search

evalcrafter / EvalCrafter

[CVPR 2024] EvalCrafter: Benchmarking and Evaluating Large Video Generation Models
http://evalcrafter.github.io
109 stars 7 forks source link

About using WordNet to divide prompt's metaclasses. #7

Open xiefan233 opened 5 months ago

xiefan233 commented 5 months ago

Hi, I would like to know how you categorized prompt into four categories: human, animal, object, and landscape using WordNet. I'm sorry that I didn't understand your description of Fig. 2 (c) very well.

For example, what is the relationship between the noun type and the four metaclasses in Fig. 2 (c)?

Yaofang-Liu commented 5 months ago

Hi, thanks for raising such a good question. The distribution shown in Fig. 2 (c) is derived from real-world user data, categorizing noun types into meta classes using WordNet. Artifacts, humans, animals, and locations are the main focuses, while communication, attribute, and cognition words are excluded. The division into four meta-subject classes—human, animal, object, and landscape—is based on this analysis. This classification ensures diversity and representativeness in the benchmark dataset. Please feel free to ask if you have any further questions.

Yaofang-Liu commented 5 months ago

BTW, the prompts are categorized using WordNet synsets.

rongpan123 commented 3 months ago

Hi, regarding prompt classification, the synonyms of each noun will correspond to many categories. How should I classify them?