Despite significant advancements in Structure-from-Motion (SfM) over the past few years, state-of-the-art systems still struggle with extreme challenges such as low overlap, low parallax, and high symmetry in scenes. These limitations make it difficult for non-professional users to capture images that work well with SfM, significantly hindering its widespread application. To address these issues, we have developed MP-SfM, a method that incorporates monocular depth and normal priors—derived from deep neural networks—into the traditional SfM framework. This integration of single-camera constraints with multi-view constraints allows our method to outperform existing techniques in scenarios with extreme perspective changes while maintaining strong performance under standard conditions. One of the key advantages of MP-SfM is its ability to resolve erroneous match associations caused by symmetry, a longstanding problem in the SfM domain. By leveraging these monocular priors, MP-SfM can reliably reconstruct challenging indoor environments from just a few images, making it the first technique capable of doing so. Our approach also includes systematic uncertainty propagation, which makes it robust to errors in the priors. This feature allows the method to adapt easily to different models without requiring extensive parameter tuning, ensuring that it can readily benefit from future advancements in monocular depth and normal estimation technologies. We have publicly released the code for MP-SfM on GitHub (https://github.com/cvg/mpsfm), making it accessible for researchers and practitioners to explore and improve upon this innovative approach. This development represents a significant step forward in the field of 3D reconstruction, particularly for applications where capturing a large number of images is impractical or challenging.