Josef Šivic

Antoine Yang, Antoine Miech, Josef Sivic et al. · deepmind

Video question answering (VideoQA) is a complex task that requires diverse multi-modal data for training. Manual annotation of question and answers for videos, however, is tedious and prohibits scalability. To tackle this problem, recent methods consider zero-shot settings with no manual annotation of visual question-answer. In particular, a promising approach adapts frozen autoregressive language models pretrained on Web-scale text-only data to multi-modal inputs. In contrast, we here build on frozen bidirectional language models (BiLM) and show that such an approach provides a stronger and cheaper alternative for zero-shot VideoQA. In particular, (i) we combine visual inputs with the frozen BiLM using light trainable modules, (ii) we train such modules using Web-scraped multi-modal data, and finally (iii) we perform zero-shot VideoQA inference through masked language modeling, where the masked text is the answer to a given question. Our proposed approach, FrozenBiLM, outperforms the state of the art in zero-shot VideoQA by a significant margin on a variety of datasets, including LSMDC-FiB, iVQA, MSRVTT-QA, MSVD-QA, ActivityNet-QA, TGIF-FrameQA, How2QA and TVQA. It also demonstrates competitive performance in the few-shot and fully-supervised setting. Our code and models are publicly available at https://github.com/antoyang/FrozenBiLM.

Antoine Yang, Arsha Nagrani, Paul Hongsuck Seo et al. · deepmind

In this work, we introduce Vid2Seq, a multi-modal single-stage dense event captioning model pretrained on narrated videos which are readily-available at scale. The Vid2Seq architecture augments a language model with special time tokens, allowing it to seamlessly predict event boundaries and textual descriptions in the same output sequence. Such a unified model requires large-scale training data, which is not available in current annotated datasets. We show that it is possible to leverage unlabeled narrated videos for dense video captioning, by reformulating sentence boundaries of transcribed speech as pseudo event boundaries, and using the transcribed speech sentences as pseudo event captions. The resulting Vid2Seq model pretrained on the YT-Temporal-1B dataset improves the state of the art on a variety of dense video captioning benchmarks including YouCook2, ViTT and ActivityNet Captions. Vid2Seq also generalizes well to the tasks of video paragraph captioning and video clip captioning, and to few-shot settings. Our code is publicly available at https://antoyang.github.io/vid2seq.html.

Antoine Yang, Antoine Miech, Josef Sivic et al. · deepmind

Recent methods for visual question answering rely on large-scale annotated datasets. Manual annotation of questions and answers for videos, however, is tedious, expensive and prevents scalability. In this work, we propose to avoid manual annotation and generate a large-scale training dataset for video question answering making use of automatic cross-modal supervision. We leverage a question generation transformer trained on text data and use it to generate question-answer pairs from transcribed video narrations. Given narrated videos, we then automatically generate the HowToVQA69M dataset with 69M video-question-answer triplets. To handle the open vocabulary of diverse answers in this dataset, we propose a training procedure based on a contrastive loss between a video-question multi-modal transformer and an answer transformer. We introduce the zero-shot VideoQA task and the VideoQA feature probe evaluation setting and show excellent results, in particular for rare answers. Furthermore, our method achieves competitive results on MSRVTT-QA, ActivityNet-QA, MSVD-QA and How2QA datasets. We also show that our VideoQA dataset generation approach generalizes to another source of web video and text data. We use our method to generate the WebVidVQA3M dataset from the WebVid dataset, i.e., videos with alt-text annotations, and show its benefits for training VideoQA models. Finally, for a detailed evaluation we introduce iVQA, a new VideoQA dataset with reduced language bias and high-quality manual annotations. Code, datasets and trained models are available at https://antoyang.github.io/just-ask.html

Antoine Yang, Arsha Nagrani, Ivan Laptev et al. · deepmind

Segmenting long videos into chapters enables users to quickly navigate to the information of their interest. This important topic has been understudied due to the lack of publicly released datasets. To address this issue, we present VidChapters-7M, a dataset of 817K user-chaptered videos including 7M chapters in total. VidChapters-7M is automatically created from videos online in a scalable manner by scraping user-annotated chapters and hence without any additional manual annotation. We introduce the following three tasks based on this data. First, the video chapter generation task consists of temporally segmenting the video and generating a chapter title for each segment. To further dissect the problem, we also define two variants of this task: video chapter generation given ground-truth boundaries, which requires generating a chapter title given an annotated video segment, and video chapter grounding, which requires temporally localizing a chapter given its annotated title. We benchmark both simple baselines and state-of-the-art video-language models for these three tasks. We also show that pretraining on VidChapters-7M transfers well to dense video captioning tasks in both zero-shot and finetuning settings, largely improving the state of the art on the YouCook2 and ViTT benchmarks. Finally, our experiments reveal that downstream performance scales well with the size of the pretraining dataset. Our dataset, code, and models are publicly available at https://antoyang.github.io/vidchapters.html.

Hugo Cisneros, Josef Sivic, Tomas Mikolov

It is common to evaluate the performance of a machine learning model by measuring its predictive power on a test dataset. This approach favors complicated models that can smoothly fit complex functions and generalize well from training data points. Although essential components of intelligence, speed and data efficiency of this learning process are rarely reported or compared between different candidate models. In this paper, we introduce a benchmark of increasingly difficult tasks together with a data efficiency metric to measure how quickly machine learning models learn from training data. We compare the learning speed of some established sequential supervised models, such as RNNs, LSTMs, or Transformers, with relatively less known alternative models based on reservoir computing. The proposed tasks require a wide range of computational primitives, such as memory or the ability to compute Boolean functions, to be effectively solved. Surprisingly, we observe that reservoir computing systems that rely on dynamically evolving feature maps learn faster than fully supervised methods trained with stochastic gradient optimization while achieving comparable accuracy scores. The code, benchmark, trained models, and results to reproduce our experiments are available at https://github.com/hugcis/benchmark_learning_efficiency/ .

Yann Labbé, Lucas Manuelli, Arsalan Mousavian et al.

