Industrial Synthetic Segment Pre-training

TL;DR

This paper tackles licensing constraints and domain gaps in industrial instance segmentation by introducing InsCore, a synthetic pre-training dataset built on formula-driven supervised learning (FDSL). InsCore generates hollow-contour instance masks with complex occlusions using a single mathematical formulation, enabling fully supervised instance-level learning without real images or manual annotations. Across five industrial benchmarks, InsCore-pre-trained models outperform real-image pre-training (ImageNet-21k, COCO) and even fine-tuned SAM, achieving an average improvement of 6.2 points in segmentation performance while using only 0.1M synthetic images (over 100x fewer than SAM's SA-1B). The results establish InsCore as a practical, license-free foundation for industry, delivering strong transferability and data efficiency for industrial segmentation tasks.

Abstract

Pre-training on real-image datasets has been widely proven effective for improving instance segmentation. However, industrial applications face two key challenges: (1) legal and ethical restrictions, such as ImageNet's prohibition of commercial use, and (2) limited transferability due to the domain gap between web images and industrial imagery. Even recent vision foundation models, including the segment anything model (SAM), show notable performance degradation in industrial settings. These challenges raise critical questions: Can we build a vision foundation model for industrial applications without relying on real images or manual annotations? And can such models outperform even fine-tuned SAM on industrial datasets? To address these questions, we propose the Instance Core Segmentation Dataset (InsCore), a synthetic pre-training dataset based on formula-driven supervised learning (FDSL). InsCore generates fully annotated instance segmentation images that reflect key characteristics of industrial data, including complex occlusions, dense hierarchical masks, and diverse non-rigid shapes, distinct from typical web imagery. Unlike previous methods, InsCore requires neither real images nor human annotations. Experiments on five industrial datasets show that models pre-trained with InsCore outperform those trained on COCO and ImageNet-21k, as well as fine-tuned SAM, achieving an average improvement of 6.2 points in instance segmentation performance. This result is achieved using only 100k synthetic images, more than 100 times fewer than the 11 million images in SAM's SA-1B dataset, demonstrating the data efficiency of our approach. These findings position InsCore as a practical and license-free vision foundation model for industrial applications.

Figures (4)

• Figure 1: We propose the Instance Core Segmentation Dataset (InsCore), a commercially usable synthetic pre-training dataset for industrial instance segmentation. InsCore synthesizes complex occlusions and dense, hierarchical masks—key characteristics of industrial data—using a single mathematical formula. This design addresses the limitations of conventional datasets such as ImageNet and SA-1B, including commercial use restrictions and poor transferability to industrial domains. Consequently, InsCore delivers superior pre-training performance compared to models trained on real-image datasets, including fine-tuned SAM, for industrial instance segmentation tasks.
• Figure 2: Comparison between InsCore and conventional synthetic segmentation pre-training with SegRCDB. (a) InsCore simulates intricate occlusions and dense, hierarchical masks observed in industrial data by randomly placing hollow masks $S_k$, defined as the region between the innermost ($R_{\text{in}}$) and outermost ($R_{\text{out}}$) contours of formula-generated shapes. (b) SegRCDB provides three types of semantic segmentation masks (M1, M2, M3), none of which reflect the visual characteristics typically found in industrial datasets. In contrast, InsCore's hollow masks better capture features such as complex occlusions, allowing models to learn representations more suitable for industrial instance segmentation.
• Figure 3: Fine-tuning loss curves on (a) Industrial-iSeg and (b) LIVECell datasets with different pre-training initialization. Models pre-trained on InsCore exhibit the steepest and most stable convergence, outperforming those initialized with ImageNet-21k deng2009imagenet, MS COCO lin2014microsoft, or random (scratch). These results highlight the effectiveness of InsCore in accelerating optimization and improving training stability in industrial segmentation tasks.
• Figure 4: Instance segmentation results on Endoscapes, SpaceNet2, LIVECell, and Industrial-iSeg using different pre-training strategies. Each row corresponds to a dataset, and each column shows results from ground truth, ImageNet-21k, COCO, SAM, and InsCore. InsCore yields more accurate and stable mask predictions across datasets. SAM results are unavailable for LIVECell due to resource constraints.