Source-linked AI summary
WireSeg-32K: A Physics-Grounded Synthetic Dataset for Wire Instance Segmentation
Zilin Dai, Lehong Wang, Yi Yang, Xiang Fei
TL;DR
Wire instance segmentation is difficult because wires are thin, deformable, often occluded, and expensive to annotate at scale. The paper presents WireSeg-32k, generated with DeformX, a Cosserat-rod and Isaac Sim co-simulation pipeline. Synthetic-only SAM3 LoRA fine-tuning improves real-world mAP@75 by 10.2% over the off-the-shelf model.
Problem
Wire instance segmentation is challenging, while large-scale real-world instance annotations are scarce and existing resources often provide tracing or semantic labels in constrained settings.
Source: §1 · p. 1Method
WireSeg-32k provides synthetic wire images with instance masks and depth plus a real-world test set, generated using DeformX, which couples Cosserat-rod dynamics with Isaac Sim.
Source: §1 · p. 1, §2 · pp. 2–3Results
10.2% improvement in real-world mAP@75 over off-the-shelf SAM3 is reported for LoRA fine-tuning trained only on WireSeg-32k.
Source: §1 · p. 1Takeaways & Limitations
Abstract
from arXiv · showhide
Deformable linear objects such as wires and cables are difficult to segment because they are thin, highly deformable, and frequently self-occluded, while large-scale instance-level annotations are expensive to obtain in real scenes. Existing resources either focus on cable tracing or semantic segmentation under constrained settings, or generate visually plausible images without physically grounded wire deformation. We present WireSeg-32k, a synthetic dataset for wire instance segmentation with 32,000 RGB images, instance masks, depth maps, and a complementary real-world test set with annotations. To generate this dataset, we develop DeformX, a co-simulation pipeline that couples Cosserat-rod dynamics with photorealistic Isaac Sim rendering, enabling physically plausible, contact-consistent wire shapes, CAD-based wire assets, and diverse visually grounded scenes. As a simple baseline, LoRA fine-tuning SAM3 on WireSeg-32k alone improves real-world mAP@75 by 10.2% over the off-the-shelf model, showing that physically grounded synthetic data can transfer to real wire perception.
1. INTRODUCTION
Wire instance segmentation is difficult and large real-world instance-labeled datasets remain scarce. WireSeg-32k pairs physically grounded synthetic data with an annotated real-world test set, and its synthetic-only SAM3 LoRA baseline improves real-world performance.
- Prior resources span cable tracing, semantic masks, or constrained scenes, leaving a need for instance labels and physically plausible wire configurations.HANDLOOM [6] focuses on trace-based grayscale, semi-planar images; other cited datasets and methods include semantic segmentation and constrained assumptions [3, 8, 1, 2].Source: §1 · p. 1
- WireSeg-32k provides 32,000 synthetic images with per-wire masks, depth maps, difficulty splits, and an annotated real-world test set.The images span tabletop, hanging-wire, and data-center scenarios.Source: §1 · p. 1
- 10.2% improvement in real-world mAP@75 over off-the-shelf SAM3 comes from LoRA fine-tuning trained only on WireSeg-32k.Source: §1 · p. 1
2. Method: DeformX
DeformX couples Cosserat-rod simulation with Isaac Sim to generate physically plausible wire shapes and visually realistic, labeled scenes. Its demonstrations qualitatively compare simulated and real wire behaviors, while the pipeline exports per-wire masks and depth alongside RGB images.