We introduce MegaPose, a method to estimate the 6D pose of novel objects, that is, objects unseen during training. At inference time, the method only assumes knowledge of (i) a region of interest displaying the object in the image and (ii) a CAD model of the observed object. The contributions of this work are threefold. First, we present a 6D pose refiner based on a render&compare strategy which can be applied to novel objects. The shape and coordinate system of the novel object are provided as inputs to the network by rendering multiple synthetic views of the object's CAD model. Second, we introduce a novel approach for coarse pose estimation which leverages a network trained to classify whether the pose error between a synthetic rendering and an observed image of the same object can be corrected by the refiner. Third, we introduce a large-scale synthetic dataset of photorealistic images of thousands of objects with diverse visual and shape properties and show that this diversity is crucial to obtain good generalization performance on novel objects. We train our approach on this large synthetic dataset and apply it without retraining to hundreds of novel objects in real images from several pose estimation benchmarks. Our approach achieves state-of-the-art performance on the ModelNet and YCB-Video datasets. An extensive evaluation on the 7 core datasets of the BOP challenge demonstrates that our approach achieves performance competitive with existing approaches that require access to the target objects during training. Code, dataset and trained models are available on the project page: https://megapose6d.github.io/.

Tomáš Souček, Jean-Baptiste Alayrac, Antoine Miech et al.

Human actions often induce changes of object states such as "cutting an apple", "cleaning shoes" or "pouring coffee". In this paper, we seek to temporally localize object states (e.g. "empty" and "full" cup) together with the corresponding state-modifying actions ("pouring coffee") in long uncurated videos with minimal supervision. The contributions of this work are threefold. First, we develop a self-supervised model for jointly learning state-modifying actions together with the corresponding object states from an uncurated set of videos from the Internet. The model is self-supervised by the causal ordering signal, i.e. initial object state $\rightarrow$ manipulating action $\rightarrow$ end state. Second, to cope with noisy uncurated training data, our model incorporates a noise adaptive weighting module supervised by a small number of annotated still images, that allows to efficiently filter out irrelevant videos during training. Third, we collect a new dataset with more than 2600 hours of video and 34 thousand changes of object states, and manually annotate a part of this data to validate our approach. Our results demonstrate substantial improvements over prior work in both action and object state-recognition in video.

Daniel McKee, Justin Salamon, Josef Sivic et al.

We propose a method to recommend music for an input video while allowing a user to guide music selection with free-form natural language. A key challenge of this problem setting is that existing music video datasets provide the needed (video, music) training pairs, but lack text descriptions of the music. This work addresses this challenge with the following three contributions. First, we propose a text-synthesis approach that relies on an analogy-based prompting procedure to generate natural language music descriptions from a large-scale language model (BLOOM-176B) given pre-trained music tagger outputs and a small number of human text descriptions. Second, we use these synthesized music descriptions to train a new trimodal model, which fuses text and video input representations to query music samples. For training, we introduce a text dropout regularization mechanism which we show is critical to model performance. Our model design allows for the retrieved music audio to agree with the two input modalities by matching visual style depicted in the video and musical genre, mood, or instrumentation described in the natural language query. Third, to evaluate our approach, we collect a testing dataset for our problem by annotating a subset of 4k clips from the YT8M-MusicVideo dataset with natural language music descriptions which we make publicly available. We show that our approach can match or exceed the performance of prior methods on video-to-music retrieval while significantly improving retrieval accuracy when using text guidance.

Antonin Vobecky, David Hurych, Oriane Siméoni et al.

This work investigates learning pixel-wise semantic image segmentation in urban scenes without any manual annotation, just from the raw non-curated data collected by cars which, equipped with cameras and LiDAR sensors, drive around a city. Our contributions are threefold. First, we propose a novel method for cross-modal unsupervised learning of semantic image segmentation by leveraging synchronized LiDAR and image data. The key ingredient of our method is the use of an object proposal module that analyzes the LiDAR point cloud to obtain proposals for spatially consistent objects. Second, we show that these 3D object proposals can be aligned with the input images and reliably clustered into semantically meaningful pseudo-classes. Finally, we develop a cross-modal distillation approach that leverages image data partially annotated with the resulting pseudo-classes to train a transformer-based model for image semantic segmentation. We show the generalization capabilities of our method by testing on four different testing datasets (Cityscapes, Dark Zurich, Nighttime Driving and ACDC) without any finetuning, and demonstrate significant improvements compared to the current state of the art on this problem. See project webpage https://vobecant.github.io/DriveAndSegment/ for the code and more.

Georgy Ponimatkin, Yann Labbé, Bryan Russell et al.

We introduce FocalPose, a neural render-and-compare method for jointly estimating the camera-object 6D pose and camera focal length given a single RGB input image depicting a known object. The contributions of this work are twofold. First, we derive a focal length update rule that extends an existing state-of-the-art render-and-compare 6D pose estimator to address the joint estimation task. Second, we investigate several different loss functions for jointly estimating the object pose and focal length. We find that a combination of direct focal length regression with a reprojection loss disentangling the contribution of translation, rotation, and focal length leads to improved results. We show results on three challenging benchmark datasets that depict known 3D models in uncontrolled settings. We demonstrate that our focal length and 6D pose estimates have lower error than the existing state-of-the-art methods.

Chun-Hsiao Yeh, Bryan Russell, Josef Sivic et al.

Large-scale vision-language models (VLM) have shown impressive results for language-guided search applications. While these models allow category-level queries, they currently struggle with personalized searches for moments in a video where a specific object instance such as ``My dog Biscuit'' appears. We present the following three contributions to address this problem. First, we describe a method to meta-personalize a pre-trained VLM, i.e., learning how to learn to personalize a VLM at test time to search in video. Our method extends the VLM's token vocabulary by learning novel word embeddings specific to each instance. To capture only instance-specific features, we represent each instance embedding as a combination of shared and learned global category features. Second, we propose to learn such personalization without explicit human supervision. Our approach automatically identifies moments of named visual instances in video using transcripts and vision-language similarity in the VLM's embedding space. Finally, we introduce This-Is-My, a personal video instance retrieval benchmark. We evaluate our approach on This-Is-My and DeepFashion2 and show that we obtain a 15% relative improvement over the state of the art on the latter dataset.

Tomáš Souček, Jean-Baptiste Alayrac, Antoine Miech et al.

