This work aims to produce enriched semantic embedding by including the knowledgesources from text and image for action classes.

1. Scraping images for each action class from Google Imageg Source. Check codes in the google images folder
2. Downloaed images according to the given keywords in the sub-folder downloads
3. Using a pre-defined RESNET101 as feature extractor. Reference
4. Averaging image representations for each action class

[1] Narayan, S., Gupta, A., Khan, F. S., Snoek, C. G., & Shao, L. (2020). Latent embedding feedback and discriminative features for zero-shot classification. In Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XXII 16 (pp. 479-495). Springer International Publishing. github repo.