PointPillars Fast Pillar Detector

Skip 3D cubes and stand pillars on the ground plane. PointPillars runs 2D convolutions on pillars for real-time lidar boxes.

Vertical pillars collapse into a pseudo image that a 2D head detects

Why Does This Exist?

VoxelNet pays for 3D convolutions even though cars never stack vertically. PointPillars exists to collapse height: it groups points into vertical pillars, encodes each with a mini-PointNet, scatters codes into a bird's-eye-view pseudo image, and detects with fast 2D convolutions. Speed jumps past $60$ Hz with small accuracy cost.

Sensing context is LiDAR based; task context is 3D object detection.

Think of It Like This

Counting crowds from a balcony

Judge a plaza crowd by looking straight down and counting heads per floor tile. Nobody stacks, so height adds nothing; tile counts plus shoe colors already find groups.

Pillars are floor tiles and pseudo images are the down view. The analogy stops at overpasses: tiles merge stacked roads while pillars keep max height features to separate them.

How It Actually Works

Pillars of $0.16 \times 0.16$ m footprint keep up to $100$ points each, augmented with offsets to pillar center and mean. A shared MLP plus max-pool gives one $64$ -dim code per pillar, scattered into a $(H, W, 64)$ pseudo image. A 2D backbone with top-down fusion plus an SSD-style head regresses boxes and heading. Height resolves from pillar codes rather than explicit $z$ convolutions.

Worked numbers

A $70 \times 70$ m scene at $0.16$ m gives about $440 \times 440$ pillars, but only $12000$ hold points ( $6$ percent dense). Inference runs near $16$ ms ( $62$ Hz) versus $250$ ms for early VoxelNet, while KITTI moderate car AP moves from $81$ to about $77$ percent.

Code

1def pillar_count(area: float, size: float, dense_frac: float) -> int:2    total = (area / size) ** 2  # grid cells over square scene3    return int(total * dense_frac)4
5
6print(pillar_count(70.0, 0.16, 0.06))  # -> 11484

Watch Out For

Pillaring tall stacks as one cell

Bridges over roads share pillars with cars below, and max-pooling can keep the wrong height. The symptom is boxes floating at deck level. Check height histograms on multi-level scenes and keep explicit $z$ features in the code.

The Quick Version

PointPillars groups points into vertical pillars, not 3D voxels.
Each pillar encodes to one code scattered into a pseudo image.
Fast 2D convolutions replace costly 3D ones.
Pillar $0.16$ m grids run past $60$ Hz near $77$ percent car AP.
Height survives through pillar features, not 3D math.
Stacked roads need extra height care.