Based on the 3D model of the city, we densely sampled points on facades and performed ray tracing to find all street view panoramas that see the same target point without occlusion. We collected images on a dense grid in the aforementioned cities. The dataset comprises 25 million google street view images forming 118 million corresponding pairs. Generic 3D Representation via Pose Estimation and Matching, Amir R Zamir, Tilman Wekel, Pulkit Agrawal, Jitendra Malik, Silvio Savarese, ECCV16.įor more detail on the data collection and the method please visit the. The datase was used in the following paper to learn a universal/generic 3D representation: The following shows the coverage areas for 6 of the 8 cities. Below you can see sample snapshots of 3D models. The models are geo-registered and manually generated. We are releasing the 3D models of these cities along with the street view images and metadata. The dataset covers the downtown and surrounding areas of New York, Chicago, Washington, Las Vegas, Florence, Amsterdam, San Francisco and Paris. For more information on how the dataset was collected, please see the. The dataset was collected automatically without human annotation by developing a system to intergrate georeferenced 3D models of cities with google street view images and their geo-metadata. You can see a few sample image bundles from the dataset below and more examples here and here. The dataset can be used for learning 6DOF camera pose estimation/visual odometry, image matching, and various 3D estimations. The data comes in bundles of matching images the content of the matching pairs show the same physical point while the camera viewpoint shows a large baseline (often >120 degrees). This repository shares a large dataset of street view images (25 million images and 118 million matching image pairs) with their 6DOF camera pose, 3D models of 8 cities, and extended metadata.

Street View Image, Pose, and 3D Cities Dataset

Download street view map