Golden Cudgel Network for Real-Time Semantic Segmentation

Guoyu Yang, Yuan Wang, Daming Shi*, Yanzhong Wang

🔥 What's New

• 2025/02/27: CVPR 2025 Accepted.

Golden Cudgel Network

Abstract

Recent real-time semantic segmentation models, whether single-branch or multi-branch, achieve good performance and speed. However, their speed is limited by multi-path blocks, and some depend on high-performance teacher models for training. To overcome these issues, we propose Golden Cudgel Network (GCNet). Specifically, GCNet uses vertical multi-convolutions and horizontal multi-paths for training, which are reparameterized into a single convolution for inference, optimizing both performance and speed. This design allows GCNet to self-enlarge during training and self-contract during inference, effectively becoming a "teacher model" without needing external ones. Experimental results show that GCNet outperforms existing state-of-the-art models in terms of performance and speed on the Cityscapes, CamVid, and Pascal VOC 2012 datasets.

Architecture

The overall architecture of GCNet. After feature flow into two branches, the upper branch corresponds to the semantic branch, while the lower branch corresponds to the detail branch. The orange box indicates that the first block within the GCBlocks has a stride of 2, while the remaining blocks have a stride of 1. The green box signifies that all GCBlocks maintain a stride of 1.

🛠️ Experiment

Environment

python==3.8.19
pytorch==1.12.1
torchvision==0.13.1
mmengine==0.10.2
mmcv==2.0.0
mmsegmentation==1.2.2

Install

Please refer to mmsegmentation for installation.

Dataset

RDRNet
├── mmsegmentation
├── figures
├── data
│   ├── cityscapes
│   │   ├── leftImg8bit
│   │   │   ├── train
│   │   │   ├── val
│   │   ├── gtFine
│   │   │   ├── train
│   │   │   ├── val
│   ├── CamVid
│   │   ├── train
│   │   ├── train_labels
│   │   ├── test
│   │   ├── test_labels
│   ├── VOCdevkit
│   │   ├── VOC2012
│   │   │   ├── JPEGImages
│   │   │   ├── SegmentationClass
│   │   │   ├── ImageSets
│   │   │   │   ├── Segmentation
├── gcnet-s_4xb3-120k_cityscapes-1024x1024.py
├── train.py
├── test.py
├── torch_speed.py

Cityscapes could be downloaded from here. Camvid could be downloaded from here. Pascal VOC 2012 could be downloaded from here.

Training

Single gpu for train:

CUDA_VISIBLE_DEVICES=0 python ./mmsegmentation/tools/train.py gcnet-s_4xb3-120k_cityscapes-1024x1024.py --work-dir ./weight/seg

Multiple gpus for train:

CUDA_VISIBLE_DEVICES=0,1,2,3 bash ./mmsegmentation/tools/dist_train.sh gcnet-s_4xb3-120k_cityscapes-1024x1024.py 4 --work-dir ./weight/seg

Train in pycharm: If you want to train in pycharm, you can run it in train.py.

see more details at mmsegmentation.

Testing

CUDA_VISIBLE_DEVICES=0 python ./mmsegmentation/tools/test.py gcnet-s_4xb3-120k_cityscapes-1024x1024.py ./weight/seg/gcnet_weight.pth

Test in pycharm: If you want to test in pycharm, you can run it in test.py.

see more details at mmsegmentation.

⚡ Results on Cityscapes

Method	Resolution	FPS	Params (M)	GFLOPs	ImageNet	val
BiSeNetV1	1024 ✕ 2048	116.8	13.3	118.0	✓	74.4
BiSeNetV2	1024 ✕ 2048	132.4	3.4	98.4	✗	73.6
DDRNet-23-Slim	1024 ✕ 2048	166.4	5.7	36.3	✗	76.3
DDRNet-23	1024 ✕ 2048	106.0	20.3	143.0	✗	78.0
PIDNet-S	1024 ✕ 2048	128.7	7.7	47.3	✗	76.4
PIDNet-M	1024 ✕ 2048	78.2	28.7	177.0	✗	78.2
PIDNet-L	1024 ✕ 2048	64.2	37.3	275.0	✗	78.8
SCTNet-S-Seg50	512 ✕ 1024	169.1	4.6	7.1	✗	71.0
SCTNet-S-Seg75	768 ✕ 1536	168.7	4.6	16.0	✗	74.7
SCTNet-B-Seg50	512 ✕ 1024	162.6	17.4	28.1	✗	75.0
SCTNet-B-Seg75	768 ✕ 1536	157.3	17.4	63.2	✗	78.5
SCTNet-B-Seg100	1024 ✕ 2048	117.0	17.4	112.3	✗	79.0
RDRNet-S	1024 ✕ 2048	182.6	7.3	43.4	✗	76.8
RDRNet-M	1024 ✕ 2048	102.8	26.0	162.0	✗	78.8
RDRNet-L	1024 ✕ 2048	76.1	36.9	238.0	✗	79.6
GCNet-S	1024 ✕ 2048	193.3	9.2	45.2	✗	76.9
GCNet-S (N=3)	1024 ✕ 2048	193.3	9.2	45.2	✗	77.3
GCNet-M	1024 ✕ 2048	105.0	34.2	178.0	✗	78.9
GCNet-L	1024 ✕ 2048	88.0	45.2	232.0	✗	79.6

The GPU used for benchmarking is the A100, and GCNet defaults to setting N to 2.

📑 Citations

If you find GCNet useful in your research, please consider citing:

@misc{yang2025goldencudgel,
      title={Golden Cudgel Network for Real-Time Semantic Segmentation}, 
      author={Guoyu Yang and Yuan Wang and Daming Shi and Yanzhong Wang},
      year={2025},
      eprint={2503.03325},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2503.03325}, 
}