Spiking Neural Networks (SNNs) become popular due to excellent energy efficiency, yet facing challenges for effective model training. Recent works improve this by introducing knowledge distillation (KD) techniques, with the pre-trained artificial neural networks (ANNs) used as teachers and the target SNNs as students. This is commonly accomplished through a straightforward element-wise alignment of intermediate features and prediction logits from ANNs and SNNs, often neglecting the intrinsic differences between their architectures. Specifically, ANN's outputs exhibit a continuous distribution, whereas SNN's outputs are characterized by sparsity and discreteness. To mitigate this issue, we introduce two innovative KD strategies. Firstly, we propose the Saliency-scaled Activation Map Distillation (SAMD), which aligns the spike activation map of the student SNN with the class-aware activation map of the teacher ANN. Rather than performing KD directly on the raw features of ANN and SNN, our SAMD directs the student to learn from saliency activation maps that exhibit greater semantic and distribution consistency. Additionally, we propose a Noise-smoothed Logits Distillation (NLD), which utilizes Gaussian noise to smooth the sparse logits of student SNN, facilitating the alignment with continuous logits from teacher ANN. Extensive experiments on multiple datasets demonstrate the effectiveness of our methods.

Vision Transformers (ViTs) have revolutionized computer vision, yet their self-attention mechanism lacks explicit spatial inductive biases, leading to suboptimal performance on spatially-structured tasks. Existing approaches introduce data-independent spatial decay based on fixed distance metrics, applying uniform attention weighting regardless of image content and limiting adaptability to diverse visual scenarios. Inspired by recent advances in large language models where content-aware gating mechanisms (e.g., GLA, HGRN2, FOX) significantly outperform static alternatives, we present the first successful adaptation of data-dependent spatial decay to 2D vision transformers. We introduce Spatial Decay Transformer (SDT), featuring a novel Context-Aware Gating (CAG) mechanism that generates dynamic, data-dependent decay for patch interactions. Our approach learns to modulate spatial attention based on both content relevance and spatial proximity. We address the fundamental challenge of 1D-to-2D adaptation through a unified spatial-content fusion framework that integrates manhattan distance-based spatial priors with learned content representations. Extensive experiments on ImageNet-1K classification and generation tasks demonstrate consistent improvements over strong baselines. Our work establishes data-dependent spatial decay as a new paradigm for enhancing spatial attention in vision transformers.

Referring Audio-Visual Segmentation (Ref-AVS) aims to segment target objects in audible videos based on given reference expressions. Prior works typically rely on learning latent embeddings via multimodal fusion to prompt a tunable SAM/SAM2 decoder for segmentation, which requires strong pixel-level supervision and lacks interpretability. From a novel perspective of explicit reference understanding, we propose TGS-Agent, which decomposes the task into a Think-Ground-Segment process, mimicking the human reasoning procedure by first identifying the referred object through multimodal analysis, followed by coarse-grained grounding and precise segmentation. To this end, we first propose Ref-Thinker, a multimodal language model capable of reasoning over textual, visual, and auditory cues. We construct an instruction-tuning dataset with explicit object-aware think-answer chains for Ref-Thinker fine-tuning. The object description inferred by Ref-Thinker is used as an explicit prompt for Grounding-DINO and SAM2, which perform grounding and segmentation without relying on pixel-level supervision. Additionally, we introduce R2-AVSBench, a new benchmark with linguistically diverse and reasoning-intensive references for better evaluating model generalization. Our approach achieves state-of-the-art results on both standard Ref-AVSBench and proposed R2-AVSBench.

The Dense Audio-Visual Event Localization (DAVEL) task aims to temporally localize events in untrimmed videos that occur simultaneously in both the audio and visual modalities. This paper explores DAVEL under a new and more challenging weakly-supervised setting (W-DAVEL task), where only video-level event labels are provided and the temporal boundaries of each event are unknown. We address W-DAVEL by exploiting cross-modal salient anchors, which are defined as reliable timestamps that are well predicted under weak supervision and exhibit highly consistent event semantics across audio and visual modalities. Specifically, we propose a Mutual Event Agreement Evaluation module, which generates an agreement score by measuring the discrepancy between the predicted audio and visual event classes. Then, the agreement score is utilized in a Cross-modal Salient Anchor Identification module, which identifies the audio and visual anchor features through global-video and local temporal window identification mechanisms. The anchor features after multimodal integration are fed into an Anchor-based Temporal Propagation module to enhance event semantic encoding in the original temporal audio and visual features, facilitating better temporal localization under weak supervision. We establish benchmarks for W-DAVEL on both the UnAV-100 and ActivityNet1.3 datasets. Extensive experiments demonstrate that our method achieves state-of-the-art performance.

