Martin Sundermeyer

Martin Sundermeyer, Tomas Hodan, Yann Labbe et al. · tsinghua

We present the evaluation methodology, datasets and results of the BOP Challenge 2022, the fourth in a series of public competitions organized with the goal to capture the status quo in the field of 6D object pose estimation from an RGB/RGB-D image. In 2022, we witnessed another significant improvement in the pose estimation accuracy -- the state of the art, which was 56.9 AR$_C$ in 2019 (Vidal et al.) and 69.8 AR$_C$ in 2020 (CosyPose), moved to new heights of 83.7 AR$_C$ (GDRNPP). Out of 49 pose estimation methods evaluated since 2019, the top 18 are from 2022. Methods based on point pair features, which were introduced in 2010 and achieved competitive results even in 2020, are now clearly outperformed by deep learning methods. The synthetic-to-real domain gap was again significantly reduced, with 82.7 AR$_C$ achieved by GDRNPP trained only on synthetic images from BlenderProc. The fastest variant of GDRNPP reached 80.5 AR$_C$ with an average time per image of 0.23s. Since most of the recent methods for 6D object pose estimation begin by detecting/segmenting objects, we also started evaluating 2D object detection and segmentation performance based on the COCO metrics. Compared to the Mask R-CNN results from CosyPose in 2020, detection improved from 60.3 to 77.3 AP$_C$ and segmentation from 40.5 to 58.7 AP$_C$. The online evaluation system stays open and is available at: \href{http://bop.felk.cvut.cz/}{bop.felk.cvut.cz}.

Manuel Stoiber, Martin Sundermeyer, Rudolph Triebel

Tracking objects in 3D space and predicting their 6DoF pose is an essential task in computer vision. State-of-the-art approaches often rely on object texture to tackle this problem. However, while they achieve impressive results, many objects do not contain sufficient texture, violating the main underlying assumption. In the following, we thus propose ICG, a novel probabilistic tracker that fuses region and depth information and only requires the object geometry. Our method deploys correspondence lines and points to iteratively refine the pose. We also implement robust occlusion handling to improve performance in real-world settings. Experiments on the YCB-Video, OPT, and Choi datasets demonstrate that, even for textured objects, our approach outperforms the current state of the art with respect to accuracy and robustness. At the same time, ICG shows fast convergence and outstanding efficiency, requiring only 1.3 ms per frame on a single CPU core. Finally, we analyze the influence of individual components and discuss our performance compared to deep learning-based methods. The source code of our tracker is publicly available.

Maximilian Ulmer, Maximilian Durner, Martin Sundermeyer et al.

We present a novel technique to estimate the 6D pose of objects from single images where the 3D geometry of the object is only given approximately and not as a precise 3D model. To achieve this, we employ a dense 2D-to-3D correspondence predictor that regresses 3D model coordinates for every pixel. In addition to the 3D coordinates, our model also estimates the pixel-wise coordinate error to discard correspondences that are likely wrong. This allows us to generate multiple 6D pose hypotheses of the object, which we then refine iteratively using a highly efficient region-based approach. We also introduce a novel pixel-wise posterior formulation by which we can estimate the probability for each hypothesis and select the most likely one. As we show in experiments, our approach is capable of dealing with extreme visual conditions including overexposure, high contrast, or low signal-to-noise ratio. This makes it a powerful technique for the particularly challenging task of estimating the pose of tumbling satellites for in-orbit robotic applications. Our method achieves state-of-the-art performance on the SPEED+ dataset and has won the SPEC2021 post-mortem competition.

Manuel Stoiber, Martin Sundermeyer, Wout Boerdijk et al.

Kinematic structures are very common in the real world. They range from simple articulated objects to complex mechanical systems. However, despite their relevance, most model-based 3D tracking methods only consider rigid objects. To overcome this limitation, we propose a flexible framework that allows the extension of existing 6DoF algorithms to kinematic structures. Our approach focuses on methods that employ Newton-like optimization techniques, which are widely used in object tracking. The framework considers both tree-like and closed kinematic structures and allows a flexible configuration of joints and constraints. To project equations from individual rigid bodies to a multi-body system, Jacobians are used. For closed kinematic chains, a novel formulation that features Lagrange multipliers is developed. In a detailed mathematical proof, we show that our constraint formulation leads to an exact kinematic solution and converges in a single iteration. Based on the proposed framework, we extend ICG, which is a state-of-the-art rigid object tracking algorithm, to multi-body tracking. For the evaluation, we create a highly-realistic synthetic dataset that features a large number of sequences and various robots. Based on this dataset, we conduct a wide variety of experiments that demonstrate the excellent performance of the developed framework and our multi-body tracker.

