SelfOccFlow: Towards end-to-end self-supervised 3D Occupancy Flow prediction

TL;DR

This paper introduces SelfOccFlow, a self-supervised approach for 3D occupancy flow prediction in autonomous driving, eliminating the need for annotated data and external flow supervision by disentangling static and dynamic scene components.

Contribution

It presents a novel self-supervised framework that learns 3D occupancy flow by disentangling static and dynamic elements and leveraging feature similarity cues, without relying on annotated datasets.

Findings

Effective on SemanticKITTI, KITTI-MOT, and nuScenes datasets.

Outperforms methods requiring annotated data or external supervision.

Demonstrates accurate 3D occupancy and motion estimation in dynamic scenes.

Abstract

Estimating 3D occupancy and motion at the vehicle's surroundings is essential for autonomous driving, enabling situational awareness in dynamic environments. Existing approaches jointly learn geometry and motion but rely on expensive 3D occupancy and flow annotations, velocity labels from bounding boxes, or pretrained optical flow models. We propose a self-supervised method for 3D occupancy flow estimation that eliminates the need for human-produced annotations or external flow supervision. Our method disentangles the scene into separate static and dynamic signed distance fields and learns motion implicitly through temporal aggregation. Additionally, we introduce a strong self-supervised flow cue derived from features' cosine similarities. We demonstrate the efficacy of our 3D occupancy flow method on SemanticKITTI, KITTI-MOT, and nuScenes.