3 ms·
How do training sets for something like this get built up? Creating a synthetic image of a foreground object pasted onto an existing background wouldn't seem to
by teruakohatu 6y ago
How do training sets for something like this get built up? Creating a synthetic image of a foreground object pasted onto an existing background wouldn't seem to work because it would not capture the the complex interaction eg. The image of an astronaut walking in a field of grass.
But hand labelling would be incredibly costly, far more than image classication.
There must be an interesting trick in generating training data.
- rland 6y agoBlender is also an option, it can be scripted and it's possible to get pixel-perfect segmentation masks that way. The downside is obviously that you have to have a suitable number of 3-D scenes pre-built. But I think typically pasting foreground objects into scenes randomly is basically how these things are trained.
- adriang133 6y agoInteresting, that makes sense. You still need a lot of diverse/realistic cropped objects that you can (realistically?) put into scenes. I'm not sure if simply dropping random objects into random scenes would work well enough. Also, what exactly do the nets predict, since the foreground objects can have very complex contours, you'd need to predict an arbitrarily shaped polygon ?
- ad404b8a372f2b9 6y agoAny existing image dataset coupled with unsupervised methods would work pretty well for this. 10 years or so ago I worked on a similar project, we used SLIC superpixels to segment the images then spectral clustering to cut out the foreground object.