react-native-vision

Library for accessing VisionKit and visual applications of CoreML from React Native. iOS Only

Incredibly super-alpha, and endeavors to provide a relatively thin wrapper between the underlying vision functionality and RN. Higher-level abstractions are @TODO and will be in a separate library.

Installation

yarn add react-native-vision react-native-swift
react-native link

Note react-native-swift is a peer dependency of react-native-vision.

If you are running on a stock RN deployment (e.g. from react-native init) you will need to make sure your app is targeting IOS 11 or higher:

yarn add react-native-fix-ios-version
react-native link

Since this module uses the camera, it will work much better on a device, and setting up permissions and codesigning in advance will help:

yarn add -D react-native-camera-ios-enable
yarn add -D react-native-setdevteam
react-native link
react-native setdevteam

Then you are ready to run!

react-native run-ios --device

Command line - adding a Machine Learning Model with add-mlmodel

react-native-vision makes it easier to bundle a pre-built machine learning model into your app.

After installing, you will find the following command available:

react-native add-mlmodel /path/to/mymodel.mlmodel

You may also refere to the model from a URL, which is handy when getting something off the interwebs. For example, to apply the pre-built mobileNet model from apple, you can:

react-native add-mlmodel https://docs-assets.developer.apple.com/coreml/models/MobileNet.mlmodel

Note that the name of your model in the code will be the same as the filename minus the "mlmodel". In the above case, the model in code can be referenced as "MobileNet"

Easy Start 1 : Full Frame Object Detection

One of the most common easy use cases is just detecting what is in front of you. For this we use the VisionCamera component that lets you apply a model and get the classification via render props.

Setup

react-native init imagedetector; cd imagedetector
yarn add react-native-swift react-native-vision
yarn add react-native-fix-ios-version react-native-camera-ios-enable react-native-setdevteam
react-native link
react-native setdevteam

Load your model with MobileNet

A free download from Apple!

react-native add-mlmodel https://docs-assets.developer.apple.com/coreml/models/MobileNet.mlmodel

Add Some App Code

import React from "react";
import { Text } from "react-native";
import { VisionCamera } from "react-native-vision";
export default () => (
  <VisionCamera style={{ flex: 1 }} classifier="MobileNet">
    {({ label, confidence }) => (
      <Text
        style={{
          width: "75%",
          fontSize: 50,
          position: "absolute",
          right: 50,
          bottom: 100
        }}
      >
        {label + " :" + (confidence * 100).toFixed(0) + "%"}
      </Text>
    )}
  </VisionCamera>
);

Easy Start 2: GeneratorView - for Style Transfer

Most machine learning application are classifiers. But generators can be useful and a lot of fun. The GeneratorView lets you look at style transfer models that show how you can use deep learning techniques for creating whole new experiences.

Setup

react-native init styletest; cd styletest
yarn add react-native-swift react-native-vision
yarn add react-native-fix-ios-version react-native-camera-ios-enable react-native-setdevteam
react-native link
react-native setdevteam

Load your model with add-mlmodel

Apple has not published a style transfer model, but there are a few locations on the web where you can download them. Here is one: https://github.com/mdramos/fast-style-transfer-coreml

So go to his github, navigate to his google drive, and then download the la_muse model to your personal Downloads directory.

react-native add-mlmodel ~/Downloads/la_muse.mlmodel

App Code

This is the insanely short part. Note that the camera view is not necessary for viewing the style-transferred view: its just for reference.

import React from "react";
import { GeneratorView, RNVCameraView } from "react-native-vision";
export default () => (
  <GeneratorView generator="FNS-The-Scream" style={{ flex: 1 }}>
    <RNVCameraView
      style={{
        position: "absolute",
        height: 200,
        width: 100,
        top: 0,
        right: 0
      }}
      resizeMode="center"
    />
  </GeneratorView>
);

Easy Start 3: Face Camera

Detect what faces are where in your camera view!

Taking a page (and the model!) from (https://github.com/gantman/nicornot)[Gant Laborde's NicOrNot app], here is the entirety of an app that discerns whether the target is nicolas cage.

