Detecting small objects in large-scale scenes remains a fundamental challenge in object detection, primarily due to scale variation, occlusion, and limited resolution. In order to contribute in this research topic, we propose Density Aided Hyper Inference (DAHI), a lightweight and detector-agnostic framework that enhances detection performance through a structured, three-stage inference process. DAHI combines: (i) Region Density Estimation (RDE), which identifies areas likely to contain overlooked objects; (ii) Density-Aided Crop Selection (DACS), which efficiently selects high-density, low-overlap regions for re-inference; and (iii) Crop Margin Aware Non-Maximum Suppression (CMA-NMS), which merges detections from full-image and region-based inferences while mitigating boundary-related errors. DAHI requires no retraining and integrates seamlessly with standard object detectors. Experiments on several aerial and driving detection benchmarks demonstrate improved detection quality and runtime efficiency compared to existing multi-inference approaches, while introducing reduced computational overhead. These results support the use of DAHI as an effective and practical enhancement for small object detection in complex visual scenes.

In this article, we propose a new planar shape representation, the medial branch graph representation (MBGR) which allows to easily generate, vary and edit all-new sorts of parametric shapes. Each MBGR shape is described by a collection of connected control circles organized as a graph. Each pair of connected circles generates a branch of the shape. The boundaries of the branches are designed using cubic Bézier curves which enable shape representation in a compact SVG format. Each circle is associated with a regularity parameter locally controlling the smoothness of the branch connections. In this way, we can manage shapes having both sharp corners and smooth boundaries. To illustrate the potential of the MBGR representation, we have created an online facility ( https://mbgrg.github.io/mbgrg/ ) where the user can create/edit MBGR shapes automatically or interactively.

An accurate estimation of the position, orientation, principal curvatures, and change in intensity of the edges in a 3D image provides highly useful information for many applications. The use of derivative operators to compute the gradient vector and the Hessian matrix in each voxel usually generates inaccurate results. This paper presents a new edge detector which is derived from the Partial Volume Effect (PVE). Instead of assuming continuity in the image values, edge features are extracted from the distribution of intensities within a neighborhood of each edge voxel. First, the influence of the intensities of the voxels in first- and second-order edges is analyzed to demonstrate that these types of edges can be precisely characterized from the intensity distribution. Afterward, this approach is extended to especially demanding situations by considering how adverse conditions can be tackled. This extension includes filtering noisy images, characterizing edges in blurred regions, and using windows with floating limits for close edges. The proposed technique has been tested on synthetic and real images, including some particularly difficult objects, and achieving a highly accurate subvoxel characterization of the edges. An open source implementation of our method is provided.

The aim of this paper is to study several strategies for the preservation of flow discontinuities in variational optical flow methods. We analyze the combination of robust functionals and diffusion tensors in the smoothness assumption. Our study includes the use of tensors based on decreasing functions, which has shown to provide good results. However, it presents several limitations and usually does not perform better than other basic approaches. It typically introduces instabilities in the computed motion fields in the form of independent blobs of vectors with large magnitude. We propose two alternatives to overcome these drawbacks: first, a simple approach that combines the decreasing function with a minimum isotropic smoothing, and second, a method that looks for the best parameter configuration that preserves the important motion contours and avoid instabilities. It relies on the input images and the regularization parameter. It is fully automatic, providing a near-optimal value for many sequences, as shown in the experiments. Both proposals allow to detect the contours of the motion field and produce more stable solutions for a large range of parameters. In the experimental results, we present a detailed study and comparison of the different strategies.

The use of decreasing functions, for mitigating the regularization at image contours, is typical in many recent optical flow methods. However, finding the correct parameter for getting the best of this strategy is challenging. Most of the methods use default parameters that are conservative, providing results that are not better than traditional approaches. Configurations that clearly enhance discontinuities may produce instabilities in the computed optical flows. This is due to the fact that the regularization process may get cancelled, yielding an ill-posed problem. In this work, we analyze the problem of instabilities and propose a method for efficiently determining the value of the parameter. We show that this approach allows us to obtain well preserved discontinuities at the same time that it avoids the ill-posed problem. The experiments with synthetic sequences demonstrate that the results are accurate and the selected parameter is close to the optimal value.