Research map / Detection and segmentation / CNN features
TransMix: Attend To Mix for Vision Transformers
Check your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Replace area-ratio label weights with attention-guided weights | ViT-based models · ImageNet-1k | top-1 accuracy: +0.9% for DeiT-S and XCiT-L | Helped |
| Pretrain the backbone with TransMix | DeiT-S with Linear decoder · Pascal Context | mIoU: 49.1 -> 49.7 | Helped |
| Pretrain the backbone with TransMix | DeiT-S with Segmenter decoder · Pascal Context | mIoU: 49.7 -> 50.6 | Helped |
| Pretrain the PVT-S backbone with TransMix | Mask R-CNN with PVT-S · COCO val2017 | box AP: 40.4 -> 40.9 | Helped |
| Pretrain the PVT-S backbone with TransMix | Mask R-CNN with PVT-S · COCO val2017 | mask AP: 37.8 -> 38.4 | Helped |
| Pretrain DeiT-S with TransMix | DeiT-S · ImageNet-A | top-1 accuracy: 19.1 -> 21.1 | Helped |
| Pretrain DeiT-S with TransMix | DeiT-S · ImageNet-O | AUPR: 20.9 -> 21.9 | Helped |
| Use attention from block 6 rather than block 12 (ablation) | TransMix-DeiT-S · ImageNet-1k | top-1 accuracy: 80.7 -> 80.3 | Hurt |
| Use attention from block 8 rather than block 12 (ablation) | TransMix-DeiT-S · ImageNet-1k | top-1 accuracy: 80.7 -> 80.3 | Hurt |
| Use attention from block 10 rather than block 12 (ablation) | TransMix-DeiT-S · ImageNet-1k | top-1 accuracy: 80.7 -> 80.4 | Hurt |
| Use attention rollout across all blocks instead of block 12 attention (ablation) | TransMix-DeiT-S · ImageNet-1k | top-1 accuracy: 80.7 -> 80.4 | Hurt |
| Use a frozen external model to provide attention instead of the model's own evolving attention (ablation) | TransMix-DeiT-S · ImageNet-1k | top-1 accuracy: 80.7 -> 80.4 | Hurt |
Closest papers by meaning
- TdAttenMix: Top-Down Attention Guided Mixup (AAAI 2025 · similarity 0.87)
- Enhance Image Classification via Inter-Class Image Mixup with Diffusion Model (CVPR 2024 · similarity 0.86)
- DiffuseMix: Label-Preserving Data Augmentation with Diffusion Models (CVPR 2024 · similarity 0.86)
- Adversarial AutoMixup (ICLR 2024 · similarity 0.86)
- Supervision Interpolation via LossMix: Generalizing Mixup for Object Detection and Beyond (AAAI 2024 · similarity 0.85)
- TiMix: Text-Aware Image Mixing for Effective Vision-Language Pre-training (AAAI 2024 · similarity 0.85)
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.