Gilfoyle727/vr-ray-pointer-landing-pose

VR Ray Pointer Landing Pose Dataset

This dataset accompanies the paper "Predicting Ray Pointer Landing Poses in VR Using Multimodal LSTM-Based Neural Networks." It contains the raw trajectory archives used for the paper's two user studies, plus the original data processing code used to prepare model inputs.

Paper link: IEEE Xplore

The data captures bare-hand raycasting selection behavior in VR with multimodal time-series signals from hand, head-mounted display (HMD), and gaze channels. The paper reports that the full dataset covers 72,096 trials across two empirical studies:

• Study 1: 55,296 trials
• Study 2: 16,800 trials

Paper Summary

The paper studies target-agnostic prediction of the final ray landing pose during VR pointing and selection. The proposed model is an LSTM-based predictor trained on time-series features derived from three modalities:

• hand movement
• HMD movement
• eye gaze movement

According to the paper:

• Study 1 recruited 16 participants
• Study 2 recruited 8 new participants
• Data was recorded at 90 Hz
• Hardware used a Meta Quest Pro
• The model achieved an average prediction error of 4.6 degrees at 50% movement progress

Included Files

• Study1_Raw.zip Raw CSV trajectories for Study 1.
• Study2_Raw.zip Raw CSV trajectories for Study 2.
• Dataprocessing_code.zip Original preprocessing scripts provided by the authors.
• data_processing_code/ Extracted copy of the preprocessing scripts for easier browsing on Hugging Face.

Data Format

Each raw archive contains per-participant CSV files with frame-level trajectories. Typical columns include:

• participant / block / trial identifiers
• error flag
• target geometry variables such as depth, theta, phi, width, and position
• task progress and distance traveled percentage
• timestamp
• HMD position and forward vector
• hand position and forward vector
• left-eye position and forward vector
• right-eye position and forward vector
• target location and target scale

The data is sampled over time during reciprocal pointing selections.

Study Design From The Paper

Study 1

The paper describes Study 1 as a within-subjects design over:

• target depth combinations: De and Ds in {3m, 6m, 9m}
• theta values: 10, 15, 20, 25, 50, 75 degrees
• phi values: 0 to 315 degrees in 45 degree steps
• target widths: 4.5 and 9 degrees

The paper reports:

• 55,296 total trials
• 16 participants
• reciprocal 3D pointing with no distractors

Study 2

The paper describes Study 2 as a validation study with:

• 8 new participants
• theta varying continuously across all integer values from 15 to 84 degrees
• 350 trial combinations
• 50 blocks
• 6 reciprocal selections per trial combination
• 2,100 trials per participant

The paper reports 16,800 total trials for Study 2.

Important Notes About The Raw Archives

This repository preserves the raw files exactly as provided by the dataset owner. A few practical details matter when using the archives:

• Study1_Raw.zip currently contains 19 CSV files
• Study2_Raw.zip currently contains 8 CSV files
• the observed raw trial counts are 64,308 trials in Study1_Raw.zip and 16,800 trials in Study2_Raw.zip
• some Study 1 CSV files do not include a ParticipantID column in the header
• some Study 1 and Study 2 files share participant-like file IDs such as 72
• raw archive contents therefore do not map one-to-one to the participant counts reported in the paper without additional curation context
• specifically, Study1_Raw.zip includes a 72_Trajectory.csv file with 2,100 trials, which matches the Study 2 per-participant protocol rather than the Study 1 per-participant total of 3,456 trials reported in the paper

For reproducibility, this repository keeps the original archives unchanged. When reconstructing participant identity for Study 1, you may need to use the filename as the participant identifier when ParticipantID is absent from the CSV header.

Recommended Usage

• Use Study1_Raw.zip and Study2_Raw.zip as the authoritative raw data sources.
• Use the scripts in data_processing_code/ to reproduce feature engineering and preprocessing.
• If you build a Hugging Face datasets loader on top of this repository, treat the raw zip files as the source of truth rather than assuming fully standardized CSV schemas.

Citation

If you use this dataset, please cite the paper:

@inproceedings{xu2025predictingray,
  title={Predicting Ray Pointer Landing Poses in VR Using Multimodal LSTM-Based Neural Networks},
  author={Xu, Wenxuan and Wei, Yushi and Hu, Xuning and Stuerzlinger, Wolfgang and Wang, Yuntao and Liang, Hai-Ning},
  booktitle={IEEE Conference on Virtual Reality and 3D User Interfaces},
  year={2025}
}

Acknowledgements

This dataset was collected for the paper above and uploaded to Hugging Face by the dataset owner.