Extracting information from audio-visual (AV) data about events, objects, sounds and scenes finds important applications in several areas such as video surveillance, multimedia indexing and robotics. Among other tasks, automatic analysis of AV scenes entails: (i) identifying events or objects, (ii) localizing them in space and time, and (iii) extracting the audio source of interest from the background. In our efforts to build a unified framework to deal with these challenging problems, we presented a first system tackling event identification and AV localization earlier [1, 2]. Continuing to build upon that study, in this paper we focus on making event/object classification robust to noisy acoustic environments and incorporating the ability to enhance or separate the object in the audio modality.
There is a long history of works on supervised event detection [3, 4, 5, 6]. However, scaling supervision to large video collections and obtaining precise annotations for multiple tasks is both time consuming and error prone [7, 8]. Hence, in our previous work  we resort to training with weak labels i.e. global video-level object labels without any timing information. Multiple instance learning (MIL) is a well-known learning paradigm central to most studies using weak supervision 
. MIL is typically applied to cases where labels are available over bags (sets of instances) instead of individual instances. The task then amounts to jointly selecting appropriate instances and estimating classifier parameters. For applying this to our case, let us begin by viewing a video as a labeled bag, containing a collection of image regions (also referred to as image proposals) and audio segments (also referred to as audio proposals) obtained by chunking the audio temporally. While such a formulation yields promising results using deep MIL[1, 10], its audio proposal design has two shortcomings with respect to our goals: it is (i) prone to erroneous classification in noisy acoustic conditions and (ii) limited to temporal localization of the audio event or object, thus does not allow for time-frequency segmentation in order to extract the audio source of interest. To address these shortcomings, we propose to generate audio proposals using non-negative matrix factorization (NMF) . Note that the term proposal refers to image or audio “parts” that may potentially contain the object of interest. For the audio modality these “parts” can be obtained through uniform chunking of the signal, as we did previously, or more sophisticated methods.
. It factorizes an audio spectrogram into two nonnegative matrices namely, so-called spectral patterns and their activations. Such a part-based decomposition is analogous to breaking up an image into constituent object regions. This motivates its use in our system. It makes it possible not only to de-noise the audio, but also to appropriately combine the parts for separation. An interesting work which has appeared recently uses NMF basis vectors with weak supervision from visual modality to perform audio source separation
. There are three key differences with our proposed approach: (i) The authors of that proposal use the NMF basis vectors and not their activations for training the system. Hence no temporal information is utilized. (ii) Unlike us, they perform a supervised dictionary construction step after training to decompose a test signal (iii) Finally, they do not consider the task of visual localization. Other recent approaches for deep learning based vision-guided audio source separation methods utilize ground-truth source masks for training[16, 17]. It is worth noting that our proposed enhancement technique is significantly different as we do not use separated ground truth sources at any stage and only rely on weak labels. This makes the problem considerably more challenging.
Contributions. Building upon our inital work 
, we show how a deep MIL framework can be flexibly used to robustly perform several audio-visual scene understanding tasks using just weak labels. In particular, in addition to temporal audio proposals as in we propose to use NMF components as audio proposals for improved classification and to allow source enhancement. We demonstrate the usefulness of such an approach on a large dataset of unconstrained musical instrument performance videos. As the data is noisy, we expect NMF decomposition to provide additional, possibly “cleaner” information about the source of interest. Moreover, scores assigned to each component by the MIL module to indicate their relevance for classification can be reliably used to enhance or separate multiple sources.
2 Proposed Approach
The proposed approach is depicted in Fig. 1
. We formulate the problem within a deep MIL framework. Each video is considered as a bag of visual and audio proposals. These proposals are fed to their respective feature extraction and scoring networks. The scores indicate relevance of each region or segment for a particular class. Their aggregation, as depicted in Fig.1, allows video-level classification. In the following section we discuss proposal generation, feature extraction, scoring and training procedures in detail.
2.1 System Details
Visual Proposals. As our goal is to localize spatially and temporally the most discriminative image region pertaining to a class, we choose to generate proposals over video frames sub-sampled at a rate of 1 frame per second. Class-agnostic bounding-box proposals are obtained using the well-known EdgeBoxes  algorithm. To reduce the computational load and redundancy, the confidence score generated by this method is used to select top proposals from each sampled image. Hence, for a 10 sec. video, such a procedure would generate a list of region proposals.
