Research map / Detection and segmentation / CNN features
MOBIUS: Big-to-Mobile Universal Instance Segmentation via Multi-modal Bottleneck Fusion and Calibrated Decoder Pruning
Mattia Segu, Marta Tintore Gazulla, Yongqin Xian, Luc Van Gool, Federico Tombari
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 |
|---|---|---|---|
| replace dot product with scaled cosine similarity for language-guided classification (ablation) | multi-modal DETR generalist (GLEE-style) · COCO-val | convergence: did not converge -> 45.9 AP_box | Improved |
| add scaled cosine only, no calibration, in joint single-stage training (ablation) | MOBIUS-H-R50 · Joint training (all datasets) | convergence: D.N.C. | No clear effect |
| add language-guided uncertainty calibration loss on top of scaled cosine (ablation) | MOBIUS-H-R50 · Joint training, single stage | convergence / AP_box: D.N.C. -> 50.0 AP_box COCO-val | Improved |
| add calibration loss on COCO-only training (ablation) | MOBIUS-H-R50 · COCO-val | AP_box: 45.9 -> 46.5 | Improved |
| decode a single scale instead of multi-scale (ablation) | MaskDINO pixel decoder · COCO-val | AP_box: 49.2 -> 47.9 | Hurt |
| decode a single scale instead of multi-scale (ablation) | MOBIUS bottleneck encoder · COCO-val | AP_box: 47.9 -> 47.5 | No clear effect |
| use stride-16 bottleneck instead of stride-8 (ablation) | MOBIUS bottleneck encoder · COCO-val | AP_box: 47.5 -> 47.5 with 200 -> 91 pixel decoder GFLOPs | No clear effect |
| reduce MaskDINO pixel decoder to 1 layer (ablation) | MaskDINO pixel decoder (single-scale) · COCO-val | AP_box: 47.9 -> 43.4 | Hurt |
| swap MaskDINO pixel decoder for RT-DETR pixel decoder (ablation) | GLEE (mobile MNv4-CM) · LVIS / ODinW | AP_box: LVIS 28.9 -> 22.8; ODinW 37.0 -> 30.7 | Hurt |
| replace MaskDINO pixel decoder with bottleneck encoder (ablation) | GLEE (mobile MNv4-CM) · ODinW | AP_box: 37.0 -> 38.6 | Improved |
| replace MaskDINO pixel decoder with bottleneck encoder | GLEE (mobile MNv4-CM) · Samsung S24 latency | latency (ms): 436.6 -> 127.1 | Improved |
| add sigmoidal-growth inference-time query pruning (ablation) | MOBIUS transformer decoder · COCO-val | AP_mask and decoder FLOPs: decoder FLOPs reduced by additional 50%, AP_mask 44.0 | Improved |
Closest papers by meaning
- MobileInst: Video Instance Segmentation on the Mobile (AAAI 2024 · similarity 0.87)
- YOLACT: Real-Time Instance Segmentation (ICCV 2019 · similarity 0.87)
- FastInst: A Simple Query-Based Model for Real-Time Instance Segmentation (CVPR 2023 · similarity 0.87)
- SOLOv2: Dynamic and Fast Instance Segmentation (NeurIPS 2020 · similarity 0.86)
- SOLO: A Simple Framework for Instance Segmentation (TPAMI 2021 · similarity 0.86)
- Segment Anything, Even Occluded (CVPR 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.
