Support for segmentation networks has been added. You can now segment images using tensorflow into grayscale images that show a heatmap per class you are looking for. The segment() method will return a list of images of these grayscale heat maps.
detect() will then do all the above but internally run find_blobs() on the heat maps to return instead a list of lists, where each sub list is the blobs detected per class.
EdgeImpulse will have support for running segmentation networks thus enabling object detection and localization on Cortex-M processors.
* PYTF now uses the optimal amount of memory for buffers versus all - buffers are placed in SRAM if they fit producing a massive speed boost.
* Custom scaled/offset outputs now work.
* Updated to the latest tensorflow library.
* You have access to all input/output model parameters.
* Person detection is now int8 and blazing fast - 20 FPS on the Arduino Portena.
* Added m55 libs (m0plus libs coming soon once EdgeImpulse adds support for them in the tensorflow make file)
* Classify/Segment/Detect work on all image types directly (JPG/BAYER/YUV/RGB565/GRAYSCALE/BINARY)
* Use a configurable number of PCM buffers in a queue to avoid overflows.
* Add option to configure whether to use pendsv or mp_scheduler for Python callbacks.
* Streaming can be started without a callback and get_buffer() returns a single buffer from the queue.
* Add precomputed LUT for lib OpenPDM stored in flash (saves about 500uS per conversion).
* Overflow detection and configurable abort on overflow.
* Remove fixed DMA channel.