We present MAGVIS, a maglev-based volumetric display platform that enables immersive, glasses-free 3D telepresence for multiple participants. The system uses a rapidly rotating LED matrix driven by magnetic levitation to generate real-time volumetric imagery, providing a realistic sense of co-presence without specialized eyewear.
We systematically review 166 exoskeleton evaluation studies published between 2015 and 2026 across industrial and clinical settings using a three-group factor framework. The analysis reveals a severe imbalance: device-oriented metrics dominate (83.1% biomechanics, 51.8% mechanical interfaces) while only 7.8% conducted thermal measurements, 45.8% omitted participant age, and none of the 123 industrial studies recruited older workers. To address these gaps, we propose an evidence-based multimodal sensor selection criterion to guide comprehensive evaluation protocols in real-world settings.
EchoVision proposes a hybrid NPU-CPU deployment strategy combining EfficientViT-SAM and YOLO for real-time assistive navigation on mobile devices. By splitting workloads across the neural processing unit and CPU, the system achieves low-latency scene understanding suitable for visually impaired users in dynamic environments.
We introduce OOKPIK, a dataset of out-of-context image-caption pairs for cheapfake detection research. The dataset comprises 545 images and 1090 real captions organized as triplets {image, caption1, caption2}, with a baseline model proposed for the MMSys'21 grand challenge on detecting out-of-context media.
We present a drone detection system combining YOLOv4 with Seq-NMS post-processing and ByteTrack object tracking to improve detection performance in video data. The approach demonstrates strong real-time detection capability on a curated drone video dataset.