Erik Learned-Miller

Megan M. Baker, Alexander New, Mario Aguilar-Simon et al.

Despite the advancement of machine learning techniques in recent years, state-of-the-art systems lack robustness to "real world" events, where the input distributions and tasks encountered by the deployed systems will not be limited to the original training context, and systems will instead need to adapt to novel distributions and tasks while deployed. This critical gap may be addressed through the development of "Lifelong Learning" systems that are capable of 1) Continuous Learning, 2) Transfer and Adaptation, and 3) Scalability. Unfortunately, efforts to improve these capabilities are typically treated as distinct areas of research that are assessed independently, without regard to the impact of each separate capability on other aspects of the system. We instead propose a holistic approach, using a suite of metrics and an evaluation framework to assess Lifelong Learning in a principled way that is agnostic to specific domains or system techniques. Through five case studies, we show that this suite of metrics can inform the development of varied and complex Lifelong Learning systems. We highlight how the proposed suite of metrics quantifies performance trade-offs present during Lifelong Learning system development - both the widely discussed Stability-Plasticity dilemma and the newly proposed relationship between Sample Efficient and Robust Learning. Further, we make recommendations for the formulation and use of metrics to guide the continuing development of Lifelong Learning systems and assess their progress in the future.

Zitian Chen, Mingyu Ding, Yikang Shen et al.

We present a model that can perform multiple vision tasks and can be adapted to other downstream tasks efficiently. Despite considerable progress in multi-task learning, most efforts focus on learning from multi-label data: a single image set with multiple task labels. Such multi-label data sets are rare, small, and expensive. We say heterogeneous to refer to image sets with different task labels, or to combinations of single-task datasets. Few have explored training on such heterogeneous datasets. General-purpose vision models are still dominated by single-task pretraining, and it remains unclear how to scale up multi-task models by leveraging mainstream vision datasets designed for different purposes. The challenges lie in managing large intrinsic differences among vision tasks, including data distribution, architectures, task-specific modules, dataset scales, and sampling strategies. To address these challenges, we propose to modify and scale up mixture-of-experts (MoE) vision transformers, so that they can simultaneously learn classification, detection, and segmentation on diverse mainstream vision datasets including ImageNet, COCO, and ADE20K. Our approach achieves comparable results to single-task state-of-the-art models and demonstrates strong generalization on downstream tasks. Due to its emergent modularity, this general-purpose model decomposes into high-performing components, efficiently adapting to downstream tasks. We can fine-tune it with fewer training parameters, fewer model parameters, and less computation. Additionally, its modularity allows for easy expansion in continual-learning-without-forgetting scenarios. Finally, these functions can be controlled and combined to meet various demands of downstream tasks.

Tejas Panambur, Deep Chakraborty, Melissa Meyer et al.

Automatic terrain recognition in Mars rover images is an important problem not just for navigation, but for scientists interested in studying rock types, and by extension, conditions of the ancient Martian paleoclimate and habitability. Existing approaches to label Martian terrain either involve the use of non-expert annotators producing taxonomies of limited granularity (e.g. soil, sand, bedrock, float rock, etc.), or rely on generic class discovery approaches that tend to produce perceptual classes such as rover parts and landscape, which are irrelevant to geologic analysis. Expert-labeled datasets containing granular geological/geomorphological terrain categories are rare or inaccessible to public, and sometimes require the extraction of relevant categorical information from complex annotations. In order to facilitate the creation of a dataset with detailed terrain categories, we present a self-supervised method that can cluster sedimentary textures in images captured from the Mast camera onboard the Curiosity rover (Mars Science Laboratory). We then present a qualitative analysis of these clusters and describe their geologic significance via the creation of a set of granular terrain categories. The precision and geologic validation of these automatically discovered clusters suggest that our methods are promising for the rapid classification of important geologic features and will therefore facilitate our long-term goal of producing a large, granular, and publicly available dataset for Mars terrain recognition.

Ashish Singh, Ashutosh Singh, Prateek Agarwal et al.

Captions are crucial for understanding scientific visualizations and documents. Existing captioning methods for scientific figures rely on figure-caption pairs extracted from documents for training, many of which fall short with respect to metrics like helpfulness, explainability, and visual-descriptiveness [15] leading to generated captions being misaligned with reader preferences. To enable the generation of high-quality figure captions, we introduce FigCaps-HF a new framework for figure-caption generation that can incorporate domain expert feedback in generating captions optimized for reader preferences. Our framework comprises of 1) an automatic method for evaluating quality of figure-caption pairs, 2) a novel reinforcement learning with human feedback (RLHF) method to optimize a generative figure-to-caption model for reader preferences. We demonstrate the effectiveness of our simple learning framework by improving performance over standard fine-tuning across different types of models. In particular, when using BLIP as the base model, our RLHF framework achieves a mean gain of 35.7%, 16.9%, and 9% in ROUGE, BLEU, and Meteor, respectively. Finally, we release a large-scale benchmark dataset with human feedback on figure-caption pairs to enable further evaluation and development of RLHF techniques for this problem.

Zitian Chen, Yikang Shen, Mingyu Ding et al.

