@techreport{Boes2021,
    Author = "Boes, Wim and Van hamme, Hugo",
    title = "Multi-Source Transformer Architectures for Audiovisual Scene Classification",
    institution = "DCASE2021 Challenge",
    year = "2021",
    month = "June",
    abstract = "In this technical report, the systems we submitted for subtask 1B of the DCASE 2021 challenge, regarding audiovisual scene classi- fication, are described in detail. They are essentially multi-source transformers employing a combination of auditory and visual fea- tures to make predictions. These models are evaluated utilizing the macro-averaged multi-class cross-entropy and accuracy metrics. In terms of the macro-averaged multi-class cross-entropy, our best model achieved a score of 0.620 on the validation data. This is slightly better than the performance of the baseline system (0.658). With regard to the accuracy measure, our best model achieved a score of 77.1\% on the validation data, which is about the same as the performance obtained by the baseline system (77.0\%)."
}

@techreport{Diez2021,
    Author = "Diez, Itxasne and Saratxaga, Ibon",
    title = "Audio Scene Classification Using Enhanced Convolutional Neural Networks for {DCASE} 2021 Challenge",
    institution = "DCASE2021 Challenge",
    year = "2021",
    month = "June",
    abstract = "This technical report describes our system proposed for Task 1B – Audio-Visual Scene Classification of the DCASE 2021 Challenge. Our system focuses in the audio signal based classification. The system has an architecture based on the combination of Convolutional Neural Networks and OpenL3 embeddings. The CNN consist of three stacked 2D convolutional layers to process the log-melspectrogram parameters obtained from the input signals. Additionally OpenL3 embeddings of the input signals are also calculated and merged with the output of the CNN stack. The resulting vector is fed to a classification block consisting of two fully connected layers. Mixup augmentation technique is applied to the training data and binaural data is also used as input to provide additional information. In this report, we describe the proposed systems in detail and com- pare them to the baseline approach using the provided development datasets."
}

@techreport{Fedorishin2021,
    Author = "Fedorishin, Dennis and Sankaran, Nishant and Mohan, Deen and Birgiolas, Justas and Schneider, Philip and Setlur, Srirangaraj and Govindaraju, Venu",
    title = "Investigating Waveform and Spectrogram Feature Fusion for Audio Classification",
    institution = "DCASE2021 Challenge",
    year = "2021",
    month = "June",
    abstract = "This technical report presents our submitted system for the DCASE 2021 Challenge Task1B: Audio-Visual Scene Classifica- tion. Focusing on the audio modality only, we investigate the use of two common feature representations within the audio understand- ing domain, the raw waveform and Mel-spectrogram, and measure their degree of complementarity when using both representations in a fusion setting. We introduce a new model paradigm for audio clas- sification by fusing features learned from Mel-spectrograms and the raw waveform from separate feature extraction branches. Our ex- perimental results show that our proposed fusion model has a 4.5\% increase in validation accuracy and a reduction of .14 in validation loss over the Task 1B baseline audio-only subnetwork. We further show that learned features of raw waveforms and Mel-spectrograms are indeed complementary to each other and that there is a consis- tent classification performance improvement over models trained on Mel-spectrograms alone."
}

@techreport{Hou2021,
    Author = "Hou, Yuanbo and Tan, Yizhou and Chang, Yue and Huang, Tianyang and Li, Shengchen and Shao, Xi and Botteldooren, Dick",
    title = "{CNN}-Based Dual-Stream Network for Audio-Visual Scene Classification",
    institution = "DCASE2021 Challenge",
    year = "2021",
    month = "June",
    abstract = "This technical report presents the CNN-based dual-stream network for audio-visual scene classification in DCASE 2021 Challenge (Task 1 Subtask B). The proposed method in this report is only trained based on the development dataset in Task 1 Subtask B and does not use any external dataset. For the performance, the model proposed in this report gets 0.318 log-loss and 90.0\% accuracy for scene classification on the development dataset, and the log-loss and accuracy in the baseline are 0.658 and 77.0\%, respectively. Our results are reproducible, source code is available here: https://github.com/Yuanbo2020/DCASE2021-T1B."
}

