Resource-constrained weight deployment is a task of immense practical importance. Recently, there has been interest in the specific task of Delta Compression, where parties each hold a common base model and only communicate compressed weight updates. However, popular parameter efficient updates such as Low Rank Adaptation (LoRA) face inherent representation limitations - which are especially pronounced when combined with aggressive quantization. To overcome this, we build on recent work that improves LoRA representation capacity by using fixed-frequency sinusoidal functions to increase stable rank without adding additional parameters. We extend this to the quantized setting and present the first theoretical analysis showing how stable rank evolves under quantization. From this, we introduce SineLoRA∆, a principled and effective method for delta compression that improves the expressivity of quantized low-rank adapters by applying a sinusoidal activation. We validate SineLoRA∆ across a diverse variety of domains - including language modeling, vision-language tasks, and text-to-image generation - achieving up to 66% memory reduction with similar performance. We additionally provide a novel application of the canonical Bjøntegaard Delta metric to consistently compare adapter compression changes across the rate-distortion curve.

We present a surface-template-free method for reconstructing dense, rigged (re-animatable) 3D models from monocular videos. By combining robust pose optimization with differentiable Gaussian splatting, this work bridges the gap between flexible template-free approaches and the visual quality of template-based methods. Starting from noisy 2D keypoints, we refine 3D poses through kinematic and temporal constraints, then attach Gaussian primitives to the optimized skeleton for differentiable supervision and rendering. We demonstrate dense, rigged (re-animatable) 3D models without surface templates across humans, animals, insects, and everyday articulated objects, and we empirically show that closing the rendering–pose loop improves 3D lifting from noisy landmarks. A key enabler of these template-free reconstructions is our kinematic optimization, which reduces 3D pose error by 20–25% relative to template-free baselines; at the same time, our results approach template-based visual metrics (PSNR/SSIM within 5%). We also demonstrate the method’s practical utility by detecting and correcting geometric inconsistencies in AI-generated videos. While limited to articulated subjects with detectable keypoints, the approach provides a practical pipeline that serves as a drop-in refinement to improve 3D lifting in existing pipelines and enables the creation of rigged 3D assets from casual captures when expensive surface templates (e.g., MoCap-derived) are unavailable.

In this work, we introduce the 3D Gaussian Point Encoder, an explicit per-point embedding built on mixtures of learned 3D Gaussians. This explicit geometric representation for 3D recognition tasks is a departure from widely used implicit representations such as PointNet. However, it is difficult to learn 3D Gaussian encoders in end-to-end fashion with standard optimizers. We develop optimization techniques based on natural gradients and distillation from PointNets to find a Gaussian Basis that can reconstruct PointNet activations. The resulting 3D Gaussian Point Encoders are faster and more parameter efficient than traditional PointNets. As in the 3D reconstruction literature where there has been considerable interest in the move from implicit (e.g., NeRF) to explicit (e.g., Gaussian Splatting) representations, we can take advantage of computational geometry heuristics to accelerate 3D Gaussian Point Encoders further. We extend filtering techniques from 3D Gaussian Splatting to construct encoders that run 2.7× faster as a comparable accuracy PointNet while using 46% less memory and 88% fewer FLOPs. Furthermore, we demonstrate the effectiveness of 3D Gaussian Point Encoders as a component in Mamba3D, running 1.27× faster and achieving a reduction in memory and FLOPs by 42% and 54% respectively. 3D Gaussian Point Encoders are lightweight enough to achieve high framerates on CPU-only devices.

Splatting-based 3D reconstruction methods have gained popularity with the advent of 3D Gaussian Splatting, efficiently synthesizing high-quality novel views. These methods commonly resort to using exponential family functions, such as the Gaussian function, as reconstruction kernels due to their anisotropic nature, ease of projection, and differentiability in rasterization. However, the field remains restricted to variations within the exponential family, leaving generalized reconstruction kernels largely underexplored, partly due to the lack of easy integrability in 3D to 2D projections. In this light, we show that a class of decaying anisotropic radial basis functions (DARBFs), which are non-negative functions of the Mahalanobis distance, supports splatting by approximating the Gaussian function's closed-form integration advantage. With this fresh perspective, we demonstrate varying performances across selected DARB reconstruction kernels, achieving comparable training convergence and memory footprints, with on-par PSNR, SSIM, and LPIPS results.

We present a method for dynamic surface reconstruction of large-scale urban scenes from LiDAR. Depth-based re-constructions tend to focus on small-scale objects or large-scale SLAM reconstructions that treat moving objects as outliers. We take a holistic perspective and optimize a compositional model of a dynamic scene that decomposes the world into rigidly-moving objects and the background. To achieve this, we take inspiration from recent novel view syn-thesis methods and frame the reconstruction problem as a global optimization over neural surfaces, ego poses, and object poses, which minimizes the error between composed spacetime surfaces and input LiDAR scans. In contrast to view synthesis methods, which typically minimize 2D errors with gradient descent, we minimize a 3D point-to-surface error by coordinate descent, which we decompose into reg-istration and surface reconstruction steps. Each step can be handled well by off-the-shelf methods without any re-training. We analyze the surface reconstruction step for rolling-shutter LiDARs, and show that deskewing operations common in continuous time SLAM can be applied to dynamic objects as well, improving results over prior art by an order of magnitude. Beyond pursuing dynamic reconstruction as a goal in and of itself, we propose that such a system can be used to auto-label partially annotated se-quences and produce ground truth annotation for hard-to-label problems such as depth completion and scene flow. Please refer https://anishmadan23.github.io/smore/ for more visual results.

We present a spatio-temporal perspective on category-agnostic 3D lifting of 2D keypoints over a temporal se-quence. Our approach differs from existing state-of-the-art methods that are either: (i) object-agnostic, but can only operate on individual frames, or (ii) can model space-time dependencies, but are only designed to work with a single object category. Our approach is grounded in two core prin-ciples. First, general information about similar objects can be leveraged to achieve better performance when there is little object-specific training data. Second, a temporally-proximate context window is advantageous for achieving consistency throughout a sequence. These two principles allow us to outperform current state-of-the-art methods on per-frame and per-sequence metrics for a variety of animal categories. Lastly, we release a new synthetic dataset con-taining 3D skeletons and motion sequences for a variety of animal categories.

Neural fields encode continuous multidimensional signals as neural networks, enabling diverse applications in computer vision, robotics, and geometry. While Adam is effective for stochastic optimization, it often requires long training times. To address this, we explore alternative optimization techniques to accelerate training without sacrificing accuracy. Traditional second-order methods like L-BFGS are unsuitable for stochastic settings. We propose a theoretical framework for training neural fields with curvatureaware diagonal preconditioners, demonstrating their effectiveness across tasks such as image reconstruction, shape modeling, and Neural Radiance Fields (NeRF)1.

We highlight a curious empirical result within modern Vision Transformers (ViTs). Specifically, self-attention catastrophically fails to train unless it is used in conjunction with a skip connection. This is in contrast to other elements of a ViT that continue to exhibit good performance (albeit suboptimal) when skip connections are removed. Further, we show that this critical dependence on skip connections is a relatively new phenomenon, with previous deep architectures (\eg, CNNs) exhibiting good performance in their absence. In this paper, we theoretically characterize that the self-attention mechanism is fundamentally ill-conditioned and is, therefore, uniquely dependent on skip connections for regularization. Additionally, we propose Token Graying -- a simple yet effective complement (to skip connections) that further improves the condition of input tokens. We validate our approach in both supervised and self-supervised training methods.

Transformers have transformed modern machine learning, driving breakthroughs in computer vision, natural language processing, and robotics. At the core of their success lies the attention mechanism, which enables the modeling of global dependencies among input tokens. However, we reveal that the attention block in transformers suffers from inherent ill-conditioning, which hampers gradient-based optimization and leads to inefficient training. To address this, we develop a theoretical framework that establishes a direct relationship between the conditioning of the attention block and that of the embedded tokenized data. Building on this insight, we introduce conditioned embedded tokens, a method that systematically modifies the embedded tokens to improve the conditioning of the attention mechanism. Our analysis demonstrates that this approach significantly mitigates ill-conditioning, leading to more stable and efficient training. We validate our methodology across various transformer architectures, achieving consistent improvements in image classification, object detection, instance segmentation, and natural language processing, highlighting its broad applicability and effectiveness.

Low-rank decomposition has emerged as a vital tool for enhancing parameter efficiency in neural network architectures, gaining traction across diverse applications in machine learning. These techniques significantly lower the number of parameters, striking a balance between compactness and performance. However, a common challenge has been the compromise between parameter efficiency and the accuracy of the model, where reduced parameters often lead to diminished accuracy compared to their full-rank counterparts. In this work, we propose a novel theoretical framework that integrates a sinusoidal function within the low-rank decomposition process. This approach not only preserves the benefits of the parameter efficiency characteristic of low-rank methods but also increases the decomposition's rank, thereby enhancing model performance. Our method proves to be a plug in enhancement for existing low-rank models, as evidenced by its successful application in Vision Transformers (ViT), Large Language Models (LLMs), Neural Radiance Fields (NeRF) and 3D shape modelling.

We present a theoretical analysis of the Jacobian of a attention block within a transformer, showing that it is governed by the query, key, and value projections that define the attention mechanism. Leveraging this insight, we introduce a method that systematically alters the spectral properties of each attention layer to reduce the Jacobian’s condition number, thereby improving the overall conditioning of the attention layers within a transformer network. We empirically show that this improved Jacobian conditioning translates to enhanced performance in practice. Our approach is simple, broadly applicable, and can be easily integrated as a drop-in replacement for a wide range of existing attention mechanisms. We validate its effectiveness across diverse transformer architectures and tasks, demonstrating consistent improvements in performance.

Convolutional Neural Networks (CNNs) inherently encode strong inductive biases, enabling effective generalization on small-scale datasets. In this paper, we propose integrating this inductive bias into ViTs, not through an architectural intervention but solely through initialization. The motivation here is to have a ViT that can enjoy strong CNN-like performance when data assets are small, but can still scale to ViT-like performance as the data expands. Our approach is motivated by our empirical results that random impulse filters can achieve commensurate performance to learned filters within a CNN. We improve upon current ViT initialization strategies, which typically rely on empirical heuristics such as using attention weights from pretrained models or focusing on the distribution of attention weights without enforcing structures. Empirical results demonstrate that our method significantly outperforms standard ViT initialization across numerous small and medium-scale benchmarks, including Food-101, CIFAR-10, CIFAR-100, STL-10, Flowers, and Pets, while maintaining comparative performance on large-scale datasets such as ImageNet-1K. Moreover, our initialization strategy can be easily integrated into various transformer-based architectures such as Swin Transformer and MLP-Mixer with consistent improvements in performance.

The test-time optimization of scene flow—using a coordinate network as a neural prior [27]—has gained popularity due to its simplicity, lack of dataset bias, and state-of-the-art performance. We observe, however, that although coordinate networks capture general motions by implicitly regularizing the scene flow predictions to be spatially smooth, the neural prior by itself is unable to identify the underlying multi-body rigid motions present in real-world data. To address this, we show that multi-body rigidity can be achieved without the cumbersome and brittle strategy of constraining the SE(3) parameters of each rigid body as done in previous works. This is achieved by regularizing the scene flow optimization to encourage isometry in flow predictions for rigid bodies. This strategy enables multi-body rigidity in scene flow while maintaining a continuous flow field, hence allowing dense long-term scene flow integration across a sequence of point clouds. We conduct extensive experiments on real-world datasets and demonstrate that our approach outperforms the state-of-the-art in 3D scene flow and long-term point-wise 4D trajectory prediction. The code is available at: https://github.com/kavisha725/MBNSF.

