CSE x0537 – Biometrics, Spring 2023, Professor Kevin W.Bowyer CSE x0537 – Biometrics, Spring 2023, Professor Kevin W.Bowyer Face recognition practical #2: One-to-many impostor matches. https://colab.research.google.com/drive/1XiszpzpGptEJ7Qo3klyyteuALIDGQp1-?usp=sharing This assignment involves an experiment to find the 1-to-many,rank-one impostor matches to some images of you, and looking at the identity of the rank-one impostor matches to your images. This assignment involves running the steps that are set up for you in this Colab notebook: here. Before you start working with this notebook, save a copy of it in your Drive, otherwise your code won’t be saved. Click “Edit” and select “Save a copy in Drive”. Enable GPU to speed up your experiment. Click ‘Runtime’,select ‘Change runtime type’ and select ‘GPU’ as ‘Hardware accelerator’. Run cells 1, 2 and 3. This gets you to the point of having the RetinaFace face detector, the ArcFace face matcher, and the set of impostor images ready for the experiment. The next step involves having some images of you. Take (1) a regular plain selfie, (2) an image wearing dark sunglasses, (3) an image wearing a covid face mask, and (4) an image with both dark sunglasses and face mask. Name your images “1”, “2”, “3”, “4”. After you run the cell 3. you should be able to see “my_images” folder by clicking the folder icon on the left panel. It might take some time for the “my_images” folder to appear. Click refresh icon to refresh files if you don’t see the folder. Right click to the folder to upload your images. After your images are uploaded to the my_images folder, run cell 4. This should detect the face in each of your images and place the cropped, normalized version of the detected face in the my_images_detected_faces folder. A “No face found” message will appear if the RetinaFace detection algorithm could not find a face in one of your images. If this happens, first check that you uploaded the correct image with the correct name. If you still have a problem, contact the TA. (Kağan Öztürk, [email protected].) If there are multiple faces in the image that you upload, the detector will likely return the largest or “strongest” face in the image; you can crop the image to get rid of a competing face. Next run cell five to extract features using ArcFace. Then run cell six. This will match your face images against the images of the other persons. You should see the highest 10 scores for all 4 * 1000 impostorpairs at the bottom of the output cell. Now, write the python code to the cell 7 that will, for each of your 4 images, find the 10 highest-similarity impostor matchesfor each one of the 4 images; also, write code in cell 7 to find the genuine scores of a total of 4 pairs of images of yourself. Your writeup for this experiment should include the followingsections: 1. Your python code for finding the 10 highest-similarity impostor pairs, for each of your 4 images. Also add thescreenshot of the output cell 2. Your python code for finding the genuine scores. Also add the screenshot of the output cell. 3. A paragraph explaining what similarities you see between your images and the highest-similarity images of other persons. (same gender? same hairstyle? also has / doesn’t have glasses? …) 4. A paragraph explaining how similar the highest-similarity impostor images are across your four images. Do the same 10 impostor images show up for each of your images? Or, does each of your 4 images have a completely different list of 10 highest-similarity impostor matches? Or something inbetween? 5. A paragraph explaining how your results for genuine pairs vary with glasses and mask. For the 4 images of you, there are six genuine pairs: plain-to-sunglasses, plain-to-masked, plain-to-(sunglasses+masked), sunglasses-to-masked, sunglasses-to-(sunglasses+masked), masked-to-(sunglasses+masked). How do they rank in terms of similarity score? 6. A paragraph explaining how your results in section 3 make sense or don’t make sense in the context of the within-gender and cross-gender distributions that you computed in face recognition practical #1.