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

• DOI: 10.48550/arxiv.2312.06106
• Publication Date: 2023-01-01
• Authors: Prithvijit Chatto...; Bharat Goyal; Boglarka Ecsedi; Viraj Prabhu; Judy Hoffman
• Keywords: Overconfidence effect; Synthetic data

Synthetic data (SIM) drawn from simulators have emerged as a popular alternative for training models where acquiring annotated real-world images is difficult. However, transferring trained on synthetic to applications can be challenging due appearance disparities. A commonly employed solution counter this SIM2REAL gap unsupervised domain adaptation, are using labeled SIM and unlabeled REAL data. Mispredictions made by such adapted often associated with miscalibration - stemming overconfident predictions real In paper, we introduce AUGCAL, simple training-time patch adaptation that improves (1) reducing overall miscalibration, (2) overconfidence in incorrect (3) improving confidence score reliability better guiding misclassification detection all while retaining or performance. Given base algorithm, at time, AUGCAL involves replacing vanilla strongly augmented views (AUG intervention) additionally optimizing time calibration loss (CAL intervention). We motivate brief analytical justification of how reduce Through our experiments, empirically show the efficacy across multiple methods, backbones, tasks shifts.