Faegheh Sardari

Tanveer Hannan, Dimitrios Mallios, Parth Pathak et al.

Large Vision-Language Models (LVLMs) have demonstrated strong multimodal reasoning capabilities on long and complex documents. However, their high memory footprint makes them impractical for deployment on resource-constrained edge devices. We present DocSLM, an efficient Small Vision-Language Model designed for long-document understanding under constrained memory resources. DocSLM incorporates a Hierarchical Multimodal Compressor that jointly encodes visual, textual, and layout information from each page into a fixed-length sequence, greatly reducing memory consumption while preserving both local and global semantics. To enable scalable processing over arbitrarily long inputs, we introduce a Streaming Abstention mechanism that operates on document segments sequentially and filters low-confidence responses using an entropy-based uncertainty calibrator. Across multiple long multimodal document benchmarks, DocSLM matches or surpasses state-of-the-art methods while using 82\% fewer visual tokens, 75\% fewer parameters, and 71\% lower latency, delivering reliable multimodal document understanding on lightweight edge devices. Code is available in the supplementary material.

Faegheh Sardari, Armin Mustafa, Philip J. B. Jackson et al.

We present PAT, a transformer-based network that learns complex temporal co-occurrence action dependencies in a video by exploiting multi-scale temporal features. In existing methods, the self-attention mechanism in transformers loses the temporal positional information, which is essential for robust action detection. To address this issue, we (i) embed relative positional encoding in the self-attention mechanism and (ii) exploit multi-scale temporal relationships by designing a novel non hierarchical network, in contrast to the recent transformer-based approaches that use a hierarchical structure. We argue that joining the self-attention mechanism with multiple sub-sampling processes in the hierarchical approaches results in increased loss of positional information. We evaluate the performance of our proposed approach on two challenging dense multi-label benchmark datasets, and show that PAT improves the current state-of-the-art result by 1.1% and 0.6% mAP on the Charades and MultiTHUMOS datasets, respectively, thereby achieving the new state-of-the-art mAP at 26.5% and 44.6%, respectively. We also perform extensive ablation studies to examine the impact of the different components of our proposed network.

Faegheh Sardari, Armin Mustafa, Philip J. B. Jackson et al.

Weakly supervised audio-visual video parsing (AVVP) methods aim to detect audible-only, visible-only, and audible-visible events using only video-level labels. Existing approaches tackle this by leveraging unimodal and cross-modal contexts. However, we argue that while cross-modal learning is beneficial for detecting audible-visible events, in the weakly supervised scenario, it negatively impacts unaligned audible or visible events by introducing irrelevant modality information. In this paper, we propose CoLeaF, a novel learning framework that optimizes the integration of cross-modal context in the embedding space such that the network explicitly learns to combine cross-modal information for audible-visible events while filtering them out for unaligned events. Additionally, as videos often involve complex class relationships, modelling them improves performance. However, this introduces extra computational costs into the network. Our framework is designed to leverage cross-class relationships during training without incurring additional computations at inference. Furthermore, we propose new metrics to better evaluate a method's capabilities in performing AVVP. Our extensive experiments demonstrate that CoLeaF significantly improves the state-of-the-art results by an average of 1.9% and 2.4% F-score on the LLP and UnAV-100 datasets, respectively.

Yaru Chen, Faegheh Sardari, Peiliang Zhang et al.

Audio-visual video parsing (AVVP) aims to detect event categories and their temporal boundaries in videos, typically under weak supervision. Existing methods mainly focus on (i) improving temporal modeling using attention-based architectures or (ii) generating richer pseudo-labels to address the absence of frame-level annotations. However, attention-based models often overfit noisy pseudo-labels, leading to cumulative training errors, while pseudo-label generation approaches distribute attention uniformly across frames, weakening temporal localization accuracy. To address these challenges, we propose TEn-CATG, a text-enriched AVVP framework that combines semantic calibration with category-aware temporal reasoning. More specifically, we design a bi-directional text fusion (BiT) module by leveraging audio-visual features as semantic anchors to refine text embeddings, which departs from conventional text-to-feature alignment, thereby mitigating noise and enhancing cross-modal consistency. Furthermore, we introduce the category-aware temporal graph (CATG) module to model temporal relationships by selecting multi-scale temporal neighbors and learning category-specific temporal decay factors, enabling effective event-dependent temporal reasoning. Extensive experiments demonstrate that TEn-CATG achieves state-of-the-art results across multiple evaluation metrics on benchmark datasets LLP and UnAV-100, highlighting its robustness and superior ability to capture complex temporal and semantic dependencies in weakly supervised AVVP tasks.

Faegheh Sardari, Armin Mustafa, Philip J. B. Jackson et al.

