Typos, some significantly affecting performance

[e841018]

I attempted to slightly modify the body estimation part in order to fine-tune the backbone network, and noticed several typos. Eventually, I refactored almost everything and noticed significant improvements on estimation accuracy, although using the same pretrained model and thresholds. Since I modified some APIs, and I can't recall what changes I made was critical to the improvement, instead of making a pull request I would like to ask the original contributors to take a look at my version and see if something can help in improving things here.

Original output for ski.jpg	My version

Some differences:

• The lower legs of the left-most person is not crossed.
• The second person from the left has lower legs recognized.
• The third person from the left has eyes recognized more accurately.
• The false positive on the upper right is gone.

Some of the changes I remember are listed below, but I think they are not the main reason of the improvement:

• Reverse the RGB order in util.draw_bodypose to align with other implementations better.
• Exchange some X and Y in util.draw_bodypose. The original version is not erroneous but just misleading.
• Remove an extra ReLU by correcting _L1 to _L2 at the end of this line
• Change the clustering behavior here.

[e841018]

I just found the main source of improvement. My default scale search option is [0.7, 1.0, 1.3], while the original scale_search is [0.5]. Sorry for the bothering, but I think correcting the typos would still help the readers in the long term, so I would keep the issue open.