Vector encoded bounding box regression for detecting remote-sensing objects with anchor-free methods

We propose a novel convolutional neural network architecture used for detecting objects in high-resolution remote-sensing images. Different from previous detectors, our method is totally anchor-free. In the architecture, we design a new regression method by encoding the bounding boxes into vectors and bring direction information into the network. We also analysed the detection head and proposed the Faster activated detector heads module to accelerate the convergence speed. Experiments were carried out on two public remote-sensing image datasets. Comparing with previous methods, our work shows the most favourable result in the detecting accuracy with no extra trainable parameters added.

Automated summary

The proposed convolutional neural network architecture for object detection in high-resolution remote-sensing images is anchor-free and outperforms previous methods in terms of detecting accuracy without adding extra trainable parameters.