Paul Voigtlaender

Lucas Beyer, Andreas Steiner, André Susano Pinto et al. · deepmind, oxford

PaliGemma is an open Vision-Language Model (VLM) that is based on the SigLIP-So400m vision encoder and the Gemma-2B language model. It is trained to be a versatile and broadly knowledgeable base model that is effective to transfer. It achieves strong performance on a wide variety of open-world tasks. We evaluate PaliGemma on almost 40 diverse tasks including standard VLM benchmarks, but also more specialized tasks such as remote-sensing and segmentation.

Ali Athar, Jonathon Luiten, Paul Voigtlaender et al.

Multiple existing benchmarks involve tracking and segmenting objects in video e.g., Video Object Segmentation (VOS) and Multi-Object Tracking and Segmentation (MOTS), but there is little interaction between them due to the use of disparate benchmark datasets and metrics (e.g. J&F, mAP, sMOTSA). As a result, published works usually target a particular benchmark, and are not easily comparable to each another. We believe that the development of generalized methods that can tackle multiple tasks requires greater cohesion among these research sub-communities. In this paper, we aim to facilitate this by proposing BURST, a dataset which contains thousands of diverse videos with high-quality object masks, and an associated benchmark with six tasks involving object tracking and segmentation in video. All tasks are evaluated using the same data and comparable metrics, which enables researchers to consider them in unison, and hence, more effectively pool knowledge from different methods across different tasks. Additionally, we demonstrate several baselines for all tasks and show that approaches for one task can be applied to another with a quantifiable and explainable performance difference. Dataset annotations and evaluation code is available at: https://github.com/Ali2500/BURST-benchmark.

Xi Chen, Xiao Wang, Lucas Beyer et al. · deepmind

This paper presents PaLI-3, a smaller, faster, and stronger vision language model (VLM) that compares favorably to similar models that are 10x larger. As part of arriving at this strong performance, we compare Vision Transformer (ViT) models pretrained using classification objectives to contrastively (SigLIP) pretrained ones. We find that, while slightly underperforming on standard image classification benchmarks, SigLIP-based PaLI shows superior performance across various multimodal benchmarks, especially on localization and visually-situated text understanding. We scale the SigLIP image encoder up to 2 billion parameters, and achieves a new state-of-the-art on multilingual cross-modal retrieval. We hope that PaLI-3, at only 5B parameters, rekindles research on fundamental pieces of complex VLMs, and could fuel a new generation of scaled-up models.

Paul Voigtlaender, Soravit Changpinyo, Jordi Pont-Tuset et al. · deepmind

We propose Video Localized Narratives, a new form of multimodal video annotations connecting vision and language. In the original Localized Narratives, annotators speak and move their mouse simultaneously on an image, thus grounding each word with a mouse trace segment. However, this is challenging on a video. Our new protocol empowers annotators to tell the story of a video with Localized Narratives, capturing even complex events involving multiple actors interacting with each other and with several passive objects. We annotated 20k videos of the OVIS, UVO, and Oops datasets, totalling 1.7M words. Based on this data, we also construct new benchmarks for the video narrative grounding and video question answering tasks, and provide reference results from strong baseline models. Our annotations are available at https://google.github.io/video-localized-narratives/.

Emanuele Bugliarello, Hernan Moraldo, Ruben Villegas et al. · tsinghua

Generating video stories from text prompts is a complex task. In addition to having high visual quality, videos need to realistically adhere to a sequence of text prompts whilst being consistent throughout the frames. Creating a benchmark for video generation requires data annotated over time, which contrasts with the single caption used often in video datasets. To fill this gap, we collect comprehensive human annotations on three existing datasets, and introduce StoryBench: a new, challenging multi-task benchmark to reliably evaluate forthcoming text-to-video models. Our benchmark includes three video generation tasks of increasing difficulty: action execution, where the next action must be generated starting from a conditioning video; story continuation, where a sequence of actions must be executed starting from a conditioning video; and story generation, where a video must be generated from only text prompts. We evaluate small yet strong text-to-video baselines, and show the benefits of training on story-like data algorithmically generated from existing video captions. Finally, we establish guidelines for human evaluation of video stories, and reaffirm the need of better automatic metrics for video generation. StoryBench aims at encouraging future research efforts in this exciting new area.

