More than the Sum: Panorama-Language Models for Adverse Omni-Scenes

TL;DR

This paper introduces a panoramic vision-language reasoning paradigm and dataset, enabling existing models to better understand 360-degree scenes with occlusions and adverse conditions.

Contribution

It proposes the Panorama-Language Modeling (PLM) paradigm, a panoramic dataset PanoVQA, and a plug-and-play attention module for improved omni-scene understanding.

Findings

PLM achieves superior robustness in challenging omni-scenes.

The panoramic attention module enables existing models to process panoramas without retraining.

Extensive experiments validate the effectiveness of the proposed approach.

Abstract

Existing vision-language models (VLMs) are tailored for pinhole imagery, stitching multiple narrow field-of-view inputs to piece together a complete omni-scene understanding. Yet, such multi-view perception overlooks the holistic spatial and contextual relationships that a single panorama inherently preserves. In this work, we introduce the Panorama-Language Modeling (PLM)paradigm, a unified $360^\circ$ vision-language reasoning that is more than the sum of its pinhole counterparts. Besides, we present PanoVQA, a large-scale panoramic VQA dataset that involves adverse omni-scenes, enabling comprehensive reasoning under object occlusions and driving accidents. To establish a foundation for PLM, we develop a plug-and-play panoramic sparse attention module that allows existing pinhole-based VLMs to process equirectangular panoramas without retraining. Extensive experiments demonstrate that our PLM achieves superior robustness and holistic reasoning under challenging omni-scenes, yielding understanding greater than the sum of its narrow parts. Project page: https://github.com/InSAI-Lab/PanoVQA.