A new audit of seven feed-forward 3D reconstruction models reveals that their per-pixel confidence scores, intended to signal reliability, are not accurate predictors of error. While the confidence scores effectively rank errors, the predicted uncertainty is consistently too low, especially when models are used under conditions different from their training data. Researchers developed a power law correction that improves the overall magnitude of uncertainty but does not address the fundamental issue of scale across predictions, suggesting a need for models to better understand scene-level context. AI
IMPACT This audit highlights a critical flaw in how current 3D reconstruction models report confidence, potentially impacting their reliability in real-world applications and guiding future research towards more accurate uncertainty estimation.
RANK_REASON The cluster contains an academic paper detailing a new audit protocol and findings on existing models. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
