MultiScene A Large-scale Dataset and Benchmark for Multi-scene Recognition in Single Aerial Images
Task
Multi-scene recognition is a challenging task due to that:
● Images are large-scale and unconstrained.
● All present scenes in an aerial image need to be exhaustively recognized.
In this work, we propose a large-scale dataset, namely MultiScene dataset, and provide extensive benchmarks.
Dataset
MultiScene dataset aims at two tasks:
● Developing algorithms for multi-scene recognition
● Network learning with noisy labels.
We collect 100k high-resolution aerial images with the size of 512x512 around the world. All of them are assigned with crowdsourced labels provided by OpenStreetMap (OSM), and 14k of them are mannually inspected yielding a subset of cleanly-annotated images (show in red), named MultiScene-Clean.
Experiment
We evaluated 22 baselines on both MultiScene-Clean and MultiScene datasets and report numerical results (%) in the following tables. Left: MultiScene-Clean. Right: MultiScene
Paper
@article{hua2021multiscene,
title={MultiScene: A Large-scale Dataset and Benchmark for Multi-scene Recognition in Single Aerial Images},
author={Hua, Y. and Mou, L. and Jin, P. and Zhu, X. X.},
journal={IEEE Transactions on Geoscience and Remote Sensing},
year={in press}}