End-to-end trained per-point embeddings are an essential ingredient of any state-of-the-art 3D point cloud processing such as detection or alignment. Methods like PointNet [17], or the more recent point cloud transformer [7]—and its variants—all employ learned per-point embeddings. Despite impressive performance, such approaches are sensitive to out-of-distribution (OOD) noise and outliers. In this paper, we explore the role of an analytical per-point embedding based on the criterion of bandwidth. The concept of bandwidth enables us to draw connections with an alternate per-point embedding—positional embedding, particularly random Fourier features. We present compelling robust results across downstream tasks such as point cloud classification and registration with several categories of OOD noise.

The lifting of a 3D structure and camera from 2D land-marks is at the cornerstone of the discipline of computer vision. Traditional methods have been confined to specific rigid objects, such as those in Perspective-n-Point (PnP) problems, but deep learning has expanded our capability to reconstruct a wide range of object classes (e.g. C3DPO [18] and PAUL [24]) with resilience to noise, occlusions, and perspective distortions. However, all these techniques have been limited by the fundamental need to establish correspondences across the 3D training data, significantly limiting their utility to applications where one has an abundance of “in-correspondence” 3D data. Our approach harnesses the inherent permutation equivariance of transformers to manage varying numbers of points per 3D data instance, withstands occlusions, and generalizes to unseen categories. We demonstrate state-of-the-art performance across 2D-3D lifting task benchmarks. Since our approach can be trained across such a broad class of structures, we refer to it simply as a 3D Lifting Foundation Model (3D-LFM) - the first of its kind.

In the realm of computer vision, Neural Fields have gained prominence as a contemporary tool harnessing neural networks for signal representation. Despite the remarkable progress in adapting these networks to solve a variety of problems, the field still lacks a comprehensive theoretical framework. This article aims to address this gap by delving into the intricate interplay between initialization and activation, providing a foundational basis for the robust optimization of Neural Fields. Our theoretical insights reveal a deep-seated connection among network initialization, architectural choices, and the optimization process, emphasizing the need for a holistic approach when designing cutting-edge Neural Fields.

We tackle semi-supervised object detection based on motion cues. Recent results suggest that heuristic-based clustering methods in conjunction with object trackers can be used to pseudo-label instances of moving objects and use these as supervisory signals to train 3D object detectors in Lidar data without manual supervision. We re-think this approach and suggest that both, object detection, as well as motion-inspired pseudo-labeling, can be tackled in a data-driven manner. We leverage recent advances in scene flow estimation to obtain point trajectories from which we extract long-term, class-agnostic motion patterns. Revisiting correlation clustering in the context of message passing networks, we learn to group those motion patterns to cluster points to object instances. By estimating the full extent of the objects, we obtain per-scan 3D bounding boxes that we use to supervise a Lidar object detection network. Our method not only outperforms prior heuristic-based approaches (57.5 AP, + 14 improvement over prior work), more importantly, we show we can pseudo-label and train object detectors across datasets.

Implicit Neural Representations (INRs) have gained popularity for encoding signals as compact, differentiable entities. While commonly using techniques like Fourier positional encodings or non-traditional activation functions (e.g., Gaussian, sinusoid, or wavelets) to capture high-frequency content, their properties lack exploration within a unified theoretical framework. Addressing this gap, we conduct a comprehensive analysis of these activations from a sampling theory perspective. Our investigation reveals that, especially in shallow INRs, $\mathrm{sinc}$ activations—previously unused in conjunction with INRs—are theoretically optimal for signal encoding. Additionally, we establish a connection between dynamical systems and INRs, leveraging sampling theory to bridge these two paradigms.

Popular benchmarks for self-supervised LiDAR scene flow (stereoKITTI, and FlyingThings3D) have unrealistic rates of dynamic motion, unrealistic correspondences, and unrealistic sampling patterns. As a result, progress on these benchmarks is misleading and may cause researchers to focus on the wrong problems. We evaluate a suite of top methods on a suite of real-world datasets (Argoverse 2.0, Waymo, and NuScenes) and report several conclusions. First, we find that performance on stereoKITTI is negatively correlated with performance on real-world data. Second, we find that one of this task’s key components –removing the dominant ego-motion –is better solved by classic ICP than any tested method. Finally, we show that despite the emphasis placed on learning, most performance gains are caused by pre- and post-processing steps: piecewise-rigid refinement and ground removal. We demonstrate this through a baseline method that combines these processing steps with a learning-free test-time flow optimization. This baseline outperforms every evaluated method.

We propose a novel method to enhance the performance of coordinate-MLPs (also referred to as neural fields) by learning instance-specific positional embeddings. End-to-end optimization of positional embedding parameters along with network weights leads to poor generalization performance. Instead, we develop a generic framework to learn the positional embedding based on the classic graph-Laplacian regularization, which can implicitly balance the trade-off between memorization and generalization. This framework is then used to propose a novel positional embedding scheme, where the hyperparameters are learned per coordinate (i.e instance) to deliver optimal performance. We show that the proposed embedding achieves better performance with higher stability compared to the well-established random Fourier features (RFF). Further, we demonstrate that the proposed embedding scheme yields stable gradients, enabling seamless integration into deep architectures as intermediate layers.

In this paper we present a new method for deformable NeRF that can directly use optical flow as supervision. We overcome the major challenge with respect to the computationally inefficiency of enforcing the flow constraints to the backward deformation field, used by deformable NeRFs. Specifically, we show that inverting the backward deformation function is actually not needed for computing scene flows between frames. This insight dramatically simplifies the problem, as one is no longer constrained to deformation functions that can be analytically inverted. Instead, thanks to the weak assumptions required by our derivation based on the inverse function theorem, our approach can be extended to a broad class of commonly used backward deformation field. We present results on monocular novel view synthesis with rapid object motion, and demonstrate significant improvements over baselines without flow supervision.

We address outdoor Neural Radiance Fields (NeRF) [23] with real-world camera views and LiDAR maps. Existing methods usually require densely-sampled source views and do not perform well with the open source camera-LiDAR datasets. In this paper, our design leverages 1) LiDAR sensors for strong 3D geometry priors that significantly improve the ray sampling locality, and 2) Conditional Adversarial Networks (cGANs) [15] to recover image details since aggregating embeddings from imperfect LiDAR maps causes artifacts. Our experiments show that while NeRF baselines produce either noisy or blurry results on Argoverse 2 [42], our system not only outperforms baselines in image quality metrics under both clean and noisy conditions, but also obtains closer Detectron2 [43] results to the ground truth images. Furthermore, this system can be used in data augmentation for training a pose regression network [3] and multi-season view synthesis. We hope this work to serve as a new LiDAR-based NeRF baseline that pushes this research direction forward (released here).

Neural Scene Flow Prior (NSFP) is of significant interest to the vision community due to its inherent robustness to out-of-distribution (OOD) effects and its ability to deal with dense lidar points. The approach utilizes a coordinate neural network to estimate scene flow at runtime, without any training. However, it is up to 100 times slower than current state-of-the-art learning methods. In other applications such as image, video, and radiance function reconstruction innovations in speeding up the runtime performance of coordinate networks have centered upon architectural changes. In this paper, we demonstrate that scene flow is different—with the dominant computational bottleneck stemming from the loss function itself (i.e., Chamfer distance). Further, we rediscover the distance transform (DT) as an efficient, correspondence-free loss function that dramatically speeds up the runtime optimization. Our fast neural scene flow (FNSF) approach reports for the first time real-time performance comparable to learning methods, without any training or OOD bias on two of the largest open autonomous driving (AV) lidar datasets Waymo Open [62] and Argoverse [8].

Coordinate networks are widely used in computer vision due to their ability to represent signals as compressed, continuous entities. However, training these networks with first-order optimizers can be slow, hindering their use in real-time applications. Recent works have opted for shallow voxel-based representations to achieve faster training, but this sacrifices memory efficiency. This work proposes a solution that leverages second-order optimization methods to significantly reduce training times for coordinate networks while maintaining their compressibility. Experiments demonstrate the effectiveness of this approach on various signal modalities, such as audio, images, videos, shape and neural radiance fields (NeRF).

Characterizing the remarkable generalization properties of over-parameterized neural networks remains an open problem. A growing body of recent literature shows that the bias of stochastic gradient descent (SGD) and architecture choice implicitly leads to better generalization. In this paper, we show on the contrary that, independently of architecture, SGD can itself be the cause of poor generalization if one does not ensure good initialization. Specifically, we prove that *any* differentiably parameterized model, trained under gradient flow, obeys a weak spectral bias law which states that sufficiently high frequencies train arbitrarily slowly. This implies that very high frequencies present at initialization will remain after training, and hamper generalization. Further, we empirically test the developed theoretical insights using practical, deep networks. Finally, we contrast our framework with that supplied by the *flat-minima* conjecture and show that Fourier analysis grants a more reliable framework for understanding the generalization of neural networks.

We introduce a general theoretical framework, designed for the study of gradient optimisation of deep neural networks, that encompasses ubiquitous architecture choices including batch normalisation, weight normalisation and skip connections. Our framework determines the curvature and regularity properties of multilayer loss landscapes in terms of their constituent layers, thereby elucidating the roles played by normalisation layers and skip connections in globalising these properties. We then demonstrate the utility of this framework in two respects. First, we give the only proof of which we are aware that a class of deep neural networks can be trained using gradient descent to global optima even when such optima only exist at infinity, as is the case for the cross-entropy cost. Second, we identify a novel causal mechanism by which skip connections accelerate training, which we verify predictively with ResNets on MNIST, CIFAR10, CIFAR100 and ImageNet.

The role of quantization within implicit/coordinate neural networks is still not fully understood. We note that using a canonical fixed quantization scheme during training produces poor performance at low bit-rates due to the network weight distributions changing over the course of training. In this work, we show that a non-uniform quantization of neural weights can lead to significant improvements. Specifically, we demonstrate that a clustered quantization enables improved reconstruction. Finally, by characterising a trade-off between quantization and network capacity, we demonstrate that it is possible (while memory inefficient) to reconstruct signals using binary neural networks. We demonstrate our findings experimentally on 2D image reconstruction and 3D radiance fields; and show that simple quantization methods and architecture search can achieve compression of NeRF to less than 16kb with minimal loss in performance (323x smaller than the original NeRF).

In comparison to classical shallow representation learning techniques, deep neural networks have achieved superior performance in nearly every application benchmark. But despite their clear empirical advantages, it is still not well understood what makes them so effective. To approach this question, we introduce deep frame approximation: a unifying framework for constrained representation learning with structured overcomplete frames. While exact inference requires iterative optimization, it may be approximated by the operations of a feed-forward deep neural network. We indirectly analyze how model capacity relates to frame structures induced by architectural hyperparameters such as depth, width, and skip connections. We quantify these structural differences with the deep frame potential, a data-independent measure of coherence linked to representation uniqueness and stability. As a criterion for model selection, we show correlation with generalization error on a variety of common deep network architectures and datasets. We also demonstrate how recurrent networks implementing iterative optimization algorithms can achieve performance comparable to their feed-forward approximations while improving adversarial robustness. This connection to the established theory of overcomplete representations suggests promising new directions for principled deep network architecture design with less reliance on ad-hoc engineering.

