search

zhmiao / OpenLongTailRecognition-OLTR

Pytorch implementation for "Large-Scale Long-Tailed Recognition in an Open World" (CVPR 2019 ORAL)
BSD 3-Clause "New" or "Revised" License
839 stars 128 forks source link

Missing very important baseline #44

Closed twangnh closed 4 years ago

twangnh commented 4 years ago

Hi! Thanks for sharing your work, I'm wondering have you tried the baseline of class balanced sampling and inverse frequecy sampling when train the model?

zhmiao commented 4 years ago

@twangnh Thanks for asking. Yes, we have tried upsampling methods although we did not report it due to space limit. We also incorporated upsampling (class-aware samping) into our code. You are very welcome to single out this module and test it out.

tonysy commented 4 years ago

I think another important baseline is just use the weight normalization or cosine normalization without memory mechanism.

liuziwei7 commented 4 years ago

All the comparison methods in Table 3 of the main paper (https://arxiv.org/pdf/1904.05160.pdf) are equipped with class-aware/balanced sampling, as elaborated in Section 4 "Competing Methods". You are more than welcome to try other variants of ablation study.