辅导COMP90086编程课程、辅导Python，c++程序
- 首页 >> Algorithm 算法 The University of Melbourne
School of Computing and Information Systems
COMP90086 Computer Vision, 2021 Semester 2
Final Project: Fine-grained localisation
Project type: Group (teams of 2)
Due: 7pm, 22 Oct 2021
Submission: Source code and written report (as .pdf)
Marks: The assignment will be marked out of 30 points, and will contribute 30% of your
total mark.
Geolocation is the problem of localising a person or device in the world using sensor data. Depending
on the device, the environment, and the level of accuracy required, geolocation may rely on GPS
coordinates, network routing addresses, or image data. Geolocation is an important problem in many
AI and computing applications, from autonomous vehicle navigation to search engine queries based
on the user’s current location (e.g., “restaurants near me”).
In this project, you will investigate the problem of fine-grained geolocation in a small indoor/outdoor
environment (an art museum). Image information is particularly important for this type of problem,
because other sources of information, like GPS, may not be accurate enough to provide fine-grained
position data and may not be able to distinguish between different floors in indoor environments.
Your task is to develop a method to recognize the location from which an image was taken. You
will be provided a dataset of images with position data to train your method. How you approach the
problem is up to you. The following are some possible approaches:
? Match each image in the test set to the most similar image in the training set, using any visual
features you wish to measure “similarity,” and assume the test image has the same position data
as its closest match. (Note that there is no guarantee that the test images will come from exactly
the same locations as the images in training set, but since they come from the same museum
environment they are likely to be from nearby locations.)
Identify key features, objects, or text in the test images and use these to locate training images
which show the same features, objects, or text.
Match each image in the test set to multiple near neighbours in the training set, and develop a
method to compute the test image’s most likely location based on multiple nearby views.
Use matching features and geometric constraints to compute the likely change in pose between
training and test views.
Or any combination of the above.
Note that these are only suggestions to help you get started; you are free to use your own ideas.
Whatever methods you choose, you are expected to evaluate these methods using the provided data,
to critically analyse the results, and to justify your design choices in your final report. Your evaluation
should include error analysis, where you attempt to understand where your method works well and
where it fails.
You are encouraged to use existing computer vision libraries your implementation. You may also use
existing models or pretrained features as part of your implementation. However, your method should
be your own; you may not simply submit an existing model for this problem.
Dataset
Figure 1: Example training images
The dataset is a collection of images taken in and around an art museum (the Getty Center in Los
Angeles, U.S.A.). Example images are shown in Figure 1. The dataset is split into 7500 training
images and 1200 test images. Each image in the training set is annotated with positional data, which
is an (x,y) value derived from a mapping algorithm. You can assume that the (x,y) values accurately
reflect position in the real world, although the units of these values are unknown. The training dataset
includes multiple views from each of several locations around the museum. Different views from the
same location are denoted with a suffix (e.g., “ 1”, “ 2”, etc.).
The images are rendered from Google Streetview images, simulating a camera with a 73.7 deg hor-
izontal x 53.1 deg vertical field of view. The optical centre of the camera is in the centre of the
image and the lens has no radial distortion. However, because the images are simulated from Google
Streetview imagery, they may contain artefacts or distortion from the Streetview panorama stitching
process. Faces in the images have been blurred for privacy. Please note that because the images were
collected in a real-world public environment, it is possible that they may contain inappropriate or
offensive content.
Scoring Predictions
You should submit your predictions for the test images on Kaggle. Your submissions for Kaggle
should follow the same format as the train.csv annotation file, with three columns: id,x,y. id
should be a string corresponding to a test image name, and x and y should be the predicted position
of that image.
The evaluation metric for this competition is the mean absolute error in x and y computed on the
test set. This can also be thought of as the Manhattan distance between the true and predicted (x,y)
coordinates, averaged over all N test images:
(Although Euclidean distance would probably make more sense for this task, Kaggle does not have
an evaluation metric which computes Euclidean distance.)
Kaggle
To join the competition on Kaggle and submit your results, you will need to register at https:
//www.kaggle.com/.
Please use the “Register with Google” option and use your @student.unimelb.edu.au email address
to make an account. Please use only your group member student IDs as your team name (e.g.,
“1234&5678”). Submissions from teams which do not correspond to valid student IDs will be treated
as fake submissions and ignored.
