Understand the `average_precision` in detectnet_v2 retraining

[music1913]

Please provide the following information when requesting support.

• Hardware (T4/V100/Xavier/Nano/etc)
x64, Ubuntu, RTX3090
• Network Type (Detectnet_v2)
• TLT Version (Please run “tlt info --verbose” and share “docker_tag” here)
• Training spec file(If have, please share here)
• How to reproduce the issue ? (This is for errors. Please share the command line and the detailed log here.)

I’m retraining based on detectnet_v2, my proprietary dataset contains only 1000 images which annotated with 3 classes of objects.
I noticed the average_precision can reach over 80%, even 90% for each class during the training.
I viewed some industry object detection performance/accuracy chart from web, seems they’re far below to this value (no matter which precision calculation algorithm).

COCO mAP:

PASCAL VOC

I looked the configuring-the-evaluator , so I believe the default using:

the ap calculation mode as used in the 2011 challenge

could you explain for the such a high precision, does is caused by small dataset?

[Morganh]

@music1913

could you explain for the such a high precision, does is caused by small dataset?

Please note that different public dataset has different SOTA.
And also your dataset may have a different mAP result from other dataset.