Yanghao Zhou, Haitian Li, Rexar Lin, Heyan Huang, Jinxing Zhou, Changsen Yuan, Tian Lan, Ziqin Zhou, Yudong Li, Jiajun Xu, Jingyun Liao, YiMing Cheng, Xuefeng Chen, Xian-Ling Mao, Yousheng Feng. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Mainstream research in audio-visual learning has focused on designing task-specific expert models, primarily implemented through sophisticated multimodal fusion approaches. Recently, a few efforts have aimed to develop more task-independent or universal audiovisual embedding networks, encoding advanced representations for use in various audiovisual downstream tasks. This is typically achieved by fine-tuning large pretrained transformers, such as Swin-V2-L and HTS-AT, in a parameter-efficient manner through techniques such as tuning only a few adapter layers inserted into the pretrained transformer backbone. Although these methods are parameter-efficient, they suffer from significant training memory consumption due to gradient backpropagation through the deep transformer backbones, which limits accessibility for researchers with constrained computational resources. In this paper, we present Meta-Token Learning (Mettle), a simple and memory-efficient method for adapting large-scale pretrained transformer models to downstream audio-visual tasks. Instead of sequentially modifying the output feature distribution of the transformer backbone, Mettle utilizes a lightweight Layer-Centric Distillation (LCD) module to distill in parallel the intact audio or visual features embedded by each transformer layer into compact meta-tokens. This distillation process considers both pretrained knowledge preservation and task-specific adaptation. The obtained meta-tokens can be directly applied to classification tasks, such as audio-visual event localization and audio-visual video parsing. To further support fine-grained segmentation tasks, such as audio-visual segmentation, we introduce a Meta-Token Injection (MTI) module, which utilizes the audio and visual meta-tokens distilled from the top transformer layer to guide feature adaptation in earlier layers. Extensive experiments on multiple audiovisual benchmarks demonstrate that our method significantly reduces memory usage and training time while maintaining parameter efficiency and competitive accuracy.

Molecular representation learning plays a crucial role in various downstream tasks, such as molecular property prediction and drug design. To accurately represent molecules, Graph Neural Networks (GNNs) and Graph Transformers (GTs) have shown potential in the realm of self-supervised pretraining. However, existing approaches often overlook the relationship between molecular structure and electronic information, as well as the internal semantic reasoning within molecules. This omission of fundamental chemical knowledge in graph semantics leads to incomplete molecular representations, missing the integration of structural and electronic data. To address these issues, we introduce MOL-Mamba, a framework that enhances molecular representation by combining structural and electronic insights. MOL-Mamba consists of an Atom & Fragment Mamba-Graph (MG) for hierarchical structural reasoning and a Mamba-Transformer (MT) fuser for integrating molecular structure and electronic correlation learning. Additionally, we propose a Structural Distribution Collaborative Training and E-semantic Fusion Training framework to further enhance molecular representation learning. Extensive experiments demonstrate that MOL-Mamba outperforms state-of-the-art baselines across eleven chemical-biological molecular datasets.

Answering questions related to audio-visual scenes, i.e., the AVQA task, is becoming increasingly popular. A critical challenge is accurately identifying and tracking sounding objects related to the question along the timeline. In this paper, we present a new Patch-level Sounding Object Tracking (PSOT) method. It begins with a Motion-driven Key Patch Tracking (M-KPT) module, which relies on visual motion information to identify salient visual patches with significant movements that are more likely to relate to sounding objects and questions. We measure the patch-wise motion intensity map between neighboring video frames and utilize it to construct and guide a motion-driven graph network. Meanwhile, we design a Sound-driven KPT (S-KPT) module to explicitly track sounding patches. This module also involves a graph network, with the adjacency matrix regularized by the audio-visual correspondence map. The M-KPT and S-KPT modules are performed in parallel for each temporal segment, allowing balanced tracking of salient and sounding objects. Based on the tracked patches, we further propose a Question-driven KPT (Q-KPT) module to retain patches highly relevant to the question, ensuring the model focuses on the most informative clues. The audio-visual-question features are updated during the processing of these modules, which are then aggregated for final answer prediction. Extensive experiments on standard datasets demonstrate the effectiveness of our method, achieving competitive performance even compared to recent large-scale pretraining-based approaches.

