Awesome
SVAM: Saliency-guided Visual Attention Modeling (To Appear at RSS 2022)
Pointers
SVAM-Net Model
- Jointly accommodate bottom-up and top-down learning in two branches sharing the same encoding layers
- Incorporates four spatial attention modules (SAMs) along these learning pathways
- Exploits coarse-level and fine-level semantic features for SOD at four stages of abstractions
- The bottom-up pipeline (SVAM-Net_Light) performs abstract saliency prediction at fast rates
- The top-down pipeline ensures fine-grained saliency estimation by aresidual refinement module (RRM)
- Pretrained weights can be downloaded from this Google-Drive link
SVAM-Net Features
- Provides SOTA performance for SOD on underwater imagery
- Exhibits significantly better generalization performance than existing solutions
- Achieves fast end-to-end inference
- The end-to-end SVAM-Net : 20.07 FPS in GTX-1080, 4.5 FPS on Jetson Xavier
- Decoupled SVAM-Net_Light: 86.15 FPS in GTX-1080, 21.77 FPS on Jetson Xavier
USOD Dataset
Bibliography entry:
@inproceedings{islam2022svam,
author={Islam, Md Jahidul and Wang, Ruobing and Sattar, Junaed},
title={{SVAM: Saliency-guided Visual Attention Modeling
by Autonomous Underwater Robots}},
booktitle={Robotics: Science and Systems (RSS)},
year={2022},
address={NY, USA}
}
Acknowledgements