Description
ASSIGNMENT 1 CS4186
You are given a collection of 5,000 images, 10 example query instances and 20 testing query
instances (you can download from this link: google driver link). Each image contains one
instance (object). Your task is to implement two methods for instance search
Specifically,
given a query instance, a method needs to find the images that contain the same query
instance from the image collection （5000 images） and then ranks them according to
similarity or confidence. The 20 testing query images are used for evaluating the performance
of your implementation.
Marking scheme
A. Report(no more than 5 pages): Write a report that briefly describes the two methods and
compares their search performance for the 10 example query instances. The performance
should be measured by mean average precision (MAP). The MAP calculation is provided.
You should also show the results for some queries by listing the retrieved images.
(30%)
B. Method (50%):
Algorithm: Implement any two methods (e.g., color histogram, LBP, SIFT, CNN) for
instance search. You can employ the methods you learn in the class, or you could choose
other methods in the literature. It is also encouraged to design the method by yourself.
(40%)
Try to combine two methods (e.g., color histogram, LBP, SIFT, CNN), which may
provide better performance than using only one method. (10%)
At least one method should be implemented by yourself, instead of calling the readymade function package.
C. Benchmarking: For each of your methods, show the retrieval result for each of the 20
testing query instances (as detailed in the submission section). The mark will be allocated
based on the retrieval performance of the method with better performance.
(20%)
DEADLINE
The submission should be done before 09-Apr 2021, 11:59pm.
Penalty on late submission: 20% of marks will be deducted per day. No submission will
be allowed after five days from the deadline. Do NOT copy code from the internet and do
not borrow other people’s code. Remember that PLAGIARISM is a serious offense for which
you may fail the class or even be expelled from the university
NOTE
1. In the provided zip file, the folders are organized as: images – 5,000 images for search;
examples – 10 example query instances; Queries – 20 testing query instances.
2. The instance of an image is given as bounding box. The bounding box is represented as
(top left corner in x, top left corner in y, width, height) in the text file. For example, the
file 01.txt specifies the bounding box of the query instance for image 01.jpg.
3. The location of instance, given as a bounding box, is also provided for 2,000 images.
4. In the examples folder, you can find the program metric_map.py for measuring search
performance in terms of MAP. The value of MAP will be high (at most 1.0) if the images
containing the query instances are ranked at top.
5. In the examples folder, the ground-truth (or answers) for the 10 examples queries can
be found in the file rank_groundtruth.txt.
6. You can use any distance or similarity function for measuring the similarity between two
images or instances.
Examples include:
Euclidean distance between a query 𝑄 = [𝑞1, 𝑞2 ⋯ 𝑞𝑛] and an image 𝐹 = [𝑓1, 𝑓2 ⋯ 𝑓𝑛]:
𝐷𝑖𝑠𝑡𝑎𝑛𝑐𝑒(𝑄, 𝐹) = √∑(𝑞𝑖 − 𝑓𝑖)
2
𝑛
𝑖=1
Cosine similarity between a query 𝑄 = [𝑞1, 𝑞2 ⋯ 𝑞𝑛] and an image 𝐹 = [𝑓1, 𝑓2 ⋯ 𝑓𝑛]:
𝑆𝑖𝑚𝑖𝑙𝑎𝑟𝑖𝑡𝑦(𝑄, 𝐹) =
∑ 𝑞𝑖𝑓𝑖
𝑛
𝑖=1
√∑ 𝑞𝑖
𝑛 2
𝑖=1 √∑ 𝑓𝑖
𝑛 2
𝑖=1
SUBMISSION
PLEASE ZIP THE FOLLOWINGS AND SUBMIT TO CANVAS:
Computer program: Python/Matlab/C++ recommended. If you want to use another
program language, please send an email to our Teaching Assistant.
Report
Rank list: A text file (rankList.txt) showing the images in descending order of similarity:
Q1: 7 12 214 350 …
Q2: 301 501 1990 2 …
Q3: 288 345 389 1290 …
Q4: 248 293 1098 2000 …
Q5: 380 287 392 478 222 …
Remark:
20 rows only (for 20 testing query instances); Each row should list the
names of the 5,000 images (named in number) in descending order of their
similarities to a query.
CONTACT PERSON
Please email Teaching Assistant Qiudan ZHANG (Email: qdzhang3-c@my.cityu.edu.hk) for
technical questions.
