question about the training dataset (versions, filtering, and availability)
Hi Evan,
Thanks for sharing this model and the detailed model card. I’m trying to reproduce your results, and I have a few questions about the dataset.
When I combine the two Roboflow Universe sources you linked, I get about 3,606 images (2,406 from nani-tmzf6/package-detection-5ozpr and 1,200 from roboflow-ngkro/package-detection-e1ssd). The model card reports ~2,787, so some images must have been removed.
Could you clarify:
Dataset versions: Which Roboflow version of each dataset did you use, and did it include Roboflow’s preprocessing or augmentation?
Filtering: Which images were removed (duplicates, bad annotations, or certain classes)? The Nani dataset also has no damaged and nodamaged labels, which aren’t in the README’s list. Were those images dropped or remapped?
Label mapping: How exactly were the original labels (Box, Package, Box_broken, Open_package, damaged) mapped to normal and damaged?
Counts: Are the class counts (1,575 normal / 1,212 damaged) images or bounding boxes? The per-class table shows 538 boxes in 499 validation images.
Split: The 20% validation split should be about 557 images, but the evaluation table shows 499. Was a different split used for evaluation?
If possible, could you share the merged and relabeled dataset (for example as a Hugging Face dataset repo, or a Roboflow fork) and the training notebook? That would make the results much easier to reproduce.
Thanks again!