Table 4.

The detection performance of average precision metrics and average recall metrics with different SAF configurations on the nuScenes validation dataset. The configuration of “1×1” stands for a convolution layer with kernel size $1\times 1\times 256\times 1$, stride (1, 1), padding [0, 0]. As for the layers of “3×3”, “5×5” and “7×7”, the configurations are {$3\times 3\times 256\times 1$, (1, 1), [1, 1]}, {$5\times 5\times 256\times 1$, (1, 1), [2, 2]}, and {$7\times 7\times 256\times 1$, (1, 1), [3, 3]}, respectively.

SAF
$1\times 1$	$3\times 3$	$5\times 5$	$7\times 7$	$\mathit{AP}$(100)	${\mathit{AP}}^{0.50}$(100)	${\mathit{AP}}^{0.75}$(100)	${\mathit{AP}}^{\mathit{s}}$(100)	${\mathit{AP}}^{\mathit{m}}$(100)	${\mathit{AP}}^{\mathit{l}}$(100)
				60.7	84.6	65.6	43.6	58.7	72.7
✓				58.6	84.5	62.9	39.6	57.5	70.9
	✓			68.8	88.7	75.3	51.7	67.2	80.1
		✓		69.8	89.8	76.2	53.2	67.9	80.6
			✓	68.1	88.8	74.0	49.8	66.8	79.0
✓	✓			69.4	89.7	76.5	52.7	67.6	80.1
	✓	✓		67.0	88.3	72.8	49.0	65.2	78.5
		✓	✓	54.7	82.5	58.0	33.6	53.3	68.8
✓	✓	✓		70.2	89.9	76.7	54.3	68.6	80.6
	✓	✓	✓	69.8	89.6	75.9	53.6	67.7	80.6
✓	✓	✓	✓	70.1	89.9	76.7	52.4	68.8	80.3
SAF
$\mathbf{1}\times \mathbf{1}$	$\mathbf{3}\times \mathbf{3}$	$\mathbf{5}\times \mathbf{5}$	$\mathbf{7}\times \mathbf{7}$	$\mathit{AR}$ (1)	$\mathit{AR}$ (10)	$\mathit{AR}$ (100)	${\mathit{AR}}^{\mathit{s}}$ (100)	${\mathit{AR}}^{\mathit{m}}$ (100)	${\mathit{AR}}^{\mathit{l}}$ (100)
				12.2	56.9	70.5	55.8	70.0	79.7
✓				12.0	55.5	68.4	50.6	68.1	78.5
	✓			12.6	63.7	77.0	64.9	76.2	85.0
		✓		12.6	64.9	77.2	66.0	76.4	84.8
			✓	12.6	63.0	76.2	64.5	75.5	83.8
✓	✓			12.7	63.8	76.6	66.5	75.1	84.7
	✓	✓		12.6	62.1	75.4	63.9	74.2	83.9
		✓	✓	11.7	52.8	64.7	45.7	63.8	76.8
✓	✓	✓		12.7	65.2	77.4	66.8	76.3	85.0
	✓	✓	✓	12.6	64.3	77.0	66.8	75.8	84.7
✓	✓	✓	✓	12.8	64.3	76.8	65.8	75.8	84.5