Yongliang Lin, Yongzhi Su, Praveen Nathan et al.

In this work, we present a novel dense-correspondence method for 6DoF object pose estimation from a single RGB-D image. While many existing data-driven methods achieve impressive performance, they tend to be time-consuming due to their reliance on rendering-based refinement approaches. To circumvent this limitation, we present HiPose, which establishes 3D-3D correspondences in a coarse-to-fine manner with a hierarchical binary surface encoding. Unlike previous dense-correspondence methods, we estimate the correspondence surface by employing point-to-surface matching and iteratively constricting the surface until it becomes a correspondence point while gradually removing outliers. Extensive experiments on public benchmarks LM-O, YCB-V, and T-Less demonstrate that our method surpasses all refinement-free methods and is even on par with expensive refinement-based approaches. Crucially, our approach is computationally efficient and enables real-time critical applications with high accuracy requirements.

Sebastian Jung, Artem Zholus, Martin Sundermeyer et al.

Tracking-Any-Point (TAP) models aim to track any point through a video which is a crucial task in AR/XR and robotics applications. The recently introduced TAPNext approach proposes an end-to-end, recurrent transformer architecture to track points frame-by-frame in a purely online fashion -- demonstrating competitive performance at minimal latency. However, we show that TAPNext struggles with longer video sequences and also frequently fails to re-detect query points that reappear after being occluded or leaving the frame. In this work, we present TAPNext++, a model that tracks points in sequences that are orders of magnitude longer while preserving the low memory and compute footprint of the architecture. We train the recurrent video transformer using several data-driven solutions, including training on long 1024-frame sequences enabled by sequence parallelism techniques. We highlight that re-detection performance is a blind spot in the current literature and introduce a new metric, Re-Detection Average Jaccard ($AJ_{RD}$), to explicitly evaluate tracking on re-appearing points. To improve re-detection of points, we introduce tailored geometric augmentations, such as periodic roll that simulates point re-entries, and supervising occluded points. We demonstrate that recurrent transformers can be substantially improved for point tracking and set a new state-of-the-art on multiple benchmarks. Model and code can be found at https://tap-next-plus-plus.github.io.

Gemini Team, Petko Georgiev, Ving Ian Lei et al. · deepmind, mila

In this report, we introduce the Gemini 1.5 family of models, representing the next generation of highly compute-efficient multimodal models capable of recalling and reasoning over fine-grained information from millions of tokens of context, including multiple long documents and hours of video and audio. The family includes two new models: (1) an updated Gemini 1.5 Pro, which exceeds the February version on the great majority of capabilities and benchmarks; (2) Gemini 1.5 Flash, a more lightweight variant designed for efficiency with minimal regression in quality. Gemini 1.5 models achieve near-perfect recall on long-context retrieval tasks across modalities, improve the state-of-the-art in long-document QA, long-video QA and long-context ASR, and match or surpass Gemini 1.0 Ultra's state-of-the-art performance across a broad set of benchmarks. Studying the limits of Gemini 1.5's long-context ability, we find continued improvement in next-token prediction and near-perfect retrieval (>99%) up to at least 10M tokens, a generational leap over existing models such as Claude 3.0 (200k) and GPT-4 Turbo (128k). Finally, we highlight real-world use cases, such as Gemini 1.5 collaborating with professionals on completing their tasks achieving 26 to 75% time savings across 10 different job categories, as well as surprising new capabilities of large language models at the frontier; when given a grammar manual for Kalamang, a language with fewer than 200 speakers worldwide, the model learns to translate English to Kalamang at a similar level to a person who learned from the same content.

Gheorghe Comanici, Eric Bieber, Mike Schaekermann et al. · amazon-science, baidu

In this report, we introduce the Gemini 2.X model family: Gemini 2.5 Pro and Gemini 2.5 Flash, as well as our earlier Gemini 2.0 Flash and Flash-Lite models. Gemini 2.5 Pro is our most capable model yet, achieving SoTA performance on frontier coding and reasoning benchmarks. In addition to its incredible coding and reasoning skills, Gemini 2.5 Pro is a thinking model that excels at multimodal understanding and it is now able to process up to 3 hours of video content. Its unique combination of long context, multimodal and reasoning capabilities can be combined to unlock new agentic workflows. Gemini 2.5 Flash provides excellent reasoning abilities at a fraction of the compute and latency requirements and Gemini 2.0 Flash and Flash-Lite provide high performance at low latency and cost. Taken together, the Gemini 2.X model generation spans the full Pareto frontier of model capability vs cost, allowing users to explore the boundaries of what is possible with complex agentic problem solving.

