ICRA 22/05/2022

SeesonDepth Prediction Chanage

Dumpling - Lingzhi 

洛杉矶大学

SwinMono : Swin Transformer based Monoclar Depth on SeasonDepth

Training Monodepth2 with Many Datasets

South Korea

Monodepth2

ICCV2019 Digging into self Supervised Monocular Depth

my

My Contribution : use my datasets

ResNet18 encoder

Per-Pixel Minimum Reprojection Loss

Auto musking

Datasets

Size 11586 images 17225 images(validation Set)

KITTI Depth Evaluation Set

No special reason ,I forget use validate data

DDAD

日天之主 : monodepth2还是强啊，什么都没改

Visual Attention Network based Monocular Depth Estimation

A self(un)-supervised learning 3D scene

Training schemes

depth cnn

pose cnn

2017

2019

2021

Back bones

Training detail

Some findings

SS-learning car

Automatic Parameters Tuning for Autoomous Drving Systems

YanChang,Planningh and Control Teeam

ZOOX

personal transportation , Purpose-Built

sensors/silding doors / on-board computer

 Sound Light

much safety

perception / prediction / planning & controls

MetaSLAM: Towards General Localization

Peng Yin

current

key challenges

Research work:

automerge

iLoc

iSimLoc : sim - to - real Uav - turn google earth into open

GPR competition

Multi-agnet objects

how to do without GPS

spherical feature -> attention enhanced <- map feature

merge

success in 120 km

sim

sattle

conditional domain /pose estimation

iLoc : viewpoint and conditon invariant additive feature extractor for visual localization

Assistance in Human Navigation

odometry camera

GPR Competition : Towards General Place Recognition

ALITA : A large-scale Incremental Dataset for Long-term Autonomy

250434km

dynmica object / noisy ?  same nothing.

How to make Trustworthy and Robust Autonoomus Vehicles ar 270 km/h ?

 Johannes Betz

Why racing ?

detecting the vehicle limits

decision making at the vehicle limits

handling at the vehicle limits

The Indy Autonomous Challenges

Radar / Lidar / Camera / CPU / GPU / Combustion Engine

Autonomous Racing:

Autonomous Racing not about control.

How we do that ?

Model the Vehicle Dynamics

Why? Less paramter , transfer to unknown environments

Detect The Vehicles Limits

Why? Predict and estimate friciton coefficient

Speed mater

Plan a Local Feasible Trajectory

Why? Derive a feasible and dynamic trajectory

Incertiaintu tube gennerated by MPC

MPC based on double track model

Adjust the Controller

Why? Reduce control errors while driving

How to verify robustness and trustworthyness at 270 km/h

Developmet Process Overview

Parameter Tuning

eexection

Adversarial Testing

simulation / real vehicle

Why? verify and cellect

Fail show in some video

people intresting and ask some qustion in 20mins

Data-Effcients

Publications and research fields

Labeled and unlabeled data with multiple modalities

2D image / 3D point / cloud trajectory / HD map

How to exploit labeled & unlabled data .etc?

The problem Semi-Supervised Learning(SSL) on paried 2D & 3D

2D & 3D modalities are good at different things

A baseline on 2D & 3D DetMatch framework

Results & Visualizaitons

SOTA accuracy

Sim2Real for Vision-Based High-Speed Flight

Antonio Loquercio

Human pilots take years ,good or better than human pilots

Why Agile Flight ?

Making drones faster increases their range(limited by battery life)

Applications:

search & rescure

Transfer Learning: Abstraction

What‘s the problem with that?

simulation /Real world :  RGB depth

Data

Asychronous Multiple-Hypothesis Action Prediction

Stereo Depth / IMU / Refrecen Direction

Controlled Experiments

Comparison with state-of-the-art navigation methods

What do next ?

Robots does not learn efficiently in the real world

Natural agents do not have something like training and deploy time.

Uncertainty in Deep Learning

A robotic system not check true.

Representation Learning for Dynamic 3D Object Tracking

Fisher Yu (Fisher Yu)

ICRA2022 Season Depth

Dynamic 3D Environment

Learning 

3D world /3D motion / heigh-fidelity instance-level reresentation

Multiple Object Tracking

Current Tracking Algorithms

直播主题： ICRA国内直播：自动驾驶中的深度估计 & 可信赖自动驾驶汽车与机器人

专题报告

直播时间： 2022年05月23日（周一） 21:30：00—5月24日（周二）04:10

直播嘉宾：

1、苏黎世联邦理工学院计算机系助理教授 Fisher Yu

2、亚马逊自动驾驶部门Zoox高级研究员, 前Lyft自动驾驶技术主管 Yan Chang 3、MIT博士 Heng Yang

4、斯坦福大学智能系统实验博士后研究员 Jiachen Li

5、加州大学伯克利人工智能研究院博后 Antonio Loquercio

6、卡内基梅隆大学机器人研究所项目科学家 Peng Yin

7、宾夕法尼亚大学安全自动系统实验室博士后Johannes Betz

8、麦吉尔大学智能交通实验室博士后Wenshuo Wang

9、加州大学伯克利分校助理教职研究员Wei Zhan

10、伊利诺伊大学厄巴纳-香槟分校博士Linyi Li

直播大纲：

1、变化环境下室外单目深度估计竞赛获奖团队（6个）技术分享

2、可信赖自动驾驶汽车与机器人专题报告

来自 <动态-哔哩哔哩>

分享机器人SLAM、三维重建、增强现实、自动驾驶、深度学习、定位导航等技术 微信公众号：计算机视觉life 进行业（SLAM、三维视觉、传感器、自动驾驶、计算摄影、检测、分割、识别、医学影像、GAN、竞赛）交流群

来自 <计算机视觉life的个人空间_哔哩哔哩_Bilibili>

ICRA直播：自动驾驶中的深度估计机器人 - 计算机视觉life - 哔哩哔哩直播，二次元弹幕直播平台 (bilibili.com)