Closed leonzgtee closed 4 years ago
@leonzgtee hi, i only train & test on ai-challenger dataset, until now, i get ai-valid dataset kps score 0.281 only just 10 epochs. Maybe its more hard for training on COCO dataset.
i have noted that the paper use centorid to represent the object instance instead of its center point. in which way did you implement this, i am curious.
I'm not 100% percent sure about if center point is equal with centroid, but just like in "Objects As Points" said, its center point can also be learned well, so i used object's center point. Maybe it is better way to ask author to make it clear.
leonzgtee notifications@github.com 于2019年10月18日周五 上午8:47写道:
i have noted that the paper use centorid to represent the object instance instead of its center point. in which way did you implement this, i am curious.
— You are receiving this because you commented. Reply to this email directly, view it on GitHub https://github.com/murdockhou/Single-Stage-Multi-person-Pose-Machines/issues/2?email_source=notifications&email_token=AEMCC7KO2WEYZXBKKSYKDUTQPEBQXA5CNFSM4JBV4BIKYY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOEBSBJNI#issuecomment-543429813, or unsubscribe https://github.com/notifications/unsubscribe-auth/AEMCC7MTYWZCXBNREXCGCI3QPEBQXANCNFSM4JBV4BIA .
thanks, i will send a email to the author to inquire about this.
@leonzgtee have you heard back about the centroid?
@wwu the author told me that he uses average of all visible keypoints as root joint.
@wwu the author told me that he uses average of all visible keypoints as root joint.
Thank you!
nice job, have you achieve a desirable result? i mean get a mAP around 0.6. i have tried, but could not achieve a good result.