As we are in the process of integrating this cool technology into our commercial software products for offering a new input possibility and productivity improvement. Accuracy and reliability of the SDK processed data has been a big concern during the prototyping stage, as the applications are used by thousands of serious professionals in the industry.
Here are some analysed and observed results we made from the latest gold release SDK, which could be the rooms for improvement in the upcoming releases:
1. Often, the head could be mistakenly interpreted as a hand even though there is at least 8-10 cm distance from the moving hand(s). This is particuarly serious when an initially detected hand suddenly moved out of the FOV, which would result the hand skeleton of the previous hand landed suddenly on the face area instead...
2. The GeoNode of a hand would only be detected if a hand is at least 40cm from the camera. Any distance beyond that would not be recognized even though the blob image clearly recognized and render the hand(s). One would expect that if the blob image can determine the hands well, the geoNode counterpart should perform the same too.
3. Regardless with which single finger one is pointing upwards with a closed palm, they are all recognized as a "Thumb's up" pose.
4. The GeoNode data (Hand skeleton nodes) could flicker a lot and gets noisy when the hand is not aligned perpendicularly to the camera, often this happen with a slight hand turning or finger folding, etc
We believe that some of these issues are known to the SDK development team and we are pretty sure that the team has been working very hard to improve the algorithms from time to time. Just in case if some of these are not known yet, we hope that it could help ;-)
Looking forward to the upcoming much improved SDK ;-) Keep up the good work!