Closed aparfenov2 closed 2 years ago
@JinkaiZheng Could you please answer this question?
@kantengri Hi~ Thank you for your attention! First, we only use 10-fold cross-validation on VehicleID dataset, because the VehicleID dose NOT provide official division documents about query and gallery. We random sample query and gallery in our implementation every inference time. In order to avoid randomness of the results, a trained model on VehicleID was tested 10 times and averaged, which is named "10-fold cross-validation". Secend, we does NOT use the test set during training. And we use the official division documents about query and gallery on VERI-Wild, do NOT adopt "10-fold cross-validation" on VERI-Wild, which means we only tset one time to get the results. Third, we got 96.4% (Rank-1) in MODEL_ZOO on small VERI-Wild test set (test_3000.txt). I think the result is high because the training set stays the same and the test set gets small. To take an extreme example, if your test set has only one ID, then when your first search result is correct, your result (Rank-1) is 100%. The larger the amount of data, the greater the uncontrollable factors, the poorer performance tends to be, and vice versa. Fourth, I can not understand what the "clustering" mean in your question. Could you please give me more details?
This issue is stale because it has been open for 30 days with no activity.
This issue was closed because it has been inactive for 14 days since being marked as stale.
Hi guys, thank you for this good framework. I tried clustering output of VERI-Wild Baseline, and got 98% matches on test set (test_3000.txt) from Veri-wild dataset. This looks too perfect. Can it be that you used that test set for cross-validation so the model over-fitted on that data? That could explain that high score. What data for cross-validation did you use here ?