Keypoint detection and matching is a fundamental task in many computer vision problems, from shape reconstruction, to structure from motion, to AR/VR applications and robotics. It is a well-studied problem with remarkable successes such as SIFT, and more recent deep learning approaches. While great robustness is exhibited by these techniques with respect to noise, illumination variation, and rigid motion transformations, less attention has been placed on image distortion sensitivity. In this work, we focus on the case when this is caused by the geometry of the cameras used for image acquisition, and consider the keypoint detection and matching problem between the hybrid scenario of a fisheye and a projective image. We build on a state-of-the-art approach and derive a self-supervised procedure that enables training an interest point detector and descriptor network. We also collected two new datasets for additional training and testing in this unexplored scenario, and we demonstrate that current approaches are suboptimal because they are designed to work in traditional projective conditions, while the proposed approach turns out to be the most effective.
Automated cellular instance segmentation is a process utilized for accelerating biological research for the past two decades, and recent advancements have produced higher quality results with less effort from the biologist. Most current endeavors focus on completely cutting the researcher out of the picture by generating highly generalized models. However, these models invariably fail when faced with novel data, distributed differently than the ones used for training. Rather than approaching the problem with methods that presume the availability of large amounts of target data and computing power for retraining, in this work we address the even greater challenge of designing an approach that requires minimal amounts of new annotated data as well as training time. We do so by designing specialized contrastive losses that leverage the few annotated samples very efficiently. A large set of results show that 3 to 5 annotations lead to models with accuracy that: 1) significantly mitigate the covariate shift effects; 2) matches or surpasses other adaptation methods; 3) even approaches methods that have been fully retrained on the target distribution. The adaptation training is only a few minutes, paving a path towards a balance between model performance, computing requirements and expert-level annotation needs.
The pandemic of these very recent years has led to a dramatic increase in people wearing protective masks in public venues. This poses obvious challenges to the pervasive use of face recognition technology that now is suffering a decline in performance. One way to address the problem is to revert to face recovery methods as a preprocessing step. Current approaches to face reconstruction and manipulation leverage the ability to model the face manifold, but tend to be generic. We introduce a method that is specific for the recovery of the face image from an image of the same individual wearing a mask. We do so by designing a specialized GAN inversion method, based on an appropriate set of losses for learning an unmasking encoder. With extensive experiments, we show that the approach is effective at unmasking face images. In addition, we also show that the identity information is preserved sufficiently well to improve face verification performance based on several face recognition benchmark datasets.
Person re-identification is a challenging task because of the high intra-class variance induced by the unrestricted nuisance factors of variations such as pose, illumination, viewpoint, background, and sensor noise. Recent approaches postulate that powerful architectures have the capacity to learn feature representations invariant to nuisance factors, by training them with losses that minimize intra-class variance and maximize inter-class separation, without modeling nuisance factors explicitly. The dominant approaches use either a discriminative loss with margin, like the softmax loss with the additive angular margin, or a metric learning loss, like the triplet loss with batch hard mining of triplets. Since the softmax imposes feature normalization, it limits the gradient flow supervising the feature embedding. We address this by joining the losses and leveraging the triplet loss as a proxy for the missing gradients. We further improve invariance to nuisance factors by adding the discriminative task of predicting attributes. Our extensive evaluation highlights that when only a holistic representation is learned, we consistently outperform the state-of-the-art on the three most challenging datasets. Such representations are easier to deploy in practical systems. Finally, we found that joining the losses removes the requirement for having a margin in the softmax loss while increasing performance.
Self supervised learning (SSL) has become a very successful technique to harness the power of unlabeled data, with no annotation effort. A number of developed approaches are evolving with the goal of outperforming supervised alternatives, which have been relatively successful. One main issue in SSL is robustness of the approaches under different settings. In this paper, for the first time, we recognize the fundamental limits of SSL coming from the use of a single-supervisory signal. To address this limitation, we leverage the power of uncertainty representation to devise a robust and general standard hierarchical learning/training protocol for any SSL baseline, regardless of their assumptions and approaches. Essentially, using the information bottleneck principle, we decompose feature learning into a two-stage training procedure, each with a distinct supervision signal. This double supervision approach is captured in two key steps: 1) invariance enforcement to data augmentation, and 2) fuzzy pseudo labeling (both hard and soft annotation). This simple, yet, effective protocol which enables cross-class/cluster feature learning, is instantiated via an initial training of an ensemble of models through invariance enforcement to data augmentation as first training phase, and then assigning fuzzy labels to the original samples for the second training phase. We consider multiple alternative scenarios with double supervision and evaluate the effectiveness of our approach on recent baselines, covering four different SSL paradigms, including geometrical, contrastive, non-contrastive, and hard/soft whitening (redundancy reduction) baselines. Extensive experiments under multiple settings show that the proposed training protocol consistently improves the performance of the former baselines, independent of their respective underlying principles.
