General Desciption

This repo contains four projects for COMP5421 - Computer Vision@HKUST. And each of these four projects involves different knowledge background about computer vision.

You may need UST CSE account to access some webpages.

Image scissor

General Idea

The general idea of this project is that after loading the image, we finished the cost computation by implementing the cost function provided in the course webpage*. Every time when the user clicks a seed, we will update the path tree according to the Dijkstra’s algorithm.

Artifacts

Original Images and Masks

We provide three images, one background image and two foreground images.

Foreground Image One and its mask:

Foreground Image Two and its mask:

Background image:

Results

The result is cool!!!

Bonus

We implemented two whistles for this project.

SnapSeed

To snap the first seed near the edge, we find the nearest neighbor of the seed on the edge after the user clicks the first seed. We used CV::Canny to get the edge and then use L2 distance to meansure the space similarity. Update the position of the first seed each time when finding out a optimal point.

Blurring

We include two gaussiFor blurring effect, and we use opencv Guassian filter. We adjust the Guassian filter kernel size and standard deviation to achieve different level of blurring effect. After opening the image, we could add this blurring effect using the interface of our IScissor. Tool -> Gaussian 3*3 Tool -> Gaussian 5*5 After click on these two buttons, the image will be blurred. We provide a sample with our school HKUST. You can see that different level of blurring effect has been achieved. Comparing two filters, we could see that larger kernel size induces stronger blurring effect. See pics.an blurring filters in the projects such that the user could apply anyone of them to compare the difference before and after blurring. Also we notice that if the blurring effect gets stronger, the cost gets lower.

Orignal Pic and its cost graph:

Blurring with 3*3 gaussian filters

Blurring with 5*5 gaussian filters

Face detection

The course project of COMP5421 Computer Vision

General Idea

The task given by this assignment is to implement a multi-scale sliding window face detector based on concepts presented in Dalal-Triggs 2005 and Viola-Jones 2001. The algorithm will be evaluated using a common benchmark for face detection (Caltech).

Methodology

Use HOG descriptor to genearte positive face images with cell size 3
Horizontally flip face images, contrast original face images with darker faces images( img*0.8 ), use HOG descriptor to generate postive images from these images and add these images into the postive face samples
Use HOG descriptor to generate negative cropped images (50000 images) from the database
Train the linear SVM Classifier
For each test image, for each position at each scale in the image, create a window and run the classifier to determine whether or not there is a face at that location
Step 5 will result in many overlapping bounding boxes for the the same face, which must then be combined or suppressed into one final bounding box (non maximum suppression).

Experiments and Results

We conducted several experiments to indentify the best approach to detect the face.

Cell Size 6, flip and contrast the faces, without hard negative mining

++The average accuracy is 0.840, and it is very quick.++

Cell Size 3, flip and contrast the faces, without hard negative mining

++The average accuracy is 0.916, and it takes about 20 minutes.++

Cell Size 3, without hard negative mining

++The average accuracy is 0.823, and it takes about 30-40 minutes.++

Cell Size 3, flip and contrast the faces, with hard negative mining

++The average accuracy is 0.901, and it takes about 90 minutes.++

Discussion

Notice that when the cell size gets smaller, the HOG descriptor can persent more details about the gradient and edge direction at each pixel, which leads to higher accuracy.
With more negative samples, the accuracy can be improved a bit, but at the same time, the detection speed becomes slow.
When we horizontally flip face images and make images darker (i.e. add more positive training data), we can largley improve the accuracy since we found that some faces in very dark background could not be detected and some side faces can not be recognized.

Bonus

Hard Negative Mining.

Train on the original datasets, collect images which are falsely detected as faces and add them into the negative samples. Retrain again.

Find and utilize alternative positive training data.

Horizontally Flip face images
Generate contrast images (image * 0.8)

Implement an interesting feature

Implement Local Binary Pattern (LBP). Instead of using HOG descriptor, we use LBP to extract features from images.

Single view metrology

Methodology

calculate the vanishing points by Bob collins’ algorithm.
Calculate the Projective matrix, where the scales are computed with reference points and vanishing points
Use Homograph Matrix to get the texture map (for simplicity, we directly “hardcode” points this time)
Mark Interesing points and get 3D coordinates
Generate VMRL