![\"\"\/](\"https:\/\/mlwhiz.com\/images\/weapons\/main.png\")
<\/p>\n
Object Detection is a helpful tool to have in your coding repository.<\/p>\n
It forms the backbone of many fantastic industrial applications. Some of them being self-driving cars, medical imaging and face detection.<\/p>\n
In my last post<\/a> on Object detection, I talked about how Object detection models evolved.<\/p>\n But what good is theory, if we can\u2019t implement it?<\/p>\n This post is about implementing and getting an object detector on our custom dataset of weapons.<\/em><\/strong><\/p>\n The problem we will specifically solve today is that of Instance Segmentation using Mask-RCNN.<\/p>\n Can we create<\/em> masks<\/em><\/strong> for each object in the image? Specifically something like:<\/em><\/p>\n The most common way to solve this problem is by using Mask-RCNN. The architecture of Mask-RCNN looks like below:<\/p>\n Essentially, it comprises of:<\/p>\n A backbone network like resnet50\/resnet101<\/p>\n<\/li>\n A Region Proposal network<\/p>\n<\/li>\n ROI-Align layers<\/p>\n<\/li>\n Two output layers \u2014 one to predict masks and one to predict class and bounding box.<\/p>\n<\/li>\n<\/ul>\n There is a lot more to it. If you want to learn more about the theory, read my last post\u2013 This post is mostly going to be about the code<\/a>.<\/p>\n The use case we will be working on is a weapon detector. A weapon detector is something that can be used in conjunction with street cameras as well as CCTV\u2019s to fight crime. So it is pretty nifty.<\/p>\n So, I started with downloading 40 images each of guns and swords from the open image dataset<\/a> and annotated them using the VIA tool. Now setting up the annotation project in VIA is petty important, so I will try to explain it step by step.<\/p>\n VIA is an annotation tool, using which you can annotate images both bounding boxes as well as masks. I found it as one of the best tools to do annotation as it is online and runs in the browser itself.<\/p>\n To use it, open http:\/\/www.robots.ox.ac.uk\/~vgg\/software\/via\/via.html<\/a><\/p>\n You will see a page like:<\/p>\n The next thing we want to do is to add the different class names in the region_attributes. Here I have added \u2018gun\u2019 and \u2018sword\u2019 as per our use case as these are the two distinct targets I want to annotate.<\/p>\n I have kept all the files in the folder data. Next step is to add the files we want to annotate. We can add files in the data folder using the \u201cAdd Files\u201d button in the VIA tool. And start annotating along with labels as shown below after selecting the polyline tool.<\/p>\n Click on save project on the top menu of the VIA tool.<\/p>\n Save file as via_region_data.json by changing the project name field. This will save the annotations in COCO format.<\/p>\n We will need to set up the data directories first so that we can do object detection. In the code below, I am creating a directory structure that is required for the model that we are going to use.<\/p>\n After running the above code, we will get the data in the below folder structure:<\/p>\n We will use the code from the matterport\/Mask_RCNN<\/a> GitHub repository. You can start by cloning the repository and installing the required libraries.<\/p>\n Once we are done with installing the dependencies and cloning the repo, we can start with implementing our project.<\/p>\n We make a copy of the samples\/balloon directory in Mask_RCNN folder and create a samples\/guns_and_swords<\/em><\/strong> directory where we will continue our work:<\/p>\n We start by renaming and changing balloon.py in the And some changes in the train function<\/p>\n<\/li>\n<\/ol>\n Showing only the changed Instance Segmentation<\/h2>\n
![\"\"\/](\"https:\/\/mlwhiz.com\/images\/weapons\/0.png\")
<\/p>\n\"\/](\"https:\/\/mlwhiz.com\/images\/weapons\/1.png\")
Source<\/a><\/em><\/p>\n\n
\nDemystifying Object Detection and Instance Segmentation for Data Scientists<\/a><\/p>\n1. Creating your Custom Dataset for Instance Segmentation<\/h2>\n
![\"Our](\"https:\/\/mlwhiz.com\/images\/weapons\/2.png\")
