Table 1 Comparison of model size, speed, and performance, ordered by top-1 accuracy. Text in bold indicates the best value in each category. Wide-ResNet101 has the highest top-1 accuracy and macro precision but was substantially slower than the other models. MnasNet-A1 was the fastest model but had relatively poor performance. InceptionV3 was relatively fast while maintaining good model performance with the highest precision and second highest accuracy and recall.

From: Assessing the potential for deep learning and computer vision to identify bumble bee species from images

Model #Params (million) Model speed (ms) Top-1 accuracy Top-2 accuracy Top-3 accuracy Top-4 accuracy Top-5 accuracy Macro recall Macro precision
Wide-ResNet101 124.9 5.46 0.9171 0.9627 0.9782 0.9850 0.9897 0.8552 0.8831
InceptionV3 24.0 3.34 0.9162 0.9610 0.9767 0.9834 0.9882 0.8519 0.8881
ResNet101 42.6 3.33 0.9133 0.9633 0.9787 0.9852 0.9892 0.8499 0.8740
MnasNet-A1 1.0 3.28 0.8579 0.9335 0.9609 0.9730 0.9814 0.7689 0.8250