Cross-View Image-to-Image Translation

The Dayton dataset is a dataset for ground-to-aerial (or aerial-to-ground) image translation, or cross-view image synthesis. It contains images of road views and aerial views of roads. There are 76,048 images in total and the train/test split is 55,000/21,048. The images in the original dataset have 354×354 resolution. Source: Multi-Channel Attention Selection GANs for Guided Image-to-Image Translation Image Source: https://arxiv.org/abs/1912.06112

Contains annotated egocentric and top-view videos.

The dataset contains two subsets of synthetic, semantically segmented road-scene images, which have been created for developing and applying the methodology described in the paper "A Sim2Real Deep Learning Approach for the Transformation of Images from Multiple Vehicle-Mounted Cameras to a Semantically Segmented Image in Bird’s Eye View" (IEEE Xplore, arXiv, YouTube) The dataset can be used through the official code implementation of the Cam2BEV methodology described on Github. | Dataset | # Training Samples | # Validation Samples | # Vehicle Cameras | # Semantic Classes | Contained Images (examples) | | --- | --- | --- | --- | --- | --- | | Dataset 1: 360° Surround | 33199 | 3731 | 4 (front, rear, left, right) | 30 (CityScapes) | front camera, rear camera, left camera, right camera, bird's eye view, [bird's eye view incl. occlusion](https://gitlab.ika.rwth-aachen.de/cam2bev/cam2bev-data/-/raw/master/1_FRLR/examples bev+occlusion.png), homography view | | Dataset 2: Front Camera only | 32246 | 3172 | 1 (front) | 30 (CityScapes) | front camera, bird's eye view, bird's eye view incl. occlusion, homography view |

This dataset builds upon the SpaGBOL dataset - a graph-based dataset covering numerous cities across the globe for the purpose of structured city-scale Cross-View Geo-Localisation (CVGL). We introduce the idea of Multi-View Geo-Localisation (MVGL) - localising in two steps within a graph. First using CVGL to retrieve the most likely primary nodes (road junctions), before performing relative pose estimation also candidate edges to increase precision. The corresponding initial network proposed reduced median error from hundreds of metres down to approximately 22m error across the entire dataset.