Search

Privacy is a crucial concern in collaborative machine vision where a part of a Deep Neural Network (DNN) model runs on the edge, and the rest is executed on the cloud. In such applications, the machine vision model does not need the exact visual content to perform its task. Taking advantage of this potential, private information could be removed from the data insofar as it does not significantly impair the accuracy of the machine vision system. In this paper, we present an autoencoder-style network integrated within an object detection pipeline, which generates a latent representation of the input image that preserves task-relevant information while removing private information. Our approach employs an adversarial training strategy that not only removes private information from the bottleneck of the autoencoder but also promotes improved compression efficiency for feature channels coded by conventional codecs like VVC-Intra. We assess the proposed system using a realistic evaluation framework for privacy, directly measuring face and license plate recognition accuracy. Experimental results show that our proposed method is able to reduce the bitrate significantly at the same object detection accuracy compared to coding the input images directly, while keeping the face and license plate recognition accuracy on the images recovered from the bottleneck features low, implying strong privacy protection. Our code is available at https://github.com/bardia-az/ppa-code.

At present, and increasingly so in the future, much of the captured visual content will not be seen by humans. Instead, it will be used for automated machine vision analytics and may require occasional human viewing. Examples of such applications include traffic monitoring, visual surveillance, autonomous navigation, and industrial machine vision. To address such requirements, we develop an end-to-end learned image codec whose latent space is designed to support scalability from simpler to more complicated tasks. The simplest task is assigned to a subset of the latent space (the base layer), while more complicated tasks make use of additional subsets of the latent space, i.e., both the base and enhancement layer(s). For the experiments, we establish a 2-layer and a 3-layer model, each of which offers input reconstruction for human vision, plus machine vision task(s), and compare them with relevant benchmarks. The experiments show that our scalable codecs offer 37%-80% bitrate savings on machine vision tasks compared to best alternatives, while being comparable to state-of-the-art image codecs in terms of input reconstruction.

Point cloud (PC) is a collection of discrete geometric samples of a physical object in 3D space. A PC video consists of temporal frames evenly spaced in time, each containing a static PC at one time instant. PCs in adjacent frames typically do not have point-to-point (P2P) correspondence, and thus exploiting temporal redundancy for PC restoration across frames is difficult. In this paper, we focus on the super-resolution (SR) problem for PC video: increase point density of PCs in video frames while preserving salient geometric features consistently across time. We accomplish this with two ideas. First, we establish partial P2P coupling between PCs of adjacent frames by interpolating interior points in a low-resolution PC patch in frame t and translating them to a corresponding patch in frame t+1 , via a motion model computed by iterative closest point (ICP). Second, we promote piecewise smoothness in 3D geometry in each patch using feature graph Laplacian regularizer (FGLR) in an easily computable quadratic form. The two ideas translate to an unconstrained quadratic programming (QP) problem with a system of linear equations as solution-one where we ensure the numerical stability by upper-bounding the condition number of the coefficient matrix. Finally, to improve the accuracy of the ICP motion model, we re-sample points in a super-resolved patch at time t to better match a low-resolution patch at time t+1 via bipartite graph matching after each SR iteration. Experimental results show temporally consistent super-resolved PC videos generated by our scheme, outperforming SR competitors that optimized on a per-frame basis, in two established PC metrics.

In recent studies, collaborative intelligence (CI) has emerged as a promising framework for deployment of Artificial Intelligence (AI)-based services on mobile/edge devices. In CI, the AI model (a deep neural network) is split between the edge and the cloud, and intermediate features are sent from the edge sub-model to the cloud sub-model. In this article, we study bit allocation for feature coding in multi-stream CI systems. We model task distortion as a function of rate using convex surfaces similar to those found in distortion-rate theory. Using such models, we are able to provide closed-form bit allocation solutions for single-task systems and scalarized multi-task systems. Moreover, we provide analytical characterization of the full Pareto set for 2-stream k -task systems, and bounds on the Pareto set for 3-stream 2-task systems. Analytical results are examined on a variety of DNN models from the literature to demonstrate wide applicability of the results.

We propose a neural network model to estimate the current frame from two reference frames, using affine transformation and adaptive spatially-varying filters. The estimated affine transformation allows for using shorter filters compared to existing approaches for deep frame prediction. The predicted frame is used as a reference for coding the current frame. Since the proposed model is available at both encoder and decoder, there is no need to code or transmit motion information for the predicted frame. By making use of dilated convolutions and reduced filter length, our model is significantly smaller, yet more accurate, than any of the neural networks in prior works on this topic. Two versions of the proposed model - one for uni-directional, and one for bi-directional prediction - are trained using a combination of Discrete Cosine Transform (DCT)-based l 1 -loss with various transform sizes, multi-scale Mean Squared Error (MSE) loss, and an object context reconstruction loss. The trained models are integrated with the HEVC video coding pipeline. The experiments show that the proposed models achieve about 7.3%, 5.4%, and 4.2% bit savings for the luminance component on average in the Low delay P, Low delay, and Random access configurations, respectively.

