为何ILSVRC'12分类任务未设置通用person（人物）类？

Great question! The absence of a broad "person/human" class in the ILSVRC'12 dataset—while it includes specific human-related categories like scuba diver, groom, and baseball player—stems from intentional design choices tied to the dataset's goals and constraints:

• Prioritizing fine-grained recognition
  ILSVRC'12 was built to push the limits of visual classification by focusing on distinguishing between highly similar subjects. A generic "person" class would be too broad to challenge models to learn subtle, meaningful differences (like the specialized attire of a scuba diver vs. the uniform of a baseball player). The dataset prioritizes these specific, distinct categories to advance fine-grained recognition capabilities.

• Ensuring annotation clarity
  Labeling images with a general "person" class would create ambiguity for annotators. For example, an image of a groom could technically fit both "groom" and "person," which conflicts with ILSVRC's single-label classification task. By using mutually exclusive, specific classes, curators ensured consistent, unambiguous labeling—critical for training robust, reliable models.

• Rooted in WordNet's hierarchy
  ILSVRC draws from ImageNet, which maps to WordNet's semantic hierarchy. While WordNet does have a general "person" synset, the ILSVRC'12 curators selected 1000 diverse classes that avoid overlapping umbrella terms. They opted for specific categories to cover a wide range of visual concepts without redundancy, ensuring the dataset offers maximum diversity in distinct entities.

• Task-specific utility
  The core ILSVRC classification task focuses on identifying the primary subject of an image. A general "person" class wouldn't align with this goal, as it doesn't capture the specific context or role of the human in the image. The fine-grained human classes, however, teach models to recognize context-specific features—far more useful for applications that require precise identification rather than just detecting a human presence.

内容的提问来源于stack exchange，提问作者Marph