Setup

react-native init nictest; cd nictest
yarn add react-native-swift react-native-vision
yarn add react-native-fix-ios-version react-native-camera-ios-enable react-native-setdevteam
react-native link
react-native setdevteam

Load your model with add-mlmodel

react-native add-mlmodel https://s3.amazonaws.com/despiteallmyrage/MegaNic50_linear_5.mlmodel

App Code

import React from "react";
import { Text, View } from "react-native";
import { FaceCamera } from "react-native-vision";
import { Identifier } from "react-native-identifier";
export default () => (
  <FaceCamera style={{ flex: 1 }} classifier="MegaNic50_linear_5">
    {({ face, faceConfidence, style }) =>
      face &&
      (face == "nic" ? (
        <Identifier style={{ ...style }} accuracy={faceConfidence} />
      ) : (
        <View
          style={{ ...style, justifyContent: "center", alignItems: "center" }}
        >
          <Text style={{ fontSize: 50, color: "red", opacity: faceConfidence }}>
            X
          </Text>
        </View>
      ))
    }
  </FaceCamera>
);

Face Detection Component Reference

FacesProvider

Context Provider that extends <RNVisionProvider /> to detect, track, and identify faces.

Props

Inherits from <RNVisionProvider />, plus:

• interval: How frequently (in ms) to run the face detection re-check. (Basically lower values here keeps the face tracking more accurate) Default: 500
• classifier: File URL to compiled MLModel (e.g. mlmodelc) that will be applied to detected faces
• updateInterval: How frequently (in ms) to update the detected faces - position, classified face, etc. Smaller values will mean smoother animation, but at the price of processor intensity. Default: 100

Example

<FacesProvider
  isStarted={true}
  isCameraFront={true}
  classifier={this.state.classifier}
>
  {/* my code for handling detected faces */}
</FacesProvider>

FacesConsumer

Consumer of <FacesProvider /> context. As such, takes no props and returns a render prop function.

Render Prop Members

• faces: Keyed object of information about the detected face. Elements of each object include:
  • region: The key associated with this object (e.g. faces[k].region === k)
  • x, y, height, width: Position and size of the bounding box for the detected face.
  • faces: Array of top-5 results from face classifier, with keys label and confidence
  • face: Label of top-scoring result from classifier (e.g. the face this is most likely to be)
  • faceConfidence: Confidence score of top-scoring result above.

Note that when there is no classifier specified, faces, face and faceConfidence are undefined

Face

Render prop generator to provision information about a single detected face. Can be instantiated by spread-propping the output of a single face value from <FacesConsumer> or by appling a faceID that maps to the key of a face. Returns null if no match.

Props

• faceID: ID of the face (corresponding to the key of the faces object in FacesConsumer)

Render Prop Members

• region: The key associated with this object (e.g. faces[k].region === k)
• x, y, height, width: Position and size of the bounding box for the detected face. Note These are adjusted for the visible camera view when you are rendering from that context.
• faces: Array of top-5 results from face classifier, with keys label and confidence
• face: Label of top-scoring result from classifier (e.g. the face this is most likely to be)
• faceConfidence: Confidence score of top-scoring result above. Note These arguments are the sam

Faces

A render-prop generator to provision information about all detected faces. Will map all detected faces into <Face> components and apply the children prop to each, so you have one function to generate all your faces. Designed to be similar to FlatMap implentation.

Required Provider Context

This component must be a descendant of a <FacesProvider>

Props

None

Render Prop Members

Same as <Face> above, but output will be mapped across all detected faces.

Example of use is in the primary Face Recognizer demo code above.

Props

• faceID: ID of the face applied.
• isCameraView: Whether the region frame information to generate should be camera-aware (e.g. is it adjusted for a preview window or not)

Render Props

This largely passes throught the members of the element that you could get from the faces collection from FaceConsumer, with the additional consideration that when isCameraView is set,

• style: A spreadable set of styling members to position the rectangle, in the same style as a RNVCameraRegion

If faceID is provided but does not map to a member of the faces collection, the function will return null.

Core Component References

The package exports a number of components to facilitate the vision process. Note that the <RNVisionProvider /> needs to be ancestors to any others in the tree. So a simple single-classifier using dominant image would look something like:

<RNVisionProvider isStarted={true}>
  <RNVDefaultRegion classifiers={[{url: this.state.FileUrlOfClassifier, max: 5}]}>
  {({classifications})=>{
    return (
      <Text>
        {classifications[this.state.FileUrlOfClassifier][0].label}
      </Text>
  }}
  </RNVDefaultRegion>
</RNVisionProvider>

RNVisionProvider

Context provider for information captured from the camera. Allows the use of regional detection methods to initialize identification of objects in the frame.

