In many fields, self-supervised learning solutions are rapidly evolving and filling the gap with supervised approaches. This fact occurs for depth estimation based on either monocular or stereo, with the latter often providing a valid source of self-supervision for the former. In contrast, to soften typical stereo artefacts, we propose a novel self-supervised paradigm reversing the link between the two. Purposely, in order to train deep stereo networks, we distill knowledge through a monocular completion network. This architecture exploits single-image clues and few sparse points, sourced by traditional stereo algorithms, to estimate dense yet accurate disparity maps by means of a consensus mechanism over multiple estimations. We thoroughly evaluate with popular stereo datasets the impact of different supervisory signals showing how stereo networks trained with our paradigm outperform existing self-supervised frameworks. Finally, our proposal achieves notable generalization capabilities dealing with domain shift issues. Code available at https://github.com/FilippoAleotti/Reversing
Supervised learning based methods for monocular depth estimation usually require large amounts of extensively annotated training data. In the case of aerial imagery, this ground truth is particularly difficult to acquire. Therefore, in this paper, we present a method for self-supervised learning for monocular depth estimation from aerial imagery that does not require annotated training data. For this, we only use an image sequence from a single moving camera and learn to simultaneously estimate depth and pose information. By sharing the weights between pose and depth estimation, we achieve a relatively small model, which favors real-time application. We evaluate our approach on three diverse datasets and compare the results to conventional methods that estimate depth maps based on multi-view geometry. We achieve an accuracy {\delta}1.25 of up to 93.5 %. In addition, we have paid particular attention to the generalization of a trained model to unknown data and the self-improving capabilities of our approach. We conclude that, even though the results of monocular depth estimation are inferior to those achieved by conventional methods, they are well suited to provide a good initialization for methods that rely on image matching or to provide estimates in regions where image matching fails, e.g. occluded or texture-less regions.
Game character customization is one of the core features of many recent Role-Playing Games (RPGs), where players can edit the appearance of their in-game characters with their preferences. This paper studies the problem of automatically creating in-game characters with a single photo. In recent literature on this topic, neural networks are introduced to make game engine differentiable and the self-supervised learning is used to predict facial customization parameters. However, in previous methods, the expression parameters and facial identity parameters are highly coupled with each other, making it difficult to model the intrinsic facial features of the character. Besides, the neural network based renderer used in previous methods is also difficult to be extended to multi-view rendering cases. In this paper, considering the above problems, we propose a novel method named "PokerFace-GAN" for neutral face game character auto-creation. We first build a differentiable character renderer which is more flexible than the previous methods in multi-view rendering cases. We then take advantage of the adversarial training to effectively disentangle the expression parameters from the identity parameters and thus generate player-preferred neutral face (expression-less) characters. Since all components of our method are differentiable, our method can be easily trained under a multi-task self-supervised learning paradigm. Experiment results show that our method can generate vivid neutral face game characters that are highly similar to the input photos. The effectiveness of our method is verified by comparison results and ablation studies.
We present a self-supervised Contrastive Video Representation Learning (CVRL) method to learn spatiotemporal visual representations from unlabeled videos. Our representations are learned using a contrastive loss, where two augmented clips from the same short video are pulled together in the embedding space, while clips from different videos are pushed away. We study what makes for good data augmentations for video self-supervised learning and find that both spatial and temporal information are crucial. We carefully design data augmentations involving spatial and temporal cues. Concretely, we propose a temporally consistent spatial augmentation method to impose strong spatial augmentations on each frame of the video while maintaining the temporal consistency across frames. We also propose a sampling-based temporal augmentation method to avoid overly enforcing invariance on clips that are distant in time. On Kinetics-600, a linear classifier trained on the representations learned by CVRL achieves 70.4% top-1 accuracy with a 3D-ResNet-50 (R3D-50) backbone, outperforming ImageNet supervised pre-training by 15.7% and SimCLR unsupervised pre-training by 18.8% using the same inflated R3D-50. The performance of CVRL can be further improved to 72.9% with a larger R3D-152 (2x filters) backbone, significantly closing the gap between unsupervised and supervised video representation learning. Our code and models will be available at https://github.com/tensorflow/models/tree/master/official/.