Dense action detection involves detecting multiple co-occurring actions while action classes are often ambiguous and represent overlapping concepts. We argue that handling the dual challenge of temporal and class overlaps is too complex to effectively be tackled by a single network. To address this, we propose to decompose the task of detecting dense ambiguous actions into detecting dense, unambiguous sub-concepts that form the action classes (i.e., action entities and action motions), and assigning these sub-tasks to distinct sub-networks. By isolating these unambiguous concepts, the sub-networks can focus exclusively on resolving a single challenge, dense temporal overlaps. Furthermore, simultaneous actions in a video often exhibit interrelationships, and exploiting these relationships can improve the method performance. However, current dense action detection networks fail to effectively learn these relationships due to their reliance on binary cross-entropy optimization, which treats each class independently. To address this limitation, we propose providing explicit supervision on co-occurring concepts during network optimization through a novel language-guided contrastive learning loss. Our extensive experiments demonstrate the superiority of our approach over state-of-the-art methods, achieving substantial improvements of 3.8% and 1.7% on average across all metrics on the challenging benchmark datasets, Charades and MultiTHUMOS.

Asmar Nadeem, Faegheh Sardari, Robert Dawes et al.

Existing video captioning benchmarks and models lack causal-temporal narrative, which is sequences of events linked through cause and effect, unfolding over time and driven by characters or agents. This lack of narrative restricts models' ability to generate text descriptions that capture the causal and temporal dynamics inherent in video content. To address this gap, we propose NarrativeBridge, an approach comprising of: (1) a novel Causal-Temporal Narrative (CTN) captions benchmark generated using a large language model and few-shot prompting, explicitly encoding cause-effect temporal relationships in video descriptions; and (2) a Cause-Effect Network (CEN) with separate encoders for capturing cause and effect dynamics, enabling effective learning and generation of captions with causal-temporal narrative. Extensive experiments demonstrate that CEN significantly outperforms state-of-the-art models in articulating the causal and temporal aspects of video content: 17.88 and 17.44 CIDEr on the MSVD-CTN and MSRVTT-CTN datasets, respectively. Cross-dataset evaluations further showcase CEN's strong generalization capabilities. The proposed framework understands and generates nuanced text descriptions with intricate causal-temporal narrative structures present in videos, addressing a critical limitation in video captioning. For project details, visit https://narrativebridge.github.io/.

Faegheh Sardari, Armin Mustafa, Philip J. B. Jackson et al.

Unlike the sparse label action detection task, where a single action occurs in each timestamp of a video, in a dense multi-label scenario, actions can overlap. To address this challenging task, it is necessary to simultaneously learn (i) temporal dependencies and (ii) co-occurrence action relationships. Recent approaches model temporal information by extracting multi-scale features through hierarchical transformer-based networks. However, the self-attention mechanism in transformers inherently loses temporal positional information. We argue that combining this with multiple sub-sampling processes in hierarchical designs can lead to further loss of positional information. Preserving this information is essential for accurate action detection. In this paper, we address this issue by proposing a novel transformer-based network that (a) employs a non-hierarchical structure when modelling different ranges of temporal dependencies and (b) embeds relative positional encoding in its transformer layers. Furthermore, to model co-occurrence action relationships, current methods explicitly embed class relations into the transformer network. However, these approaches are not computationally efficient, as the network needs to compute all possible pair action class relations. We also overcome this challenge by introducing a novel learning paradigm that allows the network to benefit from explicitly modelling temporal co-occurrence action dependencies without imposing their additional computational costs during inference. We evaluate the performance of our proposed approach on two challenging dense multi-label benchmark datasets and show that our method improves the current state-of-the-art results.

Faegheh Sardari, Björn Ommer, Majid Mirmehdi

Most recent view-invariant action recognition and performance assessment approaches rely on a large amount of annotated 3D skeleton data to extract view-invariant features. However, acquiring 3D skeleton data can be cumbersome, if not impractical, in in-the-wild scenarios. To overcome this problem, we present a novel unsupervised approach that learns to extract view-invariant 3D human pose representation from a 2D image without using 3D joint data. Our model is trained by exploiting the intrinsic view-invariant properties of human pose between simultaneous frames from different viewpoints and their equivariant properties between augmented frames from the same viewpoint. We evaluate the learned view-invariant pose representations for two downstream tasks. We perform comparative experiments that show improvements on the state-of-the-art unsupervised cross-view action classification accuracy on NTU RGB+D by a significant margin, on both RGB and depth images. We also show the efficiency of transferring the learned representations from NTU RGB+D to obtain the first ever unsupervised cross-view and cross-subject rank correlation results on the multi-view human movement quality dataset, QMAR, and marginally improve on the-state-of-the-art supervised results for this dataset. We also carry out ablation studies to examine the contributions of the different components of our proposed network.

Faegheh Sardari, Adeline Paiement, Sion Hannuna et al.

We propose a view-invariant method towards the assessment of the quality of human movements which does not rely on skeleton data. Our end-to-end convolutional neural network consists of two stages, where at first a view-invariant trajectory descriptor for each body joint is generated from RGB images, and then the collection of trajectories for all joints are processed by an adapted, pre-trained 2D CNN (e.g. VGG-19 or ResNeXt-50) to learn the relationship amongst the different body parts and deliver a score for the movement quality. We release the only publicly-available, multi-view, non-skeleton, non-mocap, rehabilitation movement dataset (QMAR), and provide results for both cross-subject and cross-view scenarios on this dataset. We show that VI-Net achieves average rank correlation of 0.66 on cross-subject and 0.65 on unseen views when trained on only two views. We also evaluate the proposed method on the single-view rehabilitation dataset KIMORE and obtain 0.66 rank correlation against a baseline of 0.62.