Object Recognition - Science topic

Dear Rahul Jain ,

Computer vision (CV) is the scientific field which defines how machines interpret the meaning of images and videos. Computer vision algorithms analyze certain criteria in images and videos, and then apply interpretations to predictive or decision making tasks.

Today, deep learning techniques are most commonly used for computer vision. This article explores different ways you can use deep learning for computer vision. In particular, you will learn about the advantages of using convolutional neural networks (CNNs), which provide a multi-layered architecture that allows neural networks to focus on the most relevant features in the image.

Regards,

Shafagat

Dear S M Mohiuddin Khan Shiam ,

Identification. Image classification using deep learning categorizes images or image regions to distinguish between similarly looking objects including those with subtle imperfections. Image classification can, for example, determine if the lips of glass bottles are safe or not.

https://imaging.matrox.com/en/technology/deep-learning#:~:text=Real%2Dworld%20examples-,Identification,bottles%20are%20safe%20or%20not.

Regards,

Shafagat

In the field of autonomous driving there is a well-known dataset KITTI ( ) which contains point clouds, if this domain works for your research.

Robotics are used in many materials handling applications because they are more efficient in certain tasks and take people out of potentially unsafe situations. To work effectively, robotics rely on sensors to interact and perceive their environment. This white paper from ifm illustrates the importance of 3D image sensors in robotics and the systems used to operate and manage these.

The interaction of the human eye with the visual centre in the brain creates a three-dimensional image of the environment. In robotics, such a three-dimensional image is important to enable robots to act independently and without causing any danger outside of safety barriers...

Dear Gustavo Carvalho ,

Look the link, maybe useful.

Regards,

Shafagat

Dear Efe Francess Ologbonori ,

Pattern/object recognition is concerned with the processes involved in the identification of images and objects. This essentially involves taking information that enters the visual system and comparing this with information stored in memory, and finding a match. There are three approaches within pattern recognition; template and prototype theories, feature comparison theories and structural theories. The focus of this essay is feature comparison theories, their advantages and disadvantages and their overall success in pattern/object recognition.

Regards,

Shafagat

Ben Harper, thanks very much for your recommendation.

Since, your problem is to identify darkest object. you can get the histogram of the image. Then apply adapive thresholding to find the threshold automatically (threshold will be different for different images but that will be automated)

Moreover, it seems that the objects shape is hexagonal that you want to detect. I think you can train a HOG+SVM classifier with bounding boxes.

It is by opposing intent that we prove its existence. (That comes obliquely here from security operations which have a duty through unambiguous warnings or non-lethal opposition to prove hostile intent in challengers before the use of force).

So, if you dangle the object above the ground but in the limits of reach, then your subjects prove their exploratory intent by both stretching up and sniffing. All sniffers are then more obviously explorers, and hiders and recliners are ruled out.

More generally, if the subjects had to overcome discomfort or fear in order to venture ever closer to an object, then it's much more likely that their intent is to explore. The greater or more numerous the obstacles, the greater the intent.

Topic is bit dusty, but I wonder if you got any answers?

Depends upon many conditions, e.g. The algorithm you want to use, any integrated cameras or sensors, and your budget.

You could try with:

SAGA SIG :https://sourceforge.net/projects/saga-gis/

or Python:

in computer vision Deep Learning is recently being used very widely by many organization including Microsoft, Google. Deep Learning uses deep net to simulated many layers of non linear processing . feature extraction and transformation and finally cascade them into a more accurate approximation of object identification. also you can check the following ideas:

Here is a similar one, which you will never be able to not see it again once you've found it:

https://www.pinterest.at/pin/567664728017501723/

You may have a look at

https://machinelearningmastery.com/singular-value-decomposition-for-machine-learning/

Hi Aravinda Kasukurthi,

The Faster R-CNN has the same number of hidden layers as the Fast R-CNN, the RPN has no hidden layers and is only used as a feature extractor. The Fast R-CNN has three fully connected layers.

Dear Rohit Yadav,

In image processing, an image is "processed", that is, transformations are applied to an input image and an output image is returned. ... Computer vision uses image processing algorithms to solve some of its tasks. The main difference betweenthese two approaches are the goals (not the methods used).