Optimization in multi-task learning (MTL) is more challenging than single-task learning (STL), as the gradient from different tasks can be contradictory. When tasks are related, it can be beneficial to share some parameters among them (cooperation). However, some tasks require additional parameters with expertise in a specific type of data or discrimination (specialization). To address the MTL challenge, we propose Mod-Squad, a new model that is Modularized into groups of experts (a 'Squad'). This structure allows us to formalize cooperation and specialization as the process of matching experts and tasks. We optimize this matching process during the training of a single model. Specifically, we incorporate mixture of experts (MoE) layers into a transformer model, with a new loss that incorporates the mutual dependence between tasks and experts. As a result, only a small set of experts are activated for each task. This prevents the sharing of the entire backbone model between all tasks, which strengthens the model, especially when the training set size and the number of tasks scale up. More interestingly, for each task, we can extract the small set of experts as a standalone model that maintains the same performance as the large model. Extensive experiments on the Taskonomy dataset with 13 vision tasks and the PASCAL-Context dataset with 5 vision tasks show the superiority of our approach.

Fabien Delattre, David Dirnfeld, Phat Nguyen et al.

We present an approach to estimating camera rotation in crowded, real-world scenes from handheld monocular video. While camera rotation estimation is a well-studied problem, no previous methods exhibit both high accuracy and acceptable speed in this setting. Because the setting is not addressed well by other datasets, we provide a new dataset and benchmark, with high-accuracy, rigorously verified ground truth, on 17 video sequences. Methods developed for wide baseline stereo (e.g., 5-point methods) perform poorly on monocular video. On the other hand, methods used in autonomous driving (e.g., SLAM) leverage specific sensor setups, specific motion models, or local optimization strategies (lagging batch processing) and do not generalize well to handheld video. Finally, for dynamic scenes, commonly used robustification techniques like RANSAC require large numbers of iterations, and become prohibitively slow. We introduce a novel generalization of the Hough transform on SO(3) to efficiently and robustly find the camera rotation most compatible with optical flow. Among comparably fast methods, ours reduces error by almost 50\% over the next best, and is more accurate than any method, irrespective of speed. This represents a strong new performance point for crowded scenes, an important setting for computer vision. The code and the dataset are available at https://fabiendelattre.com/robust-rotation-estimation.

Ke Xiao, Erik Learned-Miller, Evangelos Kalogerakis et al.

Mitral regurgitation (MR) is a heart valve disease with potentially fatal consequences that can only be forestalled through timely diagnosis and treatment. Traditional diagnosis methods are expensive, labor-intensive and require clinical expertise, posing a barrier to screening for MR. To overcome this impediment, we propose a new semi-supervised model for MR classification called CUSSP. CUSSP operates on cardiac imaging slices of the 4-chamber view of the heart. It uses standard computer vision techniques and contrastive models to learn from large amounts of unlabeled data, in conjunction with specialized classifiers to establish the first ever automated MR classification system. Evaluated on a test set of 179 labeled -- 154 non-MR and 25 MR -- sequences, CUSSP attains an F1 score of 0.69 and a ROC-AUC score of 0.88, setting the first benchmark result for this new task.

Ashish Singh, Michael J. Jones, Erik Learned-Miller

We develop a novel framework for single-scene video anomaly localization that allows for human-understandable reasons for the decisions the system makes. We first learn general representations of objects and their motions (using deep networks) and then use these representations to build a high-level, location-dependent model of any particular scene. This model can be used to detect anomalies in new videos of the same scene. Importantly, our approach is explainable - our high-level appearance and motion features can provide human-understandable reasons for why any part of a video is classified as normal or anomalous. We conduct experiments on standard video anomaly detection datasets (Street Scene, CUHK Avenue, ShanghaiTech and UCSD Ped1, Ped2) and show significant improvements over the previous state-of-the-art.

David Dirnfeld, Fabien Delattre, Pedro Miraldo et al.

Estimating camera motion from monocular video is a fundamental problem in computer vision, central to tasks such as SLAM, visual odometry, and structure-from-motion. Existing methods that recover the camera's heading under known rotation, whether from an IMU or an optimization algorithm, tend to perform well in low-noise, low-outlier conditions, but often decrease in accuracy or become computationally expensive as noise and outlier levels increase. To address these limitations, we propose a novel generalization of the Hough transform on the unit sphere (S(2)) to estimate the camera's heading. First, the method extracts correspondences between two frames and generates a great circle of directions compatible with each pair of correspondences. Then, by discretizing the unit sphere using a Fibonacci lattice as bin centers, each great circle casts votes for a range of directions, ensuring that features unaffected by noise or dynamic objects vote consistently for the correct motion direction. Experimental results on three datasets demonstrate that the proposed method is on the Pareto frontier of accuracy versus efficiency. Additionally, experiments on SLAM show that the proposed method reduces RMSE by correcting the heading during camera pose initialization.

Huaizu Jiang, Erik Learned-Miller

The cost volume, capturing the similarity of possible correspondences across two input images, is a key ingredient in state-of-the-art optical flow approaches. When sampling correspondences to build the cost volume, a large neighborhood radius is required to deal with large displacements, introducing a significant computational burden. To address this, coarse-to-fine or recurrent processing of the cost volume is usually adopted, where correspondence sampling in a local neighborhood with a small radius suffices. In this paper, we propose an alternative by constructing cost volumes with different dilation factors to capture small and large displacements simultaneously. A U-Net with skip connections is employed to convert the dilated cost volumes into interpolation weights between all possible captured displacements to get the optical flow. Our proposed model DCVNet only needs to process the cost volume once in a simple feedforward manner and does not rely on the sequential processing strategy. DCVNet obtains comparable accuracy to existing approaches and achieves real-time inference (30 fps on a mid-end 1080ti GPU). The code and model weights are available at https://github.com/neu-vi/ezflow.

Matheus Gadelha, Aruni RoyChowdhury, Gopal Sharma et al.

