Over the past few months, we’ve seen a number of questions and issues from users trying to evaluate model accuracy after compiling to a HEF file, especially with YOLO models.
Some of the common pitfalls include:
-
Using BGR instead of RGB input ordering
-
Incorrectly mapping class IDs to labels
-
Not setting NMS or confidence thresholds properly
All of these can lead to misleading mAP results and make it difficult to verify if a model is behaving correctly after compilation.
To help address this, we’ve published a guide that outlines a tool we developed to simplify and standardize the evaluation process. The tool makes it easier to correctly evaluate mAP after compilation and ensures consistency across experiments.
You can read the full guide here: Evaluating Model Accuracy After Compilation
We hope this helps streamline the workflow for the community and look forward to your feedback!