Closed boyleconnor closed 1 year ago
It's still not very scientific but I have tried playing with a few examples and 5.0
seems like a much better touching threshold than 50.0
.
More importantly, I made a significant bugfix. The mask extraction code was extracting anywhere there was an exact match along any of the (r, g, b)
values; instead it should be (and I think now is) extracting based on exact matches across all of the (r, g, b)
values. I have also added a minimum threshold of 20 pixels, below which a mask will be dropped. This helps remove a lot of trivially small masks.
Here is the current multi-object curriculum (incl. black backgrounds)
multi_obj_curriculum.zip