Parameter-Efficient fine-tuning (PEFT) has attracted significant attention due to the growth of pre-trained model sizes and the need to fine-tune (FT) them for superior downstream performance. Despite a surge in new PEFT methods, a systematic study to understand their performance and suitable application scenarios is lacking, leaving questions like "when to apply PEFT" and "which method to use" largely unanswered, especially in visual recognition. In this paper, we conduct a unifying empirical study of representative PEFT methods with Vision Transformers. We systematically tune their hyperparameters to fairly compare their accuracy on downstream tasks. Our study offers a practical user guide and unveils several new insights. First, if tuned carefully, different PEFT methods achieve similar accuracy in the low-shot benchmark VTAB-1K. This includes simple approaches like FT the bias terms that were reported inferior. Second, despite similar accuracy, we find that PEFT methods make different mistakes and high-confidence predictions, likely due to their different inductive biases. Such an inconsistency (or complementarity) opens up the opportunity for ensemble methods, and we make preliminary attempts at this. Third, going beyond the commonly used low-shot tasks, we find that PEFT is also useful in many-shot regimes, achieving comparable or better accuracy than full FT while using significantly fewer parameters. Lastly, we investigate PEFT’s ability to preserve a pre-trained model’s robustness to distribution shifts (e.g., CLIP). Perhaps not surprisingly, PEFT approaches outperform full FT alone. However, with weight-space ensembles, full FT can better balance target distribution and distribution shift performance, suggesting a future research direction for robust PEFT1.

We study the problem of how to train a “personalization-friendly” model such that given only the task descriptions, the model can be adapted to different end-users' needs, e.g., for accurately classifying different subsets of objects. One baseline approach is to train a “generic” model for classifying a wide range of objects, followed by class selection. In our experiments, we however found it suboptimal, perhaps because the model's weights are kept frozen without being personalized. To address this drawback, we propose Train-once-for-All PERsonalization (TAPER), a framework that is trained just once and can later customize a model for different end-users given their task descriptions. TAPER learns a set of “basis” models and a mixer predictor, such that given the task description, the weights (not the predictions!) of the basis models can be on the fly combined into a single “personalized” model. Via extensive experiments on multiple recognition tasks, we show that TAPER consistently outperforms the baseline methods in achieving a higher personalized accuracy. Moreover, we show that TAPER can synthesize a much smaller model to achieve comparable performance to a huge generic model, making it “deployment-friendly” to resource-limited end devices. Interestingly, even without end-users' task descriptions, TAPER can still be specialized to the deployed context based on its past predictions, making it even more “personalization-friendly”.

Data augmentation has become a de facto component for training high-performance deep image classifiers, but its potential is under-explored for object detection. Noting that most state-of-the-art object detectors benefit from fine-tuning a pre-trained classifier, we first study how the classifiers’ gains from various data augmentations transfer to object detection. The results are discouraging; the gains diminish after fine-tuning in terms of either accuracy or robustness. This work instead augments the fine-tuning stage for object detectors by exploring adversarial examples, which can be viewed as a model-dependent data augmentation. Our method dynamically selects the stronger adversarial images sourced from a detector’s classification and localization branches and evolves with the detector to ensure the augmentation policy stays current and relevant. This model-dependent augmentation generalizes to different object detectors better than AutoAugment, a model-agnostic augmentation policy searched based on one particular detector. Our approach boosts the performance of state-of-the-art EfficientDets by +1.1 mAP on the COCO object detection benchmark. It also improves the detectors’ robustness against natural distortions by +3.8 mAP and against domain shift by +1.3 mAP.

We present Mobile Video Networks (MoViNets), a family of computation and memory efficient video networks that can operate on streaming video for online inference. 3D convolutional neural networks (CNNs) are accurate at video recognition but require large computation and memory budgets and do not support online inference, making them difficult to work on mobile devices. We propose a three-step approach to improve computational efficiency while substantially reducing the peak memory usage of 3D CNNs. First, we design a video network search space and employ neural architecture search to generate efficient and diverse 3D CNN architectures. Second, we introduce the Stream Buffer technique that decouples memory from video clip duration, allowing 3D CNNs to embed arbitrary-length streaming video sequences for both training and inference with a small constant memory footprint. Third, we propose a simple ensembling technique to improve accuracy further without sacrificing efficiency. These three progressive techniques allow MoViNets to achieve state-of-the-art accuracy and efficiency on the Kinetics, Moments in Time, and Charades video action recognition datasets. For instance, MoViNet-A5-Stream achieves the same accuracy as X3D-XL on Kinetics 600 while requiring 80% fewer FLOPs and 65% less memory. Code is available at https://github.com/google-research/movinet.

