Mask2Former模型是在Cityscapes全景分割(小型版本,Swin骨干)上进行训练的。该模型在 Masked-attention Mask Transformer for Universal Image Segmentation 论文中首次介绍并于 this repository 首次发布。
声明:发布Mask2Former的团队没有为该模型撰写模型卡片,因此本模型卡片由Hugging Face团队撰写。
Mask2Former统一处理实例分割、语义分割和全景分割,采用相同的范式:预测一组掩膜和相应的标签。因此,将这3个任务都视为实例分割。Mask2Former通过(i)用更先进的多尺度可变形注意力Transformer替换像素解码器,(ii)采用具有掩膜注意力的Transformer解码器来提高性能而不引入额外的计算,(iii)通过计算子采样点上的损失而不是整个掩膜来提高训练效率,优于以往的最新结果 MaskFormer ,性能和效率都有所提升。
您可以使用此特定检查点进行全景分割。查看 model hub 以寻找您感兴趣的任务上的其他微调版本。
使用该模型的方法如下:
import requests import torch from PIL import Image from transformers import AutoImageProcessor, Mask2FormerForUniversalSegmentation # load Mask2Former fine-tuned on Cityscapes panoptic segmentation processor = AutoImageProcessor.from_pretrained("facebook/mask2former-swin-small-cityscapes-panoptic") model = Mask2FormerForUniversalSegmentation.from_pretrained("facebook/mask2former-swin-small-cityscapes-panoptic") url = "http://images.cocodataset.org/val2017/000000039769.jpg" image = Image.open(requests.get(url, stream=True).raw) inputs = processor(images=image, return_tensors="pt") with torch.no_grad(): outputs = model(**inputs) # model predicts class_queries_logits of shape `(batch_size, num_queries)` # and masks_queries_logits of shape `(batch_size, num_queries, height, width)` class_queries_logits = outputs.class_queries_logits masks_queries_logits = outputs.masks_queries_logits # you can pass them to processor for postprocessing result = processor.post_process_panoptic_segmentation(outputs, target_sizes=[image.size[::-1]])[0] # we refer to the demo notebooks for visualization (see "Resources" section in the Mask2Former docs) predicted_panoptic_map = result["segmentation"]
更多代码示例,请参阅 documentation 。