GRW: Gesture Recognition in-the-Wild

This dataset is associated with the paper "Recognizing Co-speech Gestures in-the-Wild" (ECCV 2026).

Our aim is to recognise and localize semantic gestures in real-world videos. These gestures are visually depictive and semantically linked to specific spoken words. We introduce a new large-scale benchmark, GRW (Gesture Recognition in-the-Wild), which provides word-level annotations and gesture boundaries for semantic gestures occurring in unconstrained real-world settings.

---

📋 Table of Contents

• 📚 What is GRW?
• ⚡ Quickstart
• 📦 Getting the visual data
• 📝 Dataset structure
  • Task 1: Semantic classification
  • Task 2: Gesture recognition & localization
  • Data instances
• 📊 Statistics
• 🏆 Benchmark results
• 🔖 Citation
• 📧 Contact
• ⚖️ License & data usage

---

📚 What is GRW?

GRW is a large-scale video dataset for gesture recognition in the wild, containing over 17,000 semantic gesture clips spanning 155 gesture words. Each video clip is 4 seconds long and features a single speaker performing a co-speech gesture. The dataset covers a wide range of gesture categories, including iconic, deictic, metaphoric, and beat gestures, across diverse speakers, backgrounds, and real-world settings. In addition to word-level speech boundaries, the dataset provides manually annotated temporal gesture boundaries for each video clip.

17K+	155	3
Semantic gesture clips	Word vocabulary	Modalities Video · Speech · Text

The dataset supports two tasks:

#	Task	Given	Predict
1	Semantic classification	A candidate video segment and a spoken target word	Whether the segment contains a gesture semantically related to that word
2	Gesture recognition & localization	A gesture video clip	The gestured word, and the temporal gesture boundaries

🎬 Browse the videos on the interactive dataset explorer before downloading.

---

⚡ Quickstart

This repository contains the annotation CSV files. The videos themselves are sourced from YouTube and obtained separately — see Getting the visual data.

from datasets import load_dataset

# Task 1: Semantic classification
sem = load_dataset("sindhuhegde/grw", "semantic_classification")
# splits: "train", "test", "test_unseen_words"

# Task 2: Gesture recognition & localization
rec = load_dataset("sindhuhegde/grw", "recognition_localization")
# splits: "train", "test"

print(rec["test"][0])

Or read the CSVs directly with pandas:

import pandas as pd
from huggingface_hub import hf_hub_download

path = hf_hub_download("sindhuhegde/grw",
                       "recognition_localization_test.csv",
                       repo_type="dataset")
df = pd.read_csv(path)

Files in this repository

File	Task	Split	Rows
semantic_classification_train.csv	Semantic classification	train	135,503
semantic_classification_test.csv	Semantic classification	test	4,000
semantic_classification_test_unseen_words.csv	Semantic classification	test (unseen words)	500
recognition_localization_train.csv	Recognition & localization	train	15,340
recognition_localization_test.csv	Recognition & localization	test	2,000

---

📦 Getting the data

The CSVs give you YouTube IDs, timestamps and speaker bounding boxes. There are two ways to obtain the data — pick whichever suits you.

Option A — Download pre-extracted SHuBERT features (recommended)

Our gesture models take SHuBERT features as input, and we release them directly. This skips video downloading and preprocessing entirely.

Feature set	Download
Semantic Classification — Train	Link
Semantic Classification — Test	Link
Semantic Classification — Test unseen words	Link
Word Recognition & Localization — Train	Link
Word Recognition & Localization — Test	Link

Checksums are available here and can be verified with sha512sum -c SHA512SUMS.

