GPU Rasterization-Based 3D LiDAR Simulation for Deep Learning.

High-quality data are of utmost importance for any deep-learning application. However, acquiring such data and their annotation is challenging. This paper presents a GPU-accelerated simulator that enables the generation of high-quality, perfectly labelled data for any Time-of-Flight sensor, including LiDAR. Our approach optimally exploits the 3D graphics pipeline of the GPU, significantly decreasing data generation time while preserving compatibility with all real-time rendering engines. The presented algorithms are generic and allow users to perfectly mimic the unique sampling pattern of any such sensor. To validate our simulator, two neural networks are trained for denoising and semantic segmentation. To bridge the gap between reality and simulation, a novel loss function is introduced that requires only a small set of partially annotated real data. It enables the learning of classes for which no labels are provided in the real data, hence dramatically reducing annotation efforts. With this work, we hope to provide means for alleviating the data acquisition problem that is pertinent to deep-learning applications.

MaskLayer: Enabling scalable deep learning solutions by training embedded feature sets.

Deep learning-based methods have shown to achieve excellent results in a variety of domains, however, some important assets are absent. Quality scalability is one of them. In this work, we introduce a novel and generic neural network layer, named MaskLayer. It can be integrated in any feedforward network, allowing quality scalability by design by creating embedded feature sets. These are obtained by imposing a specific structure of the feature vector during training. To further improve the performance, a masked optimizer and a balancing gradient rescaling approach are proposed. Our experiments show that the cost of introducing scalability using MaskLayer remains limited. In order to prove its generality and applicability, we integrated the proposed techniques in existing, non-scalable networks for point cloud compression and semantic hashing with excellent results. To the best of our knowledge, this is the first work presenting a generic solution able to achieve quality scalable results within the deep learning framework.