[Decoder/BoundingBox] Add option to specify TF-SSD output tensors mapping
Recent tensor flow object detections neural network models use output
tensor mapping defined by google:
https://www.tensorflow.org/lite/models/object_detection/overview#output_signature
which is different than the one currently used on bounding boxes tensor
decoder element.
This patch adds the possibility to specify any output tensors mapping
through the option-3 parameter for tf-ssd mode using the following
convention:
locations index:classes index: scores index: number of detections index.
For example, option3=0:1:2:3 would map the output tensors as defined by
previous link, .i.e. :
- Bounding boxes locations tensor on index 0
- Classes tensor on index 1
- Scores tensor on index 2
- Number of detection tensor on index 3
option3 for tf-ssd mode is optional, and if not specified, the default
tensors mapping preserve the backward compatibility (.i.e. 3:1:2:0)
Signed-off-by: Xavier Roumegue <xavier.roumegue@nxp.com>