Wrong result after retraining peopleNet on custom dataset

Dear Nvidia Team,
I have trained my model on PeopleNet unpruned model around 1000 images(person,face,bag) but I am not getting good result on those images which was detecting good on PeopleNet before training.
Like :

Re-trained PeopleNet (On Unseen Image):

Re-trained PeopleNet (on seen/training data image):

PeopleNet model provided by Nvidia:

So I am getting good result on a single image on peopleNet, but on the same image I am getting bad result after retraining the peopleNet on my custom dataset. but getting good result on my training data. what is the reason ? should I also include peopleNet training Data to get result like peopleNet.

I’m a little confused.
You mentioned that

I am getting bad result after retraining the peopleNet on my custom dataset. but getting good result on my training data

So, is it good or bad?

Hi Morganh, Thanks for replying.

Sorry for the confusion. PeopleNet is giving good results on the images which were part of the re-training data (as expected). But on the same unseen image (the third image in the original post, where the unpruned model provided by Nvidia was performing good) the re-trained model was not able to detect anything as you can see (the first image in the original post).

Could you please share your training spec and training log?
In the training log, did you see any mAP result?

There is no update from you for a period, assuming this is not an issue any more.
Hence we are closing this topic. If need further support, please open a new one.
Thanks

More, how did you evaluate " PeopleNet is giving good results on the images which were part of the re-training data (as expected)"? Using tlt-evaluate?

Hi,
I have re-trained peoplenet with my custom dataset, however when I do tlt-infer I get wrong results i.e the chair as a person, in addition to detecting person as person
How can I avoid the wrong detections ?
Is there similar negative images concept like in yolo?.
This is a folder of images where you inform yolo to consider as negative images, and that minimises the wrong detections.
Thanks

@georgesmarkuswaevu
Please create a new topic about your question. Thanks.