Recent works on remote PhotoPlethysmoGraphy (rPPG) estimation typically use techniques like CNNs and Transformers to encode implicit features from facial videos for prediction. These methods learn to directly map facial videos to the static values of rPPG signals, overlooking the inherent dynamic characteristics of rPPG sequence. Moreover, the rPPG signal is extremely weak and highly susceptible to interference from various sources of noise, including illumination conditions, head movements, and variations in skin tone. To address these limitations, we propose a Physiology-based dynamicity disentangled diffusion (PhysDiff) model particularly designed for robust rPPG estimation. PhysDiff leverages the diffusion model to learn the distribution of quasi-periodic rPPG signal and uses a dynamicity disentanglement strategy to capture two dynamic characteristics in temporal rPPG signal, i.e., trend and amplitude. This disentanglement is motivated by the underlying dynamic physiological processes of vasodilation and vasoconstriction, ensuring a more precise representation of the rPPG signal. The disentangled components are then used as pivotal conditions in the proposed spatial-temporal hybrid denoiser for rPPG reconstruction. Besides, we introduce a periodicity-based multi-hypothesis selection strategy in model inference, which compares the natural periodicity of multiple generated rPPG hypotheses and selects the most favorable one as the final prediction. Extensive experiments on four datasets demonstrate that our PhysDiff significantly outperforms prior methods on both intra-dataset and cross-dataset testing.

The Audio-Visual Video Parsing task aims to recognize and temporally localize all events occurring in either the audio or visual stream, or both. Capturing accurate event semantics for each audio/visual segment is vital. Prior works directly utilize the extracted holistic audio and visual features for intra- and cross-modal temporal interactions. However, each segment may contain multiple events, resulting in semantically mixed holistic features that can lead to semantic interference during intra- or cross-modal interactions: the event semantics of one segment may incorporate semantics of unrelated events from other segments. To address this issue, our method begins with a Class-Aware Feature Decoupling (CAFD) module, which explicitly decouples the semantically mixed features into distinct class-wise features, including multiple event-specific features and a dedicated background feature. The decoupled class-wise features enable our model to selectively aggregate useful semantics for each segment from clearly matched classes contained in other segments, preventing semantic interference from irrelevant classes. Specifically, we further design a Fine-Grained Semantic Enhancement module for encoding intra- and cross-modal relations. It comprises a Segment-wise Event Co-occurrence Modeling (SECM) block and a Local-Global Semantic Fusion (LGSF) block. The SECM exploits inter-class dependencies of concurrent events within the same timestamp with the aid of a novel event co-occurrence loss. The LGSF further enhances the event semantics of each segment by incorporating relevant semantics from more informative global video features. Extensive experiments validate the effectiveness of the proposed modules and loss functions, resulting in a new state-of-the-art parsing performance.

In the field of audio-visual learning, most research tasks focus exclusively on short videos. This paper focuses on the more practical Dense Audio-Visual Event Localization (DAVEL) task, advancing audio-visual scene understanding for longer, untrimmed videos. This task seeks to identify and temporally pinpoint all events simultaneously occurring in both audio and visual streams. Typically, each video encompasses dense events of multiple classes, which may overlap on the timeline, each exhibiting varied durations. Given these challenges, effectively exploiting the audio-visual relations and the temporal features encoded at various granularities becomes crucial. To address these challenges, we introduce a novel CCNet, comprising two core modules: the Cross-Modal Consistency Collaboration (CMCC) and the Multi-Temporal Granularity Collaboration (MTGC). Specifically, the CMCC module contains two branches: a cross-modal interaction branch and a temporal consistency-gated branch. The former branch facilitates the aggregation of consistent event semantics across modalities through the encoding of audio-visual relations, while the latter branch guides one modality's focus to pivotal event-relevant temporal areas as discerned in the other modality. The MTGC module includes a coarse-to-fine collaboration block and a fine-to-coarse collaboration block, providing bidirectional support among coarse- and fine-grained temporal features. Extensive experiments on the UnAV-100 dataset validate our module design, resulting in a new state-of-the-art performance in dense audio-visual event localization.

