search

chenyilun95 / tf-cpn

Cascaded Pyramid Network for Multi-Person Pose Estimation (CVPR 2018)
MIT License
793 stars 197 forks source link

the influence of detector #35

Closed msunming closed 6 years ago

msunming commented 6 years ago

the performance of human AP from the 57.2 to 62.9 increase extra 1% in human pose, will you release the result of 62.9 human detection result for analyze the importance of detector?

chenyilun95 commented 6 years ago

The influence of detection has been analyzed in minival dataset in paper. And now I don’t have the detection result. I’m sorry.

msunming commented 6 years ago

3Q. the conclusion may seems conflict in paper about the influence of detector. Besides , the nms thr is set one?

chenyilun95 commented 6 years ago

The detector has a greater but still limited impact on test-dev dataset which we found later. Soft-nms is applied in the given detection result.

msunming commented 6 years ago

the resnet50 var name seems not completely same as tf website, the weight is trained by youself or transform from tf?

msunming commented 6 years ago

sorry, its weight same with tf resnet50.

msunming commented 6 years ago

In the report, the batchsize also is important. Do you try implement BN cross-gpus and related experiment ?