Research map / Detection and segmentation / CNN features
Path Aggregation Network for 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 |
|---|---|---|---|
| add bottom-up path augmentation (ablation) | Mask R-CNN + FPN · COCO val-2017 | mask AP: 35.7 -> 36.4 | Helped |
| add adaptive feature pooling (ablation) | Mask R-CNN + FPN with MST and MBN, without BPA · COCO val-2017 | mask AP: 35.7 -> 36.3 | Helped |
| add adaptive feature pooling (ablation) | Mask R-CNN + FPN with BPA · COCO val-2017 | mask AP: 36.4 -> 36.9 | Helped |
| add fully-connected fusion (ablation) | PANet with BPA and AFP · COCO val-2017 | mask AP: +0.7 | Helped |
| add a heavier head (ablation) | PANet with BPA, AFP, and FF · COCO val-2017 | box AP (multi-task branch): 40.6 -> 42.1 | Helped |
| add a heavier head (ablation) | PANet · COCO val-2017 | mask AP: 37.6 -> 37.8 | No clear effect |
| place AFP fusion after the first parameter layer rather than before it (ablation) | adaptive feature pooling · COCO val-2017 | mask AP: 35.7 -> 36.3 | Helped |
| use sum instead of max to fuse AFP features (ablation) | adaptive feature pooling · COCO val-2017 | mask AP: 36.3 -> 36.2 | No clear effect |
| start the fully-connected branch at conv3 rather than conv2 (ablation) | fully-connected fusion · COCO val-2017 | mask AP: 37.5 -> 37.6 | No clear effect |
| start the fully-connected branch at conv4 rather than conv3 (ablation) | fully-connected fusion · COCO val-2017 | mask AP: 37.6 -> 37.2 | Hurt |
| use sum instead of product to fuse FC and FCN predictions (ablation) | fully-connected fusion · COCO val-2017 | mask AP: 36.9 -> 37.6 | Helped |
| add multi-GPU synchronized batch normalization (ablation) | re-implemented Mask R-CNN · Cityscapes val | AP: 33.1 -> 34.6 | Helped |
Closest papers by meaning
- Unifying Training and Inference for Panoptic Segmentation (CVPR 2020 · similarity 0.86)
- BANet: Bidirectional Aggregation Network With Occlusion Handling for Panoptic Segmentation (CVPR 2020 · similarity 0.86)
- Fully Convolutional Networks for Panoptic Segmentation (CVPR 2021 · similarity 0.86)
- InstanceCut: From Edges to Instances With MultiCut (CVPR 2017 · similarity 0.85)
- UPSNet: A Unified Panoptic Segmentation Network (CVPR 2019 · similarity 0.85)
- Progressive Neighborhood Aggregation for Semantic Segmentation Refinement (AAAI 2023 · 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.