Core components of this framework include a deformable convolution-based genarel visual backbone InternImage, the M3I-Pretraining self-supervised/semi-supervised training algorithm, the Uni-Perceiver universal decoder family, and the BEVFormer autonomous driving perception encoder suite.

Key Highlights

1. A 3-billion-parameter InternImage-G as the top-performing open-source general-purpose visual backbone
2. 90.1% Top-1 accuracy on ImageNet-1K, leading all publicly available models
3. 65.5 mAP on COCO object detection, the only open-source model surpassing the 65 mAP threshold

Core Capabilities

• Support for diverse downstream computer vision and cross-modal tasks
• Integration with CVPR 2023 Workshop on End-to-End Autonomous Driving
• Compatibility with Segment Anything
• Intermediate feature map extraction
• Cost-effective pretraining powered by DeepSpeed
• Precompiled DCNv3 .whl packages for simplified installation
• Access to InternImage-H (1B) and G (3B) heavyweight variants
• TensorRT inference support for image classification, object detection, and semantic/instance segmentation
• Clean, maintainable classification codebase
• Pre-trained InternImage-T/S/B/L/XL weights on ImageNet-1K
• Pre-trained InternImage-L/XL weights on ImageNet-22K
• Detection and instance segmentation checkpoints for all T/S/B/L/XL variants
• Semantic segmentation checkpoints for all T/S/B/L/XL variants

Image Modality Task Performance

Classification

Object Detection

Segmentation

Image-Text Cross-Modal Task Performance