louisenaud/faster_rcnn_minimum

A minimum PyTorch implementation of Faster R-CNN based on maskrcnn-benchmark

Faster R-CNN Inference with Minimum Implementation

A minimum implementation based on mask-rcnn-benchmark, aiming at easier understanding of the 2-stage object detection framework using Faster R-CNN as an example. See the Pipeline to understand how it works!

Let's make the framework more specific, and easier to understand!

What's New

• verified inference of Faster R-CNN C4 using pretrained weights

Inference

Easy enough!

python tools/demo.py --image {image file path} 

Installation

Docker

Build image with defaults (CUDA=9.0, CUDNN=7, FORCE_CUDA=1):

nvidia-docker build -t faster_rcnn_minimum docker/

Build image with other CUDA and CUDNN versions:

nvidia-docker build -t faster_rcnn_minimum --build-arg CUDA=9.2 --build-arg CUDNN=7 docker/

conda

conda create --name faster_rcnn_minimum
conda activate faster_rcnn_minimum

conda install ipython
pip install ninja yacs cython matplotlib tqdm
conda install -c pytorch pytorch-nightly torchvision cudatoolkit=9.0

python setup.py build develop