This paper proposes a deep learning architecture based on Residual Network that dynamically adjusts the number of executed layers for the regions of the image. This architecture is end-to-end trainable, deterministic and problem-agnostic. It is therefore applicable without any modifications to a wide range of computer vision problems such as image classification, object detection and image segmentation. We present experimental results showing that this model improves the computational efficiency of Residual Networks on the challenging ImageNet classification and COCO object detection datasets. Additionally, we evaluate the computation time maps on the visual saliency dataset cat2000 and find that they correlate surprisingly well with human eye fixation positions.

We propose a data-driven approach to facial landmark localization that models the correlations between each landmark and its surrounding appearance features. At runtime, each feature casts a weighted vote to predict landmark locations, where the weight is precomputed to take into account the feature's discriminative power. The feature voting-based landmark detection is more robust than previous local appearance-based detectors, we combine it with nonparametric shape regularization to build a novel facial landmark localization pipeline that is robust to scale, in-plane rotation, occlusion, expression, and most importantly, extreme head pose. We achieve state-of-the-art performance on two especially challenging in-the-wild datasets populated by faces with extreme head pose and expression.

In this work, we propose an exemplar-based face image segmentation algorithm. We take inspiration from previous works on image parsing for general scenes. Our approach assumes a database of exemplar face images, each of which is associated with a hand-labeled segmentation map. Given a test image, our algorithm first selects a subset of exemplar images from the database, Our algorithm then computes a nonrigid warp for each exemplar image to align it with the test image. Finally, we propagate labels from the exemplar images to the test image in a pixel-wise manner, using trained weights to modulate and combine label maps from different exemplars. We evaluate our method on two challenging datasets and compare with two face parsing algorithms and a general scene parsing algorithm. We also compare our segmentation results with contour-based face alignment results, that is, we first run the alignment algorithms to extract contour points and then derive segments from the contours. Our algorithm compares favorably with all previous works on all datasets evaluated.

We propose a novel method for capturing high-speed, high dynamic range video with a single low-speed camera using a coded sampling technique. Traditional video cameras use a constant full-frame exposure time, which makes temporal super-resolution difficult due to the ill-posed nature of inverting the sampling operation. Our method samples at the same rate as the traditional low-speed camera but uses random per-pixel exposure times and offsets. By exploiting temporal and spatial redundancy in the video, we can reconstruct a high-speed video from the coded input. Furthermore, the different exposure times used in our sampling scheme enable us to obtain a higher dynamic range than a traditional camera or other temporal superresolution methods. We validate our approach using simulation and provide a detailed discussion on how to make a hardware implementation. In particular, we believe that our approach can maintain a 100% light throughput similarly to existing cameras and can be implemented on a single chip, making it suitable for small form factors.

This paper extends the classical warping-based optical flow method to achieve accurate flow in the presence of spatially-varying motion blur. Our idea is to parameterize the appearance of each frame as a function of both the pixel motion and the motion-induced blur. We search for the flows that best match two consecutive frames, which amounts to finding the derivative of a blurred frame with respect to both the motion and the blur, where the blur itself is a function of the motion. We propose an efficient technique to calculate the derivatives using prefiltering. Our technique avoids performing spatially-varying filtering (which can be computationally expensive) during the optimization iterations. In the end, our derivative calculation technique can be easily incorporated with classical flow code to handle video with non-uniform motion blur with little performance penalty. Our method is evaluated on both synthetic and real videos and outperforms conventional flow methods in the presence of motion blur.

