Research map / Detection and segmentation / CNN features
Learning Equivariant Segmentation with Instance-Unique Querying
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 inter-scene instance discrimination loss (ablation) | CondInst R50 · COCO val2017 | mask AP: 35.5 -> 37.6 | Helped |
| add transformation equivariance loss (ablation) | CondInst R50 · COCO val2017 | mask AP: 35.5 -> 36.5 | Helped |
| combine both losses (ablation) | CondInst R50 · COCO val2017 | mask AP: 37.6 -> 38.1 | Helped |
| use focal loss instead of l1/l2/cross-entropy for inter-scene loss (ablation) | inter-scene loss · COCO val2017 | mask AP: 36.0 (l1) / 37.3 (CE) -> 37.6 | Helped |
| add external memory (100k pixels) vs mini-batch only (ablation) | inter-scene loss · COCO val2017 | mask AP: 36.9 -> 37.6 | Helped |
| instance-balanced sparse sampling instead of dense sampling (ablation) | inter-scene loss · COCO val2017 | mask AP: 37.1 -> 37.6 | Helped |
| equivariance on query-instance relation rather than feature only (ablation) | equivariance loss · COCO val2017 | mask AP: 36.3 -> 36.5 | Helped |
| loss coefficient 3 instead of 1 (ablation) | loss balance · COCO val2017 | mask AP: 37.0 -> 37.6 | Helped |
| add the training framework to four query-based segmenters across backbones | CondInst, SOLOv2, SOTR, Mask2Former · COCO test-dev | mask AP: +1.6 to +3.2 AP | Helped |
| add framework to SOLOv2 | SOLOv2 R50 · LVISv1 val | mask AP: 21.4 -> 24.1 | Helped |
| add framework to Mask2Former | Mask2Former Swin-B · COCO Panoptic val | PQ: 56.1 -> 57.3 | Helped |
| training cost of adding framework | CondInst R50 · COCO val2017 | hours/epoch: 1.51 -> 1.59 | Hurt |
Closest papers by meaning
- FastInst: A Simple Query-Based Model for Real-Time Instance Segmentation (CVPR 2023 · similarity 0.88)
- Training-Free Dataset Pruning for Instance Segmentation (ICLR 2025 · similarity 0.87)
- MDQE: Mining Discriminative Query Embeddings To Segment Occluded Instances on Challenging Videos (CVPR 2023 · similarity 0.87)
- SOIT: Segmenting Objects with Instance-Aware Transformers (AAAI 2022 · similarity 0.87)
- Look Before You Match: Instance Understanding Matters in Video Object Segmentation (CVPR 2023 · similarity 0.87)
- v-CLR: View-Consistent Learning for Open-World Instance Segmentation (CVPR 2025 · similarity 0.85)
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.