Details for Multi-HMR SMPL-X

AuthorsBaradel*, Fabien; Armando, Matthieu; Galaaoui, Salma; Brégier, Romain; Weinzaepfel, Philippe; Rogez, Grégory; Lucas*, Thomas
Paper titleMulti-HMR: Multi-Person Whole-Body Human Mesh Recovery in a Single Shot
Publication venuearXiv
URLhttps://github.com/naver/multi-hmr
Additional Information Single shot method (detection & regression) taking as input a RGB image at resolution 1288x1288. A ViT-L/14 is used to encode the image, then we use our HPH module for regressing human properties.

Metrics

Full setKid
F1 Score0.900F1 Score0.660
Precision0.970Precision0.570
Recall0.840Recall0.780
Body
Full setKid
MPJPE61.900MPJPE77.100
MVE61.300MVE77.100
NMJE68.800NMJE116.800
NMVE68.100NMVE116.800
Face
Full setKid
MPJPE26.000MPJPE25.700
MVE25.100MVE24.600
Full body
Full setKid
MPJPE93.200MPJPE106.800
MVE93.600MVE108.200
NMJE103.600NMJE118.700
NMVE104.000NMVE120.200
Hands
Full setKid
left hand MPJPE33.700left hand MPJPE30.200
left hand MVE35.700left hand MVE32.800
right hand MPJPE34.200right hand MPJPE33.300
right hand MVE36.100right hand MVE36.000