Open gsygsy96 opened 3 years ago
As described in our paper, following DenseBody, we evaluate our method on the SURREAL test set, which consists of 493 samples. You can find this selected part can at https://github.com/gulvarol/bodynet/issues/5. We report surface and joint errors in our paper. The surface error is calculated as per-vertex L2 distance. The joint error is obtained by computing the per-joint L2 distance. Both of them have the following calculating formulation np.average(np.sqrt(np.sum(np.square(pred_joints- gt_joints),axis=-1))) .
I see, and the checkpoint is the one you post in the readme?
The checkpoint linked in the readme is for the 3D human pose only. We have not released the 3D human body estimation models yet.
Hi @redrock303, Thanks for your work! Could you please release the evaluation script on the SURREAL?