Hi, I've spent 3 years of PhD in image segmentation and my first question is: is the problem image segmentation (like http://goo.gl/zqJ31G) or object segmentation (like http://goo.gl/qtKi8s)?

Is it automatic or interactive?

Depending on these information there are plenty of datasets with available ground truth that you can use for comparison.

If you want to try image segmentation algorithms on your images, you may be interested in LabelME, an online tool for fast image annotation: http://labelme.csail.mit.edu/Release3.0/

Several segmentation methods can be used to perform your request, among them:

Utilizing supervised classification is one of the promising method, or see my published paper used the scatterplot method for this reason.

If you want to create your own data set for segmentation, you can do it.

You need to see what are different components or parts in the image, that are of interest in your application. For example, in automated visual inspection, you may note the metal parts which need to be examined. These parts of the image are considered foreground, area of interest, and the rest of the image is consider background. Background is also important as it helps to separate the foreground parts in the image.

There's no secret for that: you should segment the images manually, and the best is to build a database by asking several persons to do a manual segmentation of the images, then try to find a consensus. This, or similar, has already been done (I think by Jitendra Malik's group); there's a lot of literature on how to evaluate segmentation results, as well as testbed databases.

hello 

i suggest you this database, where images with 3 different ground truths ( human segmentation). it is useful for segmentation evaluation. http://www.wisdom.weizmann.ac.il/~vision/Seg_Evaluation_DB/1obj/index.html

regards 

If you are using Mac OS X, you can use RectLabel.

An image annotation tool to label images for bounding box object detection and segmentation.

Key features:

- Drawing bounding box, polygon, and cubic bezier

- Export index color mask image and separated mask images

- 1-click buttons make your labeling work faster

- Customize the label dialog to combine with attributes

- Settings for objects, attributes, hotkeys, and labeling fast

- Layer order for overlapped boxes

- Quick zoom to existing boxes

- Support the PASCAL VOC format

https://rectlabel.com

I'd recommend to try the Image Labeler app in MATLAB. it is enables to label ground truth data by interactively label (ROI) for pixels for semantic segmentation. I have been used it and it gave good results.

Thanks Prof Kamil. Useful answer.

You can also use the inkscape software to do one by following the mask creation procedure.

Best Image segmentation may be performed by several ways: It may be achieved by image binarization (using useful threshold) , or by utilizing one of the well known image edge detection methods (e.g. differential masks, or band-pass filtered methods like Marr and Hilderith ... etc.) . Or segmentation can be carry on by using image classification (e.g. Mean value, parallel piped, scatter scope, etc.). For more details see my site on R.G. With best wishes

Diffgram https://diffgram.com (where I work).

• Workflow. Manage tasks at scale with Jobs.
• Validation. Pre-label images with your own model or use built in automatic training.
• Video support. Video frames with interpolation can save a ton of time!

Diffgram is web based:

• Fast to get started - create an account and start in moments.
• Share work with teammates for review

We have an integrated SDK that makes it easier to import your data.

https://medium.com/diffgram/how-to-validate-your-deep-learning-model-with-the-diffgram-sdk-tutorial-22234a9a35

https://github.com/diffgram/diffgram/tree/master/sdk

https://medium.com/diffgram/fast-annotation-net-a-framework-for-active-learning-in-2018-1c75d6b4af92

i have to use segmented food image for volume estimation and calorie prediction , I am working on Food-101 dataset ,so how can i prepare my dataset that is suiatble for segmentaion??

Hi. The easiest way to obtain a ground truth segmentation from a give image is to use thresh-holding technique.

You can simply try various thresh-hold values and ask few layman for perfect result. In this way you can get an optimal value for the thresh-hold.

Example (Matlab environment):

u= imread('image.extension');

imshow(u>T); % TRY using different values for T for which you get good result

% later on you can save your segmentation as

seg=(u>T); % where T is your obtained suitable value...

For any help you can email me at

dr.haider at uop.edu.pk

may i know the real definition og the ground truth of orignal image?

thank you

Based on the anatomical organ of interest, the standard database are there with ground truth images.

The ground truth or gold standard image can be generated by a physician extracting the ROI

1)https://www.analyticsvidhya.com/blog/2019/07/computer-vision-implementing-mask-r-cnn-image-segmentation/

2) https://stackoverflow.com/questions/64076134/how-to-display-a-ground-truth-image-segmentation-mask-image-in-python

3) https://sergilehkyi.com/image-segmentation-with-python/

4) https://imgaug.readthedocs.io/en/latest/source/examples_segmentation_maps.html

Interesting observations

Excellent question

https://towardsdatascience.com/an-image-processing-tool-to-generate-ground-truth-data-from-satellite-images-using-deep-learning-f9fd21625f6c

https://www.quora.com/How-do-I-generate-ground-truth-images-for-image-segmentation-using-deep-learning

https://www.sciencedirect.com/topics/computer-science/ground-truth-image

GROUND TRUTH PROCEDURE (MANUAL)

1. With GIMP, drag and drop the original image (radar, etc.) to open it in a new

window.

2. Add a layer or "layer" (Layer) New Layer. . ), accept the dimensions.

3. Ensure that the new layer is selected and placed on top of the original image layer.

4. Get the corresponding region from Google Maps (printscreen).

4.a. Wherever possible, try to have the same scaling to reduce the number of pre-processing steps.

pre-processing steps.

5. Copy and paste into GIMP.

5.a. Click outside the layer area with the 'Rectangle Select' tool to make sure the image is inserted into the layer.

image is inserted in the layer [Source: Gimp software].

6. Add transparency to the ground truth layer with

(Layer ) Transparency ) Add Alpha Channel). If greyed out, this means that the layer already contains transparency

Go to the next step.

7. With the selection tools [Source: Gimp software].

7. a. random contour

7. b. by colour

7. c. by continuous region

7. d. Select the areas to be removed from the ground truth (e.g. parks, lakes, etc.) and press delete/

delete (keyboard key) to remove them as you go.

8. After deleting enough regions to have only the desired regions or segments, it should be

segments, it should be easy to see the original image/ground truth overlay

overlay through the transparency.

9. Use the tools [Source: Gimp software].

Add transparency to the ground truth layer with

(Layer ) Transparency ) Add Alpha Channel). If greyed out, this means that the layer already contains transparency

Go to the next step.

7. With the selection tools

7. a. random contour

7. b. by colour

7. c. by continuous region

7. d. Select the areas to be removed from the ground truth (e.g. parks, lakes, etc.) and press delete/

delete (keyboard key) to remove them as you go.

8. After deleting enough regions to have only the desired regions or segments, it should be

segments, it should be easy to see the original image/ground truth overlay

overlay through the transparency.

9. Use the tools (see Fig. III-3)

Figure-A III-3 the selection tools [Source: Gimp software].

to "align" and resize all segments on the "ground truth layer" as closely as possible to the original image.

as closely as possible to the original image. The correct layer must be selected in the list255

selected in the list to affect only that layer and not the original image. (the modifications

changes may take a lot of incremental adjustments)

10. When satisfied with the overlay, use (Color ) Threshold to get a 0/1 of the segments

segments (the threshold value may need to be adjusted as required).

With the help of Francis Charette Migneault