Regards,

Shafagat

You need to transform the data because you are applying a parametric test that requires the fulfillment of two main assumptions: normality and homocedasticity.

The angular transformation (or arcsine transformation) is suitable for proportional data, like yours, because they are restricted in the range 0-1. In this case, one of the recommended transformations is the Arcsine transformation.

I think your confusion is regarding the value to which to compare your proportions after the transformation.

Angular transformation: W = [ASIN (SQRT (p))]

Being: W = transformed value; p = the original proportion.

W = [ASIN (SQRT (0.5))] = 0.785. That value W = 0.785 will be the one used in your

one-sample t-test.

I hope that is your doubt.

Regards !

Refer the following websites related to OpenCV useful for learning ROI:

Sowmith Reddy

Follow

https://www.mathworks.com/help/deeplearning/pattern-recognition-and-classification.html

If your data have a very high slice spacing in 3rd dimension and you are getting 2d slices results, you need to integrate neighbouring slices, with a fixed spacing of 1 mm using a linear interpolation to deal with this problem.

I've used Ctrax successful, but heard that ToxTrac is quicker at extracting data from video files

Dear Alaa,

you may find all relevant information on the ILSVRC page [1]. If you already have the original image set, you are fine. The devkit contains the ground truth annotations for the validation images, which are named like "ILSVRC2012_val_00000001.JPEG" etc.

[1] http://www.image-net.org/challenges/LSVRC/2012/nonpub-downloads

Dear Md Junayed Hasan,

Thank you for your time and answer..

I will go through it...

For Object recognition in general read this review paper

https://www.researchgate.net/publication/326199343_Recent_Advances_in_Color_Object_Recognition_A_Review?_sg=SSEynmVlLxbqluSz17N3lP2P4B3hz1VRczXMHJstCfCfuMcC1jl5v-dOCtcADHwMOVrjkeJJCc4UM20aUpMgKMgtN-DYq1qeUgUjV4In.JKKjeZL2FPgVfi19UghWLnlL8Ic8nTz7zacWTJYcMHLB0zR80agrpFLJHk6MPneeWq79Sb7OAQZoR7Yk8E3sIA

This review paper enlist recent techniques.

1. Haar Cascade Classifier

2. Template matching

3. Deep CNN

4. Contour finding after background subtraction in case of video

Fabio, is " Convex Polygon Fitting in Robot-Based Neurorehabilitation" the work you referenced which has all the examples and the code?? I have sent you a request for the article. Could you please send it across? Thanks

Xie, J., Xu, L., & Chen, E. (2012). Image denoising and inpainting with deep neural networks. In Advances in neural information processing systems (pp. 341-349).

Dong, W., Wang, P., Yin, W., Shi, G., Wu, F., & Lu, X. (2018). Denoising Prior Driven Deep Neural Network for Image Restoration. arXiv preprint arXiv:1801.06756.

Lucas, A., Iliadis, M., Molina, R., & Katsaggelos, A. K. (2018). Using Deep Neural Networks for Inverse Problems in Imaging. IEEE SIgnal ProcESSIng MagazInE, 1053(5888/18).

Traditionally, Local Binary Patterns (LBP) has been successfully used as features for facial expression recognition. It is very important to extract LBP from regions in a grid over the face image and concatenate them in a single vector.

Dear Nils Schaetti,

Look the link, maybe useful.

Regards, Shafagat

You should probably try to find a problem you can solve combining/pipelining both types of algorithms, i.e. pretend you had an ultimate goal when you worked on them both. I don't know exactly which algorithms you are talking about, but lets say you are detecting people approaching a government building and you want to identify a felon among them or something like this. Depends on your specific work, really..

Hi Ryan

Take a look at Gauthier T, Tarr MJ Becoming a Greeble Expert etc..., Vision Research 1997, 37, pp 1673-1681.

It may help.

Regards

Michael

tnx dear @Dibya jyoti Bora

 This is an important question with a number of possible answer.   In addition to the very helpful answers already given, there is a bit more to add.

Thermal imaging and thermographic measurements are the focus of the following paper:

See, for example, Fig. 1 and the ensuing discussion for an overview of the approach needed to do edge detection on thermal images.

