Table 2.

Main results on quadruped benchmarks

Method	Pre-trained weights	Data ratio	mAP	RMSE	Dataset	NE_IID	NE_OOD	Architecture
Zero-shot	SuperAnimal	–	68.038	12.971	AP-10K	–	–	HRNetw32
Zero-shot	SuperAnimal	–	66.110	12.849	AP-10K	–	–	AnimalTokenPose
Transfer learning	ImageNet	1.00	70.548	11.228	AP-10K	–	–	HRNetw32
Memory replay	SuperAnimal	1.00	80.113	11.296	AP-10K	–	–	HRNetw32
Zero-shot	AP-10K	–	79.447	5.774	AnimalPose	–	–	HRNetw32
Zero-shot	SuperAnimal	–	84.639	4.884	AnimalPose	–	–	HRNetw32
Zero-shot	SuperAnimal	–	83.043	5.154	AnimalPose	–	–	AnimalTokenPose
Transfer learning	ImageNet	1.00	86.864	5.757	AnimalPose	–	–	HRNetw32
Fine-tuning	AP-10K	1.00	86.794	4.860	AnimalPose	–	–	HRNetw32
Memory replay	SuperAnimal	1.00	87.034	4.636	AnimalPose	–	–	HRNetw32
Zero-shot	AP-10K	–	65.729	4.929	Horse-10	0.296	0.287	HRNetw32
Zero-shot	SuperAnimal	–	71.205	3.958	Horse-10	0.227	0.228	HRNetw32
Zero-shot	SuperAnimal	–	68.977	4.081	Horse-10	0.239	0.233	AnimalTokenPose
Transfer learning	ImageNet	0.01	0.934	46.255	Horse-10	2.369	2.36	HRNetw32
Transfer learning	ImageNet	1.00	90.516	1.837	Horse-10	0.036	0.135	HRNetw32
Fine-tuning	AP-10K	0.01	66.284	5.029	Horse-10	0.286	0.285	HRNetw32
Fine-tuning	AP-10K	1.00	93.973	1.220	Horse-10	0.036	0.083	HRNetw32
Memory replay	SuperAnimal	0.01	73.366	3.719	Horse-10	0.209	0.202	HRNetw32
Memory replay	SuperAnimal	1.00	95.165	1.153	Horse-10	0.040	0.073	HRNetw32
Zero-shot	AP-10K	–	40.389	37.417	iRodent	–	–	HRNetw32
Zero-shot	SuperAnimal	–	58.557	33.496	iRodent	–	–	HRNetw32
Zero-shot	SuperAnimal	–	55.415	34.666	iRodent	–	–	AnimalTokenPose
Transfer learning	AP-10K	0.01	12.910	92.649	iRodent	–	–	HRNetw32
Transfer learning	ImageNet	0.01	0.785	152.225	iRodent	–	–	HRNetw32
Transfer learning	ImageNet	1.00	58.857	35.651	iRodent	–	–	HRNetw32
Fine-tuning	AP-10K	0.01	43.144	37.704	iRodent	–	–	HRNetw32
Fine-tuning	AP-10K	1.00	61.635	26.758	iRodent	–	–	HRNetw32
Memory replay	SuperAnimal	0.01	60.853	31.801	iRodent	–	–	HRNetw32
Memory replay	SuperAnimal	1.00	72.971	24.884	iRodent	–	–	HRNetw32

Here, the base SuperAnimal-Quadruped model had none of the held-out datasets. Full results can be found in Fig. 2 for fine-tuning with different amounts of data, but the best fine-tuning performance is shown, which matches the top performance of the SuperAnimal (SA) variant as shown in Fig. 2. Cao et al.³³ do not report a unified single mAP, rather per animal, therefore we trained a model using their dataset to estimate top-line performance if only trained on AP. Number as reported in ref. ⁴¹ using the data from ref. ³¹.