We aim to learn to temporally localize object state changes and the corresponding state-modifying actions by observing people interacting with objects in long uncurated web videos. We introduce three principal contributions. First, we explore alternative multi-task network architectures and identify a model that enables efficient joint learning of multiple object states and actions such as pouring water and pouring coffee. Second, we design a multi-task self-supervised learning procedure that exploits different types of constraints between objects and state-modifying actions enabling end-to-end training of a model for temporal localization of object states and actions in videos from only noisy video-level supervision. Third, we report results on the large-scale ChangeIt and COIN datasets containing tens of thousands of long (un)curated web videos depicting various interactions such as hole drilling, cream whisking, or paper plane folding. We show that our multi-task model achieves a relative improvement of 40% over the prior single-task methods and significantly outperforms both image-based and video-based zero-shot models for this problem. We also test our method on long egocentric videos of the EPIC-KITCHENS and the Ego4D datasets in a zero-shot setup demonstrating the robustness of our learned model.

Vladimir Petrik, Mohammad Nomaan Qureshi, Josef Sivic et al.

We aim to teach robots to perform simple object manipulation tasks by watching a single video demonstration. Towards this goal, we propose an optimization approach that outputs a coarse and temporally evolving 3D scene to mimic the action demonstrated in the input video. Similar to previous work, a differentiable renderer ensures perceptual fidelity between the 3D scene and the 2D video. Our key novelty lies in the inclusion of a differentiable approach to solve a set of Ordinary Differential Equations (ODEs) that allows us to approximately model laws of physics such as gravity, friction, and hand-object or object-object interactions. This not only enables us to dramatically improve the quality of estimated hand and object states, but also produces physically admissible trajectories that can be directly translated to a robot without the need for costly reinforcement learning. We evaluate our approach on a 3D reconstruction task that consists of 54 video demonstrations sourced from 9 actions such as pull something from right to left or put something in front of something. Our approach improves over previous state-of-the-art by almost 30%, demonstrating superior quality on especially challenging actions involving physical interactions of two objects such as put something onto something. Finally, we showcase the learned skills on a Franka Emika Panda robot.

Jiri Sedlar, Karla Stepanova, Radoslav Skoviera et al.

This paper introduces a dataset for training and evaluating methods for 6D pose estimation of hand-held tools in task demonstrations captured by a standard RGB camera. Despite the significant progress of 6D pose estimation methods, their performance is usually limited for heavily occluded objects, which is a common case in imitation learning, where the object is typically partially occluded by the manipulating hand. Currently, there is a lack of datasets that would enable the development of robust 6D pose estimation methods for these conditions. To overcome this problem, we collect a new dataset (Imitrob) aimed at 6D pose estimation in imitation learning and other applications where a human holds a tool and performs a task. The dataset contains image sequences of nine different tools and twelve manipulation tasks with two camera viewpoints, four human subjects, and left/right hand. Each image is accompanied by an accurate ground truth measurement of the 6D object pose obtained by the HTC Vive motion tracking device. The use of the dataset is demonstrated by training and evaluating a recent 6D object pose estimation method (DOPE) in various setups.

Martin Cífka, Georgy Ponimatkin, Yann Labbé et al.

We introduce FocalPose++, a neural render-and-compare method for jointly estimating the camera-object 6D pose and camera focal length given a single RGB input image depicting a known object. The contributions of this work are threefold. First, we derive a focal length update rule that extends an existing state-of-the-art render-and-compare 6D pose estimator to address the joint estimation task. Second, we investigate several different loss functions for jointly estimating the object pose and focal length. We find that a combination of direct focal length regression with a reprojection loss disentangling the contribution of translation, rotation, and focal length leads to improved results. Third, we explore the effect of different synthetic training data on the performance of our method. Specifically, we investigate different distributions used for sampling object's 6D pose and camera's focal length when rendering the synthetic images, and show that parametric distribution fitted on real training data works the best. We show results on three challenging benchmark datasets that depict known 3D models in uncontrolled settings. We demonstrate that our focal length and 6D pose estimates have lower error than the existing state-of-the-art methods.

Antoine Miech, Ivan Laptev, Josef Sivic

Joint understanding of video and language is an active research area with many applications. Prior work in this domain typically relies on learning text-video embeddings. One difficulty with this approach, however, is the lack of large-scale annotated video-caption datasets for training. To address this issue, we aim at learning text-video embeddings from heterogeneous data sources. To this end, we propose a Mixture-of-Embedding-Experts (MEE) model with ability to handle missing input modalities during training. As a result, our framework can learn improved text-video embeddings simultaneously from image and video datasets. We also show the generalization of MEE to other input modalities such as face descriptors. We evaluate our method on the task of video retrieval and report results for the MPII Movie Description and MSR-VTT datasets. The proposed MEE model demonstrates significant improvements and outperforms previously reported methods on both text-to-video and video-to-text retrieval tasks. Code is available at: https://github.com/antoine77340/Mixture-of-Embedding-Experts

Antonin Vobecky, Oriane Siméoni, David Hurych et al.

We describe an approach to predict open-vocabulary 3D semantic voxel occupancy map from input 2D images with the objective of enabling 3D grounding, segmentation and retrieval of free-form language queries. This is a challenging problem because of the 2D-3D ambiguity and the open-vocabulary nature of the target tasks, where obtaining annotated training data in 3D is difficult. The contributions of this work are three-fold. First, we design a new model architecture for open-vocabulary 3D semantic occupancy prediction. The architecture consists of a 2D-3D encoder together with occupancy prediction and 3D-language heads. The output is a dense voxel map of 3D grounded language embeddings enabling a range of open-vocabulary tasks. Second, we develop a tri-modal self-supervised learning algorithm that leverages three modalities: (i) images, (ii) language and (iii) LiDAR point clouds, and enables training the proposed architecture using a strong pre-trained vision-language model without the need for any 3D manual language annotations. Finally, we demonstrate quantitatively the strengths of the proposed model on several open-vocabulary tasks: Zero-shot 3D semantic segmentation using existing datasets; 3D grounding and retrieval of free-form language queries, using a small dataset that we propose as an extension of nuScenes. You can find the project page here https://vobecant.github.io/POP3D.

Tomáš Souček, Dima Damen, Michael Wray et al.

We address the task of generating temporally consistent and physically plausible images of actions and object state transformations. Given an input image and a text prompt describing the targeted transformation, our generated images preserve the environment and transform objects in the initial image. Our contributions are threefold. First, we leverage a large body of instructional videos and automatically mine a dataset of triplets of consecutive frames corresponding to initial object states, actions, and resulting object transformations. Second, equipped with this data, we develop and train a conditioned diffusion model dubbed GenHowTo. Third, we evaluate GenHowTo on a variety of objects and actions and show superior performance compared to existing methods. In particular, we introduce a quantitative evaluation where GenHowTo achieves 88% and 74% on seen and unseen interaction categories, respectively, outperforming prior work by a large margin.

