PaperSummary24 : Oriented R-CNN for Object Detection

The paper introduces Oriented R-CNN, a two-stage framework for detecting arbitrarily oriented objects in images. It addresses the computational bottleneck by proposing a lightweight oriented Region Proposal Network (oriented RPN). The framework efficiently generates high quality oriented proposals, maintaining detection accuracy and competitive speed.

The framework has two key stages:

1. Oriented RPN: It is designed as a lightweight fully convolutional network to generate high-quality oriented proposals in computationally efficient manner. It introduces a novel midpoint offset representation to describe oriented objects using six parameters: center coordinates, width, height and offsets for the midpoints of the top and right sides. It outputs sparse, high-recall oriented proposals, minimizing redundant computation and memory usage.
2. Oriented R-CNN Head: It refines the oriented proposals using rotated RoIAlign, which extracts rotation-invariant features. It performs classification and regression to finalize the detection and orientation of objects.

Image src

The paper employs a combination of classification and Smooth L1 regression loss. Positive and negative samples are defined based on Intersection over union thresholds with ground truth boxes. It achieves 75.87% mean Average Precision on the DOTA dataset while balancing efficiency.

References:

Oriented R-CNN for Object Detection