Bin Yan, Martin Sundermeyer, David Joseph Tan et al.

In this paper, we address the challenge of performing open-vocabulary video instance segmentation (OV-VIS) in real-time. We analyze the computational bottlenecks of state-of-the-art foundation models that performs OV-VIS, and propose a new method, TROY-VIS, that significantly improves processing speed while maintaining high accuracy. We introduce three key techniques: (1) Decoupled Attention Feature Enhancer to speed up information interaction between different modalities and scales; (2) Flash Embedding Memory for obtaining fast text embeddings of object categories; and, (3) Kernel Interpolation for exploiting the temporal continuity in videos. Our experiments demonstrate that TROY-VIS achieves the best trade-off between accuracy and speed on two large-scale OV-VIS benchmarks, BURST and LV-VIS, running 20x faster than GLEE-Lite (25 FPS v.s. 1.25 FPS) with comparable or even better accuracy. These results demonstrate TROY-VIS's potential for real-time applications in dynamic environments such as mobile robotics and augmented reality. Code and model will be released at https://github.com/google-research/troyvis.

Wout Boerdijk, Martin Sundermeyer, Maximilian Durner et al.

We present a novel framework for self-supervised grasped object segmentation with a robotic manipulator. Our method successively learns an agnostic foreground segmentation followed by a distinction between manipulator and object solely by observing the motion between consecutive RGB frames. In contrast to previous approaches, we propose a single, end-to-end trainable architecture which jointly incorporates motion cues and semantic knowledge. Furthermore, while the motion of the manipulator and the object are substantial cues for our algorithm, we present means to robustly deal with distraction objects moving in the background, as well as with completely static scenes. Our method neither depends on any visual registration of a kinematic robot or 3D object models, nor on precise hand-eye calibration or any additional sensor data. By extensive experimental evaluation we demonstrate the superiority of our framework and provide detailed insights on its capability of dealing with the aforementioned extreme cases of motion. We also show that training a semantic segmentation network with the automatically labeled data achieves results on par with manually annotated training data. Code and pretrained model are available at https://github.com/DLR-RM/DistinctNet.

Tomas Hodan, Martin Sundermeyer, Bertram Drost et al.

This paper presents the evaluation methodology, datasets, and results of the BOP Challenge 2020, the third in a series of public competitions organized with the goal to capture the status quo in the field of 6D object pose estimation from an RGB-D image. In 2020, to reduce the domain gap between synthetic training and real test RGB images, the participants were provided 350K photorealistic training images generated by BlenderProc4BOP, a new open-source and light-weight physically-based renderer (PBR) and procedural data generator. Methods based on deep neural networks have finally caught up with methods based on point pair features, which were dominating previous editions of the challenge. Although the top-performing methods rely on RGB-D image channels, strong results were achieved when only RGB channels were used at both training and test time - out of the 26 evaluated methods, the third method was trained on RGB channels of PBR and real images, while the fifth on RGB channels of PBR images only. Strong data augmentation was identified as a key component of the top-performing CosyPose method, and the photorealism of PBR images was demonstrated effective despite the augmentation. The online evaluation system stays open and is available on the project website: bop.felk.cvut.cz.

Nikita Araslanov, Martin Sundermeyer, Hidenobu Matsuki et al.

One of the most exciting applications of vision models involve pixel-level reasoning. Despite the abundance of vision foundation models, we still lack representations that effectively embed spatio-temporal properties of visual scenes at the pixel level. Existing frameworks either train on image-based pretext tasks, which do not account for dynamic elements, or on video sequences for action-level reasoning, which does not scale to dense pixel-level prediction. We present a framework that learns pixel-accurate feature descriptors from videos, LILA. The core element of our training framework is linear in-context learning. LILA leverages spatio-temporal cue maps -- depth and motion -- estimated with off-the-shelf networks. Despite the noisy nature of those cues, LILA trains effectively on uncurated video datasets, embedding semantic and geometric properties in a temporally consistent manner. We demonstrate compelling empirical benefits of the learned representation across a diverse suite of vision tasks: video object segmentation, surface normal estimation and semantic segmentation.

Van Nguyen Nguyen, Stephen Tyree, Andrew Guo et al.

