2D Gaussian Splatting With Pre-Trained Dictionaries for Image Compression

Gaussian Splatting has recently shown excellent performance on 2D and 3D representation and compression tasks. It requires fewer GPU resources and achieves faster rendering compared to implicit neural representation methods. However, its large number of parameters limits its applications on image compression. To address this, we propose a new image compression method based on 2D Gaussian Splatting with attribute dictionaries. In particular, we design a dictionary matching module, which dynamically determines whether to replace, predict, or retain Gaussian parameters according to the difference between the learned parameters and dictionary items. In this way, the attributes of compressed Gaussians can be represented more efficiently with limited bits. We design scaling factors to ensure that the pre-trained dictionary can be applied to compression tasks for images of any resolution. Furthermore, we design an efficient entropy coding for compressing the dictionary indices. Experimental results show that our method achieves a 17.96% BD-Rate saving and +0.54 dB BD-PSNR gain on the Kodak dataset.