README.md

metadata

license: apache-2.0
language:
  - en
tags:
  - pose-estimation
  - human-pose-estimation
  - safetensors
  - keypoint-detection
  - computer-vision
  - pytorch
  - efficientnet
  - coco
  - mmpose
datasets:
  - coco
  - mpii
  - crowdpose
  - ochuman
metrics:
  - ap
  - ar
library_name: pytorch
pipeline_tag: keypoint-detection

EfficientNet-B5 Pose Estimation

A 2D human pose estimation model trained at DeKUT-DSAIL using the MMPose framework. Predicts 17 COCO keypoints from a single cropped person image.

Property	Value
Backbone	EfficientNet-B5
Attention Neck	None
Parameters	~40 M
Input Size	192 × 256
Output	Heatmaps (17, 64, 48)

---

Evaluation Results

Evaluated on COCO 2017 val using OKS-based metrics (top-down, GT bounding boxes).

Metric	Score
COCO AP	0.713
COCO AR	0.748

---

Repository Files

model.safetensors      # Model weights (safetensors format)
model.py               # Self-contained PoseEstimator inference helper
requirements.txt       # Python dependencies
pose.jpg               # Example test image
README.md              # This model card

---

Quick Start

Step 1 — Clone the repository

git clone https://huggingface.co/DeKUT-DSAIL/efficientnet_b5_coco_256x192
cd efficientnet_b5_coco_256x192

Step 2 — Create a virtual environment

Linux / macOS

python -m venv venv
source venv/bin/activate

Windows (Command Prompt)

python -m venv venv
venv\Scripts\activate.bat

Windows (PowerShell)

python -m venv venv
venv\Scripts\Activate.ps1

Step 3 — Install dependencies

pip install torch torchvision --index-url https://download.pytorch.org/whl/cpu
pip install -r requirements.txt

GPU users: Replace the PyTorch URL with your CUDA version. See pytorch.org/get-started.

Step 4 — Run inference

import cv2
from model import PoseEstimator

estimator = PoseEstimator("DeKUT-DSAIL/efficientnet_b5_coco_256x192")

image = cv2.imread("pose.jpg")
keypoints, scores = estimator.predict(image)

print("Keypoints shape:", keypoints.shape)  # (N, 17, 2)
print("Scores shape:   ", scores.shape)     # (N, 17, 1)

annotated = estimator.visualize(image, keypoints, scores, score_threshold=0.3)
cv2.imwrite("output.jpg", annotated)
print("Saved output.jpg")

---

Input / Output Specification

Property	Value
Input size	(1, 3, 256, 192) — RGB, channel-first
Normalisation	Mean [0.485, 0.456, 0.406] / Std [0.229, 0.224, 0.225]
Output	Heatmaps (N, 17, 64, 48)
Keypoints	COCO 17-joint format

COCO 17 Keypoints

Index	Name	Index	Name
0	nose	9	left_wrist
1	left_eye	10	right_wrist
2	right_eye	11	left_hip
3	left_ear	12	right_hip
4	right_ear	13	left_knee
5	left_shoulder	14	right_knee
6	right_shoulder	15	left_ankle
7	left_elbow	16	right_ankle
8	right_elbow

---

Training Details

Trained using MMPose on the following datasets:

Dataset	Link
COCO 2017	cocodataset.org
MPII Human Pose	mpii.is.tue.mpg.de
CrowdPose	GitHub
OCHuman	GitHub

Parameter	Value
Optimizer	AdamW
Learning rate	1 × 10⁻³
LR schedule	Multi-step decay
Batch size	64
Epochs	210
Input size	192 × 256
Loss	MSE on heatmaps + knowledge distillation loss

---

Architecture

Input Image (3, 256, 192)
        │
        ▼
  EfficientNet-B5 Backbone
        │
        │
        ▼
  HeatmapHead (3× deconv + 1×1 conv)
        │
        ▼
  Output Heatmaps (17, 64, 48)

---

Developed by

DeKUT-DSAIL — Dedan Kimathi University of Technology

• Framework: PyTorch / MMPose
• Model type: 2D Human Pose Estimation
• Task: Keypoint Detection
• License: Apache 2.0