Research map / Detection and segmentation / CNN features
MDQE: Mining Discriminative Query Embeddings To Segment Occluded Instances on Challenging Videos
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 |
|---|---|---|---|
| Initialize frame-level queries with grid-guided spatial and contextual features (ablation) | MDQE query decoder · OVIS valid | mAP: 15.4 -> 19.8 | Helped |
| Use O2I instead of I2O decoder architecture (ablation) | MDQE with initialized queries · OVIS valid | mAP: 19.8 -> 24.2 | Helped |
| Add temporal cross-attention (ablation) | MDQE with initialized queries and O2I decoder · OVIS valid | mAP: 24.2 -> 25.6 | Helped |
| Associate frame-level queries across frames using window w=5 instead of no association (ablation) | MDQE query initialization · OVIS valid | mAP: 28.5 -> 30.6 | Helped |
| Increase inter-frame association window from w=5 to w=7 (ablation) | MDQE query initialization · OVIS valid | mAP: 30.6 -> 30.5 | Hurt |
| Add inter-instance BCE repulsion with α=2 in place of typical mask prediction loss (ablation) | MDQE mask supervision · OVIS valid | mAP: 29.0 -> 30.5 | Helped |
| Add inter-instance Dice repulsion alongside inter-instance BCE (ablation) | MDQE mask supervision · OVIS valid | mAP: 30.5 -> 31.2 | Helped |
| Raise the nearby-instance threshold ϵ to 0.5 (ablation) | MDQE inter-instance mask repulsion · OVIS valid, slightly occluded objects | APso: 46.7 -> 47.2 | Hurt |
| Combine mIoU and embedding similarity for clip association rather than using either alone (ablation) | MDQE near-online tracker · OVIS valid | mAP: 29.1 -> 30.6; 28.3 -> 30.6 | Helped |
| Reduce the memory pool from Tmem=10 to Tmem=5 frames (ablation) | MDQE near-online tracker · OVIS valid | mAP: 30.6 -> 30.4 | Hurt |
| Increase clip length from T=1 to T=7 (ablation) | MDQE · OVIS valid | mask AP: 27% -> 31.0% | Helped |
| Increase clip length to T=9 (ablation) | MDQE · OVIS valid | mask AP: around 31.0% -> 29% | Hurt |
Closest papers by meaning
- InsPro: Propagating Instance Query and Proposal for Online Video Instance Segmentation (NeurIPS 2022 · similarity 0.87)
- Learning Equivariant Segmentation with Instance-Unique Querying (NeurIPS 2022 · similarity 0.87)
- SyncVIS: Synchronized Video Instance Segmentation (NeurIPS 2024 · similarity 0.86)
- Decoupled Motion Expression Video Segmentation (CVPR 2025 · similarity 0.85)
- Minimizing Labeled, Maximizing Unlabeled: An Image-Driven Approach for Video Instance Segmentation (CVPR 2025 · similarity 0.84)
- Learning Spatial-Semantic Features for Robust Video Object Segmentation (ICLR 2025 · 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.