Skip to content
Permalink
Branch: master
Find file Copy path
Find file Copy path
Fetching contributors…
Cannot retrieve contributors at this time
221 lines (141 sloc) 8.3 KB

Semantic Segmentation in the Browser: DeepLab v3 Model

This package contains a standalone implementation of the DeepLab inference pipeline, as well as a demo, for running semantic segmentation using TensorFlow.js.

DeepLab Demo

Usage

In the first step of semantic segmentation, an image is fed through a pre-trained model based on MobileNet-v2. Three types of pre-trained weights are available, trained on Pascal, Cityscapes and ADE20K datasets.

To get started, pick the model name from pascal, cityscapes and ade20k, and decide whether you want your model quantized to 1 or 2 bytes (set the quantizationBytes option to 4 if you want to disable quantization). Then, initialize the model as follows:

import * as tf from '@tensorflow-models/tfjs';
import * as deeplab from '@tensorflow-models/deeplab';
const loadModel = async () => {
  const modelName = 'pascal';   // set to your preferred model, either `pascal`, `cityscapes` or `ade20k`
  const quantizationBytes = 2;  // either 1, 2 or 4
  return await deeplab.load({base: modelName, quantizationBytes});
};

const input = tf.zeros([227, 500, 3]);
// ...

loadModel()
    .then((model) => model.segment(input))
    .then(
        ({legend}) =>
            console.log(`The predicted classes are ${JSON.stringify(legend)}`));

By default, calling load initalizes the PASCAL variant of the model quantized to 2 bytes.

If you would rather load custom weights, you can pass the URL in the config instead:

import * as deeplab from '@tensorflow-models/deeplab';
const loadModel = async () => {
  const url = 'https://tfhub.dev/tensorflow/tfjs-model/deeplab/pascal/1/default/1/model.json?tfjs-format=file';
  return await deeplab.load({modelUrl: url});
};
loadModel().then(() => console.log(`Loaded the model successfully!`));

This will initialize and return the SemanticSegmentation model.

You can set the base attribute in the argument to pascal, cityscapes or ade20k to use the corresponding colormap and labelling scheme. Otherwise, you would have to provide those yourself during segmentation.

If you require more careful control over the initialization and behavior of the model (e.g. you want to use your own labelling scheme and colormap), use the SemanticSegmentation class, passing a pre-loaded GraphModel in the constructor:

import * as tfconv from '@tensorflow/tfjs-converter';
import * as deeplab from '@tensorflow-models/deeplab';
const loadModel = async () => {
  const base = 'pascal';        // set to your preferred model, out of `pascal`,
                                // `cityscapes` and `ade20k`
  const quantizationBytes = 2;  // either 1, 2 or 4
  // use the getURL utility function to get the URL to the pre-trained weights
  const modelUrl = deeplab.getURL(base, quantizationBytes);
  const rawModel = await tfconv.loadGraphModel(modelUrl);
  const modelName = 'pascal';  // set to your preferred model, out of `pascal`,
  // `cityscapes` and `ade20k`
  return new deeplab.SemanticSegmentation(rawModel);
};
loadModel().then(() => console.log(`Loaded the model successfully!`));

Use getColormap(base) and getLabels(base) utility function to fetch the default colormap and labelling scheme.

import {getLabels, getColormap} from '@tensorflow-models/deeplab';
const model = 'ade20k';
const colormap = getColormap(model);
const labels = getLabels(model);

Segmenting an Image

The segment method of the SemanticSegmentation object covers most use cases.

Each model recognises a different set of object classes in an image:

model.segment(image, config?) inputs

  • image :: ImageData | HTMLImageElement | HTMLCanvasElement | HTMLVideoElement | tf.Tensor3D;

    The image to segment

  • config.canvas (optional) :: HTMLCanvasElement

    Pass an optional canvas element as canvas to draw the output

  • config.colormap (optional) :: [number, number, number][]

    The array of RGB colors corresponding to labels

  • config.labels (optional) :: string[]

    The array of names corresponding to labels

    By default, colormap and labels are set according to the base model attribute passed during initialization.

model.segment(image, config?) outputs

The output is a promise of a DeepLabOutput object, with four attributes:

  • legend :: { [name: string]: [number, number, number] }

    The legend is a dictionary of objects recognized in the image and their colors in RGB format.

  • height :: number

    The height of the returned segmentation map

  • width :: number

    The width of the returned segmentation map

  • segmentationMap :: Uint8ClampedArray

    The colored segmentation map as Uint8ClampedArray which can be fed into ImageData and mapped to a canvas.

model.segment(image, config?) example

const classify = async (image) => {
    return await model.segment(image);
}

Note: For more granular control, consider predict and toSegmentationImage methods described below.

Producing a Semantic Segmentation Map

To segment an arbitrary image and generate a two-dimensional tensor with class labels assigned to each cell of the grid overlayed on the image (with the maximum number of cells on the side fixed to 513), use the predict method of the SemanticSegmentation object.

model.predict(image) input

  • image :: ImageData | HTMLImageElement | HTMLCanvasElement | HTMLVideoElement | tf.Tensor3D;

    The image to segment

model.predict(image) output

  • rawSegmentationMap :: tf.Tensor2D

    The segmentation map of the image

model.predict(image) example

const getSemanticSegmentationMap = (image) => {
    return model.predict(image)
}

Translating a Segmentation Map into the Color-Labelled Image

To transform the segmentation map into a coloured image, use the toSegmentationImage method.

toSegmentationImage(colormap, labels, segmentationMap, canvas?) inputs

  • colormap :: [number, number, number][]

    The array of RGB colors corresponding to labels

  • labels :: string[]

    The array of names corresponding to labels

  • segmentationMap :: tf.Tensor2D

    The segmentation map of the image

  • canvas (optional) :: HTMLCanvasElement

    Pass an optional canvas element as canvas to draw the output

toSegmentationImage(colormap, labels, segmentationMap, canvas?) outputs

A promise resolving to the SegmentationData object that contains two attributes:

  • legend :: { [name: string]: [number, number, number] }

    The legend is a dictionary of objects recognized in the image and their colors.

  • segmentationMap :: Uint8ClampedArray

    The colored segmentation map as Uint8ClampedArray which can be fed into ImageData and mapped to a canvas.

toSegmentationImage(colormap, labels, segmentationMap, canvas?) example

const base = 'pascal';
const translateSegmentationMap = async (segmentationMap) => {
  return await toSegmentationImage(
      getColormap(base), getLabels(base), segmentationMap)
}

Contributing to the Demo

Please see the demo documentation.

Technical Details

This model is based on the TensorFlow implementation of DeepLab v3. You might want to inspect the conversion script, or download original pre-trained weights here. To convert the weights locally, run the script as follows, replacing dist with the target directory:

./scripts/convert_deeplab.sh --target_dir ./scripts/dist

Run the usage helper to learn more about the options:

./scripts/convert_deeplab.sh -h
You can’t perform that action at this time.