We address the problem of map sparsification for long-term visual localization. For map sparsification, a commonly employed assumption is that the pre-build map and the later captured localization query are consistent. However, this assumption can be easily violated in the dynamic world. Additionally, the map size grows as new data accumulate through time, causing large data overhead in the long term. In this paper, we aim to overcome the environmental changes and reduce the map size at the same time by selecting points that are valuable to future localization. Inspired by the recent progress in Graph Neural Network (GNN), we propose the first work that models SfM maps as heterogeneous graphs and predicts 3D point importance scores with a GNN, which enables us to directly exploit the rich information in the SfM map graph. Two novel supervisions are proposed: 1) a data-fitting term for selecting valuable points to future localization based on training queries; 2) a K-Cover term for selecting sparse points with full-map coverage. The experiments show that our method selected map points on stable and widely visible structures and out-performed baselines in localization performance

Although provably robust to translational perturbations, convolutional neural networks (CNNs) are known to suffer from extreme performance degradation when presented at test time with more general geometric transformations of inputs. Recently, this limitation has motivated a shift infocus from CNNs to Capsule Networks (CapsNets). However, CapsNets suffer from admitting relatively few theoretical guarantees of invariance. We introduce a rigourous mathematical framework to permit invariance to any Lie group of warps, exclusively using convolutions (over Lie groups), without the need for capsules. Previous work on group convolutions has been hampered by strong assumptions about the group, which precludes the application of such techniques to common warps in computer vision such as affine and homographic. Our framework enables the implementation of group convolutions over any finite-dimensional Lie group. We empirically validate our approach on the benchmark affine-invariant classification task, where we achieve ~30% improvement in accuracy against conventional CNNs while outperforming most CapsNets. As further illustration of the generality of our framework, we train a homography-convolutional model which achieves superior robustness on a homography-perturbed dataset, where CapsNet results degrade.

Neural networks trained with SGD were recently shown to rely preferentially on linearly-predictive features and can ignore complex, equally-predictive ones. This simplicity bias can explain their lack of robustness out of distribution (OOD). The more complex the task to learn, the more likely it is that statistical artifacts (i.e. selection biases, spurious correlations) are simpler than the mechanisms to learn. We demonstrate that the simplicity bias can be mitigated and OOD generalization improved. We train a set of similar models to fit the data in different ways using a penalty on the alignment of their input gradients. We show theoretically and empirically that this induces the learning of more com-plex predictive patterns. OOD generalization fundamentally requires information beyond i. i.d. examples, such as multiple training environ-ments, counterfactual examples, or other side information. Our approach shows that we can defer this requirement to an independent model selection stage. We obtain SOTA re-sults in visual recognition on biased data and generalization across visual domains. The method - the first to evade the simplicity bias - highlights the need for a better under-standing and control of inductive biases in deep learning.

Neural priors are a promising direction to capture low-level vision statistics without relying on handcrafted regularizers. Recent works have successfully shown the use of neural architecture biases to implicitly regularize image denoising, super-resolution, inpainting, synthesis, scene flow, among others. They do not rely on large-scale datasets to capture prior statistics and thus generalize well to out-of-the-distribution data. Inspired by such advances, we investigate neural priors for trajectory representation. Traditionally, trajectories have been represented by a set of handcrafted bases that have limited expressibility. Here, we propose a neural trajectory prior to capture continuous spatio-temporal information without the need for offline data. We demonstrate how our proposed objective is optimized during runtime to estimate trajectories for two important tasks: Non-Rigid Structure from Motion (NRSfM) and lidar scene flow integration for self-driving scenes. Our results are competitive to many state-of-the-art methods for both tasks.

Labeling articulated objects in unconstrained settings has a wide variety of applications including entertainment, neuroscience, psychology, ethology, and many fields of medicine. Large offline labeled datasets do not exist for all but the most common articulated object categories (e.g., humans). Hand labeling these landmarks within a video sequence is a laborious task. Learned landmark detectors can help, but can be error-prone when trained from only a few examples. Multi-camera systems that train fine-grained detectors have shown significant promise in detecting such errors, allowing for self-supervised solutions that only need a small percentage of the video sequence to be hand-labeled. The approach, however, is based on calibrated cameras and rigid geometry, making it expensive, difficult to manage, and impractical in real-world scenarios. In this paper, we address these bottlenecks by combining a non-rigid 3D neural prior with deep flow to obtain high-fidelity landmark estimates from videos with only two or three uncalibrated, handheld cameras. With just a few annotations (representing $1-2\%$ of the frames), we are able to produce 2D results comparable to state-of-the-art fully supervised methods, along with 3D reconstructions that are impossible with other existing approaches. Our Multi-view Bootstrapping in the Wild (MBW) approach demonstrates impressive results on standard human datasets, as well as tigers, cheetahs, fish, colobus monkeys, chimpanzees, and flamingos from videos captured casually in a zoo. We release the codebase for MBW as well as this challenging zoo dataset consisting of image frames of tail-end distribution categories with their corresponding 2D and 3D labels generated from minimal human intervention.

We show that typical implicit regularization assumptions for deep neural networks (for regression) do not hold for coordinate-MLPs, a family of MLPs that are now ubiquitous in computer vision for representing high-frequency signals. Lack of such implicit bias disrupts smooth interpolations between training samples, and hampers generalizing across signal regions with different spectra. We investigate this behavior through a Fourier lens and uncover that as the bandwidth of a coordinate-MLP is enhanced, lower frequencies tend to get suppressed unless a suitable prior is provided explicitly. Based on these insights, we propose a simple regularization technique that can mitigate the above problem, which can be incorporated into existing networks without any architectural modifications.

Multi-view triangulation is the gold standard for 3D reconstruction from 2D correspondences given known calibration and sufficient views. However in practice, expensive multi-view setups – involving tens sometimes hundreds of cameras – are required in order to obtain the high fidelity 3D reconstructions necessary for many modern applications. In this paper we present a novel approach that leverages recent advances in 2D-3D lifting using neural shape priors while also enforcing multi-view equivariance. We show how our method can achieve comparable fidelity to expensive calibrated multi-view rigs using a limited (2-3) number of uncalibrated camera views.

Despite their unmatched performance, deep neural networks remain susceptible to targeted attacks by nearly imperceptible levels of adversarial noise. While the underlying cause of this sensitivity is not well understood, theoretical analyses can be simplified by reframing each layer of a feed-forward network as an approximate solution to a sparse coding problem. Iterative solutions using basis pursuit are theoretically more stable and have improved adversarial robustness. However, cascading layer-wise pursuit implementations suffer from error accumulation in deeper networks. In contrast, our new method of deep pursuit approximates the activations of all layers as a single global optimization problem, allowing us to consider deeper, real-world architectures with skip connections such as residual networks. Experimentally, our approach demonstrates improved robustness to adversarial noise.

We address the generalization ability of recent learning-based point cloud registration methods. Despite their success, these approaches tend to have poor performance when applied to mismatched conditions that are not well-represented in the training set, such as unseen object categories, different complex scenes, or unknown depth sensors. In these circumstances, it has often been better to rely on classical non-learning methods (e.g., Iterative Closest Point), which have better generalization ability. Hybrid learning methods, that use learning for predicting point correspondences and then a deterministic step for alignment, have offered some respite, but are still limited in their generalization abilities. We revisit a recent innovation— PointNetLK [1]—and show that the inclusion of an analytical Jacobian can exhibit remarkable generalization properties while reaping the inherent fidelity benefits of a learning framework. Our approach not only outperforms the state-of-the-art in mismatched conditions but also produces results competitive with current learning methods when operating on real-world test data close to the training set.

Recent success in casting Non-rigid Structure from Motion (NRSfM) as an unsupervised deep learning problem has raised fundamental questions about what novelty in NRSfM prior could the deep learning offer. In this paper we advocate for a 3D deep auto-encoder framework to be used explicitly as the NRSfM prior. The framework is unique as: (i) it learns the 3D auto-encoder weights solely from 2D projected measurements, and (ii) it is Procrustean in that it jointly resolves the unknown rigid pose for each shape instance. We refer to this architecture as a Procustean Autoencoder for Unsupervised Lifting (PAUL), and demonstrate state-of-the-art performance across a number of benchmarks in comparison to recent innovations such as Deep NRSfM [21] and C3PDO [32].

Neural Radiance Fields (NeRF) [31] have recently gained a surge of interest within the computer vision community for its power to synthesize photorealistic novel views of real-world scenes. One limitation of NeRF, however, is its requirement of accurate camera poses to learn the scene representations. In this paper, we propose Bundle-Adjusting Neural Radiance Fields (BARF) for training NeRF from imperfect (or even unknown) camera poses — the joint problem of learning neural 3D representations and registering camera frames. We establish a theoretical connection to classical image alignment and show that coarse-to-fine registration is also applicable to NeRF. Furthermore, we show that naïvely applying positional encoding in NeRF has a negative impact on registration with a synthesis-based objective. Experiments on synthetic and real-world data show that BARF can effectively optimize the neural scene representations and resolve large camera pose misalignment at the same time. This enables view synthesis and localization of video sequences from unknown camera poses, opening up new avenues for visual localization systems (e.g. SLAM) and potential applications for dense 3D mapping and reconstruction.

We address the problem of image registration to a compressed 3D map. While this is most often performed by comparing LiDAR scans to the point cloud based map, it depends on an expensive LiDAR sensor at run time and the large point cloud based map creates overhead in data storage and transmission. Recently, efforts have been underway to replace the expensive LiDAR sensor with cheaper cameras and perform 2D-3D localization. In contrast to the previous work that learns relative pose by comparing projected depth and camera images, we propose HyperMap, a paradigm shift from online depth map feature extraction to offline 3D map feature computation for the 2D-3D camera registration task through end-to-end training. In the proposed pipeline, we first perform offline 3D sparse convolution to extract and compress the voxelwise hypercolumn features for the whole map. Then at run-time, we project and decode the compressed map features to the rough initial camera pose to form a virtual feature image. A Convolutional Neural Network (CNN) is then used to predict the relative pose between the camera image and the virtual feature image. In addition, we propose an efficient occlusion handling layer, specifically designed for large point clouds, to remove occluded points in projection. Our experiments on synthetic and real datasets show that, by moving the feature computation load offline and compressing, we reduced map size by 87−94% while maintaining comparable or better accuracy.

We aim to assess the performance of LiDAR-to-map registration on compressive maps. Modern autonomous vehicles utilize pre-built HD (High-Definition) maps to perform sensor-to-map registration, which recovers pose estimation failures and reduces drift in a large-scale environment. However, sensor-to-map registration is usually realized by registering the sensor to a dense 3D model, which occupies massive storage space in the HD map and requires much data processing overhead. Although smaller 3D models are preferable, the optimal compressive map format for preservation of the best registration performance remains unclear.In this paper, we propose a novel and challenging benchmark to evaluate existing LiDAR-to-map registration methods from three perspectives: map compressibility, robustness, and precision. We compared various map formats, including raw points, hierarchical GMMs, and feature points, and show their performance trade-offs between compressibility and robustness on real-world LiDAR datasets: KITTI Odometry Dataset and Argoverse Tracking Dataset. Our benchmark reveals that state-of-the-art deep feature point based methods outperform traditional methods significantly when the map size budget is high. However, when map size budget is low, deep methods are outperformed by the methods using simpler models in Argoverse Tracking Dataset due to poor spatial coverage. In addition, we observe that the recently published TEASER++ significantly outperforms RANSAC for the feature point methods. Our analysis provides a valuable reference for the community to design budgeted real-world systems and find potential research opportunities. We will release the benchmark for public use.

Before the deep learning revolution, many perception algorithms were based on runtime optimization in conjunction with a strong prior/regularization penalty. A prime example of this in computer vision is optical and scene flow. Supervised learning has largely displaced the need for explicit regularization. Instead, they rely on large amounts of labeled data to capture prior statistics, which are not always readily available for many problems. Although optimization is employed to learn the neural network, at runtime, the weights of this network are frozen. As a result, these learning solutions are domain-specific and do not generalize well to other statistically different scenarios. This paper revisits the scene flow problem that relies predominantly on runtime optimization and strong regularization. A central innovation here is the inclusion of a neural scene flow prior, which utilizes the architecture of neural networks as a new type of implicit regularizer. Unlike learning-based scene flow methods, optimization occurs at runtime, and our approach needs no offline datasets---making it ideal for deployment in new environments such as autonomous driving. We show that an architecture based exclusively on multilayer perceptrons (MLPs) can be used as a scene flow prior. Our method attains competitive---if not better---results on scene flow benchmarks. Also, our neural prior's implicit and continuous scene flow representation allows us to estimate dense long-term correspondences across a sequence of point clouds. The dense motion information is represented by scene flow fields where points can be propagated through time by integrating motion vectors. We demonstrate such a capability by accumulating a sequence of lidar point clouds.

