[Feature Request] Support for parsing non-square custom yolo output

[luxonis]

[Feature Request] Support for parsing non-square custom yolo output

[SzabolcsGergely]

Non-square NN resolutions are supported. There must be some confusion in communication on Discord here; sorry about that.

Examples of a variety of non-square models are used here via the API: https://docs.luxonis.com/en/latest/pages/tutorials/pretrained_openvino/#trying-other-models

For example face-detection-adas-0001 there is 384x672 resolution.

He is referring to parsing for non-square yolo output is not supported on-device.

[OmarALFA]

I built the new library off develop. Can confirm this resolves the issue on my end! Non-square yolov4-tiny NN no longer produces the error!

Note, that you will need to adjust the anchor masks depending on the NN resolution, in the case of 640x352:

detectionNetwork->setAnchorMasks({ {"side40", {0, 1, 2}}, {"side20", {3, 4, 5}} });

For other resolutions, the formula is as follows:

1. width = inputWidth/16
2. width = inputWidth/32

e.g.
640/16 = 40 --> "side40"
640/32 = 20 --> "side20"

This applies to tiny yolo. For full yolo, there will be a third:
3. width = inputWidth/8

These notes were provided by @szabi-luxonis. Thank you!

[OmarALFA]

Non-square NN resolutions are supported. There must be some confusion in communication on Discord here; sorry about that.

Examples of a variety of non-square models are used here via the API: https://docs.luxonis.com/en/latest/pages/tutorials/pretrained_openvino/#trying-other-models

For example face-detection-adas-0001 there is 384x672 resolution.

My apologies for the ambiguity.

I've edited this issue with more details and to reflect @szabi-luxonis's clarification.

[SzabolcsGergely]

Non-square NN resolutions are supported. There must be some confusion in communication on Discord here; sorry about that.
Examples of a variety of non-square models are used here via the API: https://docs.luxonis.com/en/latest/pages/tutorials/pretrained_openvino/#trying-other-models
For example face-detection-adas-0001 there is 384x672 resolution.

My apologies for the ambiguity.

I've edited this issue with more details and to reflect @szabi-luxonis's clarification.

Is it possible to share the model? Would help validating parsing support for non-squared model.
If not, sharing the parsing code on host is equally helpful.

[SzabolcsGergely]

Tested w/ network/dataset provided by @OmarALFA
640x352 YOLO network.

[Luxonis-Brandon]

Non-square NN resolutions are supported. There must be some confusion in communication on Discord here; sorry about that.

Examples of a variety of non-square models are used here via the API:
https://docs.luxonis.com/en/latest/pages/tutorials/pretrained_openvino/#trying-other-models

For example face-detection-adas-0001 there is 384x672 resolution.