Folder structure of the extracted features

shubert_features (path of the extracted shubert features)
├── semantic_classification
│   ├── split (<train>/<test>/<test_unseen_words>)
│   │   ├── *.npy
├── recognition_localization
│   ├── split (<train>/<test>)
│   │   ├── *.npy

Option B — Download and crop the videos yourself

Use the preprocessing scripts in the code repository:

git clone https://github.com/Sindhu-Hegde/grw.git
cd grw/preprocess

# Download the videos from YouTube-ids and timestamps
python download_videos.py --input_csv=<csv-file> --result_dir=<raw-video-root>

# Crop the videos with the bounding-box co-ordinates provided in the csv files
python crop_videos.py --input_csv=<csv-file> --video_dir=<raw-video-root> --output_dir=<preprocessed-video-root>

⚠️ Note: Due to new YouTube policies, downloading videos (especially for the train sets) might take a long time. If you only need to train or evaluate the gesture models, prefer Option A.

Folder structure after download and pre-processing

raw_video_root (path of the downloaded raw videos)
├── *.mp4 (raw uncropped videos)

preprocessed_video_root (path of the pre-processed gesture videos)
├── word folders
│   ├── *.mp4 (extracted person-specific gesture video)

---

📝 Dataset structure

All frame indices are given at 25 fps.

Task 1: Semantic classification

Video segment annotations for classifying whether a candidate segment contains a gesture that is semantically related to the spoken target word.

semantic_classification_train.csv · semantic_classification_test.csv · semantic_classification_test_unseen_words.csv

Column	Description
fname	unique video clip identifier
target_word	gesture word label
speech_start / speech_end	speech boundaries (in frames at 25fps), automatically extracted using WhisperX
context_start / context_end	frame boundaries (at 25fps) of the context window preceding the candidate segment
target_start / target_end	frame boundaries (at 25fps) of the candidate segment being classified
word_form	surface form of the word as spoken
gesture_label	1 if the candidate segment contains a gesture semantically related to the target word, 0 otherwise
source_file	YouTube video ID
source_video_start / source_video_end	start and end timestamps (in seconds) of the clip within the source video
source_width / source_height	resolution (in pixels) of the source video
num_frames	total number of frames in the original (long) source video; provided for preprocessing
pad / bbox	padding (in pixels) and bounding box [x1, y1, x2, y2] (in pixels) computed on the original source video; used internally to crop the speaker region during preprocessing

Task 2: Gesture recognition & localization

Video segment annotations with manually annotated gesture boundaries, for gesture recognition and temporal localization.

recognition_localization_train.csv · recognition_localization_test.csv

Column	Description
fname	unique video clip identifier
target_word	gesture word label
speech_start / speech_end	speech boundaries (in frames at 25fps), automatically extracted using WhisperX
gesture_start / gesture_end	manually annotated gesture boundaries (in frames at 25fps)
word_form	surface form of the word as spoken
source_file	YouTube video ID
source_video_start / source_video_end	start and end timestamps (in seconds) of the clip within the source video
source_width / source_height	resolution (in pixels) of the source video
num_frames	total number of frames in the original (long) source video; provided for preprocessing
pad / bbox	padding (in pixels) and bounding box [x1, y1, x2, y2] (in pixels) computed on the original source video; used internally to crop the speaker region during preprocessing

🔒 For both tasks, the train and test sets have disjoint videos.

Data instances

Recognition & localization

{
  "fname": "entire/00000",
  "target_word": "entire",
  "speech_start": 35,
  "speech_end": 44,
  "gesture_start": 25,
  "gesture_end": 67,
  "word_form": "entire",
  "source_file": "7IwTua4AaRY",
  "source_video_start": 58.36,
  "source_video_end": 62.36,
  "source_width": 1920,
  "source_height": 1080,
  "num_frames": 282,
  "pad": 454,
  "bbox": "[1259, 521, 2168, 1430]"
}

Semantic classification (a gesture_label = 0 example)

{
  "fname": "look/00012",
  "target_word": "look",
  "speech_start": 197,
  "speech_end": 201,
  "context_start": 0,
  "context_end": 149,
  "target_start": 150,
  "target_end": 249,
  "word_form": "look",
  "gesture_label": 0.0,
  "source_file": "ogCJrrvgais",
  "source_video_start": 1054.48,
  "source_video_end": 1058.48,
  "source_width": 1280,
  "source_height": 720,
  "num_frames": 1725,
  "pad": 426,
  "bbox": "[727, 384, 1579, 1236]"
}

Use the dataset viewer to explore more examples.

