Jüergen Gall

Nadine Behrmann, S. Alireza Golestaneh, Zico Kolter et al.

This paper introduces a unified framework for video action segmentation via sequence to sequence (seq2seq) translation in a fully and timestamp supervised setup. In contrast to current state-of-the-art frame-level prediction methods, we view action segmentation as a seq2seq translation task, i.e., mapping a sequence of video frames to a sequence of action segments. Our proposed method involves a series of modifications and auxiliary loss functions on the standard Transformer seq2seq translation model to cope with long input sequences opposed to short output sequences and relatively few videos. We incorporate an auxiliary supervision signal for the encoder via a frame-wise loss and propose a separate alignment decoder for an implicit duration prediction. Finally, we extend our framework to the timestamp supervised setting via our proposed constrained k-medoids algorithm to generate pseudo-segmentations. Our proposed framework performs consistently on both fully and timestamp supervised settings, outperforming or competing state-of-the-art on several datasets. Our code is publicly available at https://github.com/boschresearch/UVAST.

Shuxiao Ding, Eike Rehder, Lukas Schneider et al.

Tracking 3D objects accurately and consistently is crucial for autonomous vehicles, enabling more reliable downstream tasks such as trajectory prediction and motion planning. Based on the substantial progress in object detection in recent years, the tracking-by-detection paradigm has become a popular choice due to its simplicity and efficiency. State-of-the-art 3D multi-object tracking (MOT) approaches typically rely on non-learned model-based algorithms such as Kalman Filter but require many manually tuned parameters. On the other hand, learning-based approaches face the problem of adapting the training to the online setting, leading to inevitable distribution mismatch between training and inference as well as suboptimal performance. In this work, we propose 3DMOTFormer, a learned geometry-based 3D MOT framework building upon the transformer architecture. We use an Edge-Augmented Graph Transformer to reason on the track-detection bipartite graph frame-by-frame and conduct data association via edge classification. To reduce the distribution mismatch between training and inference, we propose a novel online training strategy with an autoregressive and recurrent forward pass as well as sequential batch optimization. Using CenterPoint detections, our approach achieves 71.2% and 68.2% AMOTA on the nuScenes validation and test split, respectively. In addition, a trained 3DMOTFormer model generalizes well across different object detectors. Code is available at: https://github.com/dsx0511/3DMOTFormer.

Peizheng Li, Shuxiao Ding, Xieyuanli Chen et al.

Accurately perceiving instances and predicting their future motion are key tasks for autonomous vehicles, enabling them to navigate safely in complex urban traffic. While bird's-eye view (BEV) representations are commonplace in perception for autonomous driving, their potential in a motion prediction setting is less explored. Existing approaches for BEV instance prediction from surround cameras rely on a multi-task auto-regressive setup coupled with complex post-processing to predict future instances in a spatio-temporally consistent manner. In this paper, we depart from this paradigm and propose an efficient novel end-to-end framework named POWERBEV, which differs in several design choices aimed at reducing the inherent redundancy in previous methods. First, rather than predicting the future in an auto-regressive fashion, POWERBEV uses a parallel, multi-scale module built from lightweight 2D convolutional networks. Second, we show that segmentation and centripetal backward flow are sufficient for prediction, simplifying previous multi-task objectives by eliminating redundant output modalities. Building on this output representation, we propose a simple, flow warping-based post-processing approach which produces more stable instance associations across time. Through this lightweight yet powerful design, POWERBEV outperforms state-of-the-art baselines on the NuScenes Dataset and poses an alternative paradigm for BEV instance prediction. We made our code publicly available at: https://github.com/EdwardLeeLPZ/PowerBEV.

Mohamad Hakam Shams Eddin, Ribana Roscher, Juergen Gall

Climate change is expected to intensify and increase extreme events in the weather cycle. Since this has a significant impact on various sectors of our life, recent works are concerned with identifying and predicting such extreme events from Earth observations. With respect to wildfire danger forecasting, previous deep learning approaches duplicate static variables along the time dimension and neglect the intrinsic differences between static and dynamic variables. Furthermore, most existing multi-branch architectures lose the interconnections between the branches during the feature learning stage. To address these issues, this paper proposes a 2D/3D two-branch convolutional neural network (CNN) with a Location-aware Adaptive Normalization layer (LOAN). Using LOAN as a building block, we can modulate the dynamic features conditional on their geographical locations. Thus, our approach considers feature properties as a unified yet compound 2D/3D model. Besides, we propose using the sinusoidal-based encoding of the day of the year to provide the model with explicit temporal information about the target day within the year. Our experimental results show a better performance of our approach than other baselines on the challenging FireCube dataset. The results show that location-aware adaptive feature normalization is a promising technique to learn the relation between dynamic variables and their geographic locations, which is highly relevant for areas where remote sensing data builds the basis for analysis. The source code is available at https://github.com/HakamShams/LOAN.

Vadim Sushko, Dan Zhang, Juergen Gall et al.

Joint synthesis of images and segmentation masks with generative adversarial networks (GANs) is promising to reduce the effort needed for collecting image data with pixel-wise annotations. However, to learn high-fidelity image-mask synthesis, existing GAN approaches first need a pre-training phase requiring large amounts of image data, which limits their utilization in restricted image domains. In this work, we take a step to reduce this limitation, introducing the task of one-shot image-mask synthesis. We aim to generate diverse images and their segmentation masks given only a single labelled example, and assuming, contrary to previous models, no access to any pre-training data. To this end, inspired by the recent architectural developments of single-image GANs, we introduce our OSMIS model which enables the synthesis of segmentation masks that are precisely aligned to the generated images in the one-shot regime. Besides achieving the high fidelity of generated masks, OSMIS outperforms state-of-the-art single-image GAN models in image synthesis quality and diversity. In addition, despite not using any additional data, OSMIS demonstrates an impressive ability to serve as a source of useful data augmentation for one-shot segmentation applications, providing performance gains that are complementary to standard data augmentation techniques. Code is available at https://github.com/ boschresearch/one-shot-synthesis

Pascal Herrmann, Maarten Bieshaar, Dennis Mack et al.

