Research map / Detection and segmentation / Open-set and few-shot detection
EVA: Exploring the Limits of Masked Visual Representation Learning at Scale
Yuxin Fang, Wen Wang, Binhui Xie, Quan Sun, Ledell Wu, Xinggang Wang, Tiejun Huang, Xinlong Wang, Yue Cao
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 masked CLIP-feature reconstruction without additional feature tokenization rather than tokenization-based masked modeling (ablation) | ViT-B pre-training · ImageNet-1K | top-1 accuracy: 85.0 -> 85.0 | No clear effect |
| Use masked CLIP-feature reconstruction without additional feature tokenization rather than tokenization-based masked modeling (ablation) | ViT-B pre-training · ADE20K | single-scale mIoU: 52.6 -> 52.7 | No clear effect |
| Extend tokenized feature pre-training from 300 to 1600 epochs (ablation) | ViT-B pre-training · ImageNet-1K | top-1 accuracy: 85.0 -> 85.5 | Improved |
| Extend tokenized feature pre-training from 300 to 1600 epochs (ablation) | ViT-B pre-training · ADE20K | single-scale mIoU: 52.7 -> 53.1 | Improved |
| Add feature distillation during 300-epoch pre-training (ablation) | ViT-B pre-training · ImageNet-1K | top-1 accuracy: 85.0 -> 85.1 | No clear effect |
| Add feature distillation during 300-epoch pre-training (ablation) | ViT-B pre-training · ADE20K | single-scale mIoU: 52.6 -> 52.5 | Hurt |
| Extend feature-distillation pre-training from 300 to 800 epochs (ablation) | ViT-B pre-training · ImageNet-1K | top-1 accuracy: 85.1 -> 85.1 | No clear effect |
| Extend feature-distillation pre-training from 300 to 800 epochs (ablation) | ViT-B pre-training · ADE20K | single-scale mIoU: 52.5 -> 52.7 | Improved |
| Use masked CLIP-feature reconstruction instead of 800-epoch feature distillation (ablation) | ViT-B pre-training · ImageNet-1K | top-1 accuracy: 85.1 -> 85.5 | Improved |
| Use masked CLIP-feature reconstruction instead of 800-epoch feature distillation (ablation) | ViT-B pre-training · ADE20K | single-scale mIoU: 52.7 -> 53.3 | Improved |
| Initialize the CLIP vision tower from pre-trained EVA rather than training it from scratch | 1.1B CLIP-g · zero-shot classification, 12-benchmark average | top-1 accuracy: +3.8 | Improved |
| Use EVA as the CLIP vision tower initialization | 1.1B CLIP-g · zero-shot classification, natural-distribution-shift benchmarks | accuracy gap: 2.5 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- Mask DINO — Mask DINO: Towards a Unified Transformer-Based Framework for Object Detection and Segmentation (CVPR 2023) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Perception Encoder: The best visual embeddings are not at the output of the network (NeurIPS 2025 · similarity 0.87)
- Closest Neighbors are Harmful for Lightweight Masked Auto-encoders (CVPR 2025 · similarity 0.86)
- ScaleKD: Strong Vision Transformers Could Be Excellent Teachers (NeurIPS 2024 · similarity 0.85)
- Beyond Encoder Accumulation: Measuring Encoder Roles in Multi-Encoder VLMs (NeurIPS 2026 · similarity 0.85)
- Efficient Vision-Language Pre-training by Cluster Masking (CVPR 2024 · similarity 0.85)
- Seeing Clearly, Reasoning Confidently: Plug-and-Play Remedies for Vision Language Model Blindness (CVPR 2026 · 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.
