Blake

Question on the way implicit objectness is combined with class logits

1

https://github.com/chensnathan/YOLOF/blob/a44b7af0efa9b89d3e91db70bc9a39e49d5d0144/yolof/modeling/decoder.py#L105 Could you please explain why implicit objectness is combined with class logits like this? Thanks.

The full version of ATSS applied to FCOS

3

The paper mentioned a full version of ATSS applied to FCOS, is it obtained by changing REGRESSION_TYPE from BBOX to POINT in atss_R_50_FPN_1x.yaml?

Is it easy to split inference code from training code?

1

Has anybody tried this and how hard is it to split, may I ask?