A fixed-length feature vector, is obtained from each image region proposal, in a video
, using a convolutional neural network altered with a region-of-interest (RoI) pooling layer. In practice, feature vectors
are passed through two fully connected layers, which are fine-tuned during training. Typically, standard CNN architectures pre-trained on ImageNet classification are used for the purpose of initializing network weights (see Section 3 for implementation details).
Audio Proposals. We study two kinds of proposals:
Temporal Segment Proposals (TSP): Herein the audio is simply decomposed into temporal segments of equal length, . These proposals are obtained by transforming the raw audio waveform into log-Mel spectrogram and subsequently chunking it by sliding a fixed-length window along the temporal axis. The dimensions of this window are chosen to be compatible with the base audio network (see Sec. 3).
NMF Component Proposals (NCP): Using NMF we decompose audio magnitude spectrogram consisting of frequency bins and
short-time Fourier transform (STFT) frames, such that,
where and are interpreted as the nonnegative audio spectral patterns and their temporal activation matrices respectively. Here is the total number of spectral patterns. To estimate and we minimize the Kullback-Leibler (KL) divergence using multiplicative update rules  where and are initialized randomly.
We now apply NMF-based Wiener filtering, as in , to an audio recording to decompose it into tracks (also referred to as NMF components) each obtained from for , where and denote spectral patterns and activations corresponding to the component, respectively. They can now be considered as proposals that may or may not belong to the class of interest. Specifically, we chunk each NMF component into temporal segments, which we call NMF Component proposals or NCPs. We denote the set of NCPs by , where each element is indexed by the component, and temporal segment number. As the same audio network is used for both kinds of audio proposals, for each NMF component or track we follow the TSP computation procedure. However, this is done with a non-overlapping window for reducing computational load.
Proposals generated by both the aforementioned methods are passed through a VGG-style deep network known as vggish  for base audio feature extraction. Hershey et al. introduced this state-of-the-art audio feature extractor as an audio counterpart to networks pre-trained on ImageNet for classification. vggish has been pre-trained on a preliminary version of YouTube-8M  for audio classification based on video tags. It generates a 128 dimensional embedding for each input log-Mel spectrogram segment with 64 Mel-bands and 96 temporal frames. We fine-tune all the layers of vggish during training.
Proposal Scoring and Fusion. Having obtained representations for each proposal in both the modalities, we now score them with respect to classes using the two-stream architecture put forth by Bilen et al. . This module consists of parallel classification and localization streams. Generically denoting audio or visual proposals by and their dimensional input representations to the scoring module by , the following sequence of operations is carried out: First, is passed through linear fully-connected layers of both classification and localization streams (shown with yellow in Fig. 1) giving transformed matrices and , respectively. This is followed by a softmax operation on in the localization stream, written as:
This allows the localization layer to choose the most relevant proposals for each class. Subsequently, the classification stream output is weighted by through element-wise multiplication: . Class scores over the video are obtained by summing the resulting weighted scores in over all the proposals.
The same set of operations is carried out for both audio and visual proposals. Before addition of global level scores from both the modalities, they are -normalized to ensure similar score range.
Classification Loss and Training. Given a set of training videos and labels, , we solve a multi-label classification problem. Here with the class presence denoted by and absence by . To recall, for each video , the network takes as input a set of image regions along with audio TSP , NCP or both. After performing the described operations on each modality separately, the normalized scores are added and represented by , with all network weights and biases denoted by . Both sub-modules are trained jointly using the multi-label hinge loss:
2.2 Source Enhancement
As noted earlier, a by-product of training the proposed system with NCPs is the ability to perform source enhancement. This can be done by aggregating the NMF component proposal relevance scores as follows:
Denoting by the score for component’s temporal segment, we compute a global score for each component as
We apply min-max scaling between [0,1]:
This is followed by soft mask based source and noise spectrogram reconstruction using complex-valued mixture spectrogram . Note that we can optionally apply a hard threshold on to choose the top ranked components for the source. This amounts to replacing by the indicator function in the following reconstruction equations:
Here and are the estimates of source of interest and of background noise, respectively.
