Experiment: curb-ramp detection performance vs. input resolution (Mapillary high-res)

Motivation

RampNet runs at a fixed 2048×4096 input. Mapillary's high-resolution 360 coverage gives us imagery well above that — in the Richmond deployment, 72% of panos are captured above 4096 px wide (median native 11000×5500; a dominant 11000×5500 rig = 4,809 panos, plus 891 at 12288×6144), all downsampled before inference. The hypothesis: more input resolution improves detection, especially recall on small / distant ramps — and it would help most exactly where deployments hurt (lower-quality consumer 360s), which ties into #20's cross-domain concern.

Experiment

Two arms, both scored with rampnet.validation (PR #23) on the HF benchmark's high-res Mapillary slice (#21):

• Frozen model, input-size sweep. The keypoint/heatmap head is fully-convolutional, so it accepts larger inputs — but it was trained at 2048×4096, so ramp scale vs. receptive field shifts. A data point, not proof: does the frozen model tolerate / benefit from more input res?
• Retrain at higher input resolution. The honest test, via the #20 harness. Plot P/R (and AP) as a function of training/eval input size against the committed v1.0-iccv2025 numbers.

Notes

• Report on the cross-domain (Mapillary) slice specifically — that's where the extra native resolution exists; GSV z3 is natively 4096×2048, no headroom.
• If it pans out, the production consequence — fetch native-res + a configurable inference input size to produce the best labels — is tracked in the deployment repo (ProjectSidewalk/sidewalk-auto-labeler#16).

Related

• Model/backbone harness: #20
• Scorer: PR #23 (#22)
• Benchmark slice: #21
• Production residue: ProjectSidewalk/sidewalk-auto-labeler#16

Start with rampnet.validation from PR #23, the HF benchmark's high-resolution Mapillary slice in #21, and the model/backbone harness in #20. Evaluate frozen and retrained models across input sizes, then report precision, recall, and AP for the Mapillary cross-domain slice against the committed v1.0-iccv2025 numbers.