WAPR, wide-angle pose refinement
WAPR

Acknowledgments and references致谢与参考资料

WAPR source code uses LGPL-2.1-only; the four first-party model weights use CC BY-ND 4.0. Dataset and third-party component licenses apply separately. See weight terms and the references below.

WAPR 源码采用 LGPL-2.1-only；四份第一方模型权重采用 CC BY-ND 4.0。数据集与第三方组件遵循各自许可。详见权重条款及下方参考资料。

The paper below is the primary reference for WAPR and SA6D. The release pipeline also uses separate pose-refinement and scoring checkpoints, while the demonstrations draw on external detectors, reconstruction models, datasets, and simulation tools. This page identifies each contribution where it enters the workflow. Consult each upstream project for its citation and license terms.

下述论文是 WAPR 与 SA6D 的主要引用来源。发布版流程还使用独立的位姿修正与评分权重；展示案例另涉及外部检测器、重建模型、数据集及仿真工具。本页按各组件进入流程的位置说明来源。其引用格式与许可条款以相应上游项目为准。

The first mention of a component or related method on each documentation page links to a footnote. The notes collect paper citations, official repositories, original sources and preserved license/notice copies; the complete inventory is licenses/manifest.json. WAPR's source remains LGPL-2.1. The four first-party WAPR checkpoints use CC BY-ND 4.0, with separate scope and attribution terms. Commercial use is free. Original redistribution requires attribution; adapted weights may not be shared. Enterprise versions and customization: Yulin Wang, shopedataset@gmail.com.

每个文档页在组件或相关方法首次出现处提供脚注，集中列出论文引用、官方仓库、原始来源与保留的许可/声明副本；完整清单见 licenses/manifest.json。WAPR 源码仍为 LGPL-2.1，四份第一方 WAPR 权重采用 CC BY-ND 4.0，见适用范围与署名声明。免费允许商用；原样再发行须署名，修改版权重不得共享。企业版本与定制需求联系 Yulin Wang：shopedataset@gmail.com。

Code and website author代码与网站作者

Yulin Wang solely developed the original WAPR release code, example scripts, project page and documentation website. School of Mechanical Engineering, Southeast University, China. Contact: yulinwang@seu.edu.cn.

Yulin Wang 独立编写了 WAPR 发布版的原创代码与示例脚本，并独立完成了宣传页及文档网站的设计与实现。中国，东南大学机械工程学院。联系方式：yulinwang@seu.edu.cn。

Copyright (c) 2026 Yulin Wang. All rights reserved, except as granted under the source license, GNU LGPL-2.1-only. Third-party code and adapted upstream portions retain their original attribution and terms; integration-file notices cover WAPR integration and modifications. Paper authorship is recorded separately in the publication citation on the project page. The four first-party checkpoints use CC BY-ND 4.0; commercial use is free, original redistribution requires attribution, and adapted weights may not be shared.

Copyright (c) 2026 Yulin Wang。版权所有，源码许可 GNU LGPL-2.1-only 授予的权利除外。第三方代码及改编的上游部分保留原作者署名与条款；集成文件中的署名对应 WAPR 的集成与修改。论文作者署名单独保留在项目主页的论文引用中。四份第一方模型权重采用 CC BY-ND 4.0；免费允许商用，原样再发行须署名，修改版权重不得共享。

Cite WAPR引用 WAPR

Wang et al., WAPR: A Foundation Model for Wide-Angle Refinement in Unseen Object Pose Estimation, ECCV 2026, presents the wide-angle refiner and the SA6D training set. All four checkpoints in this release—masked WAPR, unmasked WAPR, SAPR, and WBPS—were trained on SA6D; their inference roles are documented in the weights guide. The BibTeX entry is in the repository README and on the project page.

Wang 等人的 WAPR: A Foundation Model for Wide-Angle Refinement in Unseen Object Pose Estimation，ECCV 2026，介绍广角位姿修正模型与 SA6D 训练集。本发布包的四份权重——带掩码 WAPR、不带掩码 WAPR、SAPR 和 WBPS——均基于 SA6D 训练；各自的推理作用见权重指南。BibTeX 列于仓库 README 和项目主页。

Detection and segmentation检测与分割

The 2D pipeline obtains proposals with Grounding DINO[1], masks with the SAM 2.1[4]-L model loaded through Ultralytics[2], and visual features with DINOv2[3]. Grounding DINO's text encoder is BERT base uncased[14]. CAD-view shortlisting draws on CNOS[18]. The DINOv2 CLS + GeM[20] feature combination, pooling exponent 1.5 and absolute/relative similarity combination follow MUSE[19]; the feature scheme is described in Section 3.2. GeM pooling itself is attributed to Radenović, Tolias and Chum. These are method attributions, not a claim that the upstream projects implement this WAPR detector.

