Table 6 The effect of using different replacement rules and two widely adopted architectural changes ResNet-D and Squeeze-and-Exitation (D-SE) in the basic four phases of ResNet-50, where \(\checkmark\) denotes that we replaced the PoT module and \(\uparrow\) denotes that we replaced it using the D-SE architecture.(Experimented on VisDrone 2019 dataset).
From: End to end polysemantic cooperative mixed task trainer for UAV target detection
| Â | ResNet-2 | ResNet-3 | ResNet-4 | ResNet-5 | D-SE | param#(M) | Flops | Inference | Accuracy(ALL-top-1) | Accuracy(ALL-top-5) |
|---|---|---|---|---|---|---|---|---|---|---|
ResNet-50 | Â | Â | Â | Â | Â | 24.4M | 4.0G | 510 EX/S | 31.93 | 48.23 |
PoT-50 | Â | Â | Â | \(\checkmark\) | Â | 22.4M | 3.9G | 493 EX/S | 33.13 | 48.73 |
| Â | Â | \(\checkmark\) | \(\checkmark\) | Â | 21.3M | 3.6G | 445 EX/S | 33.63 | 48.75 | |
| Â | \(\checkmark\) | \(\checkmark\) | \(\checkmark\) | Â | 21.2M | 3.3G | 392 EX/S | 33.63 | 48.76 | |
\(\checkmark\) | \(\checkmark\) | \(\checkmark\) | \(\checkmark\) | Â | 21.1M | 3.2G | 333 EX/S | 33.62 | 48.77 | |
SE-ResNet-D-50 | Â | Â | Â | Â | \(\uparrow\) | 34.6M | 4.3G | 446 EX/S | 33.61 | 48.77 |
SE-PoT-D-50 | Â | Â | \(\checkmark\) | \(\checkmark\) | \(\uparrow\) | 22.0M | 4.0G | 416 EX/S | 33.68 | 48.79 |
