Scene Slicer is a program to automatically slice a 2D scene image into multiple layers of objects and background.
When given an image, the slicer automatically identifies foreground objects, masks them out, and inpaints the hole with reasonable imagined background.
Ziwen Chen
The purpose of Scene Slicer is to facilitate 3D scene recontruction -- by cleaning up the clutters in the foreground, we can make the geometry of the underlying environment (e.g. an indoor room) more apparent, and thus make it easier for a scene resontructor to fathom the shape of the environment. After the global coordinate is pinned down, foreground objects can be more easily localized inside the global coordinate.
Currently, I'm using Mask RCNN combined with DeepFill as the backbone of Scene Slicer. So far, RCNN-based networks are still in leading position regarding to accuracy compared to other kinds of object detectors. DeepFill is a state-of-the-art background inpainter. I trained both networks using MSCOCO dataset from scatch.
The inpainter still needs some improvement it seems...
