IC-1467DECAF achieves PVE 9.65 and f-score 89.6 on the DECAF validation set with a runtime of 19.59 seconds per image

Qingxuan Wu, Zhiyang Dou, Sirui Xu, Soshi Shimada, Chen Wang, Zhengming Yu, Yuan Liu, Cheng Lin, Zeyu Cao, Taku Komura, Vladislav Golyanik, Christian Theobalt, Wenping Wang, Lingjie Liu

SourceDICE: End-to-end Deformation Capture of Hand-Face Interactions from a Single Image

The paper reports DECAF's quantitative performance on the DECAF validation set. DECAF uses an optimization-based fitting process that takes 19.59 seconds per image on an A6000 GPU, roughly 200x slower than DICE. It achieves the best per-vertex error (9.65) and the highest overall physical plausibility f-score (89.6) among all methods tested. For contact estimation, DECAF achieves face f-score 0.57 and hand f-score 0.47. The paper notes that DECAF requires temporal information from successive frames and 3D ground-truth annotations for training, limiting its generalization to in-the-wild data.

Evidence

correlational

Key metric

PVE 9.65, col. dist. 1.03, non. col. 83.6, touchness 96.6, f-score 89.6, runtime 19.59 s; contact face f-score 0.57, hand f-score 0.47

Caveat

DECAF requires temporal information from successive frames and 3D ground-truth annotations for training; the paper notes limited generalization to in-the-wild data due to reliance on studio-collected 3D annotations

Model

DECAF

Datasets

DECAF [eval]

Methods

DECAF [primary]

Related findings

IC-1468

Extraction

automatic-extraction