Train a supervised YOLO baseline (YOLO11 / YOLO26) on the RampNet dataset — isolate architecture vs. data

Motivation

The comparison harness shows RampNet ≫ every zero-shot method (OWLv2, Grounding DINO, Molmo, Gemini, Qwen). It does not yet show RampNet > a generic supervised detector trained on the same auto-generated dataset. That's the most obvious reviewer challenge:

Is the ConvNeXt keypoint-heatmap architecture doing the work, or would any supervised detector trained on the RampNet dataset also beat the zero-shot field?

A trained YOLO answers exactly that — an architecture-vs-data ablation, and arguably the single strongest baseline the comparison is currently missing.

What to run

Train YOLO11 (mature, stable) and/or YOLO26 (Ultralytics' Jan-2026 model, NMS-free / DFL-free — gives the "latest YOLO" headline) on dataset/{train,val,test}, evaluate on the benchmark cities, add a compare.py table row. Skip v12/v13 as baselines (documented training-instability caveats).

Design frictions to resolve

• Labels are points, not boxes (per CLAUDE.md). Two paths:

  • (a) synthesize a fixed-size box around each GT point for training, then evaluate by box-center → GT-point at the existing radius 0.022 (simplest, keeps eval identical; box size is a hyperparameter); or
  • (b) YOLO pose/keypoint mode to predict points directly (closest to RampNet, less standard tooling).

  Recommend (a) first.

• Input handling. RampNet ingests full 2048×4096 panos; YOLO wants ~640–1280 tiles. Either reuse the existing reprojection rig (scripts/model_comparison/equirect_tiling.py, keeps it comparable to how VLMs are fed) or train on native pano tiles (comparable to how RampNet is fed). Pick one, hold it fixed, document it — it's a confound.

• Scoring. Reduce YOLO boxes to center points → the existing rampnet/detection_eval.py matcher at 0.022. YOLO emits calibrated scores, so it gets AP / PR curve / --sweep for free like the open detectors. Note the known caveat: RampNet's own AP is truncated (dets extracted at a 0.5 peak, a lower bound not comparable to detectors cached at 0.05), so lead with F1 at operating points and treat AP with that caveat.

Deliverable

A yolo provider path in compare.py, richmond + bend rows, and a short note on the train setup (box size, tiling choice, epochs).

Relates to #20 (harness for newer detection models/backbones). Companion zero-shot experiment: #52.

Start with CLAUDE.md, compare.py, scripts/model_comparison/equirect_tiling.py, and rampnet/detection_eval.py. Resolve the box-versus-keypoint and tiling choices, train YOLO11 or YOLO26 on dataset/{train,val,test}, and reduce detections to points for evaluation at radius 0.022. Done means a yolo provider, richmond and bend rows, and a short note documenting box size, tiling, and epochs.