RampNet
RampNet (ICCV 2025): model, dataset, Stage 1 inputs. Note: the rampnet-benchmark is later 2026 work demonstrating greater generalizability.
-
Keypoint Detection • 90.1M • Updated • 75 • 1 -
projectsidewalk/rampnet-dataset
Viewer • Updated • 214k • 234 • 5 -
projectsidewalk/rampnet-crop-model-dataset-round2
Viewer • Updated • 1.21k • 363 • 1
projectsidewalk/rampnet-crop-model
UpdatedNote Paper-era artifact. Stage 1 crop model, rounds 1 and 2: converts government curb ramp GPS coordinates into pixel keypoints. Stage 1 cannot be reproduced without round 2.
projectsidewalk/rampnet-stage1-inputs
Viewer • Updated • 526k • 79Note Paper-era artifact. The government inventories the paper actually ran on, the street centrelines, and the manifests recording exactly which panoramas Stage 1 consumed.
projectsidewalk/rampnet-benchmark
Viewer • Updated • 3.64k • 51Note POST-PUBLICATION, not part of the paper. 9 city splits built in 2026 to test the published model out of domain (non-US cities, Mapillary imagery) and against VLM detectors. The paper evaluated on a 1,000-panorama gold set from its three training cities, so numbers measured here are not comparable with the published ones.
projectsidewalk/rampnet-crop-model-dataset-round1
Viewer • Updated • 27.7k • 50Note Round-1 crop-model training set: 27,704 Project Sidewalk crops, 13.37 GB Parquet, filename labels made a real column