. Author manuscript; available in PMC: 2022 Dec 20.

Published in final edited form as: Proc IEEE Comput Soc Conf Comput Vis Pattern Recognit. 2021 Nov 2;2021:2875–2884. doi: 10.1109/cvpr46437.2021.00290

Table 2. Decoder Error Reductions.

Percentage error reduction relative to baseline keypoints and domain-specific features for training with different decoder losses for TREBA. The average is taken over all evaluated training fractions.

	Keypoint Error Reduction (%)
Decoder Loss	MARS	CRIM13	Fly
TVAE	52.2 ± 4.0	34.7 ± 1.5	15.4 ± 2.1
TVAE+Unsup. Contrast	52.6 ± 3.9	37.4 ± 2.4	20.9 ± 1.7
TVAE+Contrast+Consist	55.1 ± 3.0	41.1 ± 2.1	33.7 ± 1.2
	Features Error Reduction (%)
Decoder Loss	MARS	CRIM13	Fly
TVAE	13.7 ± 1.8	8.2 ± 4.6	11.7 ± 4.7
TVAE+Unsup. Contrast	14.3 ± 2.2	8.9 ± 4.1	16.1 ± 1.7
TVAE+Contrast+Consist	15.3 ± 2.1	9.5 ± 3.8	21.2 ± 4.5