CODEWITHSUNDEEP
neural-networkconv-neural-networkdarknet
harmegiddo
harmegiddo

Reputation: 35

Can not Yolo v3 set a number of a bounding box less than three?

In the calculation formula of Yolo v3's filter, the number of a bounding box is divided by 3 (why?).

For this reason, the number of the number of a bounding box is only allowed to be multiples of 3.

But I want to set one bounding box. Is this possible?

If it is possible, how many is a filter?

The bellow is I would like to realize the code.

[convolutional]
batch_normalize=1
size=3
stride=1
pad=1
filters=256
activation=leaky

[convolutional]
size=1
stride=1
pad=1
filters=????????? 
activation=linear

[yolo]
mask = 0,1,2
anchors = 10,13
classes=20
num=1
jitter=.3
ignore_thresh = .5
truth_thresh = 1
random=1

Upvotes: 2

Views: 2174

Answers (1)

dktrivedi
dktrivedi

Reputation: 111

There are four important variables:

anchors: predetermined set of boxes with particular height-width ratios

mask: list of ids of the bounding boxes(anchors) that the layer is responsible for predicting

num: total number of anchors

filter = (num_classes+5)*k where k=number of masks in one yolo layer

YOLOv3 predicts off-sets from a predetermined set of boxes with particular height-width ratios(anchors). Anchors are initial (width, height) sizes, some of which (the closest to the object size) will be resized to the object size.

Every [yolo] layer has to know about all of the anchor boxes but is responsible for only a subset of them. The mask tells the layer which of the bounding boxes it is supposed to use for prediction.

Default configuration of YOLOv3:

[convolutional]
size=1
stride=1
pad=1
filters=255
activation=linear


[yolo]
mask = 0,1,2
anchors = 10,13,  16,30,  33,23,  30,61,  62,45,  59,119,  116,90,  156,198,  373,326
classes=80
num=9
jitter=.3
ignore_thresh = .7
truth_thresh = 1
random=1

here mask = 0,1,2 means this [yolo] layer (last [yolo] layer in config file) is responsible for predicting bounding boxes related to anchors 0 (10,13), 1 (16,30), 2 (33,23).

Overall,

First [yolo] layer is responsible for 6,7,8 anchors.

Second [yolo] layer is responsible for 3,4,5 anchors.

Third [yolo] layer is responsible for 0,1,2 anchors.

Example configuration for predicting one bounding box at each [yolo] layer:

[convolutional]
size=1
stride=1
pad=1
filters=85
activation=linear


[yolo]
mask = 0
anchors = 10,13,  30,61,  116,90
classes=80
num=3
jitter=.3
ignore_thresh = .7
truth_thresh = 1
random=1

Upvotes: 5

Related Questions