Human motion generation has made tremendous progress in recent years, with state-of-the-art approaches surpassing ground truth data in leading evaluation benchmarks. However, visual inspection of the generated motions paints a different picture. Even state-of-the-art approaches generate motions frequently containing self-intersections, i.e., body parts interpenetrating, which are strong artifacts, severely limiting the perceived motion quality. We introduce a novel loss, which explicitly penalizes self-intersections, to the training of human motion generation methods. We base our loss on a sphere proxy of human geometry, which allows us to calculate a self-intersection loss 98% faster and uses 83% less memory than comparable methods based on triangular meshes. The loss is agnostic to the specific approach, and we add it to the training of the recent human motion generation methods human motion diffusion model (MDM) and MoMask. Our extensive experiments show a reduction of self-intersections in generated motions of up to 49% while improving other evaluation metrics. The code is available at https://github.com/boschresearch/humansphereproxy .

Zeyun Zhong, Manuel Martin, Chengzhi Wu et al.

Recent Large Multimodal Models (LMMs), primarily designed for offline settings, are ill-suited for the dynamic requirements of streaming video. While recent online adaptations improve real-time processing, they still face critical scalability challenges, with resource demands typically growing at least linearly with video duration. To overcome this bottleneck, we propose FlowNar, a novel framework for scalable streaming video narration. The core of FlowNar is a dynamic context management strategy for historical visual context removal, combined with our CLAM (Cross Linear Attentive Memory) module for streaming visual history retention, ensuring bounded visual memory usage and computational complexity, crucial for efficient streaming. We also introduce a realistic self-conditioned evaluation protocol and complementary evaluation metrics to assess streaming narration models under deployment-like conditions. Experiments on the Ego4D, EgoExo4D, and EpicKitchens100 datasets demonstrate that FlowNar substantially improves narration quality over strong baselines while being highly efficient, supporting processing of 10$\times$ longer videos and achieving 3$\times$ higher throughput (FPS). The code is available at https://github.com/zeyun-zhong/FlowNar.

Robin Burchard, Pascal-André Brückner, Marius Bock et al.

With each sensing modality exhibiting inherent strengths and limitations, multi-modal approaches for wearable Human Activity Recognition (HAR) are becoming increasingly relevant -- particularly for recognizing Activities of Daily Living (ADLs), where individual modalities often produce ambiguous signals for similar or complex activities. This work introduces HARMES, a multi-modal wearable dataset combining three wrist-recorded modalities: motion sensing via an Inertial Measurement Unit (IMU), atmospheric environmental sensors (humidity, temperature, and pressure), and audio. Collected from 20 participants performing household activities in their own homes, HARMES totals over 80 hours of recorded data, with approximately three hours of labeled activity data per participant across 15 ADL classes. To the best of our knowledge, HARMES is the first dataset to combine this particular sensor trio, and it is nearly six times larger than the previously largest wrist-inertial-acoustic HAR dataset. In an extensive benchmark, we evaluate cross-subject generalization and conduct an ablation study revealing that modality contributions are activity-dependent and can provide complementary value, particularly for activities that are ambiguous from motion data alone. HARMES is freely available at Zenodo, alongside example code for loading the dataset and training models on GitHub.

Emad Bahrami, Gianpiero Francesca, Juergen Gall

Modeling long-term context in videos is crucial for many fine-grained tasks including temporal action segmentation. An interesting question that is still open is how much long-term temporal context is needed for optimal performance. While transformers can model the long-term context of a video, this becomes computationally prohibitive for long videos. Recent works on temporal action segmentation thus combine temporal convolutional networks with self-attentions that are computed only for a local temporal window. While these approaches show good results, their performance is limited by their inability to capture the full context of a video. In this work, we try to answer how much long-term temporal context is required for temporal action segmentation by introducing a transformer-based model that leverages sparse attention to capture the full context of a video. We compare our model with the current state of the art on three datasets for temporal action segmentation, namely 50Salads, Breakfast, and Assembly101. Our experiments show that modeling the full context of a video is necessary to obtain the best performance for temporal action segmentation.

Jinhui Yi, Yanan Luo, Marion Deichmann et al.

An early, non-invasive, and on-site detection of nutrient deficiencies is critical to enable timely actions to prevent major losses of crops caused by lack of nutrients. While acquiring labeled data is very expensive, collecting images from multiple views of a crop is straightforward. Despite its relevance for practical applications, unsupervised domain adaptation where multiple views are available for the labeled source domain as well as the unlabeled target domain is an unexplored research area. In this work, we thus propose an approach that leverages multiple camera views in the source and target domain for unsupervised domain adaptation. We evaluate the proposed approach on two nutrient deficiency datasets. The proposed method achieves state-of-the-art results on both datasets compared to other unsupervised domain adaptation methods. The dataset and source code are available at https://github.com/jh-yi/MV-Match.

Olga Zatsarynna, Juergen Gall

In this paper, we address the problem of short-term action anticipation, i.e., we want to predict an upcoming action one second before it happens. We propose to harness high-level intent information to anticipate actions that will take place in the future. To this end, we incorporate an additional goal prediction branch into our model and propose a consistency loss function that encourages the anticipated actions to conform to the high-level goal pursued in the video. In our experiments, we show the effectiveness of the proposed approach and demonstrate that our method achieves state-of-the-art results on two large-scale datasets: Assembly101 and COIN.

Abdelrahman Shaker, Syed Talal Wasim, Salman Khan et al.

State-space models (SSMs) have recently shown promise in capturing long-range dependencies with subquadratic computational complexity, making them attractive for various applications. However, purely SSM-based models face critical challenges related to stability and achieving state-of-the-art performance in computer vision tasks. Our paper addresses the challenges of scaling SSM-based models for computer vision, particularly the instability and inefficiency of large model sizes. We introduce a parameter-efficient modulated group mamba layer that divides the input channels into four groups and applies our proposed SSM-based efficient Visual Single Selective Scanning (VSSS) block independently to each group, with each VSSS block scanning in one of the four spatial directions. The Modulated Group Mamba layer also wraps the four VSSS blocks into a channel modulation operator to improve cross-channel communication. Furthermore, we introduce a distillation-based training objective to stabilize the training of large models, leading to consistent performance gains. Our comprehensive experiments demonstrate the merits of the proposed contributions, leading to superior performance over existing methods for image classification on ImageNet-1K, object detection, instance segmentation on MS-COCO, and semantic segmentation on ADE20K. Our tiny variant with 23M parameters achieves state-of-the-art performance with a classification top-1 accuracy of 83.3% on ImageNet-1K, while being 26% efficient in terms of parameters, compared to the best existing Mamba design of same model size. Code and models are available at: https://github.com/Amshaker/GroupMamba.