Non-rigid structure from motion (NRSfM) refers to the problem of reconstructing cameras and the 3D point cloud of a non-rigid object from an ensemble of images with 2D correspondences. Current NRSfM algorithms are limited from two perspectives: (i) the number of images, and (ii) the type of shape variability they can handle. These difficulties stem from the inherent conflict between the condition of the system and the degrees of freedom needing to be modeled - which has hampered its practical utility for many applications within vision. In this paper we propose a novel hierarchical sparse coding model for NRSFM which can overcome (i) and (ii) to such an extent, that NRSFM can be applied to problems in vision previously thought too ill posed. Our approach is realized in practice as the training of an unsupervised deep neural network (DNN) auto-encoder with a unique architecture that is able to disentangle pose from 3D structure. Using modern deep learning computational platforms allows us to solve NRSfM problems at an unprecedented scale and shape complexity. Our approach has no 3D supervision, relying solely on 2D point correspondences. Further, our approach is also able to handle missing/occluded 2D points without the need for matrix completion. Extensive experiments demonstrate the impressive performance of our approach where we exhibit superior precision and robustness against all available state-of-the-art works in some instances by an order of magnitude. We further propose a new quality measure (based on the network weights) which circumvents the need for 3D ground-truth to ascertain the confidence we have in the reconstructability. We believe our work to be a significant advance over state-of-the-art in NRSFM.

Scene flow is the three-dimensional (3D) motion field of a scene. It provides information about the spatial arrangement and rate of change of objects in dynamic environments. Current learning-based approaches seek to estimate the scene flow directly from point clouds and have achieved state-of-the-art performance. However, supervised learning methods are inherently domain specific and require a large amount of labeled data. Annotation of scene flow on real-world point clouds is expensive and challenging, and the lack of such datasets has recently sparked interest in self-supervised learning methods. How to accurately and robustly learn scene flow representations without labeled real-world data is still an open problem. Here we present a simple and interpretable objective function to recover the scene flow from point clouds. We use the graph Laplacian of a point cloud to regularize the scene flow to be “asrigid-as-possible”. Our proposed objective function can be used with or without learning-as a self-supervisory signal to learn scene flow representations, or as a non-learning-based method in which the scene flow is optimized during runtime. Our approach outperforms related works in many datasets. We also show the immediate applications of our proposed method for two applications: motion segmentation and point cloud densification.

Point clouds have grown in importance in the way computers perceive the world. From LIDAR sensors in autonomous cars and drones to the time of flight and stereo vision systems in our phones, point clouds are everywhere. Despite their ubiquity, point clouds in the real world are often missing points because of sensor limitations or occlusions, or contain extraneous points from sensor noise or artifacts. These problems challenge algorithms that require computing correspondences between a pair of point clouds. Therefore, this paper presents a fully-convolutional neural network that identifies which points in one point cloud are most similar (inliers) to the points in another. We show improvements in learning-based and classical point cloud registration approaches when retrofitted with our network. We demonstrate these improvements on synthetic and real-world datasets. Finally, our network produces impressive results on test datasets that were unseen during training, thus exhibiting generalizability. Code and videos are available at https://github.com/vinits5/masknet.

The recovery of 3D shape and pose from 2D landmarks stemming from a large ensemble of images can be viewed as a non-rigid structure from motion (NRSfM) problem. Classical NRSfM approaches, however, are problematic as they rely on heuristic priors on the 3D structure (e.g. low rank) that do not scale well to large datasets. Learning-based methods are showing the potential to reconstruct a much broader set of 3D structures than classical methods - dramatically expanding the importance of NRSfM to a temporal unsupervised 2D to 3D lifting. Hitherto, these learning approaches have not been able to effectively model perspective cameras or handle missing/occluded points - limiting their applicability to in-the-wild datasets. In this paper, we present a generalized strategy for improving learning-based NRSfM methods [32] to tackle the above issues. Our approach, Deep NRSfM++, achieves state-of-the-art performance across numerous large-scale benchmarks, outperforming both classical and learning-based 2D-3D lifting methods.

Choosing a deep neural network architecture is a fundamental problem in applications that require balancing performance and parameter efficiency. Standard approaches rely on ad-hoc engineering or computationally expensive validation on a specific dataset. We instead attempt to quantify networks by their intrinsic capacity for unique and robust representations, enabling efficient architecture comparisons without requiring any data. Building upon theoretical connections between deep learning and sparse approximation, we propose the deep frame potential: a measure of coherence that is approximately related to representation stability but has minimizers that depend only on network structure. This provides a framework for jointly quantifying the contributions of architectural hyper-parameters such as depth, width, and skip connections. We validate its use as a criterion for model selection and demonstrate correlation with generalization error on a variety of common residual and densely connected network architectures.

Dense 3D object reconstruction from a single image has recently witnessed remarkable advances, but supervising neural networks with ground-truth 3D shapes is impractical due to the laborious process of creating paired image-shape datasets. Recent efforts have turned to learning 3D reconstruction without 3D supervision from RGB images with annotated 2D silhouettes, dramatically reducing the cost and effort of annotation. These techniques, however, remain impractical as they still require multi-view annotations of the same object instance during training. As a result, most experimental efforts to date have been limited to synthetic datasets. In this paper, we address this issue and propose SDF-SRN, an approach that requires only a single view of objects at training time, offering greater utility for real-world scenarios. SDF-SRN learns implicit 3D shape representations to handle arbitrary shape topologies that may exist in the datasets. To this end, we derive a novel differentiable rendering formulation for learning signed distance functions (SDF) from 2D silhouettes. Our method outperforms the state of the art under challenging single-view supervision settings on both synthetic and real-world datasets.

What's the most accurate 3D model of your face you can obtain while sitting at your desk? We attempt to answer this question in our work. High fidelity face reconstructions have so far been limited to either studio settings or through expensive 3D scanners. On the other hand, unconstrained reconstruction methods are typically limited by low-capacity models. Our method reconstructs accurate face geometry of a subject using a video shot from a smartphone in an unconstrained environment. Our approach takes advantage of recent advances in visual SLAM, keypoint detection, and object detection to improve accuracy and robustness. By not being constrained to a model subspace, our reconstructed meshes capture important details while being robust to noise and being topologically consistent. Our evaluations show that our method outperforms current single and multi-view baselines by a significant margin, both in terms of geometric accuracy and in capturing person-specific details important for making realistic looking models.

We present a fully data-driven method to compute depth from diverse monocular video sequences that contain large amounts of non-rigid objects, e.g., people. In order to learn reconstruction cues for non-rigid scenes, we introduce a new dataset consisting of stereo videos scraped in-the-wild. This dataset has a wide variety of scene types, and features large amounts of nonrigid objects, especially people. From this, we compute disparity maps to be used as supervision to train our approach. We propose a loss function that allows us to generate a depth prediction even with unknown camera intrinsics and stereo baselines in the dataset. We validate the use of large amounts of Internet video by evaluating our method on existing video datasets with depth supervision, including SINTEL, and KITTI, and show that our approach generalizes better to natural scenes.

PointNet has revolutionized how we think about representing point clouds. For classification and segmentation tasks, the approach and its subsequent variants/extensions are considered state-of-the-art. To date, the successful application of PointNet to point cloud registration has remained elusive. In this paper we argue that PointNet itself can be thought of as a learnable "imaging" function. As a consequence, classical vision algorithms for image alignment can be brought to bear on the problem -- namely the Lucas & Kanade (LK) algorithm. Our central innovations stem from: (i) how to modify the LK algorithm to accommodate the PointNet imaging function, and (ii) unrolling PointNet and the LK algorithm into a single trainable recurrent deep neural network. We describe the architecture, and compare its performance against state-of-the-art in several common registration scenarios. The architecture offers some remarkable properties including: generalization across shape categories and computational efficiency -- opening up new paths of exploration for the application of deep learning to point cloud registration. Code and videos are available at https://github.com/hmgoforth/PointNetLK.

We present Argoverse, a dataset designed to support autonomous vehicle perception tasks including 3D tracking and motion forecasting. Argoverse includes sensor data collected by a fleet of autonomous vehicles in Pittsburgh and Miami as well as 3D tracking annotations, 300k extracted interesting vehicle trajectories, and rich semantic maps. The sensor data consists of 360 degree images from 7 cameras with overlapping fields of view, forward-facing stereo imagery, 3D point clouds from long range LiDAR, and 6-DOF pose. Our 290km of mapped lanes contain rich geometric and semantic metadata which are not currently available in any public dataset. All data is released under a Creative Commons license at Argoverse.org. In baseline experiments, we use map information such as lane direction, driveable area, and ground height to improve the accuracy of 3D object tracking. We use 3D object tracking to mine for more than 300k interesting vehicle trajectories to create a trajectory forecasting benchmark. Motion forecasting experiments ranging in complexity from classical methods (k-NN) to LSTMs demonstrate that using detailed vector maps with lane-level information substantially reduces prediction error. Our tracking and forecasting experiments represent only a superficial exploration of the potential of rich maps in robotic perception. We hope that Argoverse will enable the research community to explore these problems in greater depth.

In this paper, we address the problem of 3D object mesh reconstruction from RGB videos. Our approach combines the best of multi-view geometric and data-driven methods for 3D reconstruction by optimizing object meshes for multi-view photometric consistency while constraining mesh deformations with a shape prior. We pose this as a piecewise image alignment problem for each mesh face projection. Our approach allows us to update shape parameters from the photometric error without any depth or mask information. Moreover, we show how to avoid a degeneracy of zero photometric gradients via rasterizing from a virtual viewpoint. We demonstrate 3D object mesh reconstruction results from both synthetic and real-world videos with our photometric mesh optimization, which is unachievable with either naive mesh generation networks or traditional pipelines of surface reconstruction without heavy manual post-processing.

Current non-rigid structure from motion (NRSfM) algorithms are mainly limited with respect to: (i) the number of images, and (ii) the type of shape variability they can handle. This has hampered the practical utility of NRSfM for many applications within vision. In this paper we propose a novel deep neural network to recover camera poses and 3D points solely from an ensemble of 2D image coordinates. The proposed neural network is mathematically interpretable as a multi-layer block sparse dictionary learning problem, and can handle problems of unprecedented scale and shape complexity. Extensive experiments demonstrate the impressive performance of our approach where we exhibit superior precision and robustness against all available state-of-the-art works in the order of magnitude. We further propose a quality measure (based on the network weights) which circumvents the need for 3D ground-truth to ascertain the confidence we have in the reconstruction.

We propose to learn a 3D pose estimator by distilling knowledge from Non-Rigid Structure from Motion (NRSfM). Our method uses solely 2D landmark annotations. No 3D data, multi-view/temporal footage, or object specific prior is required. This alleviates the data bottleneck, which is one of the major concern for supervised methods. The challenge for using NRSfM as teacher is that they often make poor depth reconstruction when the 2D projections have strong ambiguity. Directly using those wrong depth as hard target would negatively impact the student. Instead, we propose a novel loss that ties depth prediction to the cost function used in NRSfM. This gives the student pose estimator freedom to reduce depth error by associating with image features. Validated on H3.6M dataset, our learned 3D pose estimation network achieves more accurate reconstruction compared to NRSfM methods. It also outperforms other weakly supervised methods, in spite of using significantly less supervision.