@techreport{Naranjo-Alcazar2021_t1b,
    Author = "Naranjo-Alcazar, Javier and Perez-Castanos, Sergi and Cobos, Maximo and Ferri, Francesc J. and Zuccarello, Pedro",
    title = "Task {1B} {DCASE} 2021: Audio-Visual Scene Classification with Squeeze-Excitation Convolutional Recurrent Neural Networks",
    institution = "DCASE2021 Challenge",
    year = "2021",
    month = "June",
    abstract = "Automatic scene classification has always been one of the core tasks in every edition of the DCASE challenge. Until this edition, such classification was performed using only audio data, and so the problematic was defined as Acoustic Scene Classification (ASC). In this 2021 edition, audio data is accompanied with visual data, providing additional information that can be jointly exploited for achieving higher recognition accuracy. The proposed approach makes use of two separate networks which are respectively trained in isolation on audio and visual data, so that each network specializes in a given modality. After training each network, the fusion of information from the audio and visual subnetworks is performed at two different stages. The early fusion stage combines features resulting from the last convolutional block of the respective subnetworks at different time steps to feed a bidirectional recurrent structure. The late fusion stage combines the output of the early fusion stage with the independent predictions provided by the two subnetworks, resulting in the final prediction. For the visual subnetwork, a VGG16 architecture pretrained on the Places365 dataset is used, applying a fine-tuning strategy over the Challenge dataset. On the other hand, the audio subnetwork is trained from scratch and uses squeeze- excitation techniques as in previous contributions from this team. As a result, the final accuracy of the system is 92\% on development split, outperforming the baseline by 15 percentage points."
}

@techreport{Okazaki2021,
    Author = "Okazaki, Soichiro and Quan, Kong and Yoshinaga, Tomoaki",
    title = "Ldslvision Submissions to Dcase’21: A Multi-Modal Fusion Approach for Audio-Visual Scene Classification Enhanced by Clip Variants",
    institution = "DCASE2021 Challenge",
    year = "2021",
    month = "June",
    abstract = "In this report, we describe our solution for audio-visual scene classification task of DCASE2021 challenge Task1B. Our solution is based on a multi-modal fusion approach consisting of three different domain features: (1) Log-mel spectrogram audio features extracted by CNN variants from audio files. (2) Frame-wise image features extracted by CNN variants from video files. (3) Text-guided frame- wise image features extracted by CLIP variants from video files. We trained three domain models respectively and created final sub- missions by ensembling the class-wise confidences of three domain models’ outputs. With ensembling and post-processing for the confidences, our model reached 0.149 log-loss (official baseline: 0.658 log-loss) and 96.1\% accuracy (official baseline: 77.0\% accuracy) on the officially provided fold1 evaluation dataset of Task1B."
}

@techreport{Peng2021,
    Author = "Peng, Wang and Zhang, Tianyang and Zou, Zehua",
    title = "Convolutional Receptive Field Dual Selection Mechanism for Acoustic Scene Classification",
    institution = "DCASE2021 Challenge",
    year = "2021",
    month = "June",
    abstract = "Convolution neural network (CNN), which can extract rich semantic information of signal, is a representative feature learing network in acoustic scene classification (ASC). However, since that the receptive field (RF) of a CNN is fixed, it is inefficient to capture the dynamical time-frequency changing characteristic of the input Log-Mel spectrogram. In addition, although the Log-Mel spectrogram can be treated as an image, the time and frequency dimensions, which respectively represent the acoustic event duration and frequency information, have different physical meanings. Therefore, existing receptive field adaptive meth- ods, which get same-sized optimal receptive fields in two dimensions, are not suitable for ASC. To tackle this problem, we proposed a convolution receptive field dual selection mechanism (CRFDS) in this paper. Acoustic scene classification experiments conducted on DCASE 2021 subtask B with audio-only show that the accuracy of CRFDS can achieve 71.45\%."
}

@techreport{Pham2021,
    Author = "Pham, Lam and Schindler, Alexander and Schutz, Mina and Lampert, Jasmin and King, Ross",
    title = "{DCASE} 2021 Task {1B}: Technique Report",
    institution = "DCASE2021 Challenge",
    year = "2021",
    month = "June",
    abstract = "Abstract—This report shows a deep learning framework for audio-visual scene classification (SC). Our extensive experiments, which are conducted on DCASE Task 1B development dataset, achieve the best classification accuracy of 82.2\%, 91.1\%, and 93.9\% with audio input only, visual input only, and both audio- visual input, respectively."
}

@techreport{Triantafyllopoulos2021,
    Author = "Triantafyllopoulos, Andreas and Drossos, Konstantinos and Gebhard, Alexander and Alice, Baird and Björn, Schuller",
    title = "A Multimodal Wavetransformer Architecture Conditioned on Openl3 Embeddings for Audio-Visual Scene Classification",
    institution = "DCASE2021 Challenge",
    year = "2021",
    month = "June",
    abstract = "In this report, we present our submission systems to TASK1B of the DCASE2021 Challenge. We submit a total of four systems: one purely audio-based and three multimodal variants of the same architecture. The main module consists of the WaveTransformer architecture, which was recently introduced for automatic audio captioning (AAC). We first adapt the architecture to the task of acoustic scene classification (ASC), and then extend it to handle multimodal signals by globally conditioning all layers on multimodal OpenL3 embeddings. As data augmentation, we apply time- and frequency- bin masking, as well as random cropping. Our best-effort system achieves a log-loss of 0.568 and an accuracy of 79.5\%."
}

