3DOR 17

Permanent URI for this collection

https://diglib.eg.org/handle/10.2312/2631242

Browse

Browsing 3DOR 17 by Subject "Curve"

Now showing 1 - 2 of 2
  • Thumbnail Image
    Item
    Directed Curvature Histograms for Robotic Grasping
    (The Eurographics Association, 2017) Schulz, Rodrigo; Guerrero, Pablo; Bustos, Benjamin; Ioannis Pratikakis and Florent Dupont and Maks Ovsjanikov
    Three-dimensional descriptors are a common tool nowadays, used in a wide range of tasks. Most of the descriptors that have been proposed in the literature focus on tasks such as object recognition and identification. This paper proposes a novel three-dimensional local descriptor, structured as a set of histograms of the curvature observed on the surface of the object in different directions. This descriptor is designed with a focus on the resolution of the robotic grasping problem, especially on the determination of the orientation required to grasp an object. We validate our proposal following a data-driven approach using grasping information and examples generated using the Gazebo simulator and a simulated PR2 robot. Experimental results show that the proposed descriptor is well suited for the grasping problem, exceeding the performance observed with recent descriptors.
  • Thumbnail Image
    Item
    LightNet: A Lightweight 3D Convolutional Neural Network for Real-Time 3D Object Recognition
    (The Eurographics Association, 2017) Zhi, Shuaifeng; Liu, Yongxiang; Li, Xiang; Guo, Yulan; Ioannis Pratikakis and Florent Dupont and Maks Ovsjanikov
    With the rapid growth of 3D data, accurate and efficient 3D object recognition becomes a major problem. Machine learning methods have achieved the state-of-the-art performance in the area, especially for deep convolutional neural networks. However, existing network models have high computational cost and are unsuitable for real-time 3D object recognition applications. In this paper, we propose LightNet, a lightweight 3D convolutional neural network for real-time 3D object recognition. It achieves comparable accuracy to the state-of-the-art methods with a single model and extremely low computational cost. Experiments have been conducted on the ModelNet and Sydney Urban Objects datasets. It is shown that our model improves the VoxNet model by relative 17.4% and 23.1% on the ModelNet10 and ModelNet40 benchmarks with less than 67% of training parameters. It is also demonstrated that the model can be applied in real-time scenarios.