We present a method for localization of Unmanned Aerial Vehicles (UAVs) which is meant to replace an onboard GPS system in the event of a noisy or unreliable GPS signal. Our method requires only a downward-facing monocular RGB camera on the UAV, and pre-existing satellite imagery of the flight location to which the UAV imagery is compared and aligned. To overcome differences in the image capturing conditions between the satellite and UAV, such as seasonal and perspective changes, we propose the use of Convolutional Neural Network (CNN) representations trained on readily available satellite data. To increase localization accuracy, we also develop an optimization which jointly minimizes the error between adjacent UAV frames as well as the satellite map. We demonstrate how our method improves on recent systems from literature by achieving greater performance in flight environments with very few landmarks. For a GPS-denied flight at 0.2km altitude, over a flight distance of 0.85km, we achieve average localization error of less than 8 meters. We make our source code and datasets available to encourage further work on this emerging topic.

Conventional methods of 3D object generative modeling learn volumetric predictions using deep networks with 3D convolutional operations, which are direct analogies to classical 2D ones. However, these methods are computationally wasteful in attempt to predict 3D shapes, where information is rich only on the surfaces. In this paper, we propose a novel 3D generative modeling framework to efficiently generate object shapes in the form of dense point clouds. We use 2D convolutional operations to predict the 3D structure from multiple viewpoints and jointly apply geometric reasoning with 2D projection optimization. We introduce the pseudo-renderer, a differentiable module to approximate the true rendering operation, to synthesize novel depth maps for optimization. Experimental results for single-image 3D object reconstruction tasks show that we outperforms state-of-the-art methods in terms of shape similarity and prediction density.

We address the problem of finding realistic geometric corrections to a foreground object such that it appears natural when composited into a background image. To achieve this, we propose a novel Generative Adversarial Network (GAN) architecture that utilizes Spatial Transformer Networks (STNs) as the generator, which we call Spatial Transformer GANs (ST-GANs). ST-GANs seek image realism by operating in the geometric warp parameter space. In particular, we exploit an iterative STN warping scheme and propose a sequential training strategy that achieves better results compared to naive training of a single generator. One of the key advantages of ST-GAN is its applicability to high-resolution images indirectly since the predicted warp parameters are transferable between reference frames. We demonstrate our approach in two applications: (1) visualizing how indoor furniture (e.g. from product images) might be perceived in a room, (2) hallucinating how accessories like glasses would look when matched with real portraits.

The ability to predict depth from a single image - using recent advances in CNNs - is of increasing interest to the vision community. Unsupervised strategies to learning are particularly appealing as they can utilize much larger and varied monocular video datasets during learning without the need for ground truth depth or stereo. In previous works, separate pose and depth CNN predictors had to be determined such that their joint outputs minimized the photometric error. Inspired by recent advances in direct visual odometry (DVO), we argue that the depth CNN predictor can be learned without a pose CNN predictor. Further, we demonstrate empirically that incorporation of a differentiable implementation of DVO, along with a novel depth normalization strategy - substantially improves performance over state of the art that use monocular videos for training.

In this paper, we present a new approach for efficient regression-based object tracking. Our approach is closely related to the Generic Object Tracking Using Regression Networks (GOTURN) framework [1]. We make the following contributions. First, we demonstrate that there is a theoretical relationship between Siamese regression networks like GOTURN and the classical Inverse Compositional Lucas & Kanade (IC-LK) algorithm. Further, we demonstrate that unlike GOTURN, IC-LK adapts its regressor to the appearance of the current tracked frame. We argue that the lack of such property in GOTURN attributes to its poor performance on unseen objects and/or viewpoints. Second, we propose a novel framework for object tracking inspired by the IC-LK framework, which we refer to as Deep-LK. Finally, we show impressive results demonstrating that Deep-LK substantially outperforms GOTURN and demonstrate comparable tracking performance against current state-of-the-art deep trackers on high frame-rate sequences whilst being an order of magnitude (100 FPS) computationally efficient.

Reconstructing 3D shapes from a sequence of images has long been a problem of interest in computer vision. Classical Structure from Motion (SfM) methods have attempted to solve this problem through projected point displacement & bundle adjustment. More recently, deep methods have attempted to solve this problem by directly learning a relationship between geometry and appearance. There is, however, a significant gap between these two strategies. SfM tackles the problem from purely a geometric perspective, taking no account of the object shape prior. Modern deep methods more often throw away geometric constraints altogether, rendering the results unreliable. In this paper we make an effort to bring these two seemingly disparate strategies together. We introduce learned shape prior in the form of deep shape generators into Photometric Bundle Adjustment (PBA) and propose to accommodate full 3D shape generated by the shape prior within the optimization-based inference framework, demonstrating impressive results.

We propose a novel four-stage pipeline densely reconstructing depth from video sequences with small baselines with the goal of being fast. In our pipeline we make use of the sub-pixel precision of direct photometric bundle adjustment to reduce the number of tracked points required to estimate an accurate pose. Instead of using the exhaustive plane sweeping approach of existing small baseline methods, dense depth maps are calculated efficiently using an algorithm inspired by PatchMatch. Instead of a stereo matching error, our algorithm minimizes the variance over multiple frames with a robust mean estimation. The experiment results suggest that our method has the ability to cope with a wider range of baselines and sequence sizes. We also compare qualitative results on real small motion clips from Ha et al. in addition to our own, and show that our method outputs dense depth maps of similar or better quality and at least 10x faster.

3D reconstruction from 2D images is a central problem in computer vision. Recent works have been focusing on reconstruction directly from a single image. It is well known however that only one image cannot provide enough information for such a reconstruction. A prior knowledge that has been entertained are 3D CAD models due to its online ubiquity. A fundamental question is how to compactly represent millions of CAD models while allowing generalization to new unseen objects with fine-scaled geometry. We introduce an approach to compactly represent a 3D mesh. Our method first selects a 3D model from a graph structure by using a novel free-form deformation FFD 3D-2D registration, and then the selected 3D model is refined to best fit the image silhouette. We perform a comprehensive quantitative and qualitative analysis that demonstrates impressive dense and realistic 3D reconstruction from single images.

We investigate the problem of estimating the dense 3D shape of an object, given a set of 2D landmarks and silhouette in a single image. An obvious prior to employ in such a problem is a dictionary of dense CAD models. Employing a sufficiently large enough dictionary of CAD models, however, is in general computationally infeasible. A common strategy in dictionary learning to encourage generalization is to allow for linear combinations of dictionary elements. This too, however, is problematic as most CAD models cannot be readily placed in global dense correspondence. In this paper, we propose a two-step strategy. First, we employ orthogonal matching pursuit to rapidly choose the closest single CAD model in our dictionary to the projected image. Second, we employ a novel graph embedding based on local dense correspondence to allow for sparse linear combinations of CAD models. We validate our framework experimentally in both synthetic and real world scenario and demonstrate the superiority of our approach to both 3D mesh reconstruction and volumetric representation.

In this paper, we establish a theoretical connection between the classical Lucas & Kanade (LK) algorithm and the emerging topic of Spatial Transformer Networks (STNs). STNs are of interest to the vision and learning communities due to their natural ability to combine alignment and classification within the same theoretical framework. Inspired by the Inverse Compositional (IC) variant of the LK algorithm, we present Inverse Compositional Spatial Transformer Networks (IC-STNs). We demonstrate that IC-STNs can achieve better performance than conventional STNs with less model capacity, in particular, we show superior performance in pure image alignment tasks as well as joint alignment/classification problems on real-world problems.

In this paper, we present our method for enabling dense SDM to run at over 90 FPS on a mobile device. Our contributions are two-fold. Drawing inspiration from the FFT, we propose a Sparse Compositional Regression (SCR) framework, which enables a significant speed up over classical dense regressors. Second, we propose a binary approximation to SIFT features. Binary Approximated SIFT (BASIFT) features, which are a computationally efficient approximation to SIFT, a commonly used feature with SDM. We demonstrate the performance of our algorithm on an iPhone 7, and show that we achieve similar accuracy to SDM.

In this paper, we propose the first higher frame rate video dataset (called Need for Speed - NfS) and benchmark for visual object tracking. The dataset consists of 100 videos (380K frames) captured with now commonly available higher frame rate (240 FPS) cameras from real world scenarios. All frames are annotated with axis aligned bounding boxes and all sequences are manually labelled with nine visual attributes - such as occlusion, fast motion, background clutter, etc. Our benchmark provides an extensive evaluation of many recent and state-of-the-art trackers on higher frame rate sequences. We ranked each of these trackers according to their tracking accuracy and real-time performance. One of our surprising conclusions is that at higher frame rates, simple trackers such as correlation filters outperform complex methods based on deep networks. This suggests that for practical applications (such as in robotics or embedded vision), one needs to carefully tradeoff bandwidth constraints associated with higher frame rate acquisition, computational costs of real-time analysis, and the required application accuracy. Our dataset and benchmark allows for the first time (to our knowledge) systematic exploration of such issues, and will be made available to allow for further research in this space.

Correlation Filters (CFs) have recently demonstrated excellent performance in terms of rapidly tracking objects under challenging photometric and geometric variations. The strength of the approach comes from its ability to efficiently learn - on the fly - how the object is changing over time. A fundamental drawback to CFs, however, is that the background of the target is not modeled over time which can result in suboptimal performance. Recent tracking algorithms have suggested to resolve this drawback by either learning CFs from more discriminative deep features (e.g. DeepSRDCF [9] and CCOT [11]) or learning complex deep trackers (e.g. MDNet [28] and FCNT [33]). While such methods have been shown to work well, they suffer from high complexity: extracting deep features or applying deep tracking frameworks is very computationally expensive. This limits the real-time performance of such methods, even on high-end GPUs. This work proposes a Background-Aware CF based on hand-crafted features (HOG [6]) that can efficiently model how both the foreground and background of the object varies over time. Our approach, like conventional CFs, is extremely computationally efficient- and extensive experiments over multiple tracking benchmarks demonstrate the superior accuracy and real-time performance of our method compared to the state-of-the-art trackers.

Visual object tracking is a fundamental and time-critical vision task. Recent years have seen many shallow tracking methods based on real-time pixel-based correlation filters, as well as deep methods that have top performance but need a high-end GPU. In this paper, we learn to improve the speed of deep trackers without losing accuracy. Our fundamental insight is to take an adaptive approach, where easy frames are processed with cheap features (such as pixel values), while challenging frames are processed with invariant but expensive deep features. We formulate the adaptive tracking problem as a decision-making process, and learn an agent to decide whether to locate objects with high confidence on an early layer, or continue processing subsequent layers of a network. This significantly reduces the feedforward cost for easy frames with distinct or slow-moving objects. We train the agent offline in a reinforcement learning fashion, and further demonstrate that learning all deep layers (so as to provide good features for adaptive tracking) can lead to near real-time average tracking speed of 23 fps on a single CPU while achieving state-of-the-art performance. Perhaps most tellingly, our approach provides a 100X speedup for almost 50% of the time, indicating the power of an adaptive approach.

An emerging problem in computer vision is the reconstruction of 3D shape and pose of an object from a single image. Hitherto, the problem has been addressed through the application of canonical deep learning methods to regress from the image directly to the 3D shape and pose labels. These approaches, however, are problematic from two perspectives. First, they are minimizing the error between 3D shapes and pose labels - with little thought about the nature of this “label error” when reprojecting the shape back onto the image. Second, they rely on the onerous and ill-posed task of hand labeling natural images with respect to 3D shape and pose. In this paper we define the new task of pose-aware shape reconstruction from a single image, and we advocate that cheaper 2D annotations of objects silhouettes in natural images can be utilized. We design architectures of pose-aware shape reconstruction which reproject the predicted shape back on to the image using the predicted pose. Our evaluation on several object categories demonstrates the superiority of our method for predicting pose-aware 3D shapes from natural images.

