Research map / Detection and segmentation / CNN features
HyPiDecoder: Hybrid Pixel Decoder for Efficient Segmentation and Detection
Fengzhe Zhou, Humphrey Shi
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 part of MSDeformAttn pixel decoder with HyPiDecoder (3d3r) | Mask DINO SwinL · COCO panoptic val2017 | PQ: 58.3 -> 58.8 | Improved |
| replace default pixel decoder with HyPiDecoder (3d3s) | Mask2Former SwinL · COCO panoptic val2017 | PQ: 57.6 -> 58.1 | Improved |
| replace default pixel decoder with HyPiDecoder (3d3s) | Mask2Former R50 · COCO panoptic val2017 | PQ: 47.0 -> 47.5 | Improved |
| replace MSDeformAttn with MSLinearAttn plus FPN layers (3l3r) | Mask DINO R50 · COCO panoptic val2017 | PQ: 53.0 -> 53.4 | Improved |
| HyPiDecoder 3d3r instead of default pixel decoder | Mask DINO SwinL · COCO instance segmentation val2017 | Mask AP: 52.3 -> 52.3 | No clear effect |
| HyPiDecoder 3d3s instead of default pixel decoder | Mask2Former SwinL · COCO instance segmentation val2017 | Mask AP: 49.5 -> 49.9 | Improved |
| HyPiDecoder 3d3s instead of default pixel decoder | Mask2Former R50 · ADE20K val | mIoU: 47.2 -> 47.6 | Improved |
| HyPiDecoder 3d3s instead of default pixel decoder | Mask2Former R50 · Cityscapes val | mIoU: 79.4 -> 81.1 | Improved |
| add more layers (1l1r to 3l3r) (ablation) | HyPiDecoder Mask DINO R50 · COCO panoptic val2017 | PQ: 50.5 -> 53.4 | Improved |
| use MSLinearAttn instead of MSDeformAttn in hybrid layers (ablation) | HyPiDecoder Mask DINO R50 (12 ep) · COCO panoptic val2017 | PQ: 49.0 -> 49.3 | No clear effect |
| mix attention and FPN layers (3l3r) instead of linear attention only (6l0r) (ablation) | HyPiDecoder Mask DINO R50 (12 ep) · COCO panoptic val2017 | PQ: 43.7 -> 49.3 | Improved |
| mix attention and FPN layers (3l3r) instead of FPN only (0l6r) (ablation) | HyPiDecoder Mask DINO R50 (12 ep) · COCO panoptic val2017 | PQ: 48.7 -> 49.3 | Improved |
Closest papers by meaning
- DI-MaskDINO: A Joint Object Detection and Instance Segmentation Model (NeurIPS 2024 · similarity 0.85)
- Rethinking Decoders for Transformer-based Semantic Segmentation: A Compression Perspective (NeurIPS 2024 · similarity 0.84)
- Segment Anything, Even Occluded (CVPR 2025 · similarity 0.84)
- PEM: Prototype-based Efficient MaskFormer for Image Segmentation (CVPR 2024 · similarity 0.84)
- Progressive Token Length Scaling in Transformer Encoders for Efficient Universal Segmentation (ICLR 2025 · similarity 0.84)
- EfficientSAM: Leveraged Masked Image Pretraining for Efficient Segment Anything (CVPR 2024 · similarity 0.84)
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.
