Alina Roitberg

Kunyu Peng, Alina Roitberg, Kailun Yang et al.

Traditional video-based human activity recognition has experienced remarkable progress linked to the rise of deep learning, but this effect was slower as it comes to the downstream task of driver behavior understanding. Understanding the situation inside the vehicle cabin is essential for Advanced Driving Assistant System (ADAS) as it enables identifying distraction, predicting driver's intent and leads to more convenient human-vehicle interaction. At the same time, driver observation systems face substantial obstacles as they need to capture different granularities of driver states, while the complexity of such secondary activities grows with the rising automation and increased driver freedom. Furthermore, a model is rarely deployed under conditions identical to the ones in the training set, as sensor placements and types vary from vehicle to vehicle, constituting a substantial obstacle for real-life deployment of data-driven models. In this work, we present a novel vision-based framework for recognizing secondary driver behaviours based on visual transformers and an additional augmented feature distribution calibration module. This module operates in the latent feature-space enriching and diversifying the training set at feature-level in order to improve generalization to novel data appearances, (e.g., sensor changes) and general feature quality. Our framework consistently leads to better recognition rates, surpassing previous state-of-the-art results of the public Drive&Act benchmark on all granularity levels. Our code is publicly available at https://github.com/KPeng9510/TransDARC.

Yiping Wei, Kunyu Peng, Alina Roitberg et al.

Self-supervised representation learning for human action recognition has developed rapidly in recent years. Most of the existing works are based on skeleton data while using a multi-modality setup. These works overlooked the differences in performance among modalities, which led to the propagation of erroneous knowledge between modalities while only three fundamental modalities, i.e., joints, bones, and motions are used, hence no additional modalities are explored. In this work, we first propose an Implicit Knowledge Exchange Module (IKEM) which alleviates the propagation of erroneous knowledge between low-performance modalities. Then, we further propose three new modalities to enrich the complementary information between modalities. Finally, to maintain efficiency when introducing new modalities, we propose a novel teacher-student framework to distill the knowledge from the secondary modalities into the mandatory modalities considering the relationship constrained by anchors, positives, and negatives, named relational cross-modality knowledge distillation. The experimental results demonstrate the effectiveness of our approach, unlocking the efficient use of skeleton-based multi-modality data. Source code will be made publicly available at https://github.com/desehuileng0o0/IKEM.

Hao Shi, Yu Li, Kailun Yang et al.

This paper raises the new task of Fisheye Semantic Completion (FSC), where dense texture, structure, and semantics of a fisheye image are inferred even beyond the sensor field-of-view (FoV). Fisheye cameras have larger FoV than ordinary pinhole cameras, yet its unique special imaging model naturally leads to a blind area at the edge of the image plane. This is suboptimal for safety-critical applications since important perception tasks, such as semantic segmentation, become very challenging within the blind zone. Previous works considered the out-FoV outpainting and in-FoV segmentation separately. However, we observe that these two tasks are actually closely coupled. To jointly estimate the tightly intertwined complete fisheye image and scene semantics, we introduce the new FishDreamer which relies on successful ViTs enhanced with a novel Polar-aware Cross Attention module (PCA) to leverage dense context and guide semantically-consistent content generation while considering different polar distributions. In addition to the contribution of the novel task and architecture, we also derive Cityscapes-BF and KITTI360-BF datasets to facilitate training and evaluation of this new track. Our experiments demonstrate that the proposed FishDreamer outperforms methods solving each task in isolation and surpasses alternative approaches on the Fisheye Semantic Completion. Code and datasets are publicly available at https://github.com/MasterHow/FishDreamer.

Xinyu Luo, Jiaming Zhang, Kailun Yang et al.

Autonomous vehicles utilize urban scene segmentation to understand the real world like a human and react accordingly. Semantic segmentation of normal scenes has experienced a remarkable rise in accuracy on conventional benchmarks. However, a significant portion of real-life accidents features abnormal scenes, such as those with object deformations, overturns, and unexpected traffic behaviors. Since even small mis-segmentation of driving scenes can lead to serious threats to human lives, the robustness of such models in accident scenarios is an extremely important factor in ensuring safety of intelligent transportation systems. In this paper, we propose a Multi-source Meta-learning Unsupervised Domain Adaptation (MMUDA) framework, to improve the generalization of segmentation transformers to extreme accident scenes. In MMUDA, we make use of Multi-Domain Mixed Sampling to augment the images of multiple-source domains (normal scenes) with the target data appearances (abnormal scenes). To train our model, we intertwine and study a meta-learning strategy in the multi-source setting for robustifying the segmentation results. We further enhance the segmentation backbone (SegFormer) with a HybridASPP decoder design, featuring large window attention spatial pyramid pooling and strip pooling, to efficiently aggregate long-range contextual dependencies. Our approach achieves a mIoU score of 46.97% on the DADA-seg benchmark, surpassing the previous state-of-the-art model by more than 7.50%. Code will be made publicly available at https://github.com/xinyu-laura/MMUDA.

Zdravko Marinov, David Schneider, Alina Roitberg et al.

Domain shifts, such as appearance changes, are a key challenge in real-world applications of activity recognition models, which range from assistive robotics and smart homes to driver observation in intelligent vehicles. For example, while simulations are an excellent way of economical data collection, a Synthetic-to-Real domain shift leads to a > 60% drop in accuracy when recognizing activities of Daily Living (ADLs). We tackle this challenge and introduce an activity domain generation framework which creates novel ADL appearances (novel domains) from different existing activity modalities (source domains) inferred from video training data. Our framework computes human poses, heatmaps of body joints, and optical flow maps and uses them alongside the original RGB videos to learn the essence of source domains in order to generate completely new ADL domains. The model is optimized by maximizing the distance between the existing source appearances and the generated novel appearances while ensuring that the semantics of an activity is preserved through an additional classification loss. While source data multimodality is an important concept in this design, our setup does not rely on multi-sensor setups, (i.e., all source modalities are inferred from a single video only.) The newly created activity domains are then integrated in the training of the ADL classification networks, resulting in models far less susceptible to changes in data distributions. Extensive experiments on the Synthetic-to-Real benchmark Sims4Action demonstrate the potential of the domain generation paradigm for cross-domain ADL recognition, setting new state-of-the-art results. Our code is publicly available at https://github.com/Zrrr1997/syn2real_DG

Kunyu Peng, Di Wen, Kailun Yang et al.

