We present a new framework for self-supervised representation learning by positing it as a ranking problem in an image retrieval context on a large number of random views from random sets of images. Our work is based on two intuitive observations: first, a good representation of images must yield a high-quality image ranking in a retrieval task; second, we would expect random views of an image to be ranked closer to a reference view of that image than random views of other images. Hence, we model representation learning as a learning-to-rank problem in an image retrieval context, and train it by maximizing average precision (AP) for ranking. Specifically, given a mini-batch of images, we generate a large number of positive/negative samples and calculate a ranking loss term by separately treating each image view as a retrieval query. The new framework, dubbed S2R2, enables computing a global objective compared to the local objective in the popular contrastive learning framework calculated on pairs of views. A global objective leads S2R2 to faster convergence in terms of the number of epochs. In principle, by using a ranking criterion, we eliminate reliance on object-centered curated datasets (e.g., ImageNet). When trained on STL10 and MS-COCO, S2R2 outperforms SimCLR and performs on par with the state-of-the-art clustering-based contrastive learning model, SwAV, while being much simpler both conceptually and implementation-wise. Furthermore, when trained on a small subset of MS-COCO with fewer similar scenes, S2R2 significantly outperforms both SwAV and SimCLR. This indicates that S2R2 is potentially more effective on diverse scenes and decreases the need for a large training dataset for self-supervised learning.
Training deep learning models on embedded devices is typically avoided since this requires more memory, computation and power over inference. In this work, we focus on lowering the amount of memory needed for storing all activations, which are required during the backward pass to compute the gradients. Instead, during the forward pass, static Synthetic Gradient Modules (SGMs) predict gradients for each layer. This allows training the model in a feed-forward manner without having to store all activations. We tested our method on a robot grasping scenario where a robot needs to learn to grasp new objects given only a single demonstration. By first training the SGMs in a meta-learning manner on a set of common objects, during fine-tuning, the SGMs provided the model with accurate gradients to successfully learn to grasp new objects. We have shown that our method has comparable results to using standard backpropagation.
The task of visual grounding requires locating the most relevant region or object in an image, given a natural language query. So far, progress on this task was mostly measured on curated datasets, which are not always representative of human spoken language. In this work, we deviate from recent, popular task settings and consider the problem under an autonomous vehicle scenario. In particular, we consider a situation where passengers can give free-form natural language commands to a vehicle which can be associated with an object in the street scene. To stimulate research on this topic, we have organized the \emph{Commands for Autonomous Vehicles} (C4AV) challenge based on the recent \emph{Talk2Car} dataset (URL: https://www.aicrowd.com/challenges/eccv-2020-commands-4-autonomous-vehicles). This paper presents the results of the challenge. First, we compare the used benchmark against existing datasets for visual grounding. Second, we identify the aspects that render top-performing models successful, and relate them to existing state-of-the-art models for visual grounding, in addition to detecting potential failure cases by evaluating on carefully selected subsets. Finally, we discuss several possibilities for future work.
Given a really low-resolution input image of a face (say 16x16 or 8x8 pixels), the goal of this paper is to reconstruct a high-resolution version thereof. This, by itself, is an ill-posed problem, as the high-frequency information is missing in the low-resolution input and needs to be hallucinated, based on prior knowledge about the image content. Rather than relying on a generic face prior, in this paper, we explore the use of a set of exemplars, i.e. other high-resolution images of the same person. These guide the neural network as we condition the output on them. Multiple exemplars work better than a single one. To combine the information from multiple exemplars effectively, we introduce a pixel-wise weight generation module. Besides standard face super-resolution, our method allows to perform subtle face editing simply by replacing the exemplars with another set with different facial features. A user study is conducted and shows the super-resolved images can hardly be distinguished from real images on the CelebA dataset. A qualitative comparison indicates our model outperforms methods proposed in the literature on the CelebA and WebFace dataset.
As learning from non-stationary streams of data has been proven a challenging endeavour, current continual learners often strongly relax the problem, assuming balanced datasets, unlimited processing of data stream subsets, and additional availability of task information, sometimes even during inference. In contrast, our continual learner processes the data streams in an online fashion, without additional task-information, and shows solid robustness to imbalanced data streams resembling a real-world setting. Defying such challenging settings is achieved by aggregating prototypes and nearest-neighbour based classification in a shared latent space, where a Continual Prototype Evolution (CoPE) enables learning and prediction at any point in time. As the embedding network continually changes, prototypes inevitably become obsolete, which we prevent by replay of exemplars from memory. We obtain state-of-the-art performance by a significant margin on five benchmarks, including two highly unbalanced data streams.
