class_loss_weights for Object Detection

[AashishKumbashi]

Since class_loss_weights is not supported for object detection tasks by rastervision, is there any other way in which I can handle class imbalance problem for object detection?

[AdeelH]

If you have just one target class, you can use the neg_ratio argument in ObjectDetectionRandomWindowGeoDataset.

There isn't a built-in mechanism for dealing with multiple classes. You might consider subclassing and modifying ObjectDetectionRandomWindowGeoDataset to accept a list of per-class sampling probabilities instead of neg_ratio and sample according to those.