[ISSUE]: Several issues

[haavarpb]

Description

Hi,

I am a PhD student at NMBU, trying to use your repository for a baseline comparison with other segmentation methods. I have several issues:

• Pretrained models don't have any accompanying instructions on configuration. Currently, you set the num_classes from what is set under training.num_classes. This, I expect, would change based on the model used.

  aoc_fruit_detector/scripts/detectron_predictor/detectron_predictor.py

  Lines 78 to 79 in 6f50a77

  	cfg.MODEL.WEIGHTS = os.path.join(self.model_file)
  	cfg.MODEL.ROI_HEADS.NUM_CLASSES = self.num_classes

  
  Loading a different model e.g. aoc_strawberry_class_fruit.pth results in misconfigured model.

• Missing metadata catalogs for other models.
  The accompanying metadata pickle file contain two classes, ripe and unripe. So I guess this metadata file is only valid for the model for ripeness. Where are the ones for the other models? I feel there is missing description for each available model and how they are trained.

• It is not described anywhere what backbone should/could be used. Currently you have:
  

  aoc_fruit_detector/config/non_ros_params.yaml

  Line 18 in 6f50a77

  config_file: 'COCO-InstanceSegmentation/mask_rcnn_R_101_FPN_3x.yaml'

Does this insinuate freedom to choose otherwise, or should this be static?

• Broken LearnerUtils.download()
  Downloading datasets and models using LearnerUtils is not possible without configuring requests to not verify SSL certificates (verify=False)
  

  aoc_fruit_detector/scripts/utils/utils.py

  Line 78 in 6f50a77

  with requests.get(url, stream=True) as r:

Otherwise, thanks for putting this repo out there!

[tiru-venkatesh]

Hi @haavarpb, thanks for the detailed report — this is very helpful.

I’ve been working with the repo recently and can confirm several of these
points are currently under-documented or easy to misconfigure.

I’d like to help address this incrementally by:

• Clarifying pretrained model usage, num_classes, and metadata in the README
• Adding safer checks or clearer errors around model configuration
• Looking into a small, optional fix for the SSL verification issue

I’ll start with documentation + minimal fixes and share a PR soon.