Valentin Gabeur, Shangbang Long, Songyou Peng et al.

Recent works show that image and video generators exhibit zero-shot visual understanding behaviors, in a way reminiscent of how LLMs develop emergent capabilities of language understanding and reasoning from generative pretraining. While it has long been conjectured that the ability to create visual content implies an ability to understand it, there has been limited evidence that generative vision models have developed strong understanding capabilities. In this work, we demonstrate that image generation training serves a role similar to LLM pretraining, and lets models learn powerful and general visual representations that enable SOTA performance on various vision tasks. We introduce Vision Banana, a generalist model built by instruction-tuning Nano Banana Pro (NBP) on a mixture of its original training data alongside a small amount of vision task data. By parameterizing the output space of vision tasks as RGB images, we seamlessly reframe perception as image generation. Our generalist model, Vision Banana, achieves SOTA results on a variety of vision tasks involving both 2D and 3D understanding, beating or rivaling zero-shot domain-specialists, including Segment Anything Model 3 on segmentation tasks, and the Depth Anything series on metric depth estimation. We show that these results can be achieved with lightweight instruction-tuning without sacrificing the base model's image generation capabilities. The superior results suggest that image generation pretraining is a generalist vision learner. It also shows that image generation serves as a unified and universal interface for vision tasks, similar to text generation's role in language understanding and reasoning. We could be witnessing a major paradigm shift for computer vision, where generative vision pretraining takes a central role in building Foundational Vision Models for both generation and understanding.

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.

Idil Esen Zulfikar, Sabarinath Mahadevan, Paul Voigtlaender et al.

Current state-of-the-art Video Object Segmentation (VOS) methods rely on dense per-object mask annotations both during training and testing. This requires time-consuming and costly video annotation mechanisms. We propose a novel Point-VOS task with a spatio-temporally sparse point-wise annotation scheme that substantially reduces the annotation effort. We apply our annotation scheme to two large-scale video datasets with text descriptions and annotate over 19M points across 133K objects in 32K videos. Based on our annotations, we propose a new Point-VOS benchmark, and a corresponding point-based training mechanism, which we use to establish strong baseline results. We show that existing VOS methods can easily be adapted to leverage our point annotations during training, and can achieve results close to the fully-supervised performance when trained on pseudo-masks generated from these points. In addition, we show that our data can be used to improve models that connect vision and language, by evaluating it on the Video Narrative Grounding (VNG) task. We will make our code and annotations available at https://pointvos.github.io.

Divya Kothandaraman, Kihyuk Sohn, Ruben Villegas et al.

We present a method for multi-concept customization of pretrained text-to-video (T2V) models. Intuitively, the multi-concept customized video can be derived from the (non-linear) intersection of the video manifolds of the individual concepts, which is not straightforward to find. We hypothesize that sequential and controlled walking towards the intersection of the video manifolds, directed by text prompting, leads to the solution. To do so, we generate the various concepts and their corresponding interactions, sequentially, in an autoregressive manner. Our method can generate videos of multiple custom concepts (subjects, action and background) such as a teddy bear running towards a brown teapot, a dog playing violin and a teddy bear swimming in the ocean. We quantitatively evaluate our method using videoCLIP and DINO scores, in addition to human evaluation. Videos for results presented in this paper can be found at https://github.com/divyakraman/MultiConceptVideo2024.

Paul Voigtlaender, Yuning Chai, Florian Schroff et al.

Many of the recent successful methods for video object segmentation (VOS) are overly complicated, heavily rely on fine-tuning on the first frame, and/or are slow, and are hence of limited practical use. In this work, we propose FEELVOS as a simple and fast method which does not rely on fine-tuning. In order to segment a video, for each frame FEELVOS uses a semantic pixel-wise embedding together with a global and a local matching mechanism to transfer information from the first frame and from the previous frame of the video to the current frame. In contrast to previous work, our embedding is only used as an internal guidance of a convolutional network. Our novel dynamic segmentation head allows us to train the network, including the embedding, end-to-end for the multiple object segmentation task with a cross entropy loss. We achieve a new state of the art in video object segmentation without fine-tuning with a J&F measure of 71.5% on the DAVIS 2017 validation set. We make our code and models available at https://github.com/tensorflow/models/tree/master/research/feelvos.

