Koushim/detr-pose-coco50

📌 DETR + Keypoint Estimation (COCO Subset)

Author: @Koushik

---

🧠 Model Overview

This project combines:

• 🤖 facebook/detr-resnet-50 (object detector)
• 🧱 Custom PyTorch keypoint head
• 📊 Trained on 500-person subset of COCO 2017 Keypoints

The system detects people using DETR, then predicts 17 COCO-style keypoints (top-down) using heatmap regression.

---

📂 Files Included

File	Description
pytorch_model.bin	Trained PyTorch model weights
05_detr_pose_coco_colab.ipynb	Full Colab notebook (training + inference)
config.json	Basic model metadata
README.md	Project description

---

📚 Dataset

• Subset: 500 images from COCO val2017 with visible persons
• Annotations: 17 keypoints per person
• Source: COCO Keypoints

---

🏗️ Architecture

[ Input Image ]
      │
      ▼
[ DETR (Person BBox) ]
      │
      ▼
[ Crop + Resize (256×256) ]
      │
      ▼
[ CNN Keypoint Head ]
      │
      ▼
[ 17 Heatmaps (Keypoints) ]

---

🚀 Quick Start

import torch
from model import KeypointHead

model = KeypointHead()
model.load_state_dict(torch.load('pytorch_model.bin'))
model.eval()

---

🧪 Inference Demo

from PIL import Image
import cv2, numpy as np
from transformers import DetrImageProcessor, DetrForObjectDetection

img = Image.open('sample.jpg')
processor = DetrImageProcessor.from_pretrained("facebook/detr-resnet-50")
detector = DetrForObjectDetection.from_pretrained("facebook/detr-resnet-50")

inputs = processor(images=img, return_tensors="pt")
outputs = detector(**inputs)
results = processor.post_process_object_detection(outputs, target_sizes=[img.size[::-1]], threshold=0.8)[0]

# Use results['boxes'][0] to crop person
# Feed crop into model(img) to get 17 heatmaps

---

🧠 Training (optional)

To fine-tune on your own dataset:

• Convert your data to COCO format
• Use the notebook provided (05_detr_pose_coco_colab.ipynb)
• Change paths and re-train

---

✨ Credit

• Hugging Face Transformers
• COCO Dataset
• facebook/detr