While Using Peoplenet model for Transfer learning, got bad result

Docker_tag:–> v3.21.08-py3
Network Type → detectnet_v2
Configuration file → spec.txt (4.3 KB)

I am using poeplenet pretrained model for detecting people along with the phone i’e phone class samples with annotations I added in the dataset. I am using freeze blocks :4 . After training upto 200 epochs I got bad result in map as shown below :-

Result after training →

As I am using pretrained peoplenet model but this getting bad result, I am now able to understand where I am doing wrong.

I have also Attached the configuration file for your reference.Looking forward for any help from your side.

May I know

  • If your images have the same resolution? And what is the resolution?
  • Is it a public dataset for the phone and people? How many training images?

not same resolution,

for person image : -

person_image_detail

for phone image :-
phone_image_detail

Yes these are the public dataset images and annotations for phone dataset and person I have downloaded from kaggle.

For detectnet_v2, please make sure the image/labels are all with the same resolution.
You can resize the images offline and also resize labels accordingly.
More, could you share the Kaggle link? Thanks.

resize of images and labels accordingly completed. I have set 416*416. Do i need to change in the training spec file too?. I have attached the spec file in the previous comment for your reference.

Need to add

enable_auto_resize: True

can u please tell exactly where I need to put this enable_auto_resize: True. as I put this

here :-

augmentation_config {
enable_auto_resize: True
preprocessing {
output_image_width: 960
output_image_height: 544
output_image_channel: 3
min_bbox_width: 1.0
min_bbox_height: 1.0
}

but facing this issue : Message type “AugmentationConfig” has no field named “enable_auto_resize”.

Refer to DetectNet_v2 — TAO Toolkit 3.21.11 documentation

The enable_auto_resize: True should be inside preprocessing.

Still after training upto 200 epochs ,Now i am getting below result :-

not able to understand the map of person dataset, As i am using peoplenet pretrained model.

Could you upload your latest training spec , training log and an example of training image for person/phone ? Thanks.

training spec → spec.txt (4.3 KB)

training log → status.txt (54.0 KB)

example of training image →

person image →
000000001308

phone image →
phones104

How many images in your training dataset?

In training dataset there are total 316 images out of which people and phone images are 158 and in validation there are total 80 images, 40 for each class.

Can you share the link in kaggle? I need to try to reproduce.

for phone dataset → Mobile Images Dataset | Kaggle

for person I have extracted person samples along with there annotations from coco dataset.

The COCO dataset are quite different from the training dataset of peoplenet. Could you share the 158 training images along with their labels?

the trainning dataset should keep similar(background, shape, light and etc.) to the ones in inference, otherwise, the precision would be low.

I have already shared u the link of kaggle from where I have downloaded mobile dataset and for person I have extracted from coco dataset

OK, I will run some experiments against the person images from COCO.
Since the COCO dataset are quite different from the training dataset of peoplenet, it is expected that peoplenet pretrained model cannot get a good inference result against COCO images.

so is there any other way to do this. or u r going to try first?