André Kaup

Amir El-Ghoussani, André Kaup, Nassir Navab et al.

We propose a monocular depth estimation method based on visual autoregressive (VAR) priors, offering an alternative to diffusion-based approaches. Our method adapts a large-scale text-to-image VAR model and introduces a scale-wise conditional upsampling mechanism with classifier-free guidance. Our approach performs inference in ten fixed autoregressive stages, requiring only 74K synthetic samples for fine-tuning, and achieves competitive results. We report state-of-the-art performance in indoor benchmarks under constrained training conditions, and strong performance when applied to outdoor datasets. This work establishes autoregressive priors as a complementary family of geometry-aware generative models for depth estimation, highlighting advantages in data scalability, and adaptability to 3D vision tasks. Code available at "https://github.com/AmirMaEl/VAR-Depth".

Azhar Hussian, Marina Ritthaler, André Kaup et al.

We present an image-conditioned point cloud completion approach that treats images as the primary geometric source rather than a secondary guide. To this end, we introduce an Image-to-Point (I2P) module that can reconstruct complete point clouds directly from a single RGB image, with no need for 3D inputs. Additionally, we introduce a transformer-based Point-to-Point (P2P) refinement module that uses self- and cross-attention between point tokens and image features to iteratively refine the coarse I2P output. The I2P module enables the image encoder to learn rich geometric representations, while the P2P module progressively recovers fine-grained details. Unlike existing multimodal methods that rely on auxiliary losses or fusion modules, our explicit I2P task provides a strong, geometry-aware prior based on images alone. Extensive experiments on ShapeNet-ViPC demonstrate state-of-the-art completion performance with a 12.3% relative Chamfer Distance improvement over prior methods. Code is available at: https://github.com/AzharSindhi/I2PRef.git

Ming Zhao, Yongpeng Wu, Shengda Pan et al.

The implementation of automatic image registration is still difficult in various applications. In this paper, an automatic image registration approach through line-support region segmentation and geometrical outlier removal (ALRS-GOR) is proposed. This new approach is designed to address the problems associated with the registration of images with affine deformations and inconsistent content, such as remote sensing images with different spectral content or noise interference, or map images with inconsistent annotations. To begin with, line-support regions, namely a straight region whose points share roughly the same image gradient angle, are extracted to address the issues of inconsistent content existing in images. To alleviate the incompleteness of line segments, an iterative strategy with multi-resolution is employed to preserve global structures that are masked at full resolution by image details or noise. Then, Geometrical Outlier Removal (GOR) is developed to provide reliable feature point matching, which is based on affineinvariant geometrical classifications for corresponding matches initialized by SIFT. The candidate outliers are selected by comparing the disparity of accumulated classifications among all matches, instead of conventional methods which only rely on local geometrical relations. Various image sets have been considered in this paper for the evaluation of the proposed approach, including aerial images with simulated affine deformations, remote sensing optical and synthetic aperture radar images taken at different situations (multispectral, multisensor, and multitemporal), and map images with inconsistent annotations. Experimental results demonstrate the superior performance of the proposed method over the existing approaches for the whole data set.

Andrea Eichenseer, André Kaup

In video surveillance as well as automotive applications, so-called fisheye cameras are often employed to capture a very wide angle of view. As such cameras depend on projections quite different from the classical perspective projection, the resulting fisheye image and video data correspondingly exhibits non-rectilinear image characteristics. Typical image and video processing algorithms, however, are not designed for these fisheye characteristics. To be able to develop and evaluate algorithms specifically adapted to fisheye images and videos, a corresponding test data set is therefore introduced in this paper. The first of those sequences were generated during the authors' own work on motion estimation for fish-eye videos and further sequences have gradually been added to create a more extensive collection. The data set now comprises synthetically generated fisheye sequences, ranging from simple patterns to more complex scenes, as well as fisheye video sequences captured with an actual fisheye camera. For the synthetic sequences, exact information on the lens employed is available, thus facilitating both verification and evaluation of any adapted algorithms. For the real-world sequences, we provide calibration data as well as the settings used during acquisition. The sequences are freely available via www.lms.lnt.de/fisheyedataset/.

Kristian Fischer, Fabian Brand, Christian Herglotz et al.

Common state-of-the-art video codecs are optimized to deliver a low bitrate by providing a certain quality for the final human observer, which is achieved by rate-distortion optimization (RDO). But, with the steady improvement of neural networks solving computer vision tasks, more and more multimedia data is not observed by humans anymore, but directly analyzed by neural networks. In this paper, we propose a standard-compliant feature-based RDO (FRDO) that is designed to increase the coding performance, when the decoded frame is analyzed by a neural network in a video coding for machine scenario. To that extent, we replace the pixel-based distortion metrics in conventional RDO of VTM-8.0 with distortion metrics calculated in the feature space created by the first layers of a neural network. Throughout several tests with the segmentation network Mask R-CNN and single images from the Cityscapes dataset, we compare the proposed FRDO and its hybrid version HFRDO with different distortion measures in the feature space against the conventional RDO. With HFRDO, up to 5.49 % bitrate can be saved compared to the VTM-8.0 implementation in terms of Bjøntegaard Delta Rate and using the weighted average precision as quality metric. Additionally, allowing the encoder to vary the quantization parameter results in coding gains for the proposed HFRDO of up 9.95 % compared to conventional VTM.