The problems of shape classification and part segmentation from 3D point clouds have garnered increasing attention in the last few years. Both of these problems, however, suffer from relatively small training sets, creating the need for statistically efficient methods to learn 3D shape representations. In this paper, we investigate the use of Approximate Convex Decompositions (ACD) as a self-supervisory signal for label-efficient learning of point cloud representations. We show that using ACD to approximate ground truth segmentation provides excellent self-supervision for learning 3D point cloud representations that are highly effective on downstream tasks. We report improvements over the state-of-the-art for unsupervised representation learning on the ModelNet40 shape classification dataset and significant gains in few-shot part segmentation on the ShapeNetPart dataset.Code available at https://github.com/matheusgadelha/PointCloudLearningACD

Bruce Levin, Erik Learned-Miller

We show that the likelihood function for a multinomial vector observed under arbitrary interval censoring constraints on the frequencies or their partial sums is completely log-concave by proving that the constrained sample spaces comprise M-convex subsets of the discrete simplex.

Deep Chakraborty, Yann LeCun, Tim G. J. Rudner et al.

A number of different architectures and loss functions have been applied to the problem of self-supervised learning (SSL), with the goal of developing embeddings that provide the best possible pre-training for as-yet-unknown, lightly supervised downstream tasks. One of these SSL criteria is to maximize the entropy of a set of embeddings in some compact space. But the goal of maximizing the embedding entropy often depends -- whether explicitly or implicitly -- upon high dimensional entropy estimates, which typically perform poorly in more than a few dimensions. In this paper, we motivate an effective entropy maximization criterion (E2MC), defined in terms of easy-to-estimate, low-dimensional constraints. We demonstrate that using it to continue training an already-trained SSL model for only a handful of epochs leads to a consistent and, in some cases, significant improvement in downstream performance. We perform careful ablation studies to show that the improved performance is due to the proposed add-on criterion. We also show that continued pre-training with alternative criteria does not lead to notable improvements, and in some cases, even degrades performance.

Yilun Kuang, Yash Dagade, Deep Chakraborty et al.

Self-supervised learning aims to learn maximally informative representations, but explicit information maximization is hindered by the curse of dimensionality. Existing methods like VCReg address this by regularizing first and second-order feature statistics, which cannot fully achieve maximum entropy. We propose Radial-VCReg, which augments VCReg with a radial Gaussianization loss that aligns feature norms with the Chi distribution-a defining property of high-dimensional Gaussians. We prove that Radial-VCReg transforms a broader class of distributions towards normality compared to VCReg and show on synthetic and real-world datasets that it consistently improves performance by reducing higher-order dependencies and promoting more diverse and informative representations.

Ashish Singh, Michael J. Jones, Kuan-Chuan Peng et al.

Our work addresses the problem of learning to localize objects in an open-world setting, i.e., given the bounding box information of a limited number of object classes during training, the goal is to localize all objects, belonging to both the training and unseen classes in an image, during inference. Towards this end, recent work in this area has focused on improving the characterization of objects either explicitly by proposing new objective functions (localization quality) or implicitly using object-centric auxiliary-information, such as depth information, pixel/region affinity map etc. In this work, we address this problem by incorporating background information to guide the learning of the notion of objectness. Specifically, we propose a novel framework to discover background regions in an image and train an object proposal network to not detect any objects in these regions. We formulate the background discovery task as that of identifying image regions that are not discriminative, i.e., those that are redundant and constitute low information content. We conduct experiments on standard benchmarks to showcase the effectiveness of our proposed approach and observe significant improvements over the previous state-of-the-art approaches for this task.

Shifan Zhu, Zhipeng Tang, Michael Yang et al.

Our paper proposes a direct sparse visual odometry method that combines event and RGB-D data to estimate the pose of agile-legged robots during dynamic locomotion and acrobatic behaviors. Event cameras offer high temporal resolution and dynamic range, which can eliminate the issue of blurred RGB images during fast movements. This unique strength holds a potential for accurate pose estimation of agile-legged robots, which has been a challenging problem to tackle. Our framework leverages the benefits of both RGB-D and event cameras to achieve robust and accurate pose estimation, even during dynamic maneuvers such as jumping and landing a quadruped robot, the Mini-Cheetah. Our major contributions are threefold: Firstly, we introduce an adaptive time surface (ATS) method that addresses the whiteout and blackout issue in conventional time surfaces by formulating pixel-wise decay rates based on scene complexity and motion speed. Secondly, we develop an effective pixel selection method that directly samples from event data and applies sample filtering through ATS, enabling us to pick pixels on distinct features. Lastly, we propose a nonlinear pose optimization formula that simultaneously performs 3D-2D alignment on both RGB-based and event-based maps and images, allowing the algorithm to fully exploit the benefits of both data streams. We extensively evaluate the performance of our framework on both public datasets and our own quadruped robot dataset, demonstrating its effectiveness in accurately estimating the pose of agile robots during dynamic movements.

Ashutosh Singh, Ashish Singh, Aria Masoomi et al.

Subspace clustering algorithms are used for understanding the cluster structure that explains the dataset well. These methods are extensively used for data-exploration tasks in various areas of Natural Sciences. However, most of these methods fail to handle unwanted biases in datasets. For datasets where a data sample represents multiple attributes, naively applying any clustering approach can result in undesired output. To this end, we propose a novel framework for jointly removing unwanted attributes (biases) while learning to cluster data points in individual subspaces. Assuming we have information about the bias, we regularize the clustering method by adversarially learning to minimize the mutual information between the data and the unwanted attributes. Our experimental result on synthetic and real-world datasets demonstrate the effectiveness of our approach.

