3D Gaussian Splatting reconstruction with depth enhanced initialization

This thesis proposes a novel pipeline incorporating segmentation masking and depth val- ues to enhance the performance of Gaussian Splatting techniques. By leveraging 3D Gaussian Splatting’s ability to achieve high-accuracy photorealistic reconstructions, the pipeline focuses on singular object reconstruction through segmentation masking, which removes unwanted backgrounds and accelerates the optimization process. Integrating re- cent advances in fast semantic segmentation using neural networks, the pipeline produces nearly ready-to-use models. Additionally, using a depth-sensing camera during acquisi- tion allows for more accurate point cloud initialization with minimal overhead, leading to a significant boost in visual accuracy during the early optimization steps. The pipeline achieves a significant speedup, making it particularly advantageous for devices with lim- ited computational capacity, though with a slight trade-off in the accuracy of the final results. The depth-enhanced initialization is carried out by sampling and projecting mean- ingful point information into the reconstructed space, offering a better approximation of under-sampled regions. This step also ensures the reconstruction is to scale, enabling precise measurements and further analysis.