Props

• isStarted: Whether the camera should be activated for vision capture. Boolean
• isCameraFront: Facing of the camera. False for the back camera, true to use the front. Note only one camera facing can be used at a time. As of now, this is a hardware limitation.
• regions: Specified regions on the camera capture frame articulated as {x,y,width,height} that should always be returned by the consumer
• trackedObjects: Specified regions that should be tracked as objects, so that the regions returned match these object IDs and show current position.
• onRegionsChanged: Fires when the list of regions has been altered
• onDetectedFaces: Fires when the number of detected faces has changed

Class imperative member

• detectFaces: Triggers one call to detect faces based on current active frame. Directly returns locations.

RNVisionConsumer

Consumer partner of RNVisionProvider. Must be its descendant in the node tree.

Render Prop Members

• imageDimensions: Object representing size of the camera frame in {width, height}
• isCameraFront: Relaying whether camera is currently in selfie mode. This is important if you plan on displaying camera output, because in selfie mode a preview will be mirrored.
• regions: The list of detected rectangles in the most recently captured frame, where detection is driven by the RNVisionProvider props

RNVRegion

Props

• region: ID of the region (Note the default region, which is the whole frame, has an id of "" - blank.)
• classifiers: CoreML classifiers passed as file URLs to the classifier mlmodelc itself. Array
• generators: CoreML image generators passed as file URLs to the classifier mlmodelc itself. Array
• generators: CoreML models that generate a collection of output values passed as file URLs to the classifier mlmodelc itself.
• bottlenecks: A collection of CoreML models that take other CoreML model outputs as their inputs. Keys are the file URLs of the original models (that take an image as their input) and values are arrays of mdoels that generate the output passed via render props.
• onFrameCaptured: Callback to fire when a new image of the current frame in this region has been captured. Making non-null activates frame capture, setting to null turns it off. The callback passes a URL of the saved frame image file.

Render Prop members

• key: ID of the region
• x, y, width, height: the elements of the frame containing the region. All values expressed as percentages of the overall frame size, so a 50x100 frame at origin 5,10 in a 500x500 frame would come across as {x: 0.01, y: 0.02, width: .1, height: .2}. Changes in these values are often what drives the re-render of the component (and therefore re-run of the render prop)
• confidence: If set, the confidence that the object identified as key is actually at this location. Used by tracked objects API of iOS Vision. Sometimes null.
• classifications: Collection, keyed by the file URL of the classifier passed in props, of collections of labels and probabilities. (e.g. {"file:///path/to/myclassifier.mlmodelc": {"label1": 0.84, "label2": 0.84}})
• genericResults: Collection of generic results returned from generic models passed in via props to the region

RNVDefaultRegion

Convenience region that references the full frame. Same props as RNVRegion, except region is always set to "" - the full frame. Useful for simple style transfers or "dominant image" classifiers.

Props

Same as RNVRegion, with the exception that region is forced to ""

Render Prop Members

Same as RNVRegion, with the note that key will always be ""

RNVCameraView

Preview of the camera captured by the RNVisionProvider. Note that the preview is flipped in selfie mode (e.g. when isCameraFront is true)

Props

The properties of a View plus:

• gravity: how to scale the captured camera frame in the view. String. Valid values:
  • fill: Fills the rectangle much like the "cover" in an Image
  • resize: Leaves transparent (or style:{backgroundColor}) the parts of the rectangle that are left over from a resized version of the image.

RNVCameraConsumer

Render prop consumer for delivering additional context that regions will find helpful, mostly for rendering rectangles that map to the regions identified.

Render Prop Members

• viewPortDimensions: A collection of {width, height} of the view rectangle.
• viewPortGravity: A pass-through of the gravity prop to help decide how to manage the math converting coordinates.

RNVCameraRegion

A compound consumer that blends the render prop members of RNVRegion and RNVCameraConsumer and adds a style prop that can position the region on a specified camera preview

Props

Same as RNVRegion

Render Prop Members

Includes members from RNVRegion and RNVCameraConsumer and adds:

• style: A pre-built colleciton of style prop members {position, width, height, left, top} that are designed to act in the context of the RNVCameraView rectangle. Spread-prop with your other style preferences (border? backgroundColor?) for easy on-screen representation.

RNVImageView

View for displaying output of image generators. Link it to , and the resulting image will display in this view. Useful for style transfer models. More performant because there is no round trip to JavaScript notifying of each frame update.

Props

• id: the ID of an image generator model attached to a region. Usually is the file:/// URL of the .mlmodelc.

Otherwise conforms to Image and View API.