Dual-Resolution Feature Fusion Based on Curriculum Learning

In the domain of object detection, achieving consistent accuracy across varied object scales is a formidable challenge. Most existing methodologies aim to enhance model scale invariance by merging features from both high and low-resolution images. However, these methods often employ static fusion techniques that neglect the evolving significance and dynamics of features throughout different stages of the learning process. This oversight can limit the optimal employment of multi-scale information, thereby restricting potential enhancements in detection accuracy. To address this limitation, we introduce the Dual-Resolution Curriculum Fusion (DRCF) framework. DRCF innovatively conceptualizes feature fusion as a two-phase process. In the initial phase, the framework concentrates on separately training features derived from both high and low resolutions, prioritizing the stabilization of these foundational resolution features. Upon achieving stability, DRCF transitions its focus to the fusion process, adeptly integrating these dual-resolution features. This sequential approach, leveraging the principles of curriculum learning, ensures a more effective and nuanced assimilation of multi-resolution information by dynamically adjusting the fusion strategy in alignment with the learning progression. Such an approach not only enhances the feature fusion quality but also significantly improves object detection accuracy. Empirically, our framework demonstrates a notable 0.7% improvement in mean average precision (mAP) over the existing state-of-the-art, evidencing the effectiveness and innovative contribution of our proposed method.