Train ssd-mobilenet: Add background images to custom dataset

[anasri89]

Hello,

I am currently trying the object detection training (ssd-mobilenet) from the hello ai world tutorial. I created my own dataset and i have added a few background images to the training dataset. Each background images has its own xml file and contain no label. Below is an example of the xml-file of the background image.

selected images prüfstand.jpg /selected images/prüfstand.jpg Unknown 4032 3024 3 0

When I executed the train_ssd.py, I noticed there were warnings about images that has no bounding box/labels annotations. The warnings look like these:
2023-10-26 12:32:47 - Prepare training datasets.
warning - image pruefstand has no box/labels annotations, ignoring from dataset
warning - image bilder00511 has no box/labels annotations, ignoring from dataset
2023-10-26 12:32:47 - VOC Labels read from file: (‘BACKGROUND’, ‘felge’)
2023-10-26 12:32:47 - Stored labels into file models/f_v2/labels.txt.
2023-10-26 12:32:47 - Train dataset size: 26

And I realized that the background images were not included in the training. The training itself works and didnt produce any errors.
My question is how can I add background images into the training. I read that background images can help an object detection model to reduce its false positive.

[dusty_nv]

Hi @anasri89, negative training examples are automatically extracted from the dataset by train_ssd.py in a balanced way, so you needn’t manually do it. You should probably add significantly more than 25 images in your dataset though.