Pretraining boosts out-of-domain robustness for pose estimation

Neural networks are highly effective tools for pose estimation. However, asin other computer vision tasks, robustness to out-of-domain data remains achallenge, especially for small training sets that are common for real-worldapplications. Here, we probe the generalization ability with three architectureclasses (MobileNetV2s, ResNets, and EfficientNets) for pose estimation. Wedeveloped a dataset of 30 horses that allowed for both "within-domain" and"out-of-domain" (unseen horse) benchmarking - this is a crucial test forrobustness that current human pose estimation benchmarks do not directlyaddress. We show that better ImageNet-performing architectures perform betteron both within- and out-of-domain data if they are first pretrained onImageNet. We additionally show that better ImageNet models generalize betteracross animal species. Furthermore, we introduce Horse-C, a new benchmark forcommon corruptions for pose estimation, and confirm that pretraining increasesperformance in this domain shift context as well. Overall, our resultsdemonstrate that transfer learning is beneficial for out-of-domain robustness.