Ming Zhao, Bowen An, Yongpeng Wu et al.

Reliable feature point matching is a vital yet challenging process in feature-based image registration. In this paper,a robust feature point matching algorithm called Recovery and Filtering Vertex Trichotomy Matching (RFVTM) is proposed to remove outliers and retain sufficient inliers for remote sensing images. A novel affine invariant descriptor called vertex trichotomy descriptor is proposed on the basis of that geometrical relations between any of vertices and lines are preserved after affine transformations, which is constructed by mapping each vertex into trichotomy sets. The outlier removals in Vertex Trichotomy Matching (VTM) are implemented by iteratively comparing the disparity of corresponding vertex trichotomy descriptors. Some inliers mistakenly validated by a large amount of outliers are removed in VTM iterations, and several residual outliers close to correct locations cannot be excluded with the same graph structures. Therefore, a recovery and filtering strategy is designed to recover some inliers based on identical vertex trichotomy descriptors and restricted transformation errors. Assisted with the additional recovered inliers, residual outliers can also be filtered out during the process of reaching identical graph for the expanded vertex sets. Experimental results demonstrate the superior performance on precision and stability of this algorithm under various conditions, such as remote sensing images with large transformations, duplicated patterns, or inconsistent spectral content.

Ján Koloda, Jürgen Seiler, André Kaup et al.

The purpose of signal extrapolation is to estimate unknown signal parts from known samples. This task is especially important for error concealment in image and video communication. For obtaining a high quality reconstruction, assumptions have to be made about the underlying signal in order to solve this underdetermined problem. Among existent reconstruction algorithms, frequency selective extrapolation (FSE) achieves high performance by assuming that image signals can be sparsely represented in the frequency domain. However, FSE does not take into account the low-pass behaviour of natural images. In this paper, we propose a modified FSE that takes this prior knowledge into account for the modelling, yielding significant PSNR gains.

Kristian Fischer, Felix Fleckenstein, Christian Herglotz et al.

Saliency-driven image and video coding for humans has gained importance in the recent past. In this paper, we propose such a saliency-driven coding framework for the video coding for machines task using the latest video coding standard Versatile Video Coding (VVC). To determine the salient regions before encoding, we employ the real-time-capable object detection network You Only Look Once~(YOLO) in combination with a novel decision criterion. To measure the coding quality for a machine, the state-of-the-art object segmentation network Mask R-CNN was applied to the decoded frame. From extensive simulations we find that, compared to the reference VVC with a constant quality, up to 29 % of bitrate can be saved with the same detection accuracy at the decoder side by applying the proposed saliency-driven framework. Besides, we compare YOLO against other, more traditional saliency detection methods.

Andrea Eichenseer, Michel Bätz, Jürgen Seiler et al.

Capturing large fields of view with only one camera is an important aspect in surveillance and automotive applications, but the wide-angle fisheye imagery thus obtained exhibits very special characteristics that may not be very well suited for typical image and video processing methods such as motion estimation. This paper introduces a motion estimation method that adapts to the typical radial characteristics of fisheye video sequences by making use of an equisolid re-projection after moving part of the motion vector search into the perspective domain via a corresponding back-projection. By combining this approach with conventional translational motion estimation and compensation, average gains in luminance PSNR of up to 1.14 dB are achieved for synthetic fish-eye sequences and up to 0.96 dB for real-world data. Maximum gains for selected frame pairs amount to 2.40 dB and 1.39 dB for synthetic and real-world data, respectively.

Esin Koyuncu, Timofey Solovyev, Elena Alshina et al.

Learning-based image compression has improved to a level where it can outperform traditional image codecs such as HEVC and VVC in terms of coding performance. In addition to good compression performance, device interoperability is essential for a compression codec to be deployed, i.e., encoding and decoding on different CPUs or GPUs should be error-free and with negligible performance reduction. In this paper, we present a method to solve the device interoperability problem of a state-of-the-art image compression network. We implement quantization to entropy networks which output entropy parameters. We suggest a simple method which can ensure cross-platform encoding and decoding, and can be implemented quickly with minor performance deviation, of 0.3% BD-rate, from floating point model results.

Ján Koloda, Jürgen Seiler, Antonio M. Peinado et al.

