Tensorflow Faster R-CNN for Windows by using Python 3.5
This is the branch to compile Faster R-CNN on Windows. It is heavily inspired by the great work done here and here. I have not implemented anything new but I fixed the implementations for Windows and Python 3.5.
PLEASE BE AWARE: I do not have time or intention to fix all the issues for this branch as I do not use it commercially. I created this branch just for fun. If you want to make any commitment, it is more than welcome. Tensorflow has already released an object detection api. Please refer to it. https://github.com/tensorflow/models/tree/master/research/object_detection
Also, instead of trying to deal with Tensorflow, use Chainer. It is ready to be used with all the common models https://github.com/chainer/chainercv & https://github.com/chainer/chainer . I can reply all of your questions about Chainer
1- Install tensorflow, preferably GPU version. Follow instructions. If you do not install GPU version, you need to comment out all the GPU calls inside code and replace them with relavent CPU ones.
2- Install python packages (cython, python-opencv, easydict)
3- Checkout this branch
4- Go to ./data/coco/PythonAPI
Run python setup.py build_ext --inplace
Run python setup.py build_ext install
Go to ./lib/utils and run python setup.py build_ext --inplace
5- Follow this instruction to download PyCoco database. Link
I will be glad if you can contribute with a batch script to automatically download and fetch. The final structure has to look like
"data/VOCDevkit2007/annotations_cache"
"data/VOCDevkit2007/VOC2007"
6- Download pre-trained VGG16 from here and place it as "data\imagenet_weights\vgg16.ckpt"
For rest of the models, please check here
7- Run train.py
Notify me if there is any issue found. Please note that, I have compiled cython modules with sm61 architecture (GTX 1060, 1070 etc.). Compile support for other architectures will be added.