2D 流程使用 Grounding DINO[1] 产生候选框，经 Ultralytics[2] 加载 SAM 2.1[4]-L 获得掩码，再用 DINOv2[3] 提取视觉特征。Grounding DINO 的文本编码器是 BERT base uncased[14]。CAD 视角候选缩减参考 CNOS[18]。DINOv2 的 CLS 与 GeM[20] 特征组合、池化指数 1.5，以及绝对与相对相似度的组合参考 MUSE[19]，其特征组合见第 3.2 节；GeM 池化本身注明源自 Radenović、Tolias 和 Chum。这些是方法来源说明，不表示上游项目实现了本项目的完整检测器。

Single-frame reconstruction单帧重建

The reconstruction examples use SAM 2 for prompt masks and SAM 3D Objects[5] for object geometry. The local acceleration path adapts ideas from Fast-SAM3D[10]. The current stack also uses MoGe v1[6] internally, UniPose9D[7] for metric box dimensions, RoMa[8] for experimental correspondences, and Qwen2.5-VL[9] for optional language prompts. The RoMa results ranked with reference CAD are explicitly labeled as diagnostics on the reconstruction page.

重建示例使用 SAM 2 生成提示掩码，使用 SAM 3D Objects[5] 生成物体几何。本地加速路径参考 Fast-SAM3D[10]。当前流程还将 MoGe v1[6] 用作内部深度先验，以 UniPose9D[7] 预测米制包围盒边长，以 RoMa[8] 研究表面对应关系，并可用 Qwen2.5-VL[9] 处理语言提示。使用参考 CAD 排序的 RoMa 结果已在重建页明确标为诊断实验。

Datasets, comparisons, and simulation数据集、对照方法与仿真

The figures and lessons draw on the BOP benchmark[21] datasets, YCBInEOAT and TACO[16][15] tracking sequences, and ROBI[17]. The wide-angle and tracking comparisons cite the original MegaPose[12] and FoundationPose[11] papers; these are comparison methods, not required WAPR inference dependencies. The robot examples run in ManiSkill[13] and use its task assets. Each dataset page specifies the excerpt shown here, its source, and where a complete dataset can be obtained.

图示和教程采用 BOP 基准[21]中的数据集、YCBInEOAT 与 TACO[16][15] 跟踪片段，以及 ROBI[17]。广角修正与跟踪对照方法的原始论文包括 MegaPose[12] 和 FoundationPose[11]；它们不是 WAPR 推理的必需依赖。机器人示例在 ManiSkill[13] 中运行，并使用其任务资源。各数据页分别说明本站展示的摘录、来源及完整数据集的获取位置。

Visual evidence and licenses图示来源与许可

Each result figure should be read with its caption and the linked example or evaluation page. The BOP Challenge 2025 certificates on the project page name FRTPose-WAPR system submissions; the bin-picking certificates distinguish WAPR-only and FRTPose-WAPR results. Those certificates do not establish a standalone WAPR score for every displayed track. First-party source is governed by this repository's LICENSE; external code, weights, models, images, and datasets retain their own terms.

阅读结果图时，请结合图注及对应示例或评测页。项目主页的 BOP Challenge 2025 奖状署名为 FRTPose-WAPR 系统提交；料箱抓取奖状则区分了纯 WAPR 与 FRTPose-WAPR 的结果。这些奖状并不构成各赛道上 WAPR 独立方法的分数。第一方源码遵循本仓库 LICENSE；外部代码、权重、模型、图片和数据集仍遵循各自的条款。

