ProjectSidewalk/RampNet

View on GitHub

RampNet is a two-stage pipeline that addresses the scarcity of curb ramp detection datasets by using government location data to automatically generate over 210,000 annotated Google Street View panoramas. This new dataset is then used to train a state-of-the-art curb ramp detection model that significantly outperforms previous efforts.

Stars: 7
Forks: 1
Open beginner issues: 0
Indexed issues: 50
Avg merge: 4d 11h
Merged PRs (30d): 7

Dominant language: Python
License: MIT
Last GitHub push: Sep 17, 2026
Latest indexed: Sep 19, 2026
Contributing guide: No contributing guide
Code of conduct: No code of conduct
Beginner labels: No beginner labels indexed

Beginner friendly Everything

Open All

50 open issues indexed

Re-run and publish the Vistas 1024-parity and A40-control detections (#126, #137) Open

Difficulty 5/5 Over a week Newbie friendliness 35/100

ProjectSidewalk/RampNet#163 · Opened Sep 17, 2026
RampNet 1.0 closeout: merge queue, what to close, what moves to 2.0 Open

documentation

Difficulty 5/5 Over a week Newbie friendliness 25/100

ProjectSidewalk/RampNet#162 · Opened Sep 16, 2026
Add Panoramax as a third imagery domain: Bayonne (first Panoramax PS city) as a benchmark split, Lyon as a training-only city (#158) Open

enhancement

Difficulty 5/5 Over a week Newbie friendliness 35/100

ProjectSidewalk/RampNet#159 · Opened Sep 6, 2026
Make RampNet rig-native for open imagery: retrain on Mapillary/Panoramax hard positives mined from multi-view consensus, precision check first Open

enhancement

Difficulty 5/5 Over a week Newbie friendliness 38/100

ProjectSidewalk/RampNet#158 · 1 comment · Opened Sep 6, 2026
Score the frontier releases: Claude Fable 5/5.1 and GPT-6 Astra (#122 roster refresh) Open

Difficulty 5/5 Over a week Newbie friendliness 45/100

ProjectSidewalk/RampNet#156 · 6 comments · Opened Sep 4, 2026
Recall collapses on flush / minimal-reveal curb ramps: Laurens R 0.390, and the experiment that isolates it Open

Difficulty 4/5 3-5 days Newbie friendliness 45/100

ProjectSidewalk/RampNet#151 · 2 comments · Opened Aug 31, 2026
FAIR release of rampnet-dataset + rampnet-benchmark: DOI, full cards, Croissant 1.1/GeoCroissant/RAI, evaluation protocol, leaderboard (NSF FAIROS) Open

documentation enhancement FAIROS

Difficulty 5/5 Over a week Newbie friendliness 25/100

ProjectSidewalk/RampNet#150 · Opened Aug 29, 2026
Add Laurens, Iowa from Mapillary? Open

Difficulty 3/5 1-2 days Newbie friendliness 48/100

ProjectSidewalk/RampNet#149 · 1 comment · Opened Aug 25, 2026
The #140 seam fix moved three committed YOLO baseline cells, and benchmark_eval/ was never regenerated Open

Difficulty 4/5 3-5 days Newbie friendliness 68/100

ProjectSidewalk/RampNet#148 · 2 comments · Opened Aug 20, 2026
model_comparison.md and curb_ramp_data_sourcing.md are too long to be read whole: split the runbooks out and generate the tables, don't trim prose Open

documentation

Difficulty 4/5 3-5 days Newbie friendliness 48/100

ProjectSidewalk/RampNet#145 · Opened Aug 19, 2026
claude-opus-5 on manual_gold: recommend NOT running it alone, because the top challenger is not on that split either Open

question

Difficulty 4/5 3-5 days Newbie friendliness 48/100

ProjectSidewalk/RampNet#144 · Opened Aug 19, 2026
The 360 seam is unsealed end to end: RampNet finds under half the ramps on it, Stage 1 double-labels them, and the scorer can see neither Open

Difficulty 5/5 Over a week Newbie friendliness 38/100

ProjectSidewalk/RampNet#132 · 7 comments · Opened Aug 18, 2026
Replicate #46 Phase 1 on a second machine: silent_activation.json has one provenance and §0c rests on it Open

Difficulty 4/5 3-5 days Newbie friendliness 58/100

ProjectSidewalk/RampNet#131 · Opened Aug 18, 2026
360 seam, in the scoring code: 11 double-marked ramps in manual_gold, and a matcher that does not wrap Open

Difficulty 4/5 3-5 days Newbie friendliness 48/100

ProjectSidewalk/RampNet#130 · 1 comment · Opened Aug 18, 2026
Carry the review notes and Bend's train-overlap flag into the published benchmark Open

enhancement

Difficulty 4/5 3-5 days Newbie friendliness 48/100

ProjectSidewalk/RampNet#127 · Opened Aug 18, 2026
Mapillary Vistas as a second supervision source: off-the-shelf baseline, and a path to non-ramp label types Open

