Table 1. Quantitative comparison with other state-of-the-art 2D and 3D animal and human pose estimation methods.

We report the absolute 3D MPJPE in millimeters for each approach using four different fractions of training data.

Protocol 1 (absolute 3D MPJPE, mm)
	Training set fraction
	5%	10%	50%	100%
2D pose estimation methods (+ post hoc triangulation)
DLC† (Mathis et al., 2018)	11.0973	11.0512	9.8934	8.9060
SimpleBaseline† (Xiao, Wu, & Wei, 2018)	18.0990	14.6191	7.3636	5.9555
SimpleBaseline	18.5675	16.5800	8.3573	6.6957
DLC + soft argmax	11.0323	9.2244	6.3545	6.4739
DLC + 2D variant of our temporal constraint*	8.5432	9.1236	5.9526	6.0390
3D monocular pose estimation methods
Temporal Convolution* (Pavllo et al., 2019)	-	-	-	17.6337
3D multi-view pose estimation methods
Learnable Triangulation† (Iskakov et al., 2019)	18.7795	15.6614	8.9729	6.3177
DANNCE (Dunn et al., 2021)	12.8754	10.9085	4.9912	4.3614
Ours (temporal baseline)*	12.4940	7.1162	4.8347	4.3749
Ours (temporal + extra)*	8.1706	6.6927	5.0461	4.1409

The methods that use ground truth 2D bounding boxes during inference are masked by †.

The methods that use temporal information during training are masked by *.

For the monocular approach, the reported metric results are separately computed and averaged across all camera views.