senior-computer-vision_skill

Overview

This skill provides a practical, end-to-end engineering workflow for building robust computer vision systems. It covers dataset design, model selection across classification/detection/segmentation, evaluation strategies, and deployment optimizations. The guidance is focused on measurable outcomes, generalization, and realistic production constraints.

How this skill works

The skill inspects and formalizes the problem (task type, latency, target hardware) and enforces a dataset-first approach: defining taxonomies, edge cases, splits, and metrics. It prescribes disciplined training practices (baselines, ablations, error analysis) and checks inference parity during deployment using techniques like batching, quantization, and TensorRT where appropriate. An optional dataset inventory script creates a lightweight report of folder-structured image datasets to speed audits without ML dependencies.

When to use it

• Scoping a new CV project and choosing between classification, detection, or segmentation.
• Auditing a dataset for label consistency, edge-case coverage, and split bias.
• Defining evaluation metrics and split strategies to ensure reproducible benchmarks.
• Diagnosing model failures, dataset drift, or inference performance regressions.
• Preparing models for production with latency, memory, and hardware constraints in mind.

Best practices

• Specify the task and target hardware up front to guide model and optimization choices.
• Treat the dataset as a product: document taxonomy, label rules, and notable edge cases.
• Run simple, reproducible baselines and systematic ablations before major changes.
• Use holdout splits and realistic metrics aligned with business needs; avoid only train/val tweaks.
• Validate preprocessing parity between training and deployment; measure quantized/optimized models in-situ.

Example use cases

• Designing a dataset and evaluation plan for a multi-class defect detection system.
• Choosing and benchmarking segmentation backbones for real-time robotics deployment.
• Auditing a public image dataset to find class imbalance and ambiguous labels before training.
• Optimizing an object detector for mobile hardware using pruning, quantization, and batching.
• Running the dataset inventory script to generate a quick report for dataset reviewers.