Average loss is the average of various losses that is arise in a model.
Average loss varies from model to model since different types of loss arise in each model. Some examples are discussed below.
Mask RCNN involves the calculation of bounding boxes, regional proposals and masks. The usual classification is also involved. Hence average loss for MaskRCNN would be the average of the regression loss of bounding box ( L2 loss ), classification loss in regional proposals (0-1 loss), localization error in regional proposals (again L2 loss), and the final classification and mask error. The Average Error takes the average of each of these errors across the samples in a batch. The final loss is the sum of these averages across batch.
Like MaskRCNN, FasterRCNN also involves multiple losses of various kinds. The only difference is that there is no mask error in Faster RCNN since no masks are generated. But all the regression errors like bounding box and regional proposal localization and the classification errors are averaged across batch to find the average error for FasterRCNN.
Here we briefly discussed two examples explaining how the average loss might apply.
Average loss helps you get a rough estimation on how the model is performing as a whole, and it might not necessarily evaluate the end result of the model.