We present the evaluation methodology, datasets and results of the BOP Challenge 2024, the 6th in a series of public competitions organized to capture the state of the art in 6D object pose estimation and related tasks. In 2024, our goal was to transition BOP from lab-like setups to real-world scenarios. First, we introduced new model-free tasks, where no 3D object models are available and methods need to onboard objects just from provided reference videos. Second, we defined a new, more practical 6D object detection task where identities of objects visible in a test image are not provided as input. Third, we introduced new BOP-H3 datasets recorded with high-resolution sensors and AR/VR headsets, closely resembling real-world scenarios. BOP-H3 include 3D models and onboarding videos to support both model-based and model-free tasks. Participants competed on seven challenge tracks. Notably, the best 2024 method for model-based 6D localization of unseen objects (FreeZeV2.1) achieves 22% higher accuracy on BOP-Classic-Core than the best 2023 method (GenFlow), and is only 4% behind the best 2023 method for seen objects (GPose2023) although being significantly slower (24.9 vs 2.7s per image). A more practical 2024 method for this task is Co-op which takes only 0.8s per image and is 13% more accurate than GenFlow. Methods have similar rankings on 6D detection as on 6D localization but higher run time. On model-based 2D detection of unseen objects, the best 2024 method (MUSE) achieves 21--29% relative improvement compared to the best 2023 method (CNOS). However, the 2D detection accuracy for unseen objects is still -35% behind the accuracy for seen objects (GDet2023), and the 2D detection stage is consequently the main bottleneck of existing pipelines for 6D localization/detection of unseen objects. The online evaluation system stays open and is available at http://bop.felk.cvut.cz/

Junwen Huang, Jizhong Liang, Jiaqi Hu et al.

We introduce XYZ-IBD, a bin-picking dataset for 6D pose estimation that captures real-world industrial complexity, including challenging object geometries, reflective materials, severe occlusions, and dense clutter. The dataset reflects authentic robotic manipulation scenarios with millimeter-accurate annotations. Unlike existing datasets that primarily focus on household objects, which approach saturation,XYZ-IBD represents the unsolved realistic industrial conditions. The dataset features 15 texture-less, metallic, and mostly symmetrical objects of varying shapes and sizes. These objects are heavily occluded and randomly arranged in bins with high density, replicating the challenges of real-world bin-picking. XYZ-IBD was collected using two high-precision industrial cameras and one commercially available camera, providing RGB, grayscale, and depth images. It contains 75 multi-view real-world scenes, along with a large-scale synthetic dataset rendered under simulated bin-picking conditions. We employ a meticulous annotation pipeline that includes anti-reflection spray, multi-view depth fusion, and semi-automatic annotation, achieving millimeter-level pose labeling accuracy required for industrial manipulation. Quantification in simulated environments confirms the reliability of the ground-truth annotations. We benchmark state-of-the-art methods on 2D detection, 6D pose estimation, and depth estimation tasks on our dataset, revealing significant performance degradation in our setups compared to current academic household benchmarks. By capturing the complexity of real-world bin-picking scenarios, XYZ-IBD introduces more realistic and challenging problems for future research. The dataset and benchmark are publicly available at https://xyz-ibd.github.io/XYZ-IBD/.

Tomas Hodan, Martin Sundermeyer, Yann Labbe et al.

We present the evaluation methodology, datasets and results of the BOP Challenge 2023, the fifth in a series of public competitions organized to capture the state of the art in model-based 6D object pose estimation from an RGB/RGB-D image and related tasks. Besides the three tasks from 2022 (model-based 2D detection, 2D segmentation, and 6D localization of objects seen during training), the 2023 challenge introduced new variants of these tasks focused on objects unseen during training. In the new tasks, methods were required to learn new objects during a short onboarding stage (max 5 minutes, 1 GPU) from provided 3D object models. The best 2023 method for 6D localization of unseen objects (GenFlow) notably reached the accuracy of the best 2020 method for seen objects (CosyPose), although being noticeably slower. The best 2023 method for seen objects (GPose) achieved a moderate accuracy improvement but a significant 43% run-time improvement compared to the best 2022 counterpart (GDRNPP). Since 2017, the accuracy of 6D localization of seen objects has improved by more than 50% (from 56.9 to 85.6 AR_C). The online evaluation system stays open and is available at: http://bop.felk.cvut.cz/.

Martin Sundermeyer, Arsalan Mousavian, Rudolph Triebel et al.

Grasping unseen objects in unconstrained, cluttered environments is an essential skill for autonomous robotic manipulation. Despite recent progress in full 6-DoF grasp learning, existing approaches often consist of complex sequential pipelines that possess several potential failure points and run-times unsuitable for closed-loop grasping. Therefore, we propose an end-to-end network that efficiently generates a distribution of 6-DoF parallel-jaw grasps directly from a depth recording of a scene. Our novel grasp representation treats 3D points of the recorded point cloud as potential grasp contacts. By rooting the full 6-DoF grasp pose and width in the observed point cloud, we can reduce the dimensionality of our grasp representation to 4-DoF which greatly facilitates the learning process. Our class-agnostic approach is trained on 17 million simulated grasps and generalizes well to real world sensor data. In a robotic grasping study of unseen objects in structured clutter we achieve over 90% success rate, cutting the failure rate in half compared to a recent state-of-the-art method.

