ST-STORM Framework Introduces Appearance as Semantic Signal in Self-Supervised Learning

A new self-supervised learning framework called ST-STORM (Stylistic-STORM) has been introduced to address limitations in current approaches. While methods like MoCo and DINO focus on creating representations that ignore appearance variations like lighting or geometry, this strategy becomes problematic when appearance itself contains critical information. In fields such as weather analysis, visual elements like rain streaks, snow granularity, atmospheric scattering, reflections, and halos are not noise but essential discriminative signals. For safety-critical applications including autonomous driving, ignoring these appearance cues poses risks since ground conditions and atmospheric visibility directly affect vehicle grip and safety. The hybrid SSL framework treats appearance (style) as a semantic modality that should be disentangled from content rather than discarded. This approach recognizes that in many real-world scenarios, appearance variations carry meaningful information that should be preserved rather than suppressed. The research was published on arXiv under identifier 2604.16086v1 with an announcement type of cross.