I suggest to use LabelMe or GEDI

Both are annotation tools developed for creating datasets from document images.

Both are used by the scientific community and both have been presented during conferences, in papers and so on.

I have used GEDI in the past but I know that LabelMe is more famous.

Choose according your needs.

http://labelme.csail.mit.edu/Release3.0/

http://lampsrv02.umiacs.umd.edu/projdb/project.php?id=53

Depending on the concrete task there may be different tools helpful. For example if you have to segment bone structures in medical images there are semi-automatic annotation tools, which may provide a first automatic segmentation which than has to be corrected manually.

If you have to outline arbitrary shapes on 2D images you can use one of the many image editing tools. It's also not to hard to program a tool yourself which for specific tasks even may be the best solution. In a simple image editing software you could use the Lasso-tool to draw the outline and the simply export a binary mask where white or black areas mark your ground truth. Here are some tools, but I'm sure there are others.

• GIMP
• ImageJ
• Matlab (see the function getline or getpts)
• Paint.Net

Thank you Mr Gordon Böer

I suggest to use LabelMe or GEDI

Both are annotation tools developed for creating datasets from document images.

Both are used by the scientific community and both have been presented during conferences, in papers and so on.

I have used GEDI in the past but I know that LabelMe is more famous.

Choose according your needs.

http://labelme.csail.mit.edu/Release3.0/

http://lampsrv02.umiacs.umd.edu/projdb/project.php?id=53

thank you for adding your answers

You can use Photoshop. I used GIMP as well. But according to my experience(i devoloped two databases for ground truth data), photoshop is easy and less time consuming.

Try Aletheia:

http://www.primaresearch.org/tools

There's also a evaluation tool to measure accuracy

http://www.primaresearch.org/tools

Don't know which kind of images you have to segment.

For medical 3D images, I use ITK-SNAP. 

http://www.itksnap.org/pmwiki/pmwiki.php

for skin dataset. I manually used Adobe Photoshop. Its tools time to generate the ground truth. :)

Thank you for your answers ...

There are free ground-truth available at different database that helps in evaluating the efficiency of segmentation. Also ground truth can be developed by inserting manual landmarks on images. This can be done in matlab using the ginput function. 

There are standard database that you can use as part of your image database which comes with ground truth images. Say for example Berkeley image database (http://www.eecs.berkeley.edu/Research/Projects/CS/vision/grouping/segbench/), etc.

This is used by most prominent research scholars in their research.

For generating ground truth image of any image, you can use a tool named ImageJ. This is free and open source tool. Its also easy to learn. I have used it in past for my research related to segmentation. This also allows other image processing options like applying various filters, zooming, rotations, etc.

@ Rahul Patel . Thank you for your answer.

I m working on co-segmentation problem which means segmenting simultaneously a set of related images. so i need à dataset which share a common objects with a large variation of scale across images. 

Ok. So for your co-segmentation, you can use image database like Weizmann horses database, MSRC database, Oxford database, etc. Below are the links:

www.msri.org/people/members/eranb/

www.research.microsoft.com/en-us/projects/objectclassrecognition/

http://www.robots.ox.ac.uk/~vgg/data/flowers/17/

Thank you Mr Rahul Patel. 

Hi Dear

1- I hope the following link be useful

https://en.wikipedia.org/wiki/Ground_truth

2-I found the following answer in the www , I think may be useful

(((

Answer by David Young on 3 Jan 2012

 Accepted answer

"Ground truth" means a set of measurements that is known to be much more accurate than measurements from the system you are testing.

For example, suppose you are testing a stereo vision system to see how well it can estimate 3D positions. The "ground truth" might be the positions given by a laser rangefinder which is known to be much more accurate than the camera system.

Sometimes synthetic data are generated from a model, to test a system whose goal is to estimate parameters of the model. In such cases the "ground truth" is the known parameters of the model. Again stereo vision provides a good example: computer graphics can generate synthetic images from a 3D model, and then a stereo vision system can try to reconstruct the model. The original model is the "ground truth".

In the case of edge detection, it's much less clear what "ground truth" means, and in fact I don't think it's well defined, as edge detection generally involves making some (more or less) arbitrary choices. Nonetheless, I can think of two possible candidates: one is the output of an edge-detector which is generally accepted to be high-quality; the other is the edges drawn by an expert human looking at the image. Neither of these is really ground truth, however, and the term "gold standard" would be more applicable.

This comes down to a question to you: what is your ultimate goal? That is, why do you want "ground truth"?)))

Thank you Mr Faisel Mohammed

Hello everyone I have been trying to find ground truth images of popular lenna, cameraman, paper, mandrill and lake for comparing my segmentation result. I have been trying for last couple of weeks, but could not find them. There are many images with ground truth but not the ones I need. Can you help me find those? Thanks

@Stefano Trebeschi will this ITK-snap be able to segment 2d images? I want to segment 2d brain tumor MRI Images 

I recently started a project I'm calling LabelD (https://sweppner.github.io/labeld/) that enables quick and easy image annotation sourced either locally or from Imgur via keyword search. Check it out! Still very alpha, but the annotation is up and running!

Try MatLab image processing ToobBox

I'd like to thanks all for this interesting  conversation. I respect all of answers, but I still have confusion about the "ground truth". As understand, the ground truth used to evaluate the performance of our proposed work. Is it mean the ground truth present the optimal solution (much more accurate). Therefore, why we look to proposed methods to solve the problem, at the same time, we have ground truth?

Hi Mr Shehab,

For example, In the field of image segmentation a ground truth is a segementation done by an expert, and  it is used to evaluate the accuracy  of  a proposed segmentation algorithm.

Hi Ms. Rachida,

Thank you for your replying. In My research, I generated a synthetic data for signal tensor. I saw in many articles the evaluate the performance via ground truth, but they didn't mention anything about the procedures (how they get it? or how the calculate it?).

Antonio Parziale's suggestion, LabelMe or GEDI, can be a good try. But I personally think the accuracy of such a tool depends on the complexity of specific segmentation cases.

I'd recommend to try Labelbox (www.labelbox.io).

Labelbox is a cloud based tool to easily label data.

Labelbox makes it really easy to do basic image classification or segmentation tasks. Simply upload a CSV file pointing to the location of your data and choose an image classification or segmentation labeling template to get started.

In addition, It supports any kind of data such as text, images, maps, videos and even point clouds or medical DICOM. One can create a custom labeling interface with HTML and javascript and can be as simple as 30 lines of code. You can see their template code here: https://github.com/Labelbox/Labelbox

Additional feature includes project management, team management and performance metrics.

Dear Rachida , 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.

All segmentation method is "approximate". So, the answer of your question is impossible! You only can make the ground truth of an image by doing manually.

This depends on the application. In general, it is produced by an human expert who is considered the best possible result.

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

Dear Rachida Es-Salhi ,

Great discussion point. Following the updates!

if you have satellite image, you should performance a filed work using GPS, or you can use visual Interpretation,