Research map / Label-efficient and robust learning / Knowledge distillation
MobileSAM2: Lightweight Segment Anything in Images and Videos via Hypergraphical Knowledge Distillation
Kai Jiang, Jiaxing Huang, Jingyi Zhang, Weiying Xie, Yunsong Li, Yufei Wang, Aoran Xiao, Dacheng Tao
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 HyperKD distillation (ablation) | MobileSAM2-23M · LVOS val | J&F: 44.8 -> 69.0 | Improved |
| replace FAKD with HyperKD | MobileSAM2-23M · LVOS val | J&F: 63.4 -> 69.0 | Improved |
| replace FAKD with HyperKD | MobileSAM2-5M · LVOS val | J&F: 41.0 -> 48.8 | Improved |
| replace FAKD with HyperKD | MobileSAM2-10M · LVOS val | J&F: 44.0 -> 51.8 | Improved |
| use searched MobileSAM2-23M instead of SAM2-TinyViT-21M with no distillation | student architecture · LVOS val | J&F: 34.5 -> 44.8 | Improved |
| replace FitNet with HyperKD | MobileSAM2-23M · LVOS val | J&F: 60.8 -> 69.0 | Improved |
| replace two-vertex graph KD (Context Matters, IntRA-KD) with hypergraph KD (ablation) | MobileSAM2-23M · LVOS val | J&F: 60.7 -> 69.0 | Improved |
| use cosine similarity instead of Euclidean distance for hypergraph distance (ablation) | HyperKD · LVOS val | J&F: 67.2 -> 69.0 | Improved |
| use cosine similarity instead of Manhattan distance for hypergraph distance (ablation) | HyperKD · LVOS val | J&F: 65.7 -> 69.0 | Improved |
| set hyperedge threshold epsilon to 1.0 instead of 0.5 (ablation) | HyperKD · LVOS val | J&F: 65.2 -> 69.0 | Improved |
| set hyperedge threshold epsilon to 1.0 instead of 1.5 (ablation) | HyperKD · LVOS val | J&F: 62.0 -> 69.0 | Improved |
| vary epsilon from 0.75 to 1.25 (ablation) | HyperKD · LVOS val | J&F: 68.4 / 69.0 / 68.1 | No clear effect |
Closest papers by meaning
- SAM 2: Segment Anything in Images and Videos (ICLR 2025 · similarity 0.86)
- HyperGLM: HyperGraph for Video Scene Graph Generation and Anticipation (CVPR 2025 · similarity 0.85)
- MobileInst: Video Instance Segmentation on the Mobile (AAAI 2024 · similarity 0.84)
- ARK: A Dual-Axis Multimodal Retrieval Benchmark along Reasoning and Knowledge (NeurIPS 2026 · similarity 0.84)
- SAM-R1: Leveraging SAM for Reward Feedback in Multimodal Segmentation via Reinforcement Learning (NeurIPS 2025 · similarity 0.84)
- MMMG: A Massive, Multidisciplinary, Multi-Tier Generation Benchmark for Text-to-Image Reasoning (NeurIPS 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.
