Alapfogalmak
This research paper introduces a novel depthwise switchable atrous convolutional network for object detection, enhancing the detection of objects at varying scales by dynamically adjusting atrous convolution rates and incorporating global context information.
Statisztikák
The proposed method achieved a mAP of 51.32% on the MSCOCO dataset, outperforming other state-of-the-art methods.
Applying global context to the EfficientNet backbone resulted in a 1% improvement in mAP.
The Depthwise Atrous with Pointwise Switchable Convolution (DAPSC) scheme, along with global context, yielded the highest mAP values among the tested variations.