Comparison of MobileNet, ResNet50, and InceptionV3 in Keras
This is a simple example of using ResNet, MobileNet and InceptionV3 from Keras to do object detection and classification tasks.
Demo
Models
MobileNet
ResNet50
InceptionV3
1st guess
'traffic_light', 0.99999177
'nematode', 0.090073794
'traffic_light', 1.0
2nd guess
'pole', 6.4110359e-06
'dishwasher', 0.042908493
'maillot', 8.7838531e-10
3rd guess
'flagpole', 4.8521628e-07
'cleaver', 0.026148975
'walking_stick', 5.6997368e-10
In this case, both MobileNet and InceptionV3 successfully recognize the traffic light, but the ResNet50 seems to be confused by it. Also, as for Model size, we can see from the Model Comparision section, and get to know that MobileNet has really small size. MobileNet can be a great choice for real-time applications.