Shijie Li, Ming-Ming Cheng, Juergen Gall

The goal of semantic image synthesis is to generate photo-realistic images from semantic label maps. It is highly relevant for tasks like content generation and image editing. Current state-of-the-art approaches, however, still struggle to generate realistic objects in images at various scales. In particular, small objects tend to fade away and large objects are often generated as collages of patches. In order to address this issue, we propose a Dual Pyramid Generative Adversarial Network (DP-GAN) that learns the conditioning of spatially-adaptive normalization blocks at all scales jointly, such that scale information is bi-directionally used, and it unifies supervision at different scales. Our qualitative and quantitative results show that the proposed approach generates images where small and large objects look more realistic compared to images generated by state-of-the-art methods.

Olga Zatsarynna, Yazan Abu Farha, Juergen Gall

Distinguishing if an action is performed as intended or if an intended action fails is an important skill that not only humans have, but that is also important for intelligent systems that operate in human environments. Recognizing if an action is unintentional or anticipating if an action will fail, however, is not straightforward due to lack of annotated data. While videos of unintentional or failed actions can be found in the Internet in abundance, high annotation costs are a major bottleneck for learning networks for these tasks. In this work, we thus study the problem of self-supervised representation learning for unintentional action prediction. While previous works learn the representation based on a local temporal neighborhood, we show that the global context of a video is needed to learn a good representation for the three downstream tasks: unintentional action classification, localization and anticipation. In the supplementary material, we show that the learned representation can be used for detecting anomalies in videos as well.

Jifeng Wang, Kaouther Messaoud, Yuejiang Liu et al.

Recent progress in motion forecasting has been substantially driven by self-supervised pre-training. However, adapting pre-trained models for specific downstream tasks, especially motion prediction, through extensive fine-tuning is often inefficient. This inefficiency arises because motion prediction closely aligns with the masked pre-training tasks, and traditional full fine-tuning methods fail to fully leverage this alignment. To address this, we introduce Forecast-PEFT, a fine-tuning strategy that freezes the majority of the model's parameters, focusing adjustments on newly introduced prompts and adapters. This approach not only preserves the pre-learned representations but also significantly reduces the number of parameters that need retraining, thereby enhancing efficiency. This tailored strategy, supplemented by our method's capability to efficiently adapt to different datasets, enhances model efficiency and ensures robust performance across datasets without the need for extensive retraining. Our experiments show that Forecast-PEFT outperforms traditional full fine-tuning methods in motion prediction tasks, achieving higher accuracy with only 17% of the trainable parameters typically required. Moreover, our comprehensive adaptation, Forecast-FT, further improves prediction performance, evidencing up to a 9.6% enhancement over conventional baseline methods. Code will be available at https://github.com/csjfwang/Forecast-PEFT.

Yaser Souri, Yazan Abu Farha, Emad Bahrami et al.

Action segmentation is the task of predicting an action label for each frame of an untrimmed video. As obtaining annotations to train an approach for action segmentation in a fully supervised way is expensive, various approaches have been proposed to train action segmentation models using different forms of weak supervision, e.g., action transcripts, action sets, or more recently timestamps. Timestamp supervision is a promising type of weak supervision as obtaining one timestamp per action is less expensive than annotating all frames, but it provides more information than other forms of weak supervision. However, previous works assume that every action instance is annotated with a timestamp, which is a restrictive assumption since it assumes that annotators do not miss any action. In this work, we relax this restrictive assumption and take missing annotations for some action instances into account. We show that our approach is more robust to missing annotations compared to other approaches and various baselines.

Olga Zatsarynna, Emad Bahrami, Yazan Abu Farha et al.

Long-term action anticipation has become an important task for many applications such as autonomous driving and human-robot interaction. Unlike short-term anticipation, predicting more actions into the future imposes a real challenge with the increasing uncertainty in longer horizons. While there has been a significant progress in predicting more actions into the future, most of the proposed methods address the task in a deterministic setup and ignore the underlying uncertainty. In this paper, we propose a novel Gated Temporal Diffusion (GTD) network that models the uncertainty of both the observation and the future predictions. As generator, we introduce a Gated Anticipation Network (GTAN) to model both observed and unobserved frames of a video in a mutual representation. On the one hand, using a mutual representation for past and future allows us to jointly model ambiguities in the observation and future, while on the other hand GTAN can by design treat the observed and unobserved parts differently and steer the information flow between them. Our model achieves state-of-the-art results on the Breakfast, Assembly101 and 50Salads datasets in both stochastic and deterministic settings. Code: https://github.com/olga-zats/GTDA .

Rong Li, ShiJie Li, Xieyuanli Chen et al.

LiDAR semantic segmentation plays a crucial role in enabling autonomous driving and robots to understand their surroundings accurately and robustly. A multitude of methods exist within this domain, including point-based, range-image-based, polar-coordinate-based, and hybrid strategies. Among these, range-image-based techniques have gained widespread adoption in practical applications due to their efficiency. However, they face a significant challenge known as the ``many-to-one'' problem caused by the range image's limited horizontal and vertical angular resolution. As a result, around 20% of the 3D points can be occluded. In this paper, we present TFNet, a range-image-based LiDAR semantic segmentation method that utilizes temporal information to address this issue. Specifically, we incorporate a temporal fusion layer to extract useful information from previous scans and integrate it with the current scan. We then design a max-voting-based post-processing technique to correct false predictions, particularly those caused by the ``many-to-one'' issue. We evaluated the approach on two benchmarks and demonstrated that the plug-in post-processing technique is generic and can be applied to various networks.

