tao-analyze-gaps-visual-changenet

Gap analysis identifies misclassified samples by analyzing classification results from the NVIDIA TAO VCN Classify model, ranking samples by distance from the decision threshold to highlight weak data for augmentation.

A threshold sweep determines the optimal decision threshold for VCN classification by evaluating model performance across varying values, isolating misclassified samples for downstream relabeling or augmentation.

Yes, gap analysis requires Docker with NVIDIA Container Toolkit and GPU support. It utilizes the tao_toolkit.data_services container for processing and requires pynvml for GPU detection.

Weakness ranking scores misclassified samples based on their distance from the decision threshold. The lowest-scoring samples are generated as top-K candidates for targeted data augmentation.

Yes, after training a model, use this gap analysis to identify misclassified image samples that require relabeling or augmentation to improve overall model performance.

VCN gap analysis requires pynvml for GPU detection to ensure the Docker environment properly allocates GPU resources for the tao_toolkit.data_services container during image classification processing.