Offline experiment

This a 1k dataset sampled randomly from actual project data. All images are from SEA local sellers so that the difficulty level is lower than the hard-dataset. The objective is to test out the difference in the 3 following models to the best extent: baseline Siamese network, Vanilla MoCo, Enhance MoCo(our version)

Siamese v.s. Vanilla MoCo v.s. Enhanced MoCo

There are 2 metrics that we care about the most for the actual downstream applications. The total amount of identical items that we could capture through vision: reflected by Coverage. Among these recalled items, how many of them are true positive matches: reflected by Precision.

  • Both MoCo models performs better baseline in both coverage and precision
  • Enhanced MoCo consistently performance best for 2 reasons: fine-tune on our dataset and better robustness to disturbance
  • From this point onwards, we only consider Enhanced MoCo for official launching