---
id: wiki-2026-0508-pose-estimation
title: Pose Estimation
category: 10_Wiki/Topics
status: verified
canonical_id: self
aliases: [Human Pose Estimation, HPE, Keypoint Detection]
duplicate_of: none
source_trust_level: A
confidence_score: 0.9
verification_status: applied
tags: [computer-vision, pose-estimation, deep-learning, keypoints]
raw_sources: []
last_reinforced: 2026-05-10
github_commit: pending
tech_stack:
  language: python
  framework: pytorch, mmpose, mediapipe
---

# Pose Estimation

## 매 한 줄
> **"매 image/video에서 인체 keypoints (joints) 위치 detection."**. OpenPose (2017)가 multi-person bottom-up을 popularize, MediaPipe로 mobile real-time, 2024-2025 ViTPose / SAM-style transformer가 SOTA.

## 매 핵심

### 매 두 가지 paradigm
- **Top-down**: detect person bbox → crop → keypoint regression. 매 정확, slow with crowd.
- **Bottom-up**: keypoints first → group into persons (PAF / associative embedding). 매 fast at scale.
- **Single-stage** (modern): YOLO-Pose, ED-Pose — detection + keypoints joint.

### 매 표현 방식
- **2D keypoints**: (x, y, confidence) — COCO 17 keypoints standard.
- **3D pose**: (x, y, z) — single image lift 또는 multi-view.
- **SMPL / mesh**: full body parametric model — VIBE, HMR, 4D-Humans.

### 매 응용
1. AR/VR avatar driving (Meta Quest, Apple Vision Pro).
2. Fitness coaching (form correction).
3. Sports analytics (gait, biomechanics).
4. Animation mocap markerless.
5. Surveillance / fall detection.

## 💻 패턴

### MediaPipe (real-time, on-device)
```python
import mediapipe as mp
import cv2

mp_pose = mp.solutions.pose
pose = mp_pose.Pose(model_complexity=1, min_detection_confidence=0.5)

cap = cv2.VideoCapture(0)
while cap.isOpened():
    ok, frame = cap.read()
    if not ok: break
    results = pose.process(cv2.cvtColor(frame, cv2.COLOR_BGR2RGB))
    if results.pose_landmarks:
        for lm in results.pose_landmarks.landmark:
            print(lm.x, lm.y, lm.visibility)
```

### MMPose (research, ViTPose backbone)
```python
from mmpose.apis import MMPoseInferencer

inferencer = MMPoseInferencer(pose2d='vitpose-h')
result = next(inferencer('image.jpg', show=False))
keypoints = result['predictions'][0][0]['keypoints']  # (17, 2)
scores = result['predictions'][0][0]['keypoint_scores']
```

### YOLO-Pose (Ultralytics, single-stage)
```python
from ultralytics import YOLO

model = YOLO('yolo11n-pose.pt')
results = model('image.jpg')
for r in results:
    kpts = r.keypoints.xy  # (n_persons, 17, 2)
    conf = r.keypoints.conf
```

### 3D lift (VideoPose3D-style)
```python
import torch
# 2D (T, 17, 2) -> 3D (T, 17, 3) via temporal CNN
class TemporalLift(torch.nn.Module):
    def __init__(self, n_kpts=17, ch=1024):
        super().__init__()
        self.expand = torch.nn.Conv1d(n_kpts*2, ch, 3, padding=1)
        self.blocks = torch.nn.Sequential(*[
            torch.nn.Sequential(
                torch.nn.Conv1d(ch, ch, 3, padding=1, dilation=d),
                torch.nn.BatchNorm1d(ch), torch.nn.ReLU()
            ) for d in (3, 9, 27)
        ])
        self.head = torch.nn.Conv1d(ch, n_kpts*3, 1)

    def forward(self, x):  # x: (B, T, 17, 2)
        B, T = x.shape[:2]
        x = x.reshape(B, T, -1).transpose(1, 2)
        return self.head(self.blocks(self.expand(x))).transpose(1, 2).reshape(B, T, -1, 3)
```

### COCO keypoint metric (OKS / mAP)
```python
from pycocotools.coco import COCO
from pycocotools.cocoeval import COCOeval

gt = COCO('person_keypoints_val2017.json')
dt = gt.loadRes('predictions.json')
e = COCOeval(gt, dt, 'keypoints')
e.evaluate(); e.accumulate(); e.summarize()
# AP @ OKS=.50:.95 — 표준 metric
```

### SMPL mesh recovery (4D-Humans / HMR2)
```python
from hmr2.models import load_hmr2
model, model_cfg = load_hmr2('logs/checkpoints/epoch=35.ckpt')
out = model(image_tensor)
verts = out['pred_vertices']    # (B, 6890, 3)
betas = out['pred_smpl_params']['betas']
pose  = out['pred_smpl_params']['body_pose']
```

## 매 결정 기준
| 상황 | Approach |
|---|---|
| Mobile / web real-time | MediaPipe Pose |
| Highest accuracy single image | ViTPose-H (MMPose) |
| Multi-person crowd | YOLO-Pose / ED-Pose (single-stage) |
| 3D from monocular video | 4D-Humans / WHAM |
| Animation mocap | SMPL / SMPL-X based |
| Edge device < 10ms | MoveNet Lightning, RTMPose-tiny |

**기본값**: 2D는 RTMPose, 3D mesh는 4D-Humans.

## 🔗 Graph
- 부모: [[Computer_Vision]] · [[Deep_Learning]]
- 변형: [[MediaPipe]]
- Adjacent: [[Object_Detection]] · [[Keypoint_Detection]]

## 🤖 LLM 활용
**언제**: vision-action pipeline 의 input feature, fitness/AR app, mocap automation.
**언제 X**: facial keypoints는 face-specific model (MediaPipe Face Mesh, dlib), hand는 MediaPipe Hands.

## ❌ 안티패턴
- **Top-down without bbox tracking**: 매 frame redetect — temporal jitter 매 심각. ByteTrack 결합.
- **2D regression direct (x,y) without heatmap**: 매 lower accuracy. Heatmap supervision 매 표준.
- **3D from single 2D pose**: depth ambiguity — temporal context 또는 multi-view 필요.
- **Ignoring camera intrinsics for 3D**: 매 metric scale wrong.

## 🧪 검증 / 중복
- Verified (MMPose docs, Ultralytics YOLO11-pose, MediaPipe docs, COCO keypoint benchmark).
- 신뢰도 A.

## 🕓 Changelog
| 날짜 | 변경 |
|---|---|
| 2026-05-08 | Phase 1 |
| 2026-05-10 | Manual cleanup — pose estimation paradigms + modern stack (ViTPose, YOLO-Pose, 4D-Humans) |