Tensorflow Object Detection API: How to ignore regions during training?

Question

I'm using the object detection API from the models/research python repo on Ubuntu 16.04, and I wanted to fine-tune a pre-trained model (at the moment I'm interested in SSD with MobileNet or Inception backbones) on the UA-DETRAC dataset.

The problem is that there are specific regions, with their bounding boxes, which are marked as "ignored regions" and I wouldn't want the model to train on what he thinks are some false positives which are true, just not annotated (included in those regions). I thought of cropping the images to exclude those regions, but I would lose some information.

Is there the built-in possibility to mark them as "don't care" boxes or should I modify the code?

Thanks

Answer 1

If those regions to ignore remain static, as in, the contents of the region doesn't change throughout the dataset, then the model can be learnt to ignore those regions.

If you really want the model to ignore them during training, then mask them with a constant value.