2nd/10_Wiki/Topics/Computer_Vision.md

id, title, category, status, canonical_id, aliases, duplicate_of, source_trust_level, confidence_score, verification_status, tags, raw_sources, last_reinforced, github_commit, tech_stack

id	title	category	status	canonical_id	aliases	duplicate_of	source_trust_level	confidence_score	verification_status	tags	raw_sources	last_reinforced	github_commit	tech_stack
wiki-2026-0508-computer-vision	Computer Vision	10_Wiki/Topics	verified	self	CV Vision AI Visual Recognition	none	A	0.93	applied	computer-vision deep-learning multimodal perception		2026-05-10	pending	language framework Python PyTorch / Transformers / OpenCV

Computer Vision

매 한 줄

"매 CV 의 핵심: pixels → semantic understanding via learned hierarchical features". 매 1960s Roberts edge detector 로 시작, 매 2012 AlexNet 으로 deep-learning revolution, 매 2020 ViT, 매 2023 SAM. 매 2026 현재 multimodal foundation models (GPT-5V, Claude Opus 4.7, Qwen3-VL, Llama 4 Vision) 가 zero-shot 으로 detection / VQA / OCR 의 통합.

매 핵심

매 task taxonomy

• Classification: 매 image → label.
• Detection: 매 bounding boxes + classes (YOLO, DETR, RT-DETR).
• Segmentation: 매 pixel-level (semantic, instance, panoptic). SAM2 의 promptable.
• Pose / Keypoint: 매 human/object joints (RTMPose, ViTPose).
• Depth / 3D: 매 monocular depth (Depth Anything V2), NeRF, 3DGS.
• Generation: 매 diffusion (FLUX.1, SD3.5), 매 video (Sora 2, Veo 3).
• VLM (Vision-Language): 매 image+text → text (GPT-5V, Claude Opus 4.7, Qwen3-VL).

매 modern stack (2026)

• Backbone: ConvNeXt-V2, ViT-L, EVA-02, DINOv2.
• Detection: YOLOv11, RT-DETR-v2, Grounding DINO 1.5.
• Segmentation: SAM2, Mask2Former.
• Foundation: SigLIP-2, CLIP, DINOv2 (self-supervised features).
• VLM: Claude Opus 4.7 Vision, GPT-5V, InternVL3, Qwen3-VL.

매 응용

1. Autonomous driving (Waymo, Tesla FSD perception).
2. Medical imaging (MONAI, nnU-Net for segmentation).
3. Document AI / OCR (Donut, Florence-2, GPT-5V).
4. Robotics (open-vocabulary manipulation, RT-2).
5. Content moderation, retail, agriculture.

💻 패턴

Image classification (timm + finetune)

import timm, torch
model = timm.create_model("convnextv2_tiny.fcmae_ft_in22k_in1k",
                          pretrained=True, num_classes=10)
opt = torch.optim.AdamW(model.parameters(), lr=1e-4)
# train as usual

Object detection (Ultralytics YOLOv11)

from ultralytics import YOLO
model = YOLO("yolo11n.pt")
results = model("img.jpg")
for r in results:
    print(r.boxes.xyxy, r.boxes.cls, r.boxes.conf)

Promptable segmentation (SAM2)

from sam2.sam2_image_predictor import SAM2ImagePredictor
predictor = SAM2ImagePredictor.from_pretrained("facebook/sam2-hiera-large")
predictor.set_image(image)
masks, scores, _ = predictor.predict(
    point_coords=[[500, 375]], point_labels=[1])

CLIP zero-shot classification

import torch, open_clip
model, _, preprocess = open_clip.create_model_and_transforms(
    "ViT-L-14-SigLIP2", pretrained="webli")
tokenizer = open_clip.get_tokenizer("ViT-L-14-SigLIP2")
text = tokenizer(["a cat", "a dog", "a car"])
with torch.no_grad():
    img_feat = model.encode_image(preprocess(img).unsqueeze(0))
    txt_feat = model.encode_text(text)
    probs = (img_feat @ txt_feat.T).softmax(-1)

VLM via Claude (vision)

import anthropic, base64
client = anthropic.Anthropic()
img_b64 = base64.b64encode(open("img.jpg", "rb").read()).decode()
resp = client.messages.create(
    model="claude-opus-4-7-20260101",
    max_tokens=1024,
    messages=[{
        "role": "user",
        "content": [
            {"type": "image", "source": {"type": "base64",
                "media_type": "image/jpeg", "data": img_b64}},
            {"type": "text", "text": "What objects? Bounding boxes (x1,y1,x2,y2)."}
        ]
    }])

Monocular depth (Depth Anything V2)

from transformers import pipeline
pipe = pipeline("depth-estimation",
                model="depth-anything/Depth-Anything-V2-Large-hf")
depth = pipe(image)["depth"]

OpenCV preprocessing pipeline

import cv2, numpy as np
img = cv2.imread("scan.jpg")
gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
blur = cv2.GaussianBlur(gray, (5, 5), 0)
edges = cv2.Canny(blur, 50, 150)
contours, _ = cv2.findContours(edges, cv2.RETR_EXTERNAL,
                               cv2.CHAIN_APPROX_SIMPLE)

매 결정 기준

상황	Approach
Quick prototype, no labels	VLM (GPT-5V/Claude) zero-shot
Production classification	timm finetune (ConvNeXt-V2)
Real-time detection	YOLOv11 / RT-DETR
Open-vocab detection	Grounding DINO 1.5
Pixel-perfect masks	SAM2 (promptable)
Video understanding	InternVideo2 / Sora-style
Edge / mobile	MobileNetV4 + ONNX/CoreML

기본값: 매 prototype 은 VLM 로 baseline, 매 production scale 시 specialized model 의 finetune.

🔗 Graph

• 부모: Deep Learning · Multimodal AI
• 변형: CNN · Vision Transformer · Diffusion Models
• 응용: Object Detection · Image Segmentation · OCR
• Adjacent: NLP · Robotics · Generative AI

🤖 LLM 활용

언제: 매 quick image-understanding tasks (VQA, OCR, caption), 매 dataset bootstrapping (label generation), 매 vision-pipeline scaffolding. 언제 X: 매 high-throughput / low-latency production — 매 specialized model 의 use. 매 medical / safety-critical 은 validated model only.

❌ 안티패턴

• Re-inventing vs. timm/ultralytics: 매 well-tested baselines 의 무시 X.
• No domain-specific augmentation: 매 medical/satellite 의 ImageNet aug 의 그대로 사용.
• Ignoring image preprocessing: 매 wrong normalization 의 가장 흔한 bug.
• VLM 의 fine-grained 작업 의 무비판 신뢰: 매 small-object detection / counting 의 hallucination.
• No test-time augmentation for production: 매 robustness 손실.

🧪 검증 / 중복

• Verified (Szeliski Computer Vision 2nd ed., Ultralytics YOLOv11 2025, SAM2 paper 2024, Depth Anything V2 2024).
• 신뢰도 A.
• 관련: CNN, Vision Transformer.

🕓 Changelog

날짜	변경
2026-05-08	Phase 1
2026-05-10	Manual cleanup — 2026 modern stack (YOLOv11, SAM2, VLMs)