Research map / Architectures and efficiency / Attention and transformers
DropPos: Pre-Training Vision Transformers by Reconstructing Dropped Positions
Haochen Wang, Junsong Fan, Yuxi Wang, Kaiyou Song, Tong Wang, Zhaoxiang Zhang
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 patch mask ratio 0.75 instead of 0 (whole image visible) (ablation) | DropPos · ImageNet-1K fine-tune | top-1: 81.94 -> 82.96 | Improved |
| use patch mask ratio 0.75 instead of 0 (ablation) | DropPos · ADE20k | mIoU: 31.66 -> 40.68 | Improved |
| raise mask ratio from 0.75 to 0.90 (ablation) | DropPos · ADE20k | mIoU: 40.68 -> 40.30 | Hurt |
| set position mask ratio to 1.00 instead of 0.75 (no reference patch) (ablation) | DropPos · ImageNet-1K fine-tune | top-1: 82.96 -> 82.66 | Hurt |
| set position mask ratio to 1.00 instead of 0.75 (ablation) | DropPos · ADE20k | mIoU: 40.68 -> 39.98 | Hurt |
| remove position smoothing (sigma=0) versus linear decay 1->0 (ablation) | DropPos · COCO detection | AP^b: 40.32 -> 42.14 | Improved |
| use constant sigma=2 instead of decay 1->0 (ablation) | DropPos · ImageNet-1K fine-tune | top-1: 82.79 -> 82.96 | Hurt |
| remove attentive reconstruction (tau=inf) (ablation) | DropPos · COCO detection | AP^b: 40.56 -> 42.14 | Improved |
| remove attentive reconstruction (tau=inf) (ablation) | DropPos · ADE20k | mIoU: 38.38 -> 40.68 | Improved |
| use tau=0.5 instead of 0.1 (ablation) | DropPos · ADE20k | mIoU: 40.68 -> 38.80 | Hurt |
| make sin-cos position embedding learnable (ablation) | DropPos · ADE20k | mIoU: 40.68 -> 39.37 | Hurt |
| random learnable position embedding instead of fixed sin-cos (ablation) | DropPos · ImageNet-1K fine-tune | top-1: 82.96 -> 82.48 | Hurt |
Closest papers by meaning
- Cameras as Relative Positional Encoding (NeurIPS 2025 · similarity 0.84)
- Vision Transformers Need Registers (ICLR 2024 · similarity 0.84)
- Maximizing the Position Embedding for Vision Transformers with Global Average Pooling (AAAI 2025 · similarity 0.84)
- LookHere: Vision Transformers with Directed Attention Generalize and Extrapolate (NeurIPS 2024 · similarity 0.84)
- DPPE: Rethinking Camera-Based Positional Encoding for Scaling Multi-View Transformers (NeurIPS 2026 · similarity 0.84)
- Where Am I and What Will I See: An Auto-Regressive Model for Spatial Localization and View Prediction (ICLR 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.
