Directional Alignment Instance Knowledge Distillation for Arbitrary-Oriented Object Detection

Abstract

Recently, many lightweight neural networks have been deployed on airborne or satellite remote sensing platforms for real-time object detection. To bridge the performance gap between lightweight models and complex models, many knowledge distillation (KD) methods are investigated. However, existing KD methods ignore to transfer effective directional knowledge. Meanwhile, knowledge of different subtasks interfere with each other. To this end, a directional alignment instance knowledge distillation (DAIK) method for improving the performance of the lightweight object detection model is proposed. Specifically, an angle distillation (AD) module is developed to combine the circular smooth label and teacher logits to transfer effective directional knowledge. Angular-distance Aspect-ratio Look-up-table (AAL) is incorporated into label assignment and re-weighting loss to enhance the prediction sensitivity of direction and shape in a discrete manner. Sample alignment distillation (SAD) reduces the spatial misalignment by mimicking the teacher model's distribution of anchor points. Extensive experiments are performed on several public remote sensing object detection datasets, which demonstrates the effectiveness of the proposed DAIK.