Systems, methods, and computer programs for using a network of machine learning models to detect an image depicting an object of interest which can be partially occluded by another object

Abstract

Methods, systems, and apparatus, including computer programs encoded on computer-storage media, used for object detection in an input image that can include at least partial object occlusion. In some implementations, input data representing the image depicting an object can include object-based features and context-based features used for object detection. The feature is processed by a deep convolutional neural network (DCNN) model. First feature data generated by the DCNN is provided to an occlusion model and a generative compositional model. The occlusion model can detect locations where an object depicted in the image is occluded by an object of any other type. The generative compositional model detects the presence of different classes of objects that represent parts or partial components of object depicted in the image. The output of the compositional model and occlusion model is a likelihood map that shows if an object is depicted in the input image.