Evaluate AVIS under distributional shifts

Investigate the performance of Activation Variance Informative Sampling (AVIS) for INT8 post-training calibration when the deployment environment contains unfamiliar terrain types that are absent from the calibration dataset.

Background

Activation Variance Informative Sampling (AVIS) selects calibration images using activation-variance statistics from a representative offline dataset. The paper evaluates AVIS using lunar-representative imagery, but does not establish whether its calibration benefits persist when the rover encounters terrain distributions that were not represented during calibration. Determining performance under such distributional shifts is important for assessing the reliability of quantized instance segmentation throughout a lunar mission.