Investigation into 2D Gaussian Splatting for Image Representation

This work explores the use of Gaussian splatting as a method for representing images as a collection of 2D Gaussians, with a particular focus on balancing image quality and compression. Gaussian splatting, originally developed for representing 3D scenes as collections of 3D Gaussians by minimizing reconstruction loss across multiple photographic viewpoints, can be similarly applied in 2D. By optimizing the positions, colors, and shapes of 2D Gaussians, we aim to minimize reconstruction loss with respect to the target image. This 2D Gaussian splatting approach shares similarities with vectorization techniques like DiffVG and LIVE, which instead optimize the positions and colors of polygons or closed Bezier shapes. However, 2D Gaussians offer the advantage of easier differentiation of pixel values with respect to their positions and colors, resulting in faster optimization. Additionally, 2D Gaussians naturally create smoother color gradients within objects compared to uniformly-colored shapes. We observe a trade-off between compression (fewer Gaussians) and reconstructed image quality. A key finding from our preliminary investigation is that the current implementation uses significantly more 2D Gaussians than necessary. To address this, we introduce an optimization strategy that begins with a small number of Gaussians to capture broad image features, progressively adding more Gaussians to refine finer details. This approach improves the compression-quality trade-off, achieving better reconstruction quality with fewer Gaussians. Furthermore, we address a cropping artifact present in the current implementation and experiment with various reconstruction loss functions.