Once you have registered for Kaggle, you will be able to join the COMP90086 Final Project compe-
tition using the link under Final Project: Code in the Assignments tab on the Canvas LMS. After
following that link, you will need to click the “Join Competition” button and agree to the competition
rules.
Group Formation
You should complete this project in a group of 2. You are required to register your group membership
on Canvas by completing the “Project Group Registration” survey under “Quizzes.” You may modify
your group membership at any time up until the survey due date, but after the survey closes we will
consider the group membership final.
Submission
Submission will be made via the Canvas LMS. Please submit your code and written report separately
under the Final Project: Code and the Final Project: Report links on Canvas.
Your code submission should include your model code, your test predictions (in Kaggle format), a
readme file that explains how to run your code, and any additional files we would need to recreate
your results. You should not include the provided train/test images in your code submission, but your
readme file should explain where your code expects to find these images.
Your written report should be a .pdf that includes the description, analysis, and comparative assess-
ment of the method(s) you developed to solve this problem. The report should follow the style of a
short conference paper with no more than four A4 pages of content (excluding references, which
can extend to a 5th page). The report should follow the style and format of an IEEE conference
short paper. The IEEE Conference Template for Word, LaTeX, and Overleaf is available here:
https://www.ieee.org/conferences/publishing/templates.html.
Your report should explain the design choices in your method and justify these based on your un-
derstanding of computer vision theory. You should explain the experimentation steps you followed
to develop and improve on your basic method, and report your final evaluation result. Your method,
experiments, and evaluation results should be explained in sufficient detail for readers to understand
them without having to look at your code. You should include an error analysis which assesses where
your method performs well and where it fails, provide an explanation of the errors based on your un-
derstanding of the method, and give suggestions for future improvements. Your report should include
tables, graphs, figures, and/or images as appropriate to explain and illustrate your results.
Evaluation
Your submission will be marked on the follow grounds:
Component Marks Criteria
Report writing 5 Clarity of writing and report organisation; use of tables, fig-
ures, and/or images to illustrate and support results
Report method and
justification
10 Correctness of method; motivation and justification of design
choices based on computer vision theory
Report experimenta-
tion and evaluation
10 Quality of experimentation, evaluation, and error analysis;
interpretation of results and experimental conclusions
Kaggle submission 3 Kaggle performance
Team contribution 2 Group self-assessment
The report is marked out of 25 marks, distributed between the writing, method and justification, and
experimentation and evaluation as shown above.
In addition to the report marks, up to 3 marks will be given for performance on the Kaggle leaderboard.
To obtain the full 3 marks, a team must make a Kaggle submission that performs reasonably above a
simple baseline. 1-2 marks will be given for Kaggle submissions which perform at or only marginally
above the baseline, and 0 marks will be given for submissions which perform at chance. Teams which
do not submit results to Kaggle will receive 0 performance marks.
Up to 2 marks will be given for team contribution. Each group member will be asked to provide
a self-assessment of their own and their teammate’s contribution to the group project, and to mark
themselves and their teammate out of 2 (2 = contributed strongly to the project, 1 = made a small
contribution to the project, 0 = minimal or no contribution to the project). Your final team contribution
mark will be based on the mark assigned to you by your teammate (and their team contribution mark
will be based on the mark you assign to them).
Late submission
The submission mechanism will stay open for one week after the submission deadline. Late submis-
sions will be penalised at 10% of the total possible mark per 24-hour period after the original deadline.
Submissions will be closed 7 days (168 hours) after the published assignment deadline, and no further
submissions will be accepted after this point.
Updates to the assignment specifications
If any changes or clarifications are made to the project specification, these will be posted on the LMS.
Academic misconduct
You are welcome — indeed encouraged — to collaborate with your peers in terms of the conceptual-
isation and framing of the problem. For example, we encourage you to discuss what the assignment
specification is asking you to do, or what you would need to implement to be able to respond to a
question.
However, sharing materials — for example, showing other students your code or colluding in writ-
ing responses to questions — or plagiarising existing code or material will be considered cheating.
Your submission must be your own original, individual work. We will invoke University’s Academic
Misconduct policy (http://academichonesty.unimelb.edu.au/policy.html) where
inappropriate levels of plagiarism or collusion are deemed to have taken place.
