Deep Multiple Instance Learning for Airplane Detection in High Resolution Imagery

TL;DR

This paper introduces a rotation-and-scale invariant deep multiple instance learning approach for airplane detection in high-resolution aerial images, effectively handling variations in airplane appearance and orientation.

Contribution

The paper presents a novel symmetric line segment proposal generator combined with a deep multiple instance learning framework for improved airplane detection.

Findings

Effective detection on NWPU VHR-10 and DOTA datasets.

Ability to estimate airplane direction from box annotations.

Outperforms existing methods in high-resolution imagery detection.

Abstract

Automatic airplane detection in aerial imagery has a variety of applications. Two of the significant challenges in this task are variations in the scale and direction of the airplanes. To solve these challenges, we present a rotation-and-scale invariant airplane proposal generator. We call this generator symmetric line segments (SLS) that is developed based on the symmetric and regular boundaries of airplanes from the top view. Then, the generated proposals are used to train a deep convolutional neural network for removing non-airplane proposals. Since each airplane can have multiple SLS proposals, where some of them are not in the direction of the fuselage, we collect all proposals corresponding to one ground-truth as a positive bag and the others as the negative instances. To have multiple instance deep learning, we modify the loss function of the network to learn from each positive bag at least one instance as well as all negative instances. Finally, we employ non-maximum suppression to remove duplicate detections. Our experiments on NWPU VHR-10 and DOTA datasets show that our method is a promising approach for automatic airplane detection in very high-resolution images. Moreover, we estimate the direction of the airplanes using box-level annotations as an extra achievement.