Dilate and Erode

edit

My goal is to get bulding contours. This is the input image. I am using a single HSV channel My code (most important pieces for char limit on comment).

Imgproc.medianBlur(hsv, median, 1);
Imgproc.threshold(median, bin, 30, 255, Imgproc.THRESH_BINARY);
Imgproc.Sobel(bin, grad_x, CvType.CV_16S, 1, 0, 3, 1, 0, Core.BORDER_DEFAULT);
Core.convertScaleAbs(grad_x, abs_grad_x);
Imgproc.Sobel(bin, grad_y, CvType.CV_16S, 0, 1, 3, 1, 0, Core.BORDER_DEFAULT);
Core.convertScaleAbs(grad_y, abs_grad_y);

Imgproc.findContours(grad, contours, new Mat(), Imgproc.RETR_LIST, Imgproc.CHAIN_APPROX_NONE);

Processing a Sobel filter on a binarized image makes little sense.

You can try to smooth the input image a little bit using cv::GaussianBlur and than find the edges using cv::Canny on the colour image. After that you can use cv::findContours.

Otherwise you can play around with cv::adaptiveThreshold on the grayscale image and find contours right after that. If you have merged or broken contours you can try to use cv::erode or cv::dilate before finding contours.

if I well understand you are refeferring to the original image ? Do you discourage the use of HSV channel, as I am trying ? What do you think if I try to detect buildings using inRange ? (I see that on Hsv channel buildings are close to black color

For Canny it is the best using the colour image because the algorithm uses the strongest gradient for each pixel out of the 3 channels.

Using adaptiveThreshold makes the binarization more robust against inhomogeneous illumination compared to a simple threshold. For this you need a grayscale/single-channel image which is the V-channel in HSV image for example, but you can play around with the other channels, too. You can use what ever fits best, these are just some ideas you could try.