Table A1.

Comparison of different approaches. Legend:: Invasive: Approaches which requires surgery to put markers for tracking, semi-invasive: Approaches which do not need surgery for marker insertion, non-invasive: no marker needed. Real time means that the system can process frames at the same rate they are being acquired. If it needs specialized equipment apart from standard video cameras and housing setup, it is pointed out in the last column.

	Type	Code Availability	Performance	Real Time or Offline	Need Specialized Setup & Invasiveness
[59]	Commercial	Paid	Comparison with ground truth not provided. One paper reports the reproducibility: 2.65% max SD	Yes	Yes
[71]	Commercial	Paid	Comparison with ground truth not provided. One paper reports the reproducibility: 1.57% max SD	Yes	Yes
[49]	Research	data and code for demo available at http://bit.do/eTTai	tracking performance not reported, behavioral classification of 12 traits reported to be max at 71%	Tracking real time, classification offline	yes
[51]	Research	not available	tracking: SD of only 0.034% when compared with ground truth, Max SD of 1.71 degrees in estimating joint angle	real time legs and joints tracking	yes, invasive
[52]	Research	not available	tracking performance not reported explicitly	real time whisker tracking	yes, semi-invasive
[53]	Research	available on request	whisker tracking performance not reported explicitly	real time single whisker tracking	yes, semi-invasive
[54]	Research	not available	head motion tracked correctly with a max false positive of 13%	real time head and snout tracking	yes, semi-invasive
[55]	Research	not available	head motion tracked continuously with a reported SD of only 0.5 mm	real time head and snout tracking	yes, semi-invasive
[57]	Research	not available	head motion tracked with an accuracy of 96.3% and the tracking can be reproduced over multiple studies with a correlation coefficient of 0.78	real time head tracking	yes, semi-invasive
[89]	Research	code and demo data available at https://goo.gl/vYaYPy	they reported a correlation between whisking amplitude and velocity as a measure of reliability, R = 0.89	Offline head and whisker tracking	no, invasive
[90]	Research	not available	Tracking and gait prediction with confidence of 95%, deviation between human annotator and computer at 8%	Offline	yes, semi-invasive
[93]	Research	not available	Paw tracked with an accuracy of 88.5 on transparent floor and 83.2% on opaque floor	Offline	yes, semi-invasive
[95]	Research	code available at https://goo.gl/58DQij	tail and paws tracked with an accuracy >90%	Real time	yes, semi-invasive
[97]	Research	not available	5 class behavioral classification problem, accuracy in bright condition is 95.34 and in dark conditions is 89.4%	offline	yes, non-invasive
[100]	Research	not available	6 behavioral class accuracy: 66.9%, 4 behavioral class accuracy: 76.3%	offline	yes, non-invasive
[98]	Research	code available at https://goo.gl/eY2Yza	whisker detection rate: 76.9%, peak spatial error in whisker detection: 10 pixels	offline	yes, non-invasive
[101]	Research	not available	Peak deviation between human annotator and automated annotation: 0.5 mm with a camera of 6 pixel/mm resolution	offline	yes, non-invasive
[92]	Research	not available	Tracking accuracy >90% after the algorithm was assisted by human users in 3–5% of the frames	offline	yes, semi-invasive
[105]	Research	code available at https://goo.gl/Gny89o	A max deviation of 17.7% between human and automated whisker annotation	offline	yes, non-invasive
[107]	Research	not available	Maximum paw detection error: 5.9%, minimum error : 0.4%	offline	no, non-invasive
[110]	Research	Source code at https://goo.gl/zesyez, demo data at https://goo.gl/dn2L3y	Behavioral classification: 1% false positive rate	offline	no, semi-invasive
[108]	Research	Source code available at https://goo.gl/JCv3AV	Whisker tracing accuracy: max error of 0.45 pixels	offline	no, non-invasive
[116]	Research	not available	Correlation with annotated data; for whiskers r = 0.78, for limbs r = 0.85	real time	no, non-invasive
[109]	Research	code available at https://goo.gl/V54mpL	Velocity calculated by AGATHA was off from manually calculated velocity by 1.5%	real time	no, non-invasive
[117]	Research	code available at http://bit.ly/2vgJUbr	Detected pose matched ground truth with an accuracy of $4.17\pm 0.32$ pixels	real time on GPUs	no, non-invasive
[119]	Research	code available at https://bit.ly/2XuJmPv	No performance metric reported	offline	no, non-invasive