Compressed sensing is a simple and efficient technique that has a number of applications in signal processing and machine learning. In machine learning it provides answers to questions such as: "under what conditions is the sparse representation of data efficient?", "when is learning a large margin classifier directly on the compressed domain possible?", and "why does a large margin classifier learn more effectively if the data is sparse?". This work tackles the problem of feature representation from the context of sparsity and affine rank minimization by leveraging compressed sensing from the learning perspective in order to provide answers to the aforementioned questions. We show, for a full-rank signal, the high dimensional sparse representation of data is efficient because from the classifiers viewpoint such a representation is in fact a low dimensional problem. We provide practical bounds on the linear classifier to investigate the relationship between the SVM classifier in the high dimensional and compressed domains and show for the high dimensional sparse signals, when the bounds are tight, directly learning in the compressed domain is possible.

Recent work in dense monocular 3D reconstruction relies on dense pixel correspondences and assumes brightness constancy and saliency, and thus are fundamentally unable to reconstruct low-textured or non-lambertian objects such as glass or metal. Occlusion boundaries differ from texture in that each unique view generates a unique set of occlusions. By detecting and solving for the depths of occlusion boundaries, we show how dense reconstructions of challenging objects can be integrated with existing monocular reconstruction algorithms by compensating with an increasing number of unique views. In a sequence containing the Stanford Bunny (8.5 cm) the points of our reconstructed edge cloud have an RMS error of 2.3 mm.

We present an algorithm for robust and real-time visual tracking under challenging illumination conditions characterized by poor lighting as well as sudden and drastic changes in illumination. Robustness is achieved by adapting illumination-invariant binary descriptors to dense image alignment using the Lucas and Kanade algorithm. The proposed adaptation preserves the Hamming distance under least-squares minimization, thus preserving the photometric invariance properties of binary descriptors. Due to the compactness of the descriptor, the algorithm runs in excess of 400 fps on laptops and 100 fps on mobile devices.

Inferring the motion and shape of non-rigid objects from images has been widely explored by Non-Rigid Structure from Motion (NRSfM) algorithms. Despite their promising results, they often utilize additional constraints about the camera motion (e.g. temporal order) and the deformation of the object of interest, which are not always provided in real-world scenarios. This makes the application of NRSfM limited to very few deformable objects (e.g. human face and body). In this paper, we propose the concept of Structure from Category (SfC) to reconstruct 3D structure of generic objects solely from images with no shape and motion constraint (i.e. prior-less). Similar to the NRSfM approaches, SfC involves two steps: (i) correspondence, and (ii) inversion. Correspondence determines the location of key points across images of the same object category. Once established, the inverse problem of recovering the 3D structure from the 2D points is solved over an augmented sparse shape-space model. We validate our approach experimentally by reconstructing 3D structures of both synthetic and natural images, and demonstrate the superiority of our approach to the state-of-the-art low-rank NRSfM approaches.

Many non-rigid 3D structures are not modelled well through a low-rank subspace assumption. This is problematic when it comes to their reconstruction through Structure from Motion (SfM). We argue in this paper that a more expressive and general assumption can be made around compressible 3D structures. The vision community, however, has hitherto struggled to formulate effective strategies for recovering such structures after projection without the aid of additional priors (e.g. temporal ordering, rigid substructures, etc.). In this paper we present a "prior-less" approach to solve compressible SfM. Specifically, we demonstrate how the problem of SfM - assuming compressible 3D structures - can be theoretically characterized as a block sparse dictionary learning problem. We validate our approach experimentally by demonstrating reconstructions of 3D structures that are intractable using current state-of-theart low-rank SfM approaches.

Real-time target tracking is an important service provided by embedded camera networks. The first step in target tracking is to extract the moving targets from the video frames, which can be realised by using background subtraction. For a background subtraction method to be useful in embedded camera networks, it must be both accurate and computationally efficient because of the resource constraints on embedded platforms. This makes many traditional background subtraction algorithms unsuitable for embedded platforms because they use complex statistical models to handle subtle illumination changes. These models make them accurate but the computational requirement of these complex models is often too high for embedded platforms. In this paper, we propose a new background subtraction method which is both accurate and computationally efficient. We propose a baseline version which uses luminance only and then extend it to use colour information. The key idea is to use random projection matrics to reduce the dimensionality of the data while retaining most of the information. By using multiple datasets, we show that the accuracy of our proposed background subtraction method is comparable to that of the traditional background subtraction methods. Moreover, to show the computational efficiency of our methods is not platform specific, we implement it on various platforms. The real implementation shows that our proposed method is consistently better and is up to six times faster, and consume significantly less resources than the conventional approaches. Finally, we demonstrated the feasibility of the proposed method by the implementation and evaluation of an end-to-end real-time embedded camera network target tracking application.

Correlation filters take advantage of specific properties in the Fourier domain allowing them to be estimated efficiently: O(N D log D) in the frequency domain, versus O(D3+ N D2) spatially where D is signal length, and N is the number of signals. Recent extensions to correlation filters, such as MOSSE, have reignited interest of their use in the vision community due to their robustness and attractive computational properties. In this paper we demonstrate, however, that this computational efficiency comes at a cost. Specifically, we demonstrate that only 1/D proportion of shifted examples are unaffected by boundary effects which has a dramatic effect on detection/tracking performance. In this paper, we propose a novel approach to correlation filter estimation that: (i) takes advantage of inherent computational redundancies in the frequency domain, (ii) dramatically reduces boundary effects, and (iii) is able to implicitly exploit all possible patches densely extracted from training examples during learning process. Impressive object tracking and detection results are presented in terms of both accuracy and computational efficiency.

Determining dense semantic correspondences across objects and scenes is a difficult problem that underpins many higher-level computer vision algorithms. Unlike canonical dense correspondence problems which consider images that are spatially or temporally adjacent, semantic correspondence is characterized by images that share similar high-level structures whose exact appearance and geometry may differ. Motivated by object recognition literature and recent work on rapidly estimating linear classifiers, we treat semantic correspondence as a constrained detection problem, where an exemplar LDA classifier is learned for each pixel. LDA classifiers have two distinct benefits: (i) they exhibit higher average precision than similarity metrics typically used in correspondence problems, and (ii) unlike exemplar SVM, can output globally interpretable posterior probabilities without calibration, whilst also being significantly faster to train. We pose the correspondence problem as a graphical model, where the unary potentials are computed via convolution with the set of exemplar classifiers, and the joint potentials enforce smoothly varying correspondence assignment.

Trajectory basis Non-Rigid Structure from Motion (NRSfM) refers to the process of reconstructing the 3D trajectory of each point of a non-rigid object from just their 2D projected trajectories. Reconstruction relies on two factors: (i) the condition of the composed camera & trajectory basis matrix, and (ii) whether the trajectory basis has enough degrees of freedom to model the 3D point trajectory. These two factors are inherently conflicting. Employing a trajectory basis with small capacity has the positive characteristic of reducing the likelihood of an ill-conditioned system (when composed with the camera) during reconstruction. However, this has the negative characteristic of increasing the likelihood that the basis will not be able to fully model the object's "true" 3D point trajectories. In this paper we draw upon a well known result centering around the Reduced Isometry Property (RIP) condition for sparse signal reconstruction. RIP allow us to relax the requirement that the full trajectory basis composed with the camera matrix must be well conditioned. Further, we propose a strategy for learning an over-complete basis using convolutional sparse coding from naturally occurring point trajectory corpora to increase the likelihood that the RIP condition holds for a broad class of point trajectories and camera motions. Finally, we propose an l1 inspired objective for trajectory reconstruction that is able to "adaptively" select the smallest sub-matrix from an over-complete trajectory basis that balances (i) and (ii). We present more practical 3D reconstruction results compared to current state of the art in trajectory basis NRSfM.

The task of estimating complex non-rigid 3D motion through a monocular camera is of increasing interest to the wider scientific community. Assuming one has the 2D point tracks of the non-rigid object in question, the vision community refers to this problem as Non-Rigid Structure from Motion (NRSfM). In this paper we make two contributions. First, we demonstrate empirically that the current state of the art approach to NRSfM (i.e. Dai et al. [5]) exhibits poor reconstruction performance on complex motion (i.e motions involving a sequence of primitive actions such as walk, sit and stand involving a human object). Second, we propose that this limitation can be circumvented by modeling complex motion as a union of subspaces. This does not naturally occur in Dai et al.'s approach which instead makes a less compact summation of subspaces assumption. Experiments on both synthetic and real videos illustrate the benefits of our approach for the complex nonrigid motion analysis.

In this work, we propose to employ multi-channel correlation filters for recognizing human actions (e.g. waking, riding) in videos. In our framework, each action sequence is represented as a multi-channel signal (frames) and the goal is to learn a multi-channel filter for each action class that produces a set of desired outputs when correlated with training examples. The experiments on the Weizmann and UCF sport datasets demonstrate superior computational cost (real-time), memory efficiency and very competitive performance of our approach compared to the state of the arts.

Sparse coding has become an increasingly popular method in learning and vision for a variety of classification, reconstruction and coding tasks. The canonical approach intrinsically assumes independence between observations during learning. For many natural signals however, sparse coding is applied to sub-elements ( i.e. patches) of the signal, where such an assumption is invalid. Convolutional sparse coding explicitly models local interactions through the convolution operator, however the resulting optimization problem is considerably more complex than traditional sparse coding. In this paper, we draw upon ideas from signal processing and Augmented Lagrange Methods (ALMs) to produce a fast algorithm with globally optimal sub problems and super-linear convergence.

Active Appearance Models (AAMs) employ a paradigm of inverting a synthesis model of how an object can vary in terms of shape and appearance. As a result, the ability of AAMs to register an unseen object image is intrinsically linked to two factors. First, how well the synthesis model can reconstruct the object image. Second, the degrees of freedom in the model. Fewer degrees of freedom yield a higher likelihood of good fitting performance. In this paper we look at how these seemingly contrasting factors can complement one another for the problem of AAM fitting of an ensemble of images stemming from a constrained set (e.g. an ensemble of face images of the same person).

Modern descriptors like HOG and SIFT are now commonly used in vision for pattern detection within image and video. From a signal processing perspective, this detection process can be efficiently posed as a correlation/ convolution between a multi-channel image and a multi-channel detector/filter which results in a single channel response map indicating where the pattern (e.g. object) has occurred. In this paper, we propose a novel framework for learning a multi-channel detector/filter efficiently in the frequency domain, both in terms of training time and memory footprint, which we refer to as a multichannel correlation filter. To demonstrate the effectiveness of our strategy, we evaluate it across a number of visual detection/ localization tasks where we: (i) exhibit superior performance to current state of the art correlation filters, and (ii) superior computational and memory efficiencies compared to state of the art spatial detectors.

In this paper, we propose a framework for both gradient descent image and object alignment in the Fourier domain. Our method centers upon the classical Lucas & Kanade (LK) algorithm where we represent the source and template/model in the complex 2D Fourier domain rather than in the spatial 2D domain. We refer to our approach as the Fourier LK (FLK) algorithm. The FLK formulation is advantageous when one preprocesses the source image and template/model with a bank of filters (e.g., oriented edges, Gabor, etc.) as 1) it can handle substantial illumination variations, 2) the inefficient preprocessing filter bank step can be subsumed within the FLK algorithm as a sparse diagonal weighting matrix, 3) unlike traditional LK, the computational cost is invariant to the number of filters and as a result is far more efficient, and 4) this approach can be extended to the Inverse Compositional (IC) form of the LK algorithm where nearly all steps (including Fourier transform and filter bank preprocessing) can be precomputed, leading to an extremely efficient and robust approach to gradient descent image matching. Further, these computational savings translate to nonrigid object alignment tasks that are considered extensions of the LK algorithm, such as those found in Active Appearance Models (AAMs).

