跳转至

Gaussian based

绷不住了,又有人刷榜了

在这项工作中,Hu Cao 等人建立了一个高效且鲁棒的全卷积神经网络,从来自真实场景的 n 通道图像中获取抓取信息。

  • lightweight generative architecture
  • a grasping representation based on Gaussian kernel
  • a Receptive Field Block is assembled to the bottleneck of grasping detection architecture
  • Combined pixel attention and channel attention

Grasping is a challenge task for robots: perception, planning and extection.

作者贡献如下:

  • We propose a Gaussian-based grasping representation, which relects the maximum grasping score at the center point location and can signigicantly improve the grasping detection accuracy.
  • We develope a lightweight generative architecture which achieves high detection accuracy and real-time running speed with small network parameters.
  • A receptive field block module is embedded in the bottleneck of the network to enhance its feature discriminability and robustness, and a multi-dimensional attention fusion network is developed to suppress redundant features and enhance target features in the fusion process.
  • Evaluation on the public Cornell and Jacquard grasping datasets demonstrate that the proposed generative based grasping detection algorithm achieves state-of-the-art performance of both speed and detection accuracy

Oriented rectangle-based representation

  • Analytic methods: use mathematical and physical models in geometry, motion and dynamics to carry out the calculation for grasping
  • Empirical methods: deep learning
  • Classification-based: Proposals, GQ-CNN, Spatial Transformer Network
  • Regression-based: Multi-model fusion, ROI => more inclined to learn the mean value of the ground truth grasps
  • Vision and tactic sensing are fuse

Point-based Grasp representation

GGCNN

Orientation Attentive Grasping Detection

Gaussian-based

\[ G_{K}=\left\{\Phi, W, Q_{K}\right\} \in \mathbb{R}^{3 \times W \times H} \]

where,

\[ Q_{K}=K(x, y)=\exp \left(-\frac{\left(x-x_{0}\right)^{2}}{2 \sigma_{x}^{2}}-\frac{\left(y-y_{0}\right)^{2}}{2 \sigma_{y}^{2}}\right) \]

where,

\[ \sigma_{x}=T_{x}, \sigma_{y}=T_{y} \]

Questions

What about classification? In some of the implementations they also output class probabilities.

What about implementation?

Are you conducting related researches?

Multiple objects?