Trying to implement transfer learning using peoplenet model

[user86169]

Docker_tag:–> v3.21.08-py3
Network Type → detectnet_v2
spec file → spec.txt (4.2 KB)

log file →
log file.txt (26.3 KB)

Hi,

I am trying to add people images along with the phone class and trying to implement transfer learning using peoplenet (resnet18_peoplenet.tlt) model file.

after freezing initial 3 blocks I getting low map for person but the phone map is good.

Phone map → 86
person map → 14

I am not able to understand why person map is not getting increase.I want to increase person class map

[Morganh]

Hi @user86169
As we synced previously, did you ever try to run with KITTI dataset?
May I know for your current result, which is the dataset and how many training dataset is it?

[user86169]

does kitti dataset have people class also?

[Morganh]

Yes, the public KITTI dataset has people class.

[user86169]

is there any way through which i can extract some images of people along with there annotations from kitti dataset?

[Morganh]

You can download the pubic dataset according to the guide in jupyter notebook https://catalog.ngc.nvidia.com/orgs/nvidia/teams/tao/resources/cv_samples/version/v1.3.0/files/detectnet_v2/detectnet_v2.ipynb#head-2

Then, search the lable files which contain pedestrian class.
Then, move all of their corresponding images from images folder.

[user86169]

but I want person class images not pedestrian class

[Morganh]

You can view the KITTI images to check if it meets your requirement.
In KITTI dataset, there are “pedestrian” and “person_sitting” .

So, in the detectnet_v2 spec files(TAO Toolkit Quick Start Guide — TAO Toolkit 3.22.05 documentation), it set below class_mapping.

target_class_mapping {
key: “pedestrian”
value: “pedestrian”
}
target_class_mapping {
key: “person_sitting”
value: “pedestrian”
}