Evaluation for few-shot object detection about imted HOT 9 CLOSED

@nil123532 Hi, to be more clear, trainvalno5k.json is used to generate the 10/30 shot taining json and 5k.json is for evaluation. We report the performance of novel class.

from imted.

@RafLaf Hi, for the first question, yes. For the second question, you may refer to link and link. Most of previous work conduct experiments on seed 0 only.

from imted.

Hi，we use this json fro evaluation. We don't count the number of images. If you're interested, you can check the json mentioned above.

from imted.

So to evaluate the FSOD 9 seeds you used the trainvalno5k.json?

from imted.

We report the performance of the model trained on seed 0, following previous work. The evaluation is conducted on the trainvalno5k.json.

from imted.

Hi @LiewFeng,
Thanks for your answer! May I ask a couple more questions?
-Is the "seed 0" the list of classes (in several shots settings) that I see when clicking on the link of the original issue post after "seed 9" (see screenshot below)?
-Could you provide me with a paper that details completely the evaluation protocol. I could not find a proper document that highlights it all! I am particularly interested in knowing how many seeds and which seeds should be used as "few training examples".

Thanks !

from imted.

Thanks!

from imted.

Closed due to inactivity. Feel free to reopen it if necessary.

from imted.

Hi, thank you for the response on this thread.
The config file for finetuning accepts only 1 json for train but in the previous posts there are multiple json files for each class.
Can you please let me know the steps how you combined these files ?

from imted.