Learning Remote Sensing Object Detection With Single Point Supervision

Pointly supervised object detection (PSOD) has attracted considerable interest due to its lower labeling cost when compared to box-level supervised object detection. However, the complex scenes and densely packed and dynamic-scale objects in remote-sensing (RS) images hinder the development of PSOD methods in the RS field. In this article, we make the first attempt to achieve RS object detection with single-point supervision and propose a PSOD method tailored for RS images. Specifically, we design a point label upgrader (PLUG) to generate pseudo-box labels from single-point labels and then use the pseudo-boxes to supervise the optimization of existing detectors. Moreover, to handle the challenge of the densely packed objects in RS images, we propose a sparse feature-guided semantic prediction (SemPred) module that can generate high-quality semantic maps by fully exploiting informative cues from sparse objects. Extensive ablation studies on the DOTA dataset have validated the effectiveness of our method. Our method can achieve significantly better performance when compared to state-of-the-art image-level and point-level supervised detection methods and reduce the performance gap between PSOD and box-level supervised object detection. The code is available at https://github.com/heshitian/PLUG.

I. Introduction

Remote-sensing object detection (RSOD) plays an important role in many fields, such as national defense and security, resource management, and emergency rescuing. With the development of deep learning, many deep neural network (DNN)-based detection methods [1], [2], [3], [4], [5], [6], [7] were proposed and achieved promising performance. Besides, a number of remote-sensing (RS) datasets (e.g., HRSC2016 [8], NWPU VHR-10 [9], and DOTA series [10]) containing accurate and rich annotations were proposed to develop and benchmark RSOD methods. In these datasets, accurate location, scale, category, and quantity information of objects are provided and greatly facilitate the development of RSOD. However, such rich annotation formats will lead to expensive labor costs when RSOD methods are transferred to the new RS data (e.g., images captured by new satellites).