[CVPR 2023 Highlight] InternImage: Exploring Large-Scale Vision Foundation Models with Deformable Convolutions
[中文版本]
# InternImage: Large-Scale Vision Foundation Model
The official implementation of
[InternImage: Exploring Large-Scale Vision Foundation Models with Deformable Convolutions](https://arxiv.org/abs/2211.05778).
\[[Paper](https://arxiv.org/abs/2211.05778)\] \[[Blog in Chinese](https://zhuanlan.zhihu.com/p/610772005)\]
## Highlights
- :thumbsup: **The strongest open-source visual universal backbone model with up to 3 billion parameters**
- **Achieved `90.1% Top1` accuracy in ImageNet, the most accurate among open-source models**
- **Achieved `65.5 mAP` on the COCO benchmark dataset for object detection, the only model that exceeded `65.0 mAP`**
## News
- `Jan 22, 2024`: Support [DCNv4](https://github.com/OpenGVLab/DCNv4) in InternImage!
- `Feb 28, 2023`: InternImage is accepted to CVPR 2023!
- `Nov 18, 2022`: InternImage-XL merged into [BEVFormer v2](https://arxiv.org/abs/2211.10439) achieves state-of-the-art performance of `63.4 NDS` on nuScenes Camera Only.
- `Nov 10, 2022`: InternImage-H achieves a new record `65.4 mAP` on COCO detection test-dev and `62.9 mIoU` on ADE20K, outperforming previous models by a large margin.
## History
- [x] Models for other downstream tasks
- [x] Support [CVPR 2023 Workshop on End-to-End Autonomous Driving](https://opendrivelab.com/e2ead/cvpr23), see [here](https://github.com/OpenGVLab/InternImage/tree/master/autonomous_driving)
- [x] Support extracting intermediate features, see [here](classification/extract_feature.py)
- [x] Low-cost training with [DeepSpeed](https://github.com/microsoft/DeepSpeed), see [here](https://github.com/OpenGVLab/InternImage/tree/master/classification)
- [x] Compiling-free `.whl` package of DCNv3 operator, see [here](https://github.com/OpenGVLab/InternImage/releases/tag/whl_files)
- [x] InternImage-H(1B)/G(3B)
- [x] TensorRT inference for classification/detection/segmentation models
- [x] Classification code of the InternImage series
- [x] InternImage-T/S/B/L/XL ImageNet-1K pretrained model
- [x] InternImage-L/XL ImageNet-22K pretrained model
- [x] InternImage-T/S/B/L/XL detection and instance segmentation model
- [x] InternImage-T/S/B/L/XL semantic segmentation model
## Introduction
InternImage is an advanced vision foundation model developed by researchers from Shanghai AI Laboratory, Tsinghua University, and other institutions. Unlike models based on Transformers, InternImage employs DCNv3 as its core operator. This approach equips the model with dynamic and effective receptive fields required for downstream tasks like object detection and segmentation, while enabling adaptive spatial aggregation.
Some other projects related to InternImage include the pretraining algorithm "M3I-Pretraining," the general-purpose decoder series "Uni-Perceiver," and the autonomous driving perception encoder series "BEVFormer."
## Performance
- InternImage achieved an impressive Top-1 accuracy of 90.1% on the ImageNet benchmark dataset using only publicly available data for image classification. Apart from two undisclosed models trained with additional datasets by Google and Microsoft, InternImage is the only open-source model that achieves a Top-1 accuracy of over 90.0%, and it is also the largest model in scale worldwide.
- InternImage outperformed all other models worldwide on the COCO object detection benchmark dataset with a remarkable mAP of 65.5, making it the only model that surpasses 65 mAP in the world.
- InternImage also demonstrated world's best performance on 16 other important visual benchmark datasets, covering a wide range of tasks such as classification, detection, and segmentation, making it the top-performing model across multiple domains.
**Classification**
| Image Classification | Scene Classification | Long-Tail Classification |
| ImageNet | Places365 | Places 205 | iNaturalist 2018 |
| 90.1 | 61.2 | 71.7 | 92.6 |
**Detection**
| General Object Detection | Long-Tail Object Detection | Autonomous Driving Object Detection | Dense Object Detection |
| COCO | VOC 2007 | VOC 2012 | OpenImage | LVIS minival | LVIS val | BDD100K | nuScenes | CrowdHuman |
| 65.5 | 94.0 | 97.2 | 74.1 | 65.8 | 63.2 | 38.8 | 64.8 | 97.2 |
**Segmentation**
| Semantic Segmentation | Street Segmentation | RGBD Segmentation |
| ADE20K | COCO Stuff-10K | Pascal Context | CityScapes | NYU Depth V2 |
| 62.9 | 59.6 | 70.3 | 87.0 | 68.1 |
## Released Models
Open-Source Visual Pretrained Models
ImageNet-1K Image Classification
COCO Object Detection and Instance Segmentation
ADE20K Semantic Segmentation
Main Results of FPS
## Related Projects
### Foundation Models
- [Uni-Perceiver](https://github.com/fundamentalvision/Uni-Perceiver): A Pre-training unified architecture for generic perception for zero-shot and few-shot tasks
- [Uni-Perceiver v2](https://arxiv.org/abs/2211.09808): A generalist model for large-scale vision and vision-language tasks
- [M3I-Pretraining](https://github.com/OpenGVLab/M3I-Pretraining): One-stage pre-training paradigm via maximizing multi-modal mutual information
- [InternVL](https://github.com/OpenGVLab/InternVL): A leading multimodal large language model excelling in tasks such as OCR, multimodal reasoning, and dialogue
### Autonomous Driving
- [BEVFormer](https://github.com/fundamentalvision/BEVFormer): A cutting-edge baseline for camera-based 3D detection
- [BEVFormer v2](https://arxiv.org/abs/2211.10439): Adapting modern image backbones to Bird's-Eye-View recognition via perspective supervision
## Application in Challenges
- [2022 Waymo 3D Camera-Only Detection Challenge](https://waymo.com/open/challenges/2022/3d-camera-only-detection/): BEVFormer++ ranks 1st based on InternImage
- [nuScenes 3D detection](https://www.nuscenes.org/object-detection?externalData=all&mapData=all&modalities=Camera): BEVFormer v2 achieves SOTA performance of 64.8 NDS on nuScenes Camera Only
- [CVPR 2023 Workshop End-to-End Autonomous Driving](https://opendrivelab.com/e2ead/cvpr23): InternImage supports the baseline of the [3D Occupancy Prediction Challenge](https://opendrivelab.com/AD23Challenge.html#Track3) and [OpenLane Topology Challenge](https://opendrivelab.com/AD23Challenge.html#Track1)
## Citation
If this work is helpful for your research, please consider citing the following BibTeX entry.
```bibtex
@inproceedings{wang2023internimage,
title={Internimage: Exploring large-scale vision foundation models with deformable convolutions},
author={Wang, Wenhai and Dai, Jifeng and Chen, Zhe and Huang, Zhenhang and Li, Zhiqi and Zhu, Xizhou and Hu, Xiaowei and Lu, Tong and Lu, Lewei and Li, Hongsheng and others},
booktitle={Proceedings of the IEEE/CVF conference on computer vision and pattern recognition},
pages={14408--14419},
year={2023}
}
```