Next-Best-View Planning for Object-Centred 3D Gaussian Splatting

Abstract

Accurate Three-Dimensional Gaussian Splatting (3DGS) reconstruction depends strongly on the acquired images — including their viewpoints, overlap, image quality, and estimated camera poses — making image acquisition a key factor in 3DGS reconstruction quality and visual completeness. This thesis investigates an adaptive Next-Best-View (NBV)-guided image acquisition strategy for object-centred 3DGS reconstruction. The aim is to determine whether the NBV strategy improves reconstruction quality, density, and visual completeness for object-centred 3DGS compared to a predefined image acquisition strategy. A quadrotor-based proof of concept (PoC) was implemented, combining onboard ArUco-based pose estimation and image acquisition with off-board Structure-from- Motion (SfM), 3DGS training, and NBV planning. During testing, several practical limitations — pose-estimation robustness, marker visibility, indoor flight reliability, image quality, and the computational cost of training — prevented the quadrotor from reliably completing the full autonomous workflow. The quadrotor is therefore evaluated as a PoC, while the final comparison between the predefined strategy and the NBV strategy is conducted in a controlled proxy setup using the same reconstruction pipeline and identical 3DGS training settings. The two strategies are assessed using image registration rate, PSNR, SSIM, and LPIPS on held-out evaluation views, together with reconstruction density, computational cost, and qualitative visual inspection. The results show that the NBV strategy improves some reconstruction metrics, but not consistently across all comparisons. For the first test object, the NBV strategy achieved higher PSNR and lower LPIPS than the predefined strategy, while the predefined strategy achieved higher SSIM. The NBV-guided reconstructions were also denser, with a greater number of sparse COLMAP points and optimized Gaussians, but required higher computational cost and produced larger models. For the second test object, the predefined strategy outperformed the NBV strategy in most reconstruction metrics, although the NBV strategy again produced denser reconstructions at the cost of higher computational requirements and larger models. Overall, the results suggest that the NBV strategy has potential as an approach for object-centred 3DGS reconstruction. However, a fully autonomous quadrotor implementation requires more robust pose estimation, flight control, image acquisition, and computational integration before reliable indoor deployment is feasible.