In this paper, we propose a new multi-modal task, termed audio-visual instance segmentation (AVIS), which aims to simultaneously identify, segment and track individual sounding object instances in audible videos. To facilitate this research, we introduce a high-quality benchmark named AVISeg, containing over 90K instance masks from 26 semantic categories in 926 long videos. Additionally, we propose a strong baseline model for this task. Our model first localizes sound source within each frame, and condenses object-specific contexts into concise tokens. Then it builds long-range audio-visual dependencies between these tokens using window-based attention, and tracks sounding objects among the entire video sequences. Extensive experiments reveal that our method performs best on AVISeg, surpassing the existing methods from related tasks. We further conduct the evaluation on several multi-modal large models. Unfortunately, they exhibits subpar performance on instance-level sound source localization and temporal perception. We expect that AVIS will inspire the community towards a more comprehensive multi-modal understanding. Dataset and code is available at https://github.com/ruohaoguo/avis.

The Audio-Visual Event Localization (AVEL) task aims to temporally locate and classify video events that are both audible and visible. Most research in this field assumes a closed-set setting, which restricts these models’ ability to handle test data containing event categories absent (unseen) during training. Recently, a few studies have explored AVEL in an open-set setting, enabling the recognition of unseen events as "unknown", but without providing category-specific semantics. In this paper, we advance the field by introducing the Open-Vocabulary Audio-Visual Event Localization (OV-AVEL) problem, which requires localizing audio-visual events and predicting explicit categories for both seen and unseen test data at inference. To address this new task, we propose the OV-AVEBench dataset, comprising 24,800 videos across 67 real-life audiovisual scenes (seen:unseen = 46:21), each with manual segment-level annotation. We also establish three evaluation metrics for this task. Moreover, we investigate two baseline approaches, one training-free and one using a further fine-tuning paradigm. Specifically, we utilize the unified multimodal space from the pretrained ImageBind model to extract audio, visual, and textual (event classes) features. The training-free baseline then determines predictions by comparing the consistency of audio-text and visual-text feature similarities. The fine-tuning baseline incorporates lightweight temporal layers to encode temporal relations within the audio and visual modalities, using OVAVEBench training data for model fine-tuning. We evaluate these baselines on the proposed OV-AVEBench dataset and discuss potential directions for future work in this new field.

Yevheniia Kryklyvets, Mohammed Irfan Kurpath, Sahal Shaji Mullappilly, Jinxing Zhou, Fahad Shahbaz Khan, Rao Muhammad Anwer, Salman Khan, Hisham Cholakkal. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. 2025.

The Audio-visual Classification (AVC) task aims to determine video categories by integrating audio and visual signals. Traditional methods for AVC leverage Artificial Neural Networks (ANNs) that operate on floating-point features, affording large parameter counts and consuming extensive energy. Recent research has shifted towards brain-inspired Spiking Neural Networks (SNNs), which transmit audiovisual information through sparser 0/1 spike features allowing for better energy efficiency. However, a byproduct of such sparsity is the increased difficulty in effectively encoding and utilizing these spike features. Moreover, the spike firing characteristics based on neuron membrane potential cause asynchronous spike activations due to the heterogeneous distributions of different modalities in the AVC task, resulting in cross-modal asynchronization. This issue is often overlooked by prior SNN models, resulting in lower classification accuracy compared to traditional ANN models. To address these challenges, we present a new Multimodal Interaction Spiking Network (MISNet), the first to successfully balance both accuracy and efficiency for the AVC task. As the core of MISNet, we propose a Multimodal Leaky Integrate-and-fire (MLIF) neuron, which coordinates and synchronizes the spike activations of audiovisual signals within a single neuron, distinguishing it from the prior paradigm of SNNs that relies on multiple separate processing neurons. As a result, our MISNet enables to generate audio and visual spiking features with effective cross-modal fusion. Additionally, we propose to add extra loss regularizations before fusing the obtained audio-visual features for final classification, thereby benefiting unimodal spiking learning for multimodal interaction. We evaluate our method on five audio-visual datasets, demonstrating advanced performance in both accuracy and energy consumption.

