This repo is a Python3 implementation of Faster-RCNN_TF by smallcorgi which implements Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks by Shaoqing Ren, Kaiming He, Ross Girshick, Jian Sun.

The full repository has been made compaitable for python3, and if you want to train a network you can follow the following instructions. In case, if you want to do the same using python2, then you can fork or download the following repo.

https://github.com/dxyang/Faster-RCNN-COCO_TF

1. Requirements for Tensorflow (see: Tensorflow)

2. Python packages you might not have: pillow scipy cython, opencv3, easydict

This repo has been tested using python3.5, tensorflow1.4, cudNN-6.

1. Clone the PY3-Faster-RCNN-TF repository

   # Make sure to clone with --recursive
   git clone --recursive https://github.com/surajitsaikia27/PY3-Faster-RCNN-TF

2. Build the Cython modules

   cd $PY3-Faster-RCNN-TF/lib
   make

3. To run the demo:

   Download the COCO trained pretrained model from https://drive.google.com/file/d/0Bw0qMqgwZcafZlRqRDYxSnBkNFE/view

   rename VGGnet_fast_rcnn_iter_490000.ckpt.data-00000-of-00001 as VGGnet_fast_rcnn_iter_490000.ckpt

   Run the following command.

 python ./tools/demo.py --model "path to the VGGnet_fast_rcnn_iter_490000.ckpt"  --net VGGnet_test  --img-path path to image folder

The results will be saved in a directory called detections_test

Follow the instructions to train your own model.

1. Build pycocotools modules

   cd $PY3-Faster-RCNN-TF/lib
   git clone https://github.com/cocodataset/cocoapi.git
   cd cocoapi/PythonAPI
   make
   cd ../..
   mv cocoapi/PythonAPI/pycocotools pycocotools
   rm -rf cocoapi

2. Build the Cython modules

   cd $PY3-Faster-RCNN-TF/lib
   make

3. Install gsutil if you haven't already

   curl https://sdk.cloud.google.com | bash

4. Download the training, validation, test data for MS COCO

   cd $PY3-Faster-RCNN-TF/data
   mkdir coco; cd coco
   mkdir images; cd images
   mkdir train2014
   mkdir test2014
   mkdir val2014
   gsutil -m rsync gs://images.cocodataset.org/train2014 train2014
   gsutil -m rsync gs://images.cocodataset.org/test2014 test2014
   gsutil -m rsync gs://images.cocodataset.org/val2014 val2014

5. Download the annotations for MS COCO and unzip

   cd $PY3-Faster-RCNN-TF/data
   gsutil -m rsync gs://images.cocodataset.org/annotations coco
   cd coco
   unzip annotations_trainval2014.zip
   unzip image_info_test2014.zip
   rm *.zip

6. Download the annotations for the 5000 image minival subset of COCO val2014 as mentioned here

   cd $PY3-Faster-RCNN-TF/data/coco/annotations
   wget https://dl.dropboxusercontent.com/s/o43o90bna78omob/instances_minival2014.json.zip
   wget https://dl.dropboxusercontent.com/s/s3tw5zcg7395368/instances_valminusminival2014.json.zip
   unzip instances_minival2014.json.zip; rm instances_minival2014.json.zip
   unzip instances_valminusminival2014.json.zip; rm instances_valminusminival2014.json.zip

7. Download the pre-trained ImageNet model [Google Drive] [Dropbox]

   cd $PY3-Faster-RCNN-TF
   wget https://www.dropbox.com/s/po2kzdhdgl4ix55/VGG_imagenet.npy
   mkdir data/pretrain_model
   mv VGG_imagenet.npy data/pretrain_model/VGG_imagenet.npy

8. Create an output directory for log files

   cd $PY3-Faster-RCNN-TF
   mkdir experiments/logs

9. Run script to train and test model

   cd $PY3-Faster-RCNN-TF
   ./experiments/scripts/faster_rcnn_end2end.sh $DEVICE $DEVICE_ID VGG16 coco

• DEVICE is either cpu/gpu

Run the following command.

python ./tools/test_net.py \
		--device gpu \
		--device_id 0 \
		--weights output/faster_rcnn_end2end/coco_2014_train/VGGnet_fast_rcnn_iter_490000.ckpt \
		--cfg experiments/cfgs/faster_rcnn_end2end.yml \
		--imdb coco_2014_minival \
		--network VGGnet_test	\
		--vis False

• Changing vis to True will save images with all detections above 0.8 for every image in the testing set.
• The checkpoint files folder contains the following:

   cd output/faster_rcnn_end2end/coco_2014_train
   ls
   # VGGnet_fast_rcnn_iter_490000.ckpt.data-00000-of-00001
   # VGGnet_fast_rcnn_iter_490000.ckpt.index
   # VGGnet_fast_rcnn_iter_490000.ckpt.meta