Adaptive Selection of Optimal Feature for Object Detection

摘要

We propose the Adaptive Selection of Optimal Feature (ASOF) module, a new approach utilizing anchor-free is used for object detection. Most of state-of-the-art object detectors are anchor-base, which use heuristic-guided anchor boxes. Such design is not suitable for detecting objects with different scale and aspect ratios, especially with highly overlapping borders. To address this gap, we propose ASOF, which assigns each instance to the most suitable feature layers by combining pre-defined setting with automatic feature selection. Each instance pass through the forward network to select one FPN layer by manual control, then using the automatic control to revise the former result to achieve the optimal feature map. Through such strategy, we separate most of the overlapping targets. Numerical results suggest that our method has better robustness for mAP than the method with anchor-based detector.