Evaluation study / Model robustness
Overview
InspecSpider inspects structures at height. Tests with simulated fog exposed weaknesses in the vision model's segmentation, despite otherwise stable performance.
Method
The team used test logs to pinpoint failures, then added difficult examples to the training data.
Findings
After these changes, robustness improved by 23.7% in the tested scenario.
Scope & limitations
This record concerns vision segmentation under simulated fog and a comparison before and after training-data changes. The available summary does not report the robustness metric definition, test-set size, or variability. The stated 23.7% improvement should be interpreted within that scope.
Data & evidence
This page preserves the published project summary. Raw data, individual test results, and reproducible code are not supplied with this record. Contact HephaKnot with questions about the evidence or method.