New scene generation using advanced 3D Gaussian splatting
This study investigates a limitation of 3D Gaussian Splatting in real-time novel view synthesis and proposes a solution to address it. Despite its state-of-the-art performance, 3D Gaussian Splatting often results in unbalanced degrees of reconstruction, leading to blurred areas in rendered images (a...
Saved in:
主要作者: | |
---|---|
其他作者: | |
格式: | Final Year Project |
語言: | English |
出版: |
Nanyang Technological University
2024
|
主題: | |
在線閱讀: | https://hdl.handle.net/10356/175315 |
標簽: |
添加標簽
沒有標簽, 成為第一個標記此記錄!
|
總結: | This study investigates a limitation of 3D Gaussian Splatting in real-time novel view synthesis and proposes a solution to address it. Despite its state-of-the-art performance, 3D Gaussian Splatting often results in unbalanced degrees of reconstruction, leading to blurred areas in rendered images (as depicted in Fig. 1). To address this issue, a progressive frequency control technique is implemented within the training process as inspired by FreeNeRF and FreGS. Specifically, Fourier transformation is applied
to both the resulting images and ground truth to map them into frequency spectrums, facilitating the learning of different frequency levels separately through the introduction of smooth filter masks. Experimental results on widely adopted benchmarks, including Mip-NeRF360, Tanks&Temples, and Deep Blending, demonstrate enhanced performance through the proposed frequency optimization technique. The study concludes that integrating frequency optimization into 3D Gaussian Splatting holds promise for
enhancing rendering quality. Moreover, recommendations for future research include exploring additional techniques such as compression and regularization to further enhance performance during the optimization process. |
---|