Unsupervised and self-supervised learning approaches have become a crucial tool to learn representations for downstream prediction tasks. While these approaches are widely used in practice and achieve impressive empirical gains, their theoretical understanding largely lags behind. Towards bridging this gap, we present a unifying perspective where several such approaches can be viewed as imposing a regularization on the representation via a learnable function using unlabeled data. We propose a discriminative theoretical framework for analyzing the sample complexity of these approaches, which generalizes the framework of (Balcan and Blum, 2010) to allow learnable regularization functions. Our sample complexity bounds show that, with carefully chosen hypothesis classes to exploit the structure in the data, these learnable regularization functions can prune the hypothesis space, and help reduce the amount of labeled data needed. We then provide two concrete examples of functional regularization, one using auto-encoders and the other using masked self-supervision, and apply our framework to quantify the reduction in the sample complexity bound of labeled data. We also provide complementary empirical results to support our analysis.
Point clouds provide a compact and efficient representation of 3D shapes. While deep neural networks have achieved impressive results on point cloud learning tasks, they require massive amounts of manually labeled data, which can be costly and time-consuming to collect. In this paper, we leverage 3D self-supervision for learning downstream tasks on point clouds with fewer labels. A point cloud can be rotated in infinitely many ways, which provides a rich label-free source for self-supervision. We consider the auxiliary task of predicting rotations that in turn leads to useful features for other tasks such as shape classification and 3D keypoint prediction. Using experiments on ShapeNet and ModelNet, we demonstrate that our approach outperforms the state-of-the-art. Moreover, features learned by our model are complementary to other self-supervised methods and combining them leads to further performance improvement.
Convolutional Neural Networks (CNNs) are prone to overfit small training datasets. We present a novel two-phase pipeline that leverages self-supervised learning and knowledge distillation to improve the generalization ability of CNN models for image classification under the data-deficient setting. The first phase is to learn a teacher model which possesses rich and generalizable visual representations via self-supervised learning, and the second phase is to distill the representations into a student model in a self-distillation manner, and meanwhile fine-tune the student model for the image classification task. We also propose a novel margin loss for the self-supervised contrastive learning proxy task to better learn the representation under the data-deficient scenario. Together with other tricks, we achieve competitive performance in the VIPriors image classification challenge.
Self-supervised learning for visual object tracking possesses valuable advantages compared to supervised learning, such as the non-necessity of laborious human annotations and online training. In this work, we exploit an end-to-end Siamese network in a cycle-consistent self-supervised framework for object tracking. Self-supervision can be performed by taking advantage of the cycle consistency in the forward and backward tracking. To better leverage the end-to-end learning of deep networks, we propose to integrate a Siamese region proposal and mask regression network in our tracking framework so that a fast and more accurate tracker can be learned without the annotation of each frame. The experiments on the VOT dataset for visual object tracking and on the DAVIS dataset for video object segmentation propagation show that our method outperforms prior approaches on both tasks.
The objective of this paper is self-supervised learning from video, in particular for representations for action recognition. We make the following contributions: (i) We propose a new architecture and learning framework Memory-augmented Dense Predictive Coding (MemDPC) for the task. It is trained with a predictive attention mechanism over the set of compressed memories, such that any future states can always be constructed by a convex combination of the condense representations, allowing to make multiple hypotheses efficiently. (ii) We investigate visual-only self-supervised video representation learning from RGB frames, or from unsupervised optical flow, or both. (iii) We thoroughly evaluate the quality of learnt representation on four different downstream tasks: action recognition, video retrieval, learning with scarce annotations, and unintentional action classification. In all cases, we demonstrate state-of-the-art or comparable performance over other approaches with orders of magnitude fewer training data.
Within months of birth, children develop meaningful expectations about the world around them. How much of this early knowledge can be explained through generic learning mechanisms applied to sensory data, and how much of it requires more substantive innate inductive biases? Addressing this fundamental question in its full generality is currently infeasible, but we can hope to make real progress in more narrowly defined domains, such as the development of high-level visual categories, thanks to improvements in data collecting technology and recent progress in deep learning. In this paper, our goal is precisely to achieve such progress by utilizing modern self-supervised deep learning methods and a recent longitudinal, egocentric video dataset recorded from the perspective of three young children (Sullivan et al., 2020). Our results demonstrate the emergence of powerful, high-level visual representations from developmentally realistic natural videos using generic self-supervised learning objectives.