Current face image retrieval methods achieve impressive results, but lack efficient ways to refine the search, particularly for geometric face attributes. Users cannot easily find faces with slightly more furrowed brows or specific leftward pose shifts, for example. To address this problem, we propose a new face search technique based on shape manipulation that is complementary to current search engines. Users drag one or a small number of contour points, like the bottom of the chin or the corner of an eyebrow, to search for faces similar in shape to the current face, but with updated geometric attributes specific to their edits. For example, the user can drag a mouth corner to find faces with wider smiles, or the tip of the nose to find faces with a specific pose. As part of our system, we propose (1) a novel confidence score for face alignment results that automatically constructs a contour-aligned face database with reasonable alignment accuracy, (2) a simple and straightforward extension of PCA with missing data to tensor analysis, and (3) a new regularized tensor model to compute shape feature vectors for each aligned face, all built upon previous work. To the best of our knowledge, our system demonstrates the first face retrieval approach based chiefly on shape manipulation. We show compelling results on a sizable database of over 10,000 face images captured in uncontrolled environments.

New cameras such as the Canon EOS 7D and Pointgrey Grasshopper have 14-bit sensors. We present a theoretical analysis and a practical approach that exploit these new cameras with high-resolution quantization for reliable HDR imaging from a moving camera. Specifically, we propose a unified probabilistic formulation that allows us to analytically compare two HDR imaging alternatives: (1) deblurring a single blurry but clean image and (2) denoising a sequence of sharp but noisy images. By analyzing the uncertainty in the estimation of the HDR image, we conclude that multi-image denoising offers a more reliable solution. Our theoretical analysis assumes translational motion and spatially-invariant blur. For practice, we propose an approach that combines optical flow and image denoising algorithms for HDR imaging, which enables capturing sharp HDR images using handheld cameras for complex scenes with large depth variation. Quantitative evaluation on both synthetic and real images is presented.

In this paper, we present a new framework for non-rigid structure from motion (NRSFM) that simultaneously addresses three significant challenges: severe occlusion, perspective camera projection, and large non-linear deformation. We introduce a concept called a model graph, which greatly reduces the computational cost of discovering groups of input images that depict consistent 3D shapes. A 3D model is constructed for each input image by traversing the model graph along multiple evolutionary paths. A compressive shape representation is constructed, which (1) consolidates the multiple 3D models for each image reconstructed during model evolution and (2) reduces the number of models needed to represent the input image set. Assuming feature correspondences are known, we demonstrate our algorithm on both real and synthetic data sets that exemplify all three aforementioned challenges.

Most active scene recovery techniques assume that a scene point is illuminated only directly by the illumination source. Consequently, global illumination effects due to inter-reflections, sub-surface scattering and volumetric scattering introduce strong biases in the recovered scene shape. Our goal is to recover scene properties in the presence of global illumination. To this end, we study the interplay between global illumination and the depth cue of illumination defocus. By expressing both these effects as low pass filters, we derive an approximate invariant that can be used to separate them without explicitly modeling the light transport. This is directly useful in any scenario where limited depth-of-field devices (such as projectors) are used to illuminate scenes with global light transport and significant depth variations. We show two applications: (a) accurate depth recovery in the presence of global illumination, and (b) factoring out the effects of defocus for correct direct-global separation in large depth scenes. We demonstrate our approach using scenes with complex shapes, reflectances, textures and translucencies.

We propose a novel formulation of stereo matching that considers each pixel as a feature vector. Under this view, matching two or more images can be cast as matching point clouds in feature space. We build a nonparametric depth smoothness model in this space that correlates the image features and depth values. This model induces a sparse graph that links pixels with similar features, thereby converting each point cloud into a connected network. This network defines a neighborhood system that captures pixel grouping hierarchies without resorting to image segmentation. We formulate global stereo matching over this neighborhood system and use graph cuts to match pixels between two or more such networks. We show that our stereo formulation is able to recover surfaces with different orders of smoothness, such as those with high-curvature details and sharp discontinuities. Furthermore, compared to other single-frame stereo methods, our method produces more temporally stable results from videos of dynamic scenes, even when applied to each frame independently.

We present a novel multi-view denoising algorithm. Our algorithm takes noisy images taken from different viewpoints as input and groups similar patches in the input images using depth estimation. We model intensity-dependent noise in low-light conditions and use the principal component analysis and tensor analysis to remove such noise. The dimensionalities for both PCA and tensor analysis are automatically computed in a way that is adaptive to the complexity of image structures in the patches. Our method is based on a probabilistic formulation that marginalizes depth maps as hidden variables and therefore does not require perfect depth estimation. We validate our algorithm on both synthetic and real images with different content. Our algorithm compares favorably against several state-of-the-art denoising algorithms.

