SDM-Car: A Dataset for Small and Dim Moving Vehicles Detection in Satellite Videos

TL;DR

SDM-Car tackles the underrepresentation of small and dim moving vehicles in satellite-video datasets by introducing a large, dim-target–focused dataset collected from Luojia 3-01 imagery and a benchmark method that combines patch-wise image enhancement and SimAM attention within a two-stream DSFNet framework. The dataset comprises 99 videos with 16,830 annotated frames highlighting dim targets across diverse scenes, densities, and scales, enabling more robust evaluation of detection and tracking approaches. Empirical results show that training on SDM-Car boosts dim-vehicle detection performance and that the proposed enhancement plus attention method outperforms traditional baselines on SDM-Car, with notable gains in Recall, Precision, and $F_1$ Score. The work provides a valuable resource for remote sensing-based vehicle detection and offers practical improvements for reliable small-object detection under low-contrast conditions.

Abstract

Vehicle detection and tracking in satellite video is essential in remote sensing (RS) applications. However, upon the statistical analysis of existing datasets, we find that the dim vehicles with low radiation intensity and limited contrast against the background are rarely annotated, which leads to the poor effect of existing approaches in detecting moving vehicles under low radiation conditions. In this paper, we address the challenge by building a \textbf{S}mall and \textbf{D}im \textbf{M}oving Cars (SDM-Car) dataset with a multitude of annotations for dim vehicles in satellite videos, which is collected by the Luojia 3-01 satellite and comprises 99 high-quality videos. Furthermore, we propose a method based on image enhancement and attention mechanisms to improve the detection accuracy of dim vehicles, serving as a benchmark for evaluating the dataset. Finally, we assess the performance of several representative methods on SDM-Car and present insightful findings. The dataset is openly available at https://github.com/TanedaM/SDM-Car.

Figures (7)

• Figure 1: (a) Statistic analysis of the pixel intensity distribution of different datasets. The horizontal axis and vertical axis represent the pixel intensity of the center of the annotation box and the instance counts, respectively. (b) Visualization of annotations for different datasets. The green boxes are annotation boxes, while the red circles are unannotated vehicles. The interval between adjacent images is 15 frames. These unannotated vehicles are difficult to distinguish due to the low pixel intensity, but their motion information is perceptible to the human eye.
• Figure 2: Example images and their annotations in the SDM-Car dataset. Our dataset shows diversity in terms of instance density, road structure, and background, and contains a large number of dim instances.
• Figure 3: Distribution of instance density of in our dataset. We counted the number of instances in a 128$\times$128 patch as the approximation of the density. The histogram on the primary axis shows the number distribution of vehicles with different densities, and the curve chart on the secondary axis shows their accumulated proportions.
• Figure 4: Overall architecture of the proposed method. The video frames are initially enhanced and then fed into a 2-D static stream (superior) and a 3-D dynamic stream (inferior). The SimAM module is incorporated into the 2-D static stream. Features extracted by the aforementioned two streams are subsequently fused and input into the detection head, which yields the final detection outcomes.
• Figure 5: The structure of basicblock with SimAM. Bn: Batch Normalization. Relu: Rectified Linear Unit.