Repetitive Action Counting (RAC) is a critical and challenging task in video analysis, aiming to count the number of repeated actions in videos accurately. Existing methods typically generate a Temporal Self-similarity Matrix (TSM) as an intermediate representation to predict the number of repetitive actions. While this simplifies the process, it often overlooks the variable lengths between action cycles and the phenomenon of motion interruptions. The period inconsistency problem caused by the change in the action period and the motion interruption problem resulting from the motion pause are the two main challenges that affect the accuracy of RAC in complex scenes. To address these challenges, we propose a novel framework. First, we construct a boundary-aware encoder equipped with a temporal pyramid structure to build multi-scale video features, capturing the period information of different lengths of repetitive actions to solve the period inconsistency problem. Next, a cycle and boundary attention module is followed by each layer in the pyramid to enhance these multi-scale features with periodic and event boundary information. Finally, we design a gated density estimator to generate the actionness score for each frame that reflects the probability of the corresponding time point being within the motion cycle. These scores are used to weight features to reduce the impact of noise frames without actions present and solve the motion interruption problem for better density prediction. Extensive experiments conducted on public datasets demonstrate the effectiveness of our method. The source code will be available at https://github.com/zqzhang2023/TBANRAC .

This paper focuses on the Audio-Visual Question Answering (AVQA) task that aims to answer questions derived from untrimmed audible videos. To generate accurate answers, an AVQA model is expected to find the most informative audio-visual clues relevant to the given questions. In this paper, we propose to explicitly consider fine-grained visual objects in video frames (object-level clues) and explore the multi-modal relations (\textit{i.e.}, the object, audio, and question) in terms of feature interaction and model optimization. For the former, we present an end-to-end object-oriented network that adopts a question-conditioned clue discovery module to concentrate audio/visual modalities on respective keywords of the question and designs a modality-conditioned clue collection module to highlight closely associated audio segments or visual objects. For model optimization, we propose an object-aware adaptive-positivity learning strategy that selects the highly semantic-matched multi-modal pair as \textit{positivity}. Specifically, we design two object-aware contrastive loss functions to identify the highly relevant question-object pairs and audio-object pairs, respectively. These selected pairs are constrained to have larger similarity values than the mismatched pairs. The positivity-selecting process is adaptive as the positivity pairs selected in each video frame may be different. These two object-aware objectives help the model understand \textit{which objects are exactly relevant to the question} and \textit{which are making sounds}. Extensive experiments on the MUSIC-AVQA dataset demonstrate the proposed method is effective in finding favorable audio-visual clues and also achieves new state-of-the-art question-answering performance. The code is available at https://github.com/zhangbin-ai/APL.

The Text to Audible-Video Generation (TAVG) task involves generating videos with accompanying audio based on text descriptions. Achieving this requires skillful alignment of both audio and video elements. To support research in this field, we have developed a comprehensive Text to Audible-Video Generation Benchmark (TAVGBench), which contains over 1.7 million clips with a total duration of 11.8 thousand hours. We propose an automatic annotation pipeline to ensure each audible video has detailed descriptions for both its audio and video contents. We also introduce the Audio-Visual Harmoni score (AVHScore) to provide a quantitative measure of the alignment between the generated audio and video modalities. Additionally, we present a baseline model for TAVG called TAVDiffusion, which uses a two-stream latent diffusion model to provide a fundamental starting point for further research in this area. We achieve the alignment of audio and video by employing cross-attention and contrastive learning. Through extensive experiments and evaluations on TAVGBench, we demonstrate the effectiveness of our proposed model under both conventional metrics and our proposed metrics. The dataset and code can be found on this page https://npucvr.github.io/TAVGBench/ and on github https://github.com/OpenNLPLab/TAVGBench.

