[SimCC Label Confidence Score is Low for COCO_Val Evaluation]

[T-S-Liang]

Hi guys,

I'm currently testing an RTMPose-like model using a SimCC-style encoder/decoder, and I’ve been puzzled by the confidence scores produced by SimCC.

During training, I’m using the standard KLDiscreteLoss, and my labels are unnormalized Gaussian distributions. During inference, I apply a softmax to the output logits from SimCC, then pass the result to get_simcc_maximum to extract both the predicted coordinates (via the max indices) and the confidence scores (via the max values).

However, I’ve noticed that the maximum values after the softmax are quite low—usually around 0.02 to 0.05—even though the visualization results look pretty good.

I have two questions:

During inference, is it appropriate to apply a softmax to the output logits?
How should I align or interpret the confidence scores from the model in relation to COCO-style evaluation metrics?
Thanks so much in advance!