Our objective is an audio-visual model for separating a single speaker from a mixture of sounds such as other speakers and background noise. Moreover, we wish to hear the speaker even when the visual cues are temporarily absent due to occlusion. To this end we introduce a deep audio-visual speech enhancement network that is able to separate a speaker's voice by conditioning on both the speaker's lip movements and/or a representation of their voice. The voice representation can be obtained by either (i) enrollment, or (ii) by self-enrollment -- learning the representation on-the-fly given sufficient unobstructed visual input. The model is trained by blending audios, and by introducing artificial occlusions around the mouth region that prevent the visual modality from dominating. The method is speaker-independent, and we demonstrate it on real examples of speakers unheard (and unseen) during training. The method also improves over previous models in particular for cases of occlusion in the visual modality.
Simulation is an anonymous, low-bias source of data where annotation can often be done automatically; however, for some tasks, current models trained on synthetic data generalize poorly to real data. The task of 3D human pose estimation is a particularly interesting example of this sim2real problem, because learning-based approaches perform reasonably well given real training data, yet labeled 3D poses are extremely difficult to obtain in the wild, limiting scalability. In this paper, we show that standard neural-network approaches, which perform poorly when trained on synthetic RGB images, can perform well when the data is pre-processed to extract cues about the person's motion, notably as optical flow and the motion of 2D keypoints. Therefore, our results suggest that motion can be a simple way to bridge a sim2real gap when video is available. We evaluate on the 3D Poses in the Wild dataset, the most challenging modern standard of 3D pose estimation, where we show full 3D mesh recovery that is on par with state-of-the-art methods trained on real 3D sequences, despite training only on synthetic humans from the SURREAL dataset.
The study of object representations in computer vision has primarily focused on developing representations that are useful for image classification, object detection, or semantic segmentation as downstream tasks. In this work we aim to learn object representations that are useful for control and reinforcement learning (RL). To this end, we introduce Transporter, a neural network architecture for discovering concise geometric object representations in terms of keypoints or image-space coordinates. Our method learns from raw video frames in a fully unsupervised manner, by transporting learnt image features between video frames using a keypoint bottleneck. The discovered keypoints track objects and object parts across long time-horizons more accurately than recent similar methods. Furthermore, consistent long-term tracking enables two notable results in control domains -- (1) using the keypoint co-ordinates and corresponding image features as inputs enables highly sample-efficient reinforcement learning; (2) learning to explore by controlling keypoint locations drastically reduces the search space, enabling deep exploration (leading to states unreachable through random action exploration) without any extrinsic rewards.
The majority of modern deep learning models are able to interpolate the data: the empirical loss can be driven near zero on all samples simultaneously. In this work, we explicitly exploit this interpolation property for the design of a new optimization algorithm for deep learning. Specifically, we use it to compute an adaptive learning-rate given a stochastic gradient direction. This results in the Adaptive Learning-rates for Interpolation with Gradients (ALI-G) algorithm. ALI-G retains the advantages of SGD, which are low computational cost and provable convergence in the convex setting. But unlike SGD, the learning-rate of ALI-G can be computed inexpensively in closed-form and does not require a manual schedule. We provide a detailed analysis of ALI-G in the stochastic convex setting with explicit convergence rates. In order to obtain good empirical performance in deep learning, we extend the algorithm to use a maximal learning-rate, which gives a single hyper-parameter to tune. We show that employing such a maximal learning-rate has an intuitive proximal interpretation and preserves all convergence guarantees. We provide experiments on a variety of architectures and tasks: (i) learning a differentiable neural computer; (ii) training a wide residual network on the SVHN data set; (iii) training a Bi-LSTM on the SNLI data set; and (iv) training wide residual networks and densely connected networks on the CIFAR data sets. We empirically show that ALI-G outperforms adaptive gradient methods such as Adam, and provides comparable performance with SGD, although SGD benefits from manual learning rate schedules. We release PyTorch and Tensorflow implementations of ALI-G as standalone optimizers that can be used as a drop-in replacement in existing code (code available at https://github.com/oval-group/ali-g ).