We describe a method for producing a smooth, stabilized video from the shaky input of a hand-held light field video camera—specifically, a small camera array. Traditional stabilization techniques dampen shake with 2D warps, and thus have limited ability to stabilize a significantly shaky camera motion through a 3D scene. Other recent stabilization techniques synthesize novel views as they would have been seen along a virtual, smooth 3D camera path, but are limited to static scenes. We show that video camera arrays enable much more powerful video stabilization, since they allow changes in viewpoint for a single time instant. Furthermore, we point out that the straightforward approach to light field video stabilization requires computing structure-from-motion, which can be brittle for typical consumer-level video of general dynamic scenes. We present a more robust approach that avoids input camera path reconstruction. Instead, we employ a spacetime optimization that directly computes a sequence of relative poses between the virtual camera and the camera array, while minimizing acceleration of salient visual features in the virtual image plane. We validate our novel method by comparing it to state-of-the-art stabilization software, such as Apple iMovie and 2d3 SteadyMove Pro, on a number of challenging scenes.

In this paper, we propose a novel imaging system that enables the capture of photos and videos with sparse informational pixels. Our system is based on the projection and detection of 3D optical tags. We use an infrared (IR) projector to project temporally-coded (blinking) dots onto selected points in a scene. These tags are invisible to the human eye, but appear as clearly visible time-varying codes to an IR photosensor. As a proof of concept, we have built a prototype camera system (consisting of co-located visible and IR sensors) to simultaneously capture visible and IR images. When a user takes an image of a tagged scene using such a camera system, all the scene tags that are visible from the system's viewpoint are detected. In addition, tags that lie in the field of view but are occluded, and ones that lie just outside the field of view, are also automatically generated for the image. Associated with each tagged pixel is its 3D location and the identity of the object that the tag falls on. Our system can interface with conventional image recognition methods for efficient scene authoring, enabling objects in an image to be robustly identified using cheap cameras, minimal computations, and no domain knowledge. We demonstrate several applications of our system, including, photo-browsing, e-commerce, augmented reality, and objection localization.

This paper presents a novel approach for estimating the parameters for MRF-based stereo algorithms. This approach is based on a new formulation of stereo as a maximum a posterior (MAP) problem in which both a disparity map and MRF parameters are estimated from the stereo pair itself. We present an iterative algorithm for the MAP estimation that alternates between estimating the parameters while fixing the disparity map and estimating the disparity map while fixing the parameters. The estimated parameters include robust truncation thresholds for both data and neighborhood terms, as well as a regularization weight. The regularization weight can be either a constant for the whole image or spatially-varying, depending on local intensity gradients. In the latter case, the weights for intensity gradients are also estimated. Our approach works as a wrapper for existing stereo algorithms based on graph cuts or belief propagation, automatically tuning their parameters to improve performance without requiring the stereo code to be modified. Experiments demonstrate that our approach moves a baseline belief propagation stereo algorithm up six slots in the Middlebury rankings.

In order to produce bright images, projectors have large apertures and hence narrow depths of field. In this paper, we present methods for robust scene capture and enhanced image display based on projection defocus analysis. We model a projector's defocus using a linear system. This model is used to develop a novel temporal defocus analysis method to recover depth at each camera pixel by estimating the parameters of its projection defocus kemel in frequency domain. Compared to most depth recovery methods, our approach is more accurate near depth discontinuities. Furthermore, by using a coaxial projector-camera system, we ensure that depth is computed at all camera pixels, without any missing parts. We show that the recovered scene geometry can be used for refocus synthesis and for depth-based image composition. Using the same projector defocus model and estimation technique, we also propose a defocus compensation method that filters a projection image in a spatially-varying, depth-dependent manner to minimize its defocus blur after it is projected onto the scene. This method effectively increases the depth of field of a projector without modifying its optics. Finally, we present an algorithm that exploits projector defocus to reduce the strong pixelation artifacts produced by digital projectors, while preserving the quality of the projected image. We have experimentally verified each of our methods using real scenes.

