Label Relationships#
Beyond merely identifying the presence of labels, understanding the relationships between labels can be pivotal, especially in complex classification tasks. These relationships delve into the intricacies of how labels interact and influence each other, which is critical in:
Single election: Each image/frame is assigned to one label. In computer vision literature, it is called binary classification.
Multi selection: Each image/frame can be tagged with multiple labels simultaneously. In computer vision literature, it is called multiclass or multinomial classification.
Mix: Labels are organized in a hierarchical structure, reflecting the relationships and specificity among them.
More details can be found on the following page. Here, you will read about the notions of label exclusivity and label specificity.
Label exclusivity#
Label exclusivity is a principle where the assignment of one label to a dataset precludes the use of another label for the same dataset. This concept is crucial in classification tasks, where it’s necessary to delineate clear, mutually exclusive categories.
Within the Intel® Geti™ platform, label exclusivity is managed using a LabelGroup. This organizational tool groups labels to define their relationships — whether they are exclusive or can coexist.
Consider the task of categorizing ID photos based on whether the subject is wearing glasses. This scenario presents two exclusive categories: “With Glasses” and “Without Glasses”. In such cases, a subject cannot simultaneously belong to both categories, illustrating a perfect example of label exclusivity.
The platform configures this by placing all relevant labels into a single LabelGroup. This setup defines the task as a Single selection classification problem, where each piece of data can be classified into one and only one of the predefined categories.
Conversely, Multi selection classification recognizes that a single dataset can simultaneously belong to multiple categories. This approach is necessary for scenarios where categories are not mutually exclusive.
Imagine a project aimed at classifying movie genres based on their posters. Given that movies often span multiple genres, this task requires a classification system where each genre is not exclusive to others. For instance, a movie could be a “Comedy” and a “Drama”.
In practice, in the platform, this necessitates creating a separate LabelGroup for each genre. Therefore, if there are ten genres, there would be ten LabelGroups, each holding its respective label. This setup accommodates the multifaceted nature of movie genres, allowing for a more nuanced and accurate classification.
Label specificity#
Label specificity in the Intel® Geti™ platform encapsulates the hierarchical or “is-a” relationship between two or more labels. This feature is crucial for tasks where understanding the granularity of classification can significantly enhance the model’s accuracy and utility.
Consider a car classification project aiming to identify both the make (brand) and model (series) of cars. The “make” of a car (e.g., Honda, Toyota, Ford) represents a broader category, while the “model” (e.g., Civic for Honda, Yaris for Toyota) signifies a more specific classification within that make. This setup exemplifies a hierarchical classification scenario, where the make is the parent label, and the model is the child label.
Identifying the model of a car intrinsically provides information about its make, offering a two-tiered insight from a single classification act. This approach is not just about the end goal of label identification; it enriches the dataset with structured and detailed information, facilitating better algorithmic convergence and performance. Furthermore, in cases where the model may be challenging to discern while the make is more apparent, providing at least the make as information aids the deep learning algorithm significantly better than providing no information.
The Intel® Geti™ platform employs a LabelTree to represent these hierarchical relationships visually and structurally, ensuring clarity in the organization of labels.
Beyond Single Tasks: Inter-Task Label Hierarchy
Notably, label hierarchy in the Intel® Geti™ platform isn’t confined to a singular task. It extends across task chains, from Detection to Classification, introducing a nuanced layer of inter-task relationships. This mechanism enhances user experience and label validation, allowing for automated label assignments that streamline the annotation process.
Considerations and Limitations
Task Chain Hierarchy: Currently, within a task chain (e.g., Detection -> Classification), only a single label is supported for the detection phase, with subsequent classification labels acting as children. This setup simplifies the process of assigning detection labels to bounding boxes, further facilitating the classification task.
Current Restrictions: It’s important to note that the principles of label exclusivity and hierarchy do not apply to non-classification single-task projects. For instance, detection tasks are limited to a single label per bounding box, segmentation tasks to a single label per pixel. For anomaly type tasks, it is a multiclass classification (anomaly classification) or with the variation of segmentation (anomaly segmentation) or detection (anomaly detection).