Research map / Multi-agent systems and games / LLM agents

XFeat Revisited: Reproducibility and Evaluation of a Lightweight Image Matcher

NeurIPS 2026 · Multi-agent systems and games › LLM agents · approach: Agentic benchmarks and orchestration · 0 citations (OpenAlex)

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.

Reproduces XFeat from paper and supplement, re-evaluates the released checkpoint, runs architectural and loss ablations, and tests zero-shot transfer to retinal, thermal-visible, and remote-sensing matching.

Problem

Whether the lightweight XFeat local feature matcher can be reproduced from its paper and supplement, given discrepancies between paper, supplement, and released code, and whether its claims hold across tasks and image domains.

Approach

Re-implements XFeat from the paper and supplement (falling back to code only for omitted details), compares against the re-evaluated released checkpoint, tests a coupled keypoint detector and alternative skip-connection designs, and measures the effect of the code's loss formulation.

Key results

• Reproduced sparse XFeat reaches 57.9 AUC@10 on MegaDepth-1500 versus 57.1 for the re-evaluated checkpoint, and XFeat* reproduced reaches 37.1 AUC@10 on ScanNet-1500 versus 33.2 re-evaluated.
• Coupling the keypoint detector to the backbone drops XFeat* AUC@10 on MegaDepth-1500 from 65.7 to 53.8, supporting a separate keypoint branch for semi-dense matching.
• On cross-modal tests XFeat degrades sharply: the reproduced sparse model reaches only 9.01% precision on MTV thermal-visible and at most 12.86% S@10 on SRIF optical-infrared, with all variants below 2% S@10 on optical-SAR.

Surprising findings

• Aachen visual localization falls below reported results even for the authors' released checkpoint, pointing to underspecified evaluation details.
• ResNet-style and input-projection skip designs match or exceed the original single skip-connection, so the placement is not clearly preferable.

Limitations

• Aachen HLoc configuration unspecified, so localization results are only partially reproducible.
• Ablation variants showed high cross-seed variance, limiting conclusions on coupled detector and No Skip.
• Cross-modal and out-of-distribution evaluations use zero-shot settings with custom pipelines that may differ from original protocols.

Experimental setup

Datasets: MegaDepth COCO 2017 MegaDepth-1500 ScanNet-1500 HPatches Aachen Day-Night 1.0 RUBIK FIRE MTV SRIF

Benchmarks: MegaDepth-1500 ScanNet-1500 HPatches Aachen Day-Night 1.0 RUBIK FIRE MTV SRIF

Metrics: AUC@5 AUC@10 AUC@20 Acc@10 MIR #inliers FPS MHA Success rate AUC

Baselines: ALIKE DISK ORB SuperPoint SiLK ZippyPoint ALIKED+LightGlue LoFTR QuadTreeAttention RoMa MINIMA-RoMa SuperRetina

Builds on: XFeat ALIKE ResNet HLoc MAGSAC++

Closest papers by meaning

• [Re] Boosting the Visual Interpretability of CLIP via Adversarial Fine-Tuning (NeurIPS 2026 · similarity 0.97)
• Revisiting "Edit Away and My Face Will not Stay: Personal Biometric Defense against Malicious Generative Editing" (NeurIPS 2026 · similarity 0.96)
• Debate with Images: Detecting Deceptive Behaviors in Multimodal Large Language Models (ICML 2026 · similarity 0.91)
• Among Us: A Sandbox for Measuring and Detecting Agentic Deception (NeurIPS 2025 · similarity 0.88)
• DeceptionBench: A Comprehensive Benchmark for AI Deception Behaviors in Real-world Scenarios (NeurIPS 2025 · similarity 0.88)
• Multi-Agent Undercover Gaming: Hallucination Removal Through Counterfactual Test for Multimodal Reasoning (AAAI 2026 · similarity 0.87)