matrixnet
matrixnet copied to clipboard
arashwan
A question about layer range
why the the base layer (top left) is set as [24,48,24,48] but in corresonding layer range the the base layer (top left) is [0,48,0,48]
@lmomoy In layer ranges, we set it to [0,48,0,48] cause we want to assign objects with width/height less than 48 to the base layer. Note that we extended other layer ranges for the same reason.
thanks for your reply but I am still confused that if the base layer (top left) should be set as [0,48,0,48], but not [24,48,24,48], because the top left layer is [0,48,0,48] in matirxnet
@lmomoy base_layer_range variable is used to compute anchor widths/heights. We didn't want to use the extended ranges (layer_ranges) since this will introduce some inconsistencies between matrix layers (i.e. anchor width and height w.r.t. layer size will be different from layer to another).
Hello,guys,I also have some questions in layer range, as I know, the height and width of feature map is a stationary number, but why it is a range in your paper and code ?
It's a range in the sense that the height or width of the ground truth has to lie within a [min, max] set to be assigned to a layer.
It's a range in the sense that the height or width of the ground truth has to lie within a [min, max] set to be assigned to a layer.
Does it means that the feature map will be assigned every layer and the every layer will dectect the corresponding aspect ratios object, and the result bbox is the sum of all of layer result bbox?
each layer has a feature map of its own - it's downsampled from the last layers of a resnet into a large feature pyramid with asymmetric strides.
each layer will have its own set of detections - yes
then NMS is used to get the highest scoring boxes