0
点赞
收藏
分享

微信扫一扫

笔记:Few-shot learning for tackling open-set generalization

彩虹_bd07 2022-04-07 阅读 70
人工智能

笔记:Few-shot learning for tackling open-set generalization:

  • 基于点云的语义分割的应用:场景理解,给点云中每一个点赋予特点的语义标签。(如自动驾驶)

  • 小样本学习的意义:解决太过于依赖大量标定数据,减少成本;可以提高泛化能力,识别未曾见过的目标。

  • paper1:Few-shot 3D Point Cloud Semantic Segmentation

    • 提出问题:

      • rely on large amounts of labeled training data, so they are time-consuming and expensive to collect.
      • follow the closed set assumption.(训练集和测试集取自同一label space) ,泛化能力差。
    • 解决:

      • multi-prototype transductive inference method.
        • transductive inference: 转导推理;是一种通过观察特点的样本,进而预测特定的测试样本的方法,是一种从特殊到特殊的推理,适合于小样本推理。不同于归纳推理,先从训练样本中学习规则,再用规则判断测试样本。
      • architecture
        在这里插入图片描述

        • embedding network:

          • three properties:1.local geometric features; 2.global geometric features; 3. adapt to different few-shot tasks.
          • DGGNN: the backbone of feature extractor.(local)
          • SAN(self-attention network): generate semantic feature.(global)
          • MLP: adapt to different few-shot tasks.
        • multi-prototype generation:

          • It samples a subset of n seed points from a set of support points in one class using the farthest point sampling based on the embedding space.(对support set的每一类样本点farthest points sample,抽取n个seed point)
          • The farthest points represent different perspectives of one class. (farthest points sample保证足够的感受野)
        • transductive inference:

          • use transductive label propagation to construct a graph on the labeled multi-prototypes and the unlabeled query points.(用k-NN建立相关类的图)

            在这里插入图片描述

        • label propagation

        • cross-entropy loss function(交叉熵损失函数):

          • compute the cross-entropy loss with ground truth labels.
  • paper2:What Makes for Effective Few-shot Point Cloud Classification?

    • 提出问题

      • they require extensive data collection and retraining when dealing with novel classes never seen before.
      • It is hard to study from existing 2D methods when migrating to the 3D domain.
      • point clouds are more complex and have unorder structure in European space.
    • 3D point cloud classification

      • projection-based: It first converts the irregular points into a representation like voxel, pillar, and then apply typical 2D or 3D CNN to extract features.
      • point-based: It can learns point-wise features with multilayer perceptron(MLP) and aggregates global feature with a symmetric function implemented by a max-pooling layer.
    • 2D few-shot learning

      • Metric-based: It focus on learning an embedding space where similar samples pairs are closer, or designing a metric function to compare the feature similarity of samples.
      • Optimization-based: It regards meta-learning as an optimization process.
    • State-of-the-art 2D FSL on Point Cloud

      • compare the metric-based methods and optimization-based methods, and concludes that metric-based methods outperform the optimization-based methods in point cloud scenario.
    • Influence of Backbone Architecture on FSL

      • select three types of current state-of-the-art 3D point-based networks including Pointwise-basedConvolution-basedGraph-based(DGCNN). One can conclude that the graph-based network DGCNN achieves higher classification accuracy than other networks on these two datasets.
    • Cross Instance Adaption (CIA) module
      • CIA can be inserted into existing backbones and learning frameworks to learn more discriminative representations for the support set and query set.

        在这里插入图片描述

        Embedding module把support-set和query-set作为输入分别进行特征提取得到他们的prototype,然后再通过CIA模块更新support-set和query-set,然后在特征空间计算每个class prototype和query examples的欧氏距离,最后便可得到损失函数并进行优化。

      • Self-Channel Interaction Module: address the issues of subtle inter-class differences.

        在这里插入图片描述

        • 先从embedding space分别由两个线性系数φ和γ得到q向量和k向量,然后通过CIM的双线性变换得到一个channel-wise relation score map - R, 然后进行softmax操作得到权重矩阵R’,最后得到更新的向量v是有R’与开始的特征向量加权和得到,vi越大说明特供信息越大,有利于区分class之间的细小差别。
      • Cross-Instance Fusion Module: address high intra-class variances issues

      在这里插入图片描述

      • 首先将support feature和query feature 连结起来得到Z,然后用两个卷积层来解码连结后的特征得到W,将W进行softmax操作得到权值矩阵后与Z点乘来更新support feature和query feature。
    • 本文还提供了两个适用于3D FSL的数据集:ModelNet40-FS,ShapeNet70-FS

举报

相关推荐

0 条评论