ExtendedNeuralNetwork node

[tadeas0]

Purpose

Create ExtendedNeuralNetwork node that wraps ParsingNeuralNetwork and adds the following capabilities:

• Automatic input resizing to the neural network input size
• Remapping of detection coordinates from neural network output to input frame coordinates
• Neural network output filtering based on confidence threshold and labels (Only supported for ImgDetectionsExtended and ImgDetections messages)
• Input tiling

Specification

https://app.clickup.com/t/86c5mz8er

Dependencies & Potential Impact

None / not applicable

Deployment Plan

None / not applicable

Testing & Validation

Manually tested on RVC4. Currently does not work on RVC2 due to missing transformation bindings in DAI Script node.

Example usage

extended_neural_network_example.py

[aljazkonec1]

Took a quick look and left some comments. Mostly looks good, just a couple of minor fixes. Thanks!

Should we add tests aswell?

[aljazkonec1]

General comment for entire PR: Use logger to log the stages of the class. Something like logger.info("Building TilingPipeline") and logger.info("Building Tiles patcher") etc etc.

[kkeroo]

Agree, check the parsers implementation for reference.

[tadeas0]

Should we add tests aswell?

Unit tests would require mocking Script and ImageManip nodes. We can add integration tests, however, I would create a separate PR for those to get this merged ASAP, because it's blocking Stage2NeuralNetwork node PR.

[PetrNovota]

@tadeas0 FYI I am testing the new 1st and 2nd stage nodes in the focused vision PoC and have encountered an issue so we will likely not be merging either PR very quickly. I will let you know what the issue is.

[klemen1999]

Closing as this was re-implemented in #274