Can semantic labeling methods generalize to any city? the inria aerial image labeling benchmark

Abstract

New challenges in remote sensing impose the necessity of designing pixel classification methods that, once trained on a certain dataset, generalize to other areas of the earth. This may include regions where the appearance of the same type of objects is significantly different. In the literature it is common to use a single image and split it into training and test sets to train a classifier and assess its performance, respectively. However, this does not prove the generalization capabilities to other inputs. In this paper, we propose an aerial image labeling dataset that covers a wide range of urban settlement appearances, from different geographic locations. Moreover, the cities included in the test set are different from those of the training set. We also experiment with convolutional neural networks on our dataset.

Distance transform regression for spatially-aware deep semantic segmentation

“…Semantic segmentation is a longstanding task in the computer vision community. Several benchmarks have been introduced on many application domains such as COCO Lin et al (2014) and Pascal VOC Everingham et al (2014) for multimedia images, CamVid Brostow et al (2009) and Cityscapes Cordts et al (2016) for autonomous driving, the ISPRS Semantic Labeling Rottensteiner et al (2012) and INRIA Aerial Image Labeling Maggiori et al (2017) datasets for aerial image, and medical datasets Ulman et al (2017), which are now dominated by the deep fully convolutional networks. Many applications rely on a pixel-wise semantic labeling to perform scene understanding, such as objectinstance detection and segmentation ; Arnab and Torr (2017) in multimedia images, segment-beforedetect pipelines for remote sensing data processing Audebert et al (2017); Sommer et al (2017) and segmentation of medical images for neural structure detection and gland segmentation Ronneberger et al (2015); Chen et al (2016a).…”

“…Evaluation is done by splitting the datasets with a 3-fold cross-validation.INRIA Aerial Image Labeling Benchmark. The INRIA Aerial Image Labeling datasetMaggiori et al (2017) is comprised of 360 RGB tiles of 5000 × 5000px with a spatial resolution of 30cm/px on 10 cities across the globe. Half of the cities are used for training and are associated to a public ground truth of building footprints.…”

A Comparison and Strategy of Semantic Segmentation on Remote Sensing Images

“…It is collected from five cities: Austin, Chicago, Kitsap, Vienna and West Tyrol, and each of them has 36 images. As described in [21], the first 5 images of each city are used for testing and the other 155 images for training. These pixel-level labeled images contains two classes: building and non-building.…”

Strategy Synthesis for Surveillance-Evasion Games with Learning-Enabled Visibility Optimization

“…Intuitively, the shadow boundaries (or frontiers) B i are the boundaries between free space and occlusion. In order to train the neural network, we sample different environments L cropped from the INRIA Aerial Image Labeling Dataset [31]. For each L, an initial position l 0 ∈ L ⊆ L for the agent is randomly sampled.…”