Joanna Materzynska, Josef Sivic, Eli Shechtman et al.

We introduce an approach for augmenting text-to-video generation models with customized motions, extending their capabilities beyond the motions depicted in the original training data. By leveraging a few video samples demonstrating specific movements as input, our method learns and generalizes the input motion patterns for diverse, text-specified scenarios. Our contributions are threefold. First, to achieve our results, we finetune an existing text-to-video model to learn a novel mapping between the depicted motion in the input examples to a new unique token. To avoid overfitting to the new custom motion, we introduce an approach for regularization over videos. Second, by leveraging the motion priors in a pretrained model, our method can produce novel videos featuring multiple people doing the custom motion, and can invoke the motion in combination with other motions. Furthermore, our approach extends to the multimodal customization of motion and appearance of individualized subjects, enabling the generation of videos featuring unique characters and distinct motions. Third, to validate our method, we introduce an approach for quantitatively evaluating the learned custom motion and perform a systematic ablation study. We show that our method significantly outperforms prior appearance-based customization approaches when extended to the motion customization task.

Georgy Ponimatkin, Martin Cífka, Tomáš Souček et al.

We seek to extract a temporally consistent 6D pose trajectory of a manipulated object from an Internet instructional video. This is a challenging set-up for current 6D pose estimation methods due to uncontrolled capturing conditions, subtle but dynamic object motions, and the fact that the exact mesh of the manipulated object is not known. To address these challenges, we present the following contributions. First, we develop a new method that estimates the 6D pose of any object in the input image without prior knowledge of the object itself. The method proceeds by (i) retrieving a CAD model similar to the depicted object from a large-scale model database, (ii) 6D aligning the retrieved CAD model with the input image, and (iii) grounding the absolute scale of the object with respect to the scene. Second, we extract smooth 6D object trajectories from Internet videos by carefully tracking the detected objects across video frames. The extracted object trajectories are then retargeted via trajectory optimization into the configuration space of a robotic manipulator. Third, we thoroughly evaluate and ablate our 6D pose estimation method on YCB-V and HOPE-Video datasets as well as a new dataset of instructional videos manually annotated with approximate 6D object trajectories. We demonstrate significant improvements over existing state-of-the-art RGB 6D pose estimation methods. Finally, we show that the 6D object motion estimated from Internet videos can be transferred to a 7-axis robotic manipulator both in a virtual simulator as well as in a real world set-up. We also successfully apply our method to egocentric videos taken from the EPIC-KITCHENS dataset, demonstrating potential for Embodied AI applications.

Anton Bushuiev, Roman Bushuiev, Jiri Sedlar et al.

In recent years, there has been remarkable progress in machine learning for protein-protein interactions. However, prior work has predominantly focused on improving learning algorithms, with less attention paid to evaluation strategies and data preparation. Here, we demonstrate that further development of machine learning methods may be hindered by the quality of existing train-test splits. Specifically, we find that commonly used splitting strategies for protein complexes, based on protein sequence or metadata similarity, introduce major data leakage. This may result in overoptimistic evaluation of generalization, as well as unfair benchmarking of the models, biased towards assessing their overfitting capacity rather than practical utility. To overcome the data leakage, we recommend constructing data splits based on 3D structural similarity of protein-protein interfaces and suggest corresponding algorithms. We believe that addressing the data leakage problem is critical for further progress in this research area.

Evangelos Kazakos, Cordelia Schmid, Josef Sivic

We propose a novel approach for captioning and object grounding in video, where the objects in the caption are grounded in the video via temporally dense bounding boxes. We introduce the following contributions. First, we present a large-scale automatic annotation method that aggregates frame-level captions grounded with bounding boxes into temporally dense and consistent annotations. We apply this approach on the HowTo100M dataset to construct a large-scale pre-training dataset, named HowToGround1M. We also introduce a Grounded Video Caption Generation model, dubbed GROVE, and pre-train the model on HowToGround1M. Second, we introduce iGround--a dataset of 3513 videos with manually annotated captions and dense spatio-temporally grounded bounding boxes. This allows us to measure progress on this challenging problem, as well as to fine-tune our model on this small-scale but high-quality data. Third, we demonstrate that our approach achieves state-of-the-art results on the proposed iGround dataset, as well as on the VidSTG, ActivityNet-Entities, GroundingYouTube, and YouCook-Interactions datasets. Our ablations demonstrate the importance of pre-training on our automatically annotated HowToGround1M dataset followed by fine-tuning on the manually annotated iGround dataset and validate the key technical contributions of our model. The dataset and code are available at https://ekazakos.github.io/grounded_video_caption_generation/.

Jakub Kopko, David Graber, Saltuk Mustafa Eyrilmez et al.

As machine learning becomes increasingly central to molecular design, it is vital to ensure the reliability of learnable protein-ligand scoring functions on novel protein targets. While many scoring functions perform well on standard benchmarks, their ability to generalize beyond training data remains a significant challenge. In this work, we evaluate the generalization capability of state-of-the-art scoring functions on dataset splits that simulate evaluation on targets with a limited number of known structures and experimental affinity measurements. Our analysis reveals that the commonly used benchmarks do not reflect the true challenge of generalizing to novel targets. We also investigate whether large-scale self-supervised pretraining can bridge this generalization gap and we provide preliminary evidence of its potential. Furthermore, we probe the efficacy of simple methods that leverage limited test-target data to improve scoring function performance. Our findings underscore the need for more rigorous evaluation protocols and offer practical guidance for designing scoring functions with predictive power extending to novel protein targets.

Mattia Soldan, Fabian Caba Heilbron, Bernard Ghanem et al.

