The KITTI Vision Benchmark Suite

we propose a straightforward yet effective fusion
network termed RAFDet that combines the merits of
raw point and its range view to mitigate the
sparsity problem of point cloud, thus boosting the
detection performance. In particular, we introduce
a novel bidirectional attentive fusion module to
equip sparse points with interacted fine-grained
indepth point features. Then, we devise the
range-view augmented fusion module to fully
exploit the supplementary relationship between
different perspectives with the aim of enhancing
original point view features. In the end, a
single-stage detection head is utilized to predict
final 3D bounding boxes based on the enhanced
semantics.

none

none