OneFormer

OneFormer模型是在Cityscapes数据集（大尺寸版本，Swin骨干）上训练的。由Jain等人在 OneFormer: One Transformer to Rule Universal Image Segmentation 论文中介绍，并于 this repository 首次发布。

模型描述

OneFormer是第一个多任务通用图像分割框架。它只需要使用单个通用架构、单个模型和单个数据集进行训练，就能在语义分割、实例分割和全景分割任务上胜过现有的专门化模型。OneFormer使用任务令牌将模型与关注的任务相结合，使得架构在训练时具有任务引导性，而在推理时具有任务动态性，所有这些都由单个模型完成。

使用目的和限制

您可以使用这个特定的检查点进行语义分割、实例分割和全景分割。查看 model hub 以查找在不同数据集上进行微调的其他版本。

如何使用

这是如何使用这个模型的方法：

from transformers import OneFormerProcessor, OneFormerForUniversalSegmentation
from PIL import Image
import requests
url = "https://huggingface.co/datasets/shi-labs/oneformer_demo/blob/main/cityscapes.png"
image = Image.open(requests.get(url, stream=True).raw)

# Loading a single model for all three tasks
processor = OneFormerProcessor.from_pretrained("shi-labs/oneformer_cityscapes_swin_large")
model = OneFormerForUniversalSegmentation.from_pretrained("shi-labs/oneformer_cityscapes_swin_large")

# Semantic Segmentation
semantic_inputs = processor(images=image, task_inputs=["semantic"], return_tensors="pt")
semantic_outputs = model(**semantic_inputs)
# pass through image_processor for postprocessing
predicted_semantic_map = processor.post_process_semantic_segmentation(outputs, target_sizes=[image.size[::-1]])[0]

# Instance Segmentation
instance_inputs = processor(images=image, task_inputs=["instance"], return_tensors="pt")
instance_outputs = model(**instance_inputs)
# pass through image_processor for postprocessing
predicted_instance_map = processor.post_process_instance_segmentation(outputs, target_sizes=[image.size[::-1]])[0]["segmentation"]

# Panoptic Segmentation
panoptic_inputs = processor(images=image, task_inputs=["panoptic"], return_tensors="pt")
panoptic_outputs = model(**panoptic_inputs)
# pass through image_processor for postprocessing
predicted_semantic_map = processor.post_process_panoptic_segmentation(outputs, target_sizes=[image.size[::-1]])[0]["segmentation"]

更多示例，请参考 documentation 。

引用

@article{jain2022oneformer,
      title={{OneFormer: One Transformer to Rule Universal Image Segmentation}},
      author={Jitesh Jain and Jiachen Li and MangTik Chiu and Ali Hassani and Nikita Orlov and Humphrey Shi},
      journal={arXiv}, 
      year={2022}
    }