Retraining peoplenet model with own images

Thanks I’ll send you trained images and labels files from private message.

According previous experiments mentioned above, the training dataset is different from test image.
In other words, the train dataset and test dataset do not come from the same data distribution.
You can search some info via keyword “deep learning data distribution” or something others.
For example, https://www.kdnuggets.com/2019/01/when-your-training-testing-data-different-distributions.html

My suggestion:
Method 1), Adjust your label. The (x1,y1,x2,y2) is expected to be the coordinate of one person instead of the rectangle. Next, place the full body cropped images into the background image with high overlapping.

Method 2), Manual label parts of your test images, and use them to run training.

In a word, the trian/test datasets are expected to come from the same data distribution.

Thank you. Let me test.