Age | Commit message (Collapse) | Author |
|
* Remove head-pose.onnx
* Add head-pose-0.3-big-quantized.onnx, which was trained with https://github.com/opentrack/neuralnet-tracker-traincode commit 9891d60
|
|
|
|
|
|
onnxruntime 1.18
|
|
|
|
Adds support for 3x3 covariance matrices for position+size estimation.
My recent (unreleased) models output those instead of only component wise standard deviations.
Overall
Accept 3x3 covariance matrix output for position+size
Old models still worked fine in my testing
Delete commented code which was there previously to support recurrent models. (Won't be used in foreseeable future)
Delete leftover code for experiments with closed eyes detection (Likewise, I have given up on that / Is not really needed any more)
Improve comments and code quality in the internal filter code
|
|
|
|
tracker/nn: Add new models
Co-Authored-By: Stanislaw Halik <sthalik@misaki.pl>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
PR: #1638
|
|
|
|
|
|
|
|
|
|
uncertainty estimation
* Add rudimentary test for two functions .. maybe more in future
* Fix the rotation correction from vertical translation
* Move preview class to new files
* Move neural network model adapters to new files
* Add utility functions for opencv
* Query the model inputs/outputs by name to see what is available
* Supports outputs for standard deviation of the data distribution -
What you get if you let your model output the full parameters of a
gaussian distribution (depending on the inputs) and fit it with
negative log likelihood loss.
* Disabled support for sequence models
* Add support for detection of eye open/close classification.
Scale uncertainty estimate up if eyes closed
* Add a deadzone filter which activates if the model supports uncertainty
quantification. The deadzone scales becomes larger the more uncertain
the model/data are. This is mostly supposed to be useful to suppress
large estimate errors when the user blinks with the eyes
* Fix distance being twice of what it should have been
|
|
Followup from #1562.
|
|
Issue: #1559
|
|
This reverts commit a67e8630caf20e7f48151024e9e68dd9271d75c7.
|
|
|
|
This is useful not just to save on complexity in call sites, but also
because I plan on using the Verdigris library to remove needless
`valueChanged()` and `setValue()` overloads from each `value<t>`
instance.
Also fix a bug in `options/tie.hpp` where `QComboBox::setCurrentIndex`
was erroneously called as `Qt::DirectConnection`.
|
|
|
|
|
|
|
|
* The preview image is now generated with the dimensions of the widget
* The pose visualization is added afterwards adjusted to the preview size
* The fps / inference time readout is moved to the settings dialog
* The actual obtained resolution from the camera is also shown
* Dialog layout is changed
Note: Switching to using underscore to mark class member vars.
It's not consistently applied yet.
|
|
Tooltips only for half of the settings or so. When hovering over the
actual input boxes.
|
|
Regarding tweaks:
* EWA smoothing of head ROI. Smoothing strength is a UI setting.
* Adjustible zoom into the detected face. The predicted ROI is scaled by a factor the user can set. There is a sweet spot somewhere near 1.
* Adjustible number of threads
* The ROI is no longer taken as model output directly. This was actually not needed. Perhaps as auxiliary training objective for the network. But the tracker implementation now just uses a square area around the head center according to the predicted head size.
* Add comment and debug notification on face ROI model output
|
|
|
|
|
|
|
|
- Added support for MJPEG compression for neuralnet tracker
|
|
|
|
Issue: #1394
|
|
|
|
|
|
This is because multiple opentrack modules depend on different opencv
modules and whichever FIND_PACKAGE(OpenCV COMPONENTS ...) gets executed
first, wins. Following FIND_PACKAGE calls will exit and not populate the
component list any further.
This didn't break the build as the build code for tracker/nn wasn't the
one being executed first.
|
|
|
|
|
|
|
|
|
|
|
|
It crashes later though.
|
|
|