Course Project

The course project is an opportunity to investigate a computer vision problem in greater depth. Projects may focus on a method, an empirical question, or an application involving images, video, or other visual data. Topics at the intersection of computer vision with machine learning, graphics, language, human-computer interaction, or scientific and medical applications are welcome.

The project is worth 25% of the final course grade and has three deliverables:

All written submissions should use the CVPR paper format.

• Teams and Scope
• Proposal — 5%
• Presentation — 5%
• Final Report — 15%
• Collaboration and Use of AI
• Choosing a Topic
• Computing Resources

Teams and Scope

Projects will normally be completed in teams of two. Individual projects or larger teams require prior approval. If you need help finding a partner, contact the course staff.

The project should be a focused, semester-scale investigation rather than an attempt to build an entire production system. A strong project should:

• pose a clear question or technical objective;
• build on relevant prior work;
• implement or adapt an appropriate method;
• include meaningful baselines and evaluation; and
• analyze successes, failures, and limitations.

The amount of work should scale with the size of the team and should be distributed reasonably evenly. Every final report must include a brief statement describing each member’s contributions.

You may build on public datasets, code, and pretrained models, and you may use datasets you have worked with previously. However, the central work submitted for this project must be new work completed during the course. Any overlap with prior or concurrent research must be disclosed in the proposal and final report.

Proposal — 5%

The proposal should be approximately two pages, excluding references. Its purpose is to establish that the question is meaningful, the evaluation is credible, and the project is feasible within the semester.

The proposal should include:

• the project title and team members;
• the problem, motivation, and proposed contribution;
• a brief discussion of at least four relevant papers;
• the datasets, baselines, and evaluation metrics you plan to use;
• a concrete plan of work, including the division of responsibilities; and
• major technical or data risks and a reasonable fallback plan.

The proposal will be evaluated on:

The proposal is due at 11:59 PM on Thursday, October 22 through Gradescope.

Presentation — 5%

All teams will present their projects during the scheduled final-examination period. The presentation will normally take the form of a conference-style poster session; exact logistics will be announced later. Live demonstrations are encouraged when they materially help explain the work.

The presentation should clearly communicate:

• the problem and why it matters;
• the central method or hypothesis;
• the experimental setup and principal results;
• important failure cases and limitations; and
• the contribution of the project relative to the baselines and prior work.

The presentation will be evaluated on the clarity of the explanation, the quality of the technical evidence, the design of the poster or demonstration, and the team’s ability to answer questions.

Final Report — 15%

The final report should be no more than eight pages, excluding references. It should read like a conference paper and clearly distinguish completed work from proposed extensions.

A typical report will contain:

• Introduction: problem definition, motivation, and contributions;
• Related work: the most relevant prior approaches and how the project differs;
• Method: the proposed approach, design choices, and implementation details;
• Experiments: datasets, baselines, metrics, experimental protocol, and results;
• Analysis: ablations, failure cases, limitations, or other evidence that explains the results; and
• Conclusion: principal findings and promising directions for future work.

The report will be evaluated on:

The final report is due at 11:59 PM on Friday, December 18 through Gradescope. Late days do not apply to the final report.

Collaboration and Use of AI

Collaboration within the project team is expected. Assistance from people outside the team and substantive use of AI-generated code, text, images, or experimental analysis must be disclosed in the proposal and final report. Regardless of the tools used, team members are responsible for verifying and understanding everything they submit.

AI tools may assist with implementation, debugging, literature discovery, or editing, but they do not replace the project’s intellectual contribution. The problem formulation, experimental reasoning, interpretation of results, and final claims must represent the team’s own work. Fabricated citations, experiments, or results are violations of the University’s Academic Integrity Policy.

Choosing a Topic

You are encouraged to propose a topic connected to your interests. Possible directions include:

• evaluating or adapting self-supervised and vision-language models;
• open-vocabulary recognition, detection, or segmentation;
• image or video generation, editing, and consistency;
• correspondence, tracking, pose, or shape estimation;
• neural rendering and 3D scene representations;
• robustness, uncertainty, dataset shift, and evaluation; or
• computer vision for scientific, medical, ecological, or remote-sensing applications.

Recent papers and workshop proceedings from CVPR, ICCV, ECCV, NeurIPS, and ICML are useful sources of ideas. A good project does not need to train a large model from scratch: careful evaluation, adaptation, or analysis of an existing method can make a strong contribution when it asks a clear question and provides convincing evidence.

Computing Resources

Computing instructions and available course resources will be announced separately. Projects should be scoped so that their core experiments are feasible with the available resources. Whenever possible, use public datasets, pretrained models, and reproducible evaluation pipelines rather than relying on large-scale training.