Pia Bideau, Erik Learned-Miller, Cordelia Schmid et al.

Both a good understanding of geometrical concepts and a broad familiarity with objects lead to our excellent perception of moving objects. The human ability to detect and segment moving objects works in the presence of multiple objects, complex background geometry, motion of the observer and even camouflage. How humans perceive moving objects so reliably is a longstanding research question in computer vision and borrows findings from related areas such as psychology, cognitive science and physics. One approach to the problem is to teach a deep network to model all of these effects. This contrasts with the strategy used by human vision, where cognitive processes and body design are tightly coupled and each is responsible for certain aspects of correctly identifying moving objects. Similarly from the computer vision perspective, there is evidence that classical, geometry-based techniques are better suited to the "motion-based" parts of the problem, while deep networks are more suitable for modeling appearance. In this work, we argue that the coupling of camera rotation and camera translation can create complex motion fields that are difficult for a deep network to untangle directly. We present a novel probabilistic model to estimate the camera's rotation given the motion field. We then rectify the flow field to obtain a rotation-compensated motion field for subsequent segmentation. This strategy of first estimating camera motion, and then allowing a network to learn the remaining parts of the problem, yields improved results on the widely used DAVIS benchmark as well as the recently published motion segmentation data set MoCA (Moving Camouflaged Animals).

Gopal Sharma, Bidya Dash, Aruni RoyChowdhury et al.

We present PriFit, a semi-supervised approach for label-efficient learning of 3D point cloud segmentation networks. PriFit combines geometric primitive fitting with point-based representation learning. Its key idea is to learn point representations whose clustering reveals shape regions that can be approximated well by basic geometric primitives, such as cuboids and ellipsoids. The learned point representations can then be re-used in existing network architectures for 3D point cloud segmentation, and improves their performance in the few-shot setting. According to our experiments on the widely used ShapeNet and PartNet benchmarks, PriFit outperforms several state-of-the-art methods in this setting, suggesting that decomposability into primitives is a useful prior for learning representations predictive of semantic parts. We present a number of ablative experiments varying the choice of geometric primitives and downstream tasks to demonstrate the effectiveness of the method.

Cheng Gu, Erik Learned-Miller, Daniel Sheldon et al.

Event cameras, inspired by biological vision systems, provide a natural and data efficient representation of visual information. Visual information is acquired in the form of events that are triggered by local brightness changes. Each pixel location of the camera's sensor records events asynchronously and independently with very high temporal resolution. However, because most brightness changes are triggered by relative motion of the camera and the scene, the events recorded at a single sensor location seldom correspond to the same world point. To extract meaningful information from event cameras, it is helpful to register events that were triggered by the same underlying world point. In this work we propose a new model of event data that captures its natural spatio-temporal structure. We start by developing a model for aligned event data. That is, we develop a model for the data as though it has been perfectly registered already. In particular, we model the aligned data as a spatio-temporal Poisson point process. Based on this model, we develop a maximum likelihood approach to registering events that are not yet aligned. That is, we find transformations of the observed events that make them as likely as possible under our model. In particular we extract the camera rotation that leads to the best event alignment. We show new state of the art accuracy for rotational velocity estimation on the DAVIS 240C dataset. In addition, our method is also faster and has lower computational complexity than several competing methods.

Chen Qu, Hamed Zamani, Liu Yang et al.

In this work, we address multi-modal information needs that contain text questions and images by focusing on passage retrieval for outside-knowledge visual question answering. This task requires access to outside knowledge, which in our case we define to be a large unstructured passage collection. We first conduct sparse retrieval with BM25 and study expanding the question with object names and image captions. We verify that visual clues play an important role and captions tend to be more informative than object names in sparse retrieval. We then construct a dual-encoder dense retriever, with the query encoder being LXMERT, a multi-modal pre-trained transformer. We further show that dense retrieval significantly outperforms sparse retrieval that uses object expansion. Moreover, dense retrieval matches the performance of sparse retrieval that leverages human-generated captions.

Yash Chandak, Scott Niekum, Bruno Castro da Silva et al.

When faced with sequential decision-making problems, it is often useful to be able to predict what would happen if decisions were made using a new policy. Those predictions must often be based on data collected under some previously used decision-making rule. Many previous methods enable such off-policy (or counterfactual) estimation of the expected value of a performance measure called the return. In this paper, we take the first steps towards a universal off-policy estimator (UnO) -- one that provides off-policy estimates and high-confidence bounds for any parameter of the return distribution. We use UnO for estimating and simultaneously bounding the mean, variance, quantiles/median, inter-quantile range, CVaR, and the entire cumulative distribution of returns. Finally, we also discuss Uno's applicability in various settings, including fully observable, partially observable (i.e., with unobserved confounders), Markovian, non-Markovian, stationary, smoothly non-stationary, and discrete distribution shifts.

Lijun Zhang, Xiao Liu, Erik Learned-Miller et al.

When capturing images in low-light conditions, the images often suffer from low visibility, which not only degrades the visual aesthetics of images, but also significantly degenerates the performance of many computer vision algorithms. In this paper, we propose a self-supervised low-light image enhancement framework (SID-NISM), which consists of two components, a Self-supervised Image Decomposition Network (SID-Net) and a Nonlinear Illumination Saturation Mapping function (NISM). As a self-supervised network, SID-Net could decompose the given low-light image into its reflectance, illumination and noise directly without any prior training or reference image, which distinguishes it from existing supervised-learning methods greatly. Then, the decomposed illumination map will be enhanced by NISM. Having the restored illumination map, the enhancement can be achieved accordingly. Experiments on several public challenging low-light image datasets reveal that the images enhanced by SID-NISM are more natural and have less unexpected artifacts.

