Dataset: Posetrack2018
Pretrained Backbone: MMPose HRNet (heatmap-based) mAP in PCKh is 84.5
Classifier loss weight (same as original) : token_loss=1.0, joint_loss=1.0
Tokenizer loss weight (same as original) : joint_loss_w=1.0, e_loss_w=15.0,
In both stage I and stage II, the joints (keypoints) loss couldn't be stable.
In stage II, although the mAP could reach value 78.9, but the kpt_loss still get jitter. And stuck in 78.8 for a long time.
Dataset: Posetrack2018 Pretrained Backbone: MMPose HRNet (heatmap-based) mAP in PCKh is 84.5 Classifier loss weight (same as original) : token_loss=1.0, joint_loss=1.0 Tokenizer loss weight (same as original) : joint_loss_w=1.0, e_loss_w=15.0,
In both stage I and stage II, the joints (keypoints) loss couldn't be stable.
In stage II, although the mAP could reach value 78.9, but the
kpt_loss
still get jitter. And stuck in 78.8 for a long time.Also in stage I, we got quiet well accuracy 99.6, but still .... unstable
joint_loss
we got.Has anyone run into the same problem? Or how could I fix it ?