ROCOCO: Robust Outline Completion of Occluded Curvilinear Objects

• The work targets occlusion-induced discontinuities in curvilinear objects, whose variable segment states make pose estimation challenging.Source: §1 · p. 1
• The pipeline is presented as a two-stage learning-based approach for completing curvilinear objects under occlusion.Source: §1 · p. 1
• The authors report experiments ablating their approach and comparing it with relevant baselines.Source: §1 · p. 1
• ROCOCO is reported strongest when no occlusion mask is provided and competitive with mask-supervised inpainting when one is provided, using a small fraction of the parameters.Source: §1 · p. 2

• General perception methods may split long DLOs into multiple objects, while bounding-box and point-prompt approaches can struggle with sparse curvilinear targets.Source: §2 · p. 2
• Existing tracing methods often assume full visibility and a seed point, while occlusion work uses physical intervention rather than direct completion.Source: §2 · p. 2
• Diffusion inpainting completes masked pixels from known pixels, and pix2gestalt [11] focuses mainly on rigid or piecewise-rigid objects.Source: §2 · p. 2
• ROCOCO targets curvilinear completion from partial observations, using flow matching [8] for a direct noise-to-data probability path and more efficient sampling than DDPMs.Source: §2 · p. 2

• The target is the complete unoccluded feature representation recovered from an observation containing a partially occluded curvilinear feature.Source: §3 · p. 2
• The mapping from observation and optional mask to prediction is trained to minimize mean squared error, with coefficients balancing masked and global reconstruction loss.Source: §3 · p. 2
• Masked inpainting uses a provided mask to complete within the occluded region while preserving observed pixels.Source: §3 · p. 3
• Blind completion has no mask input and must jointly estimate occlusion regions and complete the feature from discontinuity patterns.Source: §3 · p. 3

4.1 Architecture Overview

• The architecture pairs a UNet mask predictor with a time-conditioned UNet flow matcher that predicts velocity fields.The networks share core building blocks; attention relates curve fragments across occluders and is placed at the coarsest level to limit quadratic token cost.Source: §4.1 · p. 3

4.2 Flow Matching Formulation

• Flow matching uses linear interpolation paths from data toward sampled noise, conditions on the observation and mask, and samples by Euler integration.Observed pixels are re-imposed after every step so visible content is preserved exactly.Source: §4.2 · p. 3

4.3 Two-Stage Inference

• For blind completion, the mask predictor estimates the occlusion mask before the flow matcher completes the curve, separating discontinuity detection from geometric completion.Repeated sampling can produce different plausible curves; averaging n completions inside the occlusion trades computation for accuracy.Source: §4.3 · p. 4

4.4 Additional Implementation Details

• The flow matcher has 7.72M parameters and the mask predictor 7.61M; synthetic samples are 64 × 64 and realistic-dataset samples are 128 × 128.The flow matcher uses different learning rates by dataset, with the higher rate found to diverge late on synthetic and crack data.Source: §4.4 · p. 4

5.1 Datasets

• Synthetic tests use procedurally generated curves and occluders, while realistic tests apply occluding polygons to resized crack and wire segmentation datasets.Synthetic samples are generated at runtime, and are used only when the occluder overlaps the curve.Source: §5.1 · p. 4

5.2 Evaluation Metrics

• Evaluation reports full-image and occluded-region MSE and ML1L, plus SSIM over the occluded crop.ML1L of 0.2 corresponds to an average 25.5 RGB-channel difference; the occluded-region losses are averaged over masked pixels.Source: §5.2 · p. 5

5.3.1 Synthetic Images

• On synthetic data, masked flow-matching lowers ML1LM to 0.089 from the naive UNet’s 0.135, while the blind pipeline reaches 0.092 and 0.844 SSIMM.Pure Flow scores 0.167 ML1LM; the blind pipeline’s mask predictor reaches IoU 0.977, nearly matching mask-supervised performance.Source: §5.3.1 · p. 5

5.3.2 Realistic Images

• On realistic wire and crack datasets, mask prediction reaches IoU 0.944 and 0.939, respectively, and ROCOCO improves on trivial completion references.ROCOCO scores 0.167 and 0.193 ML1LM on wires and cracks, versus 0.565 and 0.577 when leaving occluders in place.Source: Table 2 · p. 5, §5.3.2 · p. 6

5.3.3 Comparison with Inpainting Baselines

• With masks, ROCOCO achieves the best full-image MSE (0.0022 against LaMa’s 0.0037) and ML1L (0.0060 against 0.0297), while using a 7.7M-parameter completer.It is second to LaMa on masked ML1L (0.1615 against 0.1468) and SSIM (0.664 against 0.695); without masks, it is strongest by a wide margin.Source: Table 3 · p. 7, §5.3.3 · p. 7

5.4 Robustness and Inference Cost

• As occluded area grows from 0.8% to 10.3%, ML1LM rises from 0.191 to 0.253 and SSIMM falls from 0.677 to 0.533.The tested occluder radius spans 6% to 35% of image width.Source: §5.4 · p. 7, Figure 3 · p. 8
• At T = 5, averaging completions lowers ML1LM from 0.206 at n = 1 to 0.171 at n = 8, with little further gain at n = 16.Quality peaks at T ≈2 and then degrades slowly; the reported T = 5, n = 8 setting costs 40 network evaluations.Source: §5.4 · p. 7, Figure 4 · p. 8

6 Conclusion

• The two-stage pipeline pairs a flow-matching completer with a mask predictor, obtaining occlusion-region supervision without privileged mask information.Ablations show that flow matching improves prediction quality over a one-shot UNet and that localizing the occluded region specifically improves curve reconstruction.Source: §6 · p. 8
• ROCOCO is far stronger than the comparable mask-free inpainting alternative and remains competitive with dedicated inpainting models when a mask is provided.It achieves the best full-image error and second-best occluded-region result while using roughly a seventh of LaMa’s parameters and a hundredth of Stable Diffusion’s.Source: §6 · p. 8
• Hyperparameters, randomization ranges, thresholds, and data scenarios such as multiple curves, intersections, and more varied occlusions have not been explored exhaustively.The authors also identify color variation, such as shadows, and combining the method with curve tracing as future directions.Source: §6 · p. 8

Additional Figures

• Example-generation figures cover inpainting with a ground-truth mask, mask prediction, and a pure flow model.The listed examples also include ROCOCO on cracks and wires.Source: Figure 5 · p. 11, Figure 6 · p. 11, Figure 7 · p. 11
• A baseline-comparison example is provided for the wires dataset.Source: Figure 9 · p. 12
• Training-loss figures are provided for the naive UNet, mask prediction model, and flow-matching inpainting model.Source: Figure 10 · p. 13, Figure 11 · p. 13, Figure 12 · p. 13