When Domain Generalization
meets Generalized Category Discovery:
An Adaptive Task-Arithmetic Driven Approach

CVPR 2025
Vaibhav Rathore1,  Shubhranil B1,  Saikat Dutta1,2*,  Sarthak Mehrotra1*,  Zsolt Kira3,  Biplab Banerjee1
1 IIT Bombay,   2 IITB-Monash Research Academy,   3 Georgia Institute of Technology
* Equal contribution
arXiv Code Poster
Training
Inference Seen Inference Unseen
Full Span Image
Model Animation
Model Animation
0.00
0.00
0.00
0.00
Ties Merging
Fisher Merging
Task Arithmetic
DG²CD-Net
ViT GCD CDAD-Net DG2CD-Net DG2CD-Net LoRA Legend

t-SNE visualization of clustered embeddings from ViT baseline to our final approach DG2CD-Net* (LoRA). Categories are color-coded as shown in the legend.

BibTeX Copied! 

@inproceedings{Rathore2025DGCD,
    author    = {Rathore, Vaibhav and B, Shubhranil and Dutta, Saikat and Mehrotra, Sarthak and Kira, Zsolt and Banerjee, Biplab},
    title     = {When Domain Generalization meets Generalized Category Discovery: An Adaptive Task-Arithmetic Driven Approach},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
    year      = {2025}
  }