Taks at EESTech Challenge 2017
In this challenge, we was implementing a sliding window face detector. The sliding window model is conceptually simple: independently classify all image patches as being face or non-face. Face detection is one of the most noticeable successes of computer vision. For example, modern cameras and photo organization tools have prominent face detection capabilities.
Write your code to train a model to classify images to face vs. non-face:
- write your own code to extract this feature 2 from an image; given an image patch, the value of a feature = P (pixel values in white area(s)) − P pixel values in black area(s). The white area and black area are equal in size. You may want to extract the features at multiple scales (why?). Using an image pyramid 3 for that purpose, where the five features are extracted from each of the cells of the pyramid. For example, an 1-layer pyramid leads to a 5-dimensional feature, a 2-layer pyramid leads to a 25-dimensional feature (1+ 4) ∗ 5 = 25, and a 3-layer leads to an 105-dimensional feature (1 + 4 + 16) ∗ 5 = 105 ...
- select a good trade-off for your feature dimensions. Explain why this is good. You can do it either empirically or by techniques such as Cross-Validation.
- train a classifier of your choice for a binary classification of face vs. non-face. you are allowed to use any existing library for the classifier. SVMs, Random Forest, Adaboost, Logistic Regression are among the popular choices.
- predict the labels for the test data provided for this task; the predicted labels should be written into a ’.txt’ file, each line contains a filename and its label, separated by a space. The same format as the training data.
- (Optional) use Hard Negative Mining to further improve your classifier. A hard negative is when you take that falsely detected image, and explicitly create a negative example out of that image, and add that negative to your training set. When you retrain your classifier, it should perform better with this extra knowledge, and not make as many false positives.
The second task was to extend your face classifier to a face detector, which consists of the following tasks. In the end, your algorithm should generate similar results as Fig. 1 shows (maybe not as good).
- write your own code to sample patches at multiple scales, over different positions;
- Run your face classifier over all patches sampled from an given image, and classify whether they are faces or no-faces;
- write your own code for non-maximum suppression if multiple detections are heavily overlapped; they are actually fired by the same object. See Fig. 4. tip: for overlapping detections, keep the detection with the highest detection score.
- write your outputs as images with bounding boxes imposed onto the original images.