Several video understanding tasks, such as natural language temporal video grounding, temporal activity localization, and audio description generation, require "temporally dense" reasoning over frames sampled at high temporal resolution. However, computing frame-level features for these tasks is computationally expensive given the temporal resolution requirements. In this paper, we make three contributions to reduce the cost of computing features for temporally dense tasks. First, we introduce a vision transformer (ViT) architecture, dubbed ResidualViT, that leverages the large temporal redundancy in videos to efficiently compute temporally dense frame-level features. Our architecture incorporates (i) learnable residual connections that ensure temporal consistency across consecutive frames and (ii) a token reduction module that enhances processing speed by selectively discarding temporally redundant information while reusing weights of a pretrained foundation model. Second, we propose a lightweight distillation strategy to approximate the frame-level features of the original foundation model. Finally, we evaluate our approach across four tasks and five datasets, in both zero-shot and fully supervised settings, demonstrating significant reductions in computational cost (up to 60%) and improvements in inference speed (up to 2.5x faster), all while closely approximating the accuracy of the original foundation model.

Fiona Ryan, Josef Sivic, Fabian Caba Heilbron et al. · gatech

Personalized vision-language retrieval seeks to recognize new concepts (e.g. "my dog Fido") from only a few examples. This task is challenging because it requires not only learning a new concept from a few images, but also integrating the personal and general knowledge together to recognize the concept in different contexts. In this paper, we show how to effectively adapt the internal representation of a vision-language dual encoder model for personalized vision-language retrieval. We find that regularized low-rank adaption of a small set of parameters in the language encoder's final layer serves as a highly effective alternative to textual inversion for recognizing the personal concept while preserving general knowledge. Additionally, we explore strategies for combining parameters of multiple learned personal concepts, finding that parameter addition is effective. To evaluate how well general knowledge is preserved in a finetuned representation, we introduce a metric that measures image retrieval accuracy based on captions generated by a vision language model (VLM). Our approach achieves state-of-the-art accuracy on two benchmarks for personalized image retrieval with natural language queries - DeepFashion2 and ConCon-Chi - outperforming the prior art by 4%-22% on personal retrievals.

Kateryna Zorina, David Kovar, Mederic Fourmy et al.

This work aims to leverage instructional video to solve complex multi-step task-and-motion planning tasks in robotics. Towards this goal, we propose an extension of the well-established Rapidly-Exploring Random Tree (RRT) planner, which simultaneously grows multiple trees around grasp and release states extracted from the guiding video. Our key novelty lies in combining contact states and 3D object poses extracted from the guiding video with a traditional planning algorithm that allows us to solve tasks with sequential dependencies, for example, if an object needs to be placed at a specific location to be grasped later. We also investigate the generalization capabilities of our approach to go beyond the scene depicted in the instructional video. To demonstrate the benefits of the proposed video-guided planning approach, we design a new benchmark with three challenging tasks: (I) 3D re-arrangement of multiple objects between a table and a shelf, (ii) multi-step transfer of an object through a tunnel, and (iii) transferring objects using a tray similar to a waiter transfers dishes. We demonstrate the effectiveness of our planning algorithm on several robots, including the Franka Emika Panda and the KUKA KMR iiwa. For a seamless transfer of the obtained plans to the real robot, we develop a trajectory refinement approach formulated as an optimal control problem (OCP).

Martin Malenický, Martin Cífka, Médéric Fourmy et al.

Accurate 6D object pose estimation from images is a key problem in object-centric scene understanding, enabling applications in robotics, augmented reality, and scene reconstruction. Despite recent advances, existing methods often produce physically inconsistent pose estimates, hindering their deployment in real-world scenarios. We introduce PhysPose, a novel approach that integrates physical reasoning into pose estimation through a postprocessing optimization enforcing non-penetration and gravitational constraints. By leveraging scene geometry, PhysPose refines pose estimates to ensure physical plausibility. Our approach achieves state-of-the-art accuracy on the YCB-Video dataset from the BOP benchmark and improves over the state-of-the-art pose estimation methods on the HOPE-Video dataset. Furthermore, we demonstrate its impact in robotics by significantly improving success rates in a challenging pick-and-place task, highlighting the importance of physical consistency in real-world applications.

Alejandro Pardo, Jui-Hsien Wang, Bernard Ghanem et al.

The objective of this work is to manipulate visual timelines (e.g. a video) through natural language instructions, making complex timeline editing tasks accessible to non-expert or potentially even disabled users. We call this task Instructed visual assembly. This task is challenging as it requires (i) identifying relevant visual content in the input timeline as well as retrieving relevant visual content in a given input (video) collection, (ii) understanding the input natural language instruction, and (iii) performing the desired edits of the input visual timeline to produce an output timeline. To address these challenges, we propose the Timeline Assembler, a generative model trained to perform instructed visual assembly tasks. The contributions of this work are three-fold. First, we develop a large multimodal language model, which is designed to process visual content, compactly represent timelines and accurately interpret timeline editing instructions. Second, we introduce a novel method for automatically generating datasets for visual assembly tasks, enabling efficient training of our model without the need for human-labeled data. Third, we validate our approach by creating two novel datasets for image and video assembly, demonstrating that the Timeline Assembler substantially outperforms established baseline models, including the recent GPT-4o, in accurately executing complex assembly instructions across various real-world inspired scenarios.

Evangelos Kazakos, Cordelia Schmid, Josef Sivic

We propose a new task, dataset and model for grounded video caption generation. This task unifies captioning and object grounding in video, where the objects in the caption are grounded in the video via temporally consistent bounding boxes. We introduce the following contributions. First, we present a task definition and a manually annotated test dataset for this task, referred to as GROunded Video Caption Generation (GROC). Second, we introduce a large-scale automatic annotation method leveraging an existing model for grounded still image captioning together with an LLM for summarising frame-level captions into temporally consistent captions in video. Furthermore, we prompt the LLM to track by language -- classifying noun phrases from the frame-level captions into noun phrases of the video-level generated caption. We apply this approach to videos from the HowTo100M dataset, which results in a new large-scale training dataset, called HowToGround, with automatically annotated captions and spatio-temporally consistent bounding boxes with coherent natural language labels. Third, we introduce a new grounded video caption generation model, called VideoGround, and train the model on the new automatically annotated HowToGround dataset. Finally, results of our VideoGround model set the state of the art for the new task of grounded video caption generation. We perform extensive ablations and demonstrate the importance of key technical contributions of our model.

Anton Bushuiev, Roman Bushuiev, Olga Pimenova et al.

