Question

[Divadi]

Hello - I just had a short question. For the 100% setting, how is Efficient-CLS applied? As in, what is pseudo-labeled since there are no unlabeled frames?

[zhangjiewu]

Hi @Divadi, sorry for the delay. For the 100% setting, we use only the EMA branch and do not utilize pseudo labels since it's fully supervised.

[Divadi]

@zhangjiewu Thank you for your prompt response!
To ask - is "offline training" also done with one epoch total as well? I've been trying to train Faster RCNN on Wanderlust, but I cannot seem to get anywhere close to 48 AP50 with just one epoch, shuffled or not shuffled.

Edit for details:
I have tried Faster-RCNN + R50 + COCO pre-training with AdamW optimizer, EMA, 1-pass, to get 36.4 AP50. VOC pretrained is much worse.
I have also tried the above, except sampling 2 batches per "step" to get 45.2 AP50 which is much closer, so I am wondering if >1 batch was used per step (I see that original Wanderlust used on average 10 batches)

[zhangjiewu]

Hi @Divadi, sorry I missed this message.

No, the offline training allows unlimited access to previous data and multiple-epoch training.