notes state that tfjs graph model is just converted from tf model, but model signatures,
internal operations and sizes (when manually converted and equally quantized) do not match?
what is the correct model input size?
model input size is marked as dynamic shape: [1, -1, -1, 3],
but models model names indicate 320 for ssd320fpnlite and 640 for ssd640fpnlite ?
and notes further confuse the issue since they state that both models are trained on inputs 450 * 380 ???
those ops are not supported by tfjs, but it seems it doesn't impact model execution at all (defined but unused?)
and one request for enhancement - predict hand rotation
my use case is so the hand detection model can be comined with other models
(e.g. detailed finger tracking, gesture analysis, sign language recognition, etc.)
the thing is, all those models are trained on vertially oriented hand
if hand detection model returned approximate rotation angle, then image could be
rotated and cropped before being used for further analysis
i just stumbled on your model and so far i really like it
i do have few questions:
what is the relationship of two different models?
notes state that tfjs graph model is just converted from tf model, but model signatures,
internal operations and sizes (when manually converted and equally quantized) do not match?
what is the correct model input size?
model input size is marked as dynamic
shape: [1, -1, -1, 3]
,but models model names indicate
320
forssd320fpnlite
and640
forssd640fpnlite
?and notes further confuse the issue since they state that both models are trained on inputs
450 * 380
???tfjs graph model has unknown operations
those ops are not supported by tfjs, but it seems it doesn't impact model execution at all
(defined but unused?)
and one request for enhancement - predict hand rotation
my use case is so the hand detection model can be comined with other models
(e.g. detailed finger tracking, gesture analysis, sign language recognition, etc.)
the thing is, all those models are trained on vertially oriented hand
if hand detection model returned approximate rotation angle, then image could be
rotated and cropped before being used for further analysis