Aerial Image (Roads & Sidewalks/ Crosswalks)

Citation Author(s):
Rafi Ibn
Wayne State University
Submitted by:
Rafi Ibn Sultan
Last updated:
Tue, 02/06/2024 - 12:21
Data Format:
0 ratings - Please login to submit your rating.


This dataset is dedicated to the work " GeoSAM: Enhancing SAM with Sparse and Dense Visual Prompting for Automated Mobility Infrastructure Segmentation
", we leverage publicly available high-resolution orthorectified imagery and GIS data (contained in Shapefiles) to create this dataset. In order to train and evaluate the GeoSAM model, aimed at improving the segmentation of mobility infrastructure in urban environments, this dataset is used. Orthorectified images, provided by the U.S. Geological Survey, serve as a precise and corrected form of aerial imagery, which, along with GIS data, are used to create accurate annotation labels of two classes: road infrastructures and pedestrian (sidewalk/ crosswalk) infrastructures. Our methodology uses this dataset across two distinct regions within Washington DC for model training and evaluation. Additionally, to assess the model's generalization capabilities, we introduce image tiles from Cambridge—a city not seen during training—thereby testing GeoSAM's ability to adapt and perform in new, unseen environments. The images prepared for GeoSAM are of (1024x1024) resolution, with a zoom level of 20, assembled from adjacent tiles to meet the required dimensions. This approach underscores the robustness and wide applicability of GeoSAM across various geographical locations, demonstrating its potential to enhance the accuracy of urban infrastructure mapping through advanced image processing techniques. 


Three datasets: D_train, D_test, D_gen. These datasets have been created for three different purposes: training, testing, and evaluating the generalizability of the model respectively. All datasets have three different folders: images, ground truths, and masks. Images are aerial images, ground truths have two classes: roads and sidewalks/ crosswalks. Masks folder are the pseudo labels used for prompts generation in this project. More details about the datasets are in the readme file.