CV veteran Michael Black: ECCV papers are two years stale — academia must rethink how it works

Michael_J_Black · x · 2026-09-09

Computer vision researcher Michael J. Black (attending ECCV since 1992) reflects on academia's predicament in the era of frontier models. His ECCV paper VIGA, an agentic method converting images into 3D Blender scenes, was rejected at first submission, delayed, then surpassed first by people using Claude Code for the same task and now by GPT-6 Astra — obsolete by the time it was presented.

His argument:

Published ideas rest on literature at least a year old, so conference papers are effectively two years behind — enough to make work irrelevant in today's AI.
Many CVPR authors still pursue problems based on outdated assumptions about how vision will be "solved," or niche problems large models skip for lack of data or business interest. The impactful papers came from industry with huge author lists and compute, serving mostly as documentation of commercial systems.

His prescriptions for academics: start every project by rigorously testing existing models and analyzing why they fail; only then pursue insights fundamental enough to outlive the next model release. And reviewers should judge papers on novel insight, not novel technical contribution — otherwise people just tweak architectures.

Original post →

CV veteran Michael Black: ECCV papers are two years stale — academia must rethink how it works

More from AGI Musings