The current implementation relies on 80 classes (COCO) on the output layer. If you have a custom trained network with different classes this will not work. The dependency isn't really required, because the actual size of the output layer is available. Instead of complaining about the difference and malfuntioning it is much easier to use the correct size.
Removed all dependencies on the number of coco classes (80).
The current implementation relies on 80 classes (COCO) on the output layer. If you have a custom trained network with different classes this will not work. The dependency isn't really required, because the actual size of the output layer is available. Instead of complaining about the difference and malfuntioning it is much easier to use the correct size.
Removed all dependencies on the number of coco classes (80).