SAE‐CenterNet: Self‐attention enhanced CenterNet for small dense object detection

Abstract

The existing object detection models have not been sufficiently optimized for small dense objects. One of the most common solutions is to extract multi-scale features via feature pyramid network (FPN). However, the information loss of downsampling in multi-scale feature extraction will seriously affect the detection accuracy. Therefore, a dynamic attention convolution (DAC) for downsampling is developed, which can embed regional information for each single pixel. Besides, an attention fusion module (AFM) is also designed to alleviate the data inconsistency from different layers during multi-scale feature fusion. Based on these, the proposed model, SAE-CenterNet, has achieved optimal performance in the mainstream object detection models on the small dense rebar dataset. For example, with 6 FPS decreasing, the m A P 50 $ mAP_{50}$ , m A P 75 $ mAP_{75}$ , R e c a l l 50 $ Recall_{50}$ and R e c a l l 75 $ Recall_{75}$ of SAE-CenterNet is 87.3%, 57.6%, 88.6% and 67.2%, respectively, which are 8.0%, 13.5%, 8.3% and 10.4% higher than the baseline CenterNet, respectively.