tf-faster-rcnn

Tensorflow Faster R-CNN for Windows by using Python 3.5. By default it uses Imagenet database.

This is the repository 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.

Installation

1- Install tensorflow, preferably GPU version. Follow instructions.

2- Install python packages (cython, python-opencv, easydict)

3- Clone this repository

4- Move to data/coco/PythonAPI and launch

python setup.py build_ext --inplace
python setup.py build_ext install

Then in lib/utils

python setup.py build_ext --inplace

5- Download pre-trained VGG16 from here and place it as "data\imagenet_weights\vgg16.ckpt"

For rest of the models, please check here

6- Run train.py

Notify me if there is any issue

About the net

The train downloads from ImageNet database images of the following classes/synset (defined in lib/dataset/imagenet.py)

CLASSES = {'synthesizer':'n04376400', 'pipe organ':'n03854065', 'music box': 'n03801353', 'electric guitar':'n03272010', 'sax':'n04141076', 'ocarina':'n03840681', 'harmonica':'n03494278', 'acoustic guitar':'n02676566', 'trombone':'n04487394','gong':'n03447721', 'maraca':'n03720891', 'xylophone':'n03721384', 'pianoforte':'n03928116'}

If you want to change the classes you should change it and the tuple after it. Also you must put the extracted annotations in data/imagenet/Annotation_imagenet (pull requests are welcome).

Results

After a session of 10000 iterations (it took less than 1 day on a Nvidia GTX 980) these are the results (obtained by running demo.py)

while these are the metrics visualized on tensorboard