Error concealment is of great importance for block-based video systems, such as DVB or video streaming services. In this paper, we propose a novel scalable spatial error concealment algorithm that aims at obtaining high quality reconstructions with reduced computational burden. The proposed technique exploits the excellent reconstructing abilities of the kernel-based minimum mean square error K-MMSE estimator. We propose to decompose this approach into a set of hierarchically stacked layers. The first layer performs the basic reconstruction that the subsequent layers can eventually refine. In addition, we design a layer management mechanism, based on profiles, that dynamically adapts the use of higher layers to the visual complexity of the area being reconstructed. The proposed technique outperforms other state-of-the-art algorithms and produces high quality reconstructions, equivalent to K-MMSE, while requiring around one tenth of its computational time.

Ján Koloda, Jürgen Seiler, André Kaup

Digital images are commonly represented as regular 2D arrays, so pixels are organized in form of a matrix addressed by integers. However, there are many image processing operations, such as rotation or motion compensation, that produce pixels at non-integer positions. Typically, image reconstruction techniques cannot handle samples at non-integer positions. In this paper, we propose to use triangulation-based reconstruction as initial estimate that is later refined by a novel adaptive denoising framework. Simulations reveal that improvements of up to more than 1.8 dB (in terms of PSNR) are achieved with respect to the initial estimate.

Andreas Spruck, Maximiliane Hawesch, Anatol Maier et al.

In this paper, we propose a data augmentation framework for Optical Character Recognition (OCR). The proposed framework is able to synthesize new viewing angles and illumination scenarios, effectively enriching any available OCR dataset. Its modular structure allows to be modified to match individual user requirements. The framework enables to comfortably scale the enlargement factor of the available dataset. Furthermore, the proposed method is not restricted to single frame OCR but can also be applied to video OCR. We demonstrate the performance of our framework by augmenting a 15% subset of the common Brno Mobile OCR dataset. Our proposed framework is capable of leveraging the performance of OCR applications especially for small datasets. Applying the proposed method, improvements of up to 2.79 percentage points in terms of Character Error Rate (CER), and up to 7.88 percentage points in terms of Word Error Rate (WER) are achieved on the subset. Especially the recognition of challenging text lines can be improved. The CER may be decreased by up to 14.92 percentage points and the WER by up to 18.19 percentage points for this class. Moreover, we are able to achieve smaller error rates when training on the 15% subset augmented with the proposed method than on the original non-augmented full dataset.

Arya Rachman, Jürgen Seiler, André Kaup

Autonomous vehicles are equipped with a multi-modal sensor setup to enable the car to drive safely. The initial calibration of such perception sensors is a highly matured topic and is routinely done in an automated factory environment. However, an intriguing question arises on how to maintain the calibration quality throughout the vehicle's operating duration. Another challenge is to calibrate multiple sensors jointly to ensure no propagation of systemic errors. In this paper, we propose CaLiCa, an end-to-end deep self-calibration network which addresses the automatic calibration problem for pinhole camera and Lidar. We jointly predict the camera intrinsic parameters (focal length and distortion) as well as Lidar-Camera extrinsic parameters (rotation and translation), by regressing feature correlation between the camera image and the Lidar point cloud. The network is arranged in a Siamese-twin structure to constrain the network features learning to a mutually shared feature in both point cloud and camera (Lidar-camera constraint). Evaluation using KITTI datasets shows that we achieve 0.154 ° and 0.059 m accuracy with a reprojection error of 0.028 pixel with a single-pass inference. We also provide an ablative study of how our end-to-end learning architecture offers lower terminal loss (21% decrease in rotation loss) compared to isolated calibration

Viktoria Heimann, Andreas Spruck, André Kaup

The demand for high-resolution point clouds has increased throughout the last years. However, capturing high-resolution point clouds is expensive and thus, frequently replaced by upsampling of low-resolution data. Most state-of-the-art methods are either restricted to a rastered grid, incorporate normal vectors, or are trained for a single use case. We propose to use the frequency selectivity principle, where a frequency model is estimated locally that approximates the surface of the point cloud. Then, additional points are inserted into the approximated surface. Our novel frequency-selective geometry upsampling shows superior results in terms of subjective as well as objective quality compared to state-of-the-art methods for scaling factors of 2 and 4. On average, our proposed method shows a 4.4 times smaller point-to-point error than the second best state-of-the-art PU-Net for a scale factor of 4.

Andreas Spruck, Maximilane Gruber, Anatol Maier et al.

