Deep learning models deployed in real-world applications (e.g., medicine) face challenges because source models do not generalize well to domain-shifted target data. Many successful domain adaptation (DA) approaches require full access to source data. Yet, such requirements are unrealistic in scenarios where source data cannot be shared either because of privacy concerns or because it is too large and incurs prohibitive storage or computational costs. Moreover, resource constraints may limit the availability of labeled targets. We illustrate this challenge in a neuroscience setting where source data are unavailable, labeled target data are meager, and predictions involve continuous-valued outputs. We build upon Contradistinguisher (CUDA), an efficient framework that learns a shared model across the labeled source and unlabeled target samples, without intermediate representation alignment. Yet, CUDA was designed for unsupervised DA, with full access to source data, and for classification tasks. We develop CRAFT -- a Contradistinguisher-based Regularization Approach for Flexible Training -- for source-free (SF), semi-supervised transfer of pretrained models in regression tasks. We showcase the efficacy of CRAFT in two neuroscience settings: gaze prediction with electroencephalography (EEG) data and ``brain age'' prediction with structural MRI data. For both datasets, CRAFT yielded up to 9% improvement in root-mean-squared error (RMSE) over fine-tuned models when labeled training examples were scarce. Moreover, CRAFT leveraged unlabeled target data and outperformed four competing state-of-the-art source-free domain adaptation models by more than 3%. Lastly, we demonstrate the efficacy of CRAFT on two other real-world regression benchmarks. We propose CRAFT as an efficient approach for source-free, semi-supervised deep transfer for regression that is ubiquitous in biology and medicine.

We propose a fast and robust 2D-affine global motion estimation algorithm based on phase-correlation in the Fourier-Mellin domain and robust least square model fitting of sparse motion vector field and its application for digital image stabilization. Rotation-scale-translation (RST) approximation of affine parameters is obtained at the coarsest level of the image pyramid, thus ensuring convergence for a much larger range of motions. Despite working at the coarsest resolution level, using subpixel-accurate phase correlation provides sufficiently accurate coarse estimates for the subsequent refinement stage of the algorithm. The refinement stage consists of RANSAC based robust least-square model fitting for sparse motion vector field, estimated using block-based subpixel-accurate phase correlation at randomly selected high activity regions in finest level of image pyramid. Resulting algorithm is very robust to outliers such as foreground objects and flat regions. We investigate the robustness of the proposed method for digital image stabilization application. Experimental results show that the proposed algorithm is capable of estimating larger range of motions as compared to another phase correlation method and optical flow algorithm.

With the advent of high frame rate driving in Liquid Crystal Displays (LCD), the need for high quality frame rate up-conversion has become very important. Various motion-compensated frame rate up-conversion (MCFRC) algorithms and have been developed to interpolate moving objects along their motion trajectories. In addition to robust motion estimation, accurate occlusion handling is also very essential for high quality interpolated video frames. Incorrect handling of occlusion can lead to artifacts that manifest itself as halo around moving objects and can be extremely objectionable. Many occlusion handling algorithms have been proposed in the literature over the years. Use of cascaded median filters was proposed in to handle occlusion. Intensity mismatch and the number of projected vectors in neighboring blocks were also used to detect occlusion. In this paper, we present a motion compensated video upconversion algorithm that produces occlusion free video. The algorithm is categorized in two steps. In the first stage occlusion area is detected using motion vector clustering and segmentation and in the next stage occlusion region is handled using motion vector steered video in-painting. This algorithm does not introduce one additional frame latency that is required for occlusion handling by other methods such as.

In this paper, we propose a compression artifact reduction algorithm based on ν support vector regression. It belongs to the broad family of regularized reconstruction methods but regularization model is learned from a set of training samples of original images and corresponding noise corrupted version. As opposed to artifact reduction methods specific to each type of compression artifact (e.g. blocking, ringing etc), we treat such different artifacts as symptoms of the same problem, quantization of DCT coefficients. In the testing step, algorithm tries to undo the effect of quantization using information (relationship between original and artifact-corrupted image) learned during the training step. Experimental results exhibit significant reduction in all types of compression artifacts.

A lot of research has been conducted on motion-compensated (MC) de-interlacing, but there are very few publications that discuss the performances of de-interlacing quantatively. The various methods are compared through their performance on known video sequences. Linear system analysis of interlaced video and de-interlacer are proposed in. It is well established that the performance of the MC methods outperform the fixed or motion-adaptive methods when the motion vectors used are reliable and true to the scene content. Being an open-loop process the performance of the MC de-interlacers degrade drastically when there are motion vector errors. In this paper, a linear system analysis of MC video upconversion systems is presented and the effects of motion vector accuracy on system performance are analyzed. We investigate the various factors that contribute to the motion vector inaccuracy, such as incorrect motion modelling, acceleration between the frames, and insufficient interpolation kernel.

In this paper we investigate the application of texture synthesis and image inpainting techniques for video applications. Working in the non-parametric framework, we use 3D patches for matching and copying. This ensures temporal continuity to some extent which is not possible to obtain by working with individual frames. Since, in present application, patches might contain arbitrary shaped and multiple disconnected holes, fast Fourier transform (FFT) and summed area table based sum of squared difference (SSD) calculation (S.L. Kilthau, et al, 2002) cannot be used. We propose a modification of above scheme which allows its use in present application. This results in significant gain of efficiency since search space is typically huge for video applications.

The majority of the work on deinterlacing techniques is ad hoc in nature and lacks theoretical justification. The designs are heuristic and the results are in most cases experimental. It is very difficult to quantify the overall performance of a deinterlacing system based on its performance over a few video sequences. In this paper, we present a linear system analysis for a motion-compensated video upconversion system, analyze the effect of motion vector accuracy in system performance and investigate switching the algorithm between fixed and motion-compensated deinterlacing.

We propose a fast and robust global motion estimation (GME) algorithm, based on a 2-stage coarse-to-fine refinement strategy, which is capable of measuring large motions. A 6-parameter affine motion model has been used. Coarse estimation is carried out in the frequency domain using polar, log-polar or log-log sampling of the Fourier magnitude spectrum of the sub-sampled image. The sampling scheme is adaptively selected, based on past motion patterns. The refinement stage consists of a RANSAC based model fitting to motion vectors of randomly selected high-activity blocks, and hence is robust to outliers. The motion vector of blocks is measured using phase correlation, which offers two advantages in this context: sub-pixel accuracy without significant computational overhead; and if a particular block consists of background as well as foreground pixels, both motions are simultaneously measured. Due to its hardware-friendly nature, the proposed algorithm holds potential for real-time GME even for television images.

A DCT-based phase correlation motion estimation algorithm is proposed in this paper. By combining four real transforms, a new complex linear phase transform is obtained and is used in phase correlation motion estimation. The application in video compression is discussed in detail. The simulation results show that the proposed algorithm is robust and efficient.