The determination of the relative 6 Degree of Freedom (DoF) pose of vehicles around the ego-vehicle from monocular cameras is an important aspect of the perception problem for Autonomous Vehicles (AVs) and Driver Assist Technology (DAT). Current deep learning techniques used for tackling this problem are data hungry, driving the need for unsupervised or self-supervised methods. In this paper, we consider the domain adaptation task of fine-tuning a vehicle orientation estimator on a new domain without labels. By leveraging the ego-motion consistencies obtained from a monocular SLAM method, we show that our self-supervised fine-tuning scheme consistently improves the accuracy of the resulting network. More specifically, when transitioning from Virtual Kitti to nuScenes, up to 70% of the performance is recovered compared to the 100% of a supervised method. Our self-supervised method hence allows us to safely transfer vehicle orientation estimators to new domains without requiring expensive new labels.
In a dynamic environment, an agent with a limited field of view/resource cannot fully observe the scene before attempting to parse it. The deployment of common semantic segmentation architectures is not feasible in such settings. In this paper we propose a method to gradually segment a scene given a sequence of partial observations. The main idea is to refine an agent's understanding of the environment by attending the areas it is most uncertain about. Our method includes a self-supervised attention mechanism and a specialized architecture to maintain and exploit spatial memory maps for filling-in the unseen areas in the environment. The agent can select and attend an area while relying on the cues coming from the visited areas to hallucinate the other parts. We reach a mean pixel-wise accuracy of 78.1%, 80.9% and 76.5% on CityScapes, CamVid, and Kitti datasets by processing only 18% of the image pixels (10 retina-like glimpses). We perform an ablation study on the number of glimpses, input image size and effectiveness of retina-like glimpses. We compare our method to several baselines and show that the optimal results are achieved by having access to a very low resolution view of the scene at the first timestep.
We present MIX'EM, a novel solution for unsupervised image classification. Our model generates representations that by themselves are sufficient to drive a general-purpose clustering method to deliver high-quality classification without supervision. MIX'EM integrates an internal mixture of embeddings module into the contrastive visual representation learning framework to disentangle the representation space at the category level. It generates a set of embeddings from a visual representation and mixes them to construct the contrastive loss input. Parallel to the contrastive loss, we introduce three techniques to train MIX'EM and avoid a degenerate solution; (i) we maximize entropy across mixture components to diversify them, and (ii) minimize component entropy conditioned on instances to enforce a clustered embedding space. Applying (i) and (ii) lead to the emergence of semantic categories through the mixture coefficients, making it possible to (iii) apply an associative embedding loss to enforce semantic separability directly. Subsequently, we run K-means on the representations to acquire semantic classification, which outperforms the state-of-the-art by a large margin. We conduct extensive experiments and analyses on STL10, CIFAR10, and CIFAR100-20 datasets, achieving 78\%, 82\%, and 44\% accuracy, respectively. Essential to robust high accuracy is using MIX'EM to initialize K-means. Finally, we report impressively high accuracy baselines (70\% on STL10) achieved solely by applying K-means to the "normalized" representations learned using the contrastive loss.
Replay in neural networks involves training on sequential data with memorized samples, which counteracts forgetting of previous behavior caused by non-stationarity. We present a method where these auxiliary samples are generated on the fly, given only the model that is being trained for the assessed objective, without extraneous buffers or generator networks. Instead the implicit memory of learned samples within the assessed model itself is exploited. Furthermore, whereas existing work focuses on reinforcing the full seen data distribution, we show that optimizing for not forgetting calls for the generation of samples that are specialized to each real training batch, which is more efficient and scalable. We consider high-level parallels with the brain, notably the use of a single model for inference and recall, the dependency of recalled samples on the current environment batch, top-down modulation of activations and learning, abstract recall, and the dependency between the degree to which a task is learned and the degree to which it is recalled. These characteristics emerge naturally from the method without being controlled for.