More to the point, see the use of histograms as a means of extracting information from thermal images in

Hi Chamari,

Basically, there is no best machine learning algorithm, however, I recommend trying linear algorithms first, but if you found them not sufficient, you could give kernel methods and neural networks a try. 

HTH.

Samer

Just attach the image files with the "paper clip" below the message...

Dear Kate, if you want to take into account the initial exploration of the objects as a baseline, you can utilize the recognition index (RI), as described in:

If the RI confirms the result obtained with the discrimination index (DI), then you might really have found a neophobic phenotype. You said you have counterbalaced for both the objects and the sides, but the best way to check for possible biases is to compare the RI and the DI. If you still find a significant aversion for the new object, then it would be interesting to investigate this phenotype with a different test. As Katherine suggested, you could use the novelty-suppressed feeding test, by which you can evaluate the hyponeophagia phenomenon (https://www.ncbi.nlm.nih.gov/pmc/articles/PMC3197427/).

Go through this paper you can detect most of the object shapes using this approach

Hello Aditya,

What are the Objects you want to recognise ? Are they same object? and more in Number?

Example : 1st Case : You want to Detect Multiple objects of same class .

You want to detect Multiple Oranges in a Picture.

2nd Case : Multi Class Object detection

You want to detect Oranges, Apples, Banana's . 

If you are looking for the first case, it should not be that hard to implement, Apply the object detector multiple times on the images, instead returning when you find the first object.

If its the second case, try building feature detectors for every class and apply every detector onto the image to find the different objects.

In general there are many approaches to recognise the object. Simple features like Color, Shape, or more complex features using HaarWavelets.

Start with one object and move to multiple object recognition

It may  5-10

This is an interesting problem.

The paper "Curvelet Based Feature Extraction" might give some insight.

You can extract texture features and then run some supervised classification experiments to determine roughness. You can do this with CVIPtools and run experiments with CVIP-FEPC. CVIPtools software is available here: http://cviptools.ece.siue.edu/

Try :

Hi, 

Can you please elaborate on your context a little? What is the scene you're looking at? One person or multiple persons, static or dynamic background? Identify a particular person or detect people in general?

Simplest case of a static background and detecting a moving person in a video- 

The first possibility is a static frame looking at a particular scene with change only being a person moving. Difference of frames (absolute and thresholded) shall give you the location of a particular blob (where the moving person. You can then use OpenCV contours detecting for a maximum value (use more than one for multiple moving objects) which shall localise the moving person for you. Draw a bounding box and extract the object. For identifying, you can apply your feature based techniques to do that. For example, face-detection and so on. 

More clarity on context can help me come up with a better approach maybe. I hope I have understood the question correctly. 

I think the best way is to train a classifier to detect the stomata.

Crop from several images the example of stomata and areas with out stomata.

Use some image descriptor like Histogram of Gradients (Hog) to describe each one the patches.

Train a classifier, SVM for example, to distinguish between stomata and non stomata.

Use a sliding window and the classifier to detect the areas with the stomata.

For now if you don't have enough images  you can use the following code as a staring point. It uses one of the stomata in the image as an example in order to do correlation base detection.

Pay attention that the right low stomata is not detected because it is not a full stomata. You can change the patch to be just the upper side and also detect the right low stomata but then you would have to change some basic staff in the final part of the code.

Hope it is a good base for you good luck.

Dear Arefeh...

The proper threshold should be the valley point in the histogram... You can calulate it as the point whose value is less than the average of nearest n histogram values from the left side, and also less than the average of of n histogram values lay at the right side... You can set n=5 or 6.... Regards

your work is more aligned towards image recolonization rather then detection. you can use many features of the images containing object there is not objection on the use of feature(mean , entropy, standard deviation, correlation etc are some of the commonly   used features   ) more the number of features more will be the accuracy you can , do the add on like use of wavelets curvelets along with the features to make it more robust.

This is a very interesting question.

You might want to consider the location of the center of mass of a hole as a means of measuring the major and minor axes of a hole.    Another approach is suggested in the following: find the major and minor axes of the best fit elipse for a hole:

article: https://www.researchgate.net/publication/27248858_Nanometer_hypervelocity_dust_impacts_in_low_Earth_orbit

This problem is a little bit harder than invariant-based shape matching since you will have occlusions from the perspective projection (if I understand the problem correctly, you are trying to match views of a 3d object to a template). If the observed view does not overlap with the view used to construct the template then the problem is impossible. However, if the view has a high amount of overlap then what you need is a perspective invariant, which would fall within the realm of affine invariant.

