Adaptive YOLO: Dynamic Model Selection for Single Object Tracking

Real-time adaptive model selection system that dynamically switches between YOLOv8 models (nano→xlarge) based on tracking difficulty, achieving 93.5% tracking success with 62% computational savings.

📊 Visual Results

Comprehensive analysis showing model switching patterns, confidence evolution, and efficiency metrics

🎬 Video Demonstrations

MOT17-04: 398 Frames Tracked (High Confidence Object)

MOT17-04_adaptive_tracking.webm

• 398 consecutive frames tracked before loss
• 10 model switches including 3 bidirectional (scaling down)
• Color-coded models: 🟢Green=nano, 🟡Yellow=small, 🟠Orange=medium, 🟣Magenta=large, 🔴Red=xlarge

MOT17-02: Initial Demo (Medium Confidence Object)

adaptive_tracking_demo.webm

• 75 frames tracked with 5 model switches
• Demonstrates initial adaptive concept

🚀 Key Results

Metric	Performance
Tracking Success Rate	93.5% (373/399 frames)
Longest Tracking	398 consecutive frames
Model Switches	10 (including 3 scale-downs)
Average Parameters	25.9M (vs 68.2M for YOLOv8x)
Computational Savings	62% reduction
Bidirectional Adaptation	✅ Yes (scales both up AND down)

Project Structure

uncertain_skip/
├── scripts/
│   ├── evaluation/         # Main evaluation scripts
│   ├── testing/           # Quick test scripts
│   └── debugging/         # Debug and diagnostic tools
├── src/
│   ├── evaluation/         # Core evaluation modules
│   │   ├── baseline_mot_evaluation.py
│   │   └── scene_complexity.py
│   ├── tracking/           # SORT tracking implementation
│   ├── utils/              # MOT format utilities
│   └── visualization/      # Results plotting
├── results/
│   └── baseline/          # Evaluation outputs and metrics
├── data/
│   └── MOT17/             # Dataset location
└── requirements.txt       # Dependencies

Features

• Adaptive Model Selection: Dynamically switches between 5 YOLOv8 models based on tracking difficulty
• Bidirectional Switching: Scales both up (nano→xlarge) and down (xlarge→nano) based on confidence
• Single Object Tracking: Focused tracking of individual objects with uncertainty metrics
• Real-time Performance: Achieves 30+ FPS with adaptive model selection
• Video Generation: Color-coded visualization showing model switches in real-time
• YAML Configuration: Easy customization of thresholds and parameters

Quick Start

1. Install Dependencies

pip install -r requirements.txt

2. Dataset Setup

# MOT17 dataset already available at: data/MOT17/

3. Run Adaptive Tracking Demo

python run_adaptive_demo.py

4. Customize Configuration (Optional) Edit run_adaptive_demo.py to change:

• Dataset/sequence (MOT17-02, MOT17-04, etc.)
• Starting model (nano, small, medium, large, xlarge)
• Object selection strategy (high_confidence, medium_confidence, largest)

Adaptive Tracking Results (MOT17-04, 398 frames)

Bidirectional Model Switching Pattern

Frame 3:   nano → small    ↑ (confidence: 0.823)
Frame 28:  small → nano    ↓ (SCALE DOWN, confidence: 0.859)
Frame 50:  small → nano    ↓ (SCALE DOWN, confidence: 0.859)
Frame 70:  small → nano    ↓ (SCALE DOWN, confidence: 0.869)
Frame 222: small → medium  ↑ (confidence: 0.675)
Frame 307: medium → large  ↑ (confidence: 0.476)
Frame 317: large → xlarge  ↑ (confidence: 0.259)

Model Usage Statistics

Model	Frames Used	Percentage	When Used
YOLOv8n	46	11.5%	High confidence (>0.85)
YOLOv8s	176	44.1%	Good confidence (0.70-0.85)
YOLOv8m	85	21.3%	Medium confidence (0.50-0.70)
YOLOv8l	10	2.5%	Low confidence (0.35-0.50)
YOLOv8x	82	20.6%	Very low confidence (<0.35)

Key Components

1. Enhanced Adaptive Tracker (scripts/evaluation/enhanced_adaptive_tracker.py)

• Bidirectional model switching logic
• Uncertainty-based adaptation
• Video generation with color-coded models
• Real-time confidence tracking

2. Scene Complexity Analyzer (src/evaluation/scene_complexity.py)

• Calculates tracking difficulty metrics
• Determines optimal model based on scene
• Implements hysteresis to prevent oscillation

3. Configuration System (configs/adaptive_tracking_config.yaml)

• Easy customization of thresholds
• Model selection parameters
• Video generation settings
• Dataset configuration

How It Works

1. Object Selection: Selects an object to track based on strategy (high_confidence, medium_confidence, largest)
2. Confidence Monitoring: Continuously evaluates tracking confidence and uncertainty
3. Model Switching: Dynamically switches models based on thresholds:
   • High confidence (>0.85) → Use smaller model (nano/small)
   • Medium confidence (0.50-0.85) → Use medium models
   • Low confidence (<0.35) → Use larger models (large/xlarge)
4. Hysteresis: Prevents rapid switching with cooldown periods
5. Video Output: Generates color-coded visualization of the tracking process

Requirements

• Python 3.8+
• PyTorch with CUDA support
• Ultralytics YOLOv8
• OpenCV
• MOT17 dataset

Experimental Results Summary

• Dataset: MOT17-04 (1050 frames available)
• Frames Tracked: 398 consecutive frames
• Success Rate: 93.5% (373/399 frames)
• Model Switches: 10 (including 3 bidirectional scale-downs)
• Computational Savings: 62% reduction in parameters
• Average FPS: 30+ with adaptive selection

Future Work

• Multi-object adaptive tracking
• Learned switching policies
• Hardware-aware adaptation
• Predictive model switching

License

MIT