Research map / Detection and segmentation / CNN features
Structural Pruning via Spatial-aware Information Redundancy for Semantic Segmentation
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 SIRFP, a structured filter pruning method for semantic segmentation that scores channel redundancy with spatial-aware JS-divergence on full-resolution feature maps (EMA-accumulated), frames selection as a maximum edge weight clique problem, and solves it with a fast greedy algorithm (EHGP).
Problem
Existing filter pruning methods designed for image classification ignore spatial information, so they perform suboptimally when pruning location-sensitive segmentation (and detection) networks, and exact/heuristic solutions to the underlying clique formulation are too costly for network pruning.
Approach
Model channel selection as a maximum edge weight clique problem over a graph whose edge weights reflect spatial-aware (un-pooled feature map) redundancy measured via JS divergence and accumulated with EMA during training, then prune greedily by repeatedly removing the channel with the smallest total edge weight.
Key results
- On Cityscapes with DeepLabv3-ResNet50 at 60% FLOPs reduction, SIRFP reaches 81.3 mIoU vs 80.9 for DCFP and 81.6 unpruned.
- On ImageNet with ResNet-50 at 1.0G FLOPs, SIRFP achieves 74.67 top-1 vs 73.30 for CHIP.
- On COCO2017 SSD-ResNet50 at 50% FLOPs reduction, SIRFP reaches 26.6 AP vs 25.9 for CHEX-1.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use spatial-aware redundancy pruning rather than DCFP | Deeplabv3-ResNet50 · Cityscapes validation, 60% pruning | mIoU: 80.9 -> 81.3 | Helped |
| Use spatial-aware redundancy pruning rather than DCFP | Deeplabv3-ResNet50 · ADE20K, 60% FLOPs reduction | mIoU: 44.8 -> 44.9 | Helped |
| Use spatial-aware redundancy pruning rather than DCFP | Deeplabv3-ResNet50 · COCO stuff-10K, 60% FLOPs reduction | mIoU: 34.3 -> 34.8 | Helped |
| Use spatial-aware redundancy pruning rather than CHEX | SSD-ResNet50 · COCO2017 validation, 50% FLOPs reduction | AP: 25.9 -> 26.6 | Helped |
| Use spatial-aware redundancy pruning instead of CafeNet | ResNet-50 · ImageNet validation, 2.0G FLOPs | Top-1 accuracy: 76.90 -> 77.35 | Helped |
| Use full-resolution rather than 1×1 feature maps for redundancy (ablation) | SIRFP · Cityscapes, Deeplabv3-ResNet50 | mIoU: 80.4 -> 81.3 | Helped |
| Use full-resolution rather than quarter-resolution feature maps for redundancy (ablation) | SIRFP · Cityscapes, Deeplabv3-ResNet50 | mIoU: 80.6 -> 81.3 | Helped |
| Use full-resolution rather than half-resolution feature maps for redundancy (ablation) | SIRFP · Cityscapes, Deeplabv3-ResNet50 | mIoU: 80.8 -> 81.3 | Helped |
| Increase redundancy feature resolution from full to twice the input feature dimensions (ablation) | SIRFP · Cityscapes, Deeplabv3-ResNet50 | mIoU: 81.3 -> 81.3 | No clear effect |
| Use JS divergence instead of dot product for the redundancy metric (ablation) | SIRFP · Cityscapes, Deeplabv3-ResNet50 | mIoU: 80.5 -> 81.3 | Helped |
| Use JS divergence instead of KL divergence for the redundancy metric (ablation) | SIRFP · Cityscapes, Deeplabv3-ResNet50 | mIoU: 81.1 -> 81.3 | Helped |
| Use JS divergence instead of dice score for the redundancy metric (ablation) | SIRFP · Cityscapes, Deeplabv3-ResNet50 | mIoU: 81.0 -> 81.3 | Helped |
Surprising findings
- Dot-product redundancy metric performs notably worse (80.5 mIoU) than JS divergence (81.3), attributed to sign-related cancellation.
- Doubling the feature resolution (twice) gave no further mIoU gain over full resolution (both 81.3).
Limitations
- Requires forward inference over training samples to collect feature maps for pretrained networks, adding cost.
- Redundancy computation can be expensive for networks with very large feature map resolution.
- Gains over strong baselines are modest on some settings (e.g., ADE20K and COCO stuff-10K at 60% reduction vs DCFP: 44.9 vs 44.8).
Experimental setup
Datasets: Cityscapes ADE20K COCO stuff-10K ImageNet COCO2017
Benchmarks: Cityscapes ADE20K COCO stuff-10K COCO2017 ImageNet
Metrics: mIoU FLOPs reduction parameters AP Top-1 accuracy Top-5 accuracy FPS runtime
Baselines: Random NS Taylor DepGraph FPGM DCFP CAP CHEX DMCP
Builds on: FPGM DCFP HRank CHIP
Closest papers by meaning
- Structural Pruning via Latency-Saliency Knapsack (NeurIPS 2022 · similarity 0.87)
- Unsupervised Semantic Segmentation Through Depth-Guided Feature Correlation and Sampling (CVPR 2024 · similarity 0.86)
- Improving Semi-Supervised Semantic Segmentation with Sliced-Wasserstein Feature Alignment and Uniformity (CVPR 2025 · similarity 0.86)
- Hierarchy-Agnostic Unsupervised Segmentation: Parsing Semantic Image Structure (NeurIPS 2024 · similarity 0.85)
- Training-Free Dataset Pruning for Instance Segmentation (ICLR 2025 · similarity 0.85)
- VLM-Pruner: Buffering for Spatial Sparsity in an Efficient VLM Centrifugal Token Pruning Paradigm (CVPR 2026 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.