- Cosserat rods model stretching, shearing, bending, and twisting, providing control over gravity-driven sagging, stiffness, torsion, and shape transitions.The rods are discretized into vertices and segment frames and integrated in a dedicated engine.Source: §2 · p. 2, Figure 2 · p. 2
- Free-form mesh contact applies damped repulsive forces and reaction impulses, allowing wires to drape, slide, wrap, and form multi-point contacts.Broad-phase pruning accelerates contact queries, and a repulsion distance reduces deep penetration at larger time steps.Source: §2 · p. 2
- Simulated knot transitions qualitatively match real wire behavior, while scarf wrapping exhibits sliding, self-contact, and draping without unrealistic penetration.The demonstrations test torsion-bending coupling and geometry-conforming contact relevant to wire scenes.Source: Figure 4 · p. 3, §2 · p. 3
- A lightweight multi-rate interface couples the fine-step rod solver to Isaac Sim scene updates, supporting interactive authoring and headless generation.Rod states drive skinned tubular or CAD-derived meshes for rendering.Source: §2 · p. 3
- Randomized scenes and wire properties produce RGB images, per-wire instance masks, and depth maps, with labels tied to simulated DLO identities through deformation and occlusion.The pipeline varies objects, textures, viewpoints, lighting, wire count, lengths, radii, shapes, and material parameters.Source: Figure 3 · p. 2, §2 · p. 3
3. WireSeg-32K Dataset
WireSeg-32K combines diverse, realistic wire configurations with per-wire labels and a complementary annotated real-world test set.
3. WireSeg-32K Dataset
- Wire configurations vary in count, geometry, material properties, and initial states across tabletop, dangling-wire, and data-center scenes.Randomized viewpoints, lighting, clutter, and textures broaden visual variation, while simulated DLO identities provide consistent per-wire masks for overlapping and self-occluding structures.Source: §3 · p. 3
- Table 1 reports F1@75 and COCO mAP@75 for Easy, Medium, Hard, full synthetic, and held-out real subsets, with a final row for percent improvement.Both metrics are higher-is-better.Source: Table 1 · p. 4
- Table 2 compares DLO datasets on physics realism, instance labels, CAD support, and visually grounded scenes.Trace-only supervision counts as lacking instance labels; visually grounded scenes use realistic, image-based backgrounds rather than plain or purely synthetic backdrops.Source: Table 2 · p. 4
- 32,000 rendered images from more than 300 simulation runs include RGB, per-wire instance masks, depth, and a 300-image annotated real test set.Synthetic samples are stratified into easy, medium, and hard tiers using SAM3 performance; the thresholds are AP@75 < 0.3 for hard and AP@75 > 0.6 for easy.Source: §3 · p. 4
4. Vision Benchmark
The benchmark evaluates whether WireSeg-32K supervision improves SAM3 wire-instance segmentation on synthetic and held-out real images; LoRA fine-tuning improves performance across all subsets.
4. Vision Benchmark
- The benchmark tests training on WireSeg-32K against synthetic images and a held-out real-world test set.Performance is compared across easy, medium, and hard synthetic subsets, the full synthetic benchmark, and the held-out real set.Source: §4 · p. 4
- SAM3 is adapted with rank-16 LoRA modules in vision-encoder and mask-decoder attention projections while other parameters remain frozen.Training runs for 5 epochs on WireSeg-32K; evaluation reports dataset-level F1@75 with one-to-one matching at IoU ≥0.75 and COCO mAP@75.Source: §4 · p. 4
- LoRA fine-tuning consistently improves off-the-shelf SAM3 across all subsets, including real images, with especially clear gains on harder scenes.The harder scenes include crossings, clutter, and thin-object ambiguities.Source: §4 · p. 4
5. Conclusion
The paper presents WireSeg-32K and DeformX, which combines physically grounded wire simulation and rendering to generate labeled data; initial SAM3 experiments show measurable synthetic-to-real transfer.
5. Conclusion
- WireSeg-32K is presented as a physics-grounded synthetic dataset for wire instance segmentation, generated with the DeformX simulation tool.Source: §5 · p. 4
- DeformX combines Cosserat rod dynamics, free-form mesh contact, lightweight co-simulation, and mesh-skinned rendering to synthesize labeled data in realistic scenes.Source: §5 · p. 4
- Initial SAM3 experiments show measurable synthetic-to-real transfer, supporting WireSeg-32K as a benchmark for perception on thin deformable objects.Source: §5 · p. 4