ECNet: Edge-aware Context-aggregation Network for Transparent and Reflective Object Detection

Highly transparent and reflective objects are prevalent in life, yet their existence severely degrades the performance of existing object detectors. In this paper, we propose a novel Edge-aware Context-aggregation Network (ECNet) for transparent and reflective object detection, with three well-designed modules introduced to enhance the ability of extracting distinctive features. Specially, we develop the Edge Detection Module (EDM) to make the network pay more attention to the boundary areas; present the Depth Feature Fusion Module (DFFM) to extract content discontinuity in depth maps; and introduce the Multi-respective-field Feature Extraction Module (MFEM) which fusing multi-receptive-field features to reveal the texture discontinuity. Extensive experiments show the proposed ECNet surpasses the state-of-the-art detectors by at least 7.4% mAP in transparent and reflective object detection, which well demonstrates the effectiveness of the proposed method.