Details for NLF-L nonparam. SMPL-X

AuthorsIstvan Sarandi, Gerard Pons-Moll
Paper titleNeural Localizer Fields for Continuous 3D Human Pose and Shape Estimation
Publication venueNeurIPS
URLhttps://arxiv.org/abs/2407.07532
Additional Information The model was trained on a large collection of datasets including SMPL(-X/+H) parametric body, 2D and 3D skeleton and DensePose annotations. After that, it was fine-tuned on AGORA. YOLOv8x is used as the detector, also fine-tuned on AGORA.

Metrics

Full setKid
F1 Score0.940F1 Score0.710
Precision0.970Precision0.600
Recall0.910Recall0.880
Body
Full setKid
MPJPE58.200MPJPE56.500
MVE58.400MVE56.100
NMJE61.900NMJE79.600
NMVE62.100NMVE79.000
Face
Full setKid
MPJPE28.500MPJPE21.500
MVE27.000MVE21.400
Full body
Full setKid
MPJPE90.800MPJPE81.400
MVE92.700MVE81.900
NMJE96.600NMJE86.600
NMVE98.600NMVE87.100
Hands
Full setKid
left hand MPJPE34.400left hand MPJPE25.900
left hand MVE37.900left hand MVE27.200
right hand MPJPE34.900right hand MPJPE27.400
right hand MVE38.100right hand MVE28.700