HoyerChou/MultiAugs

Pretrained models of our method MultiAugs

Title: Boosting Semi-Supervised 2D Human Pose Estimation by Revisiting Data Augmentation and Consistency Training

Paper link: https://arxiv.org/abs/2402.11566

Code link: https://github.com/hnuzhy/MultiAugs

COCO1K / COCO5K / COCO10K

trained on partly labeled (1k, 5k or 10k) COCO train-set and left unlabeled COCO train-set

• ResNet-18 (Pose_Cons using single network)(256x192, COCO1K, 30 epochs): pose_cons_18-COCO1K_e30-model_best.pth.tar
• ResNet-18 (Pose_Cons using single network)(256x192, COCO5K, 70 epochs): pose_cons_18-COCO5K_e70-model_best.pth.tar
• ResNet-18 (Pose_Cons using single network)(256x192, COCO10K, 100 epochs): pose_cons_18-COCO10K_e100-model_best.pth.tar
• ResNet-18 (Pose_Dual using dual networks)(256x192, COCO1K, 30 epochs): pose_dual_18-COCO1K_e30-model_best.pth.tar
• ResNet-18 (Pose_Dual using dual networks)(256x192, COCO5K, 70 epochs): pose_dual_18-COCO5K_e70-model_best.pth.tar
• ResNet-18 (Pose_Dual using dual networks)(256x192, COCO10K, 100 epochs): pose_dual_18-COCO10K_e100-model_best.pth.tar

COCOall + COCOunlabel

trained on labeled COCO train-set and unlabeled COCO unlabeled-set

• ResNet-50 (Pose_Cons) (256x192, 400 epochs): pose_cons_50-COCO_COCOunlabel_e400-model_best.pth.tar
• ResNet-101 (Pose_Cons) (256x192, 400 epochs): pose_cons_101-COCO_COCOunlabel_e400-model_best.pth.tar
• HRNet-w48 (Pose_Cons) (384x288, 300 epochs): pose_cons_w48-COCO_COCOunlabel_e300-model_best.pth.tar
• ResNet-50 (Pose_Dual) (256x192, 400 epochs): pose_dual_50-COCO_COCOunlabel_e400-model_best.pth.tar
• ResNet-101 (Pose_Dual) (256x192, 400 epochs): pose_dual_101-COCO_COCOunlabel_e400-model_best.pth.tar
• HRNet-w48 (Pose_Dual) (384x288, 300 epochs): pose_dual_w48-COCO_COCOunlabel_e300-model_best.pth.tar

MPII + AIC

trained on labeled MPII train-set and unlabeled AIC train-set

• HRNet-w32 (Pose_Dual) (256x256, 400 epochs) [We are sorry that it cannot be released due to company copyright issues]