An insufficient number of training samples is a common problem in neural network applications. While data augmentation methods require at least a minimum number of samples, we propose a novel, rendering-based pipeline for synthesizing annotated data sets. Our method does not modify existing samples but synthesizes entirely new samples. The proposed rendering-based pipeline is capable of generating and annotating synthetic and partly-real image and video data in a fully automatic procedure. Moreover, the pipeline can aid the acquisition of real data. The proposed pipeline is based on a rendering process. This process generates synthetic data. Partly-real data bring the synthetic sequences closer to reality by incorporating real cameras during the acquisition process. The benefits of the proposed data generation pipeline, especially for machine learning scenarios with limited available training data, are demonstrated by an extensive experimental validation in the context of automatic license plate recognition. The experiments demonstrate a significant reduction of the character error rate and miss rate from 73.74% and 100% to 14.11% and 41.27% respectively, compared to an OCR algorithm trained on a real data set solely. These improvements are achieved by training the algorithm on synthesized data solely. When additionally incorporating real data, the error rates can be decreased further. Thereby, the character error rate and miss rate can be reduced to 11.90% and 39.88% respectively. All data used during the experiments as well as the proposed rendering-based pipeline for the automated data generation is made publicly available under (URL will be revealed upon publication).

Viktoria Heimann, Andreas Spruck, André Kaup

With the increasing demand of capturing our environment in three-dimensions for AR/ VR applications and autonomous driving among others, the importance of high-resolution point clouds rises. As the capturing process is a complex task, point cloud upsampling is often desired. We propose Frequency-Selective Upsampling (FSU), an upsampling scheme that upsamples geometry and attribute information of point clouds jointly in a sequential manner with overlapped support areas. The point cloud is partitioned into blocks with overlapping support area first. Then, a continuous frequency model is generated that estimates the point cloud's surface locally. The model is sampled at new positions for upsampling. In a subsequent step, another frequency model is created that models the attribute signal. Here, knowledge from the geometry upsampling is exploited for a simplified projection of the points in two dimensions. The attribute model is evaluated for the upsampled geometry positions. In our extensive evaluation, we evaluate geometry and attribute upsampling independently and show joint results. The geometry results show best performances for our proposed FSU in terms of point-to-plane error and plane-to-plane angular similarity. Moreover, FSU outperforms other color upsampling schemes by 1.9 dB in terms of color PSNR. In addition, the visual appearance of the point clouds clearly increases with FSU.

Ján Koloda, Jürgen Seiler, André Kaup

This paper presents a novel method for the reconstruction of images from samples located at non-integer positions, called mesh. This is a common scenario for many image processing applications, such as super-resolution, warping or virtual view generation in multi-camera systems. The proposed method relies on a set of initial estimates that are later refined by a new reliability-based content-adaptive framework that employs denoising in order to reduce the reconstruction error. The reliability of the initial estimate is computed so stronger denoising is applied to less reliable estimates. The proposed technique can improve the reconstruction quality by more than 2 dB (in terms of PSNR) with respect to the initial estimate and it outperforms the state-of-the-art denoising-based refinement by up to 0.7 dB.

Simon Grosche, Fabian Brand, André Kaup

Quarter sampling and three-quarter sampling are novel sensor concepts that enable the acquisition of higher resolution images without increasing the number of pixels. This is achieved by non-regularly covering parts of each pixel of a low-resolution sensor such that only one quadrant or three quadrants of the sensor area of each pixel is sensitive to light. Combining a properly designed mask and a high-quality reconstruction algorithm, a higher image quality can be achieved than using a low-resolution sensor and subsequent upsampling. For the latter case, the image quality can be further enhanced using super resolution algorithms such as the very deep super resolution network (VDSR). In this paper, we propose a novel end-to-end neural network to reconstruct high resolution images from non-regularly sampled sensor data. The network is a concatenation of a locally fully connected reconstruction network (LFCR) and a standard VDSR network. Altogether, using a three-quarter sampling sensor with our novel neural network layout, the image quality in terms of PSNR for the Urban100 dataset can be increased by 2.96 dB compared to the state-of-the-art approach. Compared to a low-resolution sensor with VDSR, a gain of 1.11 dB is achieved.

Andreas Spruck, Viktoria Heimann, André Kaup

Neural networks are widely used for almost any task of recognizing image content. Even though much effort has been put into investigating efficient network architectures, optimizers, and training strategies, the influence of image interpolation on the performance of neural networks is not well studied. Furthermore, research has shown that neural networks are often sensitive to minor changes in the input image leading to drastic drops of their performance. Therefore, we propose the use of keypoint agnostic frequency selective mesh-to-grid resampling (FSMR) for the processing of input data for neural networks in this paper. This model-based interpolation method already showed that it is capable of outperforming common interpolation methods in terms of PSNR. Using an extensive experimental evaluation we show that depending on the network architecture and classification task the application of FSMR during training aids the learning process. Furthermore, we show that the usage of FSMR in the application phase is beneficial. The classification accuracy can be increased by up to 4.31 percentage points for ResNet50 and the Oxflower17 dataset.

