CSDN: Context-based self-adaptive detection network for real-time object detection.

Total: 1

Convolutional Neural Networks (CNNs) have long been the cornerstones of target recognition, but are often restricted by limited reception fields, which affects their ability to capture global context information. We have re-examined the DETR-inspired recognition head and identified a considerable redundancy in its self-attention module. To solve these problems, we introduced the Context-Gated Scale-Adaptive Detection Network (CSDN), a Transformer-based recognition head inspired by human visual perception: When observing an object, we always focus on a place, perceive the surroundings and look around the object. This mechanism allows each Region of Interest (ROI), feature dimensions and scaling information to be adaptively selected and combined from different patterns. CSDN offers more powerful global context modeling functions and can adapt better to objects of different size and structure. Our proposed detection head can directly replace the native heads of different CNN-based detectors, and only a few fine tuning rounds at the previously trained weights can significantly improve the detection accuracy.