Complex-YOLO Paper - Intro - Analysis

[NicolaBernini]

Overview

Complex-YOLO: Real-time 3D Object Detection on Point Clouds

Arxiv: https://arxiv.org/abs/1803.06199

Key	Value
Type of Contribution	DNN Model
Application	Object Detection from 3D Point Cloud provided by Lidar
Domain of Application	Autonomous Car

[NicolaBernini]

Goal

Develop a pipeline connecting

• Input: Point Cloud as BEV RGB Map
  • Obtained as preprocessing
  • RGB means just 3 channels image, encoding:
    • 1. height from the ground
    • 2. laser ray intensity
    • 3. density (resulting from spatial discretization)
• Output: Detection Results (Bounding Box + Semantic Label)

• From the Paper Fig1

Related Work

PointNet

• Used for
  • Classification
  • Parts Segmentation
  • Semantic Segmentation

[NicolaBernini]

Work Challenges

Description

1. Design a NN able to take sparse data as input, which is difficult in general
2. Performing an adaptation of an existing network, designed to perform detection on a dense input

Notes

• Actually this has been performed with a pre-processing step to convert a the Point Cloud which is a Sparse Data Structure into an Image which is a Dense Data Structure
• This could possibly add some extra computational cost and be sensitive to the non trainable pre-processing related hyperparams like Grid Cell Size

[NicolaBernini]

BEV Construction

• Prior on the space with a certain regular geometry
  • Each cell is 8cm square
• Detected Points get projected on the grid according to sensor extrinsic calibration and this defines a Lidar point to cell association
• According to this association it is possible to compute cell specific statistics defining the values for the 3 channels: average height, average intensity and density as number of points in the cell

[NicolaBernini]

Procedure

1. Preprocessing

1.1 Project into Grid

Goals

• Removing Ego Perspective so to work in BEV
• Introduce Spatial Quantization

1.2 Compute Statistics on Spatial Grids

Goals

• Build an RGB like representation, more precisely a WxHx3 Tensor with channels representing
  • height from ground
  • intensity
  • cell density

2. Processing

• Use a YOLO like NN to process the representation resulting from preprocessing step

[NicolaBernini]

Processing Strategy

• This method relies on a specifically designed Region Proposal Network
  • One of the most famous improvement in Object Detection Research, introduced in Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks