computer-vision-guide

Computer Vision Guide

A skill for conducting computer vision research, covering model architectures, dataset preparation, training pipelines, evaluation metrics, and common experimental protocols for image classification, object detection, and segmentation tasks.

Core Tasks and Architectures

Computer Vision Task Taxonomy

Image Classification:
  Input: Single image
  Output: Class label(s)
  Models: ResNet, EfficientNet, ViT, ConvNeXt

Object Detection:
  Input: Single image
  Output: Bounding boxes + class labels
  Models: YOLO (v5-v9), Faster R-CNN, DETR, RT-DETR

Semantic Segmentation:
  Input: Single image
  Output: Per-pixel class label
  Models: U-Net, DeepLab, SegFormer, Mask2Former

Instance Segmentation:
  Input: Single image
  Output: Per-pixel labels distinguishing individual objects
  Models: Mask R-CNN, Mask2Former, SAM

Image Generation:
  Input: Text prompt or noise
  Output: Generated image
  Models: Stable Diffusion, DALL-E, Imagen

Model Architecture Evolution

CNNs (Convolutional Neural Networks):
  LeNet (1998) -> AlexNet (2012) -> VGG (2014) -> ResNet (2015)
  -> EfficientNet (2019) -> ConvNeXt (2022)

Vision Transformers:
  ViT (2020) -> DeiT (2021) -> Swin Transformer (2021)
  -> BEiT (2021) -> DINOv2 (2023)

Trend: Transformers are competitive with CNNs at scale.
Hybrid architectures combining convolutions and attention are common.

Dataset Preparation

Building a Research Dataset

import os
from pathlib import Path


def organize_image_dataset(source_dir: str,
                            split_ratios: dict = None) -> dict:
    """
    Organize images into train/val/test splits.

    Args:
        source_dir: Directory containing class subdirectories
        split_ratios: Dict with 'train', 'val', 'test' ratios
    """
    if split_ratios is None:
        split_ratios = {"train": 0.7, "val": 0.15, "test": 0.15}

    import random
    random.seed(42)

    stats = {}
    for class_dir in sorted(Path(source_dir).iterdir()):
        if not class_dir.is_dir():
            continue

        images = list(class_dir.glob("*.jpg")) + list(class_dir.glob("*.png"))
        random.shuffle(images)

        n = len(images)
        n_train = int(n * split_ratios["train"])
        n_val = int(n * split_ratios["val"])

        stats[class_dir.name] = {
            "total": n,
            "train": n_train,
            "val": n_val,
            "test": n - n_train - n_val
        }

    return stats