PeopleNet. Coverage output is always zero

[sby089]

Please provide the following information when requesting support.

• Hardware TITAN Xp
• Network Type Detectnet_v2
• TLT Version
Configuration of the TAO Toolkit Instance
dockers: [‘nvidia/tao/tao-toolkit-tf’, ‘nvidia/tao/tao-toolkit-pyt’, ‘nvidia/tao/tao-toolkit-lm’]
format_version: 2.0
toolkit_version: 3.22.02
published_date: 02/28/2022
• Training spec file
train sepc file:
peoplenet_spec.txt (3.2 KB)
tfrecods log file:
tfrecords.log (57.6 KB)
training log file:
train.log (334.4 KB)

Hi, i’m trying to use peoplenet with pretrained model resnet18_peoplenet.tlt to train on custom dataset, i map the mask and non-mask to face class, but the average percision is always 0.
Hope someone help me,very thanks

[Morganh]

There is no “face” class in your label file.
So, please modify below

target_class_mapping: {
key: “mask”
value: “face”
}
target_class_mapping: {
key: “non-mask”
value: “face”
}

to

target_class_mapping: {
key: “mask”
value: “mask”
}
target_class_mapping: {
key: “non-mask”
value: “non-mask”
}

[sby089]

@Morganh Thinks for your replay my issue.
your means , if i want to detect the face class i shuold modify the dataset’ label file from

to

or modify the mentioned aborve the target class map

[Morganh]

Yes, the value in target_class_mapping should not be a virtual class.

[sby089]

@Morganh Tankyou very much!
I have a little question about the relationship between enable_auto _resize and output _image _height and output _image _weight.
In this experiment, i manually preprocess the dataset image size to 544x960x3(HWC), the original image is 1296x1296x3, if i want to use enable_auto_resize param, the outpu_image_height and output_image_weight how to set? is 544x960 or 1296x1296

[Morganh]

For detectnet_v2 network, firstly, it is necessary to resize images/labels to the same resolution offline. You already did this. You mention that you resize images/labels to 960x544(WxH)

So, if you want to train a 960x544 model(set output _image _weight:960 and set output _image _height: 544) , you need not set enable_auto_resize because your images/labels are already 960x544.

If you want to train a 640x480 model(set output _image _weight:640 and set output _image _height:480), you need to set enable_auto_resize:true because your images/labels are not 640x480.

[sby089]

Thanks again!