Efficient Point-Based Rendering Using Image Reconstruction

Image-space reconstruction of continuous surfaces from scattered one-pixel projections of points is known to potentially offer an advantageous time complexity compared to surface splatting techniques. We propose a new algorithm for hardware-accelerated image-space reconstruction using pull-push interpolation and present an efficient GPU implementation. Compared to published image-space reconstruction approaches employing the pull-push interpolation, our method offers a significantly improved image quality because of the integration of elliptic boxfilters and support for deferred Phong shading. For large point-based models, our GPU implementation is capable of rendering more than 50M points per second—including image-space reconstruction and deferred shading.