Research map / Detection and segmentation / CNN features
PatchDCT: Patch Refinement for High Quality Instance Segmentation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 global DCT refinement with independent patch refinement (ablation) | DCT-Mask · COCO val2017 | AP: 36.6 to 37.2 | Helped |
| replace binary-grid refinement with patch DCT refinement (ablation) | Mask-RCNN refinement · COCO val2017 | AP: 35.7 to 37.2 | Helped |
| use a three-class classifier that also corrects foreground/background patches (ablation) | PatchDCT · COCO val2017 | AP: 36.9 to 37.2 | Helped |
| regress only mixed patches rather than all patches (ablation) | PatchDCT · COCO val2017 | AP: 36.6 to 37.2 | Helped |
| add the patch DCT regressor (ablation) | PatchDCT classifier-only variant · COCO val2017 | Boundary AP: 23.3 to 24.5 | Helped |
| add the patch DCT regressor (ablation) | PatchDCT classifier-only variant · COCO val2017 | AP: 36.7 to 37.2 | Helped |
| use 6-dimensional rather than 3-dimensional patch DCT vectors (ablation) | PatchDCT · COCO val2017 | AP: 36.8 to 37.2 | Helped |
| increase patch DCT dimension from 6 to 9 (ablation) | PatchDCT · COCO val2017 | AP: 37.2 to 36.9 | Hurt |
| add a second PatchDCT refinement stage (ablation) | PatchDCT · COCO val2017 | AP: 37.2 to 37.4 | Helped |
| increase refinement from two to three stages (ablation) | PatchDCT · COCO val2017 | AP: 37.4 to 37.3 | Hurt |
| use 8×8 rather than 4×4 patches (ablation) | PatchDCT · COCO val2017 | AP: 37.0 to 37.2 | Helped |
| use P2-only features rather than P2-P5 features (ablation) | PatchDCT · COCO val2017 | AP: 37.1 to 37.2 | Helped |
Closest papers by meaning
- Look Closer To Segment Better: Boundary Patch Refinement for Instance Segmentation (CVPR 2021 · similarity 0.90)
- DCT-Mask: Discrete Cosine Transform Mask Representation for Instance Segmentation (CVPR 2021 · similarity 0.88)
- RefineMask: Towards High-Quality Instance Segmentation With Fine-Grained Features (CVPR 2021 · similarity 0.88)
- Instance Segmentation With Mask-Supervised Polygonal Boundary Transformers (CVPR 2022 · similarity 0.87)
- Scaling Laws in Patchification: An Image Is Worth 50,176 Tokens And More (ICML 2025 · similarity 0.85)
- MaskTwins: Dual-form Complementary Masking for Domain-Adaptive Image Segmentation (ICML 2025 · 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.