Reproduced results different from paper

[ncoop57]

Some of the reproduced results are different from the paper, specifically tango_vis and tango_txt can vary in performance by 5-10%. However, tango_comb only varies around 1%.

tango_vis differences may be due to stochasticity in the kmeans cluster assignment during inference.

Differences in tango_txt might be explained due to differences in which frames are selected to have text extracted from in current implementation. However, this doesn't seem like it would cause the stark difference.

[ncoop57]

Issue does not appear to be due to stochasticity as different seeds have been tried and no change in performance has occurred. I was able to reproduce the original paper's results using an older version of the code. Therefore, the issue might be due to some change in code or data. Investigations are still underway.