Point cloud is a collection of 3D coordinates that are discrete geometric samples of an object's 2D surfaces. Imperfection in the acquisition process means that point clouds are often corrupted with noise. Building on recent advances in graph signal processing, we design local algorithms for 3D point cloud denoising. Specifically, we design a signal-dependent feature graph Laplacian regularizer (SDFGLR) that assumes surface normals computed from point coordinates are piecewise smooth with respect to a signal-dependent graph Laplacian matrix. Using SDFGLR as a signal prior, we formulate an optimization problem with a general 'p-norm fidelity term that can explicitly remove only two types of additive noise: small but non-sparse noise like Gaussian (using '2 fidelity term) and large but sparser noise like Laplacian (using '1 fidelity term). To establish a linear relationship between normals and 3D point coordinates, we first perform bipartite graph approximation to divide the point cloud into two disjoint node sets (red and blue). We then optimize the red and blue nodes' coordinates alternately. For '2-norm fidelity term, we iteratively solve an unconstrained quadratic programming (QP) problem, efficiently computed using conjugate gradient with a bounded condition number to ensure numerical stability. For '1-norm fidelity term, we iteratively minimize an '1-'2 cost function using accelerated proximal gradient (APG), where a good step size is chosen via Lipschitz continuity analysis. Finally, we propose simple mean and median filters for flat patches of a given point cloud to estimate the noise variance given the noise type, which in turn is used to compute a weight parameter trading off the fidelity term and signal prior in the problem formulation. Extensive experiments show state-of-the-art denoising performance among local methods using our proposed algorithms.

Just noticeable difference (JND) models are widely used for perceptual redundancy estimation in images and videos. A common method for measuring the accuracy of a JND model is to inject random noise in an image based on the JND model, and check whether the JND-noise-contaminated image is perceptually distinguishable from the original image or not. Also, when comparing the accuracy of two different JND models, the model that produces the JND-noise-contaminated image with better quality at the same level of noise energy is the better model. But in both of these cases, a subjective test is necessary, which is very time consuming and costly. In this paper, we present a full-reference metric called PDP (perceptual distinguishability predictor), which can be used to determine whether a given JND-noise-contaminated image is perceptually distinguishable from the reference image. The proposed metric employs the concept of sparse coding, and extracts a feature vector out of a given image pair. The feature vector is then fed to a multilayer neural network for classification. To train the network, we built a public database of 999 natural images with distinguishbility thresholds for four different JND models obtained from an extensive subjective experiment. The results indicated that PDD achieves high classification accuracy of 97.1%. The proposed method can be used to objectively compare various JND models without performing any subjective test. It can also be used to obtain proper scaling factors to improve the JND thresholds estimated by an arbitrary JND model.

A new method for video watermarking is presented in this paper. In the proposed method, data are embedded in the LL subband of wavelet coefficients, and decoding is performed based on the comparison among the elements of the first principal component resulting from empirical principal component analysis (PCA). The locations for data embedding are selected such that they offer the most robust PCA-based decoding. Data are inserted in the LL subband in an adaptive manner based on the energy of high frequency subbands and visual saliency. Extensive testing was performed under various types of attacks, such as spatial attacks (uniform and Gaussian noise and median filtering), compression attacks (MPEG-2, H. 263, and H. 264), and temporal attacks (frame repetition, frame averaging, frame swapping, and frame rate conversion). The results show that the proposed method offers improved performance compared with several methods from the literature, especially under additive noise and compression attacks.

In this paper we present joint decoding of JPEG2000 bitstreams and Reed-Solomon codes in the context of unequal loss protection. Using error resilience features of JPEG2000 bitstreams, the joint decoder helps to restore the erased symbols when the Reed-Solomon decoder fails to retrieve them on its own. However, the joint decoding process might become time-consuming due to a search through the set of possible erased symbols. We propose the use of smaller codeblocks and transmission of a relatively small amount of side information with high reliability as two approaches to accelerate the joint decoding process. The accelerated joint decoder can deliver essentially the same quality enhancement as the nonaccelerated one, while operating several times faster.

In this paper, we present a method of creating domain-based multiple descriptions of images and video. These descriptions are created by partitioning the transform domain of the signal into sets whose points are maximally separated from each other. This property enables simple error concealment methods to produce good estimates of lost signal samples. We present the approach in the context of Internet transmission of subband/wavelet-coded images and scalable motion compensated three-dimensional (3D) subband/wavelet-coded video, but applications are not limited to these scenarios. The results indicate that the proposed methods offer improvements over similar competing methods by up to 1 dB for images, and several decibels for video. Visual quality is also improved.