Values in detectnetv2 config files use by tao

[user73736]

This is a general question to understand how one goes about customizing the “canned” config files that are shipped with TAO… I’ve been studying the config files that have been provided with TAO and am trying to modify it for training custom data. I find that a lot of numbers seem to be computer generated - for example detectnet_v2/specs/detectnet_v2_train_resnet18_kitti.txt - has values such as
coverage_threshold: 0.00499999988824
dbscan_eps: 0.20000000298
dbscan_min_samples: 0.0500000007451
…
minimum_detection_ground_truth_overlap {
key: “car”
value: 0.699999988079
}

coverage_foreground_weight: 0.0500000007451

etc…

Just wondering if someone could enlighten me on the methodology to use to tune these numbers to such a great degree of accuracy… Looks like there are some tools that generate these numbers…

Thanks, nanu

[Morganh]

Please ignore the value with such a great degree of accuracy. You can refer to DetectNet_v2 — TAO Toolkit 3.22.05 documentation.

For tuning some parameters, you can have a look at Frequently Asked Questions — TAO Toolkit 3.22.05 documentation