🔍 ghostfinder

Lightweight Visual Re-Identification & Template Tracker

CPU-friendly Re-ID using color histograms and template matching — fast enough for edge devices.

---

✨ Why ghostfinder?

Feature	ghostfinder	torchreid	deep-person-reid
CPU-only (no GPU needed)	✅	❌	❌
Edge device ready (RPi, Jetson)	✅	❌	❌
< 0.5ms per comparison	✅	❌	❌
Template matching fallback	✅	❌	❌
Multi-feature scoring	✅	✅	✅
ROI-constrained search	✅	❌	❌
No training data needed	✅	❌	❌
Lightweight (~300 lines total)	✅	❌	❌

Deep learning Re-ID solutions are powerful but require GPUs and large models. ghostfinder is designed for real-time systems on resource-constrained hardware — where you need to re-identify a target in under 1ms using only CPU.

---

📦 Installation

pip install ghostfinder

Or install from source:

git clone https://github.com/ByIbos/ghostfinder.git
cd ghostfinder
pip install -e .

---

🚀 Quick Start

Visual Re-Identification

import cv2
from ghostfinder import TargetReID

reid = TargetReID(similarity_threshold=0.55)

# While tracking the target:
reid.update_fingerprint(frame, (x1, y1, x2, y2))

# When target is lost and new detections appear:
best_id, score = reid.find_best_match(frame, all_boxes, all_ids)
if best_id is not None:
    print(f"Target re-identified as ID {best_id} (score: {score:.2f})")

Template Matching Fallback

from ghostfinder import TemplateTracker

tracker = TemplateTracker(match_threshold=0.45, search_margin=150)

# While detector is working:
tracker.update_template(frame, (x1, y1, x2, y2))

# When YOLO/detector fails to find the target:
result = tracker.search(frame, predicted_center=(320, 240))
if result:
    center = result['center']
    score = result['score']
    print(f"Template found at {center} (confidence: {score:.0%})")

Combined Pipeline (Re-ID + Template)

from ghostfinder import TargetReID, TemplateTracker

reid = TargetReID()
template = TemplateTracker()

def handle_target_lost(frame, boxes, ids, kalman_prediction):
    """
    3-layer recovery strategy:
    1. Try Re-ID (color fingerprint matching)
    2. Try Template Matching (visual search)
    3. Use Kalman prediction (coast on momentum)
    """
    # Layer 1: Re-ID
    match_id, score = reid.find_best_match(frame, boxes, ids)
    if match_id is not None:
        return ("REID", match_id, score)

    # Layer 2: Template
    result = template.search(frame, predicted_center=kalman_prediction)
    if result:
        return ("TEMPLATE", result['center'], result['score'])

    # Layer 3: Kalman (handled externally)
    return ("PREDICT", kalman_prediction, 0.0)

---

🔧 API Reference

TargetReID(similarity_threshold=0.55)

A visual fingerprinting system that stores the target's color histogram, aspect ratio, and area for comparison against new detections.

Methods

Method	Returns	Description
update_fingerprint(frame, box)	None	Store/update fingerprint from current crop
compare(frame, box)	float	Compare one box against fingerprint (0-1)
find_best_match(frame, boxes, ids)	(id, score)	Find best matching detection
reset()	None	Clear all stored data

Scoring Breakdown

Component	Weight	Metric
Color similarity	60%	HSV histogram Bhattacharyya distance
Shape similarity	20%	Aspect ratio difference
Size similarity	20%	Area ratio (min/max)

---

TemplateTracker(match_threshold=0.45, search_margin=150)

A fallback tracker that stores the last known visual appearance and searches for it using cv2.matchTemplate within a constrained ROI.

Methods

Method	Returns	Description
update_template(frame, box)	None	Store current target appearance
search(frame, predicted_center)	dict or None	Search for template in frame
reset()	None	Clear all stored data

Search Result Dictionary

{
    'center': (cx, cy),                    # Match center point
    'box': (x1, y1, x2, y2),              # Bounding box
    'score': 0.72,                         # Correlation score (0-1)
    'search_region': (sx1, sy1, sx2, sy2)  # ROI that was searched
}

---

📐 How It Works

┌──────────────────────────────────────────────────┐
│                ghostfinder Pipeline                    │
│                                                  │
│  ┌─── TRACKING PHASE ─────────────────────────┐  │
│  │  Target visible → update fingerprint        │  │
│  │    • HSV histogram (30×32 bins)             │  │
│  │    • Aspect ratio (w/h)                     │  │
│  │    • Area (w×h)                             │  │
│  │    • Template crop (grayscale)              │  │
│  └─────────────────────────────────────────────┘  │
│                      │                            │
│                  Target Lost!                     │
│                      │                            │
│  ┌─── RECOVERY PHASE ─────────────────────────┐  │
│  │                                             │  │
│  │  Layer 1: Re-ID (TargetReID)                │  │
│  │    Compare all detections vs fingerprint     │  │
│  │    Score = 0.6×color + 0.2×shape + 0.2×size │  │
│  │    Match if score > threshold (0.55)         │  │
│  │                                             │  │
│  │  Layer 2: Template (TemplateTracker)         │  │
│  │    Search ROI = predicted_center ± margin    │  │
│  │    cv2.matchTemplate (TM_CCOEFF_NORMED)     │  │
│  │    Match if correlation > threshold (0.45)   │  │
│  │                                             │  │
│  └─────────────────────────────────────────────┘  │
└──────────────────────────────────────────────────┘

---

⚡ Performance

Benchmarked on a Raspberry Pi 4B (no GPU):

Operation	Time	Notes
update_fingerprint()	~0.3ms	Per frame
compare() (single box)	~0.4ms	Per candidate
find_best_match() (10 boxes)	~3ms	Worst case
search() (template)	~1-3ms	Depends on ROI size

---

🎯 Use Cases

• 🚁 Drone Tracking — Re-acquire target after occlusion
• 📹 Surveillance — Track individuals across camera blind spots
• 🤖 Robot Vision — Persistent object following
• 🏭 Industrial — Track parts on conveyor belts
• 🎯 Sports Analytics — Player tracking through crowds

---

📜 License

MIT License — use it anywhere, commercially or personally.

---

Built with ❤️ by ByIbos — extracted from the Auto-ReID Drone Tracker project.