Linking vision and motion for self-supervised object-centric perception

TL;DR

This paper presents a self-supervised vision model that learns object-centric representations from RGB video and vehicle pose, enabling tracking and scene understanding for autonomous driving without relying on labeled data.

Contribution

It adapts a self-supervised object-centric vision model for autonomous driving, demonstrating effective object decomposition and multi-view fusion using only RGB video and pose information.

Findings

Achieves promising object tracking results on Waymo dataset

Fuses multiple camera viewpoints over time effectively

Lags behind supervised methods in mask quality

Abstract

Object-centric representations enable autonomous driving algorithms to reason about interactions between many independent agents and scene features. Traditionally these representations have been obtained via supervised learning, but this decouples perception from the downstream driving task and could harm generalization. In this work we adapt a self-supervised object-centric vision model to perform object decomposition using only RGB video and the pose of the vehicle as inputs. We demonstrate that our method obtains promising results on the Waymo Open perception dataset. While object mask quality lags behind supervised methods or alternatives that use more privileged information, we find that our model is capable of learning a representation that fuses multiple camera viewpoints over time and successfully tracks many vehicles and pedestrians in the dataset. Code for our model is available at https://github.com/wayveai/SOCS.