We present KOVIS, a novel learning-based, calibration-free visual servoing method for fine robotic manipulation tasks with eye-in-hand stereo camera system. We train the deep neural network only in the simulated environment; and the trained model could be directly used for real-world visual servoing tasks. KOVIS consists of two networks. The first keypoint network learns the keypoint representation from the image using with an autoencoder. Then the visual servoing network learns the motion based on keypoints extracted from the camera image. The two networks are trained end-to-end in the simulated environment by self-supervised learning without manual data labeling. After training with data augmentation, domain randomization, and adversarial examples, we are able to achieve zero-shot sim-to-real transfer to real-world robotic manipulation tasks. We demonstrate the effectiveness of the proposed method in both simulated environment and real-world experiment with different robotic manipulation tasks, including grasping, peg-in-hole insertion with 4mm clearance, and M13 screw insertion. The demo video is available at http://youtu.be/gfBJBR2tDzA
The ability to efficiently utilize crowdsourced visual data carries immense potential for the domains of large scale dynamic mapping and autonomous driving. However, state-of-the-art methods for crowdsourced 3D mapping assume prior knowledge of camera intrinsics. In this work, we propose a framework that estimates the 3D positions of semantically meaningful landmarks such as traffic signs without assuming known camera intrinsics, using only monocular color camera and GPS. We utilize multi-view geometry as well as deep learning based self-calibration, depth, and ego-motion estimation for traffic sign positioning, and show that combining their strengths is important for increasing the map coverage. To facilitate research on this task, we construct and make available a KITTI based 3D traffic sign ground truth positioning dataset. Using our proposed framework, we achieve an average single-journey relative and absolute positioning accuracy of 39cm and 1.26m respectively, on this dataset.
3D human shape and pose estimation from monocular images has been an active area of research in computer vision, having a substantial impact on the development of new applications, from activity recognition to creating virtual avatars. Existing deep learning methods for 3D human shape and pose estimation rely on relatively high-resolution input images; however, high-resolution visual content is not always available in several practical scenarios such as video surveillance and sports broadcasting. Low-resolution images in real scenarios can vary in a wide range of sizes, and a model trained in one resolution does not typically degrade gracefully across resolutions. Two common approaches to solve the problem of low-resolution input are applying super-resolution techniques to the input images which may result in visual artifacts, or simply training one model for each resolution, which is impractical in many realistic applications. To address the above issues, this paper proposes a novel algorithm called RSC-Net, which consists of a Resolution-aware network, a Self-supervision loss, and a Contrastive learning scheme. The proposed network is able to learn the 3D body shape and pose across different resolutions with a single model. The self-supervision loss encourages scale-consistency of the output, and the contrastive learning scheme enforces scale-consistency of the deep features. We show that both these new training losses provide robustness when learning 3D shape and pose in a weakly-supervised manner. Extensive experiments demonstrate that the RSC-Net can achieve consistently better results than the state-of-the-art methods for challenging low-resolution images.
Recent learning-based approaches, in which models are trained by single-view images have shown promising results for monocular 3D face reconstruction, but they suffer from the ill-posed face pose and depth ambiguity issue. In contrast to previous works that only enforce 2D feature constraints, we propose a self-supervised training architecture by leveraging the multi-view geometry consistency, which provides reliable constraints on face pose and depth estimation. We first propose an occlusion-aware view synthesis method to apply multi-view geometry consistency to self-supervised learning. Then we design three novel loss functions for multi-view consistency, including the pixel consistency loss, the depth consistency loss, and the facial landmark-based epipolar loss. Our method is accurate and robust, especially under large variations of expressions, poses, and illumination conditions. Comprehensive experiments on the face alignment and 3D face reconstruction benchmarks have demonstrated superiority over state-of-the-art methods. Our code and model are released in https://github.com/jiaxiangshang/MGCNet.
Reducing the amount of labels required to train convolutional neural networks without performance degradation is key to effectively reduce human annotation efforts. We propose Reliable Label Bootstrapping (ReLaB), an unsupervised preprossessing algorithm which improves the performance of semi-supervised algorithms in extremely low supervision settings. Given a dataset with few labeled samples, we first learn meaningful self-supervised, latent features for the data. Second, a label propagation algorithm propagates the known labels on the unsupervised features, effectively labeling the full dataset in an automatic fashion. Third, we select a subset of correctly labeled (reliable) samples using a label noise detection algorithm. Finally, we train a semi-supervised algorithm on the extended subset. We show that the selection of the network architecture and the self-supervised algorithm are important factors to achieve successful label propagation and demonstrate that ReLaB substantially improves semi-supervised learning in scenarios of very limited supervision on CIFAR-10, CIFAR-100 and mini-ImageNet. We reach average error rates of $\boldsymbol{22.34}$ with 1 random labeled sample per class on CIFAR-10 and lower this error to $\boldsymbol{8.46}$ when the labeled sample in each class is highly representative. Our work is fully reproducible: https://github.com/PaulAlbert31/ReLaB.
