Abstract:
Quantifying deep neural classifier confidence is difficult but essential for safety-critical applications. This paper introduces the true class probability (TCP) as a model confidence target criterion. TCP outperforms maximum class probability (MCP) for confidence estimation. Since the true class is unknown at test time, we propose learning TCP criterion from data with an auxiliary model using a context-specific learning scheme. Our approach is tested on failure prediction and domain adaptation self-training with pseudo-labels, which require effective confidence estimates. Extensive experiments validate the proposed approach in each task. We test image classification and semantic segmentation network architectures on small and large datasets. We outperform strong baselines in every benchmark.
Note: Please discuss with our team before submitting this abstract to the college. This Abstract or Synopsis varies based on student project requirements.
Did you like this final year project?
To download this project Code with thesis report and project training... Click Here
