Extended Data Fig. 2: Per keypoint error and processing speed of networks using various batch sizes.
From: Facemap: a framework for modeling neural activity based on orofacial tracking
a, Error for each keypoint, averaged across 100 test frames for each network plotted against the Facemap tracker errors. Points above the diagonal indicate keypoints for which Facemap outperformed the other networks. b, Processing speed of Facemap, DeepLabCut (ResNet50), DeepLabCut (Mobilenet), SLEAP (default) and SLEAP (c = 32) models for a sample image of size 256 × 256 pixels on A100 (48 slots, 40GB/slot), V100 (48 slots, 30GB/slot), RTX 2080 Ti (40 slots, 18GB/slot), Tesla T4 (48, 15GB/slot) and Quadro RTX 8000 (40 slots, 18GB/slot). Processing speed averages shown for a total of 1,024 frames across n = 10 runs, and error bars represent SEM.