In Open-Set Domain Generalization (OSDG), the model is exposed to both new variations of data appearance (domains) and open-set conditions, where both known and novel categories are present at test time. The challenges of this task arise from the dual need to generalize across diverse domains and accurately quantify category novelty, which is critical for applications in dynamic environments. Recently, meta-learning techniques have demonstrated superior results in OSDG, effectively orchestrating the meta-train and -test tasks by employing varied random categories and predefined domain partition strategies. These approaches prioritize a well-designed training schedule over traditional methods that focus primarily on data augmentation and the enhancement of discriminative feature learning. The prevailing meta-learning models in OSDG typically utilize a predefined sequential domain scheduler to structure data partitions. However, a crucial aspect that remains inadequately explored is the influence brought by strategies of domain schedulers during training. In this paper, we observe that an adaptive domain scheduler benefits more in OSDG compared with prefixed sequential and random domain schedulers. We propose the Evidential Bi-Level Hardest Domain Scheduler (EBiL-HaDS) to achieve an adaptive domain scheduler. This method strategically sequences domains by assessing their reliabilities in utilizing a follower network, trained with confidence scores learned in an evidential manner, regularized by max rebiasing discrepancy, and optimized in a bi-level manner. The results show that our method substantially improves OSDG performance and achieves more discriminative embeddings for both the seen and unseen categories. The source code is publicly available at https://github.com/KPeng9510/EBiL-HaDS.

Kunyu Peng, Jia Fu, Kailun Yang et al.

We introduce a new task called Referring Atomic Video Action Recognition (RAVAR), aimed at identifying atomic actions of a particular person based on a textual description and the video data of this person. This task differs from traditional action recognition and localization, where predictions are delivered for all present individuals. In contrast, we focus on recognizing the correct atomic action of a specific individual, guided by text. To explore this task, we present the RefAVA dataset, containing 36,630 instances with manually annotated textual descriptions of the individuals. To establish a strong initial benchmark, we implement and validate baselines from various domains, e.g., atomic action localization, video question answering, and text-video retrieval. Since these existing methods underperform on RAVAR, we introduce RefAtomNet -- a novel cross-stream attention-driven method specialized for the unique challenges of RAVAR: the need to interpret a textual referring expression for the targeted individual, utilize this reference to guide the spatial localization and harvest the prediction of the atomic actions for the referring person. The key ingredients are: (1) a multi-stream architecture that connects video, text, and a new location-semantic stream, and (2) cross-stream agent attention fusion and agent token fusion which amplify the most relevant information across these streams and consistently surpasses standard attention-based fusion on RAVAR. Extensive experiments demonstrate the effectiveness of RefAtomNet and its building blocks for recognizing the action of the described individual. The dataset and code will be made publicly available at https://github.com/KPeng9510/RAVAR.

Ping-Cheng Wei, Kunyu Peng, Alina Roitberg et al.

Failure to timely diagnose and effectively treat depression leads to over 280 million people suffering from this psychological disorder worldwide. The information cues of depression can be harvested from diverse heterogeneous resources, e.g., audio, visual, and textual data, raising demand for new effective multi-modal fusion approaches for automatic estimation. In this work, we tackle the task of automatically identifying depression from multi-modal data and introduce a sub-attention mechanism for linking heterogeneous information while leveraging Convolutional Bidirectional LSTM as our backbone. To validate this idea, we conduct extensive experiments on the public DAIC-WOZ benchmark for depression assessment featuring different evaluation modes and taking gender-specific biases into account. The proposed model yields effective results with 0.89 precision and 0.70 F1-score in detecting major depression and 4.92 MAE in estimating the severity. Our attention-based fusion module consistently outperforms conventional late fusion approaches and achieves competitive performance compared to the previously published depression estimation frameworks, while learning to diagnose the disorder end-to-end and relying on far fewer preprocessing steps.

Yifei Chen, Kunyu Peng, Alina Roitberg et al.

To integrate action recognition into autonomous robotic systems, it is essential to address challenges such as person occlusions-a common yet often overlooked scenario in existing self-supervised skeleton-based action recognition methods. In this work, we propose IosPSTL, a simple and effective self-supervised learning framework designed to handle occlusions. IosPSTL combines a cluster-agnostic KNN imputer with an Occluded Partial Spatio-Temporal Learning (OPSTL) strategy. First, we pre-train the model on occluded skeleton sequences. Then, we introduce a cluster-agnostic KNN imputer that performs semantic grouping using k-means clustering on sequence embeddings. It imputes missing skeleton data by applying K-Nearest Neighbors in the latent space, leveraging nearby sample representations to restore occluded joints. This imputation generates more complete skeleton sequences, which significantly benefits downstream self-supervised models. To further enhance learning, the OPSTL module incorporates Adaptive Spatial Masking (ASM) to make better use of intact, high-quality skeleton sequences during training. Our method achieves state-of-the-art performance on the occluded versions of the NTU-60 and NTU-120 datasets, demonstrating its robustness and effectiveness under challenging conditions. Code is available at https://github.com/cyfml/OPSTL.

Kunyu Peng, David Schneider, Alina Roitberg et al.

In this paper, we tackle the new task of video-based Activated Muscle Group Estimation (AMGE) aiming at identifying active muscle regions during physical activity in the wild. To this intent, we provide the MuscleMap dataset featuring >15K video clips with 135 different activities and 20 labeled muscle groups. This dataset opens the vistas to multiple video-based applications in sports and rehabilitation medicine under flexible environment constraints. The proposed MuscleMap dataset is constructed with YouTube videos, specifically targeting High-Intensity Interval Training (HIIT) physical exercise in the wild. To make the AMGE model applicable in real-life situations, it is crucial to ensure that the model can generalize well to numerous types of physical activities not present during training and involving new combinations of activated muscles. To achieve this, our benchmark also covers an evaluation setting where the model is exposed to activity types excluded from the training set. Our experiments reveal that the generalizability of existing architectures adapted for the AMGE task remains a challenge. Therefore, we also propose a new approach, TransM3E, which employs a multi-modality feature fusion mechanism between both the video transformer model and the skeleton-based graph convolution model with novel cross-modal knowledge distillation executed on multi-classification tokens. The proposed method surpasses all popular video classification models when dealing with both, previously seen and new types of physical activities. The database and code can be found at https://github.com/KPeng9510/MuscleMap.

