Abstract:This paper introduces a modeling approach that employs multi-level global processing, encompassing both short-term frame-level and long-term sample-level feature scales. In the initial stage of shallow feature extraction, various scales are employed to extract multi-level features, including Mel-Frequency Cepstral Coefficients (MFCC) and pre-Fbank log energy spectrum. The construction of the identification network model involves considering the input two-dimensional temporal features from both frame and sample levels. Specifically, the model initially employs one-dimensional convolution-based Convolutional Long Short-Term Memory (ConvLSTM) to fuse spatiotemporal information and extract short-term frame-level features. Subsequently, bidirectional long Short-Term Memory (BiLSTM) is utilized to learn long-term sample-level sequential representations. The transformer encoder then performs cross-scale, multi-level processing on global frame-level and sample-level features, facilitating deep feature representation and fusion at both levels. Finally, recognition results are obtained through Softmax. Our method achieves an impressive 99.6% recognition accuracy on the CCNU_Mobile dataset, exhibiting a notable improvement of 2% to 12% compared to the baseline system. Additionally, we thoroughly investigate the transferability of our model, achieving an 87.9% accuracy in a classification task on a new dataset.
Abstract:Audio-Visual Source Localization (AVSL) aims to locate sounding objects within video frames given the paired audio clips. Existing methods predominantly rely on self-supervised contrastive learning of audio-visual correspondence. Without any bounding-box annotations, they struggle to achieve precise localization, especially for small objects, and suffer from blurry boundaries and false positives. Moreover, the naive semi-supervised method is poor in fully leveraging the information of abundant unlabeled data. In this paper, we propose a novel semi-supervised learning framework for AVSL, namely Dual Mean-Teacher (DMT), comprising two teacher-student structures to circumvent the confirmation bias issue. Specifically, two teachers, pre-trained on limited labeled data, are employed to filter out noisy samples via the consensus between their predictions, and then generate high-quality pseudo-labels by intersecting their confidence maps. The sufficient utilization of both labeled and unlabeled data and the proposed unbiased framework enable DMT to outperform current state-of-the-art methods by a large margin, with CIoU of 90.4% and 48.8% on Flickr-SoundNet and VGG-Sound Source, obtaining 8.9%, 9.6% and 4.6%, 6.4% improvements over self- and semi-supervised methods respectively, given only 3% positional-annotations. We also extend our framework to some existing AVSL methods and consistently boost their performance.
Abstract:Audio-Visual Source Localization (AVSL) is the task of identifying specific sounding objects in the scene given audio cues. In our work, we focus on semi-supervised AVSL with pseudo-labeling. To address the issues with vanilla hard pseudo-labels including bias accumulation, noise sensitivity, and instability, we propose a novel method named Cross Pseudo-Labeling (XPL), wherein two models learn from each other with the cross-refine mechanism to avoid bias accumulation. We equip XPL with two effective components. Firstly, the soft pseudo-labels with sharpening and pseudo-label exponential moving average mechanisms enable models to achieve gradual self-improvement and ensure stable training. Secondly, the curriculum data selection module adaptively selects pseudo-labels with high quality during training to mitigate potential bias. Experimental results demonstrate that XPL significantly outperforms existing methods, achieving state-of-the-art performance while effectively mitigating confirmation bias and ensuring training stability.