Road damage detection is an important part of road maintenance and management, as it allows for timely repairs and helps to extend the lifespan of roads. Recent advancements in deep learning have led to the development of deep models that can analyze images from road inspections to automatically detect and classify damage. In this work, we propose a novel three-stage learning approach for road damage detection. Our method combines multiple state-of-the-art object detection models and leverage their strengths through mutual learning and knowledge distillation. The approach achieves an Fl-score of 0.7927 and an inference speed of 0.0268 seconds per image, securing the first place in both Phase 1 and Phase 2 of the Optimized Road Damage Detection Challenge (ORDDC’2024). The extensive experiments and comparisons with existing methods demonstrate the effectiveness of our method.

We address the problem of estimating a high-quality dense depth map from a single RGB input image. We first analyze Conditional Random Field (CRF) in combination with transformers and exploit the multi-head attention mechanism to compute a potential function. Then, we propose spatial window CRFs and channel-wise CRFs to observe information in spatial and channel dimensions, and fuse them with a two-way fusion module, which is called Dual aggregation CRFs (DCRFs). Finally, the information from the multi-scale features observed by DCRFs is used for internal scene clustering by slot-attention to obtain the depth map. We call our method as MSD-CRFs. Experiments demonstrate that our method improves the performance across all metrics on the KITTI, and outperforms current SOTA results on the main ranking metrics $A b s \_$Rel on NYU Depth-v2. Further, we explore the model generalization capability via zero-shot test.

This article aims to propose a novel Analytical Tensor Voting (ATV) mechanism, which enables robust perceptual grouping and salient information extraction for noisy N-dimensional (ND) data. Firstly, the approximation of the decaying function is investigated and adopted based on the idea of penalizing the 1-tensor votes by distance and curvature, respectively, followed by the derivation of analytical solution to the 1-tensor voting in ND space from the geometric view. Secondly, a novel spherical representation mechanism is proposed to facilitate the representation of the elementary tensors in various dimensional spaces, where the high dimensional spherical coordinate system is utilized to construct the controllable unit vectors and corresponding 1-tensors. Accordingly, any elementary K-tensor is represented by the surface integration of the constructed 1-tensors over the unit K-sphere. Thirdly, the ATV mechanism is constructed using the adopted decaying function and proposed spherical representation mechanism, where the analytical solution to tensor voting in ND space is derived, which enables the robust and accurate salient information extraction from noisy ND data. Finally, several interesting properties of the proposed ATV mechanism are investigated. Experimental results on synthetic and real data validate the effectiveness, efficiency and robustness of the proposed method in perceptual grouping tasks in 3D,10D or higher dimensional spaces.

3D object detection has recently become popular due to many applications in robotics, augmented reality, autonomy, and image retrieval. We introduce the Objectron dataset to advance the state of the art in 3D object detection and foster new research and applications, such as 3D object tracking, view synthesis, and improved 3D shape representation. The dataset contains object-centric short videos with pose annotations for nine categories and includes 4 million annotated images in 14, 819 annotated videos. We also propose a new evaluation metric, 3D Intersection over Union, for 3D object detection. We demonstrate the usefulness of our dataset in 3D object detection and novel view synthesis tasks by providing baseline models trained on this dataset. Our dataset and evaluation source code are available online at Github.com/google-research-datasets/Objectron.

Many imaging applications require the implementation of space-varying convolution for accurate restoration and reconstruction of images. Here, we use the term space-varying convolution to refer to linear operators whose impulse response has slow spatial variation. In addition, these space-varying convolution operators are often dense, so direct implementation of the convolution operator is typically computationally impractical. One such example is the problem of stray light reduction in digital cameras, which requires the implementation of a dense space-varying deconvolution operator. However, other inverse problems, such as iterative tomographic reconstruction, can also depend on the implementation of dense space-varying convolution. While space-invariant convolution can be efficiently implemented with the fast Fourier transform, this approach does not work for space-varying operators. So direct convolution is often the only option for implementing space-varying convolution. In this paper, we develop a general approach to the efficient implementation of space-varying convolution, and demonstrate its use in the application of stray light reduction. Our approach, which we call matrix source coding, is based on lossy source coding of the dense space-varying convolution matrix. Importantly, by coding the transformation matrix, we not only reduce the memory required to store it; we also dramatically reduce the computation required to implement matrix-vector products. Our algorithm is able to reduce computation by approximately factoring the dense space-varying convolution operator into a product of sparse transforms. Experimental results show that our method can dramatically reduce the computation required for stray light reduction while maintaining high accuracy.

We propose a new model for event-related functional magnetic resonance imaging (fMRI), and develop a new set of tools for activation detection. A novel feature of our framework is the explicit modeling of the spatial correlation introduced by the scanner. We propose simple, efficient algorithms to estimate model parameters. We develop an activation detection algorithm which consists of two parts: image restoration and least-squares estimation of the parameters of the hemodynamic response function. During the image restoration stage, a total-variation-based approach is employed to restore each data slice, for each time index. The amplitude of the least-squares fit of the hemodynamic response function is then thresholded to yield an estimate of the activation map. We illustrate the promise of our method through several experiments with synthetic data as well as one example with real data.