Adopt HSLVision (first public HSL detection benchmark) — dataset integration + 512-res retrain

[CodeByDamianK]

Summary

Whirlwind Amsterdam (ex-Dutch Nao Team, UvA) released HSLVision — the first public object
detection dataset/benchmark for the unified Humanoid Soccer League:
https://huggingface.co/datasets/whirlwind-ams/hslvision (CC BY 4.0) ·
https://intelligentroboticslab.github.io/hslvision/
6,068 images / 65,309 boxes across 6 venues (LAB42, German Open 2026 Cologne, Beijing "Ice
Ribbon" Masters, Salvador, Sydney, Nagoya), Booster K1+T1 + other KidSize robots, splits
4,247/910/911.

Why it matters — the baseline collapses on it

Evaluating our NAO-only RF-DETR baseline on hslvision-valid (910 imgs, first real in-domain
labeled benchmark we've ever had):

benchmark	overall mAP@50:95	Ball	Robot	GoalPost	LSpot	TSpot	XSpot
NAO val (in-domain)	0.542	0.848	0.819	0.679	0.238	0.202	0.274
hslvision valid (HSL)	0.133	0.538	0.194	0.000	0.000	0.000	0.001

Ball transfers (consistent with the bake-off), Robot collapses (NAO→Booster appearance shift),
GoalPost/junctions hit ~0 — note the junction zeros are partly a labeling-convention mismatch
(we predict ~10px point-boxes; hslvision GT uses ~35px junction-arm boxes), so they overstate
pure blindness — but "useless under the benchmark's convention" is the operative reality.

Done

• Downloaded (6.19 GB) + converted to our COCO layout: parquet → data/hslvision_coco/,
  class remap onto our 7-class taxonomy (their K1 class → Robot; verified visually it
  covers ALL humanoid robot types). Conversion is lossless: 6,068 imgs / 65,309 anns exact.
• Audit: per-class counts (train: 12.5k LSpot, 9.1k TSpot, 3.5k XSpot, 10.5k Robot, 2.3k
  Ball…), box-size stats, per-class crop montages + overlays (label quality credible,
  GoalPost = per-post boxes like ours), venue/filename mapping, htwk_T1 leakage check
  (0 filename collisions; content overlap for Salvador/Beijing footage still flagged).
• Config fixes unblocked by this: stale %56 divisibility validator → %32; resolution
  448 → 512 (RFDETRSmall native); fixed the dead-code augmentation bug (YAML aug block
  was parsed but never forwarded to model.train(aug_config=...)).
• Reusable eval_coco.py per-class benchmark harness (self-test gate: reproduces the known
  NAO plateau 0.542 before any new number is trusted).
• Run A kicked off: fine-tune from the NAO best checkpoint @512 on coco_mix_v1 =
  NAO train (13,106) + hslvision train ×3 oversample (12,741 entries), 183k annotations;
  early stopping tracks hslvision-valid (deployment domain). hslvision-test (911) stays
  untouched as the final holdout.

Next

• Run A results: per-class AP on hslvision-valid + NAO-val regression check (don't lose the
  ball/robot strength) → post table here.
• Run B ablation: hslvision-only fine-tune (quantifies whether NAO data helps or hurts).
• Consider Person class (hslvision has none; its unlabeled humans-as-background already
  suppress the human-as-robot FPs from the bake-off).

Attribution: Xavier Catarrinho, D.M. et al., "HSLVision: A Multimodal Vision Dataset for
RoboCup Humanoid Soccer", RoboCup 2026 Symposium. CC BY 4.0.