Details for DIGIT-ResNet50

AuthorsZicong Fan, Adrian Spurr, Muhammed Kocabas, Siyu Tang, Michael J. Black, Otmar Hilliges
Paper titleLearning to Disambiguate Strongly Interacting Hands via Probabilistic Per-Pixel Part Segmentation
Publication venue3DV
Publication year2021
URLhttps://arxiv.org/pdf/2107.00434.pdf
Additional Information Adaptation of DIGIT to ARCTIC setting. Changes: - Estimate the object mask as well on top of the hand segmentation masks. - Keep the segmentation and fusion U-Net, but swap the pose regressor to ArcticNet-SF heads. - Use ResNet50 as backbone Code: https://github.com/XueYing126/arctic-digit/tree/Digit
Note
Use a single algorithm per method. All tasks can be submitted via the same algorithm.
Tasks:
AAE 5.58
MPJPE_RA_H 19.14
MPJPE_RA_L 0.00
MPJPE_RA_R 0.00
MRRPE_R_L 44.37
MRRPE_R_O 34.84
SUCCESS_RATE 72.48
CDEV_HO 36.81
MDEV_H 8.81
ACC_H 5.15
ACC_O 7.25
AVG_HO 0.00
AVG_OH 0.00
ACC_OH 0.00
ACC_HO 0.00
CD_H 0.00
CD_ICP 0.00
CD_L 0.00
CD_R 0.00
F10_ICP 0.00
F5_ICP 0.00
AAE 6.65
MPJPE_RA_H 18.01
MPJPE_RA_L 0.00
MPJPE_RA_R 0.00
MRRPE_R_L 26.16
MRRPE_R_O 33.22
SUCCESS_RATE 48.94
CDEV_HO 43.13
MDEV_H 9.55
ACC_H 4.13
ACC_O 8.12
AVG_HO 0.00
AVG_OH 0.00
ACC_OH 0.00
ACC_HO 0.00
CD_H 0.00
CD_ICP 0.00
CD_L 0.00
CD_R 0.00
F10_ICP 0.00
F5_ICP 0.00