SuYeon Kim

Detecting 3D defects does not require memorizing every valid shape: framing anomalies as local relational inconsistencies unlocks robust, category-agnostic point cloud inspection across unseen domains.

Class confusion in few-shot object detection can be drastically reduced, leading to a +10.1 nAP improvement over previous methods.

Stop treating concept drift as one thing: DynaME's hybrid approach, separating recurring and emergent drifts, unlocks better online time series forecasting.

Generate more realistic and nuanced human movements from text by explicitly modeling individual body parts, overcoming the limitations of existing holistic approaches.

Dataset distillation for time series can now be done efficiently and scalably by operating in the frequency domain, sidestepping the limitations of prior methods.