Abstract
Advances in Convolutional Neural Network (CNN) has spurred extensive research on computer vision as object detection is a very important part. Object detection can either be one stage, such as Single-Shot Multibox Detector (SSD) and You Look Only Once (YOLO), and two stage, such as the faster region-based CNN. Majority of these studies use RGB images. Insufficient lighting often cause detection errors, which we aimed to solve by combining depth information with RGB images. We used edge attention-based learning to extract the edge features from RGB and depth images. The extracted features were fused with the feature maps from the backbone. These features were enhanced using the enhanced feature block and upsample block for small objects. From the final layer feature maps of the RGB and depth paths, we adjust the weighting of the RGB and depth paths to fuse their features, which significantly improves the object detection performance. Finally, the results obtained by the weight fusion layer were combined with the edge images to adjust the confidence score according to the edge ratio and output the final result. We evaluated our method on the SUN RGB-D dataset.