Building on our previous work, initial experiments using frozen weights achieved an F1 score of ~65% and a subset accuracy of 15%. Before committing to a more complex fine-tuning setup, we need to conduct three sanity checks to validate our current approach.
The goal here is to verify whether the structural distances between movie posters are accurately preserved within the extracted feature space. We will add a random noise to the original poster and compare the distance between the original projection and the projection after noise injection.
| Noise Level | Distance from Original Projection (± 2 std) |
|---|---|
| $\sigma = 0.1$ | 6.72 (± 2.22) |
| $\sigma = 0.2$ | 8.65 (± 3.47) |
| $\sigma = 0.5$ | 15.84 (± 11.85) |
We apply the same approach to a larger dataset from an open Kaggle challenge (hosted by neha1703).
If these checks yield positive results, the next step will be to apply standard fine-tuning methods and compare the performance against Vision Transformers and low-rank approximations.