Abstract
Weakly supervised object detection (WSOD) in remote sensing images (RSI) only require image-level labels to detect various objects. Most of the WSOD methods incline to capture the most discriminative parts of object rather than the entire object, and the number of easy and hard samples is imbalanced. To address the first problem, a novel metric named objectness score (OS) is proposed and incorporated into the training loss of our WSOD model. The OS is consisted of the traditional class confidence score (CCS) and the object completeness prior score (OCPS). The CCS can provide the probability that a proposal belongs to a certain class, and the OCPS can quantify the completeness that a proposal covers the entire object. Therefore, the samples which cover the entire object with high class confidences will be assigned large weight in the training loss through OS. To handle the second problem, a novel metric named difficulty evaluation score (DES) is proposed and also incorporated into the training loss. The DES is calculated by using the entropy of confidence score vector of each proposal and is used to quantify how difficult a proposal can be identified correctly, consequently, the hard samples will also be assigned large weight in the training loss through DES. The ablation experiments on two RSI datasets verify the effectiveness of the proposed OS and DES. The comprehensive quantitative and subjective evaluations demonstrate that our method inclines to detect the entire object accurately, and surpasses seven state-of-the-art WSOD methods.
Highlights
O BJECT detection is a key task for the understanding of remote sensing images (RSI), and has been widely applied in various practical tasks [1]–[10].Thanks to the development of deep learning [11]–[18], the performance of object detection in RSI has made a significant improvement [19], [20], most of the existing object detection methods in RSI are fully supervised [21]–[23], which require the class and bounding box annotations of objects, apparently, Corresponding authors: Gong Cheng and Wei Wang.acquiring these annotations is expensive and time-consuming
The objectness score (OS) is consisted of traditional confidence score (CCS) and object completeness prior score (OCPS), where OCPS is used to quantify the completeness that a proposal covers the entire object, the OS can assign large weight to the samples which can cover entire object by introducing the OCPS
The OS is consisted of traditional CCS and OCPS, and is incorporated into the training loss of our weakly supervised object detection (WSOD) model, where OCPS is used to quantify the completeness that each proposal covers the entire object
Summary
O BJECT detection is a key task for the understanding of remote sensing images (RSI), and has been widely applied in various practical tasks [1]–[10].Thanks to the development of deep learning [11]–[18], the performance of object detection in RSI has made a significant improvement [19], [20], most of the existing object detection methods in RSI are fully supervised [21]–[23], which require the class and bounding box annotations of objects, apparently, Corresponding authors: Gong Cheng and Wei Wang.acquiring these annotations is expensive and time-consuming. To reduce the labeling costs, some weakly supervised object detection (WSOD) methods are proposed, which only require image-level class labels during training. Comparing with the bounding box annotations, the cost of image-level labeling is greatly reduced, the WSOD in RSI has become an important research direction [24]–[26]. The deep feature vectors of proposals were fed into the basic multiple instance detector [31] to train an image-level classifier. Multiple instance classifier refinement (ICR) branches were simultaneously trained, and the pseudo labels of each ICR branch were produced by its upper ICR branch except the first ICR branch of which the pseudo labels were derived from the basic multiple instance detector
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Free) 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call
