Understanding decision-making in autonomous driving models is essential for real-world applications. Attribution explanation is a primary research direction for interpreting neural network decisions. However, in the context of autonomous driving, numerical attributions fail to interpret the complex semantic information and often result in explanations that are difficult to understand. This paper introduces a novel semantic attribution approach that both identifies where important features appear and provides intuitive information about what they represent. To establish the semantic correspondences for attributions, we propose an interpreting framework that integrates unsupervised differentiable semantic representations with the attribution computational model. To further enhance the accuracy of the attribution computation while ensuring strong semantic correspondence, we design a Semantic-Informed Aumann-Shapley (SIAS) method, which defines a novel integration path solution using constraints from semantic scores and discrete gradients. Extensive experiments confirm that our method outperforms state-of-the-art explanation techniques both qualitatively and quantitatively in autonomous driving scenarios.

Given the critical need for more reliable autonomous driving systems, explainability has become a key focus within the research community. In autonomous driving models, even minor perception differences can significantly influence the decision-making process, and this impact often diverges markedly from human cognition. However, understanding the specific reasons why a model decides to stop or keep forward remains a significant challenge. This paper presents an attribution-guided visualization method aimed at exploring the triggers behind decision shifts, providing clear insights into the underlying “why” and “why not” of such decisions. We propose the cumulative layer fusion attribution method that identifies the parameters most critical to decision-making. These attributions are then used to inform the visualization optimization by applying attribution-guided weights to crucial generation parameters, ensuring that decision changes are driven only by modifications to critical information. Furthermore, we develop an indirect regularization method that increases visualization quality without necessitating additional hyperparameters. Experiments on large datasets demonstrate that our method produces insightful visualization explanations and outperforms state-of-the-art methods in both qualitative and quantitative evaluations.

Deep neural networks (DNNs) have advanced autonomous driving, but their lack of transparency remains a major obstacle to real-world application. Attribution methods, which aim to explain DNN decisions, offer a potential solution. However, existing methods, primarily designed for image classification models, often suffer from performance degradation and require specialized algorithmic adjustments when applied to the diverse models in autonomous driving. To address this challenge, we introduce a universally applicable representation of traffic scenes, forming the basis for our unified attribution method. Specifically, we leverage the first-order Taylor expansion at a specific hidden layer, i.e., the product of gradients and feature maps, to represent abstract traffic scene information. This representation guides both the optimization of attribution path generation and the attribution computation, enabling consistent and effective attributions for both lane-change prediction and vision-based control models. Experiments on two distinct autonomous driving models demonstrate that our approach outperforms state-of-the-art methods in explanation accuracy and robustness, advancing the interpretability of DNN-based autonomous driving models.

The restoration of digital images holds practical significance due to the fact that degradation of digital image data on the internet is common. State-of-the-art image restoration methods usually employ end-to-end trained networks. However, we argue that a network trained with diverse image pairs is not optimal for restoring line drawings which have extensive plain backgrounds. We propose a line-drawing restoration framework which takes a restoration neural network as backbone and processes an input degraded line drawing in two steps. First, a proposed mask-predicting network predicts a line mask which indicates the possible location of foreground and background in the potential original line drawing. Next, we feed the degraded input line drawing together with the predicted line mask into the backbone restoration network. The traditional L 1 loss for the backbone restoration network is substituted with a masked Mean Square Error (MSE) loss. We test our framework on two classical image restoration tasks: JPEG restoration and super-resolution, and experiments demonstrate that our framework can achieve better quantitative and visual results in most cases.

Recent research has demonstrated that Vision Transformers (ViTs) are capable of comparable or even better performance than convolutional neural network (CNN) baselines. The differences in their structural designs are obvious, but our understanding of the differences in their feature representations remains limited. In this work, we propose several techniques to achieve high-quality visualization of representations in ViTs. Both qualitative and quantitative experiments show that our technical improvements can observably improve ViT visualization quality compared to previous studies. Furthermore, we conduct visualizations to explore the disparities between ViTs and CNNs pre-trained on ImageNet1K, revealing three intriguing properties of ViTs: a) ViT feature propagation retains image detail information with minimal loss, whereas CNNs discard most image details for class discrimination. b) Different from CNNs, object-related features do not show in ViT higher layers, suggesting that class-discriminative features may not be required for ViT classification. c) Our visualization-assisted texture-bias experiment reveals that both ViTs and CNNs exhibit texture bias, of which ViTs seem to be more biased towards local textures.