Alina Roitberg, Kunyu Peng, Zdravko Marinov et al.

Visual recognition inside the vehicle cabin leads to safer driving and more intuitive human-vehicle interaction but such systems face substantial obstacles as they need to capture different granularities of driver behaviour while dealing with highly limited body visibility and changing illumination. Multimodal recognition mitigates a number of such issues: prediction outcomes of different sensors complement each other due to different modality-specific strengths and weaknesses. While several late fusion methods have been considered in previously published frameworks, they constantly feature different architecture backbones and building blocks making it very hard to isolate the role of the chosen late fusion strategy itself. This paper presents an empirical evaluation of different paradigms for decision-level late fusion in video-based driver observation. We compare seven different mechanisms for joining the results of single-modal classifiers which have been both popular, (e.g. score averaging) and not yet considered (e.g. rank-level fusion) in the context of driver observation evaluating them based on different criteria and benchmark settings. This is the first systematic study of strategies for fusing outcomes of multimodal predictors inside the vehicles, conducted with the goal to provide guidance for fusion scheme selection.

Calvin Tanama, Kunyu Peng, Zdravko Marinov et al.

Deep learning-based models are at the forefront of most driver observation benchmarks due to their remarkable accuracies but are also associated with high computational costs. This is challenging, as resources are often limited in real-world driving scenarios. This paper introduces a lightweight framework for resource-efficient driver activity recognition. The framework enhances 3D MobileNet, a neural architecture optimized for speed in video classification, by incorporating knowledge distillation and model quantization to balance model accuracy and computational efficiency. Knowledge distillation helps maintain accuracy while reducing the model size by leveraging soft labels from a larger teacher model (I3D), instead of relying solely on original ground truth data. Model quantization significantly lowers memory and computation demands by using lower precision integers for model weights and activations. Extensive testing on a public dataset for in-vehicle monitoring during autonomous driving demonstrates that this new framework achieves a threefold reduction in model size and a 1.4-fold improvement in inference time, compared to an already optimized architecture. The code for this study is available at https://github.com/calvintanama/qd-driver-activity-reco.

Alina Roitberg, Kunyu Peng, David Schneider et al.

Driver observation models are rarely deployed under perfect conditions. In practice, illumination, camera placement and type differ from the ones present during training and unforeseen behaviours may occur at any time. While observing the human behind the steering wheel leads to more intuitive human-vehicle-interaction and safer driving, it requires recognition algorithms which do not only predict the correct driver state, but also determine their prediction quality through realistic and interpretable confidence measures. Reliable uncertainty estimates are crucial for building trust and are a serious obstacle for deploying activity recognition networks in real driving systems. In this work, we for the first time examine how well the confidence values of modern driver observation models indeed match the probability of the correct outcome and show that raw neural network-based approaches tend to significantly overestimate their prediction quality. To correct this misalignment between the confidence values and the actual uncertainty, we consider two strategies. First, we enhance two activity recognition models often used for driver observation with temperature scaling-an off-the-shelf method for confidence calibration in image classification. Then, we introduce Calibrated Action Recognition with Input Guidance (CARING)-a novel approach leveraging an additional neural network to learn scaling the confidences depending on the video representation. Extensive experiments on the Drive&Act dataset demonstrate that both strategies drastically improve the quality of model confidences, while our CARING model out-performs both, the original architectures and their temperature scaling enhancement, leading to best uncertainty estimates.

Omar Moured, Jiaming Zhang, Alina Roitberg et al.

The digitization of documents allows for wider accessibility and reproducibility. While automatic digitization of document layout and text content has been a long-standing focus of research, this problem in regard to graphical elements, such as statistical plots, has been under-explored. In this paper, we introduce the task of fine-grained visual understanding of mathematical graphics and present the Line Graphics (LG) dataset, which includes pixel-wise annotations of 5 coarse and 10 fine-grained categories. Our dataset covers 520 images of mathematical graphics collected from 450 documents from different disciplines. Our proposed dataset can support two different computer vision tasks, i.e., semantic segmentation and object detection. To benchmark our LG dataset, we explore 7 state-of-the-art models. To foster further research on the digitization of statistical graphs, we will make the dataset, code, and models publicly available to the community.

Zdravko Marinov, Alina Roitberg, David Schneider et al.

Modality selection is an important step when designing multimodal systems, especially in the case of cross-domain activity recognition as certain modalities are more robust to domain shift than others. However, selecting only the modalities which have a positive contribution requires a systematic approach. We tackle this problem by proposing an unsupervised modality selection method (ModSelect), which does not require any ground-truth labels. We determine the correlation between the predictions of multiple unimodal classifiers and the domain discrepancy between their embeddings. Then, we systematically compute modality selection thresholds, which select only modalities with a high correlation and low domain discrepancy. We show in our experiments that our method ModSelect chooses only modalities with positive contributions and consistently improves the performance on a Synthetic-to-Real domain adaptation benchmark, narrowing the domain gap.

Walter Morales-Alvarez, Novel Certad, Alina Roitberg et al.

For driver observation frameworks, clean datasets collected in controlled simulated environments often serve as the initial training ground. Yet, when deployed under real driving conditions, such simulator-trained models quickly face the problem of distributional shifts brought about by changing illumination, car model, variations in subject appearances, sensor discrepancies, and other environmental alterations. This paper investigates the viability of transferring video-based driver observation models from simulation to real-world scenarios in autonomous vehicles, given the frequent use of simulation data in this domain due to safety issues. To achieve this, we record a dataset featuring actual autonomous driving conditions and involving seven participants engaged in highly distracting secondary activities. To enable direct SIM to REAL transfer, our dataset was designed in accordance with an existing large-scale simulator dataset used as the training source. We utilize the Inflated 3D ConvNet (I3D) model, a popular choice for driver observation, with Gradient-weighted Class Activation Mapping (Grad-CAM) for detailed analysis of model decision-making. Though the simulator-based model clearly surpasses the random baseline, its recognition quality diminishes, with average accuracy dropping from 85.7% to 46.6%. We also observe strong variations across different behavior classes. This underscores the challenges of model transferability, facilitating our research of more robust driver observation systems capable of dealing with real driving conditions.

