In recent years, computer vision researchers have proposed several deep learning 3D reconstruction methods, with methods based on an implicit function representation finding great success. When using an implicit function, the network learns to extract features from the input image and make predictions separately for each point. For example, the network decides whether a query point is inside an object or outside of it, or it predicts the signed distance value of that point to the objectʹs surface. The 3D object is reconstructed by running the network on batches of points in 3D space. Out of the various existing methods based on implicit functions, methods using local image features produce state-of-the-art results. However, these methods generally assume simple camera models, such as the orthogonal or weak-perspective models, in order to facilitate network convergence and speed up the training process. Also, when it comes to multi-view, implicit function based methods widely require known camera parameters at test time in order to find corresponding features for a query point across images. In this research, a single-view and multi-view 3D reconstruction method is proposed that is based on the implicit function representation and local features. For each input image, a sparse point cloud is predicted, and point cloud registration is used to find relative camera parameters. A perspective camera model is assumed, and query point depth values are normalized relative to the average point of the predicted point cloud. The model performance is eva‎luated relative to a competing method using a weak-perspective projection camera. Furthermore, multi-view reconstruction results are presented, without the need of known camera parameters at test time. This shows the practicality of our method in real-life uses.

عبدالرضا ميرزايي

مهران صفاياني،‌ فرزانه شايق