Vadim Sushko, Ruyu Wang, Juergen Gall

The task of few-shot GAN adaptation aims to adapt a pre-trained GAN model to a small dataset with very few training images. While existing methods perform well when the dataset for pre-training is structurally similar to the target dataset, the approaches suffer from training instabilities or memorization issues when the objects in the two domains have a very different structure. To mitigate this limitation, we propose a new smoothness similarity regularization that transfers the inherently learned smoothness of the pre-trained GAN to the few-shot target domain even if the two domains are very different. We evaluate our approach by adapting an unconditional and a class-conditional GAN to diverse few-shot target domains. Our proposed method significantly outperforms prior few-shot GAN adaptation methods in the challenging case of structurally dissimilar source-target domains, while performing on par with the state of the art for similar source-target domains.

Andreas Doering, Juergen Gall

Multi-person pose tracking is an important element for many applications and requires to estimate the human poses of all persons in a video and to track them over time. The association of poses across frames remains an open research problem, in particular for online tracking methods, due to motion blur, crowded scenes and occlusions. To tackle the association challenge, we propose a Gated Attention Transformer. The core aspect of our model is the gating mechanism that automatically adapts the impact of appearance embeddings and embeddings based on temporal pose similarity in the attention layers. In order to re-identify persons that have been occluded, we incorporate a pose-conditioned re-identification network that provides initial embeddings and allows to match persons even if the number of visible joints differ between frames. We further propose a matching layer based on gated attention for pose-to-track association and duplicate removal. We evaluate our approach on PoseTrack 2018 and PoseTrack21.

Shijie Li, Rong Li, Juergen Gall

Collecting diverse sets of training images for RGB-D semantic image segmentation is not always possible. In particular, when robots need to operate in privacy-sensitive areas like homes, the collection is often limited to a small set of locations. As a consequence, the annotated images lack diversity in appearance and approaches for RGB-D semantic image segmentation tend to overfit the training data. In this paper, we thus introduce semantic RGB-D image synthesis to address this problem. It requires synthesising a realistic-looking RGB-D image for a given semantic label map. Current approaches, however, are uni-modal and cannot cope with multi-modal data. Indeed, we show that extending uni-modal approaches to multi-modal data does not perform well. In this paper, we therefore propose a generator for multi-modal data that separates modal-independent information of the semantic layout from the modal-dependent information that is needed to generate an RGB and a depth image, respectively. Furthermore, we propose a discriminator that ensures semantic consistency between the label maps and the generated images and perceptual similarity between the real and generated images. Our comprehensive experiments demonstrate that the proposed method outperforms previous uni-modal methods by a large margin and that the accuracy of an approach for RGB-D semantic segmentation can be significantly improved by mixing real and generated images during training.

Felix B Mueller, Julian Tanke, Juergen Gall

Forecasting long-term 3D human motion is challenging: the stochasticity of human behavior makes it hard to generate realistic human motion from the input sequence alone. Information on the scene environment and the motion of nearby people can greatly aid the generation process. We propose a scene-aware social transformer model (SAST) to forecast long-term (10s) human motion motion. Unlike previous models, our approach can model interactions between both widely varying numbers of people and objects in a scene. We combine a temporal convolutional encoder-decoder architecture with a Transformer-based bottleneck that allows us to efficiently combine motion and scene information. We model the conditional motion distribution using denoising diffusion models. We benchmark our approach on the Humans in Kitchens dataset, which contains 1 to 16 persons and 29 to 50 objects that are visible simultaneously. Our model outperforms other approaches in terms of realism and diversity on different metrics and in a user study. Code is available at https://github.com/felixbmuller/SAST.

Rania Briq, Chuhang Zou, Leonid Pishchulin et al.

We consider the problem of synthesizing multi-action human motion sequences of arbitrary lengths. Existing approaches have mastered motion sequence generation in single action scenarios, but fail to generalize to multi-action and arbitrary-length sequences. We fill this gap by proposing a novel efficient approach that leverages expressiveness of Recurrent Transformers and generative richness of conditional Variational Autoencoders. The proposed iterative approach is able to generate smooth and realistic human motion sequences with an arbitrary number of actions and frames while doing so in linear space and time. We train and evaluate the proposed approach on PROX and Charades datasets, where we augment PROX with ground-truth action labels and Charades with human mesh annotations. Experimental evaluation shows significant improvements in FID score and semantic consistency metrics compared to the state-of-the-art.

Shuai Li, Duc Manh Vu, Juergen Gall

This paper concerns the problem of unsupervised temporal action segmentation for long, untrimmed videos. Recent successful approaches follow a joint representation learning and clustering paradigm, where optimal transport (OT) is adopted to produce pseudo labels for learning frame representations. These approaches alternate between estimating pseudo labels using OT and optimizing the parameters with gradient descent during training, where OT is used for obtaining the final temporal action segmentation. A major limitation of these works is that they learn a deterministic embedding for frame representations. The iterative procedure between learning deterministic embeddings based on pseudo labels and estimating pseudo labels from the learned embedding can thus get quickly stuck in a local optimum. As an alternative, we thus propose to learn a probabilistic embedding for frame representations. The embeddings are modeled by Gaussian distributions and we sample from the distributions before estimating the pseudo labels. We evaluate our approach on several challenging temporal action segmentation datasets and achieve results comparable to, and in some cases, better than the state of the art. Compared to baselines with deterministic embeddings, our approach improves MoF up to 20.7\% and F1-score up to 19.0\%. Our code is available at https://github.com/derkbreeze/PEOT.

Michael Halstead, Esra Guclu, Mohamed Farag et al.

In this manuscript we release two datasets for visual sensing of tomato plants grown in commercial-like settings and acquired using a robot. The first is BUTom21 which consists of still images and manual annotations. The second is BUTom-ST21 which consists of video-based data and semi-automated annotations through AI-based methods, referred to as pseudo-labels. In both cases, we provide pixel-level labels for the ripeness of the fruit. The aim is to provide the research community a challenging set of real-world imagery to explore methods to sense and estimate the state of tomato plants and their fruit, which is an important horticultural crop. Importantly, the spatial-temporal dataset provides individual fruit count and ripeness information enabling researchers to push the boundaries of field-based phenotyping.

Zeyun Zhong, Manuel Martin, Michael Voit et al.

