Research map / Detection and segmentation / CNN features
Towards All-in-One Pre-Training via Maximizing Multi-Modal Mutual Information
Read the paperCheck 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 |
|---|---|---|---|
| combine image self-supervision with category supervision and image-text supervision in single-stage pre-training | InternImage-H · COCO test-dev | AP box: 64.5 -> 65.4 | Helped |
| combine image self-supervision with category supervision and image-text supervision in single-stage pre-training | InternImage-H · LVIS minival | AP box: 59.8 -> 62.5 | Helped |
| combine image self-supervision with category supervision and image-text supervision in single-stage pre-training | InternImage-H · ADE20k val | mIoU: 60.8 -> 62.9 | Helped |
| combine image self-supervision with category supervision and image-text supervision in single-stage pre-training | InternImage-H · ImageNet val | Top1 Accuracy: 89.2 vs 89.6 | Mixed |
| add multi-input mixing to M3I (ablation) | M3I Pre-training · ImageNet-1k, COCO, LVIS, ADE20k | Top1, AP box, AP box, mIoU: 83.7 -> 83.9; 50.3 -> 50.8; 36.6 -> 37.5; 48.7 -> 49.0 | Helped |
| add multiple prediction targets without input mixing (ablation) | single-input single-target pre-training · ImageNet-1k, COCO, LVIS, ADE20k | Top1, AP box, AP box, mIoU: Image Classification: 81.8, 46.6, 33.0, 45.1; M3I w/o mix: 83.7, 50.3, 36.6, 48.7 | Helped |
| use masked rather than full input for intra-view self-supervision (ablation) | intra-view SSP · ImageNet-1k classification | Top1: 77.5 -> 83.1; 78.8 -> 83.3; 77.1 -> 83.2 | Helped |
| use masked rather than full input for inter-view self-supervision (ablation) | inter-view SSP · ImageNet-1k classification | Top1: 78.8 -> 78.9; 83.2 -> 83.7; 83.0 -> 82.9 | Mixed |
| predict dense features instead of dense pixels from a masked second view (ablation) | Siamese Image Modeling · ImageNet-1k classification | Top1: 78.9 -> 83.7 | Helped |
| predict dense features instead of dense pixels from a masked second view (ablation) | Siamese Image Modeling · COCO | AP box: 38.1 -> 49.8 | Helped |
| use dense-feature targets rather than global-feature targets (ablation) | self-supervised pre-training · COCO | AP box: 46.4 -> 50.1; 46.2 -> 49.8 | Helped |
| use full-view inter-view prediction rather than full-view intra-view reconstruction (ablation) | self-supervised pre-training · ImageNet-1k classification | Top1: 77.5 -> 78.8; 78.8 -> 83.2; 77.1 -> 83.0 | Mixed |
Closest papers by meaning
- Building a Strong Pre-Training Baseline for Universal 3D Large-Scale Perception (CVPR 2024 · similarity 0.85)
- Beyond Random Augmentations: Pretraining with Hard Views (ICLR 2025 · similarity 0.85)
- OmniSegmentor: A Flexible Multi-Modal Learning Framework for Semantic Segmentation (NeurIPS 2025 · similarity 0.85)
- Scalable Pre-training of Large Autoregressive Image Models (ICML 2024 · similarity 0.85)
- Multi-modal Vision Pre-training for Medical Image Analysis (CVPR 2025 · similarity 0.85)
- Long-tailed Object Detection Pretraining: Dynamic Rebalancing Contrastive Learning with Dual Reconstruction (NeurIPS 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.