Скачать или смотреть AdaGaR: Sharper 3D Dynamic Scene Reconstruction

In this AI Research Roundup episode, Alex discusses the paper: 'AdaGaR: Adaptive Gabor Representation for Dynamic Scene Reconstruction' AdaGaR introduces a unified framework for 3D dynamic scene reconstruction from monocular videos, addressing issues like blurry textures and temporal flickering. The method extends 3D Gaussian Splatting by using an Adaptive Gabor Representation that captures high-frequency details while maintaining energy stability. By incorporating learnable frequency weights and energy compensation, the model preserves intricate textures like fur or sharp edges. Additionally, the authors use an orthographic coordinate system to ensure smooth motion continuity across frames. This approach prevents common artifacts like motion tearing during video interpolation. Paper URL: https://arxiv.org/abs/2601.00796 #AI #MachineLearning #DeepLearning #ComputerVision #3DGaussianSplatting #DynamicScenes #GaborRepresentation