Sparse Compatibility
This model is amazing but with a single issue for my use-case. The model only works till the subject is about 2 meters (or less) away from the camera, I have successfully converted the RunBlazeFace.cs file to run with the Long-distance model(.onnx) with some adjustments, but the performance isn't nearly as good. The only option remaining now is using the BlazeFace Sparse model with full range. I am unable to convert the sparse model into an .onnx file. It would be great help if it was provided with this project to use with longer distances.
Hi, do you mean the model doesn't work well for close up faces (when the face is big?)
Thanks, I'll look into the Sparse model.
As a workaround, you could maybe shrink add a border to the image and shrink it. Then the faces will look smaller and the model may work better on that? Unfortunately that would mean running the model twice, once for the normal image and one for the resized image with border.
To get more help in converting models etc. you can try the Sentis forum here: https://discussions.unity.com/c/ai-beta/sentis/10
The opposite is the issue. Blacefacesentis works great in close range. I wanted to ask for help with long-range model compatibility for subjects 3-5 meters away from the camera, which this model fails to detect. I replicated this project with the long-distance face model (by converting the .tflite version .onnx) but the latency is poor on mobile devices. The only option I have now is to use the sparse model as it is also long distance but smaller, which might improve performance on phones. Unfortunately, I was unable to convert the .tflite sparse model into .onnx because sparse models are not supported for conversion by tf2onnx. If you can just add the long-range models in this project, for people to use it would be a great help!
Thank you for replying so quickly