Thanks for the super clean implementation!
I noticed a discrepancy in the code and what is written in the SiamRPN paper.
Specifically you have: penalty = np.exp(-(change_ratio change_sz - 1) self.cfg.penalty_k)
but the paper seems to have penalty = np.exp(-(change_ratio change_sz) self.cfg.penalty_k), where did the -1 come from? Also in the paper they write: "After these operations, the top K proposals are re-ranked after multiply the classification score by the temporal penalty." However in your code there is no "top-K" you re-rank all proposals. The performance on VOT2018 seems to agree with SiamRPN nevertheless.
Hi,
Thanks for the super clean implementation! I noticed a discrepancy in the code and what is written in the SiamRPN paper. Specifically you have: penalty = np.exp(-(change_ratio change_sz - 1) self.cfg.penalty_k) but the paper seems to have penalty = np.exp(-(change_ratio change_sz) self.cfg.penalty_k), where did the -1 come from? Also in the paper they write: "After these operations, the top K proposals are re-ranked after multiply the classification score by the temporal penalty." However in your code there is no "top-K" you re-rank all proposals. The performance on VOT2018 seems to agree with SiamRPN nevertheless.
thanks, Ross