Generalization beyond training data remains a central challenge in machine learning for biology. A common way to enhance generalization is self-supervised pre-training on large datasets. However, aiming to perform well on all possible proteins can limit a model's capacity to excel on any specific one, whereas experimentalists typically need accurate predictions for individual proteins they study, often not covered in training data. To address this limitation, we propose a method that enables self-supervised customization of protein language models to one target protein at a time, on the fly, and without assuming any additional data. We show that our Protein Test-Time Training (ProteinTTT) method consistently enhances generalization across different models, their sizes, and datasets. ProteinTTT improves structure prediction for challenging targets, achieves new state-of-the-art results on protein fitness prediction, and enhances function prediction on two tasks. Through two challenging case studies, we also show that customization via ProteinTTT achieves more accurate antibody-antigen loop modeling and enhances 19% of structures in the Big Fantastic Virus Database, delivering improved predictions where general-purpose AlphaFold2 and ESMFold struggle.

Antoine Yang, Antoine Miech, Josef Sivic et al.

We consider the problem of localizing a spatio-temporal tube in a video corresponding to a given text query. This is a challenging task that requires the joint and efficient modeling of temporal, spatial and multi-modal interactions. To address this task, we propose TubeDETR, a transformer-based architecture inspired by the recent success of such models for text-conditioned object detection. Our model notably includes: (i) an efficient video and text encoder that models spatial multi-modal interactions over sparsely sampled frames and (ii) a space-time decoder that jointly performs spatio-temporal localization. We demonstrate the advantage of our proposed components through an extensive ablation study. We also evaluate our full approach on the spatio-temporal video grounding task and demonstrate improvements over the state of the art on the challenging VidSTG and HC-STVG benchmarks. Code and trained models are publicly available at https://antoyang.github.io/tubedetr.html.

Kateryna Zorina, Justin Carpentier, Josef Sivic et al.

A seamless integration of robots into human environments requires robots to learn how to use existing human tools. Current approaches for learning tool manipulation skills mostly rely on expert demonstrations provided in the target robot environment, for example, by manually guiding the robot manipulator or by teleoperation. In this work, we introduce an automated approach that replaces an expert demonstration with a Youtube video for learning a tool manipulation strategy. The main contributions are twofold. First, we design an alignment procedure that aligns the simulated environment with the real-world scene observed in the video. This is formulated as an optimization problem that finds a spatial alignment of the tool trajectory to maximize the sparse goal reward given by the environment. Second, we describe an imitation learning approach that focuses on the trajectory of the tool rather than the motion of the human. For this we combine reinforcement learning with an optimization procedure to find a control policy and the placement of the robot based on the tool motion in the aligned environment. We demonstrate the proposed approach on spade, scythe and hammer tools in simulation, and show the effectiveness of the trained policy for the spade on a real Franka Emika Panda robot demonstration.

Zongmian Li, Jiri Sedlar, Justin Carpentier et al.

In this paper, we introduce a method to automatically reconstruct the 3D motion of a person interacting with an object from a single RGB video. Our method estimates the 3D poses of the person together with the object pose, the contact positions and the contact forces exerted on the human body. The main contributions of this work are three-fold. First, we introduce an approach to jointly estimate the motion and the actuation forces of the person on the manipulated object by modeling contacts and the dynamics of the interactions. This is cast as a large-scale trajectory optimization problem. Second, we develop a method to automatically recognize from the input video the 2D position and timing of contacts between the person and the object or the ground, thereby significantly simplifying the complexity of the optimization. Third, we validate our approach on a recent video+MoCap dataset capturing typical parkour actions, and demonstrate its performance on a new dataset of Internet videos showing people manipulating a variety of tools in unconstrained environments.

Shuang Li, Yilun Du, Antonio Torralba et al.

We introduce the task of weakly supervised learning for detecting human and object interactions in videos. Our task poses unique challenges as a system does not know what types of human-object interactions are present in a video or the actual spatiotemporal location of the human and the object. To address these challenges, we introduce a contrastive weakly supervised training loss that aims to jointly associate spatiotemporal regions in a video with an action and object vocabulary and encourage temporal continuity of the visual appearance of moving objects as a form of self-supervision. To train our model, we introduce a dataset comprising over 6.5k videos with human-object interaction annotations that have been semi-automatically curated from sentence captions associated with the videos. We demonstrate improved performance over weakly supervised baselines adapted to our task on our video dataset.

Dimitri Zhukov, Ignacio Rocco, Ivan Laptev et al.

Narrated instructional videos often show and describe manipulations of similar objects, e.g., repairing a particular model of a car or laptop. In this work we aim to reconstruct such objects and to localize associated narrations in 3D. Contrary to the standard scenario of instance-level 3D reconstruction, where identical objects or scenes are present in all views, objects in different instructional videos may have large appearance variations given varying conditions and versions of the same product. Narrations may also have large variation in natural language expressions. We address these challenges by three contributions. First, we propose an approach for correspondence estimation combining learnt local features and dense flow. Second, we design a two-step divide and conquer reconstruction approach where the initial 3D reconstructions of individual videos are combined into a 3D alignment graph. Finally, we propose an unsupervised approach to ground natural language in obtained 3D reconstructions. We demonstrate the effectiveness of our approach for the domain of car maintenance. Given raw instructional videos and no manual supervision, our method successfully reconstructs engines of different car models and associates textual descriptions with corresponding objects in 3D.

J. Macke, J. Sedlar, M. Olsak et al.

We describe a purely image-based method for finding geometric constructions with a ruler and compass in the Euclidea geometric game. The method is based on adapting the Mask R-CNN state-of-the-art image processing neural architecture and adding a tree-based search procedure to it. In a supervised setting, the method learns to solve all 68 kinds of geometric construction problems from the first six level packs of Euclidea with an average 92% accuracy. When evaluated on new kinds of problems, the method can solve 31 of the 68 kinds of Euclidea problems. We believe that this is the first time that a purely image-based learning has been trained to solve geometric construction problems of this difficulty.

Yann Labbé, Justin Carpentier, Mathieu Aubry et al.

We introduce RoboPose, a method to estimate the joint angles and the 6D camera-to-robot pose of a known articulated robot from a single RGB image. This is an important problem to grant mobile and itinerant autonomous systems the ability to interact with other robots using only visual information in non-instrumented environments, especially in the context of collaborative robotics. It is also challenging because robots have many degrees of freedom and an infinite space of possible configurations that often result in self-occlusions and depth ambiguities when imaged by a single camera. The contributions of this work are three-fold. First, we introduce a new render & compare approach for estimating the 6D pose and joint angles of an articulated robot that can be trained from synthetic data, generalizes to new unseen robot configurations at test time, and can be applied to a variety of robots. Second, we experimentally demonstrate the importance of the robot parametrization for the iterative pose updates and design a parametrization strategy that is independent of the robot structure. Finally, we show experimental results on existing benchmark datasets for four different robots and demonstrate that our method significantly outperforms the state of the art. Code and pre-trained models are available on the project webpage https://www.di.ens.fr/willow/research/robopose/.

