Computer Vision

Computer vision enables machines to interpret and reason about visual data - images, video, and multi-modal inputs. Modern CV pipelines are built on deep neural networks pretrained on large datasets (ImageNet, COCO, ADE20K) and fine-tuned for specific domains. PyTorch and its ecosystem (torchvision, timm, ultralytics, albumentations) cover the full stack from data loading through deployment. Foundation models like SAM, DINOv2, and OpenCLIP have shifted best practice toward prompt-based and zero-shot approaches before committing to full training runs.

---

When to use this skill

Trigger this skill when the user:

• Trains or fine-tunes an image classifier on a custom dataset
• Runs inference with YOLO, DETR, or other detection models
• Builds a semantic or instance segmentation pipeline
• Implements data augmentation for CV training
• Preprocesses images for model ingestion (resize, normalize, batch)
• Exports a vision model to ONNX or optimizes with TensorRT
• Evaluates a vision model (mAP, confusion matrix, per-class metrics)
• Implements a U-Net, DeepLabV3, or similar segmentation architecture

Do NOT trigger this skill for:

• Pure NLP tasks with no visual component (use a language-model skill instead)
• 3D point-cloud processing or LiDAR-only pipelines (overlap is limited; check domain)

---

Key principles

1. Start with pretrained models - Fine-tune ImageNet/COCO weights before training from scratch. Even a frozen backbone with a new head beats random init on small datasets.
2. Augment data aggressively - Real-world distribution shifts are unavoidable. Use albumentations with geometric, color, and noise transforms. Target-aware augments (mosaic, copy-paste) matter especially for detection.
3. Validate on representative data - Always hold out data from the exact deployment distribution. Benchmark on in-distribution AND out-of-distribution splits separately.
4. Optimize inference separately from training - Training precision (FP32/AMP) and inference precision (INT8/FP16) have different tradeoffs. Profile, export to ONNX, then apply TensorRT or OpenVINO post-training quantization.
5. Monitor for distribution shift - Production images drift from training data (lighting changes, new object classes, compression artifacts). Log prediction confidence distributions and trigger retraining pipelines when they degrade.

---

Core concepts

Task taxonomy

Task	Output	Typical metric
Classification	Single label per image	Top-1 / Top-5 accuracy
Detection	Bounding boxes + labels	mAP@0.5, mAP@0.5:0.95
Semantic segmentation	Per-pixel class mask	mIoU
Instance segmentation	Per-object mask + label	mask AP
Generation / synthesis	New images	FID, LPIPS

Backbone architectures

Backbone	Strengths	Typical use
ResNet-50/101	Stable, well-understood	Classification baseline, feature extractor
EfficientNet-B0..B7	Accuracy/FLOP Pareto front	Mobile + server classification
ViT-B/16, ViT-L/16	Strong with large data, attention maps	High-accuracy classification, zero-shot
ConvNeXt-T/B	CNN with transformer-like training recipe	Drop-in ResNet replacement
DINOv2 (ViT)	Strong self-supervised features	Few-shot, feature extraction

Anchor-free vs anchor-based detection

• Anchor-based (YOLOv5, Faster R-CNN) - predefined box aspect ratios per grid cell. Fast training convergence, tuning required for unusual object scales.
• Anchor-free (YOLO11/v8, FCOS, DETR) - predict box center + offsets directly. Cleaner training, no anchor hyperparameter search, now the default for new projects.

Loss functions

Loss	Used for
Cross-entropy	Classification (multi-class), segmentation pixel-wise
Focal loss	Detection classification head - down-weights easy negatives
IoU / GIoU / CIoU / DIoU	Bounding box regression
Dice loss	Segmentation - handles class imbalance better than cross-entropy
Binary cross-entropy	Multi-label classification, mask prediction

---

Common tasks

Fine-tune an image classifier

import torch
import torch.nn as nn
from torch.utils.data import DataLoader
from torchvision import datasets, transforms, models

# 1. Data transforms
train_tf = transforms.Compose([
    transforms.RandomResizedCrop(224),
    transforms.RandomHorizontalFlip(),
    transforms.ColorJitter(brightness=0.3, contrast=0.3, saturation=0.2),
    transforms.ToTensor(),
    transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
])
val_tf = transforms.Compose([
    transforms.Resize(256),
    transforms.CenterCrop(224),
    transforms.ToTensor(),
    transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
])

train_ds = datasets.ImageFolder("data/train", transform=train_tf)
val_ds   = datasets.ImageFolder("data/val",   transform=val_tf)
train_loader = DataLoader(train_ds, batch_size=32, shuffle=True,  num_workers=4)
val_loader   = DataLoader(val_ds,   batch_size=64, shuffle=False, num_workers=4)

