I am working on the comparison of Histogram of oriented gradient (HoG) and Convolutional Neural Network (CNN) for the weed detection. I have two datasets of two different weeds.
CNN architecture is 3 layer network.
1) 1st dataset contains two classes and have 18 images.
The dataset is increased using data augmentation (rotation, adding noise, illumination changes)
Using the CNN I am getting a testing accuracy of 77% and for HoG with SVM 78%.
2) Second dataset contact leaves of two different plants.
each class contain 2500 images without data augmentation.
For this dataset, using CNN I am getting a test accuracy of 94% and for HoG with SVM 80%.
My question is Why I am getting higher accuracy for HoG using first dataset? CNN should be much better than HoG.
The only reason comes to my mind is the first data has only 18 images and less diverse as compare to the 2nd dataset. is it correct?
Yes, your intuition is right, having this small data set (just 18 images before data augmentation) can cause the worse performance. In general, for CNNs you usually need at least thousands of images. SVMs do not perform that bad because of the regularization (that you most probably use) and because of the probably much lower number of parameters the model has. There are ways how to regularize deep nets, e.g., with your first data set you might want to give dropout a try, but I would rather try to acquire a substantially larger data set.
Related
I have a dataset of around 20K images that are human labelled. Labels are as follows:
Label = 1 if the image is sharp and well lit, and
Label = 0 for those blurry/out of focus/grainy images.
The images are of documents such as Identity cards.
I want to build a Computer Vision model that can do the classification task.
I tried using VGG-16 for transfer learning for this task but it did not give good results (precision .65 and recall = .73). My sense is that VGG-16 is not suitable for this task. It is trained on ImageNet and has very different low level features. Interestingly the model is under-fitting.
We also tried EfficientNet 7. Though the model was able to decently perform on training and validation, test performance remains bad.
Can someone suggest more suitable model to try for this task?
I think your problem with VGG and other NN is the resizing of images:
VGG expects as input 224x224 size image. I assume your dataset has much larger resolution, and thus you significantly downscale the input images before feeding them to your network.
What happens to blur/noise when you downscale an image?
Blurry and noisy images become sharper and cleaner as you decrease the resolution. Therefore, in many of your training examples, the net sees a perfectly good image while you label them as "corrupt". This is not good for training.
An interesting experiment would be to see what types of degradations your net can classify correctly and what types it fails: You report 65% precision # 73% recall. Can you look at the classified images at that point and group them by degradation type?
That is, what is precision/recall for only blurry images? what is it for noisy images? What about grainy images?
What can you do?
Do not resize images at all! if the network needs fixed size input - then crop rather than resize.
Taking advantage of the "resizing" effect, you can approach the problem using a "discriminator". Train a network that "discriminate" between an image and its downscaled version. If the image is sharp and clean - this discriminator will find it difficult to succeed. However, for blurred/noisy images the task should be rather easy.
For this task, I think using opencv is sufficient to solve the issue. In fact comparing the variance of Lablacien of the image with a threshold (cv2.Laplacian(image, cv2.CV_64F).var()) will generate a decision if an image is bluered or not.
You ca find an explanation of the method and the code in the following tutorial : detection with opencv
I think that training a classifier that takes the output of one of one of your neural network models and the variance of Laplacien as features will improve the classification results.
I also recommend experementing with ResNet and DenseNet.
I would look at the change in color between pixels, then rank the photos on the median delta between pixels... a sharp change from RGB (0,0,0) to (255,255,255) on each of the adjoining pixels would be the max possible score, the more blur you have the lower the score.
I have done this in the past trying to estimate areas of fields with success.
I would like to ask somebody for advice. I created programme in C++ where I am using OpenCV library (v2.4.11), especially MLP classifier.
I have had accuracy on 2 000 testing screens about 92% but only when I set number of iterations on 1. With bigger numbers like 100, 1000 it is getting worse (on 100 it is 78%, on 1000 77%).
It is possible that problem is in data model and programming part is correct? Or it has to be my fault?
Thank you very much.
It is possible that problem is in data model and programming part is
correct?
Yes, the number of iterations, like the number of neurons and the number of layers, is one of the parameters that has great impact on the overall performance of the MLP classifier. The more iterations you apply to the MLP training the more the MLP NN adapts/fits to its training data. This leads to high performance on training data but can eventually lead to poor performance on test data. In this case you have over-train/over-fit your MLP NN.
There are however methods (e.g., grid-search) for the optimization of the parameters of a classifier.
I am trying to train a pedestrian detector using dlib and the the INRIA Person Dataset.
So far I used 27 images, the training is fast but the results are unsatisfying (on other images pedestrians are rarely recognized). Here is the result of my training using the train_object_detector program that comes with dlib (in /exmaples directory) :
Saving trained detector to object_detector.svm
Testing detector on training data...
Test detector (precision,recall,AP): 1 0.653061 0.653061
Parameters used:
threads: 4
C: 1
eps: 0.01
target-size: 6400
detection window width: 47
detection window height: 137
upsample this many times : 0
I am aware that other images need to be added to the training in order to have better results but before doing that I want to be sure of the meaning of every parameter printed in the result (precision, recall, AP, c, eps, ...) I am also wondering if you have any recommandations regarding the training : what images to choose ? how many images are needed ? Do I need to annotate every object in the image ? Do I need to ignore some regions in the image ? ...