@techreport{Wang2021,
    Author = "Wang, Qing and Zheng, Siyuan and Li, Yunqing and Wang, Yajian and Wu, Yuzhong and Hu, Hu and Yang, Chao-Han Huck and Siniscalchi, Sabato Marco and Wang, Yannan and Du, Jun and Lee, Chin-Hui",
    title = "A Model Ensemble Approach for Audio-Visual Scene Classification",
    institution = "DCASE2021 Challenge",
    year = "2021",
    month = "June",
    abstract = "In this technical report, we present our approach to Task 1b - Audio- Visual Scene Classification (AVSC) in the DCASE 2021 Challenge. We employ pre-trained networks trained on image datasets to ex- tract video embedding whereas for audio embedding models trained from scratch are more appropriate for feature extraction. We pro- pose several models for the AVSC task based on different audio and video embeddings using early fusion strategy. Besides, we propose to use audio-visual segment model (AVSM) to extract text embed- ding. Data augmentation methods are used during training. Fur- thermore, a two-stage classification strategy is adopted by leverag- ing on score fusion of two classifiers. Finally, model ensemble of two-stage AVSC classifiers is used to obtained more robust predic- tions. The proposed systems are evaluated on the development set of TAU Urban Audio Visual Scenes 2021. Compared with official baseline, our approach can achieve a much lower log loss of 0.141 and a much higher accuracy of 95.3\%."
}

@techreport{Wang2021a,
    Author = "Wang, Yuxiang and liang, Shuang",
    title = "Bit Submission for {DCASE} 2020 Challenge Task1",
    institution = "DCASE2021 Challenge",
    year = "2021",
    month = "June",
    abstract = "DCASE2021 challenge task 1 contains two sub-tasks: (i) LowComplexity Acoustic Scene Classification with Multiple Devices and (ii) Audio-Visual Scene Classification. In our submission systems, different methods are used for different tasks. For task 1a, we explore the fsFCNN (frequency sub-sampling controlled fully convolution) with two-stage training. In order to reduce the model size, the knowledge distillation approach is used. For task 1b, different models are used for different modals. For audio classification, the same fsFCNN structures with two-stage training are applied. For video classification, the TimeSformer model is used. Experimental results show that our final model obtain accuracy of 64.6\% with 128kb model size. On task 1b development set, the audio modal achieve 80.6\% accuracy and 92\% for video modal."
}

@techreport{Wang2021b,
    Author = "Wang, Meng and Chen, Chengxin and Xie, Yuan and Chen, Hangting and Liu, Yuzhuo and Zhang, Pengyuan",
    title = "Audio-Visual Scene Classification Using Transfer Learning and Hybrid Fusion Strategy",
    institution = "DCASE2021 Challenge",
    year = "2021",
    month = "June",
    abstract = "In this technical report, we describe the details of our submission for DCASE2021 Task1b. This task focuses on audio-visual scene classification. We use 1D deep convolutional neural network integrated with three different acoustic features in our audio system, and perform a two-stage fine-tuning on some pre-trained models such as ResNet-50 and EfficientNet-b5 in our image system. In model-level fusion, the extracted audio and image embeddings are concatenated as input into a classifier. We also use decision-level fusion to make our system more robust. On the official train/test setup of the development dataset, our best single audio-visual system obtained a 0.159 log loss and 94.1\% accuracy compared to 0.623 and 78.5\% for the audio-only system and 0.270 and 91.8\% for the image-only system. Our final fusion system could achieve a 0.143 log loss and 95.2\% accuracy."
}

@techreport{Yang2021,
    Author = "Yang, Yujie and Luo, Yanan",
    title = "Scene Classification Using Acoustic and Visual Feature",
    institution = "DCASE2021 Challenge",
    year = "2021",
    month = "June",
    abstract = "In this report, we provide a brief overview of our submission for the audio-visual scene classification task of the DCASE 2021 challenge. This report focuses on the joint use of audio and video features to improve the performance of scene classification. In order to extract audio features, we train a convolutional neural network similar to the VGG to classify the log-mel spectra. In order to extract video features, we use ResNext to train an image classifier. Subsequently we use both features to do the classification, which can achieve better performance than using only one feature to make the classification."
}