References and licenses参考文献与许可

  1. GroundingDINO — Apache-2.0. Vendored detector source; original notices retained.随包检测源码；保留原始声明。 · GitHubGitHub · License/notice 1许可/声明 1 · License/notice 2许可/声明 2
    Liu et al. Grounding DINO: Marrying DINO with Grounded Pre-Training for Open-Set Object Detection. ECCV 2024. · Paper论文 ↩ ↩
  2. Ultralytics — AGPL-3.0. Vendored 8.3.70 loader; review AGPL obligations or obtain an upstream commercial license. Native Meta SAM 2 has separate terms.随包 8.3.70 加载器；须遵守 AGPL 条款或取得上游商业授权。Meta 原生 SAM 2 另有许可。 · GitHubGitHub · License/notice 1许可/声明 1 · License/notice 2许可/声明 2 ↩ ↩
  3. DINOv2 — Apache-2.0. Code and official DINOv2 weights; retain copyright and license.源码与官方 DINOv2 权重；保留版权和许可。 · GitHubGitHub · License/notice 1许可/声明 1
    Oquab et al. DINOv2: Learning Robust Visual Features without Supervision. arXiv:2304.07193, 2023. · Paper论文 ↩ ↩
  4. SAM 2 — Apache-2.0; cctorch: BSD-3-Clause. Native source and official checkpoints; cctorch carries an additional BSD notice. Ultralytics-converted files require checking their distributor terms.原生源码与官方权重；cctorch 另附 BSD 声明。Ultralytics 转换文件还需核对分发方条款。 · GitHubGitHub · License/notice 1许可/声明 1 · License/notice 2许可/声明 2
    Ravi et al. SAM 2: Segment Anything in Images and Videos. arXiv:2408.00714, 2024. · Paper论文 ↩ ↩
  5. SAM 3D Objects — SAM License (custom). Source and model materials follow the upstream custom agreement, including redistribution, attribution and use restrictions.源码及模型材料遵循上游自定义协议，包括再分发、署名与使用限制。 · GitHubGitHub · License/notice 1许可/声明 1
    SAM 3D Team et al. SAM 3D: 3Dfy Anything in Images. arXiv:2511.16624, 2025. · Paper论文 ↩ ↩
  6. MoGe — MIT; bundled notices also include Apache-2.0. MoGe source and the selected moge-vitl model (model card: MIT); retain the complete combined LICENSE and dependency notices.MoGe 源码与所选 moge-vitl 权重（模型卡：MIT）；保留完整组合 LICENSE 及依赖声明。 · GitHubGitHub · License/notice 1许可/声明 1
    Wang et al. MoGe: Unlocking Accurate Monocular Geometry Estimation for Open-Domain Images with Optimal Training Supervision. CVPR 2025. · Paper论文 ↩ ↩
  7. UniPose9D — Apache-2.0. Inference repository and its release; helper models retain their own terms.推理仓库及其发布材料；辅助模型保留各自条款。 · GitHubGitHub · License/notice 1许可/声明 1
    You et al. UniPose9D: Universal Category-Agnostic Object Pose Estimation. arXiv:2607.09985, 2026. · Paper论文 ↩ ↩
  8. RoMa — MIT. Matching source; separately obtained checkpoints and dependencies retain upstream terms.匹配源码；另行获取的权重与依赖保留上游条款。 · GitHubGitHub · License/notice 1许可/声明 1
    Edstedt et al. RoMa: Robust Dense Feature Matching. CVPR 2024. · Paper论文 ↩ ↩
  9. Qwen2.5-VL-3B-Instruct — Qwen Research License (custom). This exact 3B model permits non-commercial research/evaluation; commercial use requires permission from Alibaba Cloud. Do not infer its terms from other Qwen sizes.本项目选用的 3B 模型限定非商业研究/评估；商业使用须向 Alibaba Cloud 申请。不能套用其他 Qwen 规格的许可。 · Original source原始来源 · License/notice 1许可/声明 1
    Bai et al. Qwen2.5-VL Technical Report. arXiv:2502.13923, 2025. · Paper论文 ↩ ↩
  10. Fast-SAM3D — Not independently declared / 未单独声明. No root LICENSE was found in the checked upstream snapshot. SAM-derived code retains SAM terms; an independent grant for its added acceleration code remains unconfirmed.已检查的上游快照未发现根目录 LICENSE。SAM 派生代码保留 SAM 条款；新增加速代码的独立授权尚未确认。 · GitHubGitHub ↩ ↩
