Matching strategy and skip-scale head configuration guideline based traffic object detection

Abstract

The configuration of the detection head has a significant impact on detection performance. However, when the input resolution or detection scene changes, there is not a clear method for quantitatively and efficiently configuring the detection head. We find that there is a rule of matching degrees between the object scale and the detection head across different input resolutions or detection scenes by careful analysis. Based on this matching rule, we propose simple yet very effective methods for detection head configuration. The methods consist of two main parts. The first is the matching strategy of detection head and object scale, which can handily and quantitatively guide the rational configuration of detection heads to effectively detect objects at vastly different scales. The second is the skip-scale detection head configuration guideline, which instructs to replace multiple detection heads with only two detection heads to decrease model parameters as well as achieve high detection accuracy and speed. Extensive experimental results on three benchmarks, BDD100K, nuImages and our proposed ETFOD-v2, validate the effectiveness and convenience of our proposed methods, showing potential application prospect in future intelligent traffic systems. The code and ETFOD-v2 dataset are available in https://github.com/YiShi701/MR-Net.