Zitian Chen, Subhransu Maji, Erik Learned-Miller

Few-shot learning aims to build classifiers for new classes from a small number of labeled examples and is commonly facilitated by access to examples from a distinct set of 'base classes'. The difference in data distribution between the test set (novel classes) and the base classes used to learn an inductive bias often results in poor generalization on the novel classes. To alleviate problems caused by the distribution shift, previous research has explored the use of unlabeled examples from the novel classes, in addition to labeled examples of the base classes, which is known as the transductive setting. In this work, we show that, surprisingly, off-the-shelf self-supervised learning outperforms transductive few-shot methods by 3.9% for 5-shot accuracy on miniImageNet without using any base class labels. This motivates us to examine more carefully the role of features learned through self-supervision in few-shot learning. Comprehensive experiments are conducted to compare the transferability, robustness, efficiency, and the complementarity of supervised and self-supervised features.

Aruni RoyChowdhury, Xiang Yu, Kihyuk Sohn et al.

While deep face recognition has benefited significantly from large-scale labeled data, current research is focused on leveraging unlabeled data to further boost performance, reducing the cost of human annotation. Prior work has mostly been in controlled settings, where the labeled and unlabeled data sets have no overlapping identities by construction. This is not realistic in large-scale face recognition, where one must contend with such overlaps, the frequency of which increases with the volume of data. Ignoring identity overlap leads to significant labeling noise, as data from the same identity is split into multiple clusters. To address this, we propose a novel identity separation method based on extreme value theory. It is formulated as an out-of-distribution detection algorithm, and greatly reduces the problems caused by overlapping-identity label noise. Considering cluster assignments as pseudo-labels, we must also overcome the labeling noise from clustering errors. We propose a modulation of the cosine loss, where the modulation weights correspond to an estimate of clustering uncertainty. Extensive experiments on both controlled and real settings demonstrate our method's consistent improvements over supervised baselines, e.g., 11.6% improvement on IJB-A verification.

Zitian Chen, Zhiqiang Shen, Jiahui Yu et al.

After learning a new object category from image-level annotations (with no object bounding boxes), humans are remarkably good at precisely localizing those objects. However, building good object localizers (i.e., detectors) currently requires expensive instance-level annotations. While some work has been done on learning detectors from weakly labeled samples (with only class labels), these detectors do poorly at localization. In this work, we show how to build better object detectors from weakly labeled images of new categories by leveraging knowledge learned from fully labeled base categories. We call this novel learning paradigm cross-supervised object detection. We propose a unified framework that combines a detection head trained from instance-level annotations and a recognition head learned from image-level annotations, together with a spatial correlation module that bridges the gap between detection and recognition. These contributions enable us to better detect novel objects with image-level annotations in complex multi-object scenes such as the COCO dataset.

Huaizu Jiang, Ishan Misra, Marcus Rohrbach et al.

Popularized as 'bottom-up' attention, bounding box (or region) based visual features have recently surpassed vanilla grid-based convolutional features as the de facto standard for vision and language tasks like visual question answering (VQA). However, it is not clear whether the advantages of regions (e.g. better localization) are the key reasons for the success of bottom-up attention. In this paper, we revisit grid features for VQA, and find they can work surprisingly well - running more than an order of magnitude faster with the same accuracy (e.g. if pre-trained in a similar fashion). Through extensive experiments, we verify that this observation holds true across different VQA models (reporting a state-of-the-art accuracy on VQA 2.0 test-std, 72.71), datasets, and generalizes well to other tasks like image captioning. As grid features make the model design and training process much simpler, this enables us to train them end-to-end and also use a more flexible network design. We learn VQA models end-to-end, from pixels directly to answers, and show that strong performance is achievable without using any region annotations in pre-training. We hope our findings help further improve the scientific understanding and the practical application of VQA. Code and features will be made available.

Huaizu Jiang, Deqing Sun, Varun Jampani et al.

We introduce a compact network for holistic scene flow estimation, called SENSE, which shares common encoder features among four closely-related tasks: optical flow estimation, disparity estimation from stereo, occlusion estimation, and semantic segmentation. Our key insight is that sharing features makes the network more compact, induces better feature representations, and can better exploit interactions among these tasks to handle partially labeled data. With a shared encoder, we can flexibly add decoders for different tasks during training. This modular design leads to a compact and efficient model at inference time. Exploiting the interactions among these tasks allows us to introduce distillation and self-supervised losses in addition to supervised losses, which can better handle partially labeled real-world data. SENSE achieves state-of-the-art results on several optical flow benchmarks and runs as fast as networks specifically designed for optical flow. It also compares favorably against the state of the art on stereo and scene flow, while consuming much less memory.

Erik Learned-Miller, Philip S. Thomas

We present a new method for constructing a confidence interval for the mean of a bounded random variable from samples of the random variable. We conjecture that the confidence interval has guaranteed coverage, i.e., that it contains the mean with high probability for all distributions on a bounded interval, for all samples sizes, and for all confidence levels. This new method provides confidence intervals that are competitive with those produced using Student's t-statistic, but does not rely on normality assumptions. In particular, its only requirement is that the distribution be bounded on a known finite interval.

