vision-expert

Computer Vision Expert

Overview

Use this skill to turn vague "make the model see better" requests into explicit computer-vision workflows with measurable inputs, outputs, and failure modes. Prefer simple, testable pipelines and evidence-driven evaluation over fashionable models or premature complexity.

Workflow

1. Detect project state. Consult references/topic_index.md first to match the user problem to a CV task, default library, and guide file. Then inspect the target repo:

• Search the target repo for existing CV source files, model configs, pipeline scripts, and evaluation code.
• Search for existing tests or benchmarks that characterize current behavior.
• If any of these exist, follow the existing-pipeline path: inspect model weights, validation split, and failure logs; establish a per-class baseline before changing anything; tie each proposed change to a specific failing case; prove improvement with before/after metrics on the held-out set.
• If none exist, follow the greenfield path: seed a balanced 100-500 image set with a held-out test split; write a vision contract covering resolution, class taxonomy, FPS budget, and miss-vs-false-alarm tolerance; then build and validate the minimal pipeline.
• Read references/project_workflow.md for the full branching workflow, vision contract spec, and CV review checklist.

1. Define the exact vision task.

• Determine whether the task is classification, detection, segmentation, keypoints, tracking, OCR, retrieval, restoration, or measurement.
• Identify the input domain: still image, video, live camera, document scan, microscopy, satellite, industrial, or another domain.
• Define success in measurable terms such as accuracy, recall, latency, FPS, false positives, localization error, or downstream business impact.
• Read references/task_selection.md when the request is underspecified or multiple CV framings are possible.