PoseNet-CoreML icon indicating copy to clipboard operation
PoseNet-CoreML copied to clipboard

Enhance detection accuracy

Open pabloroca opened this issue 7 years ago • 2 comments

I am testing this lib, and with my iPhone 6S, I get really poor results.

I don't know what settings I can change to enhance accuracy (even at the risk of a less performance). Can you explain it?

Ideally, i would like to use the front camera.

pabloroca avatar Aug 29 '18 05:08 pabloroca

We have confirmed that coreml can not obtain the same result as TensorFlowJS which became the base. I can not identify the cause, but I believe there is a problem that the same result can not be obtained for coreml. #2

Also, the code here is an implementation of MultiPose. Depending on the application, accuracy may be improved by using singlepose implementation.

mbotsu avatar Aug 29 '18 06:08 mbotsu

You could try changing the outputStride parameter defined in posenetHandler.decodeMultiplePoses( ... ) to 8 instead of the default 16.

velociraptor111 avatar Aug 03 '19 19:08 velociraptor111