Context embedded deep neural network for indoor object detection

Indoor object detection plays an important role in home surveillance and disability support. Given multiple objects, existing methods typically detect them separately, which ignores the spatial relationship between different objects. To boost the detection accuracy for indoor objects, we propose a context embedded deep neural network in this paper. Instead of convolving each local patch independently on the feature map, we sequentially process the feature map line by line in two spatial orders, i.e., horizontal and vertical directions, which is implemented via the recurrent neural network. Since the context information preserved in previous lines is passed to the next lines, the proposed method could efficiently capture spatial relationship between different objects and delivers a high detection accuracy on a large-scale indoor object detection database, which significantly outperforms the state-of-the-art Single Shot MultiBox Detector (SSD) at the cost of moderate complexity increase.