Research map / Detection and segmentation / CNN features
Patch-Level Representation Learning for Self-Supervised 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 |
|---|---|---|---|
| Add patch-level SelfPatch supervision to DINO | DINO, ViT-S/16 · COCO val2017 object detection | APbb: 40.8 -> 42.1 | Helped |
| Add patch-level SelfPatch supervision to DINO | DINO, ViT-S/16 · COCO val2017 instance segmentation | APmk: 37.3 -> 38.5 | Helped |
| Add patch-level SelfPatch supervision to DINO | DINO, ViT-S/16 · ADE20K semantic segmentation | mIoU: 38.3 -> 41.2 | Helped |
| Add patch-level SelfPatch supervision to DINO | DINO, ViT-S/16 · DAVIS 2017 video object segmentation | (J&F)m: 60.7 to 62.7 | Helped |
| Use all image patches rather than a local neighborhood as positive candidates (ablation) | SelfPatch · DAVIS 2017 video object segmentation | (J&F)m: 55.1 -> 47.3 | Hurt |
| Use a 5×5 rather than 3×3 positive-candidate neighborhood (ablation) | SelfPatch · DAVIS 2017 video object segmentation | (J&F)m: 57.0 -> 56.5 | Hurt |
| Select k=4 positive patches instead of k=1 (ablation) | SelfPatch · DAVIS 2017 video object segmentation | (J&F)m: 56.3 -> 57.0 | Helped |
| Select k=4 positive patches instead of k=2 (ablation) | SelfPatch · DAVIS 2017 video object segmentation | (J&F)m: 56.4 -> 57.0 | Helped |
| Select k=4 rather than k=8 positive patches (ablation) | SelfPatch · DAVIS 2017 video object segmentation | (J&F)m: 56.5 -> 57.0 | Helped |
| Use attention-based aggregation instead of average pooling (ablation) | SelfPatch · DAVIS 2017 video object segmentation | (J&F)m: 51.4 -> 57.0 | Helped |
| Add negative pairs to the SelfPatch objective (ablation) | MoBY, ViT-Ti/16 · DAVIS 2017 video object segmentation | (J&F)m: 58.4 vs. 58.9 | No clear effect |
| Set the SelfPatch loss weight to λ=0.1 instead of λ=0.0 (ablation) | DINO, ViT-Ti/16 · DAVIS 2017 video object segmentation | (J&F)m: 55.3 -> 57.0 | Helped |
Closest papers by meaning
- Vision Transformers Need Registers (ICLR 2024 · similarity 0.90)
- Efficient Self-supervised Vision Transformers for Representation Learning (ICLR 2022 · similarity 0.89)
- Revisiting Self-Supervised Visual Representation Learning (CVPR 2019 · similarity 0.89)
- Self-Supervised Learning of Intertwined Content and Positional Features for Object Detection (ICML 2025 · similarity 0.88)
- Closest Neighbors are Harmful for Lightweight Masked Auto-encoders (CVPR 2025 · similarity 0.87)
- An Image is Worth More Than 16x16 Patches: Exploring Transformers on Individual Pixels (ICLR 2025 · similarity 0.87)
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.