Overview Yandex Services Car Detection Yandex.Taxi 3D Car - - PowerPoint PPT Presentation

▶

Jul 19, 2023 330 likes •575 views

Overview Yandex Services Car Detection Yandex.Taxi 3D Car Detection Yandex Technologies Windup SDC Yandex Taxi Upcoming Plans Algorithms Design Evolvement since1989 search engine technology 1997 yandex.ru

SLIDE 1

SLIDE 2

Overview

› Yandex Services › Yandex.Taxi › Yandex Technologies › SDC Yandex Taxi › Algorithms Design Evolvement › Car Detection › 3D Car Detection › Windup › Upcoming Plans

SLIDE 3

world-class information search

n-demand transport services

navigation products mobile applications

since1989

search engine technology

1997

yandex.ru

SLIDE 4

Search Personal Geo Services Informational Media

Search Ad Tech Browser Kinopoisk Afisha Music TV Program Taxi Navi Maps Parking Traffic Transport Mail Disk Auto.ru Jobs Realty Travels Trains Flights Weather Health Market Zen Launcher

SLIDE 5

Yandex Taxi – one of the largest online taxi services in Russia & CIS

2–5 min

average waiting time

>200 000

active cars

6 countries 149

cities

SLIDE 6

Technologies

› Machine learning (MatrixNet) › Speech recognition › Language recognition and

synthesis (SpeechKit)

› Antivirus technologies › Computer vision › Duplicate image detection

› Behaviour analytics technology

(Crypta)

› Real-time bidding platform › Weather forecasting

technology (Meteum)

› Personal recommendations

technology (Disco)

SLIDE 7

SLIDE 8

Algorithms Design Evolvement

Data Deployment ML Labelling constant ongoing improvement process

SLIDE 9

Car Detection Pipeline

Data Deployment ML Labelling

SLIDE 10

Data for Car Detection

DATA DATA Collection

SLIDE 11

Labelling for Car Detection

Objects detection Labelling

SLIDE 12

SLIDE 13

ML for Сar Detection

ML SSD Modifications Yolo

SLIDE 14

Deployment for Car Detection

Deployment Tensor RT FuseNet Drive PX

SLIDE 15

3D Car Detection Pipeline

Data

Deployment ML Labelling Data Recollection Lidar Detection

SLIDE 16

Data for 3D Detection

Lidar Detection Lidar Point Cloud

SLIDE 17

ML for 3D Car Detection

Lidar Bird View (BV) ROI Pooling Concatenation Prior boxes classification Prior boxes coordinates regression Image (RGB)

SLIDE 18

Method Moderate Easy Hard Runtime MV3D (lidar

nly)

0,77 0,858 0,689 240ms

MV3D

0,769 0,86 0,684 360ms

Ours (bev only)

0,749 0,821 0, 736 95ms

SLIDE 19

Deployment for 3D Car Detection

Deployment Graph freezing DrivePX TensorRT

SLIDE 20

SLIDE 21

Wind-up

Data Labelling ML Deployment

Constant improvement process

›

image segmentation

›

depth estimation

›

end-to-end

SLIDE 22

Upcoming Plans

сity test-drives >=10 cars fleet growth 2018 more sensors => more experiments

SLIDE 23

+7 926 312-89-55 slesarev@yandex-team.ru

Thank you! Questions?

Anton Slesarev Lead computer vision scientist