Description
1. Develop face recognition software using an implemented and pre-trained network for
Triplet Loss. Please note that you don’t need to implement and train it.
(a) (9 points) By this evening, zip 16 raw image files which were used for the previous
face detection assignment using the name ‘Group ID-First Name-Last Name.zip’
and upload the zip file into ‘BeachBoard – Discussions – Dataset – Face Images by
clicking ‘Start a New Thread’. Write your Group ID-First Name-Last Name
at the subject line. (All of team members should upload their own image files
separately.)
(b) Download image of all members of all groups from BeachBoard.
(c) Please review this article.
(d) Search and study an implemented and pre-trained deep learning architecture to
generate embedding vectors of Triplet Loss. You can use any open source code.
(e) Adapt one or more implementations, and import them into your workstation.
(f) (25 points) By combining one of the member’s face detection software and pretrained deep learning architecture, implement software as follow. Please use the
name image2vect.py
• Input: An image (x) which includes a human face.
• Intermediate step: Find bounding box of the face in the input.
• Output: Embedding vector (f(x) ∈ R
d
) in d-dimensional Euclidean space of
the bounding box. The embedding vector should be normalized, i.e.
P
||f(x)||2 =
d
i
fi(x)
2 = 1.
(g) (25 points) Implement group member recognition software as follow. Use the name
imageFinder.py.
• Input: A group ID
• Intermediate steps
1. Using the highest pc of the face detection software, store one image for each
member in the group. In total, you should have three images since a group
includes three members.
2. Using image2vect.py, recognize the images of the group members from
333 images. Definitely, you can’t utilize the file names.
• Output: Images of group members.
CECS 551 Assignment 9 – Page 2 of 2
(h) (20 points) At least for three distinct recognition threshold τ values and evaluate
the performance of imageFinder.py. Summarize your result into a single pdf file.
The report should include
• The average precision and recall values of all group IDs.
Precision = # of correctly recognized images
# of recognized images
Recall = # of correctly recognized images
# of actuall images in the group
• Software usage instruction. (How to run your programs)
(i) Submit the report and the software package including necessary files for the software
such as pre-trained weight values and input image files.