A Lightweight 3D Convolutional Autoencoder Architecture for Temporal Coherence in 3D Video

Abstract

We propose a novel lightweight 3D convolutional autoencoder architecture designed to efficiently encode and decode spatiotemporal information from 3D video data while preserv- ing temporal coherence between frames. We present a theoretical analysis on the stability of temporal feature representations and validate the approach on a synthetic 3D video dataset of moving volumetric shapes. Experimental results demonstrate the effectiveness of our method in reconstructing 3D videos with high fidelity and smooth temporal transitions, highlighting its potential for real-world 3D video processing applications.