Mark Weber, Jun Xie, Maxwell Collins et al.

The task of assigning semantic classes and track identities to every pixel in a video is called video panoptic segmentation. Our work is the first that targets this task in a real-world setting requiring dense interpretation in both spatial and temporal domains. As the ground-truth for this task is difficult and expensive to obtain, existing datasets are either constructed synthetically or only sparsely annotated within short video clips. To overcome this, we introduce a new benchmark encompassing two datasets, KITTI-STEP, and MOTChallenge-STEP. The datasets contain long video sequences, providing challenging examples and a test-bed for studying long-term pixel-precise segmentation and tracking under real-world conditions. We further propose a novel evaluation metric Segmentation and Tracking Quality (STQ) that fairly balances semantic and tracking aspects of this task and is more appropriate for evaluating sequences of arbitrary length. Finally, we provide several baselines to evaluate the status of existing methods on this new challenging dataset. We have made our datasets, metric, benchmark servers, and baselines publicly available, and hope this will inspire future research.

Paul Voigtlaender, Lishu Luo, Chun Yuan et al.

For further progress in video object segmentation (VOS), larger, more diverse, and more challenging datasets will be necessary. However, densely labeling every frame with pixel masks does not scale to large datasets. We use a deep convolutional network to automatically create pseudo-labels on a pixel level from much cheaper bounding box annotations and investigate how far such pseudo-labels can carry us for training state-of-the-art VOS approaches. A very encouraging result of our study is that adding a manually annotated mask in only a single video frame for each object is sufficient to generate pseudo-labels which can be used to train a VOS method to reach almost the same performance level as when training with fully segmented videos. We use this workflow to create pixel pseudo-labels for the training set of the challenging tracking dataset TAO, and we manually annotate a subset of the validation set. Together, we obtain the new TAO-VOS benchmark, which we make publicly available at www.vision.rwth-aachen.de/page/taovos. While the performance of state-of-the-art methods on existing datasets starts to saturate, TAO-VOS remains very challenging for current algorithms and reveals their shortcomings.

Paul Voigtlaender, Jonathon Luiten, Philip H. S. Torr et al.

We present Siam R-CNN, a Siamese re-detection architecture which unleashes the full power of two-stage object detection approaches for visual object tracking. We combine this with a novel tracklet-based dynamic programming algorithm, which takes advantage of re-detections of both the first-frame template and previous-frame predictions, to model the full history of both the object to be tracked and potential distractor objects. This enables our approach to make better tracking decisions, as well as to re-detect tracked objects after long occlusion. Finally, we propose a novel hard example mining strategy to improve Siam R-CNN's robustness to similar looking objects. Siam R-CNN achieves the current best performance on ten tracking benchmarks, with especially strong results for long-term tracking. We make our code and models available at www.vision.rwth-aachen.de/page/siamrcnn.

Paul Voigtlaender, Jonathon Luiten, Bastian Leibe

We approach video object segmentation (VOS) by splitting the task into two sub-tasks: bounding box level tracking, followed by bounding box segmentation. Following this paradigm, we present BoLTVOS (Box-Level Tracking for VOS), which consists of an R-CNN detector conditioned on the first-frame bounding box to detect the object of interest, a temporal consistency rescoring algorithm, and a Box2Seg network that converts bounding boxes to segmentation masks. BoLTVOS performs VOS using only the firstframe bounding box without the mask. We evaluate our approach on DAVIS 2017 and YouTube-VOS, and show that it outperforms all methods that do not perform first-frame fine-tuning. We further present BoLTVOS-ft, which learns to segment the object in question using the first-frame mask while it is being tracked, without increasing the runtime. BoLTVOS-ft outperforms PReMVOS, the previously best performing VOS method on DAVIS 2016 and YouTube-VOS, while running up to 45 times faster. Our bounding box tracker also outperforms all previous short-term and longterm trackers on the bounding box level tracking datasets OTB 2015 and LTB35. A newer version of this work can be found at arXiv:1911.12836.

Aljosa Osep, Paul Voigtlaender, Jonathon Luiten et al.

