The colorization-v2-norebal
model is one of the colorization group of models designed to perform image colorization. For details about this family of models, check out the repository.
This model differs from model colorization-v2
in that metrics did not take into account balancing of rare classes during training.
Model consumes as input L-channel of LAB-image. Model give as output predict A- and B-channels of LAB-image.
Metric | Value |
---|---|
Type | Colorization |
GFLOPs | - |
MParams | - |
Source framework | Caffe* |
The accuracy metrics calculated on ImageNet validation dataset using VGG16 caffe model and colorization as preprocessing.
For preprocessing rgb -> gray -> coloriaztion
recieved values:
Metric | Value with preprocessing | Value without preprocessing |
---|---|---|
Accuracy top-1 | 57.24% | 70.96% |
Accuracy top-5 | 80.96% | 89.88% |
Image, name - data_l
, shape - 1,1,224,224
, format is B,C,H,W
where:
B
- batch sizeC
- channelH
- heightW
- widthChannel order is L-channel. Mean values - 50.
Image, name - data_l
, shape - 1,1,224,224
, format is B,C,H,W
where:
B
- batch sizeC
- channelH
- heightW
- widthChannel order is L-channel.
Image, name - class8_ab
*, shape - 1,2,56,56
, format is B,C,H,W
where:
B
- batch sizeC
- channelH
- heightW
- widthImage, name - class8_313_rh
*, shape - 1,313,56,56
, format is B,C,H,W
where:
B
- batch sizeC
- channelH
- heightW
- widthNOTE:
class8_313_rh
layer is in front ofclass8_ab
layer,
in order for network to work, you need to reproduce class8_ab
layer with the coefficients that downloaded separately with the model.
The original model is distributed under the following license: