Research map / Architectures and efficiency / Attention and transformers
MixFormerV2: Efficient Fully Transformer Tracking
Yutao Cui, Tianhui Song, Gangshan Wu, Limin Wang
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 4 distribution-regression prediction tokens instead of a single-token direct box (ablation) | MixFormerV2 (12-layer) · LaSOT | AUC: 63.1% -> 67.5% | Improved |
| add token-based score head (ablation) | MixFormerV2-B · LaSOT | AUC: 68.9% -> 70.6% | Improved |
| dense-to-sparse distillation from MixViT-B (ablation) | MixFormerV2 12-layer · LaSOT | AUC: 67.5% -> 68.9% | Improved |
| use larger MixViT-L teacher instead of MixViT-B (ablation) | MixFormerV2 12-layer · LaSOT | AUC: 68.9% -> 69.6% | Improved |
| add logits distillation (ablation) | 4-layer MixViT student · LaSOT | AUC: 60.7% -> 62.4% | Improved |
| add feature mimicking on top of logits distillation (ablation) | 4-layer MixViT student · LaSOT | AUC: 62.4% -> 62.9% | Improved |
| progressive model depth pruning init instead of MAE first 4 layers (ablation) | 4-layer student · LaSOT | AUC: 62.9% -> 64.8% | Improved |
| add intermediate teacher (ablation) | 4-layer student · LaSOT | AUC: 64.8% -> 65.5% | Improved |
| vary eliminating epochs m from 40 to 50 (ablation) | MixFormerV2-B · LaSOT | AUC: 68.5% -> 68.5% | No clear effect |
| prune 12 layers to 8 layers (ablation) | MixFormerV2-B · LaSOT | AUC: 68.9% -> 68.5% | Hurt |
| reduce MLP ratio 4 to 1 (ablation) | 4-layer MixFormerV2-S · LaSOT | AUC: 61.0% -> 59.4% | Hurt |
Closest papers by meaning
- General Compression Framework for Efficient Transformer Object Tracking (ICCV 2025 · similarity 0.86)
- MixFormer: End-to-End Tracking With Iterative Mixed Attention (CVPR 2022 · similarity 0.86)
- TransCenter: Transformers With Dense Representations for Multiple-Object Tracking (TPAMI 2022 · similarity 0.86)
- Exploring Lightweight Hierarchical Vision Transformers for Efficient Visual Tracking (ICCV 2023 · similarity 0.85)
- UTPTrack: Towards Simple and Unified Token Pruning for Visual Tracking (CVPR 2026 · similarity 0.84)
- Target Focused Shallow Transformer Framework for Efficient Visual Tracking (AAAI 2024 · similarity 0.83)
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.
