search

ucbdrive / few-shot-object-detection

Implementations of few-shot object detection benchmarks
Apache License 2.0
1.12k stars 225 forks source link

Novel weights LVIS #162

Open s0tt opened 2 years ago

s0tt commented 2 years ago

Hi there, i was wondering about the novel weights training step of the last layer specifically for the LVIS dataset. How does this justify to be a few-shot setting when showing the novel shots twice. As far as i understand the scripts:

  1. Novel weight training: All novel (rare classes from train_rare.json) annotations of train set are used for training
  2. Fine-tuning: Again max. 10 shots of each class used for balanced fine-tuning

Isn't this invalidating the few-shot setting to some extent ? Maybe i also have a misunderstand here. Thanks!