thxplz/HowToEat-test

HowToEat: Hand-Object Interaction and Eating Action in Eating Scenarios

HowToEat is an image dataset for analysing eating behaviour. It provides:

1. Hand-object interaction + eating face detection (hand_object_detection): 95,190 images with 190,333 hand instances (box, left/right side, contact state, and the box and category of the held object) and 151,620 face instances (box, eating / not eating).
2. Eating action recognition (eating_recognition): 6,280 manually labelled faces (eating / not eating) for image classification.

The images are frames from 6,701 publicly available eating videos covering 12 eating and drinking scenarios.

• Paper: Yingcheng Wang, Junwen Chen, Keiji Yanai. HowToEat: Exploring Human-Object Interaction and Eating Action in Eating Scenarios. MADiMa '23 (8th International Workshop on Multimedia Assisted Dietary Management, in conjunction with ACM Multimedia 2023). doi:10.1145/3607828.3617790
• Institution: Department of Informatics, The University of Electro-Communications, Tokyo, Japan
• License: HowToEat Research-Only License: non-commercial research and education only (see License)
• Not included: the source videos and the trained models are not released.

---

Quick start

Access is gated: accept the license on this page, then log in with huggingface-cli login.

from datasets import load_dataset

# Task 1: hand-object interaction + eating face detection
det = load_dataset("thxplz/HowToEat-test", "hand_object_detection")
sample = det["train"][0]
sample["image"]          # PIL.Image, 1920x1080
sample["hands"]          # list of hand instances
sample["faces"]          # list of face instances

# Task 2: eating action recognition
rec = load_dataset("thxplz/HowToEat-test", "eating_recognition")
rec["train"][0]["face_crop"], rec["train"][0]["label"]   # 224x224 crop, 0 = eating

# Stream instead of downloading everything (~35 GB)
det_stream = load_dataset("thxplz/HowToEat-test", "hand_object_detection", split="test", streaming=True)

---

Dataset structure

HowToEat/
├── hand_object_detection/       # Parquet shards (images embedded), Task 1
├── eating_recognition/          # Parquet shards (images embedded), Task 2
├── annotations/                 # the same annotations in the original JSON format
│   ├── detection_train.json     #   paper split
│   ├── detection_test.json
│   ├── detection_train_balanced.json   # balanced split (see "Splits")
│   ├── detection_test_balanced.json
│   ├── eating_recognition_train.json
│   └── eating_recognition_test.json
├── metadata/
│   └── categories.json          # id <-> name maps
├── LICENSE.md
└── README.md

Image files are identified by file_name = "<verb>/<category>/<video_id>_<frame>.jpg", for example eating/pizza/v3174_011858.jpg. <video_id> is an anonymous video ID (v0001 … v6701) and <frame> is the 6-digit frame index within that video. Frames with the same video_id come from the same video.

Scene categories (12)

hamburger, beer (under drinking/), bread, pizza, pasta, noodles, with_knife_and_fork, sushi, with_spoon, with_fork, sandwich, with_chopsticks. All except beer are under eating/. The category describes the scene of the whole video; it is not a per-instance label.

Config hand_object_detection

Column	Type	Description
image	Image	Full frame, 1920×1080 JPEG
file_name	string	Key shared with the JSON annotations
video_id	string	Anonymous video ID (e.g. v0123)
frame	int	Frame index in the source video
category	ClassLabel (12)	Scene category of the video
width, height	int	Image size
hands	list of struct	One entry per hand, see below
faces	list of struct	One entry per face, see below
split_balanced	string	"train" / "test" in the balanced split (see Splits)

hands[i]:

Field	Description
hand_box	[x_min, y_min, x_max, y_max], absolute pixels
hand_side	0 = left, 1 = right
ho_exist	1 = the hand is in contact with a portable object, 0 = no contact
obj_box	Box of the held object, [] when ho_exist = 0
obj_cat	Object category (table below), -1 when ho_exist = 0

faces[j]:

Field	Description
face_box	[x_min, y_min, x_max, y_max], absolute pixels
eat	1 = eating, 0 = not eating

Object categories (obj_cat):

id	name	instances	id	name	instances
0	food	58,720	7	bottle	6,282
1	chopsticks	34,663	8	cup	10,594
2	fork	9,268	9	glass	9,742
3	spoon	10,783	10	can	875
4	knife	12,098	11	napkin	7,598
5	bowl	3,656	12	unknown	19
6	plate	1,121	−1	(no contact)	24,914

