a point cloud could be regarded as a set of disorder points. The purpose of PointNet is to extract a global invariant feature of the point cloud, regardless of the input sequence of points, constructing a set function that could map a point set to a vector.
Although the whole structure of PointNet is rather simple and readily comprehensible, splendid results have been achieved in some databases. However, for lack of local geometric information, the performance of this approach is limited. For further exploiting the local features, LSTM has been combined with PointNet, constituting our PLSTMNet.
we apply the interaction among points in a specific neighborhood relationship as the local structure information. and the high-dimensional feature vectors extracted by PointNet contain this interaction. since LSTM could attain the relationship among the input via its sequence, it could likewise capture such local geometric relationship among points.
Due to the simplification of the forgetting gate, the memorizing length of the LSTM is infinite. Unlike the original RNN, its performance will decrease with the increase of the length. Thus, we hope to use the advantage of LSTM to classify the feature vectors of PointNet.
the experiments are carried out on the Stanford LargeScale 3D Indoor Spaces Dataset (S3DIS)
avg class accuracy: 可以提高大约1.8个点
overal accuracy: 可以提高大约0.4个点