search

NVIDIA-AI-IOT / deepstream_tao_apps

Sample apps to demonstrate how to deploy models trained with TAO on DeepStream
MIT License
377 stars 97 forks source link

Update post_processor to use per-class thresholds for filtering #79

Closed Nicholas-Kastanos closed 1 year ago

Nicholas-Kastanos commented 1 year ago

In the current version of the code, all classes are filtered based on the confidence set by the first class's attributes. This removes control from the end-user, and is falsely advertising the capabilities of the bbox parsers.

Additionally, the filtering is happening before clustering takes place, therefore the thresholding attribute is being changed from perClassThreshold to perClassPreclusterThreshold.

This makes all bbox parsers pre-processing consistent with NvDsInferParseCustomDDETRTAO.