This repo is for updating news related to our NeurIPS 2021 paper "Few-Shot Segmentation via Cycle-Consistent Transformer".
In CyCTR, We design a novel Cycle-Consistent Transformer (CyCTR) module for few-shot segmentation. CyCTR aggregates pixel-wise support (i.e., the few-shot exemplars) features into query (i.e., the sample to be segmented) ones through a transformer. As there may exist unexpected irrelevant pixel-level support features, directly performing cross-attention may aggregate these features from support to query and bias the query features. Our proposed cycle-consistent attention mechanism can effectively filter out possible harmful support features and encourage query features to attend to the most informative pixels from support images.
2021/12: To be appeared in NeurIPS 2021 (View NeurIPS Page), welcome to discussion！
- PaddlePaddle
The official PaddlePaddle implementation is in preparation.
- PyTorch
Thanks YanFangCS for the Pytorch Implemnetation.