Difficulty 5/5 Over a week Newbie friendliness 35/100

ProjectSidewalk/RampNet#126 · 6 comments · Opened Aug 17, 2026
Score crop-window rules against the manual_gold boxes: containment, margin, and context ratio (crop-clarity acceptance test) Open

Difficulty 4/5 3-5 days Newbie friendliness 48/100

ProjectSidewalk/RampNet#114 · 8 comments · Opened Aug 14, 2026
RampNet 2.0 should consume the canonical cropper, not its own: Stage 1's crop supervision inherits the tilt error in Project Sidewalk's `pano_y` Open

Difficulty 5/5 Over a week Newbie friendliness 35/100

ProjectSidewalk/RampNet#113 · Opened Aug 6, 2026
The recall-by-distance axis is stretched ~25-30%: recompute it from GSV depth Open

Difficulty 5/5 Over a week Newbie friendliness 48/100

ProjectSidewalk/RampNet#112 · Opened Aug 5, 2026
GSV serves a native per-pano depth map (streetlevel already parses it): an independent anchor for #101 and the visibility test #102 needs — archive it before the endpoint changes Open

Difficulty 5/5 Over a week Newbie friendliness 35/100

ProjectSidewalk/RampNet#111 · Opened Aug 5, 2026
Swap the Stage 2 backbone to DINOv3-ConvNeXt (there is no ConvNeXt V3) - size-matched pretraining comparison, licence-gated Open

enhancement

Difficulty 5/5 Over a week Newbie friendliness 28/100

ProjectSidewalk/RampNet#107 · Opened Aug 4, 2026
Review location precision against the STREET-LEVEL imagery Stage 1 actually consumes, not an aerial proxy (#96) Open

documentation

Difficulty 5/5 Over a week Newbie friendliness 38/100

ProjectSidewalk/RampNet#103 · 2 comments · Opened Aug 3, 2026
Mine hard positives from multi-view consensus: a training-label source that does not depend on city inventories (#96, #59) Open

Difficulty 5/5 Over a week Newbie friendliness 35/100

ProjectSidewalk/RampNet#102 · Opened Aug 3, 2026
Flat-ground distance carries a ~11% range-proportional scale error, not just scatter — multi-view self-consistency measures it and DA3 can anchor it Open

Difficulty 5/5 Over a week Newbie friendliness 38/100

ProjectSidewalk/RampNet#101 · 1 comment · Opened Aug 3, 2026
Assess location precision for the unassessed candidate cities — the gate every route to 500k depends on Open

documentation

Difficulty 4/5 3-5 days Newbie friendliness 45/100

ProjectSidewalk/RampNet#96 · 11 comments · Opened Jul 31, 2026
Attribute any RampNet-vs-YOLO gap: backbone, output representation, and training budget are confounded in #51 Open

Difficulty 5/5 Over a week Newbie friendliness 35/100

ProjectSidewalk/RampNet#95 · Opened Jul 31, 2026
Tune the YOLO baseline's peak LR / warmup before reporting it — the collapse is an untuned default (#51) Open

Difficulty 5/5 Over a week Newbie friendliness 45/100

ProjectSidewalk/RampNet#90 · Opened Jul 30, 2026
RampNet 2.0: measurement, condition inference, and tagging — we already collect the supervision and discard it at ingest Open

Difficulty 4/5 3-5 days Newbie friendliness 52/100

ProjectSidewalk/RampNet#86 · 1 comment · Opened Jul 29, 2026
Aerial orthoimagery curb-ramp detection already exists in production (34k ramps at 1-inch GSD): treat it as an uncorrelated arbiter, not a competitor Open

Difficulty 5/5 Over a week Newbie friendliness 35/100

ProjectSidewalk/RampNet#85 · 1 comment · Opened Jul 29, 2026
Extend RampNet output from points to boxes/polygons (point-prompted SAM2; CenterNet-style size head; segmentation) Open

enhancement

Difficulty 5/5 Over a week Newbie friendliness 45/100

ProjectSidewalk/RampNet#83 · 3 comments · Opened Jul 29, 2026
Training augmentation is a single horizontal flip: measure downscale / photometric jitter as a transfer lever (downstream of #84) Open

Difficulty 5/5 Over a week Newbie friendliness 35/100

ProjectSidewalk/RampNet#82 · Opened Jul 29, 2026
Second rater for Budapest District V: the rubric is now confounded with the benchmark's only ranking inversion Open