Aruni RoyChowdhury, Prithvijit Chakrabarty, Ashish Singh et al.

This work addresses the unsupervised adaptation of an existing object detector to a new target domain. We assume that a large number of unlabeled videos from this domain are readily available. We automatically obtain labels on the target data by using high-confidence detections from the existing detector, augmented with hard (misclassified) examples acquired by exploiting temporal cues using a tracker. These automatically-obtained labels are then used for re-training the original model. A modified knowledge distillation loss is proposed, and we investigate several ways of assigning soft-labels to the training examples from the target domain. Our approach is empirically evaluated on challenging face and pedestrian detection tasks: a face detector trained on WIDER-Face, which consists of high-quality images crawled from the web, is adapted to a large-scale surveillance data set; a pedestrian detector trained on clear, daytime images from the BDD-100K driving data set is adapted to all other scenarios such as rainy, foggy, night-time. Our results demonstrate the usefulness of incorporating hard examples obtained from tracking, the advantage of using soft-labels via distillation loss versus hard-labels, and show promising performance as a simple method for unsupervised domain adaptation of object detectors, with minimal dependence on hyper-parameters.

Hang Su, Varun Jampani, Deqing Sun et al.

Convolutions are the fundamental building block of CNNs. The fact that their weights are spatially shared is one of the main reasons for their widespread use, but it also is a major limitation, as it makes convolutions content agnostic. We propose a pixel-adaptive convolution (PAC) operation, a simple yet effective modification of standard convolutions, in which the filter weights are multiplied with a spatially-varying kernel that depends on learnable, local pixel features. PAC is a generalization of several popular filtering techniques and thus can be used for a wide range of use cases. Specifically, we demonstrate state-of-the-art performance when PAC is used for deep joint image upsampling. PAC also offers an effective alternative to fully-connected CRF (Full-CRF), called PAC-CRF, which performs competitively, while being considerably faster. In addition, we also demonstrate that PAC can be used as a drop-in replacement for convolution layers in pre-trained networks, resulting in consistent performance improvements.

Colin Samplawski, Heesung Kwon, Erik Learned-Miller et al.

Zero-shot learning (ZSL) is one of the most extreme forms of learning from scarce labeled data. It enables predicting that images belong to classes for which no labeled training instances are available. In this paper, we present a new ZSL framework that leverages both label attribute side information and a semantic label hierarchy. We present two methods, lifted zero-shot prediction and a custom conditional random field (CRF) model, that integrate both forms of side information. We propose benchmark tasks for this framework that focus on making predictions across a range of semantic levels. We show that lifted zero-shot prediction can dramatically outperform baseline methods when making predictions within specified semantic levels, and that the probability distribution provided by the CRF model can be leveraged to yield further performance improvements when making unconstrained predictions over the hierarchy.

Marwan Mattar, Michael Ross, Erik Learned-Miller

Congealing is a flexible nonparametric data-driven framework for the joint alignment of data. It has been successfully applied to the joint alignment of binary images of digits, binary images of object silhouettes, grayscale MRI images, color images of cars and faces, and 3D brain volumes. This research enhances congealing to practically and effectively apply it to curve data. We develop a parameterized set of nonlinear transformations that allow us to apply congealing to this type of data. We present positive results on aligning synthetic and real curve data sets and conclude with a discussion on extending this work to simultaneous alignment and clustering.

SouYoung Jin, Aruni RoyChowdhury, Huaizu Jiang et al.

Important gains have recently been obtained in object detection by using training objectives that focus on {\em hard negative} examples, i.e., negative examples that are currently rated as positive or ambiguous by the detector. These examples can strongly influence parameters when the network is trained to correct them. Unfortunately, they are often sparse in the training data, and are expensive to obtain. In this work, we show how large numbers of hard negatives can be obtained {\em automatically} by analyzing the output of a trained detector on video sequences. In particular, detections that are {\em isolated in time}, i.e., that have no associated preceding or following detections, are likely to be hard negatives. We describe simple procedures for mining large numbers of such hard negatives (and also hard {\em positives}) from unlabeled video data. Our experiments show that retraining detectors on these automatically obtained examples often significantly improves performance. We present experiments on multiple architectures and multiple data sets, including face detection, pedestrian detection and other object categories.

Huaizu Jiang, Erik Learned-Miller, Gustav Larsson et al.

As an agent moves through the world, the apparent motion of scene elements is (usually) inversely proportional to their depth. It is natural for a learning agent to associate image patterns with the magnitude of their displacement over time: as the agent moves, faraway mountains don't move much; nearby trees move a lot. This natural relationship between the appearance of objects and their motion is a rich source of information about the world. In this work, we start by training a deep network, using fully automatic supervision, to predict relative scene depth from single images. The relative depth training images are automatically derived from simple videos of cars moving through a scene, using recent motion segmentation techniques, and no human-provided labels. This proxy task of predicting relative depth from a single image induces features in the network that result in large improvements in a set of downstream tasks including semantic segmentation, joint road segmentation and car detection, and monocular (absolute) depth estimation, over a network trained from scratch. The improvement on the semantic segmentation task is greater than those produced by any other automatically supervised methods. Moreover, for monocular depth estimation, our unsupervised pre-training method even outperforms supervised pre-training with ImageNet. In addition, we demonstrate benefits from learning to predict (unsupervised) relative depth in the specific videos associated with various downstream tasks. We adapt to the specific scenes in those tasks in an unsupervised manner to improve performance. In summary, for semantic segmentation, we present state-of-the-art results among methods that do not use supervised pre-training, and we even exceed the performance of supervised ImageNet pre-trained models for monocular depth estimation, achieving results that are comparable with state-of-the-art methods.