Example record (original JSON format in annotations/detection_*.json):

{
  "file_name": "eating/pizza/v3174_011858.jpg",
  "hand_obj": [
    {"hand_box": [928, 406, 1062, 519], "hand_side": 0, "ho_exist": 1,
     "obj_box": [931, 391, 1005, 453], "obj_cat": 0},
    {"hand_box": [900, 417, 979, 550], "hand_side": 1, "ho_exist": 1,
     "obj_box": [922, 390, 1006, 478], "obj_cat": 0}
  ],
  "face": [{"face_box": [903, 264, 1057, 411], "eat": 1}]
}

(In the JSON files, the hand list is called hand_obj and the face list face.)

Config eating_recognition

Column	Type	Description
image	Image	Full frame the face was taken from
face_crop	Image	224×224 square crop around face_box with 15% context (the crop used at test time in the paper)
file_name, video_id, frame, category		As above
face_box	list[int]	[x_min, y_min, x_max, y_max] of the labelled face
label	ClassLabel	0 = eating, 1 = not_eating

⚠️ Label polarity differs between the two configs. In eating_recognition, label = 0 means eating. In hand_object_detection, eat = 1 means eating. We kept the conventions of the original training code for both.

---

Splits

hand_object_detection

The train / test splits of this config are the split used in the paper (baseline results below). No video appears in both splits. Because the split was made in category order, the test set does not cover all categories: hamburger, beer, bread, pizza, pasta, noodles and sandwich appear only in train.

For a test set that covers every category, we also provide a balanced split: a random 4:1 split by video, where each category has about 11–24% of its images in test. It is available as the split_balanced column and as annotations/detection_*_balanced.json. No baseline results have been published on the balanced split.

from datasets import load_dataset, concatenate_datasets
det = load_dataset("thxplz/HowToEat-test", "hand_object_detection")
full = concatenate_datasets([det["train"], det["test"]])
train_bal = full.filter(lambda s: s == "train", input_columns="split_balanced")
test_bal  = full.filter(lambda s: s == "test",  input_columns="split_balanced")

	Paper split (train / test)	Balanced split (train / test)
Images	76,905 / 18,285	76,025 / 19,165
Videos	5,550 / 1,104	5,351 / 1,303
Hands	155,634 / 34,699	153,532 / 36,801
Faces	122,183 / 29,437	122,033 / 29,587

Images per category:

Category	Total	Paper train	Paper test	Balanced train	Balanced test
hamburger	19,239	19,239	0	15,520	3,719
sushi	15,843	2,799	13,044	12,511	3,332
pizza	14,221	14,221	0	11,820	2,401
beer	11,701	11,701	0	9,258	2,443
noodles	10,891	10,891	0	8,511	2,380
bread	7,644	7,644	0	6,081	1,563
pasta	7,465	7,465	0	5,664	1,801
sandwich	2,834	2,834	0	2,258	576
with_chopsticks	1,727	35	1,692	1,359	368
with_spoon	1,582	2	1,580	1,290	292
with_knife_and_fork	1,165	71	1,094	974	191
with_fork	878	3	875	779	99
Total	95,190	76,905	18,285	76,025	19,165

eating_recognition

Split	Faces	Eating	Not eating	Videos
train	5,033	3,108	1,925	537
test	1,247	770	477	353

The split is stratified 4:1 within each (category, label) group, as in the paper. It was made per image, not per video, so 332 videos have frames in both train and test. Keep this in mind when you interpret test accuracy.

---

Statistics (hand_object_detection, both splits combined)

• Hands: 190,333 (left 85,797 / right 104,536). In contact with an object: 165,419. No contact: 24,914.
• Faces: 151,620 (eating 69,251 / not eating 82,369). Every image has 1–5 faces. A few images contain faces but no annotated hands.
• Faces per image: 1 face 56,153 images · 2 faces 27,518 · 3 faces 7,142 · 4 faces 2,880 · 5 faces 1,497.

---

Baseline results

Results reported in the paper on the paper split (test set), with mAP at IoU > 0.5 (see the paper §5.1 for the hand-object matching rule).