Difficulty 4/5 3-5 days Newbie friendliness 58/100

ProjectSidewalk/RampNet#74 · 1 comment · Opened Jul 28, 2026
Document the YOLOv11-pano training instability as a reported caveat Open

documentation enhancement

Difficulty 3/5 1-2 days Newbie friendliness 48/100

ProjectSidewalk/RampNet#72 · 1 comment · Opened Jul 28, 2026
Rerun the collapsed YOLOv11 pano baselines with a stabilized (larger physical batch) config Open

enhancement

Difficulty 4/5 3-5 days Newbie friendliness 52/100

ProjectSidewalk/RampNet#70 · 4 comments · Opened Jul 28, 2026
Would 2-3M training labels buy recall? Measure the scaling curve and the label ceiling before a sourcing campaign Open

enhancement

Difficulty 4/5 3-5 days Newbie friendliness 48/100

ProjectSidewalk/RampNet#59 · 3 comments · Opened Jul 25, 2026
Add the 1k manual gold-standard set as a model-comparison split (bigger, independently-labeled, un-anchored GT) Open

enhancement

Difficulty 4/5 3-5 days Newbie friendliness 48/100

ProjectSidewalk/RampNet#58 · 1 comment · Opened Jul 25, 2026
Add YOLOE (real-time open-vocab) to the zero-shot detector bucket Open

enhancement

Difficulty 3/5 1-2 days Newbie friendliness 72/100

ProjectSidewalk/RampNet#52 · Opened Jul 24, 2026
Train a supervised YOLO baseline (YOLO11 / YOLO26) on the RampNet dataset — isolate architecture vs. data Open

enhancement

Difficulty 5/5 Over a week Newbie friendliness 45/100

ProjectSidewalk/RampNet#51 · 30 comments · Opened Jul 24, 2026
Add Clovis validation split to the HF benchmark dataset Open

Difficulty 4/5 3-5 days Newbie friendliness 55/100

ProjectSidewalk/RampNet#49 · Opened Jul 24, 2026
Multi-capture / multi-view fusion for per-ramp recall: correspondence, agreement-as-arbiter, 3D (RampNet 2.0) Open

enhancement

Difficulty 5/5 Over a week Newbie friendliness 25/100

ProjectSidewalk/RampNet#48 · 1 comment · Opened Jul 24, 2026
Mask the nadir/hood region in the reprojection rig; measure ΔAP Open

enhancement

Difficulty 4/5 3-5 days Newbie friendliness 48/100

ProjectSidewalk/RampNet#47 · Opened Jul 24, 2026
Qualitative failure taxonomy: bucket each model's FP/FN from cached detections Open

enhancement

Difficulty 4/5 3-5 days Newbie friendliness 52/100

ProjectSidewalk/RampNet#46 · 7 comments · Opened Jul 24, 2026
Prompt study for the challenger VLMs: image exemplars, elicited confidence, query ensemble Open

enhancement

Difficulty 5/5 Over a week Newbie friendliness 30/100

ProjectSidewalk/RampNet#45 · Opened Jul 24, 2026
Add clovis as a third benchmark city; pursue source/geographic diversity Open

enhancement

Difficulty 4/5 3-5 days Newbie friendliness 48/100

ProjectSidewalk/RampNet#44 · 1 comment · Opened Jul 24, 2026
360 pano seam: curb ramps split across the left/right edge get double-counted in GT scoring Open

documentation enhancement

Difficulty 4/5 3-5 days Newbie friendliness 48/100

ProjectSidewalk/RampNet#43 · 2 comments · Opened Jul 24, 2026
Benchmark open-weight detectors and pointing models (OWLv2, Grounding DINO, AI2 Molmo) against RampNet Open

Difficulty 4/5 3-5 days Newbie friendliness 48/100

ProjectSidewalk/RampNet#39 · 2 comments · Opened Jul 24, 2026
Measure per-ramp deployment recall; sampling density may be the cheapest recall lever Open

Difficulty 5/5 Over a week Newbie friendliness 38/100

ProjectSidewalk/RampNet#38 · Opened Jul 24, 2026
Experiment: hybrid RampNet + open VLM (Qwen3-VL) — ensemble / cascade / distillation Open

Difficulty 5/5 Over a week Newbie friendliness 42/100

ProjectSidewalk/RampNet#35 · 4 comments · Opened Jul 23, 2026
Experiment: curb-ramp detection performance vs. input resolution (Mapillary high-res) Open

enhancement

Difficulty 4/5 3-5 days Newbie friendliness 45/100

ProjectSidewalk/RampNet#25 · 1 comment · Opened Jul 22, 2026
Set up a harness for experimenting with newer detection models/backbones Open

Difficulty 5/5 Over a week Newbie friendliness 48/100

ProjectSidewalk/RampNet#20 · 6 comments · Opened Jul 21, 2026

ProjectSidewalk/RampNet

Get new issues in your inbox