Review and Systematization of Solutions for 3D Object Detection

Abstract

Since 2017 there has been an exponential growth in scientific publications regarding the field of 3D object detection (3DOD). On the one hand, this growth can be explained by the strong demand for autonomous vehicles, and on the other hand, by the wide availability of 3D sensors. Due to the strong heterogeneity of developed approaches, this paper aims to identify, analyze and systematize publications that propose end-to-end solutions for 3DOD towards the goal to provide a structured framework which can guide future development, evaluation and application activities. To carry out the research, a systematic literature review is combined with a taxonomy development approach. The resulting framework consists of six dimensions, covering the addressed domains, applied datasets, sensor properties, data representation formats, modeling techniques, and evaluation criteria. The taxonomy can help researchers and practitioners to get a quick overview about the field by decomposing 3DOD solutions into more manageable pieces.

Outdoor Mobile Mapping and AI-Based 3D Object Detection with Low-Cost RGB-D Cameras: The Use Case of On-Street Parking Statistics

“…In recent years, traditional approaches have been replaced by neural networks with increasingly deep network architectures. This allows the use of high dimensional input data and automatic recognition of structures and representations needed for detection tasks [29].…”

“…First, 2D candidates are detected within the image, before in the second step 3D bounding boxes representing the object are computed based on the candidates. Either neural networks, geometric constraints, or 3D model matching are used to predict the 3D bounding boxes [29].…”

labelCloud: A Lightweight Domain-Independent Labeling Tool for 3D Object Detection in Point Clouds

“…The unordered nature of this data type and the absence of any fixed grid (like in 2D images) has made it difficult to simply lift successful solutions from 2D computer vision into the 3D space. However, a new architecture called PointNet [4] has led to a breakthrough and enabled multiple solutions that can detect objects directly inside the point cloud data [2]. 3D object detection methods can automatically identify and locate objects with their class, position, dimension, and sometimes even rotation.…”

A Light in the Dark: Deep Learning Practices for Industrial Computer Vision

“…Deep learning (DL) has recently received increasing attention in research and industry alike, where we can find manifold examples of successful applications like machine translation [1], predictive maintenance [2], autonomous driving [3], and predictive business process monitoring [4]. Given a sufficiently large amount of training data, DL has the advantage of automatically detecting useful patterns from raw input data without the need of manually defining sophisticated features for prediction purposes [5,6].…”