I joined the competitions recently. I found that when the features are L2 normalized, the feature descriptor track score is nearly 0. However, with this same feature, the video matching score is reasonable. Did I miss anything significant in the feature track evaluation?
For example, I used the baseline features of the test set. When the original feature is used, the descriptor track score is 0.45. When the feature is L2 normalized, the score becomes 0.01.
Without looking at the specifics of your implementation, it’s difficult for me to say what might be going wrong. If you haven’t already, it might be worth ensuring you can reproduce the example solution provided in the vsc2022 repo documentation.
If you’d like, you can also try to post more details and other competitors may be able to assist!