mmaaz60/mvits_for_class_agnostic_od

Train Open-world object detection on custom dataset

kojingying opened this issue · 1 comments

kojingying commented

Hi, thanks for the great work!

I have a few questions regarding to open world object detection:

  1. how to fine tune the the pertained model for open world object detection?
  2. how should I prepare the data?
  3. do you provide test/inference and evaluation script for open world object detection?
mmaaz60 commented

Hi @kojingying,

Thank you for your interest in our work. Please note the following,

We use MAVL (MDef-DETR) class-agnostic proposals to generate pseudo labels for unknowns in ORE. We did not finetune our model directly for Open-world task, instead we propose to improve ORE by generating unknown pseudo labels using our model.

We provide details of Open-world object detection application using ORE in Sections 4.2 and A.4 of our paper. The detailed instructions to generate unknown pseudo labels and the ORE pretrained models are available at applications/README.md.

I hope this will help. Please feel free to ask if you have any questions. Thanks