library_name: transformers tags:

• 视觉
• 图像分割
• 生态学 datasets:
• coralscapes metrics:
• 平均交并比 license: apache-2.0

模型卡片：模型ID

基于MiT-B2骨干网络并在1024x1024分辨率下针对Coralscapes数据集微调的SegFormer模型，首次发表于论文《珊瑚景观数据集：珊瑚礁语义场景理解》(https://arxiv.org/abs/2503.20000)。

模型详情

模型描述

• 模型类型: SegFormer
• 微调基础模型: 仅预训练的SegFormer(b2尺寸)编码器(nvidia/mit-b2)

模型来源

• 代码仓库: coralscapesScripts
• 演示: Hugging Face Spaces

快速使用指南

最简方式对Coralscapes数据集图像进行分割：

from transformers import SegformerImageProcessor, SegformerForSemanticSegmentation
from PIL import Image
from datasets import load_dataset

# 加载Coralscapes测试集图像或自定义图像
dataset = load_dataset("EPFL-ECEO/coralscapes") 
image = dataset["test"][42]["image"]

preprocessor = SegformerImageProcessor.from_pretrained("EPFL-ECEO/segformer-b2-finetuned-coralscapes-1024-1024")
model = SegformerForSemanticSegmentation.from_pretrained("EPFL-ECEO/segformer-b2-finetuned-coralscapes-1024-1024")

inputs = preprocessor(image, return_tensors="pt")
outputs = model(**inputs)
outputs = preprocessor.post_process_semantic_segmentation(outputs, target_sizes=[(image.size[1], image.size[0])])
label_pred = outputs[0].numpy()

对于非标准尺寸(非1024x1024)图像，推荐采用滑动窗口策略：

import torch 
import torch.nn.functional as F
from transformers import SegformerImageProcessor, SegformerForSemanticSegmentation
from PIL import Image
import numpy as np
from datasets import load_dataset
device = 'cuda' if torch.cuda.is_available() else 'cpu'

def resize_image(image, target_size=1024):
    """将图像短边调整至1024像素"""
    h_img, w_img = image.size
    if h_img < w_img:
        new_h, new_w = target_size, int(w_img * (target_size / h_img))
    else:
        new_h, new_w = int(h_img * (target_size / w_img)), target_size
    return image.resize((new_h, new_w))

def segment_image(image, preprocessor, model, crop_size=(1024, 1024), num_classes=40, transform=None):
    """基于滑动窗口的语义分割实现"""
    h_crop, w_crop = crop_size
    
    img = torch.Tensor(np.array(resize_image(image, target_size=1024)).transpose(2, 0, 1)).unsqueeze(0)
    batch_size, _, h_img, w_img = img.size()
    
    if transform:
        img = torch.Tensor(transform(image=img.numpy())["image"]).to(device)    
        
    h_grids = int(np.round(3/2*h_img/h_crop)) if h_img > h_crop else 1
    w_grids = int(np.round(3/2*w_img/w_crop)) if w_img > w_crop else 1
    
    h_stride = int((h_img - h_crop + h_grids -1)/(h_grids -1)) if h_grids > 1 else h_crop
    w_stride = int((w_img - w_crop + w_grids -1)/(w_grids -1)) if w_grids > 1 else w_crop
    
    preds = img.new_zeros((batch_size, num_classes, h_img, w_img))
    count_mat = img.new_zeros((batch_size, 1, h_img, w_img))
    
    for h_idx in range(h_grids):
        for w_idx in range(w_grids):
            y1 = h_idx * h_stride
            x1 = w_idx * w_stride
            y2 = min(y1 + h_crop, h_img)
            x2 = min(x1 + w_crop, w_img)
            y1 = max(y2 - h_crop, 0)
            x1 = max(x2 - w_crop, 0)
            crop_img = img[:, :, y1:y2, x1:x2]
            with torch.no_grad():
                if(preprocessor):
                    inputs = preprocessor(crop_img, return_tensors="pt")
                    inputs["pixel_values"] = inputs["pixel_values"].to(device)
                else:
                    inputs = crop_img.to(device)
                outputs = model(**inputs)

            resized_logits = F.interpolate(
                outputs.logits[0].unsqueeze(dim=0), size=crop_img.shape[-2:], mode="bilinear", align_corners=False
            )
            preds += F.pad(resized_logits,
                            (int(x1), int(preds.shape[3] - x2), int(y1),
                            int(preds.shape[2] - y2))).cpu()
            count_mat[:, :, y1:y2, x1:x2] += 1
        
    assert (count_mat == 0).sum() == 0
    preds = preds / count_mat
    preds = preds.argmax(dim=1)
    preds = F.interpolate(preds.unsqueeze(0).type(torch.uint8), size=image.size[::-1], mode='nearest')
    label_pred = preds.squeeze().cpu().numpy()
    return label_pred

# 使用示例
dataset = load_dataset("EPFL-ECEO/coralscapes") 
image = dataset["test"][42]["image"]

preprocessor = SegformerImageProcessor.from_pretrained("EPFL-ECEO/segformer-b2-finetuned-coralscapes-1024-1024")
model = SegformerForSemanticSegmentation.from_pretrained("EPFL-ECEO/segformer-b2-finetuned-coