Fangyi-Chen
commented
1 year ago Hi,
Replacing P6 with the optimum from P1~6 is an idealized investigation - We do this to estimate the upper bound of potential when the problem is fully resolved. But clearly, it is not easy to fully solve. SQR only partially solved this, as you can see from the decrease in TP F rating and FP E rating in Table 6. So the AP improvement of SQR is still far away from the upper bound. However, as one of the reviewers claimed, "a first step to tackling this problem is worthwhile."
Regarding the distribution, maybe Table 3 in the paper can give you the expected information?
In addition, TP Fading Rate studies if a query could predict TP in early stages, but in the final stage, for some reason, the TP is worse than the former ones. As shown in fig 2 traffic light example, The TP is 'fading'. Its scientific description is detailed in the Sec.3. We also observe that more than half of the occurred cases are marginally triggered, i.e. the predictions from the triggered 1∼5 are only marginally better than the sixth. This is a further reason why the deformable DETR has those high rates - the results from the 5th and 6th stages are extremely visually close. In a dominated number of cases, the final stage is (one of) the best, after all, its mAP is the highest.
Best regards
lijain
Thanks for your work, when reading your article, I have a bit of a problem with page 4(AP grows from 44.5 AP to 51.7 (+7.2 AP)), replacing the last layer with intermediate results has greatly improved, but the experimental results have not improved much than felt. Does that mean that the optimal result may be distributed differently in different stages
In addition, can TP F Rate be more detailed?