The figure has two panels titled “(a) End to End Latency” on the left and “(b) Accuracy versus Computational Cost” on the right. In panel (a), a grouped bar chart compares latency in milliseconds for four methods—“Our H D R L (purple), Transformer (yellow), V A E P (blue), and G R F (green)—across four hardware platforms on the horizontal axis: “Server G P U (A 100)”, “Edge G P U”, “Mobile G P U”, and “C P U Only (Xeon)”. The vertical axis is labeled “Latency (milliseconds)” and ranges from 0 to 350 with an interval of 50. A horizontal red dashed line at 50 milliseconds across the plot is labeled “Real time threshold (50 milliseconds)”. For the server G P U, the bars are very short: Our H D R L is around 8.2, the transformer is slightly above that, V A E P is slightly smaller, and G R F is again slightly higher. For the edge G P U, Our H D R L reaches 28.3 and is marked 43.4 percent below; the transformer bar is slightly higher, but the bars for V A E P and G R F are higher and range between 65 and 75. For the mobile G P U, Our H D R L reaches 47.8, but the bars for transformer, V A E P, and G R F are higher and range between 85 and 110. For the C P U only, Our H D R L reaches 143.5, but the bars for transformer, V A E P, and G R F are higher and range between 225 and 310. Panel (b) is a multi-line chart with different markers. The horizontal axis is labeled “Computational cost (G FLOPs)” and ranges from 0 to 8 with an interval of 1, and the vertical axis is labeled “Accuracy (percent)” and ranges from 45 to 90 with an interval of 5 percent. Four colored curves represent the same methods: purple circles for Our H D R L, yellow squares for Transformer, blue triangles for V A E P, and green diamonds for G R F, as indicated in a legend at the bottom right. The Our H D R L curve starts around 73 percent accuracy at about 1 G F L O P, rises steeply, and ends around 7 G F L O P at 86.5 percent. The Transformer curve begins near 61 percent at 1 G F L O P and climbs steadily to the mid 70 percent range and ends around 76 percent at 7 G F L O P. The V A E P and G R F curves start between 47 and 48 percent at 1 G F L O P and rise gradually and overlap before ending around 62 percent at 7 G F L O P. A shaded vertical band from 2 to 4 G F L O P s is labeled “Optimal Region”, and a purple arrow annotation within it notes a point labeled “41.2 FLOPs reduction at 98.2 percent accuracy”. Note: All numerical data values are approximated.Computational efficiency and real-time performance analysis
Sharing content requires targeting cookies to be enabled. Please update your cookie preferences to use this feature.