Research map / Detection and segmentation / CNN features

Golden Cudgel Network for Real-Time Semantic Segmentation

CVPR 2025 · Detection and segmentation › CNN features · approach: Context modeling for semantic segmentation · 41 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes Golden Cudgel Network (GCNet), a single-branch-at-inference two-branch segmenter whose training-time multi-conv, multi-path blocks are reparameterized into single 3x3 convolutions, removing the need for an external teacher.

Problem

Real-time semantic segmentation models are slowed by multi-path blocks at inference, and some rely on high-performance teacher models for training, increasing training cost.

Approach

GCBlock uses vertical multi-convolutions and horizontal multi-paths during training that are fully reparameterized into a single 3x3 convolution at inference, so the network self-enlarges for training and self-contracts for inference, acting as its own teacher without distillation.

Key results

• GCNet-S reaches 77.3% mIoU at 193.3 FPS on Cityscapes val without ImageNet pretraining (A100).
• GCNet-L achieves 79.6% mIoU on Cityscapes val, the highest among the compared models in the table.
• On CamVid, GCNet-S gets 76.6% mIoU at 210.6 FPS; on VOC 2012, GCNet-L reaches 58.8% mIoU.

Measured effects reported in the paper

Surprising findings

• Eliminating the first 1x1 convolution in the bottleneck was needed because reparameterization losslessly fusing it with the 3x3 conv was hindered by small trained parameter values.
• Increasing Path3x3_1x1 beyond 4-5 blocks degraded performance at N=10 while training time rose by over 80%.

Limitations

• Higher parameter count and GFLOPs during training and inference than many lightweight baselines, limiting use on storage-constrained devices.
• Performance-speed comparisons rely on reimplementation of baselines on A100 hardware.

Experimental setup

Datasets: Cityscapes CamVid Pascal VOC 2012

Benchmarks: Cityscapes CamVid Pascal VOC 2012

Metrics: mIoU FPS GFLOPs Params

Baselines: ICNet Fast-SCNN CGNet BiSeNetV1 BiSeNetV2 STDC HyperSeg PP-LiteSeg DDRNet PIDNet SCTNet

Builds on: DDRNet PIDNet SCTNet BiSeNet RepVGG

Closest papers by meaning

• SCTNet: Single-Branch CNN with Transformer Semantic Information for Real-Time Segmentation (AAAI 2024 · similarity 0.85)
• PIDNet: A Real-Time Semantic Segmentation Network Inspired by PID Controllers (CVPR 2023 · similarity 0.85)
• Graph-Guided Architecture Search for Real-Time Semantic Segmentation (CVPR 2020 · similarity 0.85)
• Patch Proposal Network for Fast Semantic Segmentation of High-Resolution Images (AAAI 2020 · similarity 0.85)
• Fast Neural Architecture Search of Compact Semantic Segmentation Models via Auxiliary Cells (CVPR 2019 · similarity 0.84)
• Rethinking BiSeNet for Real-Time Semantic Segmentation (CVPR 2021 · similarity 0.84)