Capturing the `mutual gaze' of people is essential for understanding and interpreting the social interactions between them. To this end, this paper addresses the problem of detecting people Looking At Each Other (LAEO) in video sequences. For this purpose, we propose LAEO-Net, a new deep CNN for determining LAEO in videos. In contrast to previous works, LAEO-Net takes spatio-temporal tracks as input and reasons about the whole track. It consists of three branches, one for each character's tracked head and one for their relative position. Moreover, we introduce two new LAEO datasets: UCO-LAEO and AVA-LAEO. A thorough experimental evaluation demonstrates the ability of LAEONet to successfully determine if two people are LAEO and the temporal window where it happens. Our model achieves state-of-the-art results on the existing TVHID-LAEO video dataset, significantly outperforming previous approaches. Finally, we apply LAEO-Net to social network analysis, where we automatically infer the social relationship between pairs of people based on the frequency and duration that they LAEO.
In this paper, we introduce a simple and standalone manual annotation tool for images, audio and video: the VGG Image Annotator (VIA). This is a light weight, standalone and offline software package that does not require any installation or setup and runs solely in a web browser. Due to its lightness and flexibility, the VIA software has quickly become an essential and invaluable research support tool in many academic disciplines. Furthermore, it has also been very popular in several industrial sectors which have invested in adapting this open source software to their requirements. Since its public release in 2017, the VIA software has been used more than 600,000 times and has nurtured a large and thriving open source community.
Medical imaging only indirectly measures the molecular identity of the tissue within each voxel, which often produces only ambiguous image evidence for target measures of interest, like semantic segmentation. This diversity and the variations of plausible interpretations are often specific to given image regions and may thus manifest on various scales, spanning all the way from the pixel to the image level. In order to learn a flexible distribution that can account for multiple scales of variations, we propose the Hierarchical Probabilistic U-Net, a segmentation network with a conditional variational auto-encoder (cVAE) that uses a hierarchical latent space decomposition. We show that this model formulation enables sampling and reconstruction of segmenations with high fidelity, i.e. with finely resolved detail, while providing the flexibility to learn complex structured distributions across scales. We demonstrate these abilities on the task of segmenting ambiguous medical scans as well as on instance segmentation of neurobiological and natural images. Our model automatically separates independent factors across scales, an inductive bias that we deem beneficial in structured output prediction tasks beyond segmentation.
We tackle the problem of object discovery, where objects are segmented for a given input image, and the system is trained without using any direct supervision whatsoever. A novel copy-pasting GAN framework is proposed, where the generator learns to discover an object in one image by compositing it into another image such that the discriminator cannot tell that the resulting image is fake. After carefully addressing subtle issues, such as preventing the generator from `cheating', this game results in the generator learning to select objects, as copy-pasting objects is most likely to fool the discriminator. The system is shown to work well on four very different datasets, including large object appearance variations in challenging cluttered backgrounds.
While semi-supervised learning (SSL) algorithms provide an efficient way to make use of both labelled and unlabelled data, they generally struggle when the number of annotated samples is very small. In this work, we consider the problem of SSL multi-class classification with very few labelled instances. We introduce two key ideas. The first is a simple but effective one: we leverage the power of transfer learning among different tasks and self-supervision to initialize a good representation of the data without making use of any label. The second idea is a new algorithm for SSL that can exploit well such a pre-trained representation. The algorithm works by alternating two phases, one fitting the labelled points and one fitting the unlabelled ones, with carefully-controlled information flow between them. The benefits are greatly reducing overfitting of the labelled data and avoiding issue with balancing labelled and unlabelled losses during training. We show empirically that this method can successfully train competitive models with as few as 10 labelled data points per class. More in general, we show that the idea of bootstrapping features using self-supervised learning always improves SSL on standard benchmarks. We show that our algorithm works increasingly well compared to other methods when refining from other tasks or datasets.
We present a bundle-adjustment-based algorithm for recovering accurate 3D human pose and meshes from monocular videos. Unlike previous algorithms which operate on single frames, we show that reconstructing a person over an entire sequence gives extra constraints that can resolve ambiguities. This is because videos often give multiple views of a person, yet the overall body shape does not change and 3D positions vary slowly. Our method improves not only on standard mocap-based datasets like Human 3.6M -- where we show quantitative improvements -- but also on challenging in-the-wild datasets such as Kinetics. Building upon our algorithm, we present a new dataset of more than 3 million frames of YouTube videos from Kinetics with automatically generated 3D poses and meshes. We show that retraining a single-frame 3D pose estimator on this data improves accuracy on both real-world and mocap data by evaluating on the 3DPW and HumanEVA datasets.