AmeenAktharT/dawn-yolo-wbf-ensemble

🚗 Optimized Vehicle Detection using Ensemble YOLO + Weighted Boxes Fusion

M.Tech Thesis Project — Vehicle detection in adverse weather (fog, rain, snow, sand) using the DAWN dataset with a Self-Adaptive 3-Tier WBF ensemble of YOLO11m + YOLO26m.

📁 Repository Contents

File	Description
kaggle_notebook.py	Complete Kaggle-ready pipeline — copy-paste into a Kaggle notebook with dual T4 GPUs and run. Includes all 7 phases: data prep → HP search → training → fine-tuning → WBF ensemble → evaluation → results export.
prepare_data.py	Standalone data preparation script — downloads DAWN from HuggingFace, converts to YOLO format, augments minority classes, creates 60/20/20 split
train_pipeline.py	Alternative pipeline for HuggingFace Jobs execution

🏗️ Architecture

Models

• YOLO11m (20M params) — Ultralytics YOLO11 medium, COCO pretrained
• YOLO26m (20.4M params) — Ultralytics YOLO26 medium, COCO pretrained

Self-Adaptive 3-Tier WBF Ensemble

Input Image
    ├──→ YOLO11m ──→ Detections₁
    ├──→ YOLO11m_ft ──→ Detections₂  
    └──→ YOLO26m_ft ──→ Detections₃
                          │
                    ┌─────┴─────┐
                    │  Tier 1   │  Per-class F1-based weights (static)
                    │  Tier 2   │  Per-image confidence modulation (dynamic)
                    │  Tier 3   │  Log-dampened count normalization
                    └─────┬─────┘
                          │
                    Weighted Boxes Fusion
                          │
                    Fused Detections

Tier 1 — Performance-Calibrated Base Weights: For each model mᵢ and class cⱼ, compute F1 on validation set. Normalize so weights sum to 1 per class.

$$w_{i,j}^{base}=\frac{F_{i,j}}{\sum _k{F}_{k,j}}w^\{base\}\_\{i,j\}\; =\; \backslash frac\{F\_\{i,j\}\}\{\backslash sum\_\{k\}\; F\_\{k,j\}\}$$

Tier 2 — Per-Image Confidence Modulation: Dynamically shift weights based on each model's average confidence on the current image.

$${\varphi }_i(I)=1+\alpha \cdot ({\bar{s}}_i-0.5),\alpha =0.1\backslash phi\_i(I)\; =\; 1\; +\; \backslash alpha\; \backslash cdot\; (\backslash bar\{s\}\_i\; -\; 0.5),\; \backslash quad\; \backslash alpha\; =\; 0.1$$

Tier 3 — Box Count Normalization (Fixed): Log-dampened normalization prevents volume dominance without destroying score calibration.

$${\widehat{s}}_d=\frac{s_d\cdot w_{i,j}^{base}\cdot {\varphi }_i(I)}{{\log }_2(\max (n_{i,j},2))}\backslash hat\{s\}\_d\; =\; \backslash frac\{s\_d\; \backslash cdot\; w^\{base\}\_\{i,j\}\; \backslash cdot\; \backslash phi\_i(I)\}\{\backslash log\_2(\backslash max(n\_\{i,j\},\; 2))\}$$

📊 Dataset

• Source: DAWN Dataset (Kenk & Hassaballah, 2020)
• Weather: Fog (250), Rain (250), Snow (250), Sand (250) = 1000 base images
• Classes: Bicycle, Bus, Car, Motorcycle, Pedestrian, Truck
• Augmentation: Mirror + rotation for minority classes → ~1800+ total images
• Split: 60% train / 20% val / 20% test

Class Distribution (Original)

Class	Count	%
Car	6,454	82.2%
Truck	646	8.2%
Person	477	6.1%
Bus	161	2.1%
Motorcycle	81	1.0%
Bicycle	26	0.3%

🔧 Bug Fixes Applied

This pipeline fixes all 3 known bugs from the original Kaggle execution:

1. NoneType HP Search Error

Problem: T01: failed — 'NoneType' object has no attribute 'results_dict' Cause: model.train() returns None on silent crash (OOM, NaN loss) Fix: Null-check before accessing results + bounded HP ranges + torch.cuda.empty_cache() between trials

2. _thread.lock Pickling Error

Problem: ensemble.pkl failed: cannot pickle '_thread.lock' object Cause: YOLO objects contain CUDA contexts and thread locks Fix: save_config() / load_config() using JSON — only serializes paths and weights, reconstructs YOLO objects on load

3. Ensemble Metric Drop

Problem: Ensemble P/R lower than individual models despite strong per-class scores Causes & Fixes:

• sqrt(count) normalization too aggressive → changed to log2(count)
• scores / scores.max() re-normalization destroyed weight signal → removed entirely
• conf_alpha=0.3 caused over-modulation → reduced to 0.1
• Macro average included 0-support classes → now averages active classes only
• conf_type='avg' ignored model weights → changed to 'box_and_model_avg'

🚀 Quick Start (Kaggle)

1. Create a new Kaggle Notebook
2. Select GPU T4 x2 accelerator
3. Create a single code cell and paste the contents of kaggle_notebook.py
4. Run — fully autonomous, ~4-6 hours
5. Results in /kaggle/working/results/

📖 Citation

@article{kenk2020dawn,
  title={DAWN: Vehicle Detection in Adverse Weather Nature Dataset},
  author={Kenk, Mourad Ambarka and Hassaballah, M.},
  journal={arXiv preprint arXiv:2008.05402},
  year={2020}
}

@article{solovyev2021wbf,
  title={Weighted boxes fusion: Ensembling boxes from different object detection models},
  author={Solovyev, Roman and Wang, Weimin and Gabruseva, Tatiana},
  journal={Image and Vision Computing},
  year={2021}
}