Research map / Detection and segmentation / CNN features
Learning to Refine Object Segments
Pedro O. Pinheiro, Tsung-Yi Lin, Ronan Collobert, Piotr Dollár
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 |
|---|---|---|---|
| add top-down refinement modules to the feedforward net (ablation) | DeepMask-ours · COCO val segmentation proposals | AUC: 19.4 -> 20.9 | Improved |
| upgrade VGG trunk to ResNet-50 (D39) (ablation) | W160 DeepMask · COCO val dense masks | AR: 35.5 -> 37.0 | Improved |
| reduce feature channels to 128 (ablation) | W160-P4-D39 · COCO val dense masks | AR: 37.0 -> 36.9 | No clear effect |
| reduce input size from 160 to 112 (ablation) | W160-P4-D39 · COCO val dense masks | AR: 37.0 -> 30.8 | Hurt |
| use simplified head C instead of original DeepMask head (ablation) | DeepMask · COCO val segmentation proposals | AUC: 18.3 -> 19.4 | Improved |
| add horizontal connections to a skip-style averaged-output net (k=1) (ablation) | SharpMask variant · COCO val | AR | Improved |
| add vertical (top-down) convs to horizontal-free variant (ablation) | SharpMask variant · COCO val | AR | Improved |
| increase refinement channel capacity k (ablation) | SharpMask · COCO val | AUC | Improved |
| use schedule a with k<=2 (too small mask encoding) (ablation) | SharpMask · COCO val | AUC | Hurt |
| generate skip features with two 3x3 convs instead of one (ablation) | SharpMask · COCO val | AUC | Improved |
| fine-tune the whole network after refinement training (ablation) | SharpMask · COCO val | AR | No clear effect |
| add extra image scales (Zoom) to SharpMask | SharpMask · COCO val segmentation proposals | AUC small objects: 2.5 -> 6.9 | Improved |
Closest papers by meaning
- RefineMask: Towards High-Quality Instance Segmentation With Fine-Grained Features (CVPR 2021 · similarity 0.90)
- FastMask: Segment Multi-Scale Object Candidates in One Shot (CVPR 2017 · similarity 0.89)
- Segmentation-Aware Convolutional Networks Using Local Attention Masks (ICCV 2017 · similarity 0.88)
- Learning to Segment via Cut-and-Paste (ECCV 2018 · similarity 0.88)
- ShapeMask: Learning to Segment Novel Objects by Refining Shape Priors (ICCV 2019 · similarity 0.88)
- Putting the Object Back into Video Object Segmentation (CVPR 2024 · similarity 0.86)
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.
