https://bitsensing-my.sharepoint.com/:b:/g/personal/heron_bitsensing_onmicrosoft_com/EY9QQM2Uhy5JuV0AyAQ-4BEBO7E2XkDNYIh9-Kd-X9oItQ?e=P0zpeg

1. Introduction

LiDAR

high maintenance need

high cost

not ready to mass-production

Radar-camera association(fusion)

Rule-based

hard to adapt to ever-growing data

Learning-based

LiDAR-based ground-truth

proposed method

without using ground-truth labels from LiDAR

2. Related Work

2.1 Sensor Fusion

object-level fusion: main stream approach

Data association between different sensory outputs is hard

2.2. Learning-Based Radar-Camera Fusion

radar-camera fusion categories

data-level, feature-level

rely on LiDAR

object-level

under-explored

2.3. CNN for Heterogeneous Data

CNN

tremendous success

can handle heterogeneous data as a form of pseudo-image

2.4. Representation Learning

Input data → high-dimensional feature space vector

For understanding complex concept