Research
We publish technical reports on contactless sensing and self-supervised learning on radar: what we tried, what worked, and where the open problems are.
Preprint, 20269 min read
MAEPose: learning human pose from raw mmWave video without labels
MAEPose pre-trains a video Transformer on unlabelled mmWave radar video with masked autoencoding, then decodes multi-frame joint heatmaps, reducing pose error by up to 22.1% against state-of-the-art baselines across three datasets.
Read the report