Research map / Architectures and efficiency / Attention and transformers
MixFormer: End-to-End Tracking With Iterative Mixed Attention
Yutao Cui, Cheng Jiang, Limin Wang, Gangshan Wu
Check 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 iterative mixed attention to unify feature extraction and target-search integration (ablation) | MixFormer · LaSOT | AUC: 59.8 -> 68.4 | Improved |
| Increase the number of MAM blocks while reducing SAM blocks (ablation) | MixFormer · LaSOT | AUC: 65.8 -> 66.2 -> 67.4 -> 68.1 | Improved |
| Replace the corner localization head with the query-based head (ablation) | MixFormer-Base · LaSOT | AUC: 68.4 -> 66.0 | Hurt |
| Prune target-to-search cross-attention (ablation) | MixFormer-Base · LaSOT | FPS (1080Ti): 19 -> 25 | Improved |
| Prune target-to-search cross-attention (ablation) | MixFormer-Base · LaSOT | AUC: 68.4 -> 68.1 | Hurt |
| Add fixed-interval online templates without score-based selection (ablation) | MixFormer · LaSOT | AUC: 68.1 -> 66.6 | Hurt |
| Add score-based selection of online templates (ablation) | MixFormer with online templates · LaSOT | AUC: 66.6 -> 69.2 | Improved |
| Use ImageNet-1k rather than ImageNet-22k pretraining (ablation) | MixFormer · LaSOT | AUC: 69.2 -> 67.9 | Hurt |
| Train on GOT-10k only rather than the whole training datasets (ablation) | MixFormer · LaSOT | AUC: 69.2 -> 62.1 | Hurt |
| Increase model scale from MixFormer-22k to MixFormer-L | MixFormer · LaSOT test | P_Norm (%): 78.7 -> 79.9 | Improved |
Closest papers by meaning
- Target-Aware Tracking with Long-Term Context Attention (AAAI 2023 · similarity 0.86)
- ATOM: Accurate Tracking by Overlap Maximization (CVPR 2019 · similarity 0.86)
- MixFormerV2: Efficient Fully Transformer Tracking (NeurIPS 2023 · similarity 0.86)
- TrackFormer: Multi-Object Tracking With Transformers (CVPR 2022 · similarity 0.86)
- Compact Transformer Tracker with Correlative Masked Modeling (AAAI 2023 · similarity 0.86)
- SynCL: A Synergistic Training Strategy with Instance-Aware Contrastive Learning for End-to-End Multi-Camera 3D Tracking (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.
