####This learning structure construction and feature extraction concept is based on Convolutional Neural Network, the main reference paper could be found at:
####This module uses Convolutional Neural Networks (Caffe) to train and recognize 3D poses of objects using triplet networks. The main reference paper can be found at:
####I implemented the training and feature extraction codes mainly based on CAFFE project(<http://caffe.berkeleyvision.org/>) which will be compiled as libcaffe for the cnn_3dobj OpenCV module, codes are mainly concentrating on triplet and pair-wise jointed loss layer, the training data arrangement is also important which basic training information.
####Codes about my triplet version of caffe are released on Github:
####This implements training and feature extraction code mainly using CAFFE (<http://caffe.berkeleyvision.org/>) which will be compiled as libcaffe for the cnn_3dobj OpenCV module. The code mainly concentrats on triplet networks using pair-wise jointed loss function layers. The training data arrangement is also important and there is basic information about that.
####Code for the triplet version of Caffe are on my (Yilda Wang's) Github:
####Prerequisite for this module: protobuf and caffe, for the libcaffe installation, you can install it on standard system path for the aim of being able to be linked by this OpenCV module when compiling and function using. Using: -D CMAKE_INSTALL_PREFIX=/usr/local as an building option when you cmake, the building process on Caffe on system could be like this:
####Prerequisites for this module are protobuf and Caffe. For libcaffe installation, you can install it on the standard system path so that it can be linked by this OpenCV module when compiling and linking using: -D CMAKE_INSTALL_PREFIX=/usr/local as an build option when you cmake. The building process on Caffe on system could be like this:
####After all these steps, the headers and libs of CAFFE will be set on /usr/local/ path, and when you compiling opencv with opencv_contrib modules as below, the protobuf and caffe will be recognized as already installed while building. Protobuf is needed.
####After the above steps, the headers and libs of CAFFE will be set on /usr/local/ path so that when you compile opencv with opencv_contrib modules as below, the protobuf and caffe libs will be recognized as already installed while building. Obviously protobuf is also needed.
#Compiling OpenCV
```
@ -35,13 +35,13 @@ $ make -j4
$ sudo make install
```
##Tips on compiling problems:
####If you encouter the no declaration errors when you 'make', it might becaused that you have installed the older version of cnn_3dobj module and the header file changed in a newly released version of codes. This problem is the cmake and make can't detect the header should be updated and it keeps the older header remains in /usr/local/include/opencv2 whithout updating. This error could be solved by remove the installed older version of cnn_3dobj module by:
####If you encouter the no declaration errors when you 'make', it might becaused that you have installed the older version of the cnn_3dobj module and the header file changed in a newly released version of the code. This problem is cmake and make can't detect that the header should be updated and it keeps the older header in /usr/local/include/opencv2 whithout updating. This error could be solved by removing the installed older version of the cnn_3dobj module using:
```
$ cd /
$ cd usr/local/include/opencv2/
$ sudo rm -rf cnn_3dobj.hpp
```
####And redo the compiling steps above again.
####And then redo the compiling steps above again.
####Imagas generation from different pose, by default there are 4 models used, there will be 276 images in all which each class contains 69 iamges, if you want to use additional .ply models, it is necessary to change the class number parameter to the new class number and also give it a new class label. If you will train net work and extract feature from RGB images set the parameter rgb_use as 1.
##Demo1: Training set data generation
####Image generation for different poses: by default, there are 4 models used and there will be 276 images in all in which each class contains 69 iamges. If you want to use additional .ply models, it is necessary to change the class number parameter to the new class number and also give it a new class label. If you want to train the network and extract feature from RGB images, set the parameter rgb_use as 1.
####When all images are created in images_all folder as a collection of training images for network tranining and as a gallery of reference images for the classification part, then proceed on.
####After this demo, the binary files of images and labels will be stored as 'binary_image' and 'binary_label' in current path, you should copy them into the leveldb folder in Caffe triplet training, for example: copy these 2 files in <caffe_source_directory>/data/linemod and rename them as 'binary_image_train', 'binary_image_test' and 'binary_label_train', 'binary_label_train'. Here I use the same as trianing and testing data, you can use different data for training and testing the performance in the CAFFE training process. It's important to observe the loss of testing data to check whether training data is suitable for the your aim. Loss should be obseved as keep decreasing and remain on a much smaller number than the initial loss.
####You could start triplet tranining using Caffe like this:
####When all images are created in images_all folder as a collection of training images for network tranining and as a gallery of reference images for classification, then proceed onward.
####After this demo, the binary files of images and labels will be stored as 'binary_image' and 'binary_label' in current path. You should copy them into the leveldb folder for Caffe triplet training. For example: copy these 2 files in <caffe_source_directory>/data/linemod and rename them as 'binary_image_train', 'binary_image_test' and 'binary_label_train', 'binary_label_train'. Here I use the same as trianing and testing data but you can use different data for training and testing. It's important to observe the error on the testing data to check whether the training data is suitable for the your aim. Error should be obseved to keep decreasing and remain much smaller than the initial error.
####Start triplet tranining using Caffe like this:
```
$ cd
$ cd <caffe_source_directory>
$ ./examples/triplet/create_3d_triplet.sh
$ ./examples/triplet/train_3d_triplet.sh
```
####After doing this, you will get .caffemodel files as the trained parameter of network. I have already provide the net definition .prototxt files and the pretrained .caffemodel in <opencv_contrib>/modules/cnn_3dobj/testdata/cv folder, you could just use them without training in caffe.
####After doing this, you will get .caffemodel files as the trained parameters of the network. I have already provide the network definition .prototxt files and the pretrained .caffemodel in <opencv_contrib>/modules/cnn_3dobj/testdata/cv folder, so you could just use them instead of training with Caffe.
$ cd <opencv_contrib>/modules/cnn_3dobj/samples/build
```
####Classifier, this will extracting the feature of a single image and compare it with features of gallery samples for prediction. This demo uses a set of images for feature extraction in a given path, these features will be a reference for prediction on target image. The caffe model and network prototxt file is attached in <opencv_contrib>/modules/cnn_3dobj/testdata/cv. Just run:
####Classification: This will extract features of a single image and compare it with features of a gallery of samples for prediction. This demo uses a set of images for feature extraction in a given path, these features will be a reference for prediction on the target image. The Caffe model and the network prototxt file are in <opencv_contrib>/modules/cnn_3dobj/testdata/cv. Just run:
```
$ ./classify_test
```
####if the classification and pose estimation issue need to extract mean got from all training images, you can run this:
####if you want to extract mean classification and pose estimation performance from all the training images, you can run this:
####This demo will have a test on the performance of trained CNN model on several images. If the the model fail on telling different samples from seperate classes or confused on samples with similar pose but from different classes, it will give some information on the model analysis.
##Demo3: Model performance test
####This demo will run a performance test of a trained CNN model on several images. If the the model fails on telling different samples from seperate classes apart, or is confused on samples with similar pose but from different classes, this will give some information for model analysis.
Gary Bradski
9 years ago