# 2. Load pretrained backbone, replace head
NUM_CLASSES = len(train_ds.classes)
model = models.efficientnet_b0(weights=models.EfficientNet_B0_Weights.DEFAULT)
model.classifier[1] = nn.Linear(model.classifier[1].in_features, NUM_CLASSES)

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model = model.to(device)

# 3. Two-phase training: head first, then unfreeze backbone
optimizer = torch.optim.AdamW(model.classifier.parameters(), lr=1e-3)
scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(optimizer, T_max=5)
criterion = nn.CrossEntropyLoss(label_smoothing=0.1)

def train_one_epoch(loader):
    model.train()
    for imgs, labels in loader:
        imgs, labels = imgs.to(device), labels.to(device)
        optimizer.zero_grad()
        loss = criterion(model(imgs), labels)
        loss.backward()
        optimizer.step()
    scheduler.step()

# Phase 1 - head only (5 epochs)
for epoch in range(5):
    train_one_epoch(train_loader)

# Phase 2 - unfreeze everything with lower LR
for p in model.parameters():
    p.requires_grad = True
optimizer = torch.optim.AdamW(model.parameters(), lr=1e-4, weight_decay=0.01)
scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(optimizer, T_max=10)
for epoch in range(10):
    train_one_epoch(train_loader)

torch.save(model.state_dict(), "classifier.pth")

Run object detection with YOLO

from ultralytics import YOLO

# --- Inference ---
model = YOLO("yolo11n.pt")  # nano; swap for yolo11s/m/l/x for accuracy
results = model.predict("image.jpg", conf=0.25, iou=0.45, device=0)

for r in results:
    for box in r.boxes:
        cls   = int(box.cls[0])
        label = model.names[cls]
        conf  = float(box.conf[0])
        xyxy  = box.xyxy[0].tolist()   # [x1, y1, x2, y2]
        print(f"{label}: {conf:.2f}  {xyxy}")

# --- Fine-tune on custom dataset ---
# Expects data.yaml with train/val paths and class names
model = YOLO("yolo11s.pt")
results = model.train(
    data="data.yaml",
    epochs=100,
    imgsz=640,
    batch=16,
    device=0,
    optimizer="AdamW",
    lr0=1e-3,
    weight_decay=0.0005,
    augment=True,         # built-in mosaic, mixup, copy-paste
    cos_lr=True,
    patience=20,          # early stopping
    project="runs/detect",
    name="custom_v1",
)
print(results.results_dict)  # mAP50, mAP50-95, precision, recall

Implement a data augmentation pipeline

import albumentations as A
from albumentations.pytorch import ToTensorV2
import numpy as np

# Classification pipeline
clf_transform = A.Compose([
    A.RandomResizedCrop(height=224, width=224, scale=(0.6, 1.0)),
    A.HorizontalFlip(p=0.5),
    A.ShiftScaleRotate(shift_limit=0.05, scale_limit=0.1, rotate_limit=15, p=0.5),
    A.OneOf([
        A.GaussNoise(var_limit=(10, 50)),
        A.GaussianBlur(blur_limit=3),
        A.MotionBlur(blur_limit=3),
    ], p=0.3),
    A.ColorJitter(brightness=0.3, contrast=0.3, saturation=0.2, hue=0.05, p=0.5),
    A.Normalize(mean=(0.485, 0.456, 0.406), std=(0.229, 0.224, 0.225)),
    ToTensorV2(),
])

# Detection pipeline - bbox-aware transforms
det_transform = A.Compose([
    A.RandomResizedCrop(height=640, width=640, scale=(0.5, 1.0)),
    A.HorizontalFlip(p=0.5),
    A.RandomBrightnessContrast(p=0.4),
    A.HueSaturationValue(p=0.3),
    A.Normalize(mean=(0.485, 0.456, 0.406), std=(0.229, 0.224, 0.225)),
    ToTensorV2(),
], bbox_params=A.BboxParams(format="yolo", label_fields=["class_labels"]))

# Usage
image = np.random.randint(0, 255, (480, 640, 3), dtype=np.uint8)
out = clf_transform(image=image)["image"]  # torch.Tensor [3, 224, 224]

Build an image preprocessing pipeline

import torch
from torchvision.transforms import v2 as T
from PIL import Image

# Production preprocessing - deterministic, no augmentation
preprocess = T.Compose([
    T.Resize((256, 256), interpolation=T.InterpolationMode.BILINEAR, antialias=True),
    T.CenterCrop(224),
    T.ToImage(),
    T.ToDtype(torch.float32, scale=True),
    T.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
])

def load_batch(paths: list[str], device: torch.device) -> torch.Tensor:
    """Load, preprocess, and batch a list of image paths."""
    tensors = []
    for p in paths:
        img = Image.open(p).convert("RGB")
        tensors.append(preprocess(img))
    return torch.stack(tensors).to(device)

