# Tag Info

12

Put simply, and without any mathematical symbols, prior means initial beliefs about an event in terms of probability distribution. You then set up an experiment and get some data, and then "update" your belief (and hence the probability distribution) according to the outcome of the experiment, (the posteriori probability distribution). Example: Assume we ...

8

Roughly speaking, over-fitting typically occurs when the ratio $\frac{\text{complexity of the model}}{\text{training set size}}$ is too high. Think of over-fitting as a situation where your model learn the training data by heart instead of learning the big pictures which prevent it from being able to generalized to the test data: this happens when the model ...

7

There isn't an easy way to do this. In fact a recent paper by Charles Parker explains some of the problems with using ROC curve measurements (there's a friendly overview at my blog - self promotion alert!!). His paper does make some recommendations on the best ways to compare different classifiers, so you will find something useful there.

6

The simplest Face Recognition 101 would be run a PCA on the data, followed by a K-nearest neighbor classification. but now there are more advanced techniques that you pick along, it might be using "better" classifiers like SVMs or Neural Networks or apply Bayesian learning, but then, as you move up the knowledge tree, you will realize you need your data ...

5

Shannon's entropy works as a hashing or fingerprinting function here; they are not injective which means we have to be careful not to interpret too much into relations of values. If $H(I_1) \neq H(I_2)$, then images $I_1$ and $I_2$ are certainly not the same. If $H(I_1) = H(I_2)$ (or even $H(I_1) \approx H(I_2)$), however, we know nothing. The images might ...

5

I suspect machine learning is the wrong approach. Instead, I suspect you will do better to define a metric and measure the metric, or define a hypothesis and use hypothesis testing. You are not trying to predict the future evolution of these values; that's something that ML might be suitable for, but that's not what you're trying to do, so ML doesn't seem ...

5

There is quite a bit of work on this important problem. Some of the most insightful work is by Helmut Alt and collaborators. He wrote a survey in 2009: Helmut Alt. "The computational geometry of comparing shapes." Efficient Algorithms. Springer Berlin Heidelberg, 2009. 235-248. (Springer link.)           Image from Helmut Alt ...

4

First a bit about the classifier: The knn classifier works by majority voting. It takes an input record, finds the k nearest labeled data points, looks at the class labels on each data point and assigns the current record the most common class label. For instance, if I use a 3-NN classifier and my three neighbors are of class [1 1 2], then I will select ...

4

There are multiple approaches which can be taken into account. Some of them are mentioned above. From what you want I think you should apply a combination of techniques to get good results. In my opinion you should start with blob matching (affine invariant) -> followed by keypoint feature matching inside the blob to verify the result. I think this ...

4

If you have one or several images of a particular "sink" or object, then you can possibly use SIFT or SURF or other feature extraction/matching algorithms to match the objects up. You can investigate the OpenCV project (wikipedia, project); it includes many of these (possibly patented) algorithms. You can get an idea of how they work and what they can do by ...

4

Sure. There's a straightforward way to convert an unnormalized distance metric into a normalized similarity measure. Basically, use $$S(x,y) = \frac{M - D(x,y)}{M},$$ where $D(x,y)$ is the distance between $x$ and $y$, $S$ is the normalized similarity measure between $x$ and $y$, and $M$ is the maximum value that $D(x,y)$ could be. In the case of ...

4

The state of the art for digit recognition does not use collective recognition, competence areas, ensembles, or any of the other ideas you propose in your question. Instead, the state of the art for digit recognition uses convolutional neural networks. Just a convolutional neural network: no need for multiple of them, no need for any kind of other fancy ...

4

What you are looking for is called "on-line recognition". I have written my Bachelors thesis about this: Thoma, Martin. "On-line Recognition of Handwritten Mathematical Symbols." arXiv preprint arXiv:1511.09030 (2015). And I've created the online service write-math.com as a part of it. As a short introduction, you might be interested in my presentations....

