Table 3 Performance and inference speed of the pruned (P) and quantized (Q) models with respect to the baselines. Results are reported for the tiny (a), base (b), and large (c) variants on the NYU Depth V2 (left) and KITTI (right) datasets.

From: Efficient attention vision transformers for monocular depth estimation on resource-limited hardware

Model	RMSE [m] \(\downarrow\)	\(Abs_{Rel}\downarrow\)	\(\delta _1\uparrow\)	\(\delta _2\uparrow\)	\(\delta _3\uparrow\)	XG38 [s] \(\downarrow\)	Model	RMSE [m] \(\downarrow\)	\(Abs_{Rel}\downarrow\)	\(\delta _1\uparrow\)	\(\delta _2\uparrow\)	\(\delta _3\uparrow\)	XG38 [s] \(\downarrow\)
NYU Depth V2							KITTI
(a)
METER	0.544	0.175	0.778	0.946	0.983	16.47	METER	5.945	7.408	0.287	0.485	0.604	18.13
METER P	2.789	0.938	0.001	0.001	0.001	15.17	METER P	10.960	1.630	0.023	0.047	0.071	20.26
METER Q	1.334	0.346	0.278	0.560	0.793	17.86	METER Q	6.985	7.390	0.183	0.362	0.499	17.40
PXF	0.392	0.114	0.880	0.982	0.996	112.41	PXF	2.324	0.060	0.966	0.996	0.999	160.90
PXF P	1.159	0.524	0.368	0.651	0.829	136.60	PXF P	17.172	1.609	0.112	0.225	0.337	188.94
PXF Q	1.594	0.746	0.265	0.504	0.706	136.03	PXF Q	11.379	0.502	0.262	0.502	0.725	185.37
NeWCRFs	0.388	0.112	0.885	0.980	0.995	170.78	NeWCRFs	2.373	0.059	0.965	0.995	0.999	234.65
NeWCRF P	2.355	1.177	0.126	0.303	0.509	187.33	NeWCRF P	20.433	2.037	0.076	0.158	0.261	255.80
NeWCRFs Q	1.820	0.845	0.235	0.465	0.662	211.43	NeWCRFs Q	20.951	2.086	0.077	0.160	0.261	303.66
(b)
METER	0.497	0.149	0.811	0.951	0.987	17.10	METER	5.794	6.625	0.302	0.504	0.618	22.16
METER P	2.777	0.930	0.019	0.019	0.019	19.72	METER P	11.206	1.134	0.191	0.199	0.208	29.89
METER Q	1.487	0.663	0.296	0.555	0.752	19.63	METER Q	6.046	8.031	0.231	0.459	0.607	23.59
PXF	0.338	0.096	0.918	0.988	0.997	181.09	PXF	2.205	0.055	0.972	0.997	0.999	262.77
PXF P	2.573	1.287	0.109	0.262	0.463	234.45	PXF P	26.822	2.704	0.055	0.116	0.188	343.31
PXF Q	1.487	0.692	0.281	0.550	0.738	238.05	PXF Q	13.297	0.433	0.255	0.427	0.571	334.93
NeWCRFs	0.337	0.095	0.918	0.989	0.998	241.93	NeWCRFs	2.185	0.054	0.972	0.999	0.997	356.12
NeWCRF P	2.566	1.283	0.109	0.266	0.464	304.09	NeWCRFs P	26.995	2.722	0.054	0.115	0.187	368.20
NeWCRFs Q	1.870	0.908	0.202	0.428	0.635	358.69	NeWCRFs Q	18.496	1.808	0.087	0.186	0.308	488.34
(c)
METER	0.460	0.133	0.834	0.966	0.992	21.52	METER	5.726	7.299	0.332	0.524	0.630	34.50
METER P	2.207	0.649	0.039	0.093	0.182	31.30	METER P	10.732	2.291	0.034	0.068	0.103	51.86
METER Q	1.728	0.812	0.223	0.473	0.683	26.41	METER Q	10.955	4.002	0.078	0.134	0.172	35.69
PXF	0.324	0.091	0.928	0.991	0.998	294.74	PXF	2.123	0.052	0.975	0.997	0.999	426.68
PXF P	2.616	1.308	0.105	0.255	0.457	437.33	PXF P	27.298	2.752	0.053	0.113	0.184	683.33
PXF Q	1.387	0.632	0.306	0.570	0.761	430.85	PXF Q	10.530	0.516	0.273	0.544	0.766	729.44
NeWCRFs	0.322	0.091	0.929	0.992	0.998	339.68	NeWCRFs	2.072	0.052	0.975	0.997	0.999	500.17
NeWCRF P	2.656	1.328	0.105	0.248	0.445	477.79	NeWCRF P	27.389	2.761	0.053	0.113	0.183	733.47
NeWCRFs Q	2.327	1.144	0.145	0.329	0.531	565.98	NeWCRFs Q	21.612	2.164	0.070	0.146	0.244	899.43

Back to article page

Table 3 Performance and inference speed of the pruned (P) and quantized (Q) models with respect to the baselines. Results are reported for the tiny (a), base (b), and large (c) variants on the NYU Depth V2 (left) and KITTI (right) datasets.

Search

Quick links