Основні поняття
Poly Kernel Inception Network (PKINet) addresses challenges in remote sensing object detection by utilizing multi-scale convolution kernels and a Context Anchor Attention module to improve performance.
Анотація
PKINet introduces a novel approach to object detection in remote sensing images, focusing on object scale variations and contextual diversity. The network employs multi-scale convolution kernels without dilation to extract features of varying scales and capture local context. Additionally, a Context Anchor Attention module is introduced to capture long-range contextual information. Extensive experiments on benchmark datasets demonstrate the effectiveness of PKINet in improving object detection performance.
PKINet's design allows it to outperform previous methods by effectively handling challenges related to object scale variations and diverse contexts in remote sensing images. By incorporating multi-scale convolution kernels and a Context Anchor Attention mechanism, PKINet achieves superior performance on challenging benchmarks like DOTA-v1.0, DOTA-v1.5, HRSC2016, and DIOR-R.
The network's innovative architecture enables it to adaptively extract features with both local and global contextual information, leading to improved object detection accuracy in remote sensing images.
Статистика
Miss detection: 72.45
Wrong detection: 69.70
Miss detection: 74.21
Wrong detection: 74.05
Miss detection: 75.87
Wrong detection: 74.86
Miss detection: 75.89
Wrong detection: 77.83
Цитати
"Our multi-scale convolution handles scale variations well."
"PKINet represents the pioneering effort in exploring inception-style convolutions for remote sensing object detection."