alyssarosenzweig / synthpose Goto Github PK

This is import for realism in both the depth and RGB maps -- hair can obscure body parts, leading to incorrect classifications. As it is, hair sometimes gets read as skin by some algorithms -- simulating it is the least we can do!

Ideally, more models would be used, but for the moment, the paper does allow for ±10% changes.

At the time of writing, a new blender instance is created for each image generated, contributing to overall low performance.

At the time of writing, a model takes about five seconds to render on my computer. Given the shear number of images required for learning the human pose space, this is an issue -- that being said, it should not be difficult to prune the blender file.

Due to graphics (and art!) limitations, this feature should not be relied upon solely, but used in conjunction with e.g. the depth maps, it should be okay.

The render farms synthpose is expected to run on have huge numbers of cores -- it is not out of scope to take advantage of them here (as opposed to delegating that off to the user).

Enabling the slower lighting features might help. Better textures might also help.

In the paper on which this is based, a (marker-based) mocap system is used. synthpose should be able to handle this type of data