This paper addresses the problem of object discovery from unlabeled driving videos captured in a realistic automotive setting. Identifying recurring object categories in such raw video streams is a very challenging problem. Not only do object candidates first have to be localized in the input images, but many interesting object categories occur relatively infrequently. Object discovery will therefore have to deal with the difficulties of operating in the long tail of the object distribution. We demonstrate the feasibility of performing fully automatic object discovery in such a setting by mining object tracks using a generic object tracker. In order to facilitate further research in object discovery, we release a collection of more than 360,000 automatically mined object tracks from 10+ hours of video data (560,000 frames). We use this dataset to evaluate the suitability of different feature representations and clustering strategies for object discovery.

Paul Voigtlaender, Michael Krause, Aljosa Osep et al.

This paper extends the popular task of multi-object tracking to multi-object tracking and segmentation (MOTS). Towards this goal, we create dense pixel-level annotations for two existing tracking datasets using a semi-automatic annotation procedure. Our new annotations comprise 65,213 pixel masks for 977 distinct objects (cars and pedestrians) in 10,870 video frames. For evaluation, we extend existing multi-object tracking metrics to this new task. Moreover, we propose a new baseline method which jointly addresses detection, tracking, and segmentation with a single convolutional network. We demonstrate the value of our datasets by achieving improvements in performance when training on MOTS annotations. We believe that our datasets, metrics and baseline will become a valuable resource towards developing multi-object tracking approaches that go beyond 2D bounding boxes. We make our annotations, code, and models available at https://www.vision.rwth-aachen.de/page/mots.

Aljosa Osep, Paul Voigtlaender, Mark Weber et al.

Many high-level video understanding methods require input in the form of object proposals. Currently, such proposals are predominantly generated with the help of networks that were trained for detecting and segmenting a set of known object classes, which limits their applicability to cases where all objects of interest are represented in the training set. This is a restriction for automotive scenarios, where unknown objects can frequently occur. We propose an approach that can reliably extract spatio-temporal object proposals for both known and unknown object categories from stereo video. Our 4D Generic Video Tubes (4D-GVT) method leverages motion cues, stereo data, and object instance segmentation to compute a compact set of video-object proposals that precisely localizes object candidates and their contours in 3D space and time. We show that given only a small amount of labeled data, our 4D-GVT proposal generator generalizes well to real-world scenarios, in which unknown categories appear. It outperforms other approaches that try to detect as many objects as possible by increasing the number of classes in the training set to several thousand.

Aljosa Osep, Paul Voigtlaender, Jonathon Luiten et al.

We propose to leverage a generic object tracker in order to perform object mining in large-scale unlabeled videos, captured in a realistic automotive setting. We present a dataset of more than 360'000 automatically mined object tracks from 10+ hours of video data (560'000 frames) and propose a method for automated novel category discovery and detector learning. In addition, we show preliminary results on using the mined tracks for object detector adaptation.

Jonathon Luiten, Paul Voigtlaender, Bastian Leibe

We address semi-supervised video object segmentation, the task of automatically generating accurate and consistent pixel masks for objects in a video sequence, given the first-frame ground truth annotations. Towards this goal, we present the PReMVOS algorithm (Proposal-generation, Refinement and Merging for Video Object Segmentation). Our method separates this problem into two steps, first generating a set of accurate object segmentation mask proposals for each video frame and then selecting and merging these proposals into accurate and temporally consistent pixel-wise object tracks over a video sequence in a way which is designed to specifically tackle the difficult challenges involved with segmenting multiple objects across a video sequence. Our approach surpasses all previous state-of-the-art results on the DAVIS 2017 video object segmentation benchmark with a J & F mean score of 71.6 on the test-dev dataset, and achieves first place in both the DAVIS 2018 Video Object Segmentation Challenge and the YouTube-VOS 1st Large-scale Video Object Segmentation Challenge.

Sabarinath Mahadevan, Paul Voigtlaender, Bastian Leibe

Deep learning requires large amounts of training data to be effective. For the task of object segmentation, manually labeling data is very expensive, and hence interactive methods are needed. Following recent approaches, we develop an interactive object segmentation system which uses user input in the form of clicks as the input to a convolutional network. While previous methods use heuristic click sampling strategies to emulate user clicks during training, we propose a new iterative training strategy. During training, we iteratively add clicks based on the errors of the currently predicted segmentation. We show that our iterative training strategy together with additional improvements to the network architecture results in improved results over the state-of-the-art.