Conventional active learning (AL) frameworks aim to reduce the cost of data annotation by actively requesting the labeling for the most informative data points. However, introducing AL to data hungry deep learning algorithms has been a challenge. Some proposed approaches include uncertainty-based techniques, geometric methods, implicit combination of uncertainty-based and geometric approaches, and more recently, frameworks based on semi/self supervised techniques. In this paper, we address two specific problems in this area. The first is the need for efficient exploitation/exploration trade-off in sample selection in AL. For this, we present an innovative integration of recent progress in both uncertainty-based and geometric frameworks to enable an efficient exploration/exploitation trade-off in sample selection strategy. To this end, we build on a computationally efficient approximate of Thompson sampling with key changes as a posterior estimator for uncertainty representation. Our framework provides two advantages: (1) accurate posterior estimation, and (2) tune-able trade-off between computational overhead and higher accuracy. The second problem is the need for improved training protocols in deep AL. For this, we use ideas from semi/self supervised learning to propose a general approach that is independent of the specific AL technique being used. Taken these together, our framework shows a significant improvement over the state-of-the-art, with results that are comparable to the performance of supervised-learning under the same setting. We show empirical results of our framework, and comparative performance with the state-of-the-art on four datasets, namely, MNIST, CIFAR10, CIFAR100 and ImageNet to establish a new baseline in two different settings.
Plant species identification in the wild is a difficult problem in part due to the high variability of the input data, but also because of complications induced by the long-tail effects of the datasets distribution. Inspired by the most recent fine-grained visual classification approaches which are based on attention to mitigate the effects of data variability, we explore the idea of using object detection as a form of attention. We introduce a bottom-up approach based on detecting plant organs and fusing the predictions of a variable number of organ-based species classifiers. We also curate a new dataset with a long-tail distribution for evaluating plant organ detection and organ-based species identification, which is publicly available.
Autoencoder networks are unsupervised approaches aiming at combining generative and representational properties by learning simultaneously an encoder-generator map. Although studied extensively, the issues of whether they have the same generative power of GANs, or learn disentangled representations, have not been fully addressed. We introduce an autoencoder that tackles these issues jointly, which we call Adversarial Latent Autoencoder (ALAE). It is a general architecture that can leverage recent improvements on GAN training procedures. We designed two autoencoders: one based on a MLP encoder, and another based on a StyleGAN generator, which we call StyleALAE. We verify the disentanglement properties of both architectures. We show that StyleALAE can not only generate 1024x1024 face images with comparable quality of StyleGAN, but at the same resolution can also produce face reconstructions and manipulations based on real images. This makes ALAE the first autoencoder able to compare with, and go beyond the capabilities of a generator-only type of architecture.
The quest for deeper understanding of biological systems has driven the acquisition of increasingly larger multidimensional image datasets. Inspecting and manipulating data of this complexity is very challenging in traditional visualization systems. We developed syGlass, a software package capable of visualizing large scale volumetric data with inexpensive virtual reality head-mounted display technology. This allows leveraging stereoscopic vision to significantly improve perception of complex 3D structures, and provides immersive interaction with data directly in 3D. We accomplished this by developing highly optimized data flow and volume rendering pipelines, tested on datasets up to 16TB in size, as well as tools available in a virtual reality GUI to support advanced data exploration, annotation, and cataloguing.
Novelty detection is the problem of identifying whether a new data point is considered to be an inlier or an outlier. We assume that training data is available to describe only the inlier distribution. Recent approaches primarily leverage deep encoder-decoder network architectures to compute a reconstruction error that is used to either compute a novelty score or to train a one-class classifier. While we too leverage a novel network of that kind, we take a probabilistic approach and effectively compute how likely is that a sample was generated by the inlier distribution. We achieve this with two main contributions. First, we make the computation of the novelty probability feasible because we linearize the parameterized manifold capturing the underlying structure of the inlier distribution, and show how the probability factorizes and can be computed with respect to local coordinates of the manifold tangent space. Second, we improved the training of the autoencoder network. An extensive set of results show that the approach achieves state-of-the-art results on several benchmark datasets.