JuHwan_Park

Hello, I'm Juhwan

Video-based monocular depth estimation, temporal attention, causal/bidirectional masking
Stable attention KV-cache methods for streaming inference
Teacher–student distillation (feature & depth), self-supervised learning, domain adaptation

Closing the stream vs. offline gap: cache-mix/self-forcing, Kalman-like gating, causal masks
Feature-wise KD at the temporal head (post-attention)
Multi-GPU & memory-efficient pipelines (Docker/Conda, xFormers) and reproducible evaluation (scale-shift alignment)