Jonas Nasimzada, Jens Kleesiek, Ken Herrmann et al.

Recognizing pain in video is crucial for improving patient-computer interaction systems, yet traditional data collection in this domain raises significant ethical and logistical challenges. This study introduces a novel approach that leverages synthetic data to enhance video-based pain recognition models, providing an ethical and scalable alternative. We present a pipeline that synthesizes realistic 3D facial models by capturing nuanced facial movements from a small participant pool, and mapping these onto diverse synthetic avatars. This process generates 8,600 synthetic faces, accurately reflecting genuine pain expressions from varied angles and perspectives. Utilizing advanced facial capture techniques, and leveraging public datasets like CelebV-HQ and FFHQ-UV for demographic diversity, our new synthetic dataset significantly enhances model training while ensuring privacy by anonymizing identities through facial replacements. Experimental results demonstrate that models trained on combinations of synthetic data paired with a small amount of real participants achieve superior performance in pain recognition, effectively bridging the gap between synthetic simulations and real-world applications. Our approach addresses data scarcity and ethical concerns, offering a new solution for pain detection and opening new avenues for research in privacy-preserving dataset generation. All resources are publicly available to encourage further innovation in this field.

Thinesh Thiyakesan Ponbagavathi, Kunyu Peng, Alina Roitberg

Foundation models (FMs) are large neural networks trained on broad datasets, excelling in downstream tasks with minimal fine-tuning. Human activity recognition in video has advanced with FMs, driven by competition among different architectures. However, high accuracies on standard benchmarks can draw an artificially rosy picture, as they often overlook real-world factors like changing camera perspectives. Popular benchmarks, mostly from YouTube or movies, offer diverse views but only coarse actions, which are insufficient for use-cases needing fine-grained, domain-specific actions. Domain-specific datasets (e.g., for industrial assembly) typically use data from limited static perspectives. This paper empirically evaluates how perspective changes affect different FMs in fine-grained human activity recognition. We compare multiple backbone architectures and design choices, including image- and video- based models, and various strategies for temporal information fusion, including commonly used score averaging and more novel attention-based temporal aggregation mechanisms. This is the first systematic study of different foundation models and specific design choices for human activity recognition from unknown views, conducted with the goal to provide guidance for backbone- and temporal- fusion scheme selection. Code and models will be made publicly available to the community.

Kunyu Peng, Cheng Yin, Junwei Zheng et al.

In real-world scenarios, human actions often fall outside the distribution of training data, making it crucial for models to recognize known actions and reject unknown ones. However, using pure skeleton data in such open-set conditions poses challenges due to the lack of visual background cues and the distinct sparse structure of body pose sequences. In this paper, we tackle the unexplored Open-Set Skeleton-based Action Recognition (OS-SAR) task and formalize the benchmark on three skeleton-based datasets. We assess the performance of seven established open-set approaches on our task and identify their limits and critical generalization issues when dealing with skeleton information. To address these challenges, we propose a distance-based cross-modality ensemble method that leverages the cross-modal alignment of skeleton joints, bones, and velocities to achieve superior open-set recognition performance. We refer to the key idea as CrossMax - an approach that utilizes a novel cross-modality mean max discrepancy suppression mechanism to align latent spaces during training and a cross-modality distance-based logits refinement method during testing. CrossMax outperforms existing approaches and consistently yields state-of-the-art results across all datasets and backbones. The benchmark, code, and models will be released at https://github.com/KPeng9510/OS-SAR.

Yi Xu, Kunyu Peng, Di Wen et al.

Understanding human actions from body poses is critical for assistive robots sharing space with humans in order to make informed and safe decisions about the next interaction. However, precise temporal localization and annotation of activity sequences is time-consuming and the resulting labels are often noisy. If not effectively addressed, label noise negatively affects the model's training, resulting in lower recognition quality. Despite its importance, addressing label noise for skeleton-based action recognition has been overlooked so far. In this study, we bridge this gap by implementing a framework that augments well-established skeleton-based human action recognition methods with label-denoising strategies from various research areas to serve as the initial benchmark. Observations reveal that these baselines yield only marginal performance when dealing with sparse skeleton data. Consequently, we introduce a novel methodology, NoiseEraSAR, which integrates global sample selection, co-teaching, and Cross-Modal Mixture-of-Experts (CM-MOE) strategies, aimed at mitigating the adverse impacts of label noise. Our proposed approach demonstrates better performance on the established benchmark, setting new state-of-the-art standards. The source code for this study is accessible at https://github.com/xuyizdby/NoiseEraSAR.

Sarvenaz Babakhani, David Remy, Alina Roitberg

Energy expenditure estimation aims to infer human metabolic rate from physiological signals such as heart rate, respiration, or accelerometer data, and has been studied primarily with classical regression methods. The few existing deep learning approaches rarely disentangle the role of neural architecture from that of signal choice. In this work, we systematically evaluate both aspects. We compare classical baselines with newer neural architectures across single signals, signal pairs, and grouped sensor inputs for diverse physical activities. Our results show that minute ventilation is the most predictive individual signal, with a transformer model achieving the lowest root mean square error (RMSE) of 0.87 W/kg across all activities. Paired and grouped signals, such as those from the Hexoskin smart shirt (five signals), offer good alternatives for faster models like CNN and ResNet with attention. Per-activity evaluation revealed mixed outcomes: notably better results in low-intensity activities (RMSE down to 0.29 W/kg; NRMSE = 0.04), while higher-intensity tasks showed larger RMSE but more comparable normalized errors. Finally, subject-level analysis highlights strong inter-individual variability, motivating the need for adaptive modeling strategies. Our code and models will be publicly available at https://github.com/Sarvibabakhani/deeplearning-biosignals-ee .

