Faster-RCNN adaption based on tf-faster-rcnn by endernewton built partly on files by Ross Girshick. Modified by me Tara1000 to implement two new network models, ResNet-41 and Deeper MobileNet with the purpose of testing performance and accuracy of these modifications. Pre-trained ImageNet weights added from TensorFlow or endernewton. Objectfasterrcnn is tested only on Ubuntu 18-04 on NVidia GTX 1080 Ti.
More details: https://github.com/endernewton/tf-faster-rcnn
COCO 2014 data set can be downloaded from http://cocodataset.org/#download
If you lack pip or conda packages in Anaconda-3 to run the code, see packagelist.txt at the google drive for package versions, or setup a virtual environment with the yaml file included in this repository:
> conda env create -f tfgpuenv.yaml
> conda activate tensorflow\_gpuenv
> cd lib
If your GPU is not a NVidia 1080 (Ti) change the architecture accordingly
> vim setup.py
Change sm_61 in line 130 to the best fit as shown in tf-faster-rcnn
In lib folder, proceed to build the project:
> make clean
> make
> cd ..
Download ImageNet weights (data) and pre-trained COCO models (output) from Faster-RCNN folder at https://drive.google.com/open?id=1qlXjT-P6vgZk6NylAo7JeJREKY3jccNV into project root, and clone + build coco API:
> cd data
> git clone https://github.com/pdollar/coco.git
> cd coco/PythonAPI
> make
> cd ..
In coco create symlinks of 'images' and 'annotations' folders from your data set folder to here:
> ln -s your/path/to/coco2014/annotations ./annotations
> ln -s your/path/to/coco2014/images ./images
> cd ../..
To run validation run the test_faster_rcnn.sh script with parameter 0 for default GPU 0 and e.g mobile for MobileNet v.1 or res41 for ResNet-41:
> ./experiments/scripts/test\_faster\_rcnn.sh 0 coco mobile
To train a network on GPU 0, move or remove its COCO pretrained weights folder from output, e.g. res41 for ResNet-41, and run:
> ./experiments/scripts/train\_faster\_rcnn.sh 0 coco res41