Could we discuss why its recall is lower than d-fine's when dealing with fuzzy targets?

[tms2003]

However, EC's precision regarding targets is actually higher than d-fine's. I'm curious as to why it performs worse specifically on fuzzy targets.

[capsule2077]

Hi, thanks for the feedback. Could you provide more details regarding the "fuzzy targets"?

[tms2003]

for example,just like this。 only one bird(Don't care about that "Egret" tag; it's just an example.) can be detected.

[capsule2077]

Hi, I’ve looked at the sample images. Personally, I don't think this specific scenario should be overly challenging for the model to learn. Even a model pre-trained on the COCO dataset should be able to detect the "bird" class out of the box.

Could you double-check a few things to make sure the model has fully converged? Specifically, please ensure that there is a sufficient amount of training data, or try initializing with a COCO pre-trained checkpoint. It might also be worth reviewing your hyperparameters, such as the learning rate (LR).