ML Seminar Series

Language-Guided Instance-Aware Domain-Adaptive Panoptic Segmentation

by Dr Suman Saha (Senior Data Scientist, SDSC hub at PSI)

Europe/Zurich
OHSA/E13

OHSA/E13

Description

Abstract: 
The increasing relevance of panoptic segmentation is tied to the advancements in autonomous driving and AR/VR applications. However, the deployment of such models has been limited due to the expensive nature of dense data annotation, giving rise to unsupervised domain adaptation (UDA). A key challenge in panoptic UDA is reducing the domain gap between a labeled source and an unlabeled target domain while harmonizing the subtasks of semantic and instance segmentation to limit catastrophic interference. While considerable progress has been achieved, existing approaches mainly focus on the adaptation of semantic segmentation. In this work, we focus on incorporating instance-level adaptation via a novel instance-aware cross-domain mixing strategy IMix. IMix significantly enhances the panoptic quality by improving instance segmentation performance. Specifically, we propose inserting high confidence predicted instances from the target domain onto source images, retaining the exhaustiveness of the resulting pseudo-labels while reducing the injected confirmation bias. Nevertheless, such an enhancement comes at the cost of degraded semantic performance, attributed to catastrophic forgetting. To mitigate this issue, we regularize our semantic branch by employing CLIP-based domain alignment (CDA), exploiting the domain-robustness of natural language prompts. Finally, we present an end-to-end model incorporating these two mechanisms called LIDAPS, achieving state-of-the-art results on all popular panoptic UDA benchmarks.

 ZOOM ID: https://psich.zoom.us/j/61414084707

Organised by

The Laboratory for Simulation and Modelling
SDSC hub at PSI

Registration
Participants
Participants
  • Benjamín Béjar
  • Ibrahim El Kaddouri
  • Igor Malinovic
  • +3
Dr Benjamin Bejar Haro