Simon Grosche, Jürgen Seiler, André Kaup

Quarter sampling is a novel sensor design that allows for an acquisition of higher resolution images without increasing the number of pixels. When being used for video data, one out of four pixels is measured in each frame. Effectively, this leads to a non-regular spatio-temporal sub-sampling. Compared to purely spatial or temporal sub-sampling, this allows for an increased reconstruction quality, as aliasing artifacts can be reduced. For the fast reconstruction of such sensor data with a fixed mask, recursive variant of frequency selective reconstruction (FSR) was proposed. Here, pixels measured in previous frames are projected into the current frame to support its reconstruction. In doing so, the motion between the frames is computed using template matching. Since some of the motion vectors may be erroneous, it is important to perform a proper consistency checking. In this paper, we propose faster consistency checking methods as well as a novel recursive FSR that uses the projected pixels different than in literature and can handle dynamic masks. Altogether, we are able to significantly increase the reconstruction quality by + 1.01 dB compared to the state-of-the-art recursive reconstruction method using a fixed mask. Compared to a single frame reconstruction, an average gain of about + 1.52 dB is achieved for dynamic masks. At the same time, the computational complexity of the consistency checks is reduced by a factor of 13 compared to the literature algorithm.

Mahadev Prasad Panda, Purnachandra Rao Makkena, Srivatsa Prativadibhayankaram et al.

Reducing computational complexity remains a critical challenge for the widespread adoption of learning-based image compression techniques. In this work, we propose TreeNet, a novel low-complexity image compression model that leverages a binary tree-structured encoder-decoder architecture to achieve efficient representation and reconstruction. We employ attentional feature fusion mechanism to effectively integrate features from multiple branches. We evaluate TreeNet on three widely used benchmark datasets and compare its performance against competing methods including JPEG AI, a recent standard in learning-based image compression. At low bitrates, TreeNet achieves an average improvement of 4.83% in BD-rate over JPEG AI, while reducing model complexity by 87.82%. Furthermore, we conduct extensive ablation studies to investigate the influence of various latent representations within TreeNet, offering deeper insights into the factors contributing to reconstruction.

Ming Zhao, Xin Zhang, André Kaup

Detecting ships in synthetic aperture radar (SAR) images is challenging due to strong speckle noise, complex surroundings, and varying scales. This paper proposes MLDet, a multitask learning framework for SAR ship detection, consisting of object detection, speckle suppression, and target segmentation tasks. An angle classification loss with aspect ratio weighting is introduced to improve detection accuracy by addressing angular periodicity and object proportions. The speckle suppression task uses a dual-feature fusion attention mechanism to reduce noise and fuse shallow and denoising features, enhancing robustness. The target segmentation task, leveraging a rotated Gaussian-mask, aids the network in extracting target regions from cluttered backgrounds and improves detection efficiency with pixel-level predictions. The Gaussian-mask ensures ship centers have the highest probabilities, gradually decreasing outward under a Gaussian distribution. Additionally, a weighted rotated boxes fusion (WRBF) strategy combines multi-direction anchor predictions, filtering anchors beyond boundaries or with high overlap but low confidence. Extensive experiments on SSDD+ and HRSID datasets demonstrate the effectiveness and superiority of MLDet.

Xiumei Li, Alexander Kopte, André Kaup

Scalable compression is essential for bandwidth-adaptive transmission, yet most learned codecs are optimized for a fixed rate-distortion point, making rate adaptation costly due to re-encoding or maintaining multiple bitstreams. In this work, we propose TAFA-GSGC, a scalable learned point cloud geometry codec that enables multi-quality decoding from a single bitstream and a single trained model. TAFA-GSGC combines layered residual refinement with channel-group entropy coding, and introduces Target-Aligned Feature Aggregation module to reduce cross-layer redundancy in enhancement residuals. Our framework supports up to 9 decodable quality levels with monotonic quality improvement as more subbitstreams are received, while maintaining strong compression efficiency. Compared with the baseline PCGCv2, TAFA-GSGC attains comparable and slightly better RD performance, achieving average BD-Rate savings of -4.99% in D1 and -5.92% in D2.

Zengbao Sun, Ming Zhao, Gaorui Liu et al.

Remote sensing cross-modal text-image retrieval (RSCTIR) has gained attention for its utility in information mining. However, challenges remain in effectively integrating global and local information due to variations in remote sensing imagery and ensuring proper feature pre-alignment before modal fusion, which affects retrieval accuracy and efficiency. To address these issues, we propose CMPAGL, a cross-modal pre-aligned method leveraging global and local information. Our Gswin transformer block combines local window self-attention and global-local window cross-attention to capture multi-scale features. A pre-alignment mechanism simplifies modal fusion training, improving retrieval performance. Additionally, we introduce a similarity matrix reweighting (SMR) algorithm for reranking, and enhance the triplet loss function with an intra-class distance term to optimize feature learning. Experiments on four datasets, including RSICD and RSITMD, validate CMPAGL's effectiveness, achieving up to 4.65% improvement in R@1 and 2.28% in mean Recall (mR) over state-of-the-art methods.

Srivatsa Prativadibhayankaram, Mahadev Prasad Panda, Thomas Richter et al.