Large margin learning approaches, such as support vector machines (SVM), have been successfully applied to numerous classification tasks, especially for automatic facial expression recognition. The risk of such approaches however, is their sensitivity to large margin losses due to the influence from noisy training examples and outliers which is a common problem in the area of affective computing (i.e., manual coding at the frame level is tedious so coarse labels are normally assigned). In this paper, we leverage the relaxation of the parallel-hyperplanes constraint and propose the use of modified correlation filters (MCF). The MCF is similar in spirit to SVMs and correlation filters, but with the key difference of optimizing only a single hyperplane. We demonstrate the superiority of MCF over current techniques on a battery of experiments.

Trajectory basis Non-Rigid Structure From Motion (NRSFM) currently faces two problems: the limit of reconstructability and the need to tune the basis size for different sequences. This paper provides a novel theoretical bound on 3D reconstruction error, arguing that the existing definition of reconstructability is fundamentally flawed in that it fails to consider system condition. This insight motivates a novel strategy whereby the trajectory's response to a set of high-pass filters is minimised. The new approach eliminates the need to tune the basis size and is more efficient for long sequences. Additionally, the truncated DCT basis is shown to have a dual interpretation as a high-pass filter. The success of trajectory filter reconstruction is demonstrated quantitatively on synthetic projections of real motion capture sequences and qualitatively on real image sequences.

For facial expression recognition systems to be applicable in the real world, they need to be able to detect and track a previously unseen person's face and its facial movements accurately in realistic environments. A highly plausible solution involves performing a "dense" form of alignment, where 60-70 fiducial facial points are tracked with high accuracy. The problem is that, in practice, this type of dense alignment had so far been impossible to achieve in a generic sense, mainly due to poor reliability and robustness. Instead, many expression detection methods have opted for a "coarse" form of face alignment, followed by an application of a biologically inspired appearance descriptor such as the histogram of oriented gradients or Gabor magnitudes. Encouragingly, recent advances to a number of dense alignment algorithms have demonstrated both high reliability and accuracy for unseen subjects [e.g., constrained local models (CLMs)]. This begs the question: Aside from countering against illumination variation, what do these appearance descriptors do that standard pixel representations do not? In this paper, we show that, when close to perfect alignment is obtained, there is no real benefit in employing these different appearance-based representations (under consistent illumination conditions). In fact, when misalignment does occur, we show that these appearance descriptors do work well by encoding robustness to alignment error. For this work, we compared two popular methods for dense alignment-subject-dependent active appearance models versus subject-independent CLMs-on the task of action-unit detection. These comparisons were conducted through a battery of experiments across various publicly available data sets (i.e., CK+, Pain, M3, and GEMEP-FERA). We also report our performance in the recent 2011 Facial Expression Recognition and Analysis Challenge for the subject-independent task.

This paper addresses the problem of 3D motion reconstruction from a series of 2D projections under low reconstructibility. Reconstructibility defines the accuracy of a 3D reconstruction from 2D projections given a particular trajectory basis, 3D point trajectory, and 3D camera center trajectory. Reconstructibility accuracy is inherently related to the correlation between point and camera trajectories. Poor correlation leads to good reconstruction, high correlation leads to poor reconstruction. Unfortunately, in most real-world situations involving non-rigid objects (e.g. bodies), camera and point motions are highly correlated (i.e., slow and smooth) resulting in poor reconstructibility. In this paper, we propose a novel approach for 3D motion reconstruction of non-rigid body motion in the presence of real-world camera motion. Specifically we: (i) propose the inclusion of a small number of keyframes in the video sequence from which 3D coordinates are inferred/estimated to circumvent ambiguities between point and camera motion, and (ii) employ a L1penalty term to enforce a sparsity constraint on the trajectory basis coefficients so as to ensure our reconstructions are consistent with the natural compressibility of human motion. We demonstrate impressive 3D motion reconstruction for 2D projection sequences with hitherto low reconstructibility.

In automatic facial expression detection, very accurate registration is desired which can be achieved via a deformable model approach where a dense mesh of 60-70 points on the face is used, such as an active appearance model (AAM). However, for applications where manually labeling frames is prohibitive, AAMs do not work well as they do not generalize well to unseen subjects. As such, a more coarse approach is taken for person-independent facial expression detection, where just a couple of key features (such as face and eyes) are tracked using a Viola-Jones type approach. The tracked image is normally post-processed to encode for shift and illumination invariance using a linear bank of filters. Recently, it was shown that this preprocessing step is of no benefit when close to ideal registration has been obtained. In this paper, we present a system based on the Constrained Local Model (CLM) method which is a generic or person-independent face alignment algorithm which gains high accuracy. We show these results against the LBP feature extraction on the CK+ and GEMEP-FERA datasets.

A real time facial puppetry system is presented. Compared with existing systems, the proposed method requires no special hardware, runs in real time (23 frames-per-second), and requires only a single image of the avatar and user. The user's facial expression is captured through a real-time 3D non-rigid tracking system. Expression transfer is achieved by combining a generic expression model with synthetically generated examples that better capture person specific characteristics. Performance of the system is evaluated on avatars of real people as well as masks and cartoon characters.

A real time facial puppetry system is presented. Compared with existing systems, the proposed method requires no special hardware, runs in real time (23 frames-per-second), and requires only a single image of the avatar and user. The user's facial expression is captured through a real-time 3D non-rigid tracking system. Expression transfer is achieved by combining a generic expression model with synthetically generated examples that better capture person specific characteristics. Performance of the system is evaluated on avatars of real people as well as masks and cartoon characters.

Gaining invariance to camera and illumination variations has been a well investigated topic in Active Appearance Model (AAM) fitting literature. The major problem lies in the inability of the appearance parameters of the AAM to generalize to unseen conditions. An attractive approach for gaining invariance is to fit an AAM to a multiple filter response (e.g. Gabor) representation of the input image. Naively applying this concept with a traditional AAM is computationally prohibitive, especially as the number of filter responses increase. In this paper, we present a computationally efficient AAM fitting algorithm based on the Lucas-Kanade (LK) algorithm posed in the Fourier domain that affords invariance to both expression and illumination. We refer to this as a Fourier AAM (FAAM), and show that this method gives substantial improvement in person specific AAM fitting performance over traditional AAM fitting methods.

In a clinical setting, pain is reported either through patient self-report or via an observer. Such measures are problematic as they are: 1) subjective, and 2) give no specific timing information. Coding pain as a series of facial action units (AUs) can avoid these issues as it can be used to gain an objective measure of pain on a frame-by-frame basis. Using video data from patients with shoulder injuries, in this paper, we describe an active appearance model (AAM)-based system that can automatically detect the frames in video in which a patient is in pain. This pain data set highlights the many challenges associated with spontaneous emotion detection, particularly that of expression and head movement due to the patient's reaction to pain. In this paper, we show that the AAM can deal with these movements and can achieve significant improvements in both the AU and pain detection performance compared to the current-state-of-the-art approaches which utilize similarity-normalized appearance features only.

In this paper we propose a framework for gradient descent image alignment in the Fourier domain. Specifically, we propose an extension to the classical Lucas & Kanade (LK) algorithm where we represent the source and template image's intensity pixels in the complex 2D Fourier domain rather than in the 2D spatial domain. We refer to this approach as the Fourier LK (FLK) algorithm. The FLK formulation is especially advantageous, over traditional LK, when it comes to pre-processing the source and template images with a bank of filters (e.g., Gabor filters) as: (i) it can handle substantial illumination variations, (ii) the inefficient pre-processing filter bank step can be subsumed within the FLK algorithm as a sparse diagonal weighting matrix, (iii) unlike traditional LK the computational cost is invariant to the number of filters and as a result far more efficient, (iv) this approach can be extended to the inverse compositional form of the LK algorithm where nearly all steps (including Fourier transform and filter bank pre-processing) can be pre-computed leading to an extremely efficient and robust approach to gradient descent image matching. We demonstrate robust image matching performance on a variety of objects in the presence of substantial illumination differences with exactly the same computational overhead as that of traditional inverse compositional LK during fitting.

It has been argued in recent work that correlation filters are attractive for human action recognition from videos. Motivation for their employment in this classification task lies in their ability to: (i) specify where the filter should peak in contrast to all other shifts in space and time, (ii) have some degree of tolerance to noise and intra-class variation (allowing learning from multiple examples), and (iii) can be computed deterministically with low computational overhead. Specifically, Maximum Average Correlation Height (MACH) filters have exhibited encouraging results~\cite{Mikel} on a variety of human action datasets. Here, we challenge the utility of correlation filters, like the MACH filter, in these circumstances. First, we demonstrate empirically that identical performance can be attained to the MACH filter by simply taking the~\emph{average} of the same action specific training examples. Second, we characterize theoretically and empirically under what circumstances a MACH filter would become equivalent to the average of the action specific training examples. Based on this characterization, we offer an alternative type of filter, based on a discriminative paradigm, that circumvent the inherent limitations of correlation filters for action recognition and demonstrate improved action recognition performance.

Linear filters are ubiquitously used as a preprocessing step for many classification tasks in computer vision. In particular, applying Gabor filters followed by a classification stage, such as a support vector machine (SVM), is now common practice in computer vision applications like face identity and expression recognition. A fundamental problem occurs, however, with respect to the high dimensionality of the concatenated Gabor filter responses in terms of memory requirements and computational efficiency during training and testing. In this paper, we demonstrate how the preprocessing step of applying a bank of linear filters can be reinterpreted as manipulating the type of margin being maximized within the linear SVM. This new interpretation leads to sizable memory and computational advantages with respect to existing approaches. The reinterpreted formulation turns out to be independent of the number of filters, thereby allowing the examination of the feature spaces derived from arbitrarily large number of linear filters, a hitherto untestable prospect. Further, this new interpretation of filter banks gives new insights, other than the often cited biological motivations, into why the preprocessing of images with filter banks, like Gabor filters, improves classification performance.

In this paper we pursue the task of aligning an ensemble of images in an unsupervised manner. This task has been commonly referred to as “congealing” in literature. A form of congealing, using a least-squares criteria, has been recently demonstrated to have desirable properties over conventional congealing. Least-squares congealing can be viewed as an extension of the Lucas & Kanade (LK) image alignment algorithm. It is well understood that the alignment performance for the LK algorithm, when aligning a single image with another, is theoretically and empirically equivalent for additive and compositional warps. In this paper we: (i) demonstrate that this equivalence does not hold for the extended case of congealing, (ii) characterize the inherent drawbacks associated with least-squares congealing when dealing with large numbers of images, and (iii) propose a novel method for circumventing these limitations through the application of an inverse-compositional strategy that maintains the attractive properties of the original method while being able to handle very large numbers of images.

Deformable model fitting has been actively pursued in the computer vision community for over a decade. As a result, numerous approaches have been proposed with varying degrees of success. A class of approaches that has shown substantial promise is one that makes independent predictions regarding locations of the model's landmarks, which are combined by enforcing a prior over their joint motion. A common theme in innovations to this approach is the replacement of the distribution of probable landmark locations, obtained from each local detector, with simpler parametric forms. This simplification substitutes the true objective with a smoothed version of itself, reducing sensitivity to local minima and outlying detections. In this work, a principled optimization strategy is proposed where a nonparametric representation of the landmark distributions is maximized within a hierarchy of smoothed estimates. The resulting update equations are reminiscent of mean-shift but with a subspace constraint placed on the shape's variability. This approach is shown to outperform other existing methods on the task of generic face fitting.

