This paper discusses the use of low-level approaches in computer vision, focusing on three inter-related aspects: saliency, scale selection, and content description. It argues that these aspects are intrinsically linked and proposes a multiscale algorithm for selecting salient image regions, which is applied to matching tasks like tracking, object recognition, and image retrieval. The paper begins by highlighting the importance of meaningful image descriptions in computer vision tasks, noting that the Human Visual System (HVS) processes images through pre-attentive and attentive stages. Pre-attentive processing detects 'pop-out' features, while attentive processing identifies relationships and groups features. However, achieving robust vision systems based on this model has proven challenging, particularly in grouping and matching stages. The paper discusses the limitations of purely local approaches, such as Schiele's method, which relies on local appearance descriptors but lacks position information and may fail in complex scenes. It introduces a novel algorithm that creates a hierarchy of salient regions across feature space and scale, addressing issues like scale and robustness. The paper also explores the concept of visual saliency, using local complexity as a measure, and introduces a method for assessing saliency across feature space and scale. It demonstrates the algorithm's operation on various examples, emphasizing its robustness to scale and its ability to handle complex scenes without prior assumptions. The paper concludes with a discussion on the potential improvements of the method, such as integrating Kalman or Condensation trackers.This paper discusses the use of low-level approaches in computer vision, focusing on three inter-related aspects: saliency, scale selection, and content description. It argues that these aspects are intrinsically linked and proposes a multiscale algorithm for selecting salient image regions, which is applied to matching tasks like tracking, object recognition, and image retrieval. The paper begins by highlighting the importance of meaningful image descriptions in computer vision tasks, noting that the Human Visual System (HVS) processes images through pre-attentive and attentive stages. Pre-attentive processing detects 'pop-out' features, while attentive processing identifies relationships and groups features. However, achieving robust vision systems based on this model has proven challenging, particularly in grouping and matching stages. The paper discusses the limitations of purely local approaches, such as Schiele's method, which relies on local appearance descriptors but lacks position information and may fail in complex scenes. It introduces a novel algorithm that creates a hierarchy of salient regions across feature space and scale, addressing issues like scale and robustness. The paper also explores the concept of visual saliency, using local complexity as a measure, and introduces a method for assessing saliency across feature space and scale. It demonstrates the algorithm's operation on various examples, emphasizing its robustness to scale and its ability to handle complex scenes without prior assumptions. The paper concludes with a discussion on the potential improvements of the method, such as integrating Kalman or Condensation trackers.