Pyramid Point: A Multi-Level Focusing Network for Revisiting Feature Layers

We present a mcthod to lcarn diverse group of object categories from an unordcrcd point set. propose our Pyramid Point network, which uses dense pyramid structure instead thc traditional ’U’ shape, typically seen in semantic segmentation networks. This gives second look, allowing network revisit different layers creating various leveis on the for feature propagation. introduce Focused Kemel convolution (FKP Conv), expands convolutions by adding attention mcchanism kemel outputs. FKP Conv increases quality and allows us weigh outputs dynamically. These Convs are central part Recurrent Bottlcneck block, makes up backbonc encoder. With this distinct we demonstrate competitive performance threc benchmark data sets.