Research map / 3D vision / Depth and stereo
PARSAC: Accelerating Robust Multi-Model Fitting with Parallel Sample Consensus
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes PARSAC, a neural-network-guided RANSAC-like multi-model fitting method that predicts sample and inlier weights for all model instances in one forward pass, enabling parallel real-time fitting of vanishing points, fundamental matrices and homographies, plus two new synthetic datasets (HOPE-F, SMH).
Problem
Robust multi-model fitting of geometric models (vanishing points, fundamental matrices, homographies) from noisy data is slow because prior methods discover model instances sequentially, and training data for fundamental matrix and homography fitting is scarce.
Approach
A neural network predicts per-observation sample weights and inlier weights for each putative instance simultaneously, so each instance is fit RANSAC-style in parallel using weighted inlier counts; it is trained end-to-end with a task loss via a score-function (REINFORCE-style) gradient without ground-truth segmentation.
Key results
- On SU3 PARSAC reaches AUC@1° of 67.44 and AUC@3° of 85.70 in about 4.9 ms per image, versus CONSAC's 3941 ms and Progressive-X's 28.9 ms.
- On HOPE-F PARSAC attains 14.97% ME versus 24.59% for Fast-CP, with the lowest Sampson error (3.07 px) and about twice Fast-CP's speed.
- On Adelaide-F and Adelaide-H PARSAC is outperformed by baselines such as Fast-CP and Progressive-X+, which the authors attribute to lack of refinement and training on synthetic data.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use PARSAC instead of NeurVPS | vanishing point estimation · SU3 | AUC @ $1\degree$: 79.47 -> 67.44 | Hurt |
| Use PARSAC instead of NeurVPS | vanishing point estimation · YUD | AUC @ $5\degree$: 72.61 -> 75.90 | Helped |
| Use PARSAC instead of DeepVP | vanishing point estimation · NYU-VP | AUC @ $5\degree$: 55.87 -> 51.65 | Hurt |
| Use PARSAC instead of DeepVP | vanishing point estimation · YUD+ | AUC @ $5\degree$: 59.57 -> 65.48 | Helped |
| Use PARSAC instead of the strongest baseline | fundamental matrix fitting · HOPE-F | ME: 22.78 -> 14.97 | Helped |
| Use PARSAC instead of Fast-CP | fundamental matrix fitting · Adelaide-F | ME: 4.92 -> 9.83 | Hurt |
| Use PARSAC instead of Progressive-X | homography fitting · SMH | ME: 20.60 -> 20.50 | No clear effect |
| Use PARSAC instead of CONSAC | homography fitting · Adelaide-H | ME: 5.66 -> 8.63 | Hurt |
| Train self-supervised instead of supervised (ablation) | PARSAC · SU3 | AUC @ $5\degree$: 90.17 -> 86.14 | Hurt |
| Use weighted rather than unweighted inlier counting during self-supervised training (ablation) | PARSAC · SU3 | AUC @ $5\degree$: 85.15 -> 86.14 | Helped |
| Use weighted rather than unweighted inlier counting during self-supervised training (ablation) | PARSAC · NYU-VP | AUC @ $10\degree$: 59.06 -> 61.31 | Helped |
| Use weighted rather than unweighted inlier counting during self-supervised training (ablation) | PARSAC · HOPE-F | ME: 18.93 -> 16.73 | Helped |
Surprising findings
- NeurVPS, trained on SU3, performs substantially better on SU3 but ranks second-to-last on YUD, suggesting poor generalisation beyond its training domain.
- Most competitors perform much worse on the new synthetic datasets than on Adelaide, suggesting they may be overfit to Adelaide.
Limitations
- Requires a GPU to utilise its full potential.
- The number of putative model instances must be fixed before training; instances beyond this cannot be found.
- Does not apply refinement for fundamental matrices and homographies, leading to weaker results on Adelaide.
Experimental setup
Datasets: SU3 NYU-VP HOPE-F SMH YUD YUD+ Adelaide-F Adelaide-H
Benchmarks: SU3 YUD YUD+ NYU-VP HOPE-F SMH Adelaide-F Adelaide-H
Metrics: AUC @ angle threshold misclassification error Sampson error transfer error run time (ms)
Baselines: CONSAC J-Linkage T-Linkage Progressive-X Progressive-X+ Fast-CP NeurVPS DeepVP Contrario-VP
Builds on: CONSAC RANSAC Progressive-X Multi-X
Closest papers by meaning
- CONSAC: Robust Multi-Model Fitting by Conditional Sample Consensus (CVPR 2020 · similarity 0.92)
- Finding Geometric Models by Clustering in the Consensus Space (CVPR 2023 · similarity 0.88)
- Geometric Multi-Model Fitting With a Convex Relaxation Algorithm (CVPR 2018 · similarity 0.87)
- Noisy One-point Homographies are Surprisingly Good (CVPR 2024 · similarity 0.85)
- RESfM: Robust Deep Equivariant Structure from Motion (ICLR 2025 · similarity 0.84)
- Scalable 3D Registration via Truncated Entry-wise Absolute Residuals (CVPR 2024 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.