YanHan's Modification:
- TransformParamEx: can crop the image into rectangle rather than the square only
- SyncTransformationParameter: Can sync transform the data image and label image, it is a very useful layer in the image2image models. SyncTransformLayer: Transform the input blobs with the same random parameters
- FocalTransformationParameter: Used in focal length estimation algorithm, can crop the data and alter the facal label. FocalTransformLayer: Crop the image and transform the focal length accordingly.
- Unpooling layer: It can reverse the pooling to enlarge the feature map
- euclideanfcn_loss_layer: It can ignore the bad pixel caused by NaN in the label image
- ResizeLayer: This layer use opencv function can resize the feature map (include the data and label image) online.
- ConvolutionExLayer: This layer support the custom pad value instead of the default 0
- BaseConvolutionExLayer: Provide base layer for the ConvolutionExLayer
- im2col_ex: Used by BaseConvolutionExLayer to provide modified pad value
- Conv1x1Layer: The layer only do 1x1 conv and have some extra regulation on the conv parameters
- BaseConv1x1Layer: Provide the base for the Conv1x1Layer.
- UnpackBitsLayer: Split the packed pairwise data into multi channel data
- CrfLossLayer: The CRF loss layer
- ConvDropoutLayer: Perform the dropout on the feature map rather than the pixel
- Crf2LossLayer: The second version of the CrfLossLayer, it embed the Conv1x1Layer param into the loss layer, that means the bottom[2] can directly connected to the output of the UnpackBitsLayer
- SyncTransFastLayer: A simple and faster version of the SyncTransformLayer.
- L2NormLayer: Can Perform L2 normalize between the channels of certain pixel position
- PairwiseMapLayer: Generate the R matrix in CRF loss layer according to the pairwise enery from the superpixel pooling layer
- Crf3LossLayer: The CRF loss designed for multi channels prediction (e.g. normal map)
- CrfUnaryLossLayer: modified from the Crf3LossLayer, this loss layer drop the pairwise parameter learning, only learn the unary parameter. It take the R matrix directly as the input instead of the pairwise features. NOTE: The R matrix can be calculated by the PairwiseMapLayer.
- BNLayer: This is another batch normalization layer introducted in PSPnet, contains four parameters as 'slope,bias,mean,variance' while 'batch_norm_layer' contains two parameters as 'mean,variance'.
- ScaleInvariantLossLayer: A loss layer introducted by Eigen NIPS2014, but this layer DO NOT use the log space as the Eigen did.
- ScaleInvariantLogLossLayer: The loss layer introducted by Eigne NIPS2014
- ScaleInvariantBerhuLossLayer: The ScaleInvariantLossLayer combined with Berhu loss layer
- ScaleInvariantLogBerhuLossLayer: The ScaleInvariantLogLossLayer combined with Berhu loss layer
- SuperpixelCentroidLayer: This layer locate each superpixel's centroid coordinates, which can be used in other layers
- CrfNormLossLayer: The normal guided CRF loss layer.
- DistCorrMatLayer: Generate the superpixel distance correlation matrix according to the central positial of each superpixel
- NormCorrMatLayer: Generate the superpixel normal correlation matrix
- PixelwiseLossLayer: Change the superpixel wise prediction into the pixelwise prediction, and BP the surface normal diff
- GradToNormLayer: This layer convert the gradient input bottom into the surface normal, the input bottom shape is [n, 3, x, x], and the output shape is [n, 2, x, x], it contains both forward and backward functions. 32.DepthToNormLayer: This layer convert the depth map to surface normal map, the input bottom shape is [n, 1, h, w], and the output shape is [n, 3, h, w], it only contains forward function
Caffe
Caffe is a deep learning framework made with expression, speed, and modularity in mind. It is developed by the Berkeley Vision and Learning Center (BVLC) and community contributors.
Check out the project site for all the details like
- DIY Deep Learning for Vision with Caffe
- Tutorial Documentation
- BVLC reference models and the community model zoo
- Installation instructions
and step-by-step examples.
Please join the caffe-users group or gitter chat to ask questions and talk about methods and models. Framework development discussions and thorough bug reports are collected on Issues.
Happy brewing!
License and Citation
Caffe is released under the BSD 2-Clause license. The BVLC reference models are released for unrestricted use.
Please cite Caffe in your publications if it helps your research:
@article{jia2014caffe,
Author = {Jia, Yangqing and Shelhamer, Evan and Donahue, Jeff and Karayev, Sergey and Long, Jonathan and Girshick, Ross and Guadarrama, Sergio and Darrell, Trevor},
Journal = {arXiv preprint arXiv:1408.5093},
Title = {Caffe: Convolutional Architecture for Fast Feature Embedding},
Year = {2014}
}