BMVC 2015

[arxiv 1507.05726] Rule Of Thumb: Deep derotation for improved fingertip detection [PDF] [notes]

Aaron Wetzler, Ron Slossberg, Ron Kimmel

read 19/05/2017

Objective

Improve accuracy of tasks that rely on hand image analysis by 'derotating' the image

derotation : find the rotation that transforms the original image to a 'canonical' pose, with the base of the thumb at a specific location

Synthesis

Using derotation should reduce the variance in pose space, and therefore improve training

Pipeline

• segmentation using flood-fill method (determines area connected to a given node)
• depth-dependent bounding box
• derotation around the centor of mass of segmented hand according to angle produced by DeROT network
  • predict 3 DOF hand orientation, directly predict 9 coeffs of rotation matrix
  • trained without enforcing orthonormality, using Euclidian loss
  • projected into SO(3) using SVD decomposition
  • deduce angle to derotate the hand in the image

Dataset HandNet

200k depth images

created using magnetic trackers

Results

Shows that derotation improves mAP (mean Average Precision) significantly for fingertip detection tasks