Skip to content

README.md

Latest commit

 

History

History
executable file
·
57 lines (41 loc) · 2.62 KB

README.md

File metadata and controls

executable file
·
57 lines (41 loc) · 2.62 KB
Raw

The traditional downsampling (e.g., average pooling) can reduce the resolution. Nonetheless, it may suffer from the dropping of detailed information. In this work, we propose an Efficient Wavelet Attention (EWA), which injects the wavelet transform and a Mean GELU (MGELU) function. Building upon EWA, we present an effective Semantic-aware Wavelet Transformer (SWFormer), which is then employed for pyramid learning, including the feature hierarchy of Convolutional Neural Networks (CNNs) or Region of Interest (RoI) features. For the feature hierarchy, a Pyramid SWFormer (PSWFormer) incorporates SWFormer at each level to fit the bidirectional features.
For RoIs, a Recognition-Localization SWFormer (RLSWFormer) is inserted into the head to fit their features from all levels. The codes are built on the open-source toolkit detectron2.

(a) The vanilla self-attention learns the pixel interaction modeling. (b) The vanilla self-attention with the downsampling (Ds). (c) Efficient Global Semantic Attention (EGSA) with the downsampling. EGSA learns the global semantic interaction modeling. (d) Our Efficient Wavelet Attention (EWA) injects the wavelet transform and MGELU for the efficient interaction modeling. The wavelet transform is used to produce the enriched query, as well as the downsampled key and value. MGELU normalizes the downsampled value.

Installation

See INSTALL.md.

Quick Start

See GETTING_STARTED.md,

Model Zoo

We provide a large set of baseline results for download in the Dt2_SWFormer Model Zoo.

License

Dt2_SWFormer is released under the Apache 2.0 license.

Citing Dt2_SWFormer

If you use Dt2_SWFormer in your research or wish to refer to the baseline results published in the Model Zoo, please use the following BibTeX entry.

@misc{wu2019detectron2,
  author =       {Yuxin Wu and Alexander Kirillov and Francisco Massa and
                  Wan-Yen Lo and Ross Girshick},
  title =        {Detectron2},
  howpublished = {\url{https://github.com/facebookresearch/detectron2}},
  year =         {2019}
}

@article{Li2025SemanticawareWT,
  author={Li, Yang and Jiao, Licheng and Liu, Xu and Liu, Fang and Li, Lingling and Chen, Puhua},
  journal={IEEE Transactions on Multimedia}, 
  title={Semantic-aware Wavelet Transformer for Pyramid Learning Object Detection}, 
  year={2025}}