In Tenosrflow Object Detection API, you can choose different pretrained models such as Faster RCNN, SSD. Then you can specify the batch size during training.
I know that for Stochastic Gradient Descent, you feed the network successively with batches from the input.
I have few questions in mind.
How does stochastic gradient descent work in Faster RCNN? Given that it first has a feature extractor layer, then a classification layer.
I found this from https://wiki.tum.de/pages/viewpage.action?pageId=22578448
One big improvement provided by the Fast R-CNN is that it takes advantage of feature sharing during training. In taining, stochasitc gradient descent (SGD) minibatches are sampled hierachically, first by sampling N images and then by sampling R/N RoIs from each image.
but I don't exactly understand the stated process. Can someone expound on this?
since one image can contain multiple objects, when you feed it by batch, does it learn on the entire batch of images, image by image, or object by object?
How does SGD decide on which batch to feed next?