Topic: Computational-cheap hand RoI extraction
by Fabian Stockhardt
online Big Blue Button Room: D19/2.03a, January 27, 2022 (Thursday), 12.00 noon
Keywords — object-detection, hand, normalisation, 3D, RoI
The problem of hand detection and RoI extraction seems to be solved:
Freely available open source solutions like google’s media pipe offer high performance rates and high generalizability for a wide range of use cases in the field of object detection. Models trained on huge amounts of data are provided free of charge, as is the source code. Nevertheless, there are reasons to look for alternatives:
No model is better than the data it was trained on. Using these deep learning models on unusual data, e.g. camera sensors with frequency ranges that differ from standard cameras, errors can get difficult to debug.
The developer has hardly any leeway to adapt the pre-trained models to his case without great effort.
Moreover, many of these extensive modules and models represent a computational cost overkill for practical applications running on microprocessors, for example. For these and other reasons, a small survey on hand detection and RoI extraction is given, focusing on deep learning alternatives with low computational cost and no previous training involved. The procedure is supported and simplified by including 3D information to the 2D images.