Which fold dose the training result 'ultradet.pth' corresponds to?

[imzhangyd]

Hi, @xjtulyc . I noticed in your paper that you mentioned, 'We perform 4-fold cross-validation and report the mean values and standard errors on the test set to mitigate fluctuations.' I understand that you performed four-fold cross-validation on the training data, meaning you trained the model four times and obtained four models. I would like to know which fold the training result 'ultradet.pth' corresponds to.

[xjtulyc]

Apologies for the confusion. In the context of cross-validation being used for parameter tuning, the checkpoints provided at the end are trained on the entire training set rather than a specific fold. The purpose of cross-validation is to evaluate and tune the model's performance on different subsets of the data, but the final checkpoints are typically trained on the full training set for better generalization.

[xjtulyc]

Hello,

Thank you for your inquiry. In our paper, the results presented in Table 1 reflect the performance of UltraDet compared to other real-time detectors. We conducted 4-fold cross-validation, and the reported results are the mean values along with standard errors derived from these folds on the test set to reduce any potential fluctuations.

Specifically, UltraDet demonstrates superior performance in terms of precision and false positive (FP) rates. For example, the Precision at 90% recall (Pr90) of UltraDet reaches 90.8%, which is an absolute improvement of 5.4% over the next best competitor, PTSEFormer. Additionally, the false positives per sequence per minute (FP90) for UltraDet are reduced by approximately 50% compared to PTSEFormer, standing at 5.7 seq/min.

While CVA-Net shows comparable AP50 metrics to our method, UltraDet significantly improves upon both precision and FP rates.

Best regards, Youcheng Li

发件人：QiuZhichao0409 @.> 发送日期：2024-09-12 22:44:31 收件人：HaojunYu1998/UltraDet @.> 抄送人：Youcheng Li @.>,State change @.> 主题：Re: [HaojunYu1998/UltraDet] Which fold dose the training result 'ultradet.pth' corresponds to? (Issue #6)

Hello, may I ask if the results in the paper are the average of the outcomes from the four models that were trained individually through fourfold cross-validation on the test set? — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you modified the open/close state.Message ID: @.***>