Given new tasks with very little data$-$such as new classes in a classification problem or a domain shift in the input$-$performance of modern vision systems degrades remarkably quickly. In this work, we illustrate how the neural network representations which underpin modern vision systems are subject to supervision collapse, whereby they lose any information that is not necessary for performing the training task, including information that may be necessary for transfer to new tasks or domains. We then propose two methods to mitigate this problem. First, we employ self-supervised learning to encourage general-purpose features that transfer better. Second, we propose a novel Transformer based neural network architecture called CrossTransformers, which can take a small number of labeled images and an unlabeled query, find coarse spatial correspondence between the query and the labeled images, and then infer class membership by computing distances between spatially-corresponding features. The result is a classifier that is more robust to task and domain shift, which we demonstrate via state-of-the-art performance on Meta-Dataset, a recent dataset for evaluating transfer from ImageNet to many other vision datasets.
Representation learning methods for heterogeneous networks produce a low-dimensional vector embedding for each node that is typically fixed for all tasks involving the node. Many of the existing methods focus on obtaining a static vector representation for a node in a way that is agnostic to the downstream application where it is being used. In practice, however, downstream tasks such as link prediction require specific contextual information that can be extracted from the subgraphs related to the nodes provided as input to the task. To tackle this challenge, we develop SLiCE, a framework bridging static representation learning methods using global information from the entire graph with localized attention driven mechanisms to learn contextual node representations. We first pre-train our model in a self-supervised manner by introducing higher-order semantic associations and masking nodes, and then fine-tune our model for a specific link prediction task. Instead of training node representations by aggregating information from all semantic neighbors connected via metapaths, we automatically learn the composition of different metapaths that characterize the context for a specific task without the need for any pre-defined metapaths. SLiCE significantly outperforms both static and contextual embedding learning methods on several publicly available benchmark network datasets. We also interpret the semantic association matrix and provide its utility and relevance in making successful link predictions between heterogeneous nodes in the network.
Previous audio-visual speech separation methods use the synchronization of the speaker's facial movement and speech in the video to supervise the speech separation in a self-supervised way. In this paper, we propose a model to solve the speech separation problem assisted by both face and sign language, which we call the extended speech separation problem. We design a general deep learning network for learning the combination of three modalities, audio, face, and sign language information, for better solving the speech separation problem. To train the model, we introduce a large-scale dataset named the Chinese Sign Language News Speech (CSLNSpeech) dataset, in which three modalities of audio, face, and sign language coexist. Experiment results show that the proposed model has better performance and robustness than the usual audio-visual system. Besides, sign language modality can also be used alone to supervise speech separation tasks, and the introduction of sign language is helpful for hearing-impaired people to learn and communicate. Last, our model is a general speech separation framework and can achieve very competitive separation performance on two open-source audio-visual datasets. The code is available at https://github.com/iveveive/SLNSpeech
Photometric loss is widely used for self-supervised depth and egomotion estimation. However, the loss landscapes induced by photometric differences are often problematic for optimization, caused by plateau landscapes for pixels in textureless regions or multiple local minima for less discriminative pixels. In this work, feature-metric loss is proposed and defined on feature representation, where the feature representation is also learned in a self-supervised manner and regularized by both first-order and second-order derivatives to constrain the loss landscapes to form proper convergence basins. Comprehensive experiments and detailed analysis via visualization demonstrate the effectiveness of the proposed feature-metric loss. In particular, our method improves state-of-the-art methods on KITTI from 0.885 to 0.925 measured by $\delta_1$ for depth estimation, and significantly outperforms previous method for visual odometry.
The combination of policy search and deep neural networks holds the promise of automating a variety of decision-making tasks. Model Predictive Control (MPC) provides robust solutions to robot control tasks by making use of a dynamical model of the system and solving an optimization problem online over a short planning horizon. In this work, we leverage probabilistic decision-making approaches and the generalization capability of artificial neural networks to the powerful online optimization by learning a deep high-level policy for the MPC (High-MPC). Conditioning on robot's local observations, the trained neural network policy is capable of adaptively selecting high-level decision variables for the low-level MPC controller, which then generates optimal control commands for the robot. First, we formulate the search of high-level decision variables for MPC as a policy search problem, specifically, a probabilistic inference problem. The problem can be solved in a closed-form solution. Second, we propose a self-supervised learning algorithm for learning a neural network high-level policy, which is useful for online hyperparameter adaptations in highly dynamic environments. We demonstrate the importance of incorporating the online adaption into autonomous robots by using the proposed method to solve a challenging control problem, where the task is to control a simulated quadrotor to fly through a swinging gate. We show that our approach can handle situations that are difficult for standard MPC.