SynSpill: Improved Industrial Spill Detection With Synthetic Data

Aaditya Baranwal¹^†, Abdul Mueez¹, Jason Voelker², Guneet Bhatia², Shruti Vyas¹

---

SETUP

For setting up the workflow and code, follow the instructions provided in the SETUP

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.

We address this challenge by introducing a scalable framework centered on a high-quality synthetic data generation pipeline that 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.

Key Highlights

• Industrial Safety Focus: Framework for automated industrial spill detection in safety-critical domains
• Synthetic Data Pipeline: Triple-guided generative process using Stable Diffusion XL with IP adapters and LoRA
• VLM Adaptation: Parameter-efficient fine-tuning with LoRA for vision-language models
• Dual Approach: Benefits both Vision-Language Models and object detectors (YOLO, DETR)
• Zero-Shot Capability: VLMs generalize better to unseen spill scenarios than traditional detectors
• Real-World Validation: Evaluated on both public and proprietary industrial datasets

Architecture Overview

graph TB
    A[Factory Images] --> B[Stage 1: Scene Generation]
    B --> C[Stage 2: Expert Annotation]
    C --> D[Stage 3: Spill Inpainting]
    D --> E[Synthetic Dataset]
    E --> F[PEFT Training]
    F --> G[Adapted VLM]
    
    H[SDXL + IP Adapter + LoRA] --> B
    I[Human-in-the-loop] --> C
    J[Differential Inpainting] --> D

Loading

Results

Dataset Composition:

• Synthetic images: 2,000 (Generated via Our Pipeline)
• Public dataset: 1,520 images (Deduplicated Roboflow Images)
• Proprietary dataset: 150 images (Seimens Energy)

Performance (mAP@50):

Method	Public Dataset	Proprietary Dataset
Qwen-VL 7B (Zero-Shot)	0.35	0.15
Qwen-VL 32B (Zero-Shot)	0.42	0.24
Baselines (Fine-Tuning w/ Synthetic + Public Data)
YOLOv11	0.81	0.64
RF-DETR	0.83	0.67
Proposed Method (PEFT w/ Synthetic + Public Data)
Qwen-VL 7B + LoRA (V+L)	0.78	0.66
Qwen-VL 32B + LoRA (V+L)	0.84	0.71

Key Findings:

• PEFT-adapted VLMs match or exceed detector performance
• Our proposed method achieves state-of-the-art results on both datasets
• Qwen-VL 32B + LoRA (V+L) outperforms all baselines including fine-tuned detectors
• Synthetic data enables effective domain adaptation for vision-language models

Citation

@inproceedings{baranwal2025synspill,
    title={SynSpill: Improved Industrial Spill Detection With Synthetic Data},
    author={Aaditya Baranwal, Abdul Mueez, Jason Voelker, Guneet Bhatia and Sruti Vyas},
    booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision - Workshops (ICCV-W)},
    year={2025},
    url={url={https://arxiv.org/abs/2508.10171},}
}

License

This project is licensed under the MIT License - see the LICENSE file for details.

Acknowledgments

• University of Central Florida - Research Support and Academic Guidance
• Siemens Energy - Industrial Partnership

---