Research map / Detection and segmentation / CNN features
BlendMask: Top-Down Meets Bottom-Up for Instance Segmentation
Check 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 |
|---|---|---|---|
| Use the learned attention-guided blender instead of weighted-sum assembly (ablation) | BlendMask · COCO val2017, 28_4_4 | AP: 29.7 -> 31.6 | Helped |
| Use learned attentions instead of fixed position-sensitive assembly (ablation) | BlendMask · COCO val2017, 28_4_4 | AP: 30.3 -> 31.6 | Helped |
| Increase top attention resolution from M=2 to M=7 at R=28 (ablation) | BlendMask · COCO val2017 | AP: 30.6 -> 32.0 | Helped |
| Increase top attention resolution from M=4 to M=14 at R=56 (ablation) | BlendMask · COCO val2017 | AP: 32.5 -> 33.3 | Helped |
| Increase basis count from 1 to 4 (ablation) | BlendMask · COCO val2017, 56_K_7 | AP: 30.6 -> 33.1 | Helped |
| Increase basis count from 4 to 8 (ablation) | BlendMask · COCO val2017, 56_K_7 | AP: 33.1 -> 33.0 | No clear effect |
| Use FPN P3/P5 instead of backbone C3/C5 as bottom features (ablation) | BlendMask · COCO val2017 | AP: 33.1 -> 33.3 | Helped |
| Use bilinear instead of nearest-neighbour top interpolation (ablation) | BlendMask · COCO val2017, M=14 | AP: 33.4 -> 33.6 | Helped |
| Use bilinear RoIAlign instead of nearest RoIPool for bottom sampling (ablation) | BlendMask · COCO val2017, R=56, M=14 | AP: 31.9 -> 33.6 | Helped |
| Add auxiliary semantic segmentation supervision (ablation) | BlendMask · COCO val2017 | AP: 33.6 -> 34.2 | Helped |
| Increase bottom-module channel width from 128 to 256 (ablation) | BlendMask · COCO val2017 | AP: 34.2 -> 34.3 | Helped |
| Use stride-4 bottom features from P2/P5 (ablation) | BlendMask · COCO val2017 | AP: 34.3 -> 34.4 | Helped |
Closest papers by meaning
- RefineMask: Towards High-Quality Instance Segmentation With Fine-Grained Features (CVPR 2021 · similarity 0.90)
- Mask Encoding for Single Shot Instance Segmentation (CVPR 2020 · similarity 0.90)
- SOLOv2: Dynamic and Fast Instance Segmentation (NeurIPS 2020 · similarity 0.89)
- BoxInst: High-Performance Instance Segmentation With Box Annotations (CVPR 2021 · similarity 0.89)
- PolarMask: Single Shot Instance Segmentation With Polar Representation (CVPR 2020 · similarity 0.89)
- MaskDiff: Modeling Mask Distribution with Diffusion Probabilistic Model for Few-Shot Instance Segmentation (AAAI 2024 · similarity 0.88)
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.