Open Kiainio opened 8 months ago
Hello, thank you for your interest in our work. We are excited to announce that our code, datasets, corresponding videos, extracted features, weights, and more will be made open source within the next two weeks! Regarding how we gauged the GPT's confidence in the options, our approach was straightforward: we simply assigned a confidence of 1 to the option GPT deemed correct and 0 to the others. We then merged this confidence with the output from the VQA model by multiplying it by a coefficient. Although this method is quite rudimentary, it still led to a notable improvement in performance. Nevertheless, we recognize this as a weakness and are actively working to refine this aspect in our ongoing research. Stay tuned for our future work!
Hi, I read from your paper that you require the confidence distribution from GPT to do commonsense reasoning. I wonder if you could give more details about the usage, e.g. did you require from OpenAI API directly? If not, how to acquire the GPT result? This would help me to reproduce your results, since no code and only dataset is provided.