Di Wen, Kunyu Peng, Kailun Yang et al.

Human-Object Interaction (HOI) detection is crucial for robot-human assistance, enabling context-aware support. However, models trained on clean datasets degrade in real-world conditions due to unforeseen corruptions, leading to inaccurate predictions. To address this, we introduce the first robustness benchmark for HOI detection, evaluating model resilience under diverse challenges. Despite advances, current models struggle with environmental variability, occlusions, and noise. Our benchmark, RoHOI, includes 20 corruption types based on the HICO-DET and V-COCO datasets and a new robustness-focused metric. We systematically analyze existing models in the HOI field, revealing significant performance drops under corruptions. To improve robustness, we propose a Semantic-Aware Masking-based Progressive Learning (SAMPL) strategy to guide the model to be optimized based on holistic and partial cues, thus dynamically adjusting the model's optimization to enhance robust feature learning. Extensive experiments show that our approach outperforms state-of-the-art methods, setting a new standard for robust HOI detection. Benchmarks, datasets, and code are available at https://github.com/KratosWen/RoHOI.

Kunyu Peng, Di Wen, Sarfraz M. Saquib et al.

Open-Set Domain Generalization (OSDG) is a challenging task requiring models to accurately predict familiar categories while minimizing confidence for unknown categories to effectively reject them in unseen domains. While the OSDG field has seen considerable advancements, the impact of label noise--a common issue in real-world datasets--has been largely overlooked. Label noise can mislead model optimization, thereby exacerbating the challenges of open-set recognition in novel domains. In this study, we take the first step towards addressing Open-Set Domain Generalization under Noisy Labels (OSDG-NL) by constructing dedicated benchmarks derived from widely used OSDG datasets, including PACS and DigitsDG. We evaluate baseline approaches by integrating techniques from both label denoising and OSDG methodologies, highlighting the limitations of existing strategies in handling label noise effectively. To address these limitations, we propose HyProMeta, a novel framework that integrates hyperbolic category prototypes for label noise-aware meta-learning alongside a learnable new-category agnostic prompt designed to enhance generalization to unseen classes. Our extensive experiments demonstrate the superior performance of HyProMeta compared to state-of-the-art methods across the newly established benchmarks. The source code of this work is released at https://github.com/KPeng9510/HyProMeta.

Thinesh Thiyakesan Ponbagavathi, Constantin Seibold, Alina Roitberg

Adapting image-pretrained backbones to video typically relies on time-domain adapters tuned to a single temporal scale. Our experiments show that these modules pick up static image cues and very fast flicker changes, while overlooking medium-speed motion. Capturing dynamics across multiple time-scales is, however, crucial for fine-grained temporal analysis (i.e., opening vs. closing bottle). To address this, we introduce Frame2Freq -- a family of frequency-aware adapters that perform spectral encoding during image-to-video adaptation of pretrained Vision Foundation Models (VFMs), improving fine-grained action recognition. Frame2Freq uses Fast Fourier Transform (FFT) along time and learns frequency-band specific embeddings that adaptively highlight the most discriminative frequency ranges. Across five fine-grained activity recognition datasets, Frame2Freq outperforms prior PEFT methods and even surpasses fully fine-tuned models on four of them. These results provide encouraging evidence that frequency analysis methods are a powerful tool for modeling temporal dynamics in image-to-video transfer. Code is available at https://github.com/th-nesh/Frame2Freq.

Thinesh Thiyakesan Ponbagavathi, Kunyu Peng, Alina Roitberg

Changes of camera perspective are a common obstacle in driver monitoring. While deep learning and pretrained foundation models show strong potential for improved generalization via lightweight adaptation of the final layers ('probing'), their robustness to unseen viewpoints remains underexplored. We study this challenge by adapting image foundation models to driver monitoring using a single training view, and evaluating them directly on unseen perspectives without further adaptation. We benchmark simple linear probes, advanced probing strategies, and compare two foundation models (DINOv2 and CLIP) against parameter-efficient fine-tuning (PEFT) and full fine-tuning. Building on these insights, we introduce T-MASK -- a new image-to-video probing method that leverages temporal token masking and emphasizes more dynamic video regions. Benchmarked on the public Drive&Act dataset, T-MASK improves cross-view top-1 accuracy by $+1.23\%$ over strong probing baselines and $+8.0\%$ over PEFT methods, without adding any parameters. It proves particularly effective for underrepresented secondary activities, boosting recognition by $+5.42\%$ under the trained view and $+1.36\%$ under cross-view settings. This work provides encouraging evidence that adapting foundation models with lightweight probing methods like T-MASK has strong potential in fine-grained driver observation, especially in cross-view and low-data settings. These results highlight the importance of temporal token selection when leveraging foundation models to build robust driver monitoring systems. Code and models will be made available at https://github.com/th-nesh/T-MASK to support ongoing research.

Kunyu Peng, Di Wen, David Schneider et al.

Domain adaptation is essential for activity recognition to ensure accurate and robust performance across diverse environments, sensor types, and data sources. Unsupervised domain adaptation methods have been extensively studied, yet, they require large-scale unlabeled data from the target domain. In this work, we focus on Few-Shot Domain Adaptation for Activity Recognition (FSDA-AR), which leverages a very small amount of labeled target videos to achieve effective adaptation. This approach is appealing for applications because it only needs a few or even one labeled example per class in the target domain, ideal for recognizing rare but critical activities. However, the existing FSDA-AR works mostly focus on the domain adaptation on sports videos, where the domain diversity is limited. We propose a new FSDA-AR benchmark using five established datasets considering the adaptation on more diverse and challenging domains. Our results demonstrate that FSDA-AR performs comparably to unsupervised domain adaptation with significantly fewer labeled target domain samples. We further propose a novel approach, RelaMiX, to better leverage the few labeled target domain samples as knowledge guidance. RelaMiX encompasses a temporal relational attention network with relation dropout, alongside a cross-domain information alignment mechanism. Furthermore, it integrates a mechanism for mixing features within a latent space by using the few-shot target domain samples. The proposed RelaMiX solution achieves state-of-the-art performance on all datasets within the FSDA-AR benchmark. To encourage future research of few-shot domain adaptation for activity recognition, our code will be publicly available at https://github.com/KPeng9510/RelaMiX.

