Automatic Dense Annotation for Monocular 3D Scene Understanding

Abstract

Deep neural networks have revolutionized many areas of computer vision, but they require notoriously large amounts of labeled training data. For tasks such as semantic segmentation and monocular 3d scene layout estimation, collecting high-quality training data is extremely laborious because dense, pixellevel ground truth is required and must be annotated by hand. In this paper, we present two techniques for significantly reducing the manual annotation effort involved in collecting large training datasets. The tools are designed to allow rapid annotation of entire videos collected by RGBD cameras, thus generating thousands of ground-truth frames to use for training. First, we propose a fully-automatic approach to produce dense pixel-level semantic segmentation maps. The technique uses noisy evidence from pre-trained object detectors and scene layout estimators and incorporates spatial and temporal context in a conditional random field formulation. Second, we propose a semi-automatic technique for dense annotation of 3d geometry, and in particular, the 3d poses of planes in indoor scenes. This technique requires a human to quickly annotate just a handful of keyframes per video, and then uses the camera poses and geometric reasoning to propagate these labels through an entire video sequence. Experimental results indicate that the technique could be used as an alternative or complementary source of training data, allowing large-scale data to be collected with minimal human effort.