Local experts have been used to great effect for fitting deformable models to images. Typically, the best location in an image for the deformable model's landmarks are found through a locally exhaustive search using these experts. In order to achieve efficient fitting, these experts should afford an efficient evaluation, which often leads to forms with restricted discriminative capacity. In this work, a framework is proposed in which multiple simple experts can be utilized to increase the capacity of the detections overall. In particular, the use of a mixture of linear classifiers is proposed, the computational complexity of which scales linearly with the number of mixture components. The fitting objective is maximized using the expectation maximization (EM) algorithm, where approximations to the true objective are made in order to facilitate efficient and numerically stable fitting. The efficacy of the proposed approach is evaluated on the task of generic face fitting where performance improvement is observed over two existing methods.

Pain is typically assessed by patient self-report. Self-reported pain, however, is difficult to interpret and may be impaired or in some circumstances (i.e., young children and the severely ill) not even possible. To circumvent these problems behavioral scientists have identified reliable and valid facial indicators of pain. Hitherto, these methods have required manual measurement by highly skilled human observers. In this paper we explore an approach for automatically recognizing acute pain without the need for human observers. Specifically, our study was restricted to automatically detecting pain in adult patients with rotator cuff injuries. The system employed video input of the patients as they moved their affected and unaffected shoulder. Two types of ground truth were considered. Sequence-level ground truth consisted of Likert-type ratings by skilled observers. Frame-level ground truth was calculated from presence/absence and intensity of facial actions previously associated with pain. Active appearance models (AAM) were used to decouple shape and appearance in the digitized face images. Support vector machines (SVM) were compared for several representations from the AAM and of ground truth of varying granularity. We explored two questions pertinent to the construction, design and development of automatic pain detection systems. First, at what level (i.e., sequence- or frame-level) should datasets be labeled in order to obtain satisfactory automatic pain detection performance? Second, how important is it, at both levels of labeling, that we non-rigidly register the face?

Variation due to viewpoint is one of the key challenges that stand in the way of a complete solution to the face recognition problem. It is easy to note that local regions of the face change differently in appearance as the viewpoint varies. Recently, patch-based approaches, such as those of Kanade and Yamada, have taken advantage of this effect resulting in improved viewpoint invariant face recognition. In this paper we propose a data-driven extension to their approach, in which we not only model how a face patch varies in appearance, but also how it deforms spatially as the viewpoint varies. We propose a novel alignment strategy which we refer to as “stack flow” that discovers viewpoint induced spatial deformities undergone by a face at the patch level. One can then view the spatial deformation of a patch as the correspondence of that patch between two viewpoints. We present improved identification and verification results to demonstrate the utility of our technique.

In this paper, we present an approach we refer to as "least squares congealing" which provides a solution to the problem of aligning an ensemble of images in an unsupervised manner. Our approach circumvents many of the limitations existing in the canonical "congealing" algorithm. Specifically, we present an algorithm that:- (i) is able to simultaneously, rather than sequentially, estimate warp parameter updates, (ii) exhibits fast convergence and (iii) requires no pre-defined step size. We present alignment results which show an improvement in performance for the removal of unwanted spatial variation when compared with the related work of Learned-Miller on two datasets, the MNIST hand written digit database and the MultiPIE face database.

In this paper we demonstrate that the support vector tracking (SVT) framework first proposed by Avidan is equivalent to the canonical Lucas-Kanade (LK) algorithm with a weighted Euclidean norm. From this equivalence we empirically demonstrate that in many circumstances the canonical SVT approach is unstable, and characterize these circumstances theoretically. We then propose a novel ldquononpositive support kernel machinerdquo (NSKM) to circumvent this limitation and allow the effective use of discriminative classification within the weighted LK framework. This approach ensures that the pseudo-Hessian realized within the weighted LK algorithm is positive semidefinite which allows for fast convergence and accurate alignment/tracking. A further benefit of our proposed method is that the NSKM solution results in a much sparser kernel machine than the canonical SVM leading to sizeable computational savings and much improved alignment performance.

Constrained local models (CLMs) have recently demonstrated good performance in non-rigid object alignment/tracking in comparison to leading holistic approaches (e.g., AAMs). A major problem hindering the development of CLMs further, for non-rigid object alignment/tracking, is how to jointly optimize the global warp update across all local search responses. Previous methods have either used general purpose optimizers (e.g., simplex methods) or graph based optimization techniques. Unfortunately, problems exist with both these approaches when applied to CLMs. In this paper, we propose a new approach for optimizing the global warp update in an efficient manner by enforcing convexity at each local patch response surface. Furthermore, we show that the classic Lucas-Kanade approach to gradient descent image alignment can be viewed as a special case of our proposed framework. Finally, we demonstrate that our approach receives improved performance for the task of non-rigid face alignment/tracking on the MultiPIE database and the UNBC-McMaster archive.

Despite significant progress in deformable model fitting over the last decade, the problem of efficient and accurate person-independent face fitting remains a challenging problem. In this work, a reformulation of the generative fitting objective is presented, where only soft correspondences between the model and the image are enforced. This has the dual effect of improving robustness to unseen faces as well as affording fitting time which scales linearly with the model's complexity. This approach is compared with three state-of-the-art fitting methods on the problem of person independent face fitting, where it is shown to closely approach the accuracy of the currently best performing method while affording significant computational savings.

A common problem that affects object alignment algorithms is when they have to deal with objects with unseen intra-class appearance variation. Several variants based on gradient-decent algorithms, such as the Lucas-Kanade (or forward-additive) and inverse-compositional algorithms, have been proposed to deal with this issue by solving for both alignment and appearance simultaneously. In [1], Baker and Matthews showed that without appearance variation, the inverse-compositional (IC) algorithm was theoretically and empirically equivalent to the forward-additive (FA) algorithm, whilst achieving significant improvement in computational efficiency. With appearance variation, it would be intuitive that a similar benefit of the IC algorithm would be experienced over the FA counterpart. However, to date no such comparison has been performed. In this paper we remedy this situation by performing such a comparison. In this comparison we show that the two algorithms are not equivalent due to the inclusion of the appearance variation parameters. Through a number of experiments on the MultiPIE face database, we show that we can gain greater refinement using the FA algorithm due to it being a truer solution than the IC approach.

Non-rigid object alignment is especially challenging when only a single appearance template is available and target and template images fail to match. Two sources of discrepancy between target and template are changes in illumination and non-rigid motion. Because most existing methods rely on a holistic representation for the alignment process, they require multiple training images to capture appearance variance. We developed a patch-based method that requires only a single appearance template of the object. Specifically, we fit the patch-based face model to an unseen image using an exhaustive local search and constrain the local warp updates within a global warping space. Our approach is not limited to intensity values or gradients, and therefore offers a natural framework to integrate multiple local features, such as filter responses, to increase robustness to large initialization error, illumination changes and non-rigid deformations. This approach was evaluated experimentally on more than 100 subjects for multiple illumination conditions and facial expressions. In all the experiments, our patch-based method outperforms the holistic gradient descent method in terms of accuracy and robustness of feature alignment and image registration.

Pain is typically assessed by patient self-report. Self-reported pain, however, is difficult to interpret and may be impaired or not even possible, as in young children or the severely ill. Behavioral scientists have identified reliable and valid facial indicators of pain. Until now they required manual measurement by highly skilled observers. We developed an approach that automatically recognizes acute pain. Adult patients with rotator cuff injury were video-recorded while a physiotherapist manipulated their affected and unaffected shoulder. Skilled observers rated pain expression from the video on a 5-point Likert-type scale. From these ratings, sequences were categorized as no-pain (rating of 0), pain (rating of 3, 4, or 5), and indeterminate (rating of 1 or 2). We explored machine learning approaches for pain-no pain classification. Active Appearance Models (AAM) were used to decouple shape and appearance parameters from the digitized face images. Support vector machines (SVM) were used with several representations from the AAM. Using a leave-one-out procedure, we achieved an equal error rate of 19% (hit rate = 81%) using canonical appearance and shape features. These findings suggest the feasibility of automatic pain detection from video.

Most pose robust face verification algorithms, which employ 2D appearance, rely heavily on statistics gathered from offline databases containing ample facial appearance variation across many views. Due to the high dimensionality of the face images being employed, the validity of the assumptions employed in obtaining these statistics are essential for good performance. In this paper we assess three common approaches in 2D appearance pose mismatched face recognition literature. In our experiments we demonstrate where these approaches work and fail. As a result of this analysis, we additionally propose a new algorithm that attempts to learn the statistical dependency between gallery patches (i.e. local regions of pixels) and the whole appearance of the probe image. We demonstrate improved performance over a number of leading 2D appearance face recognition algorithms.

10.5244/C.19.73

In this paper, an in-depth analysis is undertaken into effective strategies for integrating the audio-visual speech modalities with respect to two major questions. Firstly, at what level should integration occur? Secondly, given a level of integration how should this integration be implemented? Our work is based around the well-known hidden Markov model (HMM) classifier framework for modeling speech. A novel framework for modeling the mismatch between train and test observation sets is proposed, so as to provide effective classifier combination performance between the acoustic and visual HMM classifiers. From this framework, it can be shown that strategies for combining independent classifiers, such as the weighted product or sum rules, naturally emerge depending on the influence of the mismatch. Based on the assumption that poor performance in most audio-visual speech processing applications can be attributed to train/test mismatches we propose that the main impetus of practical audio-visual integration is to dampen the independent errors, resulting from the mismatch, rather than trying to model any bimodal speech dependencies. To this end a strategy is recommended, based on theory and empirical evidence, using a hybrid between the weighted product and weighted sum rules in the presence of varying acoustic noise for the task of text-dependent speaker recognition.

Rapid speaker adaptation is becoming more important in emerging applications where storage, computation and training utterances are at a premium (e.g. PDAs, cell phones). Effective adaptation can be achieved for the task of speaker verification, based on a maximum a posteriori (MAP) learning framework, by restricting the client's parametric model to be a linear combination of parameters estimated from training observations and a speaker independent "world" model (i.e. relevance adaptation (RA)). Subspace adaptation (SA) attempts to restrict a client's parametric representation to a pre-defined subspace during estimation. In this paper we elucidate where subspace adaptation outperforms world adaptation, demonstrate where and why subspace adaptation is sometimes not as effective and give insights into what cost criteria should be used to construct the adaptation parametric subspace. Results are presented on the acoustic portion of the XM2VTS database for the task of Gaussian mixture model (GMM) based text-independent speaker verification.

The weighted product rule has been shown empirically to be of great benefit in audio-visual speech recognition (AVSR), for isolated word recognition tasks. A firm theoretical ba-sis for the selection of effective weights is of considerable interest to the audio-visual speech processing community. In this paper a clear link is established between the selec-tion of effective weightings and the approximately isotropic shrinkage that the distribution of acoustic cepstral features undergo in the presence of additive noise. An elucidation of the theoretical relationship between the cepstral shrinkage and the variance of the HMM audio log-likelihoods is then explored. 1.

Recently, the use of chromatic segmentation has come very much into vogue for mouth tracking. Our recent work has endeavored to show under what conditions and representations chromatic segmentation works. Results are presented showing that, for some members of the population, chromatic segmentation does not work satisfactorily, irrespective of recording conditions. A suitability metric is proposed that can give a quantitative measure on how well chromatic mouth tracking will work for a given subject.

Multimodal speech processing in which visual facial features are jointly processed with audio features is a rapidly advancing field. Lip movements and configurations provide useful information to improve speech and speaker recognition. However, the use of this visual information requires accurate and fast lip tracking algorithms. A new technique is outlined that is able to estimate the outer lip contour directly from a given lip intensity image via linear regression. This estimate can be improved by a active shape model that is able to track a speakers lips without requiring time consuming iterative energy minimization techniques. Results of performance are presented against known tracking algorithms using the M2VTS database.