no code implementations • 25 Jul 2022 • Anique Akhtar, Zhu Li, Geert Van der Auwera
We employ convolution on target coordinates to map the latent representation of the previous frame to the downsampled coordinates of the current frame to predict the current frame's feature embedding.