More efficient prediction strategy for Object Detection

[lewfish]

Currently we use a sliding window with 50% overlap for making predictions in object detection in order to make sure that objects will appear wholly (ie. untruncated) within some window. But maybe it isn't so bad to predict truncated bounding boxes, and we can run 4x faster by using no overlap. And also avoid having to run the post-processing step to remove duplicate predictions.

[lewfish]

Other ways of speeding this up:

• Read in whole image to RAM instead of doing windowed reads off of disk?
• Test if the post-processing step in object detection is a bottleneck (I think it's an n^2 operation)
• Do chip generation on the CPU in parallel with prediction on the GPU. This is what Keras does.
• Feed in batches with more than one image at a time.