There is a growing belief that deep neural networks with a shape bias may exhibit better generalization capabilities than models with a texture bias, because shape is a more reliable indicator of the object category. However, we show experimentally that existing measures of shape bias are not stable predictors of generalization and argue that shape discrimination should not come at the expense of texture discrimination. Thus, we propose a pseudo-task to explicitly boost both shape and texture discriminability in models trained via self-supervised learning. For this purpose, we train a ViT to detect which input token has been combined with an incorrect positional embedding. To retain texture discrimination, the ViT is also trained as in MoCo with a student-teacher architecture and a contrastive loss over an extra learnable class token. We call our method DILEMMA, which stands for Detection of Incorrect Location EMbeddings with MAsked inputs. We evaluate our method through fine-tuning on several datasets and show that it outperforms MoCoV3 and DINO. Moreover, we show that when downstream tasks are strongly reliant on shape (such as in the YOGA-82 pose dataset), our pre-trained features yield a significant gain over prior work. Code will be released upon publication.
We propose a solution to the novel task of rendering sharp videos from new viewpoints from a single motion-blurred image of a face. Our method handles the complexity of face blur by implicitly learning the geometry and motion of faces through the joint training on three large datasets: FFHQ and 300VW, which are publicly available, and a new Bern Multi-View Face Dataset (BMFD) that we built. The first two datasets provide a large variety of faces and allow our model to generalize better. BMFD instead allows us to introduce multi-view constraints, which are crucial to synthesizing sharp videos from a new camera view. It consists of high frame rate synchronized videos from multiple views of several subjects displaying a wide range of facial expressions. We use the high frame rate videos to simulate realistic motion blur through averaging. Thanks to this dataset, we train a neural network to reconstruct a 3D video representation from a single image and the corresponding face gaze. We then provide a camera viewpoint relative to the estimated gaze and the blurry image as input to an encoder-decoder network to generate a video of sharp frames with a novel camera viewpoint. We demonstrate our approach on test subjects of our multi-view dataset and VIDTIMIT.
We introduce a novel self-supervised contrastive learning method to learn representations from unlabelled videos. Existing approaches ignore the specifics of input distortions, e.g., by learning invariance to temporal transformations. Instead, we argue that video representation should preserve video dynamics and reflect temporal manipulations of the input. Therefore, we exploit novel constraints to build representations that are equivariant to temporal transformations and better capture video dynamics. In our method, relative temporal transformations between augmented clips of a video are encoded in a vector and contrasted with other transformation vectors. To support temporal equivariance learning, we additionally propose the self-supervised classification of two clips of a video into 1. overlapping 2. ordered, or 3. unordered. Our experiments show that time-equivariant representations achieve state-of-the-art results in video retrieval and action recognition benchmarks on UCF101, HMDB51, and Diving48.
We present VPN - a content attribution method for recovering provenance information from videos shared online. Platforms, and users, often transform video into different quality, codecs, sizes, shapes, etc. or slightly edit its content such as adding text or emoji, as they are redistributed online. We learn a robust search embedding for matching such video, invariant to these transformations, using full-length or truncated video queries. Once matched against a trusted database of video clips, associated information on the provenance of the clip is presented to the user. We use an inverted index to match temporal chunks of video using late-fusion to combine both visual and audio features. In both cases, features are extracted via a deep neural network trained using contrastive learning on a dataset of original and augmented video clips. We demonstrate high accuracy recall over a corpus of 100,000 videos.
Current state-of-the-art methods cast monocular 3D human pose estimation as a learning problem by training neural networks on large data sets of images and corresponding skeleton poses. In contrast, we propose an approach that can exploit small annotated data sets by fine-tuning networks pre-trained via self-supervised learning on (large) unlabeled data sets. To drive such networks towards supporting 3D pose estimation during the pre-training step, we introduce a novel self-supervised feature learning task designed to focus on the 3D structure in an image. We exploit images extracted from videos captured with a multi-view camera system. The task is to classify whether two images depict two views of the same scene up to a rigid transformation. In a multi-view data set, where objects deform in a non-rigid manner, a rigid transformation occurs only between two views taken at the exact same time, i.e., when they are synchronized. We demonstrate the effectiveness of the synchronization task on the Human3.6M data set and achieve state-of-the-art results in 3D human pose estimation.
