Pose Metrics, PoseProcessors, and DTW

[cleong110]

Working on Pose Metrics, which use PoseProcessors, which will allow us to run DTW first, then call distance measures

Code broken out from #4

[cleong110]

@AmitMY Could use your thoughts on the following:

• PoseMetric's main distinction from BaseMetric is that it runs poseprocessors first
• However some poseprocessors operate on exactly 1, exactly 2, or all of the poses. For example the reduce to intersection one only makes sense if all hypotheses and references go through at once.
• fastdtw requires you to pass it exactly two 1-D arrays, you cannot pass it the whole hyp_data and ref_data. I added code to iterate accordingly over keypoint trajectories.
• fastdtw seems pretty slow! It takes nearly a minute to run on my laptop for just two poses.
• DTWMeasure is currently inheriting from AggregatedPowerDistancesMeasure but that doesn't seem quite right.

Currently the score with signatures looks like this:

DTWPowerDistance|hb:no|pre:[normalize_poses|s:1,reduce_poses_to_intersection,remove_world_landmarks,hide_legs|remove:yes]|distance_measure:{power_distance|pow:2.0|dflt:0.0|agg:mean} = 5.923352071881241

[AmitMY]

quite an easy merge. just run black, and some minor comments

[cleong110]

@AmitMY before making any code changes, what are your thoughts on #19 (comment)

[AmitMY]

PoseMetric's main distinction from BaseMetric is that it runs poseprocessors first

Good

However some poseprocessors operate on exactly 1, exactly 2, or all of the poses. For example the reduce to intersection one only makes sense if all hypotheses and references go through at once.

I think that's fine, we can always call the processor with all available poses we have, and then process them either in a batch or separately depending on the processor (you wrote that code)

fastdtw requires you to pass it exactly two 1-D arrays, you cannot pass it the whole hyp_data and ref_data. I added code to iterate accordingly over keypoint trajectories.

I see what you did. There's advantages and disadvantages to working in 1D. working in 178D is probably more accurate for the entire body, but doesn't allow different phonemes to happen at different times (face movement and hand movement for example) - but single keypoint trajectory can also be optimized for strange unaligned outputs, where there is no structure...

fastdtw seems pretty slow! It takes nearly a minute to run on my laptop for just two poses.

Strange. should be much faster.

DTWMeasure is currently inheriting from AggregatedPowerDistancesMeasure but that doesn't seem quite right.

You don't have to inherit there. can extend Mesaure if you prefer.