Computer vision internships: roles and an evaluation project

Choose the task before the model

An image label, an object boundary and a tracked sequence require different tests. For a first project, classify nonsensitive objects in openly licensed images. Define the labels and how you will handle ambiguous examples. Keep images from the same original scene together when creating training and evaluation groups so nearly identical frames do not appear on both sides.

Use a simple baseline before a more elaborate model. Record conditions that could change performance, such as lighting. The invented results below are image-classification counts; they are not measurements from an employer’s model.

Identical headline accuracy can hide different weaknesses

Both models classify 105 of 120 images correctly. The overall score is 87.5%, but their low-light behavior differs.

• B trades five bright-image successes for five dim-image successes. Whether that is useful depends on the intended setting and which errors matter.
• For a hypothetical future mix of half bright and half dim images, holding each condition’s rates fixed gives A 72.5% and B 82.5%. This is reweighting an assumption, not measuring future performance.
• Only 20 dim images were tested. Inspect those errors and collect more relevant examples before treating the estimated rates as stable.

Build a portfolio around one error you understand

Select errors by a stated rule, such as the first five misclassified images within each condition, rather than choosing only memorable examples. Check whether each is a label problem, an ambiguous image or a model failure. Keep the original prediction and explain the evidence for your interpretation.

Your repository should identify the dataset permissions, split rule, baseline, condition-level counts and how to reproduce the evaluation. Measure runtime on named hardware when speed matters; avoid comparing times gathered under different conditions as though they were directly comparable.

Distinguish training a model from fine-tuning one or evaluating an existing model. All can provide useful evidence when described accurately. In the interview, explain which error you would investigate next and what observation would change your current explanation. Ask whether the internship’s deliverable is a model, an evaluation, an inference improvement or a paper.

Before you apply

Eligibility

Check the individual degree requirement. A program accepting several degree levels can still have PhD-only perception-research roles.

Timing

Use the current university requisition for the intended term. A research-group page does not establish an application window.

Location

Confirm the product or research site and how the required images and compute are accessed.

Pay

Compare offers for the same degree level, term and location, using the employer’s stated amount.

Sources: Waymo early careers · NVIDIA university recruiting

Job titles to search

Computer vision engineering intern

Improve an image-based component in a working system.

Perception research intern

Investigate a defined method or research question.

Vision inference or evaluation intern

Measure errors, latency or resource use around a model.

Where to look

Search routes, not confirmed vacancies.

Compare the work and try an example

Use a role comparison or worked exercise to choose your preparation before tailoring an application.

• Internship skills: distinguish required, preferred and taught on the job
• Build an economic consulting work sample that another person can reproduce
• Citadel versus Citadel Securities internships: choose the employer and role