Maximilian Durner, Wout Boerdijk, Martin Sundermeyer et al.

Although instance-aware perception is a key prerequisite for many autonomous robotic applications, most of the methods only partially solve the problem by focusing solely on known object categories. However, for robots interacting in dynamic and cluttered environments, this is not realistic and severely limits the range of potential applications. Therefore, we propose a novel object instance segmentation approach that does not require any semantic or geometric information of the objects beforehand. In contrast to existing works, we do not explicitly use depth data as input, but rely on the insight that slight viewpoint changes, which for example are provided by stereo image pairs, are often sufficient to determine object boundaries and thus to segment objects. Focusing on the versatility of stereo sensors, we employ a transformer-based architecture that maps directly from the pair of input images to the object instances. This has the major advantage that instead of a noisy, and potentially incomplete depth map as an input, on which the segmentation is computed, we use the original image pair to infer the object instances and a dense depth map. In experiments in several different application domains, we show that our Instance Stereo Transformer (INSTR) algorithm outperforms current state-of-the-art methods that are based on depth maps. Training code and pretrained models will be made available.

Wout Boerdijk, Martin Sundermeyer, Maximilian Durner et al.

Accurate object segmentation is a crucial task in the context of robotic manipulation. However, creating sufficient annotated training data for neural networks is particularly time consuming and often requires manual labeling. To this end, we propose a simple, yet robust solution for learning to segment unknown objects grasped by a robot. Specifically, we exploit motion and temporal cues in RGB video sequences. Using optical flow estimation we first learn to predict segmentation masks of our given manipulator. Then, these annotations are used in combination with motion cues to automatically distinguish between background, manipulator and unknown, grasped object. In contrast to existing systems our approach is fully self-supervised and independent of precise camera calibration, 3D models or potentially imperfect depth data. We perform a thorough comparison with alternative baselines and approaches from literature. The object masks and views are shown to be suitable training data for segmentation networks that generalize to novel environments and also allow for watertight 3D reconstruction.

Maximilian Denninger, Martin Sundermeyer, Dominik Winkelbauer et al.

BlenderProc is a modular procedural pipeline, which helps in generating real looking images for the training of convolutional neural networks. These can be used in a variety of use cases including segmentation, depth, normal and pose estimation and many others. A key feature of our extension of blender is the simple to use modular pipeline, which was designed to be easily extendable. By offering standard modules, which cover a variety of scenarios, we provide a starting point on which new modules can be created.

Martin Sundermeyer, Maximilian Durner, En Yen Puang et al.

We introduce a scalable approach for object pose estimation trained on simulated RGB views of multiple 3D models together. We learn an encoding of object views that does not only describe an implicit orientation of all objects seen during training, but can also relate views of untrained objects. Our single-encoder-multi-decoder network is trained using a technique we denote "multi-path learning": While the encoder is shared by all objects, each decoder only reconstructs views of a single object. Consequently, views of different instances do not have to be separated in the latent space and can share common features. The resulting encoder generalizes well from synthetic to real data and across various instances, categories, model types and datasets. We systematically investigate the learned encodings, their generalization, and iterative refinement strategies on the ModelNet40 and T-LESS dataset. Despite training jointly on multiple objects, our 6D Object Detection pipeline achieves state-of-the-art results on T-LESS at much lower runtimes than competing approaches.

Martin Sundermeyer, Zoltan-Csaba Marton, Maximilian Durner et al.

We propose a real-time RGB-based pipeline for object detection and 6D pose estimation. Our novel 3D orientation estimation is based on a variant of the Denoising Autoencoder that is trained on simulated views of a 3D model using Domain Randomization. This so-called Augmented Autoencoder has several advantages over existing methods: It does not require real, pose-annotated training data, generalizes to various test sensors and inherently handles object and view symmetries. Instead of learning an explicit mapping from input images to object poses, it provides an implicit representation of object orientations defined by samples in a latent space. Our pipeline achieves state-of-the-art performance on the T-LESS dataset both in the RGB and RGB-D domain. We also evaluate on the LineMOD dataset where we can compete with other synthetically trained approaches. We further increase performance by correcting 3D orientation estimates to account for perspective errors when the object deviates from the image center and show extended results.