3D Daisy model classification demo
Use mouse to control 3D daisy model (mouse wheel will zoom it).
Press X,Y,Z buttons to get rotation scans around corresponding axes.
To test translation/scaling invariance just Copy/Paste images from the
3D procedural daisy editor into
the CNN translation/scaling invariance demo.
You can set new daisy age [0 - 1].
CNNs trained on ImageNet-2012 dataset do not recognize procedural roses
and tulips (they like daisy only). We can fine tune them to classify flowers next.
Blind Zone (BZ) effect
Use X,Y- scans to see BZ in the daisy side view.
daisy X-scan with BZ.
See console for X,Y-scans integration rezults (IntX, IntY, Int = (IntX + IntY)/2).
Axial daisy symmetry is assumed. φ angle integration with weight Sin(φ)dφ
should be made from the daisy top view.
Model Pmax IntX IntY Int
__________________________________
v2_1.00 0.94 0.53 0.64 0.58
v2_0.35 0.78 0.37 0.32 0.35
ResNet 1.00 0.89 0.78 0.83
Averaging in the "Daisy" Dataset test
corresponds to Monte Carlo integration and correlates well with this data.
Questions remain about the BZ causes.
E.g. why new daisy age doesn't help me to improve predictions.
TFjs notes
updated 6 Dec 2019