Research map / Detection and segmentation / CNN features
SegViT v2: Exploring Efficient and Continual Semantic Segmentation with Plain Vision Transformers
Bowen Zhang, Liyang Liu, Vu Minh Hieu Phan, Zhi Tian, Chunhua Shen, Yifan Liu
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 |
|---|---|---|---|
| replace SETR decoder with ATM decoder (CE loss) (ablation) | SegViT · ADE20K val (ViT-Base) | mIoU (ss): 46.5 -> 47.0 | Improved |
| supervise ATM with decoupled mask loss instead of CE loss (ablation) | SegViT ATM · ADE20K val (ViT-Base) | mIoU (ss): 47.0 -> 49.6 | Improved |
| add lower-layer features (cascade [6,12]) (ablation) | SegViT · ADE20K val (ViT-Base) | mIoU (ss): 49.6 -> 50.9 | Improved |
| use three layers [6,8,12] instead of single layer (ablation) | SegViT · ADE20K val (ViT-Base) | mIoU (ss): 49.6 -> 51.3 | Improved |
| use four layers [3,6,9,12] instead of three (ablation) | SegViT · ADE20K val (ViT-Base) | mIoU (ss): 51.3 -> 51.2 | No clear effect |
| add query downsampling (QD) to single structure without upsampling (ablation) | SegViT · ADE20K val (ViT-Base) | mIoU (ss): 49.6 -> 46.9 | Hurt |
| add query upsampling (QU) after QD (Shrunk) (ablation) | SegViT Shrunk · ADE20K val (ViT-Base) | mIoU (ss): 46.9 -> 50.0 | Improved |
| use nearest token squeezing instead of edge-aware downsampling (Shrunk++) (ablation) | SegViT Shrunk++ · ADE20K val (ViT-Base) | mIoU (ss): 49.9 -> 43.3 | Hurt |
| add edge-aware downsampling to token squeezing (3x3) (ablation) | SegViT Shrunk++ · ADE20K val (ViT-Base) | mIoU (ss): 38.9 -> 47.9 | Improved |
| replace FPN decoder with ATM token merge (ablation) | MaskFormer-style decoder · ADE20K val (ViT-Base) | mIoU (ss): 46.7 -> 50.6 | Improved |
| add attention mask loss instead of dot product loss (ablation) | SegViT · ADE20K val (ViT-Base) | mIoU (ss): 50.6 -> 51.2 | Improved |
| downsample only query with nearest instead of learnable stride-2 conv (ablation) | SegViT QD · ADE20K val (ViT-Large) | mIoU (ss): 44.5 -> 53.9 | Improved |
Closest papers by meaning
- SegViT: Semantic Segmentation with Plain Vision Transformers (NeurIPS 2022 · similarity 0.96)
- SED++: A Simple Encoder-Decoder for Improved Open-Vocabulary Semantic Segmentation (TPAMI 2025 · similarity 0.89)
- SED: A Simple Encoder-Decoder for Open-Vocabulary Semantic Segmentation (CVPR 2024 · similarity 0.88)
- Your ViT is Secretly an Image Segmentation Model (CVPR 2025 · similarity 0.88)
- Rethinking Decoders for Transformer-based Semantic Segmentation: A Compression Perspective (NeurIPS 2024 · similarity 0.88)
- OMG-Seg: Is One Model Good Enough For All Segmentation? (CVPR 2024 · similarity 0.87)
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.
