The paper starts by analyzing the limitations of the commonly used AP50 metric for oriented object detection. It argues that AP50 is inherently unsuitable for this task due to its large tolerance in angle deviation, and advocates the use of more stringent metrics like AP75 to better evaluate high-precision performance.
The authors then propose the ARS-DETR model, which includes several key components:
Aspect Ratio aware Circle Smooth Label (AR-CSL): A new angle classification method that smooths the angle label in a more reasonable way by considering the object's aspect ratio, eliminating the need for hyperparameters.
Rotated Deformable Attention Module: A module that rotates the sampling points according to the embedded angle information to align the features with the objects.
Denoising training strategy: A modified version of the DINO denoising training that adds noise to the angle predictions.
Aspect Ratio Sensitive Weighting and Matching: Modifications to the angle loss function and matching cost to account for the sensitivity of objects with different aspect ratios to angle deviation.
Comprehensive experiments on DOTA-v1.0, DIOR-R, and OHD-SJTU datasets demonstrate that ARS-DETR achieves state-of-the-art performance on the high-precision AP75 metric, outperforming other advanced oriented object detectors.
To Another Language
from source content
arxiv.org
Key Insights Distilled From
by Ying Zeng,Xu... at arxiv.org 04-04-2024
https://arxiv.org/pdf/2303.04989.pdfDeeper Inquiries