The file yoeo.names.yaml is called yoeo_names.yaml in bitbots_vision. We should make the naming scheme consistent.

Context

As discussed in #61 (#61 (comment)), the YOEO dataloader should not convert segmentation images from TORSO-21 format (0, 127, 255, ...) to YOEO format (0,1,2, ...) itself.
This is the task of the TORSO-21 convert-script.
Instead, we should use the correct path yoeo_segmentations.

https://github.com/bit-bots/YOEO/blob/main/yoeo/utils/utils.py#L365

When running the simulation under secondary state = STATE_TIMEOUT, invalid values are encountered in this and similar subtractions.

https://github.com/bit-bots/YOEO/blob/main/yoeo/utils/transforms.py#L83
Here an overflow happens

• YOEO expects segmentation mask images inside a directory called segmentations (see here).
  • Mask are machine-seeable: very dark colors (0, 1, 2, ...)
  • createYoeoLabelsTORSO.py script (before #60) outputs them at yoeo_masks directory (see here).
• TORSO-21 provides segmentation mask images inside a directory called segmentaions
  • Masks are human-seeable: colors are distinct (0, 127, 255, ...)

Expected behavior

The directory names should not collide, i.e., should not be identical.

Current behavior

Converted masks are in yoeo_masks directory.
User has to rename directory and resolve name conflict manually.

Possible Solution

• YOEO's dataloader reads from a differently called directory like yoeo_masks
• ...

https://github.com/bit-bots/YOEO/blob/0aaecc2359c619db8f8666b0f3890044308578ba/yoeo/utils/datasets.py#L143C9-L143C55

Here are some ideas what to do in our CI

• Test for determinism (training produces the same output?)
• Test conversion scripts

Originally posted by @jaagut in #44 (comment)

The output of the create_custom_model highly differs from the current yoeo.cfg.
The script is still identical to the original PyTorchYolo script.
We should update this. In the same breath, we could convert this bash script to a python script and create a custom command (update Readme!)

We might adapt other scaling losses. TBD

Originally posted by @jaagut in #26 (comment)

It would be nice to return the class index of the segmentation as an unit8 for performance reasons. This should lower USB copy times from NCS2 and improve overall performance, as well as simplify the code in the bitbots_vision.

Proposed change

Add an uint8 typecast to the following line.

We also need to adapt the README.md section that describes the expected data structure. Especially this section: https://github.com/bit-bots/YOEO/tree/main#classes
A sample yaml for TORSO could be added.

This should be another PR

Originally posted by @jaagut in #24 (comment)