StayStill: a large-scale 3D idle animation dataset

How to read this

Category

motion capture dataset and evaluation protocol paper

Contributions

• Releases about six hours of 3D idle motion from 50 subjects, captured with a markerless four camera GoPro rig (Freemocap plus Mediapipe pose estimation and triangulation)
• Provides both natural two minute idle sequences and 18 annotated idle action classes (scratching, stretching, checking a watch, and similar), manually cleaned
• Proposes a first standardised evaluation protocol combining numerical metrics with a 118 participant user study for idle motion generation
• Releases a baseline idle animation synthesiser built on transition concatenation, plus the evaluation code and response data, all publicly available

Context

The authors explicitly model this on LaFAN1's role in motion in betweening research: a dedicated dataset that let a whole subfield standardise its benchmarks. It also extends the same group's earlier ReActIdle dataset, which was only 45 minutes and built to show that genuine and acted idle motion are perceptually indistinguishable, into a much larger resource meant for training deep models. It positions itself against Mixamo and other commercial idle libraries, which are not licensed for machine learning training and are too small anyway.

Correctness

The motion comes from markerless mocap (four synced 1080p GoPro cameras plus Mediapipe pose estimation), which is more accessible than marker based capture but noisier and lower fidelity. The proposed evaluation protocol, numerical metrics plus a 118 person user study, is a first attempt at standardisation rather than an established benchmark. At about six hours from 50 subjects, the dataset is diverse but still modest next to large scale motion or video corpora, so downstream models trained on it will be limited by that scale.

Clarity

A clearly structured Eurographics style paper, readable without a deep machine learning background. Useful directly for anyone who wants an idle animation dataset or a simple baseline model to build on.

How to read it

First pass, skim the abstract, introduction and figures to see what StayStill actually contains and why idle motion has been under served by datasets. Second pass, read the dataset section closely for the hardware setup, the idle action taxonomy, and the evaluation protocol, since those are the parts a rigger or tech animator would reuse directly. Third pass, if you plan to train on it, dig into the baseline transition concatenation model and the released project page code to see exactly how the numerical and user based metrics are computed.