Research map / Detection and segmentation / CNN features
Embedding-Free Transformer with Inference Spatial Reduction for Efficient Semantic Segmentation
Hyunwoo Yu, Yubin Cho, Beoungwoo Kang, Seunghun Moon, Kyeongbo Kong, Suk-Ju Kang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| train with small reduction and apply ISR at inference instead of the same reduction in training (ablation) | EDAFormer-T · COCO-Stuff | mIoU (%): 39.8 -> 40.3 | Improved |
| train with small reduction and apply ISR at inference (ablation) | EDAFormer-B · COCO-Stuff | mIoU (%): 45.3 -> 45.8 | Improved |
| use embedding-free attention instead of QKV-embedding attention in decoder (ablation) | EDAFormer-T · ADE20K val | mIoU (%): 41.5 -> 42.3 | Improved |
| use global attention instead of depth-wise convolution in decoder (ablation) | EDAFormer-T · ADE20K val | mIoU (%): 40.7 -> 42.3 | Improved |
| allocate more EFT blocks to high-level decoder stages {3-2-1} (ablation) | EDAFormer-T · ADE20K val | mIoU (%): 41.5 -> 42.3 | Improved |
| apply ISR [16,8,2,1]-[2,4,8] at inference without fine-tuning | EDAFormer-T · Cityscapes val | GFLOPs: 151.7 -> 94.9 | Improved |
| apply ISR [16,8,2,1]-[2,4,8] at inference | EDAFormer-T · ADE20K val | mIoU (%): 42.3 -> 42.1 | No clear effect |
| push ISR to larger reduction [16,8,4,2]-[2,4,8] without fine-tuning (ablation) | EDAFormer-T · COCO-Stuff | mIoU (%): 40.3 -> 39.1 | Hurt |
| fine-tune 40K iterations after larger-reduction ISR (ablation) | EDAFormer-T · COCO-Stuff | mIoU (%): 39.1 -> 40.2 | Improved |
| remove QKV embeddings so ISR degrades less at high reduction (ablation) | EDAFormer-T decoder with ISR · Cityscapes val | mIoU (%): 77.9 -> 78.6 | Improved |
| use max pooling instead of average pooling for key-value reduction (ablation) | EDAFormer-T with ISR · Cityscapes val | mIoU (%): 78.7 -> 78.4 | Hurt |
| use bipartite matching instead of average pooling for ISR (ablation) | EDAFormer-T with ISR · Cityscapes val | FPS (img/s): 10.2 -> 10.5 | No clear effect |
Closest papers by meaning
- Rethinking Decoders for Transformer-based Semantic Segmentation: A Compression Perspective (NeurIPS 2024 · similarity 0.90)
- Rethinking Semantic Segmentation From a Sequence-to-Sequence Perspective With Transformers (CVPR 2021 · similarity 0.88)
- Head-Free Lightweight Semantic Segmentation with Linear Transformer (AAAI 2023 · similarity 0.88)
- SED: A Simple Encoder-Decoder for Open-Vocabulary Semantic Segmentation (CVPR 2024 · similarity 0.85)
- Learning Content-Enhanced Mask Transformer for Domain Generalized Urban-Scene Segmentation (AAAI 2024 · similarity 0.85)
- SegMAN: Omni-scale Context Modeling with State Space Models and Local Attention for Semantic Segmentation (CVPR 2025 · 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.
