Auditing Three RF-DETR Nano Checkpoints for Volleyball Detection on a Separate Roboflow Export
We audit three pre-existing RF-DETR Nano checkpoints on a separate Roboflow ex-port of 1,246 volleyball frames with 13,369 player and 589 ball annotations. The primary384-pixel protocol uses official COCO metrics, explicit Player output rather than action-labelled people, and paired conditional bootstrap comparisons. Player AP50 is 80.28%,80.91% and 81.26% for Original, Augmentation and BlurNoise; ball AP50 is 52.32%, 52.44%and 50.31%. Both numerical leaders change under one representative-per-group rule, butdiameter-bounded complete linkage preserves the full-collection ordering. Conditional pairedintervals do not distinguish the swapped leaders. Most fixed-threshold ball false positivesoccur on frames without ball annotations, and the largest similarity component containsdifferent opponent pairings. Supplementary full-query person aggregation, a shared 100-frame resolution/antialiasing check, official TIDE diagnostics, conditional calibration andinstrumented CPU measurements expose additional endpoint sensitivity. These checks sup-port a bounded observational audit, not causal augmentation benefits, matched-trainingarchitecture superiority or tablet suitability. Correlated footage, incomplete source/trainingprovenance and unresolved annotation coverage limit generalization.
Authors
- Dennis Koch
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-21
- DOI
- https://doi.org/10.5281/zenodo.24429836
- Primary Topic
- Advanced Neural Network Applications
- Type
- preprint