One last question, is there any trained detector (svm file) that I can use to compare my results ?
Thank you for your answers
I am not familiar with dlib in particular, but let me tell you that you will not get good results with 27 images. In order to generalize well, your classifier needs to see many images with a variety of data. It won't do you any good to supply it with 10,000 images of the same person, wearing the same outfit. You want different people, clothing, settings, angles, and lighting. The INRIA dataset should cover most of those.
Your detection window dimensions and upsample settings will determine how large people must look in the image in order for your trained classifier to detect them reliably. Your settings will detect only people at 1 scale where they are around 137/47 pixels tall/wide. If you upsample even once, you'll be able to detect people at a smaller scale (upsampling makes the person look bigger than they are). I suggest you use a larger dataset and increase the upsampling number (by how much you upsample is another discussion - that appears to be built into the library). Things will take longer, but that is the nature of training classifiers - tweak parameters, retrain, compare the results.
For precision/recall I'll refer you to this wikipedia article. These are not parameters, but results of your classifier. You want both to be as close to 1 as possible.
Currently I am following the caffe imagenet example but apply it on my own training data set. My dataset is about 2000 classes and about 10 ~ 50 images each class. Actually I was classifying vehicle images and the images were cropped to the front, so the images within each class have the same size, the same view angle(almost).
I've tried the imagenet schema but looks like it didn't work well and after about 3000 iterations the accuracy was down to 0. So I am wondering is there a practical guide on how to tune the schema?
You can delete the last layer in imagenet, add your own last layer with a different name(to fit the number of classes), specify it with a higher learning rate, and specify a lower overall learning rate. There does exist an official example here: http://caffe.berkeleyvision.org/gathered/examples/finetune_flickr_style.html
However, if the accuracy was 0 you should check the model parameters first, perhaps it's an overflow
My objective is to train an SVM and get support vectors which i can plug into opencv's HOGdescriptor for object detection.
I have gathered 4000~ positives and 15000~ negatives and I train using the SVM provided by opencv. the results give me too many false positives.(up to 20 per image) I would clip out the false positives and add them into the pool of negatives to retrain. and I would end up with even more false positives at times! I have tried adjusting L2HysThreshold of my hogdescriptor upwards to 300 without significant improvement. is my pool of positives and negatives large enough?
the SVM training is also much faster than expected. I have tried with a feature vector size of 2916 and 12996, using grayscale images and color images on separate tries. SVM training has never taken longer than 20 minutes. I use auto_train. I am new to machine learning but from what i hear training with a dataset as large as mine should take at least a day no?
I believe cvSVM is not doing much learning and according to http://opencv-users.1802565.n2.nabble.com/training-a-HOG-descriptor-td6363437.html, it is not suited for this purpose. does anyone with experience with cvSVM have more input on this?
I am considering using SVMLight http://svmlight.joachims.org/ but it looks like there isn't a way to visualize the SVM hyperplane. What are my options?
I use opencv2.4.3 and have tried the following setsups for hogdescriptor
hog.winSize = cv::Size(100,100);
hog.cellSize = cv::Size(5,5);
hog.blockSize = cv::Size(10,10);
hog.blockStride = cv::Size(5,5); //12996 feature vector
hog.winSize = cv::Size(100,100);
hog.cellSize = cv::Size(10,10);
hog.blockSize = cv::Size(20,20);
hog.blockStride = cv::Size(10,10); //2916 feature vector
Your first descriptor dimension is way too large to be any useful. To form any reliable SVM hyperplane, you need at least the same number of positive and negative samples as your descriptor dimensions. This is because ideally you need separating information in every dimension of the hyperplane.
The number of positive and negative samples should be more or less the same unless you provide your SVM trainer with a bias parameter (may not be available in cvSVM).
There is no guarantee that HOG is a good descriptor for the type of problem you are trying to solve. Can you visually confirm that the object you are trying to detect has a distinct shape with similar orientation in all samples? A single type of flower for example may have a unique shape, however many types of flowers together don't have the same unique shape. A bamboo has a unique shape but may not be distinguishable from other objects easily, or may not have the same orientation in all sample images.
cvSVM is normally not the tool used to train SVMs for OpenCV HOG. Use the binary form of SVMLight (not free for commercial purposes) or libSVM (ok for commercial purposes). Calculate HOGs for all samples using your C++/OpenCV code and write it to a text file in the correct input format for SVMLight/libSVM. Use either of the programs to train a model using linear kernel with the optimal C. Find the optimal C by searching for the best accuracy while changing C in a loop. Calculate the detector vector (a N+1 dimensional vector where N is the dimension of your descriptor) by finding all the support vectors, multiplying alpha values by each corresponding support vector, and then for each dimension adding all the resulting alpha * values to find an ND vector. As the last element add -b where b is the hyperplane bias (you can find it in the model file coming out of SVMLight/libSVM training). Feed this N+1 dimensional detector to HOGDescriptor::setSVMDetector() and use HOGDescriptor::detect() or HOGDescriptor::detectMultiScale() for detection.
I have had successful results using SVMLight to learn SVM models when training from OpenCV, but haven't used cvSVM, so can't compare.
The hogDraw function from http://vision.ucsd.edu/~pdollar/toolbox/doc/index.html will visualise your descriptor.