Written Homework 5

CSSE 461 - Computer Vision

Instructions

This assignment prepares you to write a final project proposal. You’ll practice a workflow for finding recent computer vision methods: identifying problems, locating papers, and evaluating whether a method is feasible for a class project.

Your homework solutions must be typed. Upload a single PDF to Gradescope. Include screenshots where requested.

AI Use

This assignment includes a section where you use AI for research. You don’t need to use the course AI prompt while working on this assignment. If you have a principled objection to personally using AI, please talk to me! I’ll give you a modified version of the assignment that doesn’t involve using AI.

Problems

Problem 1: Finding Methods for a CV Problem

We’re going to find out what the latest and best techniques are for a specific computer vision problem: object tracking. Even more specifically, let’s say that we’d looking for something that can run in real-time, or close to it. As you’re doing this assignment, any method for object tracking is relevant, but methods that run fast are more relevant.

Part a: LLM Search.

Using an LLM of your choice, ask for a list of the best methods currently known for real-time object tracking.

For this task, you’ll want to use an LLM with web search enabled—this grounds responses in real sources and reduces hallucinations. These tools cite their sources, which makes verification much easier. Some common free options include Perplexity, Claude, or ChatGPT with browsing enabled.

LLM prompting:

Taking a moment to prepare your prompt pays off. Be specific about what you are looking for. Specify that you are requesting a research task requiring web search, and that you require cited, linked evidence to validate the model’s recommendations.

In your writeup, show me the prompt you used, tell me which specific LLM model you used, and give me the names of at least three methods that the LLM recommended. For each one, do enough background research to tell me:

• The name of the method
• Whether the method really exists (i.e., was it an LLM hallucination?)
• In which year was the method published

Part b: Google Scholar.

Let’s try again with Google Scholar. Run a search for “real-time object tracking”. Find three papers with these properties:

• They were published within the last three years
• The title sounds on-topic for our purposes
• They have a relatively high number of citations

(Remember that it’s normal for a paper published three years ago to have several times more citations than a paper published this year.)

For each of these three papers, download the paper and open it. Looking only at the figures, try to screenshot an image that shows what this paper can do.

In your writeup, give me:

• The name of the paper
• A screenshot showing what this paper does

Part c: HuggingFace Papers.

Once more! This time we’ll use HuggingFace Papers. Search for “object tracking” and browse the results. I’d like you to pick three promising papers based on recency, popularity, and relevance.

HuggingFace is useful because papers are often linked directly to model implementations and demos. When a paper has an associated model, you’ll see links to the model page, which tells you about code, weights, and hardware requirements.

Include these in your writeup:

• The name of the paper
• A screenshot showing what this paper does
• A one-sentence justification for why you picked this paper
• A link to code or a model page (if available)

Part d: Feasibility Check.

Based on all of your research above, pick the method you think would be the best choice for a class project around real-time object tracking.

Now that you’ve made your choice, lets do one final check for feasibility. Do enough research to answer the following, and put your answers in your writeup:

• Is source code available? (Provide a link.)
• Are pretrained model weights available, or would you need to train from scratch?
• What hardware does it require? (Look for GPU memory requirements in the README or paper.) Sometimes you won’t find this information.
• Could you realistically run this on the hardware available to you?

It’s okay if your top choice turns out to be infeasible—that’s useful information! Just tell me what you found and what you would try next.

Problem 2: Brainstorm Your Interests

Based on what you learned today, list 2-3 computer vision problems that you might want to explore for your final project. You don’t need to research them yet. I’m asking you to brainstorm areas that sound interesting.

For each one, write one sentence about why it caught your attention.