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
batch = load_batch(["a.jpg", "b.jpg", "c.jpg"], device)
print(batch.shape)  # [3, 3, 224, 224]

Deploy a vision model

import torch
import torch.onnx
import onnxruntime as ort
import numpy as np

# --- Export to ONNX ---
model = torch.load("classifier.pth", map_location="cpu")
model.eval()

dummy = torch.randn(1, 3, 224, 224)
torch.onnx.export(
    model,
    dummy,
    "classifier.onnx",
    input_names=["image"],
    output_names=["logits"],
    dynamic_axes={"image": {0: "batch"}, "logits": {0: "batch"}},
    opset_version=17,
)

# --- ONNX Runtime inference (CPU or CUDA EP) ---
providers = ["CUDAExecutionProvider", "CPUExecutionProvider"]
session = ort.InferenceSession("classifier.onnx", providers=providers)
input_name = session.get_inputs()[0].name

def infer_onnx(batch_np: np.ndarray) -> np.ndarray:
    return session.run(None, {input_name: batch_np})[0]

# --- TensorRT optimization (requires tensorrt package) ---
# Run once offline to build the engine:
#   trtexec --onnx=classifier.onnx --saveEngine=classifier.trt \
#           --fp16 --minShapes=image:1x3x224x224 \
#           --optShapes=image:8x3x224x224 \
#           --maxShapes=image:32x3x224x224

Evaluate model performance

import torch
import numpy as np
from torchmetrics.classification import (
    MulticlassAccuracy,
    MulticlassConfusionMatrix,
    MulticlassPrecision,
    MulticlassRecall,
    MulticlassF1Score,
)
from torchmetrics.detection import MeanAveragePrecision

# --- Classification metrics ---
def evaluate_classifier(model, loader, num_classes, device):
    model.eval()
    metrics = {
        "acc":  MulticlassAccuracy(num_classes=num_classes, top_k=1).to(device),
        "prec": MulticlassPrecision(num_classes=num_classes, average="macro").to(device),
        "rec":  MulticlassRecall(num_classes=num_classes, average="macro").to(device),
        "f1":   MulticlassF1Score(num_classes=num_classes, average="macro").to(device),
        "cm":   MulticlassConfusionMatrix(num_classes=num_classes).to(device),
    }
    with torch.no_grad():
        for imgs, labels in loader:
            imgs, labels = imgs.to(device), labels.to(device)
            preds = model(imgs)
            for m in metrics.values():
                m.update(preds, labels)
    return {k: v.compute() for k, v in metrics.items()}

# --- Detection metrics (COCO mAP) ---
map_metric = MeanAveragePrecision(iou_type="bbox")
# preds and targets follow torchmetrics dict format
preds = [{"boxes": torch.tensor([[10, 20, 100, 200]]), "scores": torch.tensor([0.9]), "labels": torch.tensor([0])}]
tgts  = [{"boxes": torch.tensor([[12, 22, 102, 202]]), "labels": torch.tensor([0])}]
map_metric.update(preds, tgts)
result = map_metric.compute()
print(f"mAP@0.5: {result['map_50']:.4f}  mAP@0.5:0.95: {result['map']:.4f}")

Implement semantic segmentation

import torch
import torch.nn as nn
from torchvision.models.segmentation import deeplabv3_resnet50, DeepLabV3_ResNet50_Weights

# --- DeepLabV3 fine-tuning ---
NUM_CLASSES = 21  # e.g. PASCAL VOC
model = deeplabv3_resnet50(weights=DeepLabV3_ResNet50_Weights.DEFAULT)
model.classifier[4] = nn.Conv2d(256, NUM_CLASSES, kernel_size=1)
model.aux_classifier[4] = nn.Conv2d(256, NUM_CLASSES, kernel_size=1)

# Training step
def seg_train_step(model, imgs, masks, optimizer, device):
    model.train()
    imgs, masks = imgs.to(device), masks.long().to(device)
    out = model(imgs)
    # main loss + auxiliary loss
    loss = nn.functional.cross_entropy(out["out"], masks)
    loss += 0.4 * nn.functional.cross_entropy(out["aux"], masks)
    optimizer.zero_grad()
    loss.backward()
    optimizer.step()
    return loss.item()

# Inference - returns per-pixel class index
def seg_predict(model, img_tensor, device):
    model.eval()
    with torch.no_grad():
        out = model(img_tensor.unsqueeze(0).to(device))
    return out["out"].argmax(dim=1).squeeze(0).cpu()  # [H, W]

