A Synthetic Dataset for Manometry Recognition in Robotic Applications

TL;DR

The paper tackles data scarcity and safety concerns in industrial perception by proposing a hybrid synthetic data pipeline that combines BlenderProc-rendered images with domain randomization and AI-driven Cosmos-Predict2 video generation. A large, richly annotated dataset of manometers is created and used to train a YOLO detector, demonstrating that a 1:1 real-to-synthetic mix yields the best performance, while synthetic data remains beneficial even with limited real data. The study highlights the cost, safety, and practicality advantages of synthetic augmentation for safety-critical, resource-constrained environments, and discusses limitations and avenues for future work, including broader asset classes and hardware-enabled validation. Overall, the work provides a concrete, scalable approach to close the sim-to-real gap in industrial perception through hybrid data generation and carefully designed dataset composition. The results support adopting synthetic augmentation as a viable, efficient path for deploying robust autonomous inspection systems in hazardous settings.

Abstract

This paper addresses the challenges of data scarcity and high acquisition costs in training robust object detection models for complex industrial environments, such as offshore oil platforms. Data collection in these hazardous settings often limits the development of autonomous inspection systems. To mitigate this issue, we propose a hybrid data synthesis pipeline that integrates procedural rendering and AI-driven video generation. The approach uses BlenderProc to produce photorealistic images with domain randomization and NVIDIA's Cosmos-Predict2 to generate physically consistent video sequences with temporal variation. A YOLO-based detector trained on a composite dataset, combining real and synthetic data, outperformed models trained solely on real images. A 1:1 ratio between real and synthetic samples achieved the highest accuracy. The results demonstrate that synthetic data generation is a viable, cost-effective, and safe strategy for developing reliable perception systems in safety-critical and resource-constrained industrial applications.

Figures (10)

• Figure 1: Proposed methodology pipeline for dataset creation and training. The framework integrates three sources: (i) real data acquisition and annotation (manual via CVAT (Computer Vision Annotation Tool) and semi-automatic via SAM2 (Segment Anything Model 2)), (ii) synthetic renders generated in BlenderProc with domain randomization (backgrounds, illumination, pose, post-processing), and (iii) AI-extended video frames generated through ComfyUI with NVIDIA Cosmos-Predict2 (relighting, motion, occlusion). All data sources are reviewed, merged, and used for training the YOLO-based object detector.
• Figure 2: Examples of real data annotations: (a) manual boxes curated in CVAT; (b) semi-automatic pixel masks using SAM2, later spot-checked by experts.
• Figure 3: BlenderProc synthetic samples.
• Figure 4: AI-extended video frames produced via ComfyUI + NVIDIA Cosmos-Predict2.
• Figure 5: These results indicate that adding synthetic data is beneficial, but excessive synthetic proportion (e.g., 1:3) does not yield further gains over Real-only and incurs additional compute cost.