The ability to anticipate possible future human actions is essential for a wide range of applications, including autonomous driving and human-robot interaction. Consequently, numerous methods have been introduced for action anticipation in recent years, with deep learning-based approaches being particularly popular. In this work, we review the recent advances of action anticipation algorithms with a particular focus on daily-living scenarios. Additionally, we classify these methods according to their primary contributions and summarize them in tabular form, allowing readers to grasp the details at a glance. Furthermore, we delve into the common evaluation metrics and datasets used for action anticipation and provide future directions with systematical discussions.

Hamid Suleman, Syed Talal Wasim, Muzammal Naseer et al.

State-space models (SSMs), exemplified by S4, have introduced a novel context modeling method by integrating state-space techniques into deep learning. However, they struggle with global context modeling due to their data-independent matrices. The Mamba model addressed this with data-dependent variants via the S6 selective-scan algorithm, enhancing context modeling, especially for long sequences. However, Mamba-based architectures are difficult to scale with respect to the number of parameters, which is a major limitation for vision applications. This paper addresses the scalability issue of large SSMs for image classification and action recognition without requiring additional techniques like knowledge distillation. We analyze the distinct characteristics of Mamba-based and Attention-based models, proposing a Mamba-Attention interleaved architecture that enhances scalability, robustness, and performance. We demonstrate that the stable and efficient interleaved architecture resolves the scalability issue of Mamba-based architectures for images and videos and increases robustness to common artifacts like JPEG compression. Our thorough evaluation on the ImageNet-1K, Kinetics-400 and Something-Something-v2 benchmarks demonstrates that our approach improves the accuracy of state-of-the-art Mamba-based architectures by up to $+1.7$.

Zeyun Zhong, Chengzhi Wu, Manuel Martin et al.

Anticipating future actions is inherently uncertain. Given an observed video segment containing ongoing actions, multiple subsequent actions can plausibly follow. This uncertainty becomes even larger when predicting far into the future. However, the majority of existing action anticipation models adhere to a deterministic approach, neglecting to account for future uncertainties. In this work, we rethink action anticipation from a generative view, employing diffusion models to capture different possible future actions. In this framework, future actions are iteratively generated from standard Gaussian noise in the latent space, conditioned on the observed video, and subsequently transitioned into the action space. Extensive experiments on four benchmark datasets, i.e., Breakfast, 50Salads, EpicKitchens, and EGTEA Gaze+, are performed and the proposed method achieves superior or comparable results to state-of-the-art methods, showing the effectiveness of a generative approach for action anticipation. Our code and trained models will be published on GitHub.

Yanan Luo, Jinhui Yi, Yazan Abu Farha et al.

Counting repetitive actions in long untrimmed videos is a challenging task that has many applications such as rehabilitation. State-of-the-art methods predict action counts by first generating a temporal self-similarity matrix (TSM) from the sampled frames and then feeding the matrix to a predictor network. The self-similarity matrix, however, is not an optimal input to a network since it discards too much information from the frame-wise embeddings. We thus rethink how a TSM can be utilized for counting repetitive actions and propose a framework that learns embeddings and predicts action start probabilities at full temporal resolution. The number of repeated actions is then inferred from the action start probabilities. In contrast to current approaches that have the TSM as an intermediate representation, we propose a novel loss based on a generated reference TSM, which enforces that the self-similarity of the learned frame-wise embeddings is consistent with the self-similarity of repeated actions. The proposed framework achieves state-of-the-art results on three datasets, i.e., RepCount, UCFRep, and Countix.

Olga Zatsarynna, Emad Bahrami, Yazan Abu Farha et al.

Long-term dense action anticipation is very challenging since it requires predicting actions and their durations several minutes into the future based on provided video observations. To model the uncertainty of future outcomes, stochastic models predict several potential future action sequences for the same observation. Recent work has further proposed to incorporate uncertainty modelling for observed frames by simultaneously predicting per-frame past and future actions in a unified manner. While such joint modelling of actions is beneficial, it requires long-range temporal capabilities to connect events across distant past and future time points. However, the previous work struggles to achieve such a long-range understanding due to its limited and/or sparse receptive field. To alleviate this issue, we propose a novel MANTA (MAmba for ANTicipation) network. Our model enables effective long-term temporal modelling even for very long sequences while maintaining linear complexity in sequence length. We demonstrate that our approach achieves state-of-the-art results on three datasets - Breakfast, 50Salads, and Assembly101 - while also significantly improving computational and memory efficiency. Our code is available at https://github.com/olga-zats/DIFF_MANTA .

Jinhui Yi, Syed Talal Wasim, Yanan Luo et al.

We present an efficient encoder-free approach for video-language understanding that achieves competitive performance while significantly reducing computational overhead. Current video-language models typically rely on heavyweight image encoders (300M-1.1B parameters) or video encoders (1B-1.4B parameters), creating a substantial computational burden when processing multi-frame videos. Our method introduces a novel Spatio-Temporal Alignment Block (STAB) that directly processes video inputs without requiring pre-trained encoders while using only 45M parameters for visual processing - at least a 6.5$\times$ reduction compared to traditional approaches. The STAB architecture combines Local Spatio-Temporal Encoding for fine-grained feature extraction, efficient spatial downsampling through learned attention and separate mechanisms for modeling frame-level and video-level relationships. Our model achieves comparable or superior performance to encoder-based approaches for open-ended video question answering on standard benchmarks. The fine-grained video question-answering evaluation demonstrates our model's effectiveness, outperforming the encoder-based approaches Video-ChatGPT and Video-LLaVA in key aspects like correctness and temporal understanding. Extensive ablation studies validate our architectural choices and demonstrate the effectiveness of our spatio-temporal modeling approach while achieving 3-4$\times$ faster processing speeds than previous methods. Code is available at https://jh-yi.github.io/Video-Panda.

