This file contains models trained for the experiments in Align and Distill: Unifying and Improving Domain Adaptive Object Detection. All models were trained on 8 NVIDIA Tesla V100s with PyTorch 1.13.1 and CUDA 11.6.
For compatibility with the config files we provide, download any models here to the models
directory in this repo.
Here we provide links to models from the Detectron2 model zoo that we use for pre-training. All baseline and oracle models in our experiments start with these weights. Note we did not train these models, but provide links here for convenience.
COCO pretrained Mask R-CNN w/ Res50-FPN backbone and 3x schedule: Config Model
Here we provide checkpoints for baselines trained on source-only data for each benchmark (Cityscapes → Foggy Cityscapes, Sim10k → Cityscapes, CFC Kenai → Channel).
We also use these checkpoints to initialize domain adaptive training, i.e. they also represent the end of the "burn-in" period.
Backbone | Download links |
---|---|
Resnet50 w/ FPN | Config Model |
VitDet-B | Config Model |
Backbone | Download links |
---|---|
Resnet50 w/ FPN | Config Model |
VitDet-B | Config Model |
Backbone | Download links |
---|---|
Resnet50 w/ FPN | Config Model |
VitDet-B | Config Model |
Here we provide the models trained using ALDI++.
Backbone | Download links |
---|---|
Resnet50 w/ FPN | Config Model Log |
VitDet-B | Config Model--TODO |
Backbone | Download links |
---|---|
Resnet50 w/ FPN | Config Model Log |
Backbone | Download links |
---|---|
Resnet50 w/ FPN | Config Model Log |