Research map / Architectures and efficiency / Pruning and efficient tuning
MuSViT: A Foundation Vision Model for Sheet Music Representation
Carlos Penarrubia, Antonio Rios-Vila, Eliseo Fuentes-Martinez, Juan Martinez-Sevilla, Francisco Castellanos, María Alfaro-Contreras, Jorge Calvo-Zaragoza
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 |
|---|---|---|---|
| use music-specific MAE pre-training (MuSViT) instead of the best general-purpose encoder, frozen | MuSViT · Full-page score recognition (linear probing) | SER (%): 48.6 -> 16.4 | Improved |
| fine-tune MuSViT instead of the prior state of the art on full-page recognition | MuSViT · Full-page score recognition (fine-tuning) | SER (%): 20.0 -> 10.9 | Improved |
| music-specific pre-training for staff-level recognition under linear probing versus Qwen3-VL | MuSViT · Staff-level score recognition (linear probing) | SER (%): 21.0 -> 18.4 | Improved |
| fine-tune MuSViT versus task-specific state of the art on staff-level recognition | MuSViT · Staff-level score recognition (fine-tuning) | SER (%): 8.0 -> 8.6 | No clear effect |
| use MuSViT instead of DINOv3-7B for frozen symbol detection | MuSViT · DeepScoresV2 (linear probing) | mAP: 70.4 -> 79.7 | Improved |
| fine-tune MuSViT with LoRA and Faster R-CNN versus Transformer-based state of the art | MuSViT · DeepScoresV2 (fine-tuning) | mAP50: 90.5 -> 97.0 | Improved |
| use MuSViT instead of prior state of the art for difficulty classification (fine-tuning) | MuSViT · Score difficulty classification (fine-tuning) | Acc0 (%): 38.4 -> 54.2 | Improved |
| use full MuSViT instead of MuSViT-Light for fine-tuned difficulty classification | MuSViT · Score difficulty classification (fine-tuning) | Acc0 (%): 54.0 -> 54.2 | No clear effect |
| add synthetic warm-up stage (two-stage curriculum) before IMSLP MAE training (ablation) | MuSViT pre-training · Pre-training stability (effective rank) | embedding effective rank | Improved |
| use MuSViT embeddings instead of general-purpose encoders for embedding-transcription alignment | MuSViT · Mozarteum and Polish Digital Scores | Spearman correlation (histogram): -0.009 -> 0.714 | Improved |
Closest papers by meaning
- Toward Multimodal Sheet Music Recognition and Understanding (NeurIPS 2026 · similarity 0.88)
- Transcoda: End-to-End Zero-Shot Optical Music Recognition via Data-Centric Synthetic Training (NeurIPS 2026 · similarity 0.85)
- Rubato: Transcribing Piano Music with Timestamps (NeurIPS 2026 · similarity 0.85)
- Aria-MIDI: A Dataset of Piano MIDI Files for Symbolic Music Modeling (ICLR 2025 · similarity 0.85)
- CMI-RewardBench: Evaluating Music Reward Models with Compositional Multimodal Instruction (ICML 2026 · similarity 0.84)
- From Isolated feature to Orbits:\\Discovering Music Concepts via Multi-SAE Alignment (NeurIPS 2026 · 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.
