Table A3.

Error rates for subsets of OS-YCB and all model architectures.

	OS-YCB
	% of occlusion	B	B-F	B-K	B-D	BT	BL	BLT	GLM
Mono	20	$.152\pm .000$	$.071\pm .000$	$.074\pm .001$	$.\mathbf{031}\pm .\mathbf{000}$	$.048\pm .001$	$.039\pm .000$	$.\mathbf{036}\pm .\mathbf{000}$	$.046\pm .000$
	40	$.286\pm .002$	$.152\pm .001$	$.155\pm .001$	$.\mathbf{069}\pm .\mathbf{000}$	$.106\pm .002$	$.088\pm .001$	$.\mathbf{083}\pm .\mathbf{002}$	$.090\pm .000$
	60	$.515\pm .004$	$.339\pm .002$	$.349\pm .001$	$.\mathbf{185}\pm .\mathbf{001}$	$.256\pm .001$	$.222\pm .002$	$.\mathbf{203}\pm .\mathbf{000}$	$.215\pm .000$
	80	$.745\pm .002$	$.658\pm .001$	$.673\pm .001$	$.\mathbf{530}\pm .\mathbf{003}$	$.599\pm .002$	$.576\pm .002$	$.\mathbf{560}\pm .\mathbf{001}$	$.556\pm .000$
Stereo	20	$.072\pm .001$	$.039\pm .001$	$.042\pm .001$	$.\mathbf{016}\pm .\mathbf{000}$	$.025\pm .000$	$.\mathbf{019}\pm .\mathbf{000}$	$.\mathbf{019}\pm .\mathbf{000}$	$.033\pm .000$
	40	$.140\pm .001$	$.068\pm .001$	$.075\pm .002$	$.\mathbf{031}\pm .\mathbf{001}$	$.051\pm .000$	$.041\pm .001$	$.\mathbf{038}\pm .\mathbf{001}$	$.052\pm .000$
	60	$.260\pm .003$	$.137\pm .002$	$.146\pm .002$	$.\mathbf{072}\pm .\mathbf{001}$	$.104\pm .000$	$.085\pm .001$	$.\mathbf{083}\pm .\mathbf{001}$	$.095\pm .001$
	80	$.433\pm .003$	$.279\pm .002$	$.285\pm .002$	$.\mathbf{167}\pm .\mathbf{002}$	$.213\pm .001$	$.182\pm .001$	$.\mathbf{175}\pm .\mathbf{001}$	$.198\pm .001$

Standard error based on five independent training runs. Training occurred for 100 epochs, batchsize 500. Best two performances per dataset are highlighted in bold.