Huaizu Jiang, Deqing Sun, Varun Jampani et al.

Given two consecutive frames, video interpolation aims at generating intermediate frame(s) to form both spatially and temporally coherent video sequences. While most existing methods focus on single-frame interpolation, we propose an end-to-end convolutional neural network for variable-length multi-frame video interpolation, where the motion interpretation and occlusion reasoning are jointly modeled. We start by computing bi-directional optical flow between the input images using a U-Net architecture. These flows are then linearly combined at each time step to approximate the intermediate bi-directional optical flows. These approximate flows, however, only work well in locally smooth regions and produce artifacts around motion boundaries. To address this shortcoming, we employ another U-Net to refine the approximated flow and also predict soft visibility maps. Finally, the two input images are warped and linearly fused to form each intermediate frame. By applying the visibility maps to the warped images before fusion, we exclude the contribution of occluded pixels to the interpolated intermediate frame to avoid artifacts. Since none of our learned network parameters are time-dependent, our approach is able to produce as many intermediate frames as needed. We use 1,132 video clips with 240-fps, containing 300K individual video frames, to train our network. Experimental results on several datasets, predicting different numbers of interpolated frames, demonstrate that our approach performs consistently better than existing methods.

SouYoung Jin, Hang Su, Chris Stauffer et al.

We present an end-to-end system for detecting and clustering faces by identity in full-length movies. Unlike works that start with a predefined set of detected faces, we consider the end-to-end problem of detection and clustering together. We make three separate contributions. First, we combine a state-of-the-art face detector with a generic tracker to extract high quality face tracklets. We then introduce a novel clustering method, motivated by the classic graph theory results of Erdős and Rényi. It is based on the observations that large clusters can be fully connected by joining just a small fraction of their point pairs, while just a single connection between two different people can lead to poor clustering results. This suggests clustering using a verification system with very few false positives but perhaps moderate recall. We introduce a novel verification method, rank-1 counts verification, that has this property, and use it in a link-based clustering scheme. Finally, we define a novel end-to-end detection and clustering evaluation metric allowing us to assess the accuracy of the entire end-to-end system. We present state-of-the-art results on multiple video data sets and also on standard face databases.

Haw-Shiuan Chang, Erik Learned-Miller, Andrew McCallum

Self-paced learning and hard example mining re-weight training instances to improve learning accuracy. This paper presents two improved alternatives based on lightweight estimates of sample uncertainty in stochastic gradient descent (SGD): the variance in predicted probability of the correct class across iterations of mini-batch SGD, and the proximity of the correct class probability to the decision threshold. Extensive experimental results on six datasets show that our methods reliably improve accuracy in various network architectures, including additional gains on top of other popular training techniques, such as residual learning, momentum, ADAM, batch normalization, dropout, and distillation.

Pia Bideau, Erik Learned-Miller

Motion segmentation is currently an active area of research in computer Vision. The task of comparing different methods of motion segmentation is complicated by the fact that researchers may use subtly different definitions of the problem. Questions such as "Which objects are moving?", "What is background?", and "How can we use motion of the camera to segment objects, whether they are static or moving?" are clearly related to each other, but lead to different algorithms, and imply different versions of the ground truth. This report has two goals. The first is to offer a precise definition of motion segmentation so that the intent of an algorithm is as well-defined as possible. The second is to report on new versions of three previously existing data sets that are compatible with this definition. We hope that this more detailed definition, and the three data sets that go with it, will allow more meaningful comparisons of certain motion segmentation methods.

Li Yang Ku, Erik Learned-Miller, Rod Grupen

In this work, we provide a solution for posturing the anthropomorphic Robonaut-2 hand and arm for grasping based on visual information. A mapping from visual features extracted from a convolutional neural network (CNN) to grasp points is learned. We demonstrate that a CNN pre-trained for image classification can be applied to a grasping task based on a small set of grasping examples. Our approach takes advantage of the hierarchical nature of the CNN by identifying features that capture the hierarchical support relations between filters in different CNN layers and locating their 3D positions by tracing activations backwards in the CNN. When this backward trace terminates in the RGB-D image, important manipulable structures are thereby localized. These features that reside in different layers of the CNN are then associated with controllers that engage different kinematic subchains in the hand/arm system for grasping. A grasping dataset is collected using demonstrated hand/object relationships for Robonaut-2 to evaluate the proposed approach in terms of the precision of the resulting preshape postures. We demonstrate that this approach outperforms baseline approaches in cluttered scenarios on the grasping dataset and a point cloud based approach on a grasping task using Robonaut-2.

Huaizu Jiang, Erik Learned-Miller

The Faster R-CNN has recently demonstrated impressive results on various object detection benchmarks. By training a Faster R-CNN model on the large scale WIDER face dataset, we report state-of-the-art results on two widely used face detection benchmarks, FDDB and the recently released IJB-A.

Pia Bideau, Erik Learned-Miller

The human ability to detect and segment moving objects works in the presence of multiple objects, complex background geometry, motion of the observer, and even camouflage. In addition to all of this, the ability to detect motion is nearly instantaneous. While there has been much recent progress in motion segmentation, it still appears we are far from human capabilities. In this work, we derive from first principles a new likelihood function for assessing the probability of an optical flow vector given the 3D motion direction of an object. This likelihood uses a novel combination of the angle and magnitude of the optical flow to maximize the information about the true motions of objects. Using this new likelihood and several innovations in initialization, we develop a motion segmentation algorithm that beats current state-of-the-art methods by a large margin. We compare to five state-of-the-art methods on two established benchmarks, and a third new data set of camouflaged animals, which we introduce to push motion segmentation to the next level.

