当前位置：锐英源 / 开源技术 / OpenCV / 2026年CVPR高光重点推荐项目三维物体感知转换器3PT-摘自OpenCV实时新闻

服务方向

人工智能数据处理
人工智能培训
小语种语音识别
语音识别标注
语音识别系统
语音识别转文字
软件开发
视觉检测
建筑视觉检测
养殖视觉检测: 运动控制卡上位机; 机械加工软件
软件开发培训: Java 安卓移动开发; VC++; C#软件; 汇编和破解; 驱动开发

联系方式

固话：0371-63888850
手机：138-0381-0136
Q Q：396806883
微信：ryysoft

锐英源精品原创，禁止全文或局部转载，禁止任何形式的非法使用，侵权必究。

2026年CVPR高光重点推荐项目三维物体感知转换器3PT-摘自OpenCV实时新闻

背景

本文是翻译文章，内容来自OpenCV邮件列表，因为锐英源软件一直在从事精细特征识别，3PT项目是同一个方向，基于技术方向的关注，翻译下文章。同时内容也列入了“语言之美英语爱好社群”，进行了英语学习价值的提高。

本次翻译收获：3PT用3D CAD训练，不是用二维图片来训练，所以在三维上有成果，正是用3D CAD训练，不用3D相机图片用二维图片就可以识别出结果。但不是细节精细特征模型，不过的确很高光。如果国内客户想快速部署，锐英源软件可以提供服务。3D数据3D算法，锐英源软件都有实际的项目经验，3D数据Context上下文有初步的研究，OpenGL、WebGL和Three.js开发过实际项目，比如3D钻井等等。

3PT确实很强大，但是训练压力上应该很大，毕竟3D数据的数据量，超过二维太多了。

中文

opencv-3pt

3PT：3D物体感知Transformer——CVPR 2026 Highlight

时间： 2026年8月27日（周四）上午9点（太平洋时间）

嘉宾： Agastya Kalra——Google（Intrinsic），夏威夷大学

3D物体感知Transformer（3PT）用两个基于多视角RGB图像的Transformer模型实现检测、分割和6DoF位姿估计任务。该模型在ICCV 2025 BOP挑战赛的工业机器人和AR/VR两个赛道中均以显著优势斩获第一名，展现了卓越的准确性和跨领域鲁棒性。如今，3PT架构作为Intrinsic视觉模型（IVM），已实际部署于全球范围内的真实工业机器人工作单元中。

Agastya将详细讲解该模型的架构、BOP 2025的竞赛成果，以及将模型从基准测试推向实际生产机器人应用所需的各项工作。

通过将检测和6D姿态估计统一为单一基础模型，我们消除了目前限制机器人灵活性的复杂“脆弱”模型集合的需求，实现了3D对象提示的二维检测、分割和6D姿态估计的先进成果，涵盖所有基准测试。这也使得许多具有挑战性的机器人应用成为可能，如拣货和精密插入。

在所有三维感知任务中实现最先进的结果

当前的三维感知流程往往缺乏对三维几何的基本理解，使其在高风险工业环境中显得脆弱。3PT（3D-物体感知转换器）是一个统一的基础模型，直接基于3D CAD提示训练。它用简化的双变压器流水线取代复杂的多模型启发式方法，仅使用现成的RGB摄像头实现亚毫米级精度

二维检测

与那些事后将图像与模板进行匹配的方法不同，3PT-D 在编码过程中直接以 3D 模型为条件进行建模。这使得它能够联合学习物体表征，并以更高的置信度对其进行定位。

6D姿态估计

它执行迭代的“渲染-比较”循环，将实时摄像头画面与内部的 3D 假设进行比对。该方案原生支持多视角设置，能够融合来自不同角度的信息，以消除视觉上的歧义。

周期时间

3PT 使用极线匹配（epipolar matching）和核密度估计（Kernel Density Estimation, KDE）来融合来自多个视角的观测，从而得出一个统一的空间真值。通过对二维检测结果进行三角测量以映射到三维空间，它能够滤除歧义，并在所有可用数据和视角之间达成共识。

无深度精度

传统上，高精度姿态估计需要高端深度传感器。3PT打破了这种依赖。通过采用一种新颖的多视角RGB细化环路，它实现了超越许多RGB-D方法的最先进精度。它消除了反光表面或强烈光线造成的深度伪影，使用标准相机硬件实现“工业级”效果。