Jiacheng Lin, Jiajun Chen, Kailun Yang et al.

Interactive Image Segmentation (IIS) has emerged as a promising technique for decreasing annotation time. Substantial progress has been made in pre- and post-processing for IIS, but the critical issue of interaction ambiguity, notably hindering segmentation quality, has been under-researched. To address this, we introduce AdaptiveClick -- a click-aware transformer incorporating an adaptive focal loss that tackles annotation inconsistencies with tools for mask- and pixel-level ambiguity resolution. To the best of our knowledge, AdaptiveClick is the first transformer-based, mask-adaptive segmentation framework for IIS. The key ingredient of our method is the Click-Aware Mask-adaptive transformer Decoder (CAMD), which enhances the interaction between click and image features. Additionally, AdaptiveClick enables pixel-adaptive differentiation of hard and easy samples in the decision space, independent of their varying distributions. This is primarily achieved by optimizing a generalized Adaptive Focal Loss (AFL) with a theoretical guarantee, where two adaptive coefficients control the ratio of gradient values for hard and easy pixels. Our analysis reveals that the commonly used Focal and BCE losses can be considered special cases of the proposed AFL. With a plain ViT backbone, extensive experimental results on nine datasets demonstrate the superiority of AdaptiveClick compared to state-of-the-art methods. The source code is publicly available at https://github.com/lab206/AdaptiveClick.

Ruiping Liu, Kailun Yang, Alina Roitberg et al.

Semantic segmentation benchmarks in the realm of autonomous driving are dominated by large pre-trained transformers, yet their widespread adoption is impeded by substantial computational costs and prolonged training durations. To lift this constraint, we look at efficient semantic segmentation from a perspective of comprehensive knowledge distillation and aim to bridge the gap between multi-source knowledge extractions and transformer-specific patch embeddings. We put forward the Transformer-based Knowledge Distillation (TransKD) framework which learns compact student transformers by distilling both feature maps and patch embeddings of large teacher transformers, bypassing the long pre-training process and reducing the FLOPs by >85.0%. Specifically, we propose two fundamental modules to realize feature map distillation and patch embedding distillation, respectively: (1) Cross Selective Fusion (CSF) enables knowledge transfer between cross-stage features via channel attention and feature map distillation within hierarchical transformers; (2) Patch Embedding Alignment (PEA) performs dimensional transformation within the patchifying process to facilitate the patch embedding distillation. Furthermore, we introduce two optimization modules to enhance the patch embedding distillation from different perspectives: (1) Global-Local Context Mixer (GL-Mixer) extracts both global and local information of a representative embedding; (2) Embedding Assistant (EA) acts as an embedding method to seamlessly bridge teacher and student models with the teacher's number of channels. Experiments on Cityscapes, ACDC, NYUv2, and Pascal VOC2012 datasets show that TransKD outperforms state-of-the-art distillation frameworks and rivals the time-consuming pre-training method. The source code is publicly available at https://github.com/RuipingL/TransKD.

Kunyu Peng, Alina Roitberg, Kailun Yang et al.

We explore the problem of automatically inferring the amount of kilocalories used by human during physical activity from his/her video observation. To study this underresearched task, we introduce Vid2Burn -- an omni-source benchmark for estimating caloric expenditure from video data featuring both, high- and low-intensity activities for which we derive energy expenditure annotations based on models established in medical literature. In practice, a training set would only cover a certain amount of activity types, and it is important to validate, if the model indeed captures the essence of energy expenditure, (e.g., how many and which muscles are involved and how intense they work) instead of memorizing fixed values of specific activity categories seen during training. Ideally, the models should look beyond such category-specific biases and regress the caloric cost in videos depicting activity categories not explicitly present during training. With this property in mind, Vid2Burn is accompanied with a cross-category benchmark, where the task is to regress caloric expenditure for types of physical activities not present during training. An extensive evaluation of state-of-the-art approaches for video recognition modified for the energy expenditure estimation task demonstrates the difficulty of this problem, especially for new activity types at test-time, marking a new research direction. Dataset and code are available at https://github.com/KPeng9510/Vid2Burn.

Kunyu Peng, Alina Roitberg, David Schneider et al.

Human affect recognition is a well-established research area with numerous applications, e.g., in psychological care, but existing methods assume that all emotions-of-interest are given a priori as annotated training examples. However, the rising granularity and refinements of the human emotional spectrum through novel psychological theories and the increased consideration of emotions in context brings considerable pressure to data collection and labeling work. In this paper, we conceptualize one-shot recognition of emotions in context -- a new problem aimed at recognizing human affect states in finer particle level from a single support sample. To address this challenging task, we follow the deep metric learning paradigm and introduce a multi-modal emotion embedding approach which minimizes the distance of the same-emotion embeddings by leveraging complementary information of human appearance and the semantic scene context obtained through a semantic segmentation network. All streams of our context-aware model are optimized jointly using weighted triplet loss and weighted cross entropy loss. We conduct thorough experiments on both, categorical and numerical emotion recognition tasks of the Emotic dataset adapted to our one-shot recognition problem, revealing that categorizing human affect from a single example is a hard task. Still, all variants of our model clearly outperform the random baseline, while leveraging the semantic scene context consistently improves the learnt representations, setting state-of-the-art results in one-shot emotion recognition. To foster research of more universal representations of human affect states, we will make our benchmark and models publicly available to the community under https://github.com/KPeng9510/Affect-DML.

Alina Roitberg, David Schneider, Aulia Djamal et al.