We propose the structure and color based learned image codec (SLIC) in which the task of compression is split into that of luminance and chrominance. The deep learning model is built with a novel multi-scale architecture for Y and UV channels in the encoder, where the features from various stages are combined to obtain the latent representation. An autoregressive context model is employed for backward adaptation and a hyperprior block for forward adaptation. Various experiments are carried out to study and analyze the performance of the proposed model, and to compare it with other image codecs. We also illustrate the advantages of our method through the visualization of channel impulse responses, latent channels and various ablation studies. The model achieves Bjøntegaard delta bitrate gains of 7.5% and 4.66% in terms of MS-SSIM and CIEDE2000 metrics with respect to other state-of-the-art reference codecs.

Alexander Kopte, André Kaup

To manage the complexity of transformers in video compression, local attention mechanisms are a practical necessity. The common approach of partitioning frames into patches, however, creates architectural flaws like irregular receptive fields. When adapted for temporal autoregressive models, this paradigm, exemplified by the Video Compression Transformer (VCT), also necessitates computationally redundant overlapping windows. This work introduces 3D Sliding Window Attention (SWA), a patchless form of local attention. By enabling a decoder-only architecture that unifies spatial and temporal context processing, and by providing a uniform receptive field, our method significantly improves rate-distortion performance, achieving Bjørntegaard Delta-rate savings of up to 18.6 % against the VCT baseline. Simultaneously, by eliminating the need for overlapping windows, our method reduces overall decoder complexity by a factor of 2.8, while its entropy model is nearly 3.5 times more efficient. We further analyze our model's behavior and show that while it benefits from long-range temporal context, excessive context can degrade performance.

Xiumei Li, Marc Windsheimer, Misha Sadeghi et al.

Efficient data compression is crucial for the storage and transmission of visual data. However, in facial expression recognition (FER) tasks, lossy compression often leads to feature degradation and reduced accuracy. To address these challenges, this study proposes an end-to-end model designed to preserve critical features and enhance both compression and recognition performance. A custom loss function is introduced to optimize the model, tailored to balance compression and recognition performance effectively. This study also examines the influence of varying loss term weights on this balance. Experimental results indicate that fine-tuning the compression model alone improves classification accuracy by 0.71% and compression efficiency by 49.32%, while joint optimization achieves significant gains of 4.04% in accuracy and 89.12% in efficiency. Moreover, the findings demonstrate that the jointly optimized classification model maintains high accuracy on both compressed and uncompressed data, while the compression model reliably preserves image details, even at high compression rates.

Ayman A. Ameen, Thomas Richter, André Kaup

Current learned image compression models typically exhibit high complexity, which demands significant computational resources. To overcome these challenges, we propose an innovative approach that employs hierarchical feature extraction transforms to significantly reduce complexity while preserving bit rate reduction efficiency. Our novel architecture achieves this by using fewer channels for high spatial resolution inputs/feature maps. On the other hand, feature maps with a large number of channels have reduced spatial dimensions, thereby cutting down on computational load without sacrificing performance. This strategy effectively reduces the forward pass complexity from \(1256 \, \text{kMAC/Pixel}\) to just \(270 \, \text{kMAC/Pixel}\). As a result, the reduced complexity model can open the way for learned image compression models to operate efficiently across various devices and pave the way for the development of new architectures in image compression technology.

Ayman A. Ameen, Thomas Richter, André Kaup

Current image compression models often require separate models for each quality level, making them resource-intensive in terms of both training and storage. To address these limitations, we propose an innovative approach that utilizes latent variables from pre-existing trained models (such as the Stable Diffusion Variational Autoencoder) for perceptual image compression. Our method eliminates the need for distinct models dedicated to different quality levels. We employ overfitted learnable functions to compress the latent representation from the target model at any desired quality level. These overfitted functions operate in the latent space, ensuring low computational complexity, around $25.5$ MAC/pixel for a forward pass on images with dimensions $(1363 \times 2048)$ pixels. This approach efficiently utilizes resources during both training and decoding. Our method achieves comparable perceptual quality to state-of-the-art learned image compression models while being both model-agnostic and resolution-agnostic. This opens up new possibilities for the development of innovative image compression methods.

Srivatsa Prativadibhayankaram, Mahadev Prasad Panda, Jürgen Seiler et al.

In this work, we present a comparison between color spaces namely YUV, LAB, RGB and their effect on learned image compression. For this we use the structure and color based learned image codec (SLIC) from our prior work, which consists of two branches - one for the luminance component (Y or L) and another for chrominance components (UV or AB). However, for the RGB variant we input all 3 channels in a single branch, similar to most learned image codecs operating in RGB. The models are trained for multiple bitrate configurations in each color space. We report the findings from our experiments by evaluating them on various datasets and compare the results to state-of-the-art image codecs. The YUV model performs better than the LAB variant in terms of MS-SSIM with a Bjøntegaard delta bitrate (BD-BR) gain of 7.5\% using VTM intra-coding mode as the baseline. Whereas the LAB variant has a better performance than YUV model in terms of CIEDE2000 having a BD-BR gain of 8\%. Overall, the RGB variant of SLIC achieves the best performance with a BD-BR gain of 13.14\% in terms of MS-SSIM and a gain of 17.96\% in CIEDE2000 at the cost of a higher model complexity.