# --- Lightweight U-Net-style architecture (custom) ---
class DoubleConv(nn.Module):
    def __init__(self, in_ch, out_ch):
        super().__init__()
        self.net = nn.Sequential(
            nn.Conv2d(in_ch, out_ch, 3, padding=1, bias=False),
            nn.BatchNorm2d(out_ch), nn.ReLU(inplace=True),
            nn.Conv2d(out_ch, out_ch, 3, padding=1, bias=False),
            nn.BatchNorm2d(out_ch), nn.ReLU(inplace=True),
        )
    def forward(self, x): return self.net(x)

class UNet(nn.Module):
    def __init__(self, in_channels=3, num_classes=2, features=(64, 128, 256, 512)):
        super().__init__()
        self.downs = nn.ModuleList()
        self.ups   = nn.ModuleList()
        self.pool  = nn.MaxPool2d(2, 2)
        ch = in_channels
        for f in features:
            self.downs.append(DoubleConv(ch, f)); ch = f
        self.bottleneck = DoubleConv(features[-1], features[-1] * 2)
        for f in reversed(features):
            self.ups.append(nn.ConvTranspose2d(f * 2, f, 2, 2))
            self.ups.append(DoubleConv(f * 2, f))
        self.head = nn.Conv2d(features[0], num_classes, 1)

    def forward(self, x):
        skips = []
        for down in self.downs:
            x = down(x); skips.append(x); x = self.pool(x)
        x = self.bottleneck(x)
        for i in range(0, len(self.ups), 2):
            x = self.ups[i](x)
            skip = skips[-(i // 2 + 1)]
            if x.shape != skip.shape:
                x = torch.nn.functional.interpolate(x, size=skip.shape[2:])
            x = self.ups[i + 1](torch.cat([skip, x], dim=1))
        return self.head(x)

---

Anti-patterns / common mistakes

Anti-pattern	What goes wrong	Correct approach
Training from scratch on small datasets	Model memorizes noise, poor generalization	Always start from pretrained weights; freeze backbone initially
Normalizing with wrong mean/std	Silent accuracy drop when ImageNet stats misapplied to non-ImageNet data	Compute dataset statistics or use the exact stats that match the pretrained model
Leaking augmentation into validation	Inflated validation metrics; surprises in production	Apply only deterministic transforms (resize, normalize) to val/test splits
Skipping anchor/stride tuning for custom scale objects	Model misses very small or very large objects	Analyse object scale distribution; adjust anchor sizes or use anchor-free models
Exporting to ONNX without dynamic axes	Batch-size-1 locked model; crashes on larger batches in production	Always set dynamic_axes for batch dimension (and optionally spatial dims)
Evaluating detection with IoU threshold 0.5 only	Misses regression quality; mAP@0.5:0.95 is 2-3x harder	Report both mAP@0.5 and mAP@0.5:0.95 to COCO convention

---

Gotchas

1. Normalizing with wrong mean/std silently degrades accuracy - If you pretrain with ImageNet weights but normalize with different mean/std at inference, predictions silently degrade. The values [0.485, 0.456, 0.406] / [0.229, 0.224, 0.225] are ImageNet-specific; compute your own stats if your data is not RGB photos (e.g., medical images, satellite imagery).

2. loading="lazy" on the LCP image - This applies to CV deployment: never lazy-load the first above-fold image in a web app. Use fetchpriority="high" on the primary visual.

3. IV/nonce reuse destroys GCM security - This applies when encrypting model weights or inference results: reusing an IV with the same AES-256-GCM key is catastrophic. Generate fresh randomBytes(12) for every encrypt call.

4. Augmentation leaking into validation - Applying RandomResizedCrop or ColorJitter to the validation split inflates metrics. Only deterministic transforms (resize, center crop, normalize) belong in the val/test transforms.

5. ONNX export without dynamic axes locks batch size - Exporting with a fixed batch size of 1 causes runtime crashes in production when the batch size changes. Always set dynamic_axes={"image": {0: "batch"}} during export.

6. Anchor tuning for unusual object scales - If your objects are very small (satellite imagery, cell microscopy) or very large relative to the image, default YOLO anchor sizes will miss them. Run model.analyze_anchor_fitness() or use anchor-free models for unusual scale distributions.

---

References

For detailed content on model selection and architecture comparisons, read:

• references/model-zoo.md - backbone and detector architecture comparison, pretrained weight sources, speed/accuracy tradeoffs, hardware considerations

Key external resources:

• PyTorch Vision docs
• Ultralytics YOLO docs
• Albumentations docs
• timm model zoo
• Papers With Code - CV benchmarks