Computing confidence mask

[alexandru-dinu]

In build_targets function, at the beginning, there's a part that calculates the confidence mask tensor.
Initially it is set to a tensor of ones, but the update rule:

    # Objects with higher confidence than threshold are set to zero
    conf_mask[b][cur_ious.view_as(conf_mask[b]) > ignore_thres] = 0

doesn't make sense to me. This basically ignores any ious better than ignore_thres (currently set to 0.5).

I'd think that:

• either start with a tensor of zeros and use the update rule:
  conf_mask[b][cur_ious.view_as(conf_mask[b]) > thres] = 1
• either change the sign: conf_mask[b][cur_ious.view_as(conf_mask[b]) <= thres] = 0

Thanks

[ludovic-carre]

I found the part you are refering to in the commit history but I am not sure why this filtering was removed. Can you explain @eriklindernoren ?

[alexandru-dinu]

I guess because it was redundant / wrong. The mask is computed as:
mask[b, best_anchor_idx, gj, gi] = 1, basically saying that for batch b, the anchor best_anchor_idx of cell (gj, gi) is responsible for predicting some ground truth object.

[eriklindernoren]

This should be fixed in 959e0ff. Before this commit the confidence mask was computed as in YOLOv2. With the latest commit the confidence mask is set as in version 3, i.e. where the anchor has an overlap with a target box larger than 0.5 but is not the anchor with the largest overlap the confidence prediction is ignored during training.