PIXELBANKv8.2.1
Menu
Back to 3D Computer Vision Study Plan
Week 8

Chapter 8: 3D Object Detection

Master the techniques for detecting and localizing objects in 3D space using LiDAR point clouds, camera images, and sensor fusion. From voxel-based encoders to bird's eye view transformers, learn the full detection pipeline used in autonomous driving and robotics.

Chapter Overview

3D object detection is one of the most critical perception tasks in autonomous systems. Unlike 2D detection which produces bounding boxes on images, 3D detection estimates the full 3D bounding box of each object --- including its position (x,y,z)(x, y, z), dimensions (l,w,h)(l, w, h), and heading angle θ\theta in the real world.

The primary sensor for 3D detection is LiDAR, which produces sparse point clouds with precise depth measurements. Camera-based methods are also gaining traction due to their lower cost. State-of-the-art systems typically fuse both modalities for the best accuracy.

This chapter traces the evolution of 3D detection architectures: from early voxel-based methods (VoxelNet) and efficient pillar encodings (PointPillars), through center-based detectors (CenterPoint), to modern bird's eye view (BEV) approaches that unify camera and LiDAR in a shared representation. We also cover sensor fusion strategies and the specialized evaluation metrics unique to 3D detection.

Key topics include:

  • 3D bounding box parameterization and detection problem formulation
  • Voxel-based and pillar-based point cloud encoding for efficient processing
  • Center-based detection with heatmap predictions for anchor-free design
  • Bird's eye view representations that project 3D scenes to a top-down plane
  • Sensor fusion strategies for combining LiDAR and camera data
  • Evaluation metrics specific to 3D detection benchmarks

Chapter Roadmap

Click any topic to jump in

1
3D Detection Overview

Problem formulation — 7-DOF bounding boxes, input representations, and anchor-based vs anchor-free paradigms.

Voxel vs pillar encoding

Two paradigms for converting point clouds into structured representations

2
VoxelNet

End-to-end voxel-based detection with learned Voxel Feature Encoding and sparse 3D convolutions.

3
PointPillars

Fast pillar-based encoding that converts point clouds to pseudo-images for efficient 2D backbone processing.

Modern detection and BEV

Anchor-free detection and top-down scene representations

4
CenterPoint

Anchor-free center-based detection using heatmaps with two-stage refinement for improved localization.

5
BEV Detection

Bird's eye view representations from cameras and LiDAR — LSS, BEVFormer, and multi-camera fusion.

Combining sensors and measuring performance

Multi-modal fusion and standardized benchmarks

6
Sensor Fusion

Combining LiDAR and camera modalities via decoration, feature-level, and unified BEV fusion strategies.

7
Evaluation Metrics

3D AP, BEV AP, range-stratified evaluation, and the nuScenes Detection Score (NDS).

Sign up to unlock this chapter

This chapter is part of PixelBank Premium. Create a free account, then upgrade to read the full lesson — concepts, walkthroughs, and exercises.

3D Object Detection — 3D Computer Vision | PixelBank