BEVANet: Bilateral Efficient Visual Attention Network for Real-Time Semantic Segmentation (ICIP 2025 spotlight)

Ping-Mao Huang, I-Tien Chao, Ping-Chia Huang, Jia-Wei Liao, Yung-Yu Chuang
National Taiwan University

Model Description

Task: Real-Time Semantic Segmentation
Dataset: Cityscapes
Model: BEVANet

BEVANet semantic segmentation model trained on Cityscapes dataset. Achieves 81.0% mIoU with 32.8 FPS on RTX3090.

Performance

mIoU: 81.0%
FPS: 32.8 (RTX3090)
Parameters: 58.6M

Usage

from models.BEVANet import BEVANet_SEG

# Load model from specific branch
model = BEVANet_SEG.from_pretrained(
    "maomao0819/BEVANet", 
    revision="main"
)

# For main branch (no revision needed)
# model = BEVANet_SEG.from_pretrained("maomao0819/BEVANet")

All Model Variants

Model	Branch	Dataset	Task	Performance	FPS	Parameters
BEVANet-S	`imagenet-bevanet-s`	ImageNet	Classification	71.1% Top-1	198.6	16.3M
BEVANet	`imagenet-bevanet`	ImageNet	Classification	76.3% Top-1	82.3	57.4M
BEVANet-S	`cityscapes-bevanet-s`	Cityscapes	Segmentation	78.2% mIoU	70.0	15.2M
BEVANet	`main`	Cityscapes	Segmentation	81.0% mIoU	32.8	58.6M
BEVANet-S	`camvid-bevanet-s`	CamVid	Segmentation	83.1% mIoU	79.4	15.2M
BEVANet	`ade20k-bevanet`	ADE20K	Segmentation	39.8% mIoU	73.3	58.9M

Citation

If you use this model, please cite:

@inproceedings{huang2025bevanet,
  title={Bevanet: Bilateral Efficient Visual Attention Network for Real-Time Semantic Segmentation},
  author={Huang, Ping-Mao and Chao, I-Tien and Huang, Ping-Chia and Liao, Jia-Wei and Chuang, Yung-Yu},
  booktitle={2025 IEEE International Conference on Image Processing (ICIP)},
  pages={2778--2783},
  year={2025},
  organization={IEEE}
}