it probably should be blob in the input, however, you wont even get happy with that, since the model expects [NHWC] input, while the blob has it in [NCHW] order.
checked different input shapes:
(128,128,3) // what you do above
(-215:Assertion failed) (int)_numAxes == inputs[0].size() in function 'getMemoryShapes'
(1,3,128,128) // like dnn blob
(-2:Unspecified error) Number of input channels should be multiple of 3 but got 128 in function 'getMemoryShapes'
(1,128,128,3) // like main.py (this should be the correct shape, imo)
(-215:Assertion failed) start <= (int)shape.size() && end <= (int)shape.size() && start <= end in function 'total'
in the end, it’s a long way from tf to opencv. saved_models from tf2 are already not properly supported (you need to make a 1.x graphdef from it first) this might need some (tf) preprocessing, before you can export a onnx.
personally, i’d put this on hold, and look out for an alternative landmark model, - opencv has a lot to offer here, already:
- OpenCV: cv::face::Facemark Class Reference
- openpose also has face landmarks: opencv/samples/dnn/openpose.cpp at 4.x · opencv/opencv · GitHub
- new face detector dnn also has 5 facial marks (same as in mark_detector.py) , good enough for solvePnP() ! opencv/samples/dnn/face_detect.cpp at 4.x · opencv/opencv · GitHub