The KITTI Vision Benchmark Suite

We propose Voxel-Dense Fusion (VDF), a 3D target
detection algorithm based on point cloud and image
fusion. The voxelised point cloud data is first
converted into a meaningful high-dimensional
feature representation through a voxel set
attention coding process. At the same time, a
region of interest (ROI) can be provided for the
dense point cloud branch.
The dense point cloud branch uses PointMLP, a
dense point cloud feature encoding scheme based on
neighbourhood feature aggregation, as the backbone
network to obtain the output features of the dense
point cloud branch.
To fuse the features from the voxel branch and the
dense point cloud branch, the fusion module
performs feature alignment using Region of
Interest Grid Pooling (ROI Grid Pool). Finally,
the aligned features are fed into the 3D voxel
Recursive Gated Convolution fusion network to
achieve an efficient fusion of the features from
both data sources.

Adam，lr=0.01，Experiments were conducted using an
NMS threshold of 0.8 to generate 128 object
suggestions with a positive and negative sample of
1:1. During testing, an NMS threshold of 0.1 was
used to remove redundant boxes after suggestion
refinement.