---

📊 Statistics

Every clip is 4 seconds long. All counts below are computed directly from the released CSV files.

Task 1: Semantic classification

Split	# Rows	# Words	# Source videos	Gestured (1)	Not gestured (0)
train	135,503	155	34,132	15,340	120,163
test	4,000	100	3,297	2,000	2,000
test_unseen_words	500	10	468	201	299

The test_unseen_words split evaluates generalization to words never seen during training — its 10 words have zero overlap with the 155 training words: crawl, direct, enlarge, great, ingest, loads, proximity, rearrange, synchronize, uniform.

Task 2: Gesture recognition & localization

Split	# Clips	# Words	# Source videos	Clips per word (median)
train	15,340	155	10,407	65
test	2,000	100	1,435	14

All 100 test words are contained within the 155 training words. In the training set, the annotated gestures last 40 frames on average (≈1.6 s) out of the 4-second clip.

Vocabulary

The 155 gesture words

above, absorb, angle, arc, around, ascend, back, balance, barrier, beautiful, begin,
below, big, block, boost, bottom, bounce, branch, break, broad, build, bundle, bye,
call, catch, circle, close, collect, collide, combine, compress, condense, connect,
count, cross, cup, curve, decrease, deep, descend, develop, direction, down, eat,
elevate, embrace, engage, entire, evolve, expand, explode, few, fight, five, flip,
flow, focus, force, four, front, full, gigantic, global, grab, grasp, grow, hashtag,
heavy, height, hello, her, high, hold, horizontal, hug, huge, increase, interaction,
join, knock, large, layer, less, lift, link, little, long, look, loop, lower, many,
merge, mix, move, narrow, no, open, overlap, pause, peak, perfect, pieces, point,
press, process, push, quick, raise, reduce, roll, rotate, round, run, separate,
shake, she, short, shrink, slide, small, specific, spin, spiral, stack, stop,
straight, stretch, strong, switch, three, throw, tie, tight, tilt, tiny, together,
top, track, transform, transition, trap, turn, twist, two, unify, us, various, wait,
walk, wave, whole, wide, wrap, yes, zoom

The full list is also browsable on the Word List page.

target_word is the canonical class label, while word_form records how the word was actually spoken — so inflections are preserved (e.g. quick → quickly, wave → waving, rotate → rotating). The training set contains 381 distinct target_word / word_form pairs.

---

🏆 Benchmark results

Results of our models on the GRW test sets, reproducible with the evaluation scripts.

Semantic gesture classification on the GRW test set

Accuracy	Precision	Recall	High-confidence Accuracy
75.83	79.91	69.00	93.20

Word recognition and localization on the GRW test set

Acc@1	Acc@5	Acc@10	mIoU
18.35	37.30	51.70	0.67

---

🔖 Citation

If you find this dataset helpful, please consider starring ⭐ the repository and citing our work.

@inproceedings{hegde_eccv_2026,
  title={Recognizing Co-Speech Gestures in-the-Wild},
  author={Hegde, Sindhu and Prajwal, K R and Zisserman, Andrew},
  booktitle={European Conference on Computer Vision (ECCV)},
  year={2026}
}

---

📧 Contact

For questions about the dataset, access requests, or collaboration inquiries, please email sindhu@robots.ox.ac.uk.

Author	Affiliation
Sindhu Hegde	University of Oxford
K R Prajwal	University of Oxford
Andrew Zisserman	University of Oxford

Visual Geometry Group (VGG) · Department of Engineering Science · University of Oxford

---

⚖️ License & data usage

The annotations in this repository are released under the Apache 2.0 licence. The underlying videos are sourced from YouTube and are not redistributed here — they remain subject to their original terms of use. The dataset is intended for research purposes.