Research map / Probabilistic and scientific modeling / Variational and generative sampling
Generalized Differentiable RANSAC
Tong Wei, Yash Patel, Alexander Shekhovtsov, Jiri Matas, Daniel Barath
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| train sampling weights end-to-end with differentiable RANSAC instead of using the pretrained CLNet, OANet or NG-RANSAC | ∇-RANSAC · PhotoTourism F estimation (authors' pretrained baselines) | F1 score (%): 45.80 -> 48.12 | Improved |
| train with normalized 8PC solver instead of 7PC (ablation) | ∇-RANSAC (F estimation) · PhotoTourism F estimation | F1 score (%): 46.04 -> 47.21 | Improved |
| train end-to-end instead of KL initialization only (ablation) | ∇-RANSAC (F estimation) · PhotoTourism F estimation | median epipolar error (px): 1.73 -> 1.00 | Improved |
| use 5PC solver for E training rather than 8PC (ablation) | ∇-RANSAC (E estimation) · PhotoTourism E estimation | AUC | Improved |
| use epipolar error loss instead of pose error loss for E estimation (ablation) | ∇-RANSAC · PhotoTourism E estimation | AUC | Improved |
| use learned GS sampler weights instead of uniform sampling or SNN-guided GS (ablation) | ∇-RANSAC with MAGSAC++ · PhotoTourism E estimation | AUC | Improved |
| replace 8PC with 7PC for F training (ablation) | ∇-RANSAC · PhotoTourism F estimation | F1 score (%) | No clear effect |
| replace Gumbel-softmax objective with DSAC soft model selection (ablation) | ∇-RANSAC · PhotoTourism F estimation | F1 score (%): 48.12 -> 40.95 | Hurt |
| replace objective with SoftAM (ablation) | ∇-RANSAC · PhotoTourism F estimation | F1 score and run-time: 48.12 -> 48.17 F1; 28.66 -> 47.46 ms | No clear effect |
| use NG-RANSAC gradient approximation instead of GS relaxation (ablation) | ∇-RANSAC · PhotoTourism F estimation | F1 score (%): 48.12 -> 41.26 | Hurt |
| fine-tune LoFTR through ∇-RANSAC | LoFTR with MAGSAC-PROSAC · PhotoTourism F estimation | avg. epipolar error (px): 13.06 -> 10.31 | Improved |
| fine-tune LoFTR through ∇-RANSAC | LoFTR with OpenCV-RANSAC · PhotoTourism F estimation | F1 score (%): 64.07 -> 64.43 | No clear effect |
Closest papers by meaning
- Graph-Cut RANSAC (CVPR 2018 · similarity 0.88)
- BANSAC: A Dynamic BAyesian Network for Adaptive SAmple Consensus (ICCV 2023 · similarity 0.88)
- RANSAC Scoring Done Right (NeurIPS 2026 · similarity 0.87)
- Monte Carlo Diffusion for Generalizable Learning-Based RANSAC (AAAI 2026 · similarity 0.87)
- DSAC - Differentiable RANSAC for Camera Localization (CVPR 2017 · similarity 0.86)
- PARSAC: Accelerating Robust Multi-Model Fitting with Parallel Sample Consensus (AAAI 2024 · similarity 0.82)
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.