Hugo Cisneros, Josef Sivic, Tomas Mikolov

Emergent processes in complex systems such as cellular automata can perform computations of increasing complexity, and could possibly lead to artificial evolution. Such a feat would require scaling up current simulation sizes to allow for enough computational capacity. Understanding complex computations happening in cellular automata and other systems capable of emergence poses many challenges, especially in large-scale systems. We propose methods for coarse-graining cellular automata based on frequency analysis of cell states, clustering and autoencoders. These innovative techniques facilitate the discovery of large-scale structure formation and complexity analysis in those systems. They emphasize interesting behaviors in elementary cellular automata while filtering out background patterns. Moreover, our methods reduce large 2D automata to smaller sizes and enable identifying systems that behave interestingly at multiple scales.

Antoine Miech, Jean-Baptiste Alayrac, Ivan Laptev et al.

Our objective is language-based search of large-scale image and video datasets. For this task, the approach that consists of independently mapping text and vision to a joint embedding space, a.k.a. dual encoders, is attractive as retrieval scales and is efficient for billions of images using approximate nearest neighbour search. An alternative approach of using vision-text transformers with cross-attention gives considerable improvements in accuracy over the joint embeddings, but is often inapplicable in practice for large-scale retrieval given the cost of the cross-attention mechanisms required for each sample at test time. This work combines the best of both worlds. We make the following three contributions. First, we equip transformer-based models with a new fine-grained cross-attention architecture, providing significant improvements in retrieval accuracy whilst preserving scalability. Second, we introduce a generic approach for combining a Fast dual encoder model with our Slow but accurate transformer-based model via distillation and re-ranking. Finally, we validate our approach on the Flickr30K image dataset where we show an increase in inference speed by several orders of magnitude while having results competitive to the state of the art. We also extend our method to the video domain, improving the state of the art on the VATEX dataset.

Antonín Vobecký, David Hurych, Michal Uřičář et al.

Existing datasets for training pedestrian detectors in images suffer from limited appearance and pose variation. The most challenging scenarios are rarely included because they are too difficult to capture due to safety reasons, or they are very unlikely to happen. The strict safety requirements in assisted and autonomous driving applications call for an extra high detection accuracy also in these rare situations. Having the ability to generate people images in arbitrary poses, with arbitrary appearances and embedded in different background scenes with varying illumination and weather conditions, is a crucial component for the development and testing of such applications. The contributions of this paper are three-fold. First, we describe an augmentation method for controlled synthesis of urban scenes containing people, thus producing rare or never-seen situations. This is achieved with a data generator (called DummyNet) with disentangled control of the pose, the appearance, and the target background scene. Second, the proposed generator relies on novel network architecture and associated loss that takes into account the segmentation of the foreground person and its composition into the background scene. Finally, we demonstrate that the data generated by our DummyNet improve performance of several existing person detectors across various datasets as well as in challenging situations, such as night-time conditions, where only a limited amount of training data is available. In the setup with only day-time data available, we improve the night-time detector by $17\%$ log-average miss rate over the detector trained with the day-time data only.

Antoine Yang, Antoine Miech, Josef Sivic et al.

Recent methods for visual question answering rely on large-scale annotated datasets. Manual annotation of questions and answers for videos, however, is tedious, expensive and prevents scalability. In this work, we propose to avoid manual annotation and generate a large-scale training dataset for video question answering making use of automatic cross-modal supervision. We leverage a question generation transformer trained on text data and use it to generate question-answer pairs from transcribed video narrations. Given narrated videos, we then automatically generate the HowToVQA69M dataset with 69M video-question-answer triplets. To handle the open vocabulary of diverse answers in this dataset, we propose a training procedure based on a contrastive loss between a video-question multi-modal transformer and an answer transformer. We introduce the zero-shot VideoQA task and show excellent results, in particular for rare answers. Furthermore, we demonstrate our method to significantly outperform the state of the art on MSRVTT-QA, MSVD-QA, ActivityNet-QA and How2QA. Finally, for a detailed evaluation we introduce iVQA, a new VideoQA dataset with reduced language biases and high-quality redundant manual annotations. Our code, datasets and trained models are available at https://antoyang.github.io/just-ask.html.

Vladimír Petrík, Makarand Tapaswi, Ivan Laptev et al.

Humans are adept at learning new tasks by watching a few instructional videos. On the other hand, robots that learn new actions either require a lot of effort through trial and error, or use expert demonstrations that are challenging to obtain. In this paper, we explore a method that facilitates learning object manipulation skills directly from videos. Leveraging recent advances in 2D visual recognition and differentiable rendering, we develop an optimization based method to estimate a coarse 3D state representation for the hand and the manipulated object(s) without requiring any supervision. We use these trajectories as dense rewards for an agent that learns to mimic them through reinforcement learning. We evaluate our method on simple single- and two-object actions from the Something-Something dataset. Our approach allows an agent to learn actions from single videos, while watching multiple demonstrations makes the policy more robust. We show that policies learned in a simulated environment can be easily transferred to a real robot.

Yann Labbé, Justin Carpentier, Mathieu Aubry et al.

We introduce an approach for recovering the 6D pose of multiple known objects in a scene captured by a set of input images with unknown camera viewpoints. First, we present a single-view single-object 6D pose estimation method, which we use to generate 6D object pose hypotheses. Second, we develop a robust method for matching individual 6D object pose hypotheses across different input images in order to jointly estimate camera viewpoints and 6D poses of all objects in a single consistent scene. Our approach explicitly handles object symmetries, does not require depth measurements, is robust to missing or incorrect object hypotheses, and automatically recovers the number of objects in the scene. Third, we develop a method for global scene refinement given multiple object hypotheses and their correspondences across views. This is achieved by solving an object-level bundle adjustment problem that refines the poses of cameras and objects to minimize the reprojection error in all views. We demonstrate that the proposed method, dubbed CosyPose, outperforms current state-of-the-art results for single-view and multi-view 6D object pose estimation by a large margin on two challenging benchmarks: the YCB-Video and T-LESS datasets. Code and pre-trained models are available on the project webpage https://www.di.ens.fr/willow/research/cosypose/.

