Fast 3D Object Detection of Shoes in the Wild

This work is dedicated to the 3D object detection problem in the wild conditions. The main goal of this research is to design a neural network architecture capable of quickly and accurately enough predicting spatial information about the location of an object using a single RGB image. This task is called "Monocular 3D Object Detection". It was chosen to use the Objectron dataset as a training set, as it is the state-of-the-art and largest dataset for 3D detection. Objectron consists of household items, which currently receive too little attention in the field of 3D detection. This paper presents three 3D detectors capable of predicting a 3D bounding box of an object in real time. These detectors are based on state-of-the-art convolutional neural network architectures, each having a different strength in terms of quality or speed. The results of this work contribute to the development of the 3D object detection field, and can find its application, for example, in robotics and augmented reality tasks.