DNN.net.forward() in Android

edit

odd image sizes are not allowed with enet, try like Size(512,256) instead (also needs to be a multiple of 4)

since you cannot access a 4d tensor from java directly, you'll have to reshape() the network output to a 2d strip of probability maps per class, like:

---
|A| 
---
|B|
---
|C|
---

now you have to find the highest probability for each pixel, and knowing which class it came from, assign a pixel color to your segmentation image:

import org.opencv.core.*;
import org.opencv.dnn.*;
import org.opencv.imgcodecs.*;
import org.opencv.imgproc.*;

public class SimpleSample {
    public static void main(String[] args) {
        System.loadLibrary(Core.NATIVE_LIBRARY_NAME);
        Scalar[] colors = new Scalar[]{ // obvious TODO: find better colours ;)
            new Scalar(1,22,123), new Scalar(28,53,4), new Scalar(81,2,243),
            new Scalar(2,223,244), new Scalar(12,2,3), new Scalar(2,3,4), 
            new Scalar(161,162,3), new Scalar(124,13,144), new Scalar(112,2,223),
            new Scalar(2,3,4), new Scalar(81,82,3), new Scalar(222,223,24),
            new Scalar(12,222,32), new Scalar(82,3,4), new Scalar(221,12,3), 
            new Scalar(182,83,84), new Scalar(221,12,3), new Scalar(52,3,4), 
            new Scalar(91,162,23), new Scalar(211,3,49)
        };
        Mat img = Imgcodecs.imread("c:/data/img/warp.png"); // your data here !
        Net net = Dnn.readNetFromTorch("c:/data/mdl/enet-model-best.net");
        Mat inputBlob = Dnn.blobFromImage(img, 
                    1.0/255, new Size(512,256), new Scalar(0, 0, 0), true, false);
        net.setInput(inputBlob);
        Mat res = net.forward();
        // the net's output is a list of "probability maps", one per class 
        //   (size(1) is, how many classes there are, 2 and 3 are H,W)
        System.out.println(res.size(0)+" "+res.size(1)+" "+res.size(2)+" "+res.size(3));
        // make a long, vertical "image strip" of it 
        //   (since we can't access 4d tensors from java)
        Mat strip = res.reshape(1, res.size(1) * res.size(2));
        // will keep "highest probability" per pixel
        Mat probs = new Mat(res.size(2), res.size(3), res.type(), Scalar.all(0));
        // color overlay
        Mat segm  = new Mat(probs.size(), CvType.CV_8UC3);
        // check each map, keep the pixel with highest probability 
        //   and (re)assign a color to it
        for (int i=0; i<res.size(1); i++) {
            // probs for class i
            Mat sub = strip.submat(i*res.size(2), (i+1)*res.size(2), 0, res.size(3));
            // find out, which pixels int the current map had a higher probability,
            //   than all of the prev. ones
            Mat gt = new Mat(); // 'greater than all' mask
            Core.compare(sub, probs, gt, Core.CMP_GT);
            // update probs and segmentation
            segm.setTo(colors[i], gt);
            sub.copyTo(probs, gt);
        }
        Mat result = new Mat();
        Imgproc.resize(segm, segm, img.size());
        Core.addWeighted(img, 0.8, segm, 0.2, 0, result); // alpa + beta == 1
        Imgcodecs.imwrite("segm.png", result);
    }
}