Overview

XRFv2 Plus is a synchronized multimodal dataset update for sensor-vision-language action understanding. It is built from the original XRF V2 recording corpus, but reorganizes the release around a common cropped-video timeline so that sensing streams, body representations, temporal labels, captions, and action QA share the same sequence IDs and relative-time axis.

The public update was released on June 6, 2026. The README reports that the Kaggle release was verified through version 4 on June 5, 2026, 23:39 CST.

XRFv2 Plus sensing setup and synchronized release content — Overview of the XRFv2 Plus synchronized release content, cropped from the update manuscript.

What Is New

Category	Release content
Sensing streams	WiFi CSI, five-position IMU, AirPods IMU, RGB video embeddings, Kinect depth videos, Kinect infrared videos
Body information	Human 2D pose, depth-assisted 3D pose, SMPL mesh, DensePose-style surface information
Action labels	Temporal action localization, action captioning, action question answering
Time convention	All labels and synchronized tensors use relative time within each sequence

Key differences between XRF V2 and XRFv2 Plus — Key differences between the original XRF V2 release and the XRFv2 Plus synchronized sensor-vision-language benchmark.

Release Summary

Standardized samples	853
Original candidates	880
Excluded candidates	27 with incomplete five-position IMU coverage
Volunteers	16
Scenes	Diningroom, Studyroom, bedroom
Action classes	34
Action segments	8,157
Sequence duration range	43-89 seconds
Actions per sequence	7-11
Total video-aligned duration	16 h 14 min 38 s

Summary table for the XRFv2 Plus release — Release summary from the manuscript, including standardized samples, scenes, actions, and total video-aligned duration.

Temporal statistics for XRFv2 Plus sequences and action segments — Temporal statistics for the 853 video-aligned sequences and 8,157 TAL segments.

Samples with shortened video-aligned annotations — Samples where video-aligned annotations removed the last one or two actions due to shortened RGB video coverage.

Modalities

For a sequence of duration T seconds, the release provides synchronized tensors and files including WiFi CSI amplitude/phase at 50 Hz, five-position IMU at 50 Hz, AirPods Pro IMU at 25 Hz, one 1024-D RGB video feature per second, COCO-17 2D pose, depth-camera 3D pose, compact ROMP/SMPL mesh parameters, DensePose-style IUV information, and cropped Kinect depth/infrared videos at 15 FPS and 512 x 512 resolution.

The five-position IMU order is fixed as left wrist, right wrist, phone case in the left pants pocket, phone case in the right pants pocket, and glasses temple.

Main data files and tensor conventions in XRFv2 Plus — Main data files and tensor conventions for the standardized XRFv2 Plus release.

Synchronized RGB infrared depth pose mesh and DensePose visual example — Synchronized visual example showing RGB, infrared, depth, 2D pose, 3D pose, mesh, and DensePose-style outputs generated from the same frame.

Annotation Tasks

XRFv2 Plus includes ActivityNet-style temporal action localization labels, Chinese-only, English-only, and bilingual caption files, and action QA files. The QA release contains 108,929 QA pairs across 21 question types, including existence, counting, segment duration, temporal order, and time-point action questions.

English action question answering examples from XRFv2 Plus — English action QA examples from sample 0_Diningroom_1.

Action QA type distribution in XRFv2 Plus — Distribution of action QA types across the XRFv2 Plus release.

Release Batches

The README lists four public Kaggle versions: version 1 for the fast no-video release, version 2 for 853 cropped Kinect depth videos, version 3 for 853 cropped Kinect infrared videos, and version 4 for the DensePose H5 file. Raw Kinect recordings and cropped RGB videos are not included in the Kaggle release for privacy reasons.

Resources

Citation

@misc{wang2026xrfv2plus,
  author    = {Wang, Fei},
  title     = {XRFv2 Plus: A Multimodal Sensor-Vision-Language Dataset for Action Understanding},
  year      = {2026},
  publisher = {Zenodo},
  doi       = {10.5281/zenodo.20564312},
  url       = {https://doi.org/10.5281/zenodo.20564312}
}