SynSpill: Improved Industrial Spill Detection With Synthetic Data

1. Problem Statement

The paper addresses the critical challenge of industrial hazard detection (specifically fluid leaks, chemical spills, and stains) in safety-critical environments.

• Data Scarcity: Real-world industrial incident data is extremely rare, proprietary, and difficult to annotate due to privacy concerns and the infrequency of hazardous events.
• Limitations of Current Methods:
  • Traditional Object Detectors (e.g., YOLO, DETR): Require large-scale annotated datasets. Without them, they overfit to training domains and fail to generalize to variable lighting, facility layouts, or unseen anomaly types.
  • Zero-Shot Vision-Language Models (VLMs): While they possess strong generalization capabilities, their ability to localize specific objects (bounding boxes) and understand domain-specific visual cues is limited without adaptation.
• The Gap: There is a lack of scalable solutions that can bridge the "sim-to-real" gap for safety-critical applications where collecting real incident data is impractical or dangerous.

2. Methodology

The authors propose SynSpill, a framework combining high-fidelity synthetic data generation with Parameter-Efficient Fine-Tuning (PEFT) of Vision-Language Models (VLMs).

A. Synthetic Data Generation Pipeline ("AnomalInfusion")

Instead of relying on game engines or simple texture overlays, the authors developed a three-stage pipeline to generate 2,000 photorealistic industrial spill images:

1. Stage 1: Domain-Anchored Scene Generation:
   • Uses Stable Diffusion XL (SDXL) to generate diverse factory backgrounds.
   • Guidance Mechanisms:
     • Text Prompts: Define semantics (e.g., "concrete floor," "control room").
     • IP-Adapters: Transfer stylistic priors (textures, colors) from 150 unlabelled real factory images to ensure domain consistency.
     • LoRA Style Injection: Reinforces industrial realism.
2. Stage 2: Expert-Guided Anomaly Localization:
   • Human annotators identify semantically plausible locations for hazards (e.g., near valves, pipe junctions) rather than random placement. This ensures "causal logic" in the synthetic data.
3. Stage 3: Physically-Plausible Spill Inpainting:
   • Uses Differential Diffusion and SDXL inpainting to render the spill within the annotated bounding box.
   • Conditions the generation on material descriptions (e.g., "glistening oil") and real-world spill textures to ensure seamless blending with the background.

B. Model Adaptation Strategies

The study evaluates Qwen2.5-VL (3B, 7B, and 32B parameter versions) using three adaptation strategies:

1. Zero-Shot Inference: Using the pre-trained model without tuning.
2. In-Context Learning (ICL): Providing 5–15 labeled examples in the prompt (few-shot).
3. Parameter-Efficient Fine-Tuning (PEFT) via LoRA:
   • Adapts the model using Low-Rank Adaptation (LoRA) on the synthetic + public web dataset.
   • Three configurations tested: Language-only (LoRA-L), Vision-only (LoRA-V), and Joint (LoRA-V+L).
   • Only a small fraction of weights (<1%) are updated, preventing catastrophic forgetting.

C. Baselines

The proposed method is compared against:

• State-of-the-Art (SOTA) Detectors: YOLOv11 and RF-DETR (trained on the same hybrid dataset).
• Zero-Shot VLMs: The unadapted Qwen models.

3. Key Contributions

1. Novel Synthetic Pipeline: Introduction of AnomalInfusion, a controllable, human-in-the-loop pipeline that generates diverse, photorealistic industrial spill imagery using guided diffusion and expert localization, effectively bridging the sim-to-real gap.
2. Comprehensive VLM Evaluation: A systematic study evaluating VLMs (Qwen2.5-VL) for industrial spill detection, demonstrating that LoRA-based PEFT significantly outperforms zero-shot and few-shot (ICL) approaches.
3. Synthetic Data Efficacy: Demonstration that a purely synthetic dataset (augmented with public web data) enables VLMs to achieve performance comparable to, or exceeding, fully fine-tuned object detectors (YOLO/DETR) on real-world proprietary data.
4. Open Framework: Release of the SynSpill dataset and a modular framework designed for rapid deployment in data-scarce industrial environments.

4. Key Results

Experiments were conducted on Public Web Data and a Proprietary Factory Dataset (150 images with subtle leaks/occlusions).

• Performance of VLMs (Public Dataset @ IoU=0.5):

  • Zero-Shot: 0.42 (32B model).
  • ICL (15 shots): 0.63.
  • LoRA (V+L) on Synthetic Data: 0.71.
  • Insight: LoRA-V (vision-only) outperformed LoRA-L, indicating visual texture adaptation is more critical than language adaptation for this task.

• Performance on Proprietary Data (Real-World Test):

  • Zero-Shot: 0.24 (32B).
  • LoRA (V+L) on Synthetic Data: 0.49.
  • Comparison with SOTA Detectors:
    • RF-DETR (Fine-tuned): 0.67 mAP.
    • Qwen2.5-VL 32B + LoRA: 0.71 mAP.
  • Result: The PEFT-adapted VLM outperformed the specialized object detector (RF-DETR) by 4 points on the proprietary set, despite being trained on synthetic data.

• Qualitative Findings:

  • Adapted VLMs showed superior robustness to occlusions, lighting shifts, and subtle anomalies (e.g., semi-transparent oil stains) that traditional detectors missed.
  • In low-data regimes (without synthetic data), VLMs generalized better than detectors; with synthetic data, both became competitive, with VLMs taking the lead.

5. Significance and Implications

• Solving the Data Bottleneck: The paper proves that high-fidelity synthetic data can replace the need for collecting rare, dangerous, or proprietary real-world incident data.
• Scalability & Cost-Effectiveness: The PEFT approach allows for overnight adaptation on a single GPU, making it feasible for industrial sites with limited computational resources.
• Future-Proofing: The modular LoRA strategy allows the system to adapt to new foundation models as they emerge without retraining from scratch.
• Safety Impact: By enabling reliable detection of rare hazards, this framework supports proactive maintenance and prevents economic/environmental damage and human injury.

Conclusion: The authors conclude that "generate what you cannot collect, adapt what you cannot retrain" is a viable blueprint for industrial AI, offering a scalable path to safety-critical deployment where traditional data collection fails.