Gender Classification in Video

This project aims to assign a gender to the humans detected by a video. The input is a video taken from a moving car on the streets of Philadelphia. As a result, the human objects are full body and the resolution of facial expressions is low. Features used are presence of bags, length of clothes and presence of pink among others to arrive at a hypothesis on gender. This information could be useful for media and advertising companies who by using such software could change electronic ads on billboards to suit the gender around the board. They could do this on the fly by taking constant readings of the people in the surrounding. It could also prove invaluable in study Sample Female Frame: Features Implemented: Leg Ratio: We calculate the amount of skin showing in the region where the legs should be. In our data sample, there were several men wearing shorts but the fraction of skin showing on their leg compared to their height is less than what females show wearing skirts or shirts. After several trials on a small training set, we calculated this fraction. On the other hand, if there was no skin showing at all, we emphasized on this picture being a male since there weren’t many females wearing full length pants in the training set. of sociology to study changing gender trends in a specific region. To help determine the gender of this picture, two major features came into account. a)Leg Exposure: Such a large amount of leg exposure is indicative of skirts and thereby, females. To calculate this, we took a sample of the color from her face and searched downwards. Once a similar color was found it was traced further downward to the feet and the resulting ratio was tabulated. b)Low Shirt: This is another indication of a woman. The color of the face k d d ill i h d If hi di i ifi h Issues: Often on bright sunny days, the road or the objects around the person in question would appear bright enough to match the color of the skin tracking downwards. This gave an error. We made sure that while tracing downwards if there was a gap, it should not exceed a certain fraction of the image’s height. Neck: After we found the head, we extracted the color of the face by sampling the surrounding data. The neck area was explored in the downward, as Alignment: was trac e own t t was c ange . t s stance was s gn cant, t e result was a higher chance of being a woman. Sample Comparison Frames: well as left and right directions. This way we get two values for depth and width. Based on the reasoning that females tend to wear lower shirts, we incorporated this into our code. Issues: These were a lot harder to handle when compared to the leg ratio implementation. We haven’t solved the bright sunny day problem. Person’s Actual Height: We noticed the pixels above an obviously taller person’s head is lesser in fraction to a shorter individual’s pixel fraction. Using this tried to ti t l ti h i ht f th It’ f t th t th l i This wasn't a crucial part of our assignment and as such was not given much attention. The main purpose of this was to extract the pixel height of the human in the frame and to get an understanding of where the head, hand and feet exist. It also helps in marking the end point for the leg ratio feature.