noelcodes / mask_rcnn

Exploring Mask_RCNN. Credits to matterport for codes and mark jay for the tutorial.

Home Page

Mask_RCNN

Curious what is Mask_RCNN, so I followed tutorial by Mark Jay and made it working on my laptop. So it is actually Faster RCNN (object detection with bounding boxes) + Masks

Reference

Paper by the Facebook AI Research (FAIR) : https://arxiv.org/pdf/1703.06870.pdf
Theory: https://www.youtube.com/watch?v=g7z4mkfRjI4
Tutorial : https://www.youtube.com/playlist?list=PLX-LrBk6h3wRAF22jBUxDgOvyhIgLN4Cg
Repo: https://github.com/matterport/Mask_RCNN
Repo: https://github.com/markjay4k/Mask-RCNN-series

Run on my laptop, see my demo video below:

Installation

Step 1: create a conda virtual environment with python 3.6

conda create -n MaskRCNN python=3.6 pip

Step 2: install the dependencies

pip install -r requirements.txt

Step 3: Clone the Mask_RCNN repo

git clone https://github.com/matterport/Mask_RCNN.git

Step 4: install pycocotools NOTE: pycocotools requires Visual C++ 2015 Build Tools

git clone https://github.com/philferriere/cocoapi.git
use pip to install pycocotools
pip install git+https://github.com/philferriere/cocoapi.git#subdirectory=PythonAPI

Step 5: download the pre-trained weights Go here https://github.com/matterport/Mask_RCNN/releases download the mask_rcnn_coco.h5 file place the file in the Mask_RCNN directory

Test on static images

open up the demo.ipynb and run it. On static images it works pretty well.

Test it on webcam

Credits to Mark Jay's youtube tutorial. Goto sample folder, Run "python visualize_cv2.py" in terminal. Stream on webcam is quite lagging. I think only 2fps. I read that the creator only got 5fps. I'm not sure if I use a light model will help or not. But I guess I can't change it since Mask is built on Faster-RCNN. Click below for my video demo.