CUHK-X brings together multimodal sensing, datasets, and benchmarks to study how people act in everyday environments. The goal is to help models recognize actions, understand their context, and reason about what may happen next.
The dataset contains 64,267 samples of 40 actions from 30 participants across two indoor environments. It combines seven synchronized modalities: RGB, depth, infrared, thermal, IMU, mmWave radar, and skeleton. Its benchmarks cover human action recognition (HAR), understanding (HAU), and reasoning (HARn).
ACM MobiSys 2026
Introduces the CUHK-X dataset and benchmarks for recognizing, understanding, and reasoning about human activities.
PDF | Code | Project PageACM MobiSys 2026 Demo
A portable system that synchronizes multimodal activity recordings and supports annotation during data collection.
Paper© Siyang Jiang