Research map / Detection and segmentation / CNN features
FastInst: A Simple Query-Based Model for Real-Time 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 |
|---|---|---|---|
| Use IA-guided queries instead of zero-initialized queries (ablation) | FastInst · COCO val2017 | AP: 31.5 -> 35.6 | Helped |
| Use dual-path pixel-then-query updates instead of single-path query updates (ablation) | FastInst · COCO val2017 | AP: 36.9 -> 37.9 | Helped |
| Add GT mask-guided learning (ablation) | FastInst · COCO val2017 | AP: 37.4 -> 37.9 | Helped |
| Use PPM-FPN instead of FPN as pixel decoder (ablation) | FastInst · COCO val2017 | FPS: 35.7 -> 35.5 | Helped |
| Increase decoder depth from 3 to 9 layers (ablation) | FastInst · COCO val2017 | AP: 37.9 -> 39.1 | Helped |
| Increase IA-guided query count from 100 to 300 (ablation) | FastInst · COCO val2017 | AP: 37.9 -> 38.8 | Helped |
| Add 8 auxiliary learnable queries rather than using none (ablation) | FastInst · COCO val2017 | AP: 37.7 -> 37.9 | Helped |
| Select IA-guided queries from E3 instead of E4 (ablation) | FastInst · COCO val2017 | AP: 37.9 -> 38.0 | Helped |
| Use bipartite matching in instance activation loss (ablation) | FastInst · COCO val2017 | AP: 35.7 -> 37.9 | Helped |
| Add location cost to instance activation loss (ablation) | FastInst · COCO val2017 | AP: 37.1 -> 37.9 | Helped |
| Use learnable rather than sinusoidal positional embeddings (ablation) | FastInst-D3 · COCO val2017 | FPS: 32.9 -> 35.5 | Helped |
| Use local-maximum-first query selection (ablation) | FastInst-D1 with 10 IA-guided queries · COCO val2017 | AP: 28.6 -> 30.0 (+1.4) | Helped |
Closest papers by meaning
- Sparse Instance Activation for Real-Time Instance Segmentation (CVPR 2022 · similarity 0.90)
- SOIT: Segmenting Objects with Instance-Aware Transformers (AAAI 2022 · similarity 0.89)
- BoxInst: High-Performance Instance Segmentation With Box Annotations (CVPR 2021 · similarity 0.89)
- Learning Equivariant Segmentation with Instance-Unique Querying (NeurIPS 2022 · similarity 0.88)
- MobileInst: Video Instance Segmentation on the Mobile (AAAI 2024 · similarity 0.88)
- Training-Free Dataset Pruning for Instance Segmentation (ICLR 2025 · similarity 0.87)
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.