Why some class of custom dataset for yolov4 give me zero mAP?

[elyaabbas9]

Hi, Hope you all doing good.

I have changed my dataset labels for my previous version. so new data in training folder is in this format.

we have label 1 image twice but

1. we labeled 3 objects in one image
2. Again we use the same image for the other two object labels only no
   other objects are labeled

• The first image has 3 labeled objects and unlabeled persons objects.
• Second image show, has 2 objects labeled and remains other unlabeled.

3 class object pic but not label persons

2 class object labeled only other object is ideal.

All this happened because we were changing labels for the person objects so we just filter those images that have persons then we only label person_with_helmet and prson_without_helmet.

and the previous data now has 3 objects labeled in every picture and not labeled the persons objects.

so we combine images and kitti labels for 3 objects images and 2 objects images.
after training on TAO tool we got the result shown as under.

*******************************
invalid_balconyAP    0.0
person_with_helmetAP    0.0
person_without_helmetAP    0.42694
rail          AP    0.5299
valid_balcony AP    0.35055
              mAP   0.26148
*******************************

Is the problem because we label 3 objects in one image and unlabeled 2 objects in the same image and again in another picture we label 2 objects and the other 3 objects remain unlabeled??

[Fiona.Chen]

How did you train your models? Are you working with TAO toolkit?

[elyaabbas9]

yes, I have already mentioned the TAO tool kit.

[Fiona.Chen]

There is TAO forum. Latest Intelligent Video Analytics/TAO Toolkit topics - NVIDIA Developer Forums

Seems it is not DeepStream related. Please raise the topic to TAO forum.