Table 3. Comparison of rate-distortion... | Emerald Publishing

Table 3.

Comparison of rate-distortion performance (BD-rate) and computational complexity (kMACs/pixel and parameter size) for all token mixers evaluated within the ATEM framework. BD-rate values are calculated on the UVG data set relative to the VCT baseline. Encoding and decoding times (Enc, Dec) are calculated per-frame, evaluated on 1080p videos with RTX3090 GPU

Model	BD-rate (⁠ $↓$ ⁠)	kMACs/pix (⁠ $↓$ ⁠)	Parameters(M) (⁠ $↓$ ⁠)	Enc(s) (⁠ $↓$ ⁠)	Dec(s) (⁠ $↓$ ⁠)
VCT	0	3085.7	154	2.13	1.08
ATEM-Pool	12.56%	2693.02	142.41	2.07	1.05
ATEM-MLP	8.09%	4014.48	183.88	2.59	1.2
ATEM-Conv	−8.03%	3192.17	158.198	2.16	1.07
ATEM-Hybrid AC	−13.63%	3139.04	156.21	2.14	1.09
ATEM-Hybrid CA	−16.64%	3139.04	156.21	2.14	1.09
ATEM-Swin	−14.86%	2542.5	155.36	1.11	0.46

Model	BD-rate ( $↓$ )	kMACs/pix ( $↓$ )	Parameters(M) ( $↓$ )	Enc(s) ( $↓$ )	Dec(s) ( $↓$ )
VCT	0	3085.7	154	2.13	1.08
ATEM-Pool	12.56%	2693.02	142.41	2.07	1.05
ATEM-MLP	8.09%	4014.48	183.88	2.59	1.2
ATEM-Conv	−8.03%	3192.17	158.198	2.16	1.07
ATEM-Hybrid AC	−13.63%	3139.04	156.21	2.14	1.09
ATEM-Hybrid CA	−16.64%	3139.04	156.21	2.14	1.09
ATEM-Swin	−14.86%	2542.5	155.36	1.11	0.46

Source(s): Authors’ own work