The Role of Camera Intrinsic Initialization in End-to-End SfM/MVS Reconstruction: A Comparative Evaluation of Metadata, Fallback, and Learned Calibration

For structure-from-motion (SfM), an important step is provided through camera intrinsics, especially when camera metadata are not available. Although subsequent SfM optimization and bundle adjustment (BA) refine the camera parameters using multi-view observations, it remains unclear whether differences in the initial intrinsics have a lasting effect on the optimization process and, ultimately, on the quality of the reconstructed scene. This study investigates the role of camera intrinsic initialization in an end-to-end AliceVision/Meshroom SfM and multi-view stereo (MVS) pipeline, with particular focus on whether similar optimized intrinsics necessarily lead to similar final reconstructions. Here, three initialization strategies are evaluated within an AliceVision/Meshroom pipeline: camera intrinsics obtained from image metadata, AliceVision’s 45° field-of-view fallback, and single-image calibration using a pretrained GeoCalib model. GeoCalib is integrated only at the camera-initialization stage, where its predicted intrinsics are written into the AliceVision SfMData structure. All subsequent stages of the reconstruction pipeline remain unchanged, including feature extraction, image matching, feature matching, SfM, bundle adjustment, dense reconstruction, meshing, and texturing. Experiments are conducted on the 41-image Monstree smartphone collection and the 27-image ETH3D Observatory sequence. ETH3D additionally provides reference camera calibration and sparse reconstruction data, allowing the effect of initialization to be measured against ground truth. Here, the evaluation is extended to selected six-view, eight-view, and twelve-view Observatory subsets, the 26-view Boulders scene, and paired dense reconstructions of the 23-view ETH3D Terrace training scene. The experiments show that the three methods can produce considerably different initial camera parameters while still converging to similar optimized focal length. For the evaluated datasets, this suggests that the final intrinsic parameters are strongly influenced by multi-view optimization rather than by the initialization alone. The experiments also show that more accurate single-image calibration does not necessarily result in a correspondingly different or improved optimized focal length. However, similarity in the final intrinsic parameters does not, by itself, demonstrate that the resulting dense reconstructions are equivalent. The six-view and Boulders comparisons show that the convergence observed on the full Observatory sequence does not extend to every tested configuration. Terrace achieves F1-scores of 0.8273 for fallback and 0.8261 for learned initialization at 2 cm against public laser scans. These findings highlight the need to distinguish between the quality of the initial camera calibration, the recovery of camera parameters during SfM optimization, and the quality of the final dense reconstruction. These findings remain limited to the selected subsets and tested scenes.