We propose a novel feature adaptation method named PuzzleBoost, to address the transferring problem between two domains by introducing a pretext task of solving image jigsaw puzzles. PuzzleBoost has a concise structure that facilitates its seamless integration into existing feature embedding-based approaches.
Full-shot:
Few-shot:
This project was built under Python 3.8, with package versions are as follows:
torch == 1.13.1+cu117
torchvision == 0.14.1+cu117
faiss-gpu == 1.7.2
click == 8.1.7
numpy == 1.24.3
opencv-python == 4.8.0.74
pillow == 9.5.0
The experimental data was obtained by running the code on a single V100 GPU.
Our method is committed to improving the effectiveness of existingfeature embedding-based methods, so our code is modified basedon the code of other authors.
We are heavily borrowed codes from PatchCore, PaDiM and SPADE