SOV-STG-H2E-S (multi-task, single model):

Left hand: no contact	Left hand: portable object	Right hand: no contact	Right hand: portable object	Hand-object mAP	Face: not eating	Face: eating	Face mAP
61.91	87.79	47.98	88.56	71.56	57.43	73.89	65.66

Eating action recognition (eating_recognition), ResNet-50 (ImageNet-1K pre-trained, fine-tuned): 86.4% test accuracy.

The trained models are not released.

---

How the dataset was built

1. Video collection. Eating and drinking videos were collected for 12 scenarios (e.g. eating hamburger, drinking beer, eating with a spoon). The videos themselves are not distributed.
2. Frame extraction. A PPDM hand-object interaction detector trained on 100DOH and a RetinaFace (ResNet-50) face detector were run at 1 frame per second. Frames where a hand-held object overlapped the mouth landmarks were kept: 99,903 frames.
3. Eating labels for faces. 6,280 face crops were labelled manually (eating / not eating). This is the eating_recognition config. A ResNet-50 classifier trained on them then labelled all faces automatically.
4. Hand-object annotation. An SOV-STG-Hand model (trained on 100DOH, re-categorised into no contact / portable object) produced hand, side, contact and object boxes. Object categories were added and all annotations were then checked and corrected manually with the VIA annotation tool. Frames that could not be annotated reliably were marked invalid.
5. Filtering. We removed images without faces, images with more than 5 faces, and images whose largest face is smaller than 400 px. Faces smaller than 300 px were removed. Together with the invalid frames, this reduced the 97,484 verified frames to 95,190 images.

Labelling rules for eating faces: a face is eating if the person is performing the act of eating (mouth open with food or utensil entering it, or the face clearly shows eating while an object covers the mouth). An object in front of a closed mouth counts as not eating. Images that cannot be judged were not labelled.

---

Limitations and biases

• Partly automatic labels. Face boxes come from RetinaFace. Eating labels on detection faces come from a classifier, and hand-object boxes were pre-annotated by a model before manual checking. Some errors remain (see the paper, Fig. 6c).
• Domain. The videos are mostly eating vlogs filmed for an audience: frontal faces, often a single person, good lighting. The foods are limited to the 12 scenario categories, and the people and regions shown are not representative of the world population.
• Boxes outside the image. Some boxes extend beyond the image border (for example, negative coordinates for a face cut off at the top), mostly face boxes produced by the face detector. They are kept exactly as used in the paper; clip them to [0, width] × [0, height] if your code requires it.
• Category imbalance. unknown (id 12) has only 19 instances and can 875. The with_* categories are much smaller than the food categories.
• Paper split coverage. See Splits: the paper's test set covers only 5 of the 12 categories in meaningful numbers.

Ethical considerations

The images show real, identifiable people taken from publicly available videos. The dataset is intended only for research on eating behaviour and dietary assessment. It must not be used for face recognition, identification, tracking or profiling of individuals (see the license).

Removal requests. If you appear in the dataset or own one of the source videos and want content removed, contact <contact email> with the file_name or video_id. We will remove it from this repository.

---

License

The annotations, metadata and scripts are released under the HowToEat Research-Only License. In short:

• ✅ Non-commercial research and education
• ❌ Commercial use of any kind, including training commercial models
• ❌ Redistribution or re-hosting (a few example images in publications are fine)
• ❌ Face recognition, identification or surveillance
• 📌 Citation of the paper is required

The images are frames from publicly available online videos. Their copyright belongs to the original video owners, and the authors do not grant any rights to them. They are made available for research use only, to the extent permitted by applicable law.

---

Citation

@inproceedings{wang2023howtoeat,
  title     = {{HowToEat}: Exploring Human Object Interaction and Eating Action in Eating Scenarios},
  author    = {Wang, Yingcheng and Chen, Junwen and Yanai, Keiji},
  booktitle = {Proceedings of the 8th International Workshop on Multimedia Assisted Dietary Management},
  series    = {MADiMa '23},
  pages     = {71--78},
  year      = {2023},
  publisher = {Association for Computing Machinery},
  address   = {New York, NY, USA},
  location  = {Ottawa, ON, Canada},
  doi       = {10.1145/3607828.3617790},
  url       = {https://doi.org/10.1145/3607828.3617790}
}

Acknowledgments

This work was supported by JSPS KAKENHI Grant Numbers 21H05812, 22H00540, 22H00548, and 22K19808.