Marc Windsheimer, Fabian Brand, André Kaup

Most learning-based image compression methods lack efficiency for high image quality due to their non-invertible design. The decoding function of the frequently applied compressive autoencoder architecture is only an approximated inverse of the encoding transform. This issue can be resolved by using invertible latent variable models, which allow a perfect reconstruction if no quantization is performed. Furthermore, many traditional image and video coders apply dynamic block partitioning to vary the compression of certain image regions depending on their content. Inspired by this approach, hierarchical latent spaces have been applied to learning-based compression networks. In this paper, we present a novel concept, which adapts the hierarchical latent space for augmented normalizing flows, an invertible latent variable model. Our best performing model achieved average rate savings of more than 7% over comparable single-scale models.

Andy Regensky, Christian Herglotz, André Kaup

Although fisheye cameras are in high demand in many application areas due to their large field of view, many image and video signal processing tasks such as motion compensation suffer from the introduced strong radial distortions. A recently proposed projection-based approach takes the fisheye projection into account to improve fisheye motion compensation. However, the approach does not consider the large field of view of fisheye lenses that requires the consideration of different motion planes in 3D space. We propose a novel viewport-adaptive motion compensation technique that applies the motion vectors in different perspective viewports in order to realize these motion planes. Thereby, some pixels are mapped to so-called virtual image planes and require special treatment to obtain reliable mappings between the perspective viewports and the original fisheye image. While the state-of-the-art ultra wide-angle compensation is sufficiently accurate, we propose a virtual image plane compensation that leads to perfect mappings. All in all, we achieve average gains of +2.40 dB in terms of PSNR compared to the state of the art in fisheye motion compensation.

Andy Regensky, Christian Herglotz, André Kaup

Inter prediction is one of the key technologies enabling the high compression efficiency of modern video coding standards. 360-degree video needs to be mapped to the 2D image plane prior to coding in order to allow compression using existing video coding standards. The distortions that inevitably occur when mapping spherical data onto the 2D image plane, however, impair the performance of classical inter prediction techniques. In this paper, we propose a motion-plane-adaptive inter prediction technique (MPA) for 360-degree video that takes the spherical characteristics of 360-degree video into account. Based on the known projection format of the video, MPA allows to perform inter prediction on different motion planes in 3D space instead of having to work on the - in theory arbitrarily mapped - 2D image representation directly. We furthermore derive a motion-plane-adaptive motion vector prediction technique (MPA-MVP) that allows to translate motion information between different motion planes and motion models. Our proposed integration of MPA together with MPA-MVP into the state-of-the-art H.266/VVC video coding standard shows significant Bjontegaard Delta rate savings of 1.72% with a peak of 3.97% based on PSNR and 1.56% with a peak of 3.40% based on WS-PSNR compared to the VTM-14.2 baseline on average.

Fabian Brand, Kristian Fischer, Alexander Kopte et al.

Even though rate-distortion optimization is a crucial part of traditional image and video compression, not many approaches exist which transfer this concept to end-to-end-trained image compression. Most frameworks contain static compression and decompression models which are fixed after training, so efficient rate-distortion optimization is not possible. In a previous work, we proposed RDONet, which enables an RDO approach comparable to adaptive block partitioning in HEVC. In this paper, we enhance the training by introducing low-complexity estimations of the RDO result into the training. Additionally, we propose fast and very fast RDO inference modes. With our novel training method, we achieve average rate savings of 19.6% in MS-SSIM over the previous RDONet model, which equals rate savings of 27.3% over a comparable conventional deep image coder.

Simon Grosche, Andy Regensky, Jürgen Seiler et al.

For modern high-resolution imaging sensors, pixel binning is performed in low-lighting conditions and in case high frame rates are required. To recover the original spatial resolution, single-image super-resolution techniques can be applied for upscaling. To achieve a higher image quality after upscaling, we propose a novel binning concept using tetromino-shaped pixels. It is embedded into the field of compressed sensing and the coherence is calculated to motivate the sensor layouts used. Next, we investigate the reconstruction quality using tetromino pixels for the first time in literature. Instead of using different types of tetrominoes as proposed elsewhere, we show that using a small repeating cell consisting of only four T-tetrominoes is sufficient. For reconstruction, we use a locally fully connected reconstruction (LFCR) network as well as two classical reconstruction methods from the field of compressed sensing. Using the LFCR network in combination with the proposed tetromino layout, we achieve superior image quality in terms of PSNR, SSIM, and visually compared to conventional single-image super-resolution using the very deep super-resolution (VDSR) network. For PSNR, a gain of up to \SI[retain-explicit-plus]{+1.92}{dB} is achieved.

