About the batch size in training detection network

[yw155]

The batch size is set to 32 in the paper. Is this size indicates the number of input images?

In the SSD method, the network can generate a lot of detection windows and using these windows to compute the training loss. Since the number of windows is large enough, is it can further reduce the batch size in the training, even to 1? What is the effect of batch size to the training?

Thank you!

[weiliu89]

32 is the number of images in a batch. Although there is a lot of sampled patches from an image, we only select one from each image.

[yw155]

I got that. Thank you!