Aljoša Ošep, Paul Voigtlaender, Jonathon Luiten et al.

We explore object discovery and detector adaptation based on unlabeled video sequences captured from a mobile platform. We propose a fully automatic approach for object mining from video which builds upon a generic object tracking approach. By applying this method to three large video datasets from autonomous driving and mobile robotics scenarios, we demonstrate its robustness and generality. Based on the object mining results, we propose a novel approach for unsupervised object discovery by appearance-based clustering. We show that this approach successfully discovers interesting objects relevant to driving scenarios. In addition, we perform self-supervised detector adaptation in order to improve detection performance on the KITTI dataset for existing categories. Our approach has direct relevance for enabling large-scale object learning for autonomous driving.

Aljoša Ošep, Wolfgang Mehner, Paul Voigtlaender et al.

The most common paradigm for vision-based multi-object tracking is tracking-by-detection, due to the availability of reliable detectors for several important object categories such as cars and pedestrians. However, future mobile systems will need a capability to cope with rich human-made environments, in which obtaining detectors for every possible object category would be infeasible. In this paper, we propose a model-free multi-object tracking approach that uses a category-agnostic image segmentation method to track objects. We present an efficient segmentation mask-based tracker which associates pixel-precise masks reported by the segmentation. Our approach can utilize semantic information whenever it is available for classifying objects at the track level, while retaining the capability to track generic unknown objects in the absence of such information. We demonstrate experimentally that our approach achieves performance comparable to state-of-the-art tracking-by-detection methods for popular object categories such as cars and pedestrians. Additionally, we show that the proposed method can discover and robustly track a large variety of other objects.

Paul Voigtlaender, Bastian Leibe

We tackle the task of semi-supervised video object segmentation, i.e. segmenting the pixels belonging to an object in the video using the ground truth pixel mask for the first frame. We build on the recently introduced one-shot video object segmentation (OSVOS) approach which uses a pretrained network and fine-tunes it on the first frame. While achieving impressive performance, at test time OSVOS uses the fine-tuned network in unchanged form and is not able to adapt to large changes in object appearance. To overcome this limitation, we propose Online Adaptive Video Object Segmentation (OnAVOS) which updates the network online using training examples selected based on the confidence of the network and the spatial configuration. Additionally, we add a pretraining step based on objectness, which is learned on PASCAL. Our experiments show that both extensions are highly effective and improve the state of the art on DAVIS to an intersection-over-union score of 85.7%.

Patrick Doetsch, Albert Zeyer, Paul Voigtlaender et al.

In this work we release our extensible and easily configurable neural network training software. It provides a rich set of functional layers with a particular focus on efficient training of recurrent neural network topologies on multiple GPUs. The source of the software package is public and freely available for academic research purposes and can be used as a framework or as a standalone tool which supports a flexible configuration. The software allows to train state-of-the-art deep bidirectional long short-term memory (LSTM) models on both one dimensional data like speech or two dimensional data like handwritten text and was used to develop successful submission systems in several evaluation campaigns.

Albert Zeyer, Patrick Doetsch, Paul Voigtlaender et al.

We present a comprehensive study of deep bidirectional long short-term memory (LSTM) recurrent neural network (RNN) based acoustic models for automatic speech recognition (ASR). We study the effect of size and depth and train models of up to 8 layers. We investigate the training aspect and study different variants of optimization methods, batching, truncated backpropagation, different regularization techniques such as dropout and $L_2$ regularization, and different gradient clipping variants. The major part of the experimental analysis was performed on the Quaero corpus. Additional experiments also were performed on the Switchboard corpus. Our best LSTM model has a relative improvement in word error rate of over 14\% compared to our best feed-forward neural network (FFNN) baseline on the Quaero task. On this task, we get our best result with an 8 layer bidirectional LSTM and we show that a pretraining scheme with layer-wise construction helps for deep LSTMs. Finally we compare the training calculation time of many of the presented experiments in relation with recognition performance. All the experiments were done with RETURNN, the RWTH extensible training framework for universal recurrent neural networks in combination with RASR, the RWTH ASR toolkit.