Beyond Benchmark Accuracy:
A Hands-On Robust Vision Tutorial
ACCV 2026 Tutorial · December 14, 2026 · Osaka, Japan

Tutorial Overview
Modern vision systems move beyond curated benchmarks and operate in the wild, where inputs in the real world are noisy, blurred, rescaled, occluded, domain-shifted, or paired with ambiguous prompts. A model can post high average accuracy on a benchmark yet fail silently the moment conditions change. This half-day, fully on-site tutorial teaches robustness as a practical engineering discipline through one repeatable workflow: diagnose failures with standardized corruption, shift, and worst-case probes; understand the architectural and training factors behind them; harden the model with lightweight design-level fixes plus an out-of-distribution abstention gate; and validate the result with adaptive attacks and effective-robustness reporting. The workflow is practiced end to end on a vision transformer and then replayed compactly on an open vision-language foundation model using two no-install Colab notebooks, with slides, checkpoints, and a benchmark harness released openly.

Keywords
Robust computer vision, Distribution shift, Vision transformers, Vision-language models, Robustness evaluation

Schedule
Afternoon Session · Room 1009
Half-day · Fully on-site · Hands-on Colab

Who Should Attend
The tutorial mainly targets graduate students and practitioners who want to understand failures outside benchmark settings, especially when deploying or evaluating vision models in quality-sensitive areas such as autonomous driving, industrial inspection, medical imaging, and VLM products. No prior background in adversarial machine learning is required; attendees should be comfortable reading PyTorch-style code, but core ideas are explained visually.

Materials
Coming soon

Acknowledgements