RampNet: Automated Curb Ramp Detection

• The paper introduces a two-stage pipeline that uses government geolocation metadata to auto-generate dense curb ramp annotations in Google Street View images.
• It employs a geo-to-pixel translation method and a modified ConvNeXt V2 detector, achieving an average precision of 0.9236 on manually labeled benchmarks.
• RampNet enables large-scale urban accessibility auditing by transforming sparse municipal data into accurate image-based detections without reliance on geo-metadata at inference.

RampNet is a two-stage pipeline for bootstrapping curb ramp detection in Google Street View (GSV) images from open government metadata. It couples automatic dataset generation from government-provided curb ramp locations with a learned detector trained on the resulting annotations. In the reported formulation, Stage 1 translates curb ramp latitude–longitude records into pixel coordinates in panoramic imagery, and Stage 2 trains a modified ConvNeXt V2 detector that predicts curb ramp locations from images alone, without requiring geo-metadata at inference. The system is presented as a response to the scarcity of large-scale, high-quality curb ramp datasets and is evaluated on both auto-generated and manually labeled benchmarks, with the reported detector reaching an Average Precision (AP) of 0.9236 (O'Meara et al., 13 Aug 2025).

1. Problem setting and motivation

Curb ramps are treated as critical urban accessibility infrastructure, but robust image-based detection is described as an open problem because large-scale, high-quality datasets have been lacking. Earlier efforts based on crowdsourced or manually labeled data are characterized as limited in either quality or scale. RampNet is positioned specifically to address that data bottleneck by using open government curb ramp location datasets as supervision for large-scale image annotation (O'Meara et al., 13 Aug 2025).

The pipeline is organized around a division of labor between annotation generation and detector learning. Stage 1 uses government metadata to create pixel-level labels in GSV panoramas. Stage 2 then uses those labels to train a model that no longer depends on government metadata at test time. This separation is central to the design: the metadata is used to bootstrap supervision, while the deployed detector operates on imagery alone.

The motivating application domain is city-scale accessibility auditing. Because curb ramp inventories are often published by municipalities, the method exploits an already existing public-data layer rather than requiring exhaustive manual annotation of street imagery. A plausible implication is that RampNet belongs to a broader class of weakly supervised or automatically supervised urban-computer-vision pipelines, but its specific contribution is the use of curb ramp geolocation metadata to produce near-exhaustive point annotations in panoramas.

2. Stage 1: dataset generation from government metadata

Stage 1 begins with government curb ramp location datasets from New York City, Portland, Oregon, and Bend, Oregon, selected on the basis of visual or manual checks for geolocation precision. Each source dataset is described as a table containing a unique identifier for each curb ramp together with a <lat, long> tuple. For each curb ramp location, all GSV panoramas within a 10-meter radius are downloaded at resolution 4096×2048 pixels. Each panorama is then considered for labeling all curb ramps within a 35-meter radius, and only ramps whose installation predates the panorama capture date are retained (O'Meara et al., 13 Aug 2025).

Negative examples are incorporated explicitly. Twenty percent of the panoramas are “null images,” defined as panoramas with no ramps visible and sampled at least 60 meters away from any known ramp. After this initial selection phase, the dataset contains 219,170 panoramas and 959,442 curb ramp label candidates.

After the auto-translation stage, the resulting dataset contains 214,376 fully labeled panoramas and 849,895 curb ramp labels. The paper describes this as the largest known curb ramp dataset to date, and reports that an average ramp appears in approximately 4.5 panoramas. The dataset is intended to be near-exhaustive at the image level, in contrast to sparser labeling regimes.

This construction strategy is significant because it transforms sparse municipal point records into dense visual supervision over many panoramas. The design also exploits the redundancy of street-level imagery: the same physical ramp can appear in multiple panoramas, which improves coverage and provides multiple visual contexts for the same infrastructure instance.

3. Geo-to-pixel translation and crop-level localization

The central technical problem in Stage 1 is converting a curb ramp’s geographic coordinates into a pixel coordinate in a panorama. For each curb ramp label candidate in a panorama, the angle from the panorama’s GPS position to the curb ramp’s latitude–longitude is computed. A directional crop is then extracted from the panorama along that angle. This crop is a 1024×1024 pixel perspective crop with a 90° field of view and a 30° downward pitch; only the center third, measuring 341×1024 pixels, is retained for focus (O'Meara et al., 13 Aug 2025).

Automated localization within the crop is performed with a ConvNeXt V2 (base) model trained to output a heatmap of likely curb ramp positions. That model is first pre-trained on crops derived from 20,698 Project Sidewalk panoramas and then fine-tuned on 312 manually labeled crops to improve accuracy and account for sparse or missing labels in Project Sidewalk. The reported crop-level detection performance is 89.0% recall and 87.0% precision.

The detected curb ramp center in the crop is subsequently projected back to an $(x,y)$ pixel coordinate in the original 4096×2048 equirectangular panorama. This completes the metadata-to-pixel translation process and yields point annotations suitable for detector training.