School of Computing and Information Systems
COMP90086 Computer Vision, 2021 Semester 2
Final Project: Fine-grained localisation
Project type: Group (teams of 2)
Due: 7pm, 22 Oct 2021
Submission: Source code and written report (as .pdf)
Marks: The assignment will be marked out of 30 points, and will contribute 30% of your
total mark.
Geolocation is the problem of localising a person or device in the world using sensor data. Depending
on the device, the environment, and the level of accuracy required, geolocation may rely on GPS
coordinates, network routing addresses, or image data. Geolocation is an important problem in many
AI and computing applications, from autonomous vehicle navigation to search engine queries based
on the user’s current location (e.g., “restaurants near me”).
In this project, you will investigate the problem of fine-grained geolocation in a small indoor/outdoor
environment (an art museum). Image information is particularly important for this type of problem,
because other sources of information, like GPS, may not be accurate enough to provide fine-grained
position data and may not be able to distinguish between different floors in indoor environments.
Your task is to develop a method to recognize the location from which an image was taken. You
will be provided a dataset of images with position data to train your method. How you approach the
problem is up to you. The following are some possible approaches:
? Match each image in the test set to the most similar image in the training set, using any visual
features you wish to measure “similarity,” and assume the test image has the same position data
as its closest match. (Note that there is no guarantee that the test images will come from exactly
the same locations as the images in training set, but since they come from the same museum
environment they are likely to be from nearby locations.)
Identify key features, objects, or text in the test images and use these to locate training images
which show the same features, objects, or text.
Match each image in the test set to multiple near neighbours in the training set, and develop a
method to compute the test image’s most likely location based on multiple nearby views.
Use matching features and geometric constraints to compute the likely change in pose between
training and test views.
Or any combination of the above.
Note that these are only suggestions to help you get started; you are free to use your own ideas.
Whatever methods you choose, you are expected to evaluate these methods using the provided data,
to critically analyse the results, and to justify your design choices in your final report. Your evaluation
should include error analysis, where you attempt to understand where your method works well and
where it fails.
You are encouraged to use existing computer vision libraries your implementation. You may also use
existing models or pretrained features as part of your implementation. However, your method should
be your own; you may not simply submit an existing model for this problem.
Dataset
Figure 1: Example training images
The dataset is a collection of images taken in and around an art museum (the Getty Center in Los
Angeles, U.S.A.). Example images are shown in Figure 1. The dataset is split into 7500 training
images and 1200 test images. Each image in the training set is annotated with positional data, which
is an (x,y) value derived from a mapping algorithm. You can assume that the (x,y) values accurately
reflect position in the real world, although the units of these values are unknown. The training dataset
includes multiple views from each of several locations around the museum. Different views from the
same location are denoted with a suffix (e.g., “ 1”, “ 2”, etc.).
The images are rendered from Google Streetview images, simulating a camera with a 73.7 deg hor-
izontal x 53.1 deg vertical field of view. The optical centre of the camera is in the centre of the
image and the lens has no radial distortion. However, because the images are simulated from Google
Streetview imagery, they may contain artefacts or distortion from the Streetview panorama stitching
process. Faces in the images have been blurred for privacy. Please note that because the images were
collected in a real-world public environment, it is possible that they may contain inappropriate or
offensive content.
Scoring Predictions
You should submit your predictions for the test images on Kaggle. Your submissions for Kaggle
should follow the same format as the train.csv annotation file, with three columns: id,x,y. id
should be a string corresponding to a test image name, and x and y should be the predicted position
of that image.
The evaluation metric for this competition is the mean absolute error in x and y computed on the
test set. This can also be thought of as the Manhattan distance between the true and predicted (x,y)
coordinates, averaged over all N test images:
(Although Euclidean distance would probably make more sense for this task, Kaggle does not have
an evaluation metric which computes Euclidean distance.)
Kaggle
To join the competition on Kaggle and submit your results, you will need to register at https:
//www.kaggle.com/.
Please use the “Register with Google” option and use your @student.unimelb.edu.au email address
to make an account. Please use only your group member student IDs as your team name (e.g.,
“1234&5678”). Submissions from teams which do not correspond to valid student IDs will be treated
as fake submissions and ignored.