<\/p>\n1. Set up VIA<\/h3>\n
![\"\"\/](\"https:\/\/mlwhiz.com\/images\/weapons\/3.png\")
<\/p>\n![\"\"\/](\"https:\/\/mlwhiz.com\/images\/weapons\/4.png\")
<\/p>\n2. Annotate the Images<\/h3>\n
![\"Click,](\"https:\/\/mlwhiz.com\/images\/weapons\/5.png\")
<\/p>\n3. Download the annotation file<\/h3>\n
![\"\"\/](\"https:\/\/mlwhiz.com\/images\/weapons\/6.png\")
<\/p>\n4. Set up the data directory structure<\/h3>\n
from<\/span> random import<\/span> random\nimport<\/span> os\nfrom<\/span> glob import<\/span> glob\nimport<\/span> json\n# Path to your images<\/span>\nimage_paths =<\/span> glob(\"data\/*\"<\/span>)\n#Path to your annotations from VIA tool<\/span>\nannotation_file =<\/span> 'via_region_data.json'<\/span>\n#clean up the annotations a little<\/span>\nannotations =<\/span> json.<\/span>load(open(annotation_file))\ncleaned_annotations =<\/span> {}\nfor<\/span> k,v in<\/span> annotations['_via_img_metadata'<\/span>].<\/span>items():\n cleaned_annotations[v['filename'<\/span>]] =<\/span> v\n# create train and validation directories<\/span>\n!<\/span> mkdir procdata\n!<\/span> mkdir procdata\/<\/span>val\n!<\/span> mkdir procdata\/<\/span>train\ntrain_annotations =<\/span> {}\nvalid_annotations =<\/span> {}\n# 20% of images in validation folder<\/span>\nfor<\/span> img in<\/span> image_paths:\n # Image goes to Validation folder<\/span>\n if<\/span> random()<<\/span>0.2<\/span>:\n os.<\/span>system(\"cp \"<\/span>+<\/span> img +<\/span> \" procdata\/val\/\"<\/span>)\n img =<\/span> img.<\/span>split(\"\/\"<\/span>)[-<\/span>1<\/span>]\n valid_annotations[img] =<\/span> cleaned_annotations[img]\n else<\/span>:\n os.<\/span>system(\"cp \"<\/span>+<\/span> img +<\/span> \" procdata\/train\/\"<\/span>)\n img =<\/span> img.<\/span>split(\"\/\"<\/span>)[-<\/span>1<\/span>]\n train_annotations[img] =<\/span> cleaned_annotations[img]\n# put different annotations in different folders<\/span>\nwith<\/span> open('procdata\/val\/via_region_data.json'<\/span>, 'w'<\/span>) as<\/span> fp:\n json.<\/span>dump(valid_annotations, fp)\nwith<\/span> open('procdata\/train\/via_region_data.json'<\/span>, 'w'<\/span>) as<\/span> fp:\n json.<\/span>dump(train_annotations, fp)<\/code><\/pre>\n<\/div>\n- procdata\n - train\n - img1.jpg\n - img2.jpg\n - via_region_data.json\n - val\n - img3.jpg\n - img4.jpg\n - via_region_data.json\n<\/code><\/pre>\n2. Setup the Coding Environment<\/h2>\n
git clone https:\/\/github.com\/matterport\/Mask_RCNN\ncd Mask_RCNN\npip install -r requirements.txt<\/code><\/pre>\n<\/div>\ncp -r samples\/balloon samples\/guns_and_swords<\/code><\/pre>\n<\/div>\nSetting up the Code<\/h3>\n
samples\/guns_and_swords<\/code> directory to gns.py<\/code>. The balloon.py<\/code> file right now trains for one target. I have extended it to use multiple targets. In this file, we change:<\/p>\n\n
balloonconfig<\/code> to gnsConfig<\/code><\/p>\n<\/li>\nBalloonDataset<\/code> to gnsDataset<\/code> : We changed some code here to get the target names from our annotation data and also give multiple targets.<\/p>\n<\/li>\ngnsConfig<\/code> here to get you an idea. You can take a look at the whole gns.py<\/a> code here.<\/p>\nclass<\/span> gnsConfig<\/span>(Config):\n \"\"\"Configuration for training on the toy dataset.\n<\/span> Derives from the base Config class and overrides some values.\n<\/span> \"\"\"<\/span>\n # Give the configuration a recognizable name<\/span>\n NAME =<\/span> \"gns\"<\/span>\n # We use a GPU with 16GB memory, which can fit three image.<\/span>\n # Adjust down if you use a smaller GPU.<\/span>\n IMAGES_PER_GPU =<\/span> 3<\/span>\n # Number of classes (including background)<\/span>\n NUM_CLASSES =<\/span> 1<\/span> +<\/span> 2<\/span> # Background + sword + gun<\/span>\n # Number of training steps per epoch<\/span><\/code><\/pre>\n<\/div>\n3. Visualizing Images and Masks<\/h2>\n
![\"\"\/](\"https:\/\/mlwhiz.com\/images\/weapons\/a.png\")
<\/p>\n![\"\"\/](\"https:\/\/mlwhiz.com\/images\/weapons\/12.png\")
<\/p>\n![\"\"\/](\"https:\/\/mlwhiz.com\/images\/weapons\/13.png\")
<\/p>\n![\"\"\/](\"https:\/\/mlwhiz.com\/images\/weapons\/b.png\")
<\/p>\n