Recognizing Activities of Daily Living (ADL) is a vital process for intelligent assistive robots, but collecting large annotated datasets requires time-consuming temporal labeling and raises privacy concerns, e.g., if the data is collected in a real household. In this work, we explore the concept of constructing training examples for ADL recognition by playing life simulation video games and introduce the SIMS4ACTION dataset created with the popular commercial game THE SIMS 4. We build Sims4Action by specifically executing actions-of-interest in a "top-down" manner, while the gaming circumstances allow us to freely switch between environments, camera angles and subject appearances. While ADL recognition on gaming data is interesting from the theoretical perspective, the key challenge arises from transferring it to the real-world applications, such as smart-homes or assistive robotics. To meet this requirement, Sims4Action is accompanied with a GamingToReal benchmark, where the models are evaluated on real videos derived from an existing ADL dataset. We integrate two modern algorithms for video-based activity recognition in our framework, revealing the value of life simulation video games as an inexpensive and far less intrusive source of training data. However, our results also indicate that tasks involving a mixture of gaming and real data are challenging, opening a new research direction. We will make our dataset publicly available at https://github.com/aroitberg/sims4action.

Thinesh Thiyakesan Ponbagavathi, Alina Roitberg

We study parameter-efficient image-to-video probing for the unaddressed challenge of recognizing nearly symmetric actions - visually similar actions that unfold in opposite temporal order (e.g., opening vs. closing a bottle). Existing probing mechanisms for image-pretrained models, such as DinoV2 and CLIP, rely on attention mechanism for temporal modeling but are inherently permutation-invariant, leading to identical predictions regardless of frame order. To address this, we introduce Self-attentive Temporal Embedding Probing (STEP), a simple yet effective approach designed to enforce temporal sensitivity in parameter-efficient image-to-video transfer. STEP enhances self-attentive probing with three key modifications: (1) a learnable frame-wise positional encoding, explicitly encoding temporal order; (2) a single global CLS token, for sequence coherence; and (3) a simplified attention mechanism to improve parameter efficiency. STEP outperforms existing image-to-video probing mechanisms by 3-15% across four activity recognition benchmarks with only 1/3 of the learnable parameters. On two datasets, it surpasses all published methods, including fully fine-tuned models. STEP shows a distinct advantage in recognizing nearly symmetric actions, surpassing other probing mechanisms by 9-19%. and parameter-heavier PEFT-based transfer methods by 5-15%. Code and models will be made publicly available.

Thinesh Thiyakesan Ponbagavathi, Chengzheng Yang, Alina Roitberg

Group Activity Detection (GAD) involves recognizing social groups and their collective behaviors in videos. Vision Foundation Models (VFMs), like DinoV2, offer excellent features, but are pretrained primarily on object-centric data and remain underexplored for modeling group dynamics. While they are a promising alternative to highly task-specific GAD architectures that require full fine-tuning, our initial investigation reveals that simply swapping CNN backbones used in these methods with VFMs brings little gain, underscoring the need for structured, group-aware reasoning on top. We introduce Prompt-driven Group Activity Detection (ProGraD) -- a method that bridges this gap through 1) learnable group prompts to guide the VFM attention toward social configurations, and 2) a lightweight two-layer GroupContext Transformer that infers actor-group associations and collective behavior. We evaluate our approach on two recent GAD benchmarks: Cafe, which features multiple concurrent social groups, and Social-CAD, which focuses on single-group interactions. While we surpass state-of-the-art in both settings, our method is especially effective in complex multi-group scenarios, where we yield a gain of 6.5\% (Group mAP\@1.0) and 8.2\% (Group mAP\@0.5) using only 10M trainable parameters. Furthermore, our experiments reveal that ProGraD produces interpretable attention maps, offering insights into actor-group reasoning. Code and models will be released.

Kunyu Peng, Alina Roitberg, Kailun Yang et al.

Occlusions are universal disruptions constantly present in the real world. Especially for sparse representations, such as human skeletons, a few occluded points might destroy the geometrical and temporal continuity critically affecting the results. Yet, the research of data-scarce recognition from skeleton sequences, such as one-shot action recognition, does not explicitly consider occlusions despite their everyday pervasiveness. In this work, we explicitly tackle body occlusions for Skeleton-based One-shot Action Recognition (SOAR). We mainly consider two occlusion variants: 1) random occlusions and 2) more realistic occlusions caused by diverse everyday objects, which we generate by projecting the existing IKEA 3D furniture models into the camera coordinate system of the 3D skeletons with different geometric parameters. We leverage the proposed pipeline to blend out portions of skeleton sequences of the three popular action recognition datasets and formalize the first benchmark for SOAR from partially occluded body poses. Another key property of our benchmark are the more realistic occlusions generated by everyday objects, as even in standard recognition from 3D skeletons, only randomly missing joints were considered. We re-evaluate existing state-of-the-art frameworks for SOAR in the light of this new task and further introduce Trans4SOAR - a new transformer-based model which leverages three data streams and mixed attention fusion mechanism to alleviate the adverse effects caused by occlusions. While our experiments demonstrate a clear decline in accuracy with missing skeleton portions, this effect is smaller with Trans4SOAR, which outperforms other architectures on all datasets. Although we specifically focus on occlusions, Trans4SOAR additionally yields state-of-the-art in the standard SOAR without occlusion, surpassing the best published approach by 2.85% on NTU-120.

Jiaming Zhang, Chaoxiang Ma, Kailun Yang et al.

Autonomous vehicles clearly benefit from the expanded Field of View (FoV) of 360-degree sensors, but modern semantic segmentation approaches rely heavily on annotated training data which is rarely available for panoramic images. We look at this problem from the perspective of domain adaptation and bring panoramic semantic segmentation to a setting, where labelled training data originates from a different distribution of conventional pinhole camera images. To achieve this, we formalize the task of unsupervised domain adaptation for panoramic semantic segmentation and collect DensePASS - a novel densely annotated dataset for panoramic segmentation under cross-domain conditions, specifically built to study the Pinhole-to-Panoramic domain shift and accompanied with pinhole camera training examples obtained from Cityscapes. DensePASS covers both, labelled- and unlabelled 360-degree images, with the labelled data comprising 19 classes which explicitly fit the categories available in the source (i.e. pinhole) domain. Since data-driven models are especially susceptible to changes in data distribution, we introduce P2PDA - a generic framework for Pinhole-to-Panoramic semantic segmentation which addresses the challenge of domain divergence with different variants of attention-augmented domain adaptation modules, enabling the transfer in output-, feature-, and feature confidence spaces. P2PDA intertwines uncertainty-aware adaptation using confidence values regulated on-the-fly through attention heads with discrepant predictions. Our framework facilitates context exchange when learning domain correspondences and dramatically improves the adaptation performance of accuracy- and efficiency-focused models. Comprehensive experiments verify that our framework clearly surpasses unsupervised domain adaptation- and specialized panoramic segmentation approaches.