Subsurface scattering involves the complicated behavior of light beneath the surfaces of translucent objects that includes scattering and absorption inside the object’s volume. Physically accurate numerical representation of subsurface scattering requires a large number of parameters because of the complex nature of this phenomenon. The large amount of data restricts the use of the data on memory-limited devices such as video game consoles and mobile phones. To address this problem, this paper proposes an efficient data compression method for heterogeneous subsurface scattering. The key insight of this study is that heterogeneous materials often comprise a limited number of base materials, and the size of the subsurface scattering data can be significantly reduced by parameterizing only a few base materials. In the proposed compression method, we represent the scattering property of a base material using a function referred to as the base scattering profile. A small subset of the base materials is assigned to each surface position, and the local scattering property near the position is described using a linear combination of the base scattering profiles in the log scale. The proposed method reduces the data by a factor of approximately 30 compared to a state-of-the-art method, without significant loss of visual quality in the rendered graphics. In addition, the compressed data can also be used as bidirectional scattering surface reflectance distribution functions (BSSRDF) without incurring much computational overhead. These practical aspects of the proposed method also facilitate the use of higher-resolution BSSRDFs in devices with large memory capacity.

We propose a method to extract palette colors from an image and edit the image for palette-based photo recoloring. A previous study proposed an editing method that used a color palette. However, that method had some constraints in changing the palette color, and it was also necessary for the user to determine the palette size. Undesirable luminance changes during color editing can be avoided by performing color editing using intrinsic color decomposition. The palette size is estimated using the Bayesian information criterion for the palette colors extracted from the reflectance image; this can later be arbitrarily changed by the user. We evaluated our approach via a user study, showing that it is more precise and faster at matching the image to the target than a previously reported method.

Although many photorealistic relighting methods provide a way to change the illumination of objects in a digital photograph, it is currently difficult to relight digital illustrations having a cartoon shading style. The main difference between photorealistic and cartoon shading styles is that cartoon shading is characterized by soft color quantization and nonlinear color variations that cause noticeable reconstruction errors under a physical reflectance assumption, such as Lambertian reflection. To handle this non-photorealistic shading property, we focus on shading analysis of the most fundamental cartoon shading technique. Based on the color map shading representation, we propose a simple method to determine the input shading as that of a smooth shape with a nonlinear reflectance property. We have conducted simple ground-truth evaluations to compare our results to those obtained by other approaches.

One of the most effective techniques for developing efficient isosurfacing algorithms is the reduction of visits to nonisosurface cells. Recent algorithms have drastically reduced the unnecessary cost of visiting nonisosurface cells. The experimental results show almost optimal performance in their isosurfacing processes. However, most of them have a bottleneck in that they require more than O(n) computation time for their preprocessing, where n denotes the total number of cells. We propose an efficient isosurfacing technique, which can be applied to unstructured as well as structured volumes and which does not require more than O(n) computation time for its preprocessing. A preprocessing step generates an extrema skeleton, which consists of cells and connects all extremum points, by the volume thinning algorithm. All disjoint parts of every isosurface intersect at least one cell in the extrema skeleton. Our implementation generates isosurfaces by searching for isosurface cells in the extrema skeleton and then recursively visiting their adjacent isosurface cells, while it skips most of the nonisosurface cells. The computation time of the preprocessing is estimated as O(n). The computation time of the isosurfacing process is estimated as O(n/sup 1/3/m+k), where k denotes the number of isosurface cells and m denotes the number of extremum points since the number of cells in an extrema skeleton is estimated as O(n/sup 1/3/m).

One of the essential properties of a surface is its normal vector. Many applications, i.e., surface rendering, surface-surface intersection, and offset surface generation, require normal vectors. A normal vector at a point on a tensor product surface is usually obtained by taking a cross product of the two partial derivatives. The paper discusses a Bezier normal vector surface which is a locus of an unnormalized cross product normal vector. It also explains several applications of the Bezier normal vector surface, such as detection and computation of degenerate normal vectors which cannot be calculated by the cross product, and an algorithm to find all critical points which are key points to solve the problems on surface-surface intersection.

An isosurface can be efficiently generated by visiting adjacent intersected cells in order, as if the isosurface were propagating itself. We previously proposed an extrema graph method (T. Itoh and K. Koyamada, 1995), which generates a graph connecting extremum points. The isosurface propagation starts from some of the intersected cells that are found both by visiting the cells through which arcs of the graph pass and by visiting the cells on the boundary of a volume. We propose an efficient method of searching for cells intersected by an isosurface. This method generates a volumetric skeleton. consisting of cells, like an extrema graph, by applying a thinning algorithm used in the image recognition area. Since it preserves the topological features of the volume and the connectivity of the extremum points, it necessarily intersects every isosurface. The method is more efficient than the extrema graph method, since it does not require that cells on the boundary be visited.

and the presence of calmodulin (CaM) leads to the activation of calcineurin (CaN), a protein with serine-threonine phosphatase activity. The signaling pathway linked with CaN and transcription factors like the nuclear factor of activated T cells (NFAT) is essential for skeletal muscle development and reprogramming of fast-twitch to slow-twitch fibers. CaN activation may promote metabolic adaptations in muscle cells, resulting in better insulin-stimulated glucose transport. The molecular mechanisms underlying the altered insulin response remain unclear. The role of the CaN/NFAT pathway in regulating skeletal muscle hypertrophy is better described than its involvement in the pathogenesis of insulin resistance. Thus, there are opportunities for future research in that field. This review presents the role of CaN/NFAT signaling and suggests the relationship with insulin-resistant muscles.