Why do we need Annotation files for unlabelled dataset?

[rishabhsharma22]

In tools/train.py on line 175 you create a cfg_u for unlabelled files. This function uses X_U and all_anns.

Now all annotations contains the annotation files for both the labeled set and unlabeled set. Why do I need the annotations for the unlabelled set ? Isn't this against the idea of active learning that you dynamically create the labels for unlabelled set and use them to enhance your model ?

[yuantn]

You can have a look at the loss function in lines 479 and 565 in mmdet/models/dense_head/MIAOD_head.py, which are belongs to the step re-weighting and minimizing/maximizing instance uncertainty. In these two steps, if y_loc_img is negative, the corresponding prediction of the model will be set to 0, which prevent it from participating in network training. And if y_loc_img is negative, it means that the corresponding data is from the unlabeled set X_U, referring to the lines 74 and 92 in epoch_based_runner.py. The normal y_loc_img, which is the annotation of the bounding box, should be positive in the original dataset. Therefore, we didn’t use any annotation of the unlabeled set X_U in model training.