to 25.2s)
• FAM is sub-second in 67.4% of instances (median 0.79s)
• 69.3% of videos begin mid-eating-cycle (naturalistic recording)
• Near-deterministic action chain: CF→FAM (94%) →FIM (99%)

Dataset Structure

EatBench-2.7K/
├── README.md
├── eatbench_annotation_full.json   # Full annotations for all 525 videos
└── videos/                         # 525 MP4 video clips
    ├── <video_id>_eating_<category>.mp4
    └── ...

Annotation Format

eatbench_annotation_full.json is a JSON array of 525 entries, one per video:

[
  {
    "Video Name": "Nhj0LHx5BZc_138665_26.50_51.50_eating_cake.mp4",
    "Category": "Cake",
    "Actions": [
      {
        "Label": "Contacting Food",
        "Start": 10.417,
        "End": 12.417,
        "Duration": 2.0,
        "Video_Duration": 26.0
      },
      {
        "Label": "Food Approaching Mouth",
        "Start": 12.417,
        "End": 12.917,
        "Duration": 0.5,
        "Video_Duration": 26.0
      },
      {
        "Label": "Food in Mouth",
        "Start": 12.917,
        "End": 13.583,
        "Duration": 0.667,
        "Video_Duration": 26.0
      }
    ]
  },
  ...
]

Fields:

• Video Name (str): filename of the video clip in videos/
• Category (str): food category (one of 10 categories above)
• Actions (list): temporally annotated micro-action instances
  • Label (str): one of "Contacting Food", "Food Approaching Mouth", "Food in Mouth"
  • Start (float): start timestamp in seconds
  • End (float): end timestamp in seconds
  • Duration (float): End - Start in seconds
  • Video_Duration (float): total duration of the video clip in seconds

Evaluation Protocol

EatBench-2.7K is formulated as a fine-grained temporal action localization task. Given a video, a model predicts a set of temporal segments {(start, end, class)} for the three micro-action classes.

Matching: For each video and class independently, predictions are matched to ground truth via Hungarian matching with tIoU threshold τ. Segments below threshold τ are forbidden from matching.

Metrics: Per-class precision, recall, and F1; macro-averaged F1 across CF/FAM/FIM; and matched mean tIoU (mIoU). Primary evaluation threshold: tIoU = 0.1; also reported at tIoU ∈ {0.3, 0.5}.

Evaluation is score-free: predictions are matched on temporal overlap and class label only, without confidence ranking.

Usage

import json

with open("eatbench_annotation_full.json") as f:
    dataset = json.load(f)

# Each entry
for entry in dataset:
    video_name = entry["Video Name"]
    category = entry["Category"]
    actions = entry["Actions"]  # list of {Label, Start, End, Duration, Video_Duration}
    video_path = f"videos/{video_name}"

Source Data

Video clips are sourced from Kinetics-400, a large-scale public action-recognition benchmark of YouTube videos collected under a Creative Commons license. We retain eating-related action categories and provide new fine-grained temporal annotations (CF, FAM, FIM) created by expert annotators.

Benchmark Results

State-of-the-art zero-shot VLMs evaluated on EatBench-2.7K (Macro-F1 at tIoU=0.1):

Model	Frame Selection	CF F1	FAM F1	FIM F1	Macro-F1
VideoChat-Flash	Uniform	27.6	17.0	37.5	27.4
OneThinker	Uniform	30.4	22.2	43.9	32.2
OneThinker	SAFR	31.5	24.3	44.1	33.3
VAPO-Thinker-7B	Uniform	18.8	8.1	24.8	17.3
VAPO-Thinker-7B	SAFR	19.4	12.1	31.1	20.9
Qwen2.5-VL-7B	Uniform	19.5	11.6	25.0	18.7
Qwen2.5-VL-7B	SAFR	20.9	12.6	26.4	20.0
InternVL3-8B	Uniform	14.9	17.4	28.4	20.2
InternVL3-8B	SAFR	14.1	19.7	32.4	22.1

All models suffer >74% relative performance drop from tIoU=0.1 to tIoU=0.5, highlighting EatBench-2.7K as an open challenge for fine-grained temporal grounding.

License

The annotations are released under CC BY 4.0. The video clips are derived from Kinetics-400 and subject to their original Creative Commons license terms.