Thomas Köhler, Michel Bätz, Farzad Naderi et al.

Capturing ground truth data to benchmark super-resolution (SR) is challenging. Therefore, current quantitative studies are mainly evaluated on simulated data artificially sampled from ground truth images. We argue that such evaluations overestimate the actual performance of SR methods compared to their behavior on real images. Toward bridging this simulated-to-real gap, we introduce the Super-Resolution Erlangen (SupER) database, the first comprehensive laboratory SR database of all-real acquisitions with pixel-wise ground truth. It consists of more than 80k images of 14 scenes combining different facets: CMOS sensor noise, real sampling at four resolution levels, nine scene motion types, two photometric conditions, and lossy video coding at five levels. As such, the database exceeds existing benchmarks by an order of magnitude in quality and quantity. This paper also benchmarks 19 popular single-image and multi-frame algorithms on our data. The benchmark comprises a quantitative study by exploiting ground truth data and qualitative evaluations in a large-scale observer study. We also rigorously investigate agreements between both evaluations from a statistical perspective. One interesting result is that top-performing methods on simulated data may be surpassed by others on real data. Our insights can spur further algorithm development, and the publicy available dataset can foster future evaluations.

Huynh Van Luong, Nikos Deligiannis, Søren Forchhammer et al.

We consider a decomposition method for compressive streaming data in the context of online compressive Robust Principle Component Analysis (RPCA). The proposed decomposition solves an $n$-$\ell_1$ cluster-weighted minimization to decompose a sequence of frames (or vectors), into sparse and low-rank components, from compressive measurements. Our method processes a data vector of the stream per time instance from a small number of measurements in contrast to conventional batch RPCA, which needs to access full data. The $n$-$\ell_1$ cluster-weighted minimization leverages the sparse components along with their correlations with multiple previously-recovered sparse vectors. Moreover, the proposed minimization can exploit the structures of sparse components via clustering and re-weighting iteratively. The method outperforms the existing methods for both numerical data and actual video data.

Thomas Köhler, Michel Bätz, Farzad Naderi et al.

Over the past decades, various super-resolution (SR) techniques have been developed to enhance the spatial resolution of digital images. Despite the great number of methodical contributions, there is still a lack of comparative validations of SR under practical conditions, as capturing real ground truth data is a challenging task. Therefore, current studies are either evaluated 1) on simulated data or 2) on real data without a pixel-wise ground truth. To facilitate comprehensive studies, this paper introduces the publicly available Super-Resolution Erlangen (SupER) database that includes real low-resolution images along with high-resolution ground truth data. Our database comprises image sequences with more than 20k images captured from 14 scenes under various types of motions and photometric conditions. The datasets cover four spatial resolution levels using camera hardware binning. With this database, we benchmark 15 single-image and multi-frame SR algorithms. Our experiments quantitatively analyze SR accuracy and robustness under realistic conditions including independent object and camera motion or photometric variations.

Huynh Van Luong, Nikos Deligiannis, Søren Forchhammer et al.

In support of applications involving multiview sources in distributed object recognition using lightweight cameras, we propose a new method for the distributed coding of sparse sources as visual descriptor histograms extracted from multiview images. The problem is challenging due to the computational and energy constraints at each camera as well as the limitations regarding inter-camera communication. Our approach addresses these challenges by exploiting the sparsity of the visual descriptor histograms as well as their intra- and inter-camera correlations. Our method couples distributed source coding of the sparse sources with a new joint recovery algorithm that incorporates multiple side information signals, where prior knowledge (low quality) of all the sparse sources is initially sent to exploit their correlations. Experimental evaluation using the histograms of shift-invariant feature transform (SIFT) descriptors extracted from multiview images shows that our method leads to bit-rate saving of up to 43% compared to the state-of-the-art distributed compressed sensing method with independent encoding of the sources.

Huynh Van Luong, Jürgen Seiler, André Kaup et al.

This work considers reconstructing a target signal in a context of distributed sparse sources. We propose an efficient reconstruction algorithm with the aid of other given sources as multiple side information (SI). The proposed algorithm takes advantage of compressive sensing (CS) with SI and adaptive weights by solving a proposed weighted $n$-$\ell_{1}$ minimization. The proposed algorithm computes the adaptive weights in two levels, first each individual intra-SI and then inter-SI weights are iteratively updated at every reconstructed iteration. This two-level optimization leads the proposed reconstruction algorithm with multiple SI using adaptive weights (RAMSIA) to robustly exploit the multiple SIs with different qualities. We experimentally perform our algorithm on generated sparse signals and also correlated feature histograms as multiview sparse sources from a multiview image database. The results show that RAMSIA significantly outperforms both classical CS and CS with single SI, and RAMSIA with higher number of SIs gained more than the one with smaller number of SIs.