I saw the following as a valid sample in the training data:
Clearly, the squeezed buzz saws do not look right. What I realized is that it was probably supposed to look like this:
I made the above image manually in paint, but the problem is that in the original data, there is a discrete object that is being cut because of how we are defining the window, and as a result, we are teaching the model that weird object slices are possible. We probably shouldn't do this.
So, for any object that takes up multiple tiles, let's never allow samples that slice those objects in the middle. Since we are sliding the window across the level scene, we can just stick to the samples that don't chop objects up, and we essentially have as much data as we want, so there shouldn't be any problems with this.
I saw the following as a valid sample in the training data:
Clearly, the squeezed buzz saws do not look right. What I realized is that it was probably supposed to look like this:
I made the above image manually in paint, but the problem is that in the original data, there is a discrete object that is being cut because of how we are defining the window, and as a result, we are teaching the model that weird object slices are possible. We probably shouldn't do this.
So, for any object that takes up multiple tiles, let's never allow samples that slice those objects in the middle. Since we are sliding the window across the level scene, we can just stick to the samples that don't chop objects up, and we essentially have as much data as we want, so there shouldn't be any problems with this.