We explore a new task for audio-visual-language modeling called fine-grained audible video description (FAVD). It aims to provide detailed textual descriptions for the given audible videos, including the appearance and spatial locations of each object, the actions of moving objects, and the sounds in videos. Existing visual-language modeling tasks often concentrate on visual cues in videos while undervaluing the language and audio modalities. On the other hand, FAVD requires not only audio-visual-language modeling skills but also paragraph-level language generation abilities. We construct the first fine-grained audible video description benchmark (FAVDBench) to facilitate this research. For each video clip, we first provide a one-sentence summary of the video, i.e., the caption, followed by 4–6 sentences describing the visual details and 1–2 audio-related descriptions at the end. The descriptions are provided in both English and Chinese. We create two new metrics for this task: an EntityScore to gauge the completeness of entities in the visual descriptions, and an AudioScore to assess the audio descriptions. As a preliminary approach to this task, we propose an audio-visual-language transformer that extends existing video captioning model with an additional audio branch. We combine the masked language modeling and auto-regressive language modeling losses to optimize our model so that it can produce paragraph-level descriptions. We illustrate the efficiency of our model in audio-visual-language modeling by evaluating it against the proposed benchmark using both conventional captioning metrics and our proposed metrics. We further put our benchmark to the test in video generation models, demonstrating that employing fine-grained video descriptions can create more intricate videos than using captions. Code and dataset are available at https://github.com/OpenNLPLab/FAVDBench. Our online benchmark is available at www.avlbench.opennlplab.cn.

Visual and audio signals often coexist in natural environments, forming audio-visual events (AVEs). Given a video, we aim to localize video segments containing an AVE and identify its category. It is pivotal to learn the discriminative features for each video segment. Unlike existing work focusing on audio-visual feature fusion, in this paper, we propose a new contrastive positive sample propagation (CPSP) method for better deep feature representation learning. The contribution of CPSP is to introduce the available full or weak label as a prior that constructs the exact positive-negative samples for contrastive learning. Specifically, the CPSP involves comprehensive contrastive constraints: pair-level positive sample propagation (PSP), segment-level and video-level positive sample activation (PSA$_{S}$and PSA$_{V}$). Three new contrastive objectives are proposed (i.e.,$\mathcal {L}_{\text{avpsp}}$,$\mathcal {L}_ \text{spsa}$, and$\mathcal {L}_\text{vpsa}$) and introduced into both the fully and weakly supervised AVE localization. To draw a complete picture of the contrastive learning in AVE localization, we also study the self-supervised positive sample propagation (SSPSP). As a result, CPSP is more helpful to obtain the refined audio-visual features that are distinguishable from the negatives, thus benefiting the classifier prediction. Extensive experiments on the AVE and the newly collected VGGSound-AVEL100k datasets verify the effectiveness and generalization ability of our method.

Most works of image captioning are implemented under the full supervision of paired image-caption data. Limited to expensive cost of data collection, the task of unpaired image captioning has attracted researchers' attention. In this article, we propose a novel memorial GAN (MemGAN) with the joint semantic optimization for unpaired image captioning. The core idea is to explore implicit semantic correlation between disjointed images and sentences through building a multimodal semantic-aware space (SAS). Concretely, each modality is mapped into a unified multimodal SAS, where SAS includes the semantic vectors of image I , visual concepts O , unpaired sentence S , and the generated caption C . We adopt the memory unit based on multihead attention and relational gate as a backbone to preserve and transit crucial multimodal semantics in the SAS for image caption generation and sentence reconstruction. Then, the memory unit is embedded into a GAN framework to exploit the semantic similarity and relevance in SAS, that is, imposing a joint semantic-aware optimization on SAS without supervision clues. To summarize, the proposed MemGAN learns the latent semantic relevance of SAS's multimodalities in an adversarial manner. Extensive experiments and qualitative results demonstrate the effectiveness of MemGAN, achieving improvements over state of the arts on unpaired image captioning benchmarks.

Visual and audio signals often coexist in natural environments, forming audio-visual events (AVEs). Given a video, we aim to localize video segments containing an AVE and identify its category. In order to learn discriminative features for a classifier, it is pivotal to identify the helpful (or positive) audio-visual segment pairs while filtering out the irrelevant ones, regardless whether they are synchronized or not. To this end, we propose a new positive sample propagation (PSP) module to discover and exploit the closely related audio-visual pairs by evaluating the relationship within every possible pair. It can be done by constructing an all-pair similarity map between each audio and visual segment, and only aggregating the features from the pairs with high similarity scores. To encourage the network to extract high correlated features for positive samples, a new audio-visual pair similarity loss is proposed. We also propose a new weighting branch to better exploit the temporal correlations in weakly supervised setting. We perform extensive experiments on the public AVE dataset and achieve new state-of-the-art accuracy in both fully and weakly supervised settings, thus verifying the effectiveness of our method.