Setup for OSCD and compatibility to SeCo repository

[pvmilk]

Nice work on the paper and thank you for making the source code/ pretrained model public.

I am trying to recreate the result on the paper on change detection. The paper mentioned that the model architecture of this work on OSCD data followed SeCo paper. I would like to know whether the pretrained models are compatible with the main_oscd.py on SeCo repository.

• It seems that moco2_module.py of SeCo and CACo are difference. Is it meant that to use CACo pretrained model, I simple have to load the model using moco2_module.py from CACO, and copy it backbone with the following code and use the SeCo main_oscd.py? Or since we only use MocoV2 code to load the backbone, can I use moco2_module.py from SeCo to load CACo pretrained model also?

  model = MocoV2.load_from_checkpoint(model_path)
  backbone = deepcopy(model.encoder_q)

• SeCo main_oscd.py seems to report only result on resnet-18, but in your paper both on resnet-18 and resnet-50. Any modification on the SeCo main_oscd.py source code?

Thank you.

[utkarshmall13]

Hi, yes the model should be compatible with the main_oscd.py. You can also check the models/checkpoints without the lightning wrapper here and refer to this notebook for model loading instructions. We did not make any changes to the inference and testing code of main_oscd.py. However since the training data is different between seco and CACo, we recommend using the retrained seco checkpoint for a fair comparison.

[pvmilk]

@utkarshmall13 Thanks for your reply.

I did try to use main_oscd.py with CACo pretrained weight (using the method provided in the notebook without pl wrapper). However, I cannot recreate the result (actually the result seems to be worse than seco; both from this paper and the original paper).

I would like to know a few configurations.

• main_oscd.py fixed the seeding to 42. In your experiment, is it the same?
• main_oscd.py by default, trained for 100 epochs and use the last model. Does it same in your case?
• main_oscd.py didn't really provide the test code. Does the reported values in the paper are from the validation during training? or do you have a separate code for testing?