  11. FoundationPose — NVIDIA custom source license. Comparison method source has a custom license; do not describe it as MIT or presume weights share the same grant.对照方法源码使用自定义许可；不能标为 MIT，也不能推定权重有相同授权。 · GitHubGitHub · License/notice 1许可/声明 1
    Wen et al. FoundationPose: Unified 6D Pose Estimation and Tracking of Novel Objects. CVPR 2024. · Paper论文 ↩ ↩
  12. MegaPose — Apache-2.0. Comparison source; upstream model/data materials may have separate terms.对照源码；上游模型与数据可能有独立条款。 · GitHubGitHub · License/notice 1许可/声明 1
    Labbé et al. MegaPose: 6D Pose Estimation of Novel Objects via Render & Compare. CoRL 2022; PMLR 205, 2023. · Paper论文 ↩ ↩
  13. mani-skill — Apache-2.0. Runtime dependency; source license verified against the preserved LICENSE, with upstream copyright and third-party notices retained. Model/data assets keep separate terms.运行依赖；源码许可按已保留的 LICENSE 原文核实；保留上游版权及第三方声明，模型与数据资源另有条款。 · GitHubGitHub · License/notice 1许可/声明 1 · License/notice 2许可/声明 2
    Tao et al. ManiSkill3: GPU Parallelized Robotics Simulation and Rendering for Generalizable Embodied AI. RSS 2025. · Paper论文 ↩ ↩
  14. BERT base uncased — Apache-2.0. Official selected text model; preserve its model-repository terms.所选官方文本模型；保留其模型仓库条款。 · Original source原始来源 · License/notice 1许可/声明 1
    Devlin et al. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. NAACL-HLT 2019. · Paper论文 ↩ ↩
  15. TACO data — Not separately verified / 未单独核实. Data terms have not been separately verified. The official repository is the original source; the referenced Hugging Face acquisition mirror has no explicit license field. Confirm the data owner's terms for redistribution or commercial use.数据条款尚未单独核实。官方仓库为原始出处；实际获取数据的 Hugging Face 镜像数据卡未声明明确许可字段。再分发或商业使用须核实数据权利方的条款。 · GitHubGitHub
    Liu et al. TACO: Benchmarking Generalizable Bimanual Tool-ACtion-Object Understanding. CVPR 2024. · Paper论文 ↩ ↩
  16. YCBInEOAT data — Not separately verified / 未单独核实. Tracking data permission must be checked at the original archive; the comparison code license does not establish the data license.跟踪数据许可须在原始归档核实；对照代码的许可不能作为数据许可。 · GitHubGitHub · GitHubGitHub
    Wen et al. se(3)-TrackNet: Data-driven 6D Pose Tracking by Calibrating Image Residuals in Synthetic Domains. IROS 2020. Introduces YCBInEOAT. · Paper论文 ↩ ↩
  17. ROBI data — Not separately verified / 未单独核实. The saved public poses and dataset are credited to ROBI; an independent grant to redistribute them has not been verified.保存的公开位姿和数据均注明 ROBI 来源；其再分发授权尚未独立核实。 · GitHubGitHub
    Yang et al. ROBI: A Multi-View Dataset for Reflective Objects in Robotic Bin-Picking. IROS 2021. · Paper论文 ↩ ↩
  18. CNOS · GitHubGitHub
    Nguyen et al. CNOS: A Strong Baseline for CAD-based Novel Object Segmentation. ICCV Workshops 2023. · Paper论文 ↩ ↩
  19. MUSE
    Cho, Park and Oh. MUSE: Model-based Uncertainty-aware Similarity Estimation for zero-shot 2D Object Detection and Segmentation. arXiv:2510.17866, 2025. · Paper论文 ↩ ↩
  20. GeM pooling
    Radenović, Tolias and Chum. Fine-tuning CNN Image Retrieval with No Human Annotation. arXiv:1711.02512, 2017 (revised 2018). Introduces generalized-mean (GeM) pooling for image retrieval. · Paper论文 ↩ ↩
  21. BOP benchmark · GitHubGitHub
    Hodaň et al. BOP: Benchmark for 6D Object Pose Estimation. ECCV 2018. · Paper论文 ↩ ↩

Model weights and dataset/task assets may have separate terms. WAPR's source license does not replace them.模型权重、数据集与任务资源可能有独立条款；WAPR 源码许可不替代这些许可。