3

Number the pixels from $0$ to $n^2-1$. Then create $\lceil \log_2n^2\rceil$ many pictures. In the $i$-th picture all pixels are black, if the $i$-th bit in their number is $1$, white otherwise. Note: If you have $2^k$ colors available, you can choose the color for each pixel based on $k$ successive bits of the number. This will reduce the number of images ...

3

If it is important for you to get a high accuracy, then use a convolutional neural network (ConvNet). These ConvNets hold the state of the art for most visual recognition tasks. If your training set size is small, you should use a pretrained ConvNet as a feature extractor and then apply a support vector machine (SVM) on top of the extracted features. I can ...

3

There appears to be some work in that area. See for example this paper and the references therein. There are also demonstrations of implemented systems on youtube, see for example this video

3

I think the key is to abstract your problem by finding appropriate features. For example a feature could be the minimal distance between the two curves. Or how many times do they cross over. Or how many steps do they stay less than K units away from each other. Think about features that will capture convergence and divergence. Once you have the right ...

3

Since you asked for algorithms, I suggest you look at the Hough transform and other techniques for finding lines in an image. I suspect this will help you determine where the lines are, where the intersections are, and how much to rotate to recover the original image.

3

I think that you are interested in MFC - Multiple Foreground Cosegmentation. MFC article Awesome material: Articulated Motion and Deformable Objects : 5th International Conference, AMDO 2008, Port d'Andratx, Mallorca, Spain, July 9-11, 2008, Proceedings In advance I warn you, these are not all out of the box working solutions, but with small changes all ...

3

Neural nets don't "understand", they are trained. Despite the fancy term, a neural net is simple a regression model on steroid - often high-dimensional. It's a bunch of weights (vectors) connected in a graph, with an input side and some updating rule (for example, gradient descent). You'll need a lot of labeled data for the training set. The neural net is ...

3

The best way to get an orientation is to read a textbook on computer vision or image processing. There are lots of techniques known, and that's often the best way to get an introduction to a broad topic like this. Anyway, for some specific techniques, I would recommend that you learn about the following candidate approaches: Learn about morphological ...

3

Yes, your data is "time-series data", since it's a sequence of measurements of the same variable collected over time. Time-series data can be collected continuously or at discrete intervals. Your sample data can be expressed as a function of time - maybe it helps to think of the "function" as the process that produces the measured output, the input to the ...

3

It depends on what you're recognizing, but I'm familiar with two basic techniques: Apply the recognition algorithm to each frame of the video. Be fast enough that you can run on every frame of video. This is probably the most common approach. Recognize the object once, in one frame. Then use motion tracking to keep track of the location of the object in ...

2

One related approach would be to compute a histogram of the pixel intensities for each of the two images, then compare those two histograms. More typically, we'd compute a histogram of some measure of color (e.g., the hue), instead of pixel intensities. This is a crude measure that can sometimes nonetheless be helpful: e.g., for distinguishing a picture of ...

2

Ycbcr just has one component (y) for luminance and two for chrominance and it is less light variation sensitive than HSI, and for face detection Ycbcr comes to better result

2

I'm not an expert, but judging from the Wikipedia articles, YCbCr is used for image compression (JPEG and television [YPbPr]), while HSI is used in computer vision. Since you're doing computer vision, if we believe Wikipedia then you should use HSI. Of course, you could always experiment with both representations and all six dimensions. See what works best. ...

2

The trick here is that this works well provided that your rules can be expressed in predicate form. Is golf still fun if it's raining, or if you play poorly? If you need something more flexible, you might want to look at some statistical/Baysean tools. There, you'd say that golf had a high probability of being fun, not that it was always fun all the time ...

2

The appropriate technique is machine learning. Some keywords you could search for are "music speech discrimination", and you could look at this survey. (These pointers came from Vor's comment.)

2

Rifat, you have to ask your supervisor. If (s)he accepts, then go for it, otherwise go back to the drawing board. If this is, as you say, an area for which an OCR does not yet exist, then it sounds perfectly fine for an undergraduate project, but do not expect any reputable conference or journal to publish your results.

Only top voted, non community-wiki answers of a minimum length are eligible