mirror of
https://github.com/qgis/QGIS.git
synced 2025-10-08 00:05:09 -04:00
91100f816b
* Initial implementation * - Add disabling of blur pass - Change default parameter for shading factor - fix some tests * - Implement bilateral filtering - Optimize textures - Fix layout test * rename SSAO to ambient occlusion fix ssao settings widget margin * Remove unused variable * fix naming and add tooltips * - Refactor quad entities - Address Stefanos's suggestions * Rework SSAO implementation Previously we based the code on CloudCompare's implementation, however that did not work too well for us for a couple of reasons: - the code does not deal well with perspective projection, causing incorrect shading (CC uses orthographic projection) - there was no range check, so we would be getting false ambient occlusion on larger depth discontinuities (silhouttes) - banding artifacts as the sampling kernel was not getting rotated - parameters (shading radius, distance attenuation) that are difficult to understand The new implementation is based on John Chapman's tutorial and LearnOpenGL page (derived from the original tutorial): https://john-chapman-graphics.blogspot.com/2013/01/ssao-tutorial.html https://learnopengl.com/Advanced-Lighting/SSAO The general approach of the SSAO is the following: - for each pixel, we pick a couple of random points nearby (64 samples currently) and check with the depth buffer whether they are visible from the camera or not. The nearby points that are occluded contribute to darkening of the pixel, this is saved to a texture - in the next rendering pass, we blur the texture using 4x4 box. This is because in the first step we use 4x4 random noise pattern and it leaves a noticeable noise pattern on the screen. This pass gets rid of that noise - in the post-processing step, the blurred texture is blended with the rendered scene There are few differences to J.C.'s tutorial and LearnOpenGL page: - the approches above use normal maps (a texture with a normal vector for each pixel), but we don't because we also want to support point clouds that do not have normals (at least not by default) - we use full sphere for sampling instead of hemisphere (which is possible when you have normals), so maybe we are getting a bit lower quality / performance - LearnOpenGL also uses a texture with positions of all pixels - we only use depth map to get the original positions (like JC's original code does) * Clean up ssao parameters and GUI, add intensity parameter * Add missing Q_OBJECT macro * Add more missing Q_OBJECT macros * Add occlusion threshold parameter to control the darkening With the default threshold of 50%, pixels only start to get darker when more than half of the neighborhood samples are occluded. That means flat surfaces should not get any darker. (What we had previously is an equivalent of having threshold set at 0%) The downside is that with increased threshold, more subtle occlusions get lost. * Review from Stefanos * More review and better defaults * Clear button fix Co-authored-by: NEDJIMAbelgacem <gb_nedjima@esi.dz>