Digital Content

Computer Vision Project - Intro

Select a facial recognition and comparison app that provides both “metric” results, such as probability of gender or race, and some kind of comparison result, such as “Celebrity Look Alikes” or “Face Searching”. A few options are listed below. If you select your own program or app, be sure to research any potential privacy issues and do not download / install anything directly from an unknown / un-trusted website.

https://www.betafaceapi.com/demo.html#page-header

https://www.faceplusplus.com/attributes/
https://www.faceplusplus.com/face-searching/

Project Steps

For the remainder of this project, each time you run the facial analysis program on an image, take a screenshot of the result details for both the metrics and the photo comparison results (which should include the original photo uploaded). These apps only work with standard image files such as JPG or PNG - you cannot upload photoshop files directly, so, you will need to export JPG files when working with the composite images.

1. Run the analysis on each of the original photos used in your selfie composite image (at least 3 different images that you took), and then run it on your composite image. Record all of the results via screenshots (this will be how you record all results from this point on - don’t forget to record both sets of results.)

2. If there was a high degree of variation in the results for each of your original photos, apply the levels + color adjustment processes to each of these photos to try to produce a result more consistent with the results from one image. Export these as new images (do not overwrite the original files) and run the program again.

If the results from the first round of images were relatively uniform, apply the levels and color adjustment processes to the photos to try to elicit different results and increase the variation to each of the photos. Export these as new images (do not overwrite the original files) and run the program again.

3. Simulate lower resolution by pixelating at least 3 of the images you have already run through the program. Filter > Pixelate > Mosaic > between 10 - 30 pixels based on the size of the image. Export these as new images (do not overwrite the original files) and run the program again.

4. Working with only your composite image, make adjustments to the order, levels, effects and blending modes of the different layers in order to produce 3 additional unique composite images - run each through the program. Try to make one of these composite images “match” how you identity as closely as possible, given the constraints of the program.

Questions

1. How variable were the results from your first group of images? Were you able to diversify or your results with the second round of images? What components of the images did you alter, and what was the reasoning behind these decisions?

2. Based on the results from the first two rounds (both in terms of the “metrics” and comparison results), how do you think the algorithm “read” / interpreted your images? What visual aspects do you think the program was looking for in order to determine specific outcomes or results?

3. After pixelating the images, what changes were there in results? Which specific metrics / comparisons were significantly less or more accurate, in terms of how you identify? What kind of conclusions do you draw from these results.

4. Based on the outcomes, did you encounter any biases within the system, both in terms of how the system interpreted images and/or the images the system has available for comparisons? Describe these.