SROSDA (ICCV 2021)
implementation of the ICCV work Towards Novel Target Discovery Through Open-Set Domain Adaptation [Paper].
⚡ Please check the extension journal work "Interpretable Novel Target Discovery Through Open-Set Domain Adaptation" (XSR-OSDA).
Data Preparation
| Dataset | Domain | Role | #Images | #Attributes | #Classes |
|---|---|---|---|---|---|
| D2AwA | A P R |
source / target | 9,343 / 16,306 3,441 / 5,760 5,251 / 10,047 |
85 | 10 / 17 |
| I2AwA | I Aw |
source / target | 2,970 / 37,322 | 85 | 40 / 50 |
(1) To extract pre-trained ResNet-50 features, check script:
./data/N2AwA/features/extract_resnet_features.ipynb(2) Collect attributes for all samples based on their labels, check script:
./data/N2AwA/attributes/check_N2AwA_data.ipynbDependencies
- Python 3.6
- Pytorch 1.1
Training
Step 1: Initialization clustering on target data (Seen/Unseen Initialization)
./data/N2AwA/refine_cluster-samples.ipynbNote: Or use our clustering initialization results ./data/N2AwA/ directly.
Step 2: Train with the initialized clustering and pseudo labels on the extracted features.
python main.pyEvaluation
- Open-set Domain Adaptation Task
$OS^*$ : class-wise average accuracy on the seen categories.
$OS^\diamond$ : class-wise average accuracy on the unseen categories correctly classified as "unknown".
$OS$ :$\frac{OS^* \times C_{shr} + OS^\diamond}{C_{shr} + 1}$
$C_{shr}$ is the number of shared categories between the source and target domains.
- Semantic-Recovery Open-Set Domain Adaptation Task
$S$ : class-wise average accuracy on shared classes
$U$ : class-wise average accuracy on unknown classes
$H = \frac{2 \times S \times U}{ S + U}$
Citation
If you think this work is interesting, please cite:
@InProceedings{Jing_2021_ICCV,
author = {Jing, Taotao and Liu, Hongfu and Ding, Zhengming},
title = {Towards Novel Target Discovery Through Open-Set Domain Adaptation},
booktitle = {International Conference on Computer Vision (ICCV)},
year = {2021}
}
Contact
If you have any questions about this work, feel free to contact