So Predictable! Continuous 3D Hand Trajectory Prediction in Virtual Reality
2022.07.15
So Predictable! Continuous 3D Hand Trajectory Prediction in Virtual
Reality
Nisal Menuka Gamage, Deepana Ishtaweera, Martin Weigel, and Anusha Withana
UIST 2021
SeongOuk Kim
Contents
• Introduction
• Related Works
• Design Goals
• User Study
• Hybrid Kinematic Regressive
Model
• Verification of the Model
• Discussion
• Limitations and Future Work
• Conclusion
3
Introduction
Motivations
• Detecting & Tracking user interaction is important in VR
• Prediction can help it
• Current prediction is usually focused on particular event
Contributions
• A kinematics-based prediction approach for structured and
unstructured ballistic 3D hand movements in VR activities
• A user- and activity-independent model with similar performance to
dependent models
• Evaluation of the model through cross-validation and a secondary
study with new participants and new activities
4
Related Works
Human Motion Prediction Techniques
• Template matching, HMM – hard for arbitrary movements
• Regression(end-point, EMG)
• Deep learning(RNN, CNN, GAN) – too heavy
Motion Prediction for VR
• Reducing lag, foveated rendering, or haptic retargeting
• Head motion, Hand motion
Kinematics of Hand Movements
• Dynamic end-effector models – Minimum jerk model [Hogan, et al]
• Dynamic models for specific task
5
Design Goals
Continuous Prediction
• Can predict at arbitrary time points in the future
Structured & Unstructured Motion
• Not limited to a specific task
User & Activity-Independent
• No re-training for new users & activities
Explainable Prediction
• No black-box
• Can explain
6
User Study
Participants
• 7 female, 13 male
• Mean age = 22.4y(SD= 5.1y)
• 1 left-handed
Apparatus
• Oculus Quest
• OptiTrack with 8 cameras
to record hand movement
• Sensor data + VR screen
7
User Study
T1: Structured Movement via 3D Pointing
• Move hands towards virtual points
• 2 distances(20cm, 40cm), and 2 angular
deviations(30˚, 45 ˚)
8
User Study
T2: Unstructured Movement via VR Gameplay
• 3 minutes for each game
• Beatsaber - slashing, highest average speed (0.72m/s), highest
horizontal & vertical span (0.85m, 0.95m)
• FitXR – Boxing, lower horizontal span (0.57m) & highest frontal
span (0.75m)
• Eleven – varied among users
9
User Study
Data Preparation and Presentation
• Data from Optitrack
• Gaussian filter to reduce noise
• 30s (<10% of data from T1 and 15% of data of T2) as the training
set
10
Hybrid Kinematic Regressive Model
𝑡0: initial time
𝑡: prediction time interval
k: identified value for kinematics
Classical Kinematics of Motion
17
Verification of the Model
Cross-Validation
• 4-folds
New Users & Activities
• 3 more participants & 2 new Activities(Sweeping,
dancing)
18
Discussion
• Better performance than non-naïve
baseline with low overhead
• Can be used for countering delay, error
correction, etc
• Well performed on new users
• Also worked when using Oculus data
19
Limitations and Future Work
Only voluntary ballistic movements
• Need to work on other types of tasks(e.g. steering)
Only participants in 18~39 and not disabled
• Need to test with other age groups
Only wrist data for prediction
• Can use other body data
• Hand ≠ wrist
3D motion is not only for VR
• Can be applied to other fields.
20
Conclusion
• Contributed to novel hybrid classical-regressive kinematic model for
continuous 3D hand trajectory
• Can be used for many areas, even not in VR
My thought
• Impressive for not using deep learning
• Can it be applied for 2-handed cases?