You could use our method (attached) if the views are provided as point-sets. You could sample edges or use some other feature point selector to produce the point-sets. It is a little bit sensitive to outliers.

If you try our method and find it has trouble with outliers produce by your feature point selection routine then try Shape Contexts along with VFC. I believe that Ji Zhao has code somewhere. There are literally dozens of point-matching algorithms you can try once you get the data into this format.

If you don't want to do feature-points but can reduce the views to silhouettes then you could try the dijkstra-based method on the cross product graph of the two outlines (ok you have to reduce the silhouettes to curves this time). You would want to use an affine invariant signature in this setup.

Good luck, hope this helps 

Hi there, 

You can try this paper

Best.

You could calculate the angle between each successive line-segment pair. Depending on how the line-segments run (clockwise or anti-clockwise), angles greater than (or less than)180 degrees are concavities. However, you then get all concavities...

In this case, following heuristic might also work: evaluate all points that are between two successive convex-hull points. The point that has the longest perpendicular distance to the line connecting the two convex-hull points, probably is the point you are looking for. 

salam

have a good day. can you take a look to this code program https://github.com/rishisij/workhouse

Hi,

I am trying to shed some light here:

1. what you get from SIFT is an 128-dim. descriptor for each detected keypoint position, orientation and scale.

2. In the matching step, only the 128-dim descriptor is involved. Please note that the descriptor is already normalized to orientation and scale. The detected position gives you the center point of the image patch that you describe with the 128 dim. descriptor. The scale tells you how large this patch is and the orientation tells how much the patch has to be rotated before descriptor computation.

3. This depends on the application... if you just want to match points between images you compare the 128.dim descriptors and just look at the keypoint position, when corresponding descriptors have been detected.

4. see 2. and 3.

5. Yes, if you rotate an image by 90 degree you should get the same number of keypoints.

6. This is optional... if you implement the standard matching described in Lowe's paper then you have a one-to-many matching.

 https://d41911866d1ccb50a887d0de529d97f1808c59dc.googledrive.com/host/0B4bm3usSX8o2THRVV0JRYTNTcHc/FacultyStaff/K0140/Paper/Biometric_Verification_Using_Thermal_Images_of_Palm-Dorsa_Vein_Patterns_-_J.pdf

oups! this is far away from competencies, sorry.

Dear Bo Li,

At first you need to identify the interest finger regions. After Extracting the minutiae save the data on correspondent matrix. Then just match it by correlation with other finger print.

Check the following links.

Regards,

Ehtesham

Hi,

Graham pointed out the fact you must analyze what kind of evaluation is required.

In some MTT methods, if track labels or theoretical positions of pedestrians are known, you can use the OSPA-T distance. It can both combine the tracking accuracy and the cardinality estimation. Full details of this metric can be found in Ristic's paper :

Hi Patil,

If you have a CAD model of your target you may extract the edges and find those edge in image.

Not so common, you may also use texture if your CAD model contains texture information.

I used a commercial software to find the pose of a CAD model in images, but if you pretend to get more data and fully control the process is may be better to implement the methods by yourself

Hi Mat,

I use Halcon, it's a very powerful tool mainly for industrial purposes. For investigation it may be used in processes that aren't your focus because some functions are like a black box (its their knowledge and marketing advantage).

There is a group on LinkedIn about Halcon with experienced users that gives faster answers than Halcon support.

And, YES you develop all your code and export it to other languages or use hdevengine in which any modification require just to replace an Halcon file and not to compile again all app

Hi Laura

You may treat this problem as an ordinal variable. If in your example, the student marks a and b, he gets the maximum score (eg. 3) . If he marks c he obtain the minimum score (eg. 0), if he marks only a or b he obtains a score in the midle (eg. 2). If he marks b and c, or a and c, he gets a regular score (eg. 1). Then you may use an IRT method for an ordinal response or for a mix of binary and ordinal responses. Also it may be that to mark only a was better than to mark only b, then you may give different ordinal value to this selections. For example you may put a value +2 for a, +1 for b, -2 for c and your ordinal variable may be the sum of this values. You should also take account of the lack of response.