Dataset. We use Kinetics-Instruments (KI), a subset of the Kinetics dataset  that contains 10s Youtube videos from 15 music instrument classes. From a total of 10,267 videos, we create training and testing sets that contain 9199 and 1023 videos, respectively. For source enhancement evaluation, we handpicked 45 “clean” instrument recordings, 3 per class. Due to their unconstrained nature, the audio recordings are mostly noisy, i.e. videos are either shot with accompanying music/instruments or in acoustic environments containing other background events. In that context, “clean” refers to solo instrument samples with minimal amount of such noise.
Systems. Based on the configuration depicted in Fig. 1, we propose to evaluate audio-only, A, and audio-visual (multimodal), V + A, systems with different audio proposal types, namely:
A (TSP): temporal segment proposals,
A (NCP): NMF component proposals,
A (TSP, NCP): all TSPs and NCPs are put together into the same bag and fed to the audio network.
While systems using only TSP give state-of-the-art results , they serve as a strong baseline for establishing the usefulness of NCPs in classification. For source enhancement we compare with the following NMF related methods:
Supervised NMF : We use the class labels to train separate dictionaries of size 100 for each music instrument with stochastic mini-batch updates. At test time, depending on the label, the mixture is projected onto the appropriate dictionary for source reconstruction.
Implementation Details.and a batch size of 1. We use the matlab implementation of EdgeBoxes for generating image region proposals, obtaining approximately 100 regions per video with . Base visual features are extracted using caffenet with pre-trained ImageNet weights and RoI pooling layer modification . The fully connected layers, namely and , are fine-tuned with 50% dropout.
For audio, each recording is resampled to 16 kHz before processing. We use the official Tensorflow implementation of vggish 
. The whole audio network is fine-tuned during training. For TSP generation we first compute log-Mel spectrum over the whole file with a window size of 25ms and 10ms hop length. The resulting spectrum is chunked into segment proposals using a 960ms window with a 480ms stride. For log-Mel spectrum computation we use the accompanyingvggish code implementation. For a 10 second recording, this yields 20 segments of size . For NCP, we consider components with KL divergence and multiplicative updates. As stated in Sec. 2.1, each NMF component is passed through the TSP computation pipeline with a non-overlapping window, giving a total of 200 () NCPs for a 10s audio recording.
Classification: Kinetics-Instruments is a multi-class dataset. Hence, we consider of the score vector to be the predicted class and report the overall accuracy
Source enhancement: We corrupt the original audio with background noise corresponding to recordings of environments such as bus, busy street, park, etc. using one audio file per scene from the DCASE 2013 scene classification dataset . The system can be utilized in two modes: label known and label unknown. For the former, where the source of interest is known, we simply use the proposal ranking given by the corresponding classifier for reconstruction. For the latter, the system’s classification output is used to infer the source.
3.2 Classification Results
In Table 1 we show classification results on KI for all systems explained previously. For methods using NMF decomposition, the accuracy is averaged over 5 runs to account for changes due to random initialization. We observe that the accuracies are consistent across runs i.e.
the standard deviation does not exceed 0.5 for any of the proposed systems.
|(c)||A (NCP, TSP)||76.7|
|(d)||(a) + (b)||77.3|
|(e)||V + A (TSP)||84.5|
|(f)||V + A (NCP)||80.9|
|(g)||V + A (NCP, TSP)||84.6|
|(h)||(e) + (f)||84.6|
First, we note an evident increase in performance for all the AV systems when contrasted with audio-only methods. Indeed, the image sequence provides strong complementary information about an instrument’s presence when audio is noisy. We also see that using NCP in conjunction with TSP results in a noticable improvement for the audio-only systems. In comparison, this relative difference is negligible for multimodal methods. A possible explanation is that NCPs are expected to provide complementary information in noisy acoustic conditions. Thus, their contribution in assisting TSP is visible for audio-only classification. On the other hand, vision itself serves as a strong supporting cue for classification, unaffected by noise in audio and its presence limits the reliance on NCP. The accuracy drop when using NCP alone is expected as whole audio segments could often be easier to classify than individual components.
|SNR (dB)||V + A (TSP)||V + A (NCP, TSP)|
To further test the usefulness of NCP, we corrupt the test set audio with additional noise at different SNRs using samples from DCASE 2013 scene classification data. Average classification scores over this noisy test set are reported in Table 2. We observe a clear improvement even for the multimodal system when used with NCPs.
