Research map / Label-efficient and robust learning / Domain adaptation and generalization

AUGCAL: Improving Sim2Real Adaptation by Uncertainty Calibration on Augmented Synthetic Images

ICLR 2024 · Label-efficient and robust learning › Domain adaptation and generalization · approach: Geometric alignment and synthetic-to-real adaptation · 1 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes AugCal, a training-time patch for sim-to-real unsupervised domain adaptation that combines strong augmentation (PASTA/RandAugment) of synthetic images with a DCA calibration loss to reduce miscalibration and overconfident mispredictions on real data while retaining adaptation performance.

Problem

Sim2Real adapted models (via unsupervised domain adaptation) make overconfident mispredictions on real data, leading to miscalibration and unreliable confidence scores, which matters for safety-critical deployment.

Approach

Replace synthetic training images with strongly augmented views that reduce the synthetic-to-real distribution distance (Aug), and add a training-time calibration loss (DCA: difference between confidence and accuracy) on augmented synthetic predictions (Cal); an importance-weighting analysis motivates that this tightens an upper bound on target calibration error.

Key results

• On GTAV to Cityscapes with DAFormer, EntMin + MIC improves from 65.71 to 70.31 mIoU with AugCal while ECE drops from 5.34 to 3.43 and PRR rises from 45.93 to 62.66.
• HRDA + MIC with AugCal reaches 75.90 mIoU vs 75.56 baseline, with ECE reduced from 2.86 to 2.45.
• On VisDA with ViT-B, SDAT + MIC with AugCal improves ECE from 7.67 to 6.84 and PRR from 63.78 to 67.80 while mAcc rises from 92.53 to 92.87.

Measured effects reported in the paper

Surprising findings

• Temperature scaling tuned on synthetic validation data increases overconfidence and miscalibration on real data.
• Overly high calibration loss weight (lambda >= 5 or 100) degrades adaptation performance substantially.

Limitations

• Gains in calibration are smaller for stronger adaptation methods such as HRDA + MIC.
• Validity of the Aug transform depends on an empirical MMD check using a trained model; analysis relies on covariate shift assumptions.
• Evaluated only on GTAV to Cityscapes and VisDA with a few UDA methods.

Experimental setup

Datasets: GTAV Cityscapes VisDA

Benchmarks: GTAV to Cityscapes VisDA

Metrics: mIoU mAcc ECE IC-ECE OC PRR

Baselines: EntMin + MIC HRDA + MIC SDAT + MIC Temperature scaling MDCA MbLS Source-only Pasta

Builds on: EntMin HRDA MIC SDAT PASTA DCA

Closest papers by meaning

• Pseudo-Calibration: Improving Predictive Uncertainty Estimation in Unsupervised Domain Adaptation (ICML 2024 · similarity 0.85)
• Ultra-Resolution Adaptation with Ease (ICML 2025 · similarity 0.84)
• AugGen: Synthetic Augmentation using Diffusion Models Can Improve Recognition (NeurIPS 2025 · similarity 0.84)
• LoRA-X: Bridging Foundation Models with Training-Free Cross-Model Adaptation (ICLR 2025 · similarity 0.84)
• Addressing Misspecification in Simulation-based Inference through Data-driven Calibration (ICML 2025 · similarity 0.84)
• The Unmet Promise of Synthetic Training Images: Using Retrieved Real Images Performs Better (NeurIPS 2024 · similarity 0.84)