Abstract:Deep-learning-based object detectors have substantially improved state-of-the-art object detection in remote sensing images in terms of precision and degree of automation. Nevertheless, the large variation of the object scales makes it difficult to achieve high-quality detection across multiresolution remote sensing images, where the quality is defined by the Intersection over Union (IoU) threshold used in training. In addition, the imbalance between the positive and negative samples across multiresolution images worsens the detection precision. Recently, it was found that a Cascade region-based convolutional neural network (R-CNN) can potentially achieve a higher quality of detection by introducing a cascaded three-stage structure using progressively improved IoU thresholds. However, the performance of Cascade R-CNN degraded when the fourth stage was added. We investigated the cause and found that the mismatch between the ROI features and the classifier could be responsible for the degradation of performance. Herein, we propose a Cascade R-CNN++ structure to address this issue and extend the three-stage architecture to multiple stages for general use. Specifically, for cascaded classification, we propose a new ensemble strategy for the classifier and region of interest (RoI) features to improve classification accuracy at inference. In localization, we modified the loss function of the bounding box regressor to obtain higher sensitivity around zero. Experiments on the DOTA dataset demonstrated that Cascade R-CNN++ outperforms Cascade R-CNN in terms of precision and detection quality. We conducted further analysis on multiresolution remote sensing images to verify model transferability across different object scales.

Rotated Cascade R-CNN: A Shape Robust Detector with Coordinate Regression

RINet: Efficient 3D Lidar-Based Place Recognition Using Rotation Invariant Neural Network

CRRCNN: Cascade Rotational RCNN for Dense Arbitrary-Oriented Object Detection

R2CNN: Rotational Region CNN for Orientation Robust Scene Text Detection

SCRDet: Towards More Robust Detection for Small, Cluttered and Rotated Objects

A Three-Stage Cascade Rotating Regression Network for SAR Target Rotation Detection.

Sliding Line Point Regression for Shape Robust Scene Text Detection.

R2CNN++: Multi-Dimensional Attention Based Rotation Invariant Detector with Robust Anchor Strategy.

Point RCNN: An Angle-Free Framework for Rotated Object Detection

R-YOLO: A Real-Time Text Detector for Natural Scenes with Arbitrary Rotation

A Rotated Object Detection Strategy for Remote Sensing Images Using Misaligned Cross-Fusion Structures

Adaptive Rotated Convolution for Rotated Object Detection

Parallel Cascade R-CNN for Object Detection in Remote Sensing Imagery

Learning Modulated Loss for Rotated Object Detection

High Quality Object Detection for Multiresolution Remote Sensing Imagery Using Cascaded Multi-Stage Detectors.

Improved Oriented Object Detection in Remote Sensing Images Based on a Three-Point Regression Method

RSDet++: Point-based Modulated Loss for More Accurate Rotated Object Detection

Effective Rotate: Learning Rotation-Robust Prototype for Aerial Object Detection

Rotation-Sensitive Regression for Oriented Scene Text Detection

R3Det: Refined Single-Stage Detector with Feature Refinement for Rotating Object

Multiple Attention Encoded Cascade R-CNN for Scene Text Detection