MC-PanDA: Mask Confidence for Panoptic Domain Adaptation

TL;DR

This paper introduces MC-PanDA, a novel method for panoptic domain adaptation that uses mask confidence to improve segmentation accuracy by selectively guiding learning based on prediction certainty.

Contribution

It leverages mask transformers' ability to estimate prediction uncertainty, improving domain adaptation performance without noise amplification.

Findings

Achieves 47.4 PQ on Synthia to Cityscapes, surpassing previous methods by 6.2 points.

Utilizes mask-wide confidence to modulate loss and reduce noise in training.

Demonstrates significant improvements on standard benchmarks.

Abstract

Domain adaptive panoptic segmentation promises to resolve the long tail of corner cases in natural scene understanding. Previous state of the art addresses this problem with cross-task consistency, careful system-level optimization and heuristic improvement of teacher predictions. In contrast, we propose to build upon remarkable capability of mask transformers to estimate their own prediction uncertainty. Our method avoids noise amplification by leveraging fine-grained confidence of panoptic teacher predictions. In particular, we modulate the loss with mask-wide confidence and discourage back-propagation in pixels with uncertain teacher or confident student. Experimental evaluation on standard benchmarks reveals a substantial contribution of the proposed selection techniques. We report 47.4 PQ on Synthia to Cityscapes, which corresponds to an improvement of 6.2 percentage points over the state of the art. The source code is available at https://github.com/helen1c/MC-PanDA.