We introduce a novel self-supervised learning approach to learn representations of videos that are responsive to changes in the motion dynamics. Our representations can be learned from data without human annotation and provide a substantial boost to the training of neural networks on small labeled data sets for tasks such as action recognition, which require to accurately distinguish the motion of objects. We promote an accurate learning of motion without human annotation by training a neural network to discriminate a video sequence from its temporally transformed versions. To learn to distinguish non-trivial motions, the design of the transformations is based on two principles: 1) To define clusters of motions based on time warps of different magnitude; 2) To ensure that the discrimination is feasible only by observing and analyzing as many image frames as possible. Thus, we introduce the following transformations: forward-backward playback, random frame skipping, and uniform frame skipping. Our experiments show that networks trained with the proposed method yield representations with improved transfer performance for action recognition on UCF101 and HMDB51.
We introduce a novel principle for self-supervised feature learning based on the discrimination of specific transformations of an image. We argue that the generalization capability of learned features depends on what image neighborhood size is sufficient to discriminate different image transformations: The larger the required neighborhood size and the more global the image statistics that the feature can describe. An accurate description of global image statistics allows to better represent the shape and configuration of objects and their context, which ultimately generalizes better to new tasks such as object classification and detection. This suggests a criterion to choose and design image transformations. Based on this criterion, we introduce a novel image transformation that we call limited context inpainting (LCI). This transformation inpaints an image patch conditioned only on a small rectangular pixel boundary (the limited context). Because of the limited boundary information, the inpainter can learn to match local pixel statistics, but is unlikely to match the global statistics of the image. We claim that the same principle can be used to justify the performance of transformations such as image rotations and warping. Indeed, we demonstrate experimentally that learning to discriminate transformations such as LCI, image warping and rotations, yields features with state of the art generalization capabilities on several datasets such as Pascal VOC, STL-10, CelebA, and ImageNet. Remarkably, our trained features achieve a performance on Places on par with features trained through supervised learning with ImageNet labels.
We propose a novel method to perform extreme (16x) face super-resolution by exploiting audio. Super-resolution is the task of recovering a high-resolution image from a low-resolution one. When the resolution of the input image is too low (e.g., 8x8 pixels), the loss of information is so dire that the details of the original identity have been lost. However, when the low-resolution image is extracted from a video, the audio track is also available. Because the audio carries information about the face identity, we propose to exploit it in the face reconstruction process. Towards this goal, we propose a model and a training procedure to extract information about the identity of a person from her audio track and to combine it with the information extracted from the low-resolution input image, which relates more to pose and colors of the face. We demonstrate that the combination of these two inputs yields high-resolution images that better capture the correct identity of the face. In particular, we show that audio can assist in recovering attributes such as the gender and the identity, and thus improve the correctness of the image reconstruction process. Our procedure does not make use of human annotation and thus can be easily trained with existing video datasets. Moreover, we show that our model allows one to mix low-resolution images and audio from different videos and to generate realistic faces with semantically meaningful combinations.
We present a novel method and analysis to train generative adversarial networks (GAN) in a stable manner. As shown in recent analysis, training is often undermined by the probability distribution of the data being zero on neighborhoods of the data space. We notice that the distributions of real and generated data should match even when they undergo the same filtering. Therefore, to address the limited support problem we propose to train GANs by using different filtered versions of the real and generated data distributions. In this way, filtering does not prevent the exact matching of the data distribution, while helping training by extending the support of both distributions. As filtering we consider adding samples from an arbitrary distribution to the data, which corresponds to a convolution of the data distribution with the arbitrary one. We also propose to learn the generation of these samples so as to challenge the discriminator in the adversarial training. We show that our approach results in a stable and well-behaved training of even the original minimax GAN formulation. Moreover, our technique can be incorporated in most modern GAN formulations and leads to a consistent improvement on several common datasets.
We present a novel regularization approach to train neural networks that enjoys better generalization and test error than standard stochastic gradient descent. Our approach is based on the principles of cross-validation, where a validation set is used to limit the model overfitting. We formulate such principles as a bilevel optimization problem. This formulation allows us to define the optimization of a cost on the validation set subject to another optimization on the training set. The overfitting is controlled by introducing weights on each mini-batch in the training set and by choosing their values so that they minimize the error on the validation set. In practice, these weights define mini-batch learning rates in a gradient descent update equation that favor gradients with better generalization capabilities. Because of its simplicity, this approach can be integrated with other regularization methods and training schemes. We evaluate extensively our proposed algorithm on several neural network architectures and datasets, and find that it consistently improves the generalization of the model, especially when labels are noisy.