SynSpill
Improved Industrial Spill Detection With Synthetic Data
Aaditya Baranwal¹†,Abdul Mueez¹,
Jason Voelker²,Guneet Bhatia²,Shruti Vyas¹
¹University of Central Florida • ²Siemens Energy
Abstract
Large-scale Vision-Language Models (VLMs) have transformed general-purpose visual recognition through strong zero-shot capabilities. However, their performance degrades significantly in niche, safety-critical domains such as industrial spill detection, where hazardous events are rare, sensitive, and difficult to annotate.
This scarcity—driven by privacy concerns, data sensitivity, and the infrequency of real incidents—renders conventional fine-tuning of detectors infeasible for most industrial settings.
We address this challenge by introducing a scalable framework centered on a high-quality synthetic data generation pipeline. We demonstrate that this synthetic corpus enables effective Parameter-Efficient Fine-Tuning (PEFT) of VLMs and substantially boosts the performance of state-of-the-art object detectors such as YOLO and DETR.
Notably, in the absence of synthetic data (SynSpill dataset), VLMs still generalize better to unseen spill scenarios than these detectors. When SynSpill is used, both VLMs and detectors achieve marked improvements, with their performance becoming comparable.
Our results underscore that a high-fidelity synthetic data is a powerful means to bridge the domain gap in safety-critical applications. The combination of synthetic generation and lightweight adaptation offers a cost-effective, scalable pathway for deploying vision systems in industrial environments where real data is scarce/impractical to obtain.
Key Highlights
Breakthrough innovations in synthetic data generation and model adaptation for industrial safety applications
Industrial Safety Focus
First comprehensive framework for automated industrial spill detection using computer vision
Synthetic Data Pipeline
Novel AnomalInfusion technique using Stable Diffusion XL + IP adapters for realistic spill generation
VLM Adaptation
Parameter-efficient fine-tuning with LoRA for domain specialization without full model retraining
Dual Approach
Benefits both Vision-Language Models and traditional object detectors (YOLO, DETR)
Zero-Shot Capability
Strong generalization to unseen spill scenarios even without synthetic data training
Real-World Validation
Tested and validated on actual industrial CCTV footage from manufacturing facilities
System Architecture
End-to-end pipeline for industrial spill detection using synthetic data and Vision-Language Models
Input Stage
VLM Processing
Alert System
AnomalInfusion Pipeline
Stable Diffusion XL
Base image generation with controlled prompts
IP Adapters
Style and content conditioning for realism
Inpainting
Precise spill placement and anomaly insertion
Data Flow
Methodology Comparison
Comprehensive evaluation of different approaches to industrial spill detection
Zero-Shot VLM
Baseline performance without any adaptation
- No training required
- General capabilities
- Limited domain knowledge
PEFT + SynSpill
Parameter-efficient fine-tuning with synthetic data
- LoRA adaptation
- Synthetic data training
- Domain specialization
Traditional Detectors
YOLO/DETR baseline without synthetic data
- Object detection focus
- Limited generalization
- Real-world challenges
Detectors + SynSpill
Traditional detectors trained with synthetic data
- Improved accuracy
- Better robustness
- Enhanced performance
PEFT Training Process
Generate
Create synthetic spill images using diffusion models
Annotate
Automatically label synthetic data with ground truth
Adapt
Fine-tune models using LoRA on synthetic dataset
Deploy
Deploy adapted model for real-world detection
💡 Key Innovation
Our synthetic data generation pipeline bridges the gap between scarce real-world data and the need for robust industrial spill detection, enabling effective model adaptation with minimal computational overhead.
Experimental Results
Comprehensive evaluation demonstrates significant improvements with synthetic data
Detection Accuracy
Best mAP@50 with Qwen-VL 32B + LoRA
Synthetic Images
High-quality samples generated via pipeline
Model Adaptation
Parameter-efficient fine-tuning approach
State-of-the-Art
Outperforms fine-tuned detectors
Quantitative Comparison
| Method | Public Dataset | Proprietary Dataset |
|---|---|---|
Qwen-VL 7B (Zero-Shot) | 35% | 15% |
Qwen-VL 32B (Zero-Shot) | 42% | 24% |
YOLOv11 (Fine-tuned) | 81% | 64% |
RF-DETR (Fine-tuned) | 83% | 67% |
Qwen-VL 7B + LoRA (V+L) | 78% | 66% |
Qwen-VL 32B + LoRA (V+L)Best | 84% | 71% |
🔍 Key Findings
PEFT VLMs achieve state-of-the-art performance - Qwen-VL 32B + LoRA (V+L) outperforms all baselines with 84% mAP@50
Synthetic data enables effective adaptation - 2,000 synthetic images bridge the domain gap for industrial spill detection
Joint vision-language adaptation optimal - LoRA (V+L) provides the best performance across both datasets
🚀 Impact
First scalable solution for industrial spill detection using synthetic data
Enables deployment in data-scarce industrial environments
Provides cost-effective alternative to manual monitoring
Research Collaboration
Collaborative effort between academic researchers and industry experts to advance Smart Sensing
🤝 Academic-Industry Collaboration
This work represents a successful collaboration between academic research and industrial application, combining cutting-edge computer vision research with real-world safety requirements in industrial environments. Our partnership ensures that research innovations translate directly into practical solutions for industrial safety.
Get In Touch
Citation
If you use our work, please cite our paper
📝 BibTeX Citation
@inproceedings{baranwal2025synspill,
title={SynSpill: Improved Industrial Spill Detection With Synthetic Data},
author={Baranwal, Aaditya and Mueez, Abdul and Voelker, Jason and Bhatia, Guneet and Vyas, Shruti},
booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision - Workshops (ICCV-W)},
year={2025}
}