Chaoxiang Ma, Jiaming Zhang, Kailun Yang et al.

Intelligent vehicles clearly benefit from the expanded Field of View (FoV) of the 360-degree sensors, but the vast majority of available semantic segmentation training images are captured with pinhole cameras. In this work, we look at this problem through the lens of domain adaptation and bring panoramic semantic segmentation to a setting, where labelled training data originates from a different distribution of conventional pinhole camera images. First, we formalize the task of unsupervised domain adaptation for panoramic semantic segmentation, where a network trained on labelled examples from the source domain of pinhole camera data is deployed in a different target domain of panoramic images, for which no labels are available. To validate this idea, we collect and publicly release DensePASS - a novel densely annotated dataset for panoramic segmentation under cross-domain conditions, specifically built to study the Pinhole-to-Panoramic transfer and accompanied with pinhole camera training examples obtained from Cityscapes. DensePASS covers both, labelled- and unlabelled 360-degree images, with the labelled data comprising 19 classes which explicitly fit the categories available in the source domain (i.e. pinhole) data. To meet the challenge of domain shift, we leverage the current progress of attention-based mechanisms and build a generic framework for cross-domain panoramic semantic segmentation based on different variants of attention-augmented domain adaptation modules. Our framework facilitates information exchange at local- and global levels when learning the domain correspondences and improves the domain adaptation performance of two standard segmentation networks by 6.05% and 11.26% in Mean IoU.

Kunyu Peng, Juncong Fei, Kailun Yang et al.

At the heart of all automated driving systems is the ability to sense the surroundings, e.g., through semantic segmentation of LiDAR sequences, which experienced a remarkable progress due to the release of large datasets such as SemanticKITTI and nuScenes-LidarSeg. While most previous works focus on sparse segmentation of the LiDAR input, dense output masks provide self-driving cars with almost complete environment information. In this paper, we introduce MASS - a Multi-Attentional Semantic Segmentation model specifically built for dense top-view understanding of the driving scenes. Our framework operates on pillar- and occupancy features and comprises three attention-based building blocks: (1) a keypoint-driven graph attention, (2) an LSTM-based attention computed from a vector embedding of the spatial input, and (3) a pillar-based attention, resulting in a dense 360-degree segmentation mask. With extensive experiments on both, SemanticKITTI and nuScenes-LidarSeg, we quantitatively demonstrate the effectiveness of our model, outperforming the state of the art by 19.0% on SemanticKITTI and reaching 30.4% in mIoU on nuScenes-LidarSeg, where MASS is the first work addressing the dense segmentation task. Furthermore, our multi-attention model is shown to be very effective for 3D object detection validated on the KITTI-3D dataset, showcasing its high generalizability to other tasks related to 3D vision.

Alina Roitberg, Monica Haurilet, Manuel Martinez et al.

Beyond assigning the correct class, an activity recognition model should also be able to determine, how certain it is in its predictions. We present the first study of how welthe confidence values of modern action recognition architectures indeed reflect the probability of the correct outcome and propose a learning-based approach for improving it. First, we extend two popular action recognition datasets with a reliability benchmark in form of the expected calibration error and reliability diagrams. Since our evaluation highlights that confidence values of standard action recognition architectures do not represent the uncertainty well, we introduce a new approach which learns to transform the model output into realistic confidence estimates through an additional calibration network. The main idea of our Calibrated Action Recognition with Input Guidance (CARING) model is to learn an optimal scaling parameter depending on the video representation. We compare our model with the native action recognition networks and the temperature scaling approach - a wide spread calibration method utilized in image classification. While temperature scaling alone drastically improves the reliability of the confidence values, our CARING method consistently leads to the best uncertainty estimates in all benchmark settings.

Robin Ruede, Verena Heusser, Lukas Frank et al.

A rapidly growing amount of content posted online, such as food recipes, opens doors to new exciting applications at the intersection of vision and language. In this work, we aim to estimate the calorie amount of a meal directly from an image by learning from recipes people have published on the Internet, thus skipping time-consuming manual data annotation. Since there are few large-scale publicly available datasets captured in unconstrained environments, we propose the pic2kcal benchmark comprising 308,000 images from over 70,000 recipes including photographs, ingredients and instructions. To obtain nutritional information of the ingredients and automatically determine the ground-truth calorie value, we match the items in the recipes with structured information from a food item database. We evaluate various neural networks for regression of the calorie quantity and extend them with the multi-task paradigm. Our learning procedure combines the calorie estimation with prediction of proteins, carbohydrates, and fat amounts as well as a multi-label ingredient classification. Our experiments demonstrate clear benefits of multi-task learning for calorie estimation, surpassing the single-task calorie regression by 9.9%. To encourage further research on this task, we make the code for generating the dataset and the models publicly available.

Alina Roitberg, Ziad Al-Halah, Rainer Stiefelhagen

Novelty detection is crucial for real-life applications. While it is common in activity recognition to assume a closed-set setting, i.e. test samples are always of training categories, this assumption is impractical in a real-world scenario. Test samples can be of various categories including those never seen before during training. Thus, being able to know what we know and what we do not know is decisive for the model to avoid what can be catastrophic consequences. We present in this work a novel approach for identifying samples of activity classes that are not previously seen by the classifier. Our model employs a voting-based scheme that leverages the estimated uncertainty of the individual classifiers in their predictions to measure the novelty of a new input sample. Furthermore, the voting is privileged to a subset of informed classifiers that can best estimate whether a sample is novel or not when it is classified to a certain known category. In a thorough evaluation on UCF-101 and HMDB-51, we show that our model consistently outperforms state-of-the-art in novelty detection. Additionally, by combining our model with off-the-shelf zero-shot learning (ZSL) approaches, our model leads to a significant improvement in action classification accuracy for the generalized ZSL setting.