computer-vision

Computer Vision

Overview

Computer vision workflows require careful attention at every stage: understanding dataset characteristics first, building a sound preprocessing and augmentation pipeline, selecting an architecture matched to dataset size and task, and evaluating with task-appropriate metrics. This skill covers the full pipeline from raw images to model evaluation.

When to Use This Skill

Use this skill when:

• Working with image or video datasets (classification, detection, segmentation, generation)
• Designing or debugging a preprocessing/augmentation pipeline
• Selecting a model architecture for a vision task
• Computing vision-specific metrics (mAP, IoU, FID, SSIM, LPIPS)
• Transfer learning decisions (freeze vs. fine-tune, learning rate schedule)

Run the EDA skill first to understand file formats, directory structure, and basic counts. Use this skill for vision-specific analysis.

---

Image Dataset Characterization

Before writing any training code, profile your dataset thoroughly.

from pathlib import Path
from PIL import Image
import numpy as np
import matplotlib.pyplot as plt
from collections import Counter
import cv2

def characterize_dataset(image_dir, extensions=('.jpg', '.jpeg', '.png', '.tiff', '.bmp')):
    image_paths = [p for p in Path(image_dir).rglob('*') if p.suffix.lower() in extensions]
    print(f"Total images: {len(image_paths)}")

    widths, heights, channels_list, aspect_ratios = [], [], [], []
    channel_means, channel_stds = [], []

    for path in image_paths:
        with Image.open(path) as img:
            w, h = img.size
            c = len(img.getbands())
            widths.append(w)
            heights.append(h)
            channels_list.append(c)
            aspect_ratios.append(w / h)

            # Per-image channel stats (sample every Nth image to stay fast)
            if len(channel_means) < 500:
                arr = np.array(img.convert('RGB'), dtype=np.float32) / 255.0
                channel_means.append(arr.mean(axis=(0,1)))
                channel_stds.append(arr.std(axis=(0,1)))

    print(f"\nWidth  — min: {min(widths)}, max: {max(widths)}, mean: {np.mean(widths):.0f}")
    print(f"Height — min: {min(heights)}, max: {max(heights)}, mean: {np.mean(heights):.0f}")
    print(f"Aspect ratio — min: {min(aspect_ratios):.2f}, max: {max(aspect_ratios):.2f}, "
          f"mean: {np.mean(aspect_ratios):.2f}")
    print(f"Channels: {Counter(channels_list)}")

    means = np.array(channel_means).mean(axis=0)
    stds  = np.array(channel_stds).mean(axis=0)
    print(f"\nChannel means (RGB): {means.round(4)}")
    print(f"Channel stds  (RGB): {stds.round(4)}")

    # Class distribution (assumes ImageFolder structure: dir/class/image.jpg)
    classes = [p.parent.name for p in image_paths]
    class_counts = Counter(classes)
    print(f"\nClass distribution ({len(class_counts)} classes):")
    for cls, cnt in sorted(class_counts.items(), key=lambda x: -x[1]):
        print(f"  {cls}: {cnt}")

    # Plot size scatter
    fig, axes = plt.subplots(1, 2, figsize=(12, 4))
    axes[0].scatter(widths, heights, alpha=0.2, s=5)
    axes[0].set_xlabel('Width'); axes[0].set_ylabel('Height')
    axes[0].set_title('Image size distribution')
    axes[1].hist(aspect_ratios, bins=50)
    axes[1].set_xlabel('Aspect ratio (W/H)'); axes[1].set_title('Aspect ratio distribution')
    plt.tight_layout(); plt.savefig('dataset_profile.png', dpi=120)

    return {'widths': widths, 'heights': heights, 'means': means, 'stds': stds}