My preference is to perform Novel Object Recognition in an open area, rather than Y-maze, separate compartments etc. This will mean the objects are continually visible to the animal and hence when they tend to the object it is an active choice of one over the other, rather than simply exploring the object in closest proximity. The other risk with more compartmentalised arenas, or more distance between the objects, is that a spatial aspect to the task is included, rather than purely visual recognition memory.

Dear Giovanni

Do you have any related practical example or project, that has used sparse representation?

Thank you Nada. I appreciate your help.

Dear Zubair Please find the following article which may help you.

Maybe the attached publication by

Ma et al. (2014). Changing concepts of working memory. nature neuroscience, 17(3): 347-356

could give you some orientation?

Regards,

Klaus Blischke

Man it has been a while, I had completely forgotten about the paper that you linked to in your question.

Okay, here is my understanding of what the Kinect device is doing based upon the patent literature, and I'd love to hear from someone else if they have info. that contradicts this.

First, there is a huge difference between what people do during the R&D phase and what they do during the production phase. The reason is that during R&D you want to know as much as possible while during production you want to meet your clients demands as well as possible.  For most of the "Kinect" clients, all they want to know is where is the skeleton, and they want to know this at real time speeds.

I say this because from the patent literature it sounds as if this point-to-bone assignment is only transient inside the Kinect device.  Rather, the Kinect device focuses on what this point-to-bone information tells it about where the joints should be, and only reports that information.

Similarly, from what I have seen with respect to the Kinect SDK, it appears that the only information that is streamed from the Kinect device is {color, depth, skeletal pose}, not {color, depth, pixel assignment, skeletal pose}.

If you think about the size of the pixel assignment information there is a good reason for not reporting it.  That is, the skeletal pose information is basically non-existent with respect to the size of the color and depth information.  Since the pixel assignment information would require at least as much time/memory as the color and depth information this means including the pixel assignment information in the Kinect camera data stream would increase the amount of time required for Kinect to report a single frame by rough 50%.  Since Kinect is already operating at the abysmal frame rate ~30 fps, the final frame rate with pixel assignment would put them somewhere around 15 fps.  I have tracking libraries that I have written that can do as much as Kinect can with any depth data stream that will get you the pixel assignment in about 30 fps without optimization through parallelization. If Kinect was forced to operate at 15 fps no one would buy the device.  So, I don't think that you are going to get the per-pixel information from Kinect.

All that said, unless you want to drop Kinect and write your own tracking algorithm, I would just take the skeletal pose and depth data returned by Kinect and use Linear blend skinning to assign pixel data to skeletal pose.  If you don't want to mess with skinning and you don't want to have to write your own skeletal tracker then you could just license my companies libraries for cheap, but you would still have to set up all of real time image buffers and concurrency infrastructure yourself in order to get real time processing.  So, I would just use LBS.

The open field test (OFT) is not quite a good test to evaluate anxiety, I would be careful with this kind of interpretation. OFT is more useful for locomotor measurements than it is to anxiety. The novel object recognition is better fitted for memory evaluation, so the interpretation about anxiety in this test should be also done with good care, maybe the novelty suppressed feeding would be better for anxiety measurements. Elevated plus-maze is a considerably good test and widely used for anxiety behavior, so if you do not see any difference at this test it indicates better that you drug or treatment/manipulation does not have a strong effect on anxiety behavior control. 

Therefore, you can consider run another behavior test which is better to evaluate anxiety-like behavior such as novelty suppressed feeding or the radial maze to have a better profile of your animals behavior. As your results do not show an effect with the EPM it might be better to run more tests to confirm the effects you saw with OFT and NOR. 

 For the question 1, I think what you should keep in mind is that, classical Haar filters are sensitive to position changes. So, for the evaluated windows, those Haar filters selected in training stages should be placed in the same positions with that in training images. However, because the sliding windows tend to have different scales from training images, so you need to 'align' the sliding window with training scale via some techniques before you place those Haar filters.

and how about those formulas?

what is the formulas mean? what is is the correlation of each parts in notation with all of Dog that we count before?

example: Detect pedestrians.

You need to do two main blocks:

 1. Training

 2. Detection