Үүнийг мессежээр илгээх: Effective representations for road scene understanding with scalable learning