Learnable Loss Balancing in Anchor-Free Oriented Detectors for Aerial Object

Oriented object detection plays an important role in aerial image interpretation. Image processing speed is also essential due to massive amounts of aerial images. Anchor-free oriented detectors with fast processing speed are generally accepted despite the absence of preset anchors, contributing to their performance gap with anchor-based detectors. Most anchor-free oriented detectors are carefully designed by defining samples according to target characteristics, which require substantial prior knowledge, to realize improved performance. This study proposes an anchor-free oriented detector (termed rfpoint) that requires minimal prior knowledge. Moreover, this study mainly aims to introduce a dynamic sample definition strategy. This strategy is modeled as a dynamic regulating process, wherein the classification and box regression interact until the model converges. A rotating quality-driven loss (RQDL) and an adaptive-weight box loss (AWBL) are also proposed to realize the aforementioned process. RQDL redefines positive and negative attributes of samples according to the distribution of rotating intersection-over-unit (IoU) between predictions and ground truth. AWBL adjusts the importance degree of candidate samples in the box regression through classification scores. The proposed method is then tested on three mainstream aerial image datasets (DOTA, DIOR, and HRSC2016). Results reveal that the proposed method achieves the best performance compared with other oriented detectors, whose mean average precision (mAP) is 79.92%, 70.88%, and 90.67%. Moreover, the method maintains the inference speed advantage of anchor-free detectors. An effective sample definition method can bridge the performance gap of anchor-free oriented detectors without minimizing inference speed.

Automated summary

This study proposes an anchor-free oriented detector called rfpoint that requires minimal prior knowledge and introduces a dynamic sample definition strategy. The method achieves sample definition by dynamically adjusting the process of classification and box regression using rotating quality-driven loss and adaptive-weight box loss. Experimental results show that the proposed method outperforms other oriented detectors on three mainstream aerial image datasets, with mean average precision (mAP) of 79.92%, 70.88%, and 90.67%, while maintaining the inference speed advantage of anchor-free detectors.