The KITTI Vision Benchmark Suite

VIQC-Net is a LiDAR-only two-stage 3D object
detector built on a sparse VoxelNeXt backbone.
FICR aggregates local cross-scale voxel
information to enhance sparse bird's-eye-view
features. SmallLQA estimates localization quality
for class-specific confidence fusion. Multi-scale
voxel RoI pooling refines proposals, followed by
class-wise size calibration using training-set
statistics. The submitted single model was trained
from scratch for 80 epochs on all 7,481 labeled
KITTI training frames, without external training
data or RGB image features.

Class order: Car, Pedestrian, Cyclist.
Fixed size calibration alpha: [0.10, 0.50, 0.75].
Fixed quality fusion beta: [0.75, 0.75, 0.75].
RoI pooling grid: 6 x 6 x 6.
Training RoIs per frame: 160.
Pre-refinement NMS threshold: 0.7.
Final NMS threshold: 0.1.
Class size priors are computed from all 7,481
labeled training frames.

Runtime: 0.1259 seconds per frame on one NVIDIA
L40, measured over the original test loop,
including data loading and prediction export.
Subsequent offline angle normalization and
geometric visibility correction are excluded from
this timing.