gesture-recognition

I've provided this answer in the past.

Personally I've used the Xbox360 sensor with OpenNI the most (because it's cross platform). Also the NITE middleware on alongside OpenNI provides some basic hand detection and even gesture detection (swipes, circle gesture, "button" push, etc.).

While OpenNI is opensource, NITE isn't so you'd be limited to what they provide.

The links you've shared use OpenCV. You can install OpenNI and compile OpenCV from source with OpenNI support. Alternatively, you can manually wrap the OpenNI frame data into an OpenCV cv::Mat and carry on with the OpenCV operations from there.

Here's a basic example that uses OpenNI to get the depth data and passes that to OpenCV:

There is nothing wrong with your model, it's just a trivial problem for the models to solve. These letters look nothing alike when you consider all of the features you have. If you had chosen all of the letters, or ones that all looked the same, you might see some error.

Rerun the model using only index_pitch and index_roll. You will still get like 95% AUC. At least by doing that you can guess that the only loss comes from B,D, and K, which by looking at an image of what those look like are the only 3 that could remotely be confused if you only looked at the index finger. This turns out to be the case.

It's just a problem that given your data set is actually solvable