3.3 Source Enhancement Results and Visual Localization Examples
Following the testing protocol stated in Sec. 3.1, we report, in Table 3, average Source to Distortion Ratio (SDR)  over 450 audio mixtures created by mixing each of the 45 clean samples from the dataset with 10 noisy audio scenes. The results look promising but not state-of-the-art. This performance gap can be explained by noting that the audio network is trained for the task of audio event detection and thus does not yield optimal performance for source enhancement. The network focuses on discriminative components, failing to separate some source components from the noise by a larger margin, possibly requiring manual thresholding for best results. Also, performance for the proposed systems does not degrade when used in “Label Unknown” mode, indicating that despite incorrect classification the system is able to cluster acoustically similar sounds. Performance of supervised NMF seems to suffer due to training on a noisy dataset. We present some visual localization examples in Fig. 2. Examples and supplementary material are available on our companion website.111https://perso.telecom-paristech.fr/sparekh/icassp2019.html
|System||Label Known||Label Unknown|
|V + A (NCP), soft||3.29||3.29|
|V + A (NCP),||3.77||3.85|
|V + A (NCP),||3.56||3.56|
|V + A (NCP, TSP), soft||2.11||2.15|
We have presented a novel system for robust AV object extraction under weak supervision. Unlike previous multimodal studies, we only use weak labels for training. The central idea is to perform MIL over a set of audio and visual proposals. In particular, we propose the use of NMF for generating audio proposals. Its advantage for robust AV object classification in noisy acoustic conditions and source enhancement capability is demonstrated over a large dataset of musical instrument videos.
-  Sanjeel Parekh, Slim Essid, Alexey Ozerov, Ngoc Q. K. Duong, Patrick Pérez, and Gaël Richard, “Weakly supervised representation learning for unsynchronized audio-visual events,” CoRR, vol. abs/1804.07345, 2018.
-  Sanjeel Parekh, Slim Essid, Alexey Ozerov, Ngoc Q. K. Duong, Patrick Pérez, and Gaël Richard, “Weakly supervised representation learning for unsynchronized audio-visual events,” Workshop on Sight and Sound, CVPR, 2018.
-  Annamaria Mesaros, Toni Heittola, Onur Dikmen, and Tuomas Virtanen, “Sound event detection in real life recordings using coupled matrix factorization of spectral representations and class activity annotations,” in ICASSP. IEEE, 2015, pp. 151–155.
-  Xiaodan Zhuang, Xi Zhou, Mark A Hasegawa-Johnson, and Thomas S Huang, “Real-world acoustic event detection,” Pattern Recognition Letters, vol. 31, no. 12, pp. 1543–1551, 2010.
Sharath Adavanne, Pasi Pertilä, and Tuomas Virtanen,
“Sound event detection using spatial features and convolutional recurrent neural network,”in ICASSP. IEEE, 2017, pp. 771–775.
-  Victor Bisot, Romain Serizel, Slim Essid, and Gaël Richard, “Feature learning with matrix factorization applied to acoustic scene classification,” IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 25, no. 6, pp. 1216–1229, 2017.
-  Jia Xu, Alexander G Schwing, and Raquel Urtasun, “Learning to segment under various forms of weak supervision,” in CVPR, 2015, pp. 3781–3790.
-  Mingchen Gao, Ziyue Xu, Le Lu, Aaron Wu, Isabella Nogues, Ronald M Summers, and Daniel J Mollura, “Segmentation label propagation using deep convolutional neural networks and dense conditional random field,” in ISBI. IEEE, 2016, pp. 1265–1268.
-  Thomas G Dietterich, Richard H Lathrop, and Tomás Lozano-Pérez, “Solving the multiple instance problem with axis-parallel rectangles,” Artificial intelligence, vol. 89, no. 1-2, pp. 31–71, 1997.
-  Yapeng Tian, Jing Shi, Bochen Li, Zhiyao Duan, and Chenliang Xu, “Audio-visual event localization in unconstrained videos,” in ECCV, September 2018.
-  Daniel D Lee and H Sebastian Seung, “Algorithms for non-negative matrix factorization,” in Advances in neural information processing systems, 2001, pp. 556–562.
