Hi, please see the replies to this thread. The script runs on Linux, and only supports simple Haar cascades (the old format) and the output (*.cascade) needs to be copied to flash or SD.
I am equally struggling with implementation of my own HaarCascade
https://docs.opencv.org/3.4/d1/de5/classcv_1_1CascadeClassifier.html
One thing I noticed, is that it is next to impossible to replicate results of:
OpenCV: cascade.xml + CascadeClassifier.detectMultiScale
and
OpenMV: cascade.cascade + image.find_features
Given just the function description, the OpenCV version has more and different parameters to tune.
Are there plans to use the OpenCV implementation of āHaarInferenceā in the near future?
Itās very hard to build a pipeline where you train and test with OpenCV, but cascade format and inference algorithm on OpenMV are different.
Not really, weāre about to release tensor flow support for object detection in a few months. So, HaarCascades need not be used anymore.