Lars Doorenbos, Federico Spurio, Juergen Gall

Recent Video-Language Models (VLMs) achieve promising results on long-video understanding, but their performance still lags behind that achieved on tasks involving images or short videos. This has led to great interest in improving the long context modeling of VLMs by introducing novel modules and additional complexity. % additional training time. In this paper, we take a different approach: rather than fine-tuning VLMs with the limited data available, we attempt to maximize the performance of existing models. To this end, we propose a novel visual prompting strategy specifically designed for long-video understanding. By combining multiple frames as panels into one image, we effectively trade off spatial details for temporal resolution. Our approach is training-free, parameter-free, and model-agnostic, and can be seamlessly integrated into existing VLMs. Extensive experiments on five established benchmarks across a wide range of model architectures, sizes, and context windows confirm the consistency of our approach. For the TimeScope (Long) dataset, which has the longest videos, the accuracy for video question answering is improved by up to 19.4\%. Overall, our method raises the bar for long video understanding models. We will make our code available upon acceptance.

Dimitrios Tzionas, Juergen Gall

Although commercial and open-source software exist to reconstruct a static object from a sequence recorded with an RGB-D sensor, there is a lack of tools that build rigged models of articulated objects that deform realistically and can be used for tracking or animation. In this work, we fill this gap and propose a method that creates a fully rigged model of an articulated object from depth data of a single sensor. To this end, we combine deformable mesh tracking, motion segmentation based on spectral clustering and skeletonization based on mean curvature flow. The fully rigged model then consists of a watertight mesh, embedded skeleton, and skinning weights.

Shuxiao Ding, Lukas Schneider, Marius Cordts et al.

Many query-based approaches for 3D Multi-Object Tracking (MOT) adopt the tracking-by-attention paradigm, utilizing track queries for identity-consistent detection and object queries for identity-agnostic track spawning. Tracking-by-attention, however, entangles detection and tracking queries in one embedding for both the detection and tracking task, which is sub-optimal. Other approaches resemble the tracking-by-detection paradigm and detect objects using decoupled track and detection queries followed by a subsequent association. These methods, however, do not leverage synergies between the detection and association task. Combining the strengths of both paradigms, we introduce ADA-Track++, a novel end-to-end framework for 3D MOT from multi-view cameras. We introduce a learnable data association module based on edge-augmented cross-attention, leveraging appearance and geometric features. We also propose an auxiliary token in this attention-based association module, which helps mitigate disproportionately high attention to incorrect association targets caused by attention normalization. Furthermore, we integrate this association module into the decoder layer of a DETR-based 3D detector, enabling simultaneous DETR-like query-to-image cross-attention for detection and query-to-query cross-attention for data association. By stacking these decoder layers, queries are refined for the detection and association task alternately, effectively harnessing the task dependencies. We evaluate our method on the nuScenes dataset and demonstrate the advantage of our approach compared to the two previous paradigms.

Federico Spurio, Emad Bahrami, Gianpiero Francesca et al.

In this work, we address unsupervised temporal action segmentation, which segments a set of long, untrimmed videos into semantically meaningful segments that are consistent across videos. While recent approaches combine representation learning and clustering in a single step for this task, they do not cope with large variations within temporal segments of the same class. To address this limitation, we propose a novel method, termed Hierarchical Vector Quantization (HVQ), that consists of two subsequent vector quantization modules. This results in a hierarchical clustering where the additional subclusters cover the variations within a cluster. We demonstrate that our approach captures the distribution of segment lengths much better than the state of the art. To this end, we introduce a new metric based on the Jensen-Shannon Distance (JSD) for unsupervised temporal action segmentation. We evaluate our approach on three public datasets, namely Breakfast, YouTube Instructional and IKEA ASM. Our approach outperforms the state of the art in terms of F1 score, recall and JSD.

Shijie Li, Farhad G. Zanjani, Haitam Ben Yahia et al.

Novel View Synthesis (NVS), which tries to produce a realistic image at the target view given source view images and their corresponding poses, is a fundamental problem in 3D Vision. As this task is heavily under-constrained, some recent work, like Zero123, tries to solve this problem with generative modeling, specifically using pre-trained diffusion models. Although this strategy generalizes well to new scenes, compared to neural radiance field-based methods, it offers low levels of flexibility. For example, it can only accept a single-view image as input, despite realistic applications often offering multiple input images. This is because the source-view images and corresponding poses are processed separately and injected into the model at different stages. Thus it is not trivial to generalize the model into multi-view source images, once they are available. To solve this issue, we try to process each pose image pair separately and then fuse them as a unified visual representation which will be injected into the model to guide image synthesis at the target-views. However, inconsistency and computation costs increase as the number of input source-view images increases. To solve these issues, the Multi-view Cross Former module is proposed which maps variable-length input data to fix-size output data. A two-stage training strategy is introduced to further improve the efficiency during training time. Qualitative and quantitative evaluation over multiple datasets demonstrates the effectiveness of the proposed method against previous approaches. The code will be released according to the acceptance.

Santosh Thoduka, Nico Hochgeschwender, Juergen Gall et al.

An object handover between a robot and a human is a coordinated action which is prone to failure for reasons such as miscommunication, incorrect actions and unexpected object properties. Existing works on handover failure detection and prevention focus on preventing failures due to object slip or external disturbances. However, there is a lack of datasets and evaluation methods that consider unpreventable failures caused by the human participant. To address this deficit, we present the multimodal Handover Failure Detection dataset, which consists of failures induced by the human participant, such as ignoring the robot or not releasing the object. We also present two baseline methods for handover failure detection: (i) a video classification method using 3D CNNs and (ii) a temporal action segmentation approach which jointly classifies the human action, robot action and overall outcome of the action. The results show that video is an important modality, but using force-torque data and gripper position help improve failure detection and action segmentation accuracy.

Shuxiao Ding, Yutong Yang, Julian Wiederer et al.

Query denoising has become a standard training strategy for DETR-based detectors by addressing the slow convergence issue. Besides that, query denoising can be used to increase the diversity of training samples for modeling complex scenarios which is critical for Multi-Object Tracking (MOT), showing its potential in MOT application. Existing approaches integrate query denoising within the tracking-by-attention paradigm. However, as the denoising process only happens within the single frame, it cannot benefit the tracker to learn temporal-related information. In addition, the attention mask in query denoising prevents information exchange between denoising and object queries, limiting its potential in improving association using self-attention. To address these issues, we propose TQD-Track, which introduces Temporal Query Denoising (TQD) tailored for MOT, enabling denoising queries to carry temporal information and instance-specific feature representation. We introduce diverse noise types onto denoising queries that simulate real-world challenges in MOT. We analyze our proposed TQD for different tracking paradigms, and find out the paradigm with explicit learned data association module, e.g. tracking-by-detection or alternating detection and association, benefit from TQD by a larger margin. For these paradigms, we further design an association mask in the association module to ensure the consistent interaction between track and detection queries as during inference. Extensive experiments on the nuScenes dataset demonstrate that our approach consistently enhances different tracking methods by only changing the training process, especially the paradigms with explicit association module.