This paper presents a novel approach for estimating parameters for MRF-based stereo algorithms. This approach is based on a new formulation of stereo as a maximum a posterior (MAP) problem, in which both a disparity map and MRF parameters are estimated from the stereo pair itself. We present an iterative algorithm for the MAP estimation that alternates between estimating the parameters while fixing the disparity map and estimating the disparity map while fixing the parameters. The estimated parameters include robust truncation thresholds, for both data and neighborhood terms, as well as a regularization weight. The regularization weight can be either a constant for the whole image, or spatially-varying, depending on local intensity gradients. In the latter case, the weights for intensity gradients are also estimated. Experiments indicate that our approach, as a wrapper for existing stereo algorithms, moves a baseline belief propagation stereo algorithm up six slots in the middlebury rankings.

We present an end-to-end system that goes from video sequences to high resolution, editable, dynamically controllable face models. The capture system employs synchronized video cameras and structured light projectors to record videos of a moving face from multiple viewpoints. A novel spacetime stereo algorithm is introduced to compute depth maps accurately and overcome over-fitting deficiencies in prior work. A new template fitting and tracking procedure fills in missing data and yields point correspondence across the entire sequence without using markers. We demonstrate a data-driven, interactive method for inverse kinematics that draws on the large set of fitted templates and allows for posing new expressions by dragging surface points directly. Finally, we describe new tools that model the dynamics in the input sequence to enable new animations, created via key-framing or texture-synthesis techniques.

This paper extends the traditional binocular stereo problem into the spacetime domain, in which a pair of video streams is matched simultaneously instead of matching pairs of images frame by frame. Almost any existing stereo algorithm may be extended in this manner simply by replacing the image matching term with a spacetime term. By utilizing both spatial and temporal appearance variation, this modification reduces ambiguity and increases accuracy. Three major applications for spacetime stereo are proposed in this paper. First, spacetime stereo serves as a general framework for structured light scanning and generates high quality depth maps for static scenes. Second, spacetime stereo is effective for a class of natural scenes, such as waving trees and flowing water, which have repetitive textures and chaotic behaviors and are challenging for existing stereo algorithms. Third, the approach is one of very few existing methods that can robustly reconstruct objects that are moving and deforming over time, achieved by use of oriented spacetime windows in the matching procedure. Promising experimental results in the above three scenarios are demonstrated.

We present an algorithm for computing optical flow, shape, motion, lighting, and albedo from an image sequence of a rigidly-moving Lambertian object under distant illumination. The problem is formulated in a manner that subsumes structure from motion, multiview stereo, and photometric stereo as special cases. The algorithm utilizes both spatial and temporal intensity variation as cues: the former constrains flow and the latter constrains surface orientation; combining both cues enables dense reconstruction of both textured and textureless surfaces. The algorithm works by iteratively estimating affine camera parameters, illumination, shape, and albedo in an alternating fashion. Results are demonstrated on videos of hand-held objects moving in front of a fixed light and camera.

Abstract This paper presents a novel approach for reconstructing free‐form, texture‐mapped, 3D scene models from a single painting or photograph. Given a sparse set of user‐specified constraints on the local shape of the scene, a smooth 3D surface that satisfies the constraints is generated. This problem is formulated as a constrained variational optimization problem. In contrast to previous work in single‐view reconstruction, our technique enables high‐quality reconstructions of free‐form curved surfaces with arbitrary reflectance properties. A key feature of the approach is a novel hierarchical transformation technique for accelerating convergence on a non‐uniform, piecewise continuous grid. The technique is interactive and updates the model in real time as constraints are added, allowing fast reconstruction of photorealistic scene models. The approach is shown to yield high‐quality results on a large variety of images. Copyright © 2002 John Wiley & Sons, Ltd.

This paper presents a novel approach for reconstructing free-form, texture-mapped, 3D scene models from a single painting or photograph. Given a sparse set of user-specified constraints on the local shape of the scene, a smooth 3D surface that satisfies the constraints is generated This problem is formulated as a constrained variational optimization problem. In contrast to previous work in single view reconstruction, our technique enables high quality reconstructions of free-form curved surfaces with arbitrary reflectance properties. A key feature of the approach is a novel hierarchical transformation technique for accelerating convergence on a non-uniform, piecewise continuous grid. The technique is interactive and updates the model in real time as constraints are added, allowing fast reconstruction of photorealistic scene models. The approach is shown to yield high quality results on a large variety of images.