I am using a pretrained model from this tutorial. https://pytorch.org/tutorials/intermediate/torchvision_tutorial.html#defining-your-model
The model is pytorch's Faster RCNN ResNet 50 FPN model. Does anyone know what the classification loss, loss, and objectness loss functions are (i.e. Cross Entropy or?). Thanks in advance, Sriram A.
Objectness is a binary cross entropy loss term over 2 classes (object/not object) associated with each anchor box in the first stage (RPN), and classication loss is normal cross-entropy term over C classes. Both first stage region proposals and second stage bounding boxes are also penalized with a smooth L1 loss term.
It should also be noted that the authors train the first and second stage alternately since both rely on the same features computed with convolutional layers + FPN to aid in training convergence.
Not a very clear description? I'd recommend reading the original Faster-RCNN paper as it is pretty foundational and will probably do a better job describing the loss terms than me.
