Improvements in Compression Techniques for 3D Gaussian Splattering Data

3D Gaussian Splatting (3DGS) is an emerging technology in novel view synthesis. To address the large storage space required by synthesized scenes, this work proposes an efficient compression method. The original adaptive density control strategy is modified to enhance the gradient weight progressively with increasing iterations, allowing for more precise density control. Sensitivity-aware vector clustering compresses multiple attributes of Gaussian points, including color, covariance matrices, and opacity. Finally, small Gaussian points with high similarity in color, shape, and opacity are merged to optimize storage further. Experiments across multiple datasets demonstrate that the proposed method achieves an average compression ratio of 39.5 times compared to the original 3DGS, with an average increase of 0.307 dB in PSNR.