Once you have registered for Kaggle, you will be able to join the COMP90086 Final Project compe-
tition using the link under Final Project: Code in the Assignments tab on the Canvas LMS. After
following that link, you will need to click the “Join Competition” button and agree to the competition
rules.
Group Formation
You should complete this project in a group of 2. You are required to register your group membership
on Canvas by completing the “Project Group Registration” survey under “Quizzes.” You may modify
your group membership at any time up until the survey due date, but after the survey closes we will
consider the group membership final.
Submission
Submission will be made via the Canvas LMS. Please submit your code and written report separately
under the Final Project: Code and the Final Project: Report links on Canvas.
Your code submission should include your model code, your test predictions (in Kaggle format), a
readme file that explains how to run your code, and any additional files we would need to recreate
your results. You should not include the provided train/test images in your code submission, but your
readme file should explain where your code expects to find these images.
Your written report should be a .pdf that includes the description, analysis, and comparative assess-
ment of the method(s) you developed to solve this problem. The report should follow the style of a
short conference paper with no more than four A4 pages of content (excluding references, which
can extend to a 5th page). The report should follow the style and format of an IEEE conference
short paper. The IEEE Conference Template for Word, LaTeX, and Overleaf is available here:
https://www.ieee.org/conferences/publishing/templates.html.
Your report should explain the design choices in your method and justify these based on your un-
derstanding of computer vision theory. You should explain the experimentation steps you followed
to develop and improve on your basic method, and report your final evaluation result. Your method,
experiments, and evaluation results should be explained in sufficient detail for readers to understand
them without having to look at your code. You should include an error analysis which assesses where
your method performs well and where it fails, provide an explanation of the errors based on your un-
derstanding of the method, and give suggestions for future improvements. Your report should include
tables, graphs, figures, and/or images as appropriate to explain and illustrate your results.
Evaluation
Your submission will be marked on the follow grounds:
Component Marks Criteria
Report writing 5 Clarity of writing and report organisation; use of tables, fig-
ures, and/or images to illustrate and support results
Report method and
justification
10 Correctness of method; motivation and justification of design
choices based on computer vision theory
Report experimenta-
tion and evaluation
10 Quality of experimentation, evaluation, and error analysis;
interpretation of results and experimental conclusions
Kaggle submission 3 Kaggle performance
Team contribution 2 Group self-assessment
The report is marked out of 25 marks, distributed between the writing, method and justification, and
experimentation and evaluation as shown above.
In addition to the report marks, up to 3 marks will be given for performance on the Kaggle leaderboard.
To obtain the full 3 marks, a team must make a Kaggle submission that performs reasonably above a
simple baseline. 1-2 marks will be given for Kaggle submissions which perform at or only marginally
above the baseline, and 0 marks will be given for submissions which perform at chance. Teams which
do not submit results to Kaggle will receive 0 performance marks.
Up to 2 marks will be given for team contribution. Each group member will be asked to provide
a self-assessment of their own and their teammate’s contribution to the group project, and to mark
themselves and their teammate out of 2 (2 = contributed strongly to the project, 1 = made a small
contribution to the project, 0 = minimal or no contribution to the project). Your final team contribution
mark will be based on the mark assigned to you by your teammate (and their team contribution mark
will be based on the mark you assign to them).
Late submission
The submission mechanism will stay open for one week after the submission deadline. Late submis-
sions will be penalised at 10% of the total possible mark per 24-hour period after the original deadline.
Submissions will be closed 7 days (168 hours) after the published assignment deadline, and no further
submissions will be accepted after this point.
Updates to the assignment specifications
If any changes or clarifications are made to the project specification, these will be posted on the LMS.
Academic misconduct
You are welcome — indeed encouraged — to collaborate with your peers in terms of the conceptual-
isation and framing of the problem. For example, we encourage you to discuss what the assignment
specification is asking you to do, or what you would need to implement to be able to respond to a
question.
However, sharing materials — for example, showing other students your code or colluding in writ-
ing responses to questions — or plagiarising existing code or material will be considered cheating.
Your submission must be your own original, individual work. We will invoke University’s Academic
Misconduct policy (http://academichonesty.unimelb.edu.au/policy.html) where
inappropriate levels of plagiarism or collusion are deemed to have taken place.