Manjunath Narayana, Allen Hanson, Erik Learned-Miller

Recent work on background subtraction has shown developments on two major fronts. In one, there has been increasing sophistication of probabilistic models, from mixtures of Gaussians at each pixel [7], to kernel density estimates at each pixel [1], and more recently to joint domainrange density estimates that incorporate spatial information [6]. Another line of work has shown the benefits of increasingly complex feature representations, including the use of texture information, local binary patterns, and recently scale-invariant local ternary patterns [4]. In this work, we use joint domain-range based estimates for background and foreground scores and show that dynamically choosing kernel variances in our kernel estimates at each individual pixel can significantly improve results. We give a heuristic method for selectively applying the adaptive kernel calculations which is nearly as accurate as the full procedure but runs much faster. We combine these modeling improvements with recently developed complex features [4] and show significant improvements on a standard backgrounding benchmark.

Manjunath Narayana, Allen Hanson, Erik Learned-Miller

In its early implementations, background modeling was a process of building a model for the background of a video with a stationary camera, and identifying pixels that did not conform well to this model. The pixels that were not well-described by the background model were assumed to be moving objects. Many systems today maintain models for the foreground as well as the background, and these models compete to explain the pixels in a video. In this paper, we argue that the logical endpoint of this evolution is to simply use Bayes' rule to classify pixels. In particular, it is essential to have a background likelihood, a foreground likelihood, and a prior at each pixel. A simple application of Bayes' rule then gives a posterior probability over the label. The only remaining question is the quality of the component models: the background likelihood, the foreground likelihood, and the prior. We describe a model for the likelihoods that is built by using not only the past observations at a given pixel location, but by also including observations in a spatial neighborhood around the location. This enables us to model the influence between neighboring pixels and is an improvement over earlier pixelwise models that do not allow for such influence. Although similar in spirit to the joint domain-range model, we show that our model overcomes certain deficiencies in that model. We use a spatially dependent prior for the background and foreground. The background and foreground labels from the previous frame, after spatial smoothing to account for movement of objects,are used to build the prior for the current frame.

Manjunath Narayana, Allen Hanson, Erik Learned-Miller

In moving camera videos, motion segmentation is commonly performed using the image plane motion of pixels, or optical flow. However, objects that are at different depths from the camera can exhibit different optical flows even if they share the same real-world motion. This can cause a depth-dependent segmentation of the scene. Our goal is to develop a segmentation algorithm that clusters pixels that have similar real-world motion irrespective of their depth in the scene. Our solution uses optical flow orientations instead of the complete vectors and exploits the well-known property that under camera translation, optical flow orientations are independent of object depth. We introduce a probabilistic model that automatically estimates the number of observed independent motions and results in a labeling that is consistent with real-world motion in the scene. The result of our system is that static objects are correctly identified as one segment, even if they are at different depths. Color features and information from previous frames in the video sequence are used to correct occasional errors due to the orientation-based segmentation. We present results on more than thirty videos from different benchmarks. The system is particularly robust on complex background scenes containing objects at significantly different depths

Aruni RoyChowdhury, Tsung-Yu Lin, Subhransu Maji et al.

The recent explosive growth in convolutional neural network (CNN) research has produced a variety of new architectures for deep learning. One intriguing new architecture is the bilinear CNN (B-CNN), which has shown dramatic performance gains on certain fine-grained recognition problems [15]. We apply this new CNN to the challenging new face recognition benchmark, the IARPA Janus Benchmark A (IJB-A) [12]. It features faces from a large number of identities in challenging real-world conditions. Because the face images were not identified automatically using a computerized face detection system, it does not have the bias inherent in such a database. We demonstrate the performance of the B-CNN model beginning from an AlexNet-style network pre-trained on ImageNet. We then show results for fine-tuning using a moderate-sized and public external database, FaceScrub [17]. We also present results with additional fine-tuning on the limited training data provided by the protocol. In each case, the fine-tuned bilinear model shows substantial improvements over the standard CNN. Finally, we demonstrate how a standard CNN pre-trained on a large face database, the recently released VGG-Face model [20], can be converted into a B-CNN without any additional feature training. This B-CNN improves upon the CNN performance on the IJB-A benchmark, achieving 89.5% rank-1 recall.

Hang Su, Subhransu Maji, Evangelos Kalogerakis et al.

A longstanding question in computer vision concerns the representation of 3D shapes for recognition: should 3D shapes be represented with descriptors operating on their native 3D formats, such as voxel grid or polygon mesh, or can they be effectively represented with view-based descriptors? We address this question in the context of learning to recognize 3D shapes from a collection of their rendered views on 2D images. We first present a standard CNN architecture trained to recognize the shapes' rendered views independently of each other, and show that a 3D shape can be recognized even from a single view at an accuracy far higher than using state-of-the-art 3D shape descriptors. Recognition rates further increase when multiple views of the shapes are provided. In addition, we present a novel CNN architecture that combines information from multiple views of a 3D shape into a single and compact shape descriptor offering even better recognition performance. The same architecture can be applied to accurately recognize human hand-drawn sketches of shapes. We conclude that a collection of 2D views can be highly informative for 3D shape recognition and is amenable to emerging CNN architectures and their derivatives.