-  Alexey Ozerov and Cédric Févotte, “Multichannel nonnegative matrix factorization in convolutive mixtures for audio source separation,” IEEE Transactions on Audio, Speech, and Language Processing, vol. 18, no. 3, pp. 550–563, 2010.
-  Victor Bisot, Slim Essid, and Gaël Richard, “Overlapping sound event detection with supervised nonnegative matrix factorization,” in ICASSP. IEEE, 2017, pp. 31–35.
-  Toni Heittola, Annamaria Mesaros, Tuomas Virtanen, and Antti Eronen, “Sound event detection in multisource environments using source separation,” in Machine Listening in Multisource Environments, 2011.
-  Ruohan Gao, Rogerio Feris, and Kristen Grauman, “Learning to separate object sounds by watching unlabeled video,” in ECCV, September 2018.
-  Hang Zhao, Chuang Gan, Andrew Rouditchenko, Carl Vondrick, Josh McDermott, and Antonio Torralba, “The sound of pixels,” in ECCV, September 2018.
-  Ariel Ephrat, Inbar Mosseri, Oran Lang, Tali Dekel, Kevin Wilson, Avinatan Hassidim, William T. Freeman, and Michael Rubinstein, “Looking to listen at the cocktail party: A speaker-independent audio-visual model for speech separation,” ACM Trans. Graph., vol. 37, no. 4, pp. 112:1–112:11, July 2018.
-  C Lawrence Zitnick and Piotr Dollár, “Edge boxes: Locating object proposals from edges,” in ECCV. Springer, 2014, pp. 391–405.
-  Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei, “Imagenet: A large-scale hierarchical image database,” in Computer Vision and Pattern Recognition, 2009. CVPR 2009. IEEE Conference on. IEEE, 2009, pp. 248–255.
-  Cédric Févotte, Nancy Bertin, and Jean-Louis Durrieu, “Nonnegative matrix factorization with the itakura-saito divergence: With application to music analysis,” Neural computation, vol. 21, no. 3, pp. 793–830, 2009.
-  Shawn Hershey, Sourish Chaudhuri, Daniel PW Ellis, Jort F Gemmeke, Aren Jansen, R Channing Moore, Manoj Plakal, Devin Platt, Rif A Saurous, Bryan Seybold, et al., “CNN architectures for large-scale audio classification,” in ICASSP. IEEE, 2017, pp. 131–135.
-  Sami Abu-El-Haija, Nisarg Kothari, Joonseok Lee, Apostol (Paul) Natsev, George Toderici, Balakrishnan Varadarajan, and Sudheendra Vijayanarasimhan, “Youtube-8M: A large-scale video classification benchmark,” in arXiv:1609.08675, 2016.
-  Hakan Bilen and Andrea Vedaldi, “Weakly supervised deep detection networks,” in CVPR, 2016, pp. 2846–2854.
-  Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, et al., “The kinetics human action video dataset,” arXiv preprint arXiv:1705.06950, 2017.
-  Cédric Févotte, Emmanuel Vincent, and Alexey Ozerov, “Single-channel audio source separation with NMF: divergences, constraints and algorithms,” in Audio Source Separation, pp. 1–24. Springer, 2018.
-  Martin Spiertz and Volker Gnann, “Source-filter based clustering for monaural blind source separation,” in in Proceedings of International Conference on Digital Audio Effects DAFx’09, 2009.
-  NMF Mel Clustering Code, http://www.ient.rwth-aachen.de/cms/dafx09/.
-  Ross Girshick, “Fast R-CNN,” in ICCV. IEEE, 2015, pp. 1440–1448.
-  Vggish Code, https://github.com/tensorflow/models/tree/master/research/audioset.
-  Dan Stowell, Dimitrios Giannoulis, Emmanouil Benetos, Mathieu Lagrange, and Mark D Plumbley, “Detection and classification of acoustic scenes and events,” IEEE Transactions on Multimedia, vol. 17, no. 10, pp. 1733–1746, 2015.
-  Emmanuel Vincent, Rémi Gribonval, and Cédric Févotte, “Performance measurement in blind audio source separation,” IEEE transactions on audio, speech, and language processing, vol. 14, no. 4, pp. 1462–1469, 2006.