OccFeat: Self-supervised Occupancy Feature Prediction for Pretraining BEV Segmentation Networks

TL;DR

OccFeat is a self-supervised pretraining approach for BEV segmentation networks that combines occupancy prediction with semantic feature distillation, enhancing scene understanding especially with limited labeled data.

Contribution

The paper introduces a novel pretraining method that integrates 3D occupancy prediction and feature distillation, improving BEV segmentation performance.

Findings

Improved BEV segmentation accuracy, especially in low-data settings.

Effective integration of occupancy prediction with feature distillation.

Enhanced 3D scene understanding through self-supervised pretraining.

Abstract

We introduce a self-supervised pretraining method, called OccFeat, for camera-only Bird's-Eye-View (BEV) segmentation networks. With OccFeat, we pretrain a BEV network via occupancy prediction and feature distillation tasks. Occupancy prediction provides a 3D geometric understanding of the scene to the model. However, the geometry learned is class-agnostic. Hence, we add semantic information to the model in the 3D space through distillation from a self-supervised pretrained image foundation model. Models pretrained with our method exhibit improved BEV semantic segmentation performance, particularly in low-data scenarios. Moreover, empirical results affirm the efficacy of integrating feature distillation with 3D occupancy prediction in our pretraining approach. Repository: https://github.com/valeoai/Occfeat