Divya Velayudhan, Abdelfatah Ahmed, Mohamad Alansari et al.

Advancements in Computer-Aided Screening (CAS) systems are essential for improving the detection of security threats in X-ray baggage scans. However, current datasets are limited in representing real-world, sophisticated threats and concealment tactics, and existing approaches are constrained by a closed-set paradigm with predefined labels. To address these challenges, we introduce STCray, the first multimodal X-ray baggage security dataset, comprising 46,642 image-caption paired scans across 21 threat categories, generated using an X-ray scanner for airport security. STCray is meticulously developed with our specialized protocol that ensures domain-aware, coherent captions, that lead to the multi-modal instruction following data in X-ray baggage security. This allows us to train a domain-aware visual AI assistant named STING-BEE that supports a range of vision-language tasks, including scene comprehension, referring threat localization, visual grounding, and visual question answering (VQA), establishing novel baselines for multi-modal learning in X-ray baggage security. Further, STING-BEE shows state-of-the-art generalization in cross-domain settings. Code, data, and models are available at https://divs1159.github.io/STING-BEE/.

Syed Talal Wasim, Hamid Suleman, Olga Zatsarynna et al.

We present MixANT, a novel architecture for stochastic long-term dense anticipation of human activities. While recent State Space Models (SSMs) like Mamba have shown promise through input-dependent selectivity on three key parameters, the critical forget-gate ($\textbf{A}$ matrix) controlling temporal memory remains static. We address this limitation by introducing a mixture of experts approach that dynamically selects contextually relevant $\textbf{A}$ matrices based on input features, enhancing representational capacity without sacrificing computational efficiency. Extensive experiments on the 50Salads, Breakfast, and Assembly101 datasets demonstrate that MixANT consistently outperforms state-of-the-art methods across all evaluation settings. Our results highlight the importance of input-dependent forget-gate mechanisms for reliable prediction of human behavior in diverse real-world scenarios.

Santosh Thoduka, Sebastian Houben, Juergen Gall et al.

Robust robotic task execution hinges on the reliable detection of execution failures in order to trigger safe operation modes, recovery strategies, or task replanning. However, many failure detection methods struggle to provide meaningful performance when applied to a variety of real-world scenarios. In this paper, we propose a video-based failure detection approach that uses spatio-temporal knowledge in the form of the actions the robot performs and task-relevant objects within the field of view. Both pieces of information are available in most robotic scenarios and can thus be readily obtained. We demonstrate the effectiveness of our approach on three datasets that we amend, in part, with additional annotations of the aforementioned task-relevant knowledge. In light of the results, we also propose a data augmentation method that improves performance by applying variable frame rates to different parts of the video. We observe an improvement from 77.9 to 80.0 in F1 score on the ARMBench dataset without additional computational expense and an additional increase to 81.4 with test-time augmentation. The results emphasize the importance of spatio-temporal information during failure detection and suggest further investigation of suitable heuristics in future implementations. Code and annotations are available.

Xuying Huang, Sicong Pan, Olga Zatsarynna et al.

User privacy in mobile robotics has become a critical concern. Existing methods typically prioritize either the performance of downstream robotic tasks or privacy protection, with the latter often constraining the effectiveness of task execution. To jointly address both objectives, we study semantic-based robot navigation in an ultra-low-resolution setting to preserve visual privacy. A key challenge in such scenarios is recovering semantic segmentation from ultra-low-resolution RGB images. In this work, we introduce a novel fully joint-learning method that integrates an agglomerative feature extractor and a segmentation-aware discriminator to solve ultra-low-resolution semantic segmentation, thereby enabling privacy-preserving, semantic object-goal navigation. Our method outperforms different baselines on ultra-low-resolution semantic segmentation and our improved segmentation results increase the success rate of the semantic object-goal navigation in a real-world privacy-constrained scenario.

Shuai Li, Michael Burke, Subramanian Ramamoorthy et al.

This paper introduces a novel framework to learn data association for multi-object tracking in a self-supervised manner. Fully-supervised learning methods are known to achieve excellent tracking performances, but acquiring identity-level annotations is tedious and time-consuming. Motivated by the fact that in real-world scenarios object motion can be usually represented by a Markov process, we present a novel expectation maximization (EM) algorithm that trains a neural network to associate detections for tracking, without requiring prior knowledge of their temporal correspondences. At the core of our method lies a neural Kalman filter, with an observation model conditioned on associations of detections parameterized by a neural network. Given a batch of frames as input, data associations between detections from adjacent frames are predicted by a neural network followed by a Sinkhorn normalization that determines the assignment probabilities of detections to states. Kalman smoothing is then used to obtain the marginal probability of observations given the inferred states, producing a training objective to maximize this marginal probability using gradient descent. The proposed framework is fully differentiable, allowing the underlying neural model to be trained end-to-end. We evaluate our approach on the challenging MOT17, MOT20, and BDD100K datasets and achieve state-of-the-art results in comparison to self-supervised trackers using public detections.

Mohamad Hakam Shams Eddin, Juergen Gall

The spatio-temporal relations of impacts of extreme events and their drivers in climate data are not fully understood and there is a need of machine learning approaches to identify such spatio-temporal relations from data. The task, however, is very challenging since there are time delays between extremes and their drivers, and the spatial response of such drivers is inhomogeneous. In this work, we propose a first approach and benchmarks to tackle this challenge. Our approach is trained end-to-end to predict spatio-temporally extremes and spatio-temporally drivers in the physical input variables jointly. By enforcing the network to predict extremes from spatio-temporal binary masks of identified drivers, the network successfully identifies drivers that are correlated with extremes. We evaluate our approach on three newly created synthetic benchmarks, where two of them are based on remote sensing or reanalysis climate data, and on two real-world reanalysis datasets. The source code and datasets are publicly available at the project page https://hakamshams.github.io/IDE.