单场、极具多样性的游戏

没有再培训。没有每个对象的微调。3PT真的是“一次性游戏”。通过一个3D CAD模型（“提示”），它可以检测、切割并估计该物体在任何场景中的6D姿态。它适用于各种不同的环境，比如从杂乱的家庭垃圾桶到复杂的工业“纸牌屋”结构。

工业应用

3PT 并非为基准测试而生，而是为了在工业自动化中实现真实的性能表现，以及在实际车间和工厂环境中落地使用。该模型可靠且专为实际应用打造。在接近真实制造的机器人任务中进行的 100 多次成功且可重复的实际操作，此操作验证了这一点我们的可靠和实用性，操作包括：

DIMM 插装：仅依靠视觉完成插入任务，配合间隙仅为 0.57mm。
钣金料箱拣选：清空 100 多个密集堆叠的薄型零件。

opencv-3pt-circle

英语

3PT: 3D-Object Perception Transformer — CVPR 2026 Highlight

Date/Time: Thursday, August 27th, 2026 (9am Pacific)

Guest: Agastya Kalra — Google (Intrinsic), University of Hawaii

The 3D-Object Perception Transformer (3PT) unifies detection, segmentation, and 6DoF pose estimation into two multi-view, RGB-only transformers. Demonstrating exceptional accuracy and cross-domain robustness, it placed first by significant margins in both the Industrial Robotics and AR/VR tracks of the BOP 2025 challenge at ICCV. Today, the 3PT architecture is actively deployed in real-world industrial robotic workcells across the world as the Intrinsic Vision Model (IVM).

Agastya will walk through the architecture, the BOP 2025 results, and what it took to move the model from benchmark to production robotics.

3D-Object Perception Transformer (3PT)
Architecture of the Intrinsic Vision Model
Pushing the boundary of 3D-Object Perception Tasks.Awarded a Highlight at CVPR 2026 - top 3% of submissions.

About the research

By unifying detection and 6D pose estimation into a single foundation model, we eliminate the need for the complex, "brittle" ensembles of disjointed models that currently limit robot flexibility, achieving state-of-the-art results in 3D-object prompted 2D detection, segmentation, and 6D pose estimation across all benchmarks. This also enables a host of challenging robotic applications such as bin picking and precision insertion.

Achieving state-of-the-art results across all 3D-Perception Tasks

Current 3D perception pipelines are often lack a fundamental understanding of 3D geometry, making them brittle in high-stakes industrial environments. 3PT (3D-Object Perception Transformer) is a unified foundation model trained directly on 3D CAD prompts. It replaces complex multi-model heuristics with a streamlined, two-transformer pipeline that achieves sub-millimeter precision using only off-the-shelf RGB cameras

2D-Detection

‍

Unlike methods that match images to templates after the fact, 3PT-D is directly conditioned on 3D models during the encoding process. This allows it to jointly learn object representations and localize them with superior confidence.

6D-Pose Estimation

‍

It performs an iterative render-and-compare loop, matching the live camera feed against internal 3D hypotheses. It natively scales to multi-view setups, fusing information from different angles to resolve visual ambiguities.

Cycle Time

‍

3PT uses epipolar matching and Kernel Density Estimation (KDE) to fuse observations from multiple angles and arrive at one spatial truth. Triangulating 2D detections into 3D space, filters out ambiguities, and achieving consensus across all available data and views.

Depth-free precision

Traditionally, high-accuracy pose estimation required high-end depth sensors. 3PT breaks this dependency. By utilizing a novel multi-view RGB refinement loop, it achieves state-of-the-art accuracy that surpasses many RGB-D methods. It eliminates depth artifacts caused by reflective surfaces or harsh lighting, delivering "industrial-grade" results with standard camera hardware.

One-shot, highly versatile

No retraining. No per-object fine-tuning. 3PT is truly "one-shot. With one 3D CAD model (the "prompt") it can detect, segment, and estimate the 6D pose of that object in any scene. It generalizes across diverse environments, from cluttered household bins to complex industrial "house-of-cards" structures.

Industrially robust

3PT isn't built for benchmarks, but to achieve real world performance for industrial automation, and real world use on the shop and factory floor. The model is reliable and built for real world use. We demonstrate this through over 100 successful and repeatable real-world attempts in realistic-to-manufacturing robotics tasks, including:

- DIMM-Insertion: Vision-only insertion with a 0.57mm clearance.
- Sheet Metal Bin Picking: Clearing 100+ densely cluttered, thin parts.

友情链接