Methodologically, this stage is not a direct geometric projection from map coordinates to image coordinates. Instead, it is a hybrid procedure: coarse geometric alignment narrows the search region, and a learned crop-level localizer resolves the final image-space position. This suggests that the authors treat panorama pose and curb ramp metadata as sufficiently informative to restrict search, but not sufficiently precise to eliminate the need for visual localization.

4. Stage 2: panorama-level curb ramp detection model

Stage 2 trains a curb ramp detector directly on the generated panorama annotations. The model uses a ConvNeXt V2 (base) backbone pretrained on ImageNet-1k. Its head is replaced with a sequence consisting of a 3×3 convolution, ReLU, bilinear upsampling, and a 1×1 convolution, producing a single-channel 1024×512 heatmap for each full 4096×2048 panorama (O'Meara et al., 13 Aug 2025).

Supervision is formulated as heatmap regression. Each curb ramp label is represented as a 2D Gaussian centered at the point annotation with $\sigma = 10.0$, following a pose-estimation-style target representation. At inference, peaks in the predicted heatmap above a threshold of 0.55 are taken as curb ramp detections.

The dataset is split into training, validation, and test partitions in a 70%/20%/10% ratio, with the additional constraint that no curb ramp appears in more than one split; panoramas within 60 meters are grouped to enforce this separation. The resulting splits are 150,063 training panoramas, 42,875 validation panoramas, and 21,438 test panoramas.

Training uses random horizontal flips as augmentation, pixel-wise mean squared error between predicted and target heatmaps as the loss, and Adam with learning rate $1 \times 10^{-5}$. The reported hardware configuration is 16 NVIDIA L40s GPUs with batch size 1 due to VRAM limits, and training is run for a single epoch because of computational burden and dataset size.

A defining property of Stage 2 is that the final detector uses only image input. The government metadata that enabled supervision in Stage 1 is discarded at inference time. This makes the learned detector operationally distinct from the bootstrapping mechanism that produced its labels.

5. Evaluation protocol and reported performance

Evaluation is carried out against manually labeled panoramas as well as the auto-generated test split. For the manual benchmark, 1,000 panoramas from the test set are manually labeled, yielding 3,919 curb ramp points. A prediction is counted as correct if it lies within an 88-pixel radius of a ground-truth point; when multiple predictions match the same ground-truth point, only one is counted as a true positive (O'Meara et al., 13 Aug 2025).

The reported quantitative results are as follows:

Component	Dataset / benchmark	Reported result
Stage 1 auto-translation	Manual benchmark	94.0% precision, 92.5% recall, F1-score 0.932
Stage 2 detector	Manual benchmark	AP 0.9236; precision 93.9% at 92.5% recall
Stage 2 detector	Auto-generated test split	AP 0.873
Prior work (Weld et al., 2019)	Project Sidewalk-based setting	AP 0.3803

The detector is described as state-of-the-art on the manually labeled benchmark. The comparison to prior work is particularly notable because the earlier model is summarized as a ResNet using image, depth, and geo-data, whereas RampNet is reported to use image input only at inference while achieving AP 0.9236 versus AP 0.3803.

The relation between Stage 1 and Stage 2 performance is also important. The paper states that Stage 2 essentially matches the auto-labeling performance of Stage 1, indicating little error introduced through model learning. In other words, the automatically generated labels appear sufficiently accurate to support a detector whose real-world performance is close to the quality of the label-generation pipeline itself.

Error analysis attributes most discrepancies in Stage 1 labeling to visual ambiguity, occlusions, disagreement with government data, driveways, and adjacent ramps. This is relevant because it locates residual error partly in annotation ontology and scene ambiguity, not only in model failure.

6. Contributions, applications, and relation to similarly named systems

RampNet is presented as contributing the first large-scale, high-quality curb ramp detection dataset, benchmark, and model (O'Meara et al., 13 Aug 2025). The reported outputs include more than 210,000 panoramas and more than 840,000 curb ramp labels, together with code, data, and benchmarks released for reuse. The intended downstream use is large-scale accessibility assessment in urban environments.

The paper emphasizes that the detector requires only image input and can be applied to urban scenes with GSV data. This makes it suitable for city-scale, low-cost, and repeatable curb ramp audits for accessibility studies and municipal planning. The stated use case is not merely recognition for its own sake, but operational auditing of infrastructure relevant to accessibility law, compliance, and advocacy.

Within the broader literature, the name “RampNet” should be distinguished from similarly named systems that address unrelated domains. “RAPNet: Accelerating Algebraic Multigrid with Learned Sparse Corrections” concerns learned sparse corrections for algebraic multigrid solvers in scientific computing (Fink et al., 26 May 2026), while “RAMP-CNN: A Novel Neural Network for Enhanced Automotive Radar Object Recognition” concerns automotive millimeter-wave radar object recognition from range-velocity-angle heatmap sequences (Gao et al., 2020). These works share lexical similarity in naming but not problem formulation, data modality, or application area.

In that sense, RampNet occupies a specific niche at the intersection of urban accessibility, street-level imagery, and automated supervision from civic open data. Its distinctive feature is the end-to-end coupling of metadata-based label bootstrapping with panorama-level curb ramp detection, rather than a new sensing modality or a general-purpose detection architecture.