Antoine Miech, Jean-Baptiste Alayrac, Ivan Laptev et al.

This paper introduces a manually annotated video dataset of unusual actions, namely RareAct, including actions such as "blend phone", "cut keyboard" and "microwave shoes". RareAct aims at evaluating the zero-shot and few-shot compositionality of action recognition models for unlikely compositions of common action verbs and object nouns. It contains 122 different actions which were obtained by combining verbs and nouns rarely co-occurring together in the large-scale textual corpus from HowTo100M, but that frequently appear separately. We provide benchmarks using a state-of-the-art HowTo100M pretrained video and text model and show that zero-shot and few-shot compositionality of actions remains a challenging and unsolved task.

Ronan Riochet, Josef Sivic, Ivan Laptev et al.

To reach human performance on complex tasks, a key ability for artificial systems is to understand physical interactions between objects, and predict future outcomes of a situation. This ability, often referred to as intuitive physics, has recently received attention and several methods were proposed to learn these physical rules from video sequences. Yet, most of these methods are restricted to the case where no, or only limited, occlusions occur. In this work we propose a probabilistic formulation of learning intuitive physics in 3D scenes with significant inter-object occlusions. In our formulation, object positions are modeled as latent variables enabling the reconstruction of the scene. We then propose a series of approximations that make this problem tractable. Object proposals are linked across frames using a combination of a recurrent interaction network, modeling the physics in object space, and a compositional renderer, modeling the way in which objects project onto pixel space. We demonstrate significant improvements over state-of-the-art in the intuitive physics benchmark of IntPhys. We apply our method to a second dataset with increasing levels of occlusions, showing it realistically predicts segmentation masks up to 30 frames in the future. Finally, we also show results on predicting motion of objects in real videos.

Ignacio Rocco, Relja Arandjelović, Josef Sivic

In this work we target the problem of estimating accurately localised correspondences between a pair of images. We adopt the recent Neighbourhood Consensus Networks that have demonstrated promising performance for difficult correspondence problems and propose modifications to overcome their main limitations: large memory consumption, large inference time and poorly localised correspondences. Our proposed modifications can reduce the memory footprint and execution time more than $10\times$, with equivalent results. This is achieved by sparsifying the correlation tensor containing tentative matches, and its subsequent processing with a 4D CNN using submanifold sparse convolutions. Localisation accuracy is significantly improved by processing the input images in higher resolution, which is possible due to the reduced memory footprint, and by a novel two-stage correspondence relocalisation module. The proposed Sparse-NCNet method obtains state-of-the-art results on the HPatches Sequences and InLoc visual localisation benchmarks, and competitive results in the Aachen Day-Night benchmark.

Antoine Miech, Jean-Baptiste Alayrac, Lucas Smaira et al.

Annotating videos is cumbersome, expensive and not scalable. Yet, many strong video models still rely on manually annotated data. With the recent introduction of the HowTo100M dataset, narrated videos now offer the possibility of learning video representations without manual supervision. In this work we propose a new learning approach, MIL-NCE, capable of addressing misalignments inherent to narrated videos. With this approach we are able to learn strong video representations from scratch, without the need for any manual annotation. We evaluate our representations on a wide range of four downstream tasks over eight datasets: action recognition (HMDB-51, UCF-101, Kinetics-700), text-to-video retrieval (YouCook2, MSR-VTT), action localization (YouTube-8M Segments, CrossTask) and action segmentation (COIN). Our method outperforms all published self-supervised approaches for these tasks as well as several fully supervised baselines.

Hugo Cisneros, Josef Sivic, Tomas Mikolov

In this paper we propose an approach for measuring growth of complexity of emerging patterns in complex systems such as cellular automata. We discuss several ways how a metric for measuring the complexity growth can be defined. This includes approaches based on compression algorithms and artificial neural networks. We believe such a metric can be useful for designing systems that could exhibit open-ended evolution, which itself might be a prerequisite for development of general artificial intelligence. We conduct experiments on 1D and 2D grid worlds and demonstrate that using the proposed metric we can automatically construct computational models with emerging properties similar to those found in the Conway's Game of Life, as well as many other emergent phenomena. Interestingly, some of the patterns we observe resemble forms of artificial life. Our metric of structural complexity growth can be applied to a wide range of complex systems, as it is not limited to cellular automata.

Hajime Taira, Ignacio Rocco, Jiri Sedlar et al.

Visual localization in large and complex indoor scenes, dominated by weakly textured rooms and repeating geometric patterns, is a challenging problem with high practical relevance for applications such as Augmented Reality and robotics. To handle the ambiguities arising in this scenario, a common strategy is, first, to generate multiple estimates for the camera pose from which a given query image was taken. The pose with the largest geometric consistency with the query image, e.g., in the form of an inlier count, is then selected in a second stage. While a significant amount of research has concentrated on the first stage, there is considerably less work on the second stage. In this paper, we thus focus on pose verification. We show that combining different modalities, namely appearance, geometry, and semantics, considerably boosts pose verification and consequently pose accuracy. We develop multiple hand-crafted as well as a trainable approach to join into the geometric-semantic verification and show significant improvements over state-of-the-art on a very challenging indoor dataset.

Robin Strudel, Alexander Pashevich, Igor Kalevatykh et al.

Manipulation tasks such as preparing a meal or assembling furniture remain highly challenging for robotics and vision. Traditional task and motion planning (TAMP) methods can solve complex tasks but require full state observability and are not adapted to dynamic scene changes. Recent learning methods can operate directly on visual inputs but typically require many demonstrations and/or task-specific reward engineering. In this work we aim to overcome previous limitations and propose a reinforcement learning (RL) approach to task planning that learns to combine primitive skills. First, compared to previous learning methods, our approach requires neither intermediate rewards nor complete task demonstrations during training. Second, we demonstrate the versatility of our vision-based task planning in challenging settings with temporary occlusions and dynamic scene changes. Third, we propose an efficient training of basic skills from few synthetic demonstrations by exploring recent CNN architectures and data augmentation. Notably, while all of our policies are learned on visual inputs in simulated environments, we demonstrate the successful transfer and high success rates when applying such policies to manipulation tasks on a real UR5 robotic arm.