David T. Hoffmann, Nadine Behrmann, Juergen Gall et al.

This paper introduces Ranking Info Noise Contrastive Estimation (RINCE), a new member in the family of InfoNCE losses that preserves a ranked ordering of positive samples. In contrast to the standard InfoNCE loss, which requires a strict binary separation of the training pairs into similar and dissimilar samples, RINCE can exploit information about a similarity ranking for learning a corresponding embedding space. We show that the proposed loss function learns favorable embeddings compared to the standard InfoNCE whenever at least noisy ranking information can be obtained or when the definition of positives and negatives is blurry. We demonstrate this for a supervised classification task with additional superclass labels and noisy similarity scores. Furthermore, we show that RINCE can also be applied to unsupervised training with experiments on unsupervised representation learning from videos. In particular, the embedding yields higher classification accuracy, retrieval rates and performs better in out-of-distribution detection than the standard InfoNCE loss.

Mohsen Fayyaz, Soroush Abbasi Koohpayegani, Farnoush Rezaei Jafari et al.

While state-of-the-art vision transformer models achieve promising results in image classification, they are computationally expensive and require many GFLOPs. Although the GFLOPs of a vision transformer can be decreased by reducing the number of tokens in the network, there is no setting that is optimal for all input images. In this work, we therefore introduce a differentiable parameter-free Adaptive Token Sampler (ATS) module, which can be plugged into any existing vision transformer architecture. ATS empowers vision transformers by scoring and adaptively sampling significant tokens. As a result, the number of tokens is not constant anymore and varies for each input image. By integrating ATS as an additional layer within the current transformer blocks, we can convert them into much more efficient vision transformers with an adaptive number of tokens. Since ATS is a parameter-free module, it can be added to the off-the-shelf pre-trained vision transformers as a plug and play module, thus reducing their GFLOPs without any additional training. Moreover, due to its differentiable design, one can also train a vision transformer equipped with ATS. We evaluate the efficiency of our module in both image and video classification tasks by adding it to multiple SOTA vision transformers. Our proposed module improves the SOTA by reducing their computational costs (GFLOPs) by 2X, while preserving their accuracy on the ImageNet, Kinetics-400, and Kinetics-600 datasets.

Di Chen, Andreas Doering, Shanshan Zhang et al.

Video-based person re-identification (re-ID) is an important technique in visual surveillance systems which aims to match video snippets of people captured by different cameras. Existing methods are mostly based on convolutional neural networks (CNNs), whose building blocks either process local neighbor pixels at a time, or, when 3D convolutions are used to model temporal information, suffer from the misalignment problem caused by person movement. In this paper, we propose to overcome the limitations of normal convolutions with a human-oriented graph method. Specifically, features located at person joint keypoints are extracted and connected as a spatial-temporal graph. These keypoint features are then updated by message passing from their connected nodes with a graph convolutional network (GCN). During training, the GCN can be attached to any CNN-based person re-ID model to assist representation learning on feature maps, whilst it can be dropped after training for better inference speed. Our method brings significant improvements over the CNN-based baseline model on the MARS dataset with generated person keypoints and a newly annotated dataset: PoseTrackReID. It also defines a new state-of-the-art method in terms of top-1 accuracy and mean average precision in comparison to prior works.

Saber Pourheydari, Emad Bahrami, Mohsen Fayyaz et al.

While recurrent neural networks (RNNs) demonstrate outstanding capabilities for future video frame prediction, they model dynamics in a discrete time space, i.e., they predict the frames sequentially with a fixed temporal step. RNNs are therefore prone to accumulate the error as the number of future frames increases. In contrast, partial differential equations (PDEs) model physical phenomena like dynamics in a continuous time space. However, the estimated PDE for frame forecasting needs to be numerically solved, which is done by discretization of the PDE and diminishes most of the advantages compared to discrete models. In this work, we, therefore, propose to approximate the motion in a video by a continuous function using the Taylor series. To this end, we introduce TaylorSwiftNet, a novel convolutional neural network that learns to estimate the higher order terms of the Taylor series for a given input video. TaylorSwiftNet can swiftly predict future frames in parallel and it allows to change the temporal resolution of the forecast frames on-the-fly. The experimental results on various datasets demonstrate the superiority of our model.

Nadine Behrmann, Mohsen Fayyaz, Juergen Gall et al.

Self-supervised video representation methods typically focus on the representation of temporal attributes in videos. However, the role of stationary versus non-stationary attributes is less explored: Stationary features, which remain similar throughout the video, enable the prediction of video-level action classes. Non-stationary features, which represent temporally varying attributes, are more beneficial for downstream tasks involving more fine-grained temporal understanding, such as action segmentation. We argue that a single representation to capture both types of features is sub-optimal, and propose to decompose the representation space into stationary and non-stationary features via contrastive learning from long and short views, i.e. long video sequences and their shorter sub-sequences. Stationary features are shared between the short and long views, while non-stationary features aggregate the short views to match the corresponding long view. To empirically verify our approach, we demonstrate that our stationary features work particularly well on an action recognition downstream task, while our non-stationary features perform better on action segmentation. Furthermore, we analyse the learned representations and find that stationary features capture more temporally stable, static attributes, while non-stationary features encompass more temporally varying ones.

Yaser Souri, Yazan Abu Farha, Fabien Despinoy et al.

We introduce FIFA, a fast approximate inference method for action segmentation and alignment. Unlike previous approaches, FIFA does not rely on expensive dynamic programming for inference. Instead, it uses an approximate differentiable energy function that can be minimized using gradient-descent. FIFA is a general approach that can replace exact inference improving its speed by more than 5 times while maintaining its performance. FIFA is an anytime inference algorithm that provides a better speed vs. accuracy trade-off compared to exact inference. We apply FIFA on top of state-of-the-art approaches for weakly supervised action segmentation and alignment as well as fully supervised action segmentation. FIFA achieves state-of-the-art results on most metrics on two action segmentation datasets.