Factory Extraction from Satellite Images: Benchmark and Baseline

Abstract

Factory extraction from satellite images is a key step in urban factory planning, and plays a crucial role in ecological protection and land-use optimization. However, factory extraction is greatly underexplored in the existing literature due to the lack of large-scale benchmarks. In this paper, we contribute a challenging benchmark dataset named SFE4395, which consists of 4395 satellite images acquired from Google Earth. The features of SFE4395 include rich multiscale factory instances and a wide variety of factory types, with diverse challenges. To provide a strong baseline for this task, we propose a novel bidirectional feature aggregation and compensation network called BACNet. In particular, we design a bidirectional feature aggregation module to sufficiently integrate multiscale features in a bidirectional manner, which can improve the extraction ability for targets of different sizes. To recover the detailed information lost due to multiple instances of downsampling, we design a feature compensation module. The module adds the detailed information of low-level features to high-level features in a guidance of attention manner. In additional, a point-rendering module is introduced in BACNet to refine results. Experiments using SFE4395 and public datasets demonstrate the effectiveness of the proposed BACNet against state-of-the-art methods.