The KITTI Vision Benchmark Suite

This article proposes a 3D object detection
network (SpaA) with spatial perception capability.
On the basis of 3D sparse convolutional networks,
we designed a variable sparse convolutional
network (VSC Conv) that can perceive position
importance. We propose a local aggregation method
based on spatial perception and density
information (SDLA) to address the problem of
completely ignoring spatial structure when
aggregating local features using set abstraction
operations

We use Adam optimizer combined with One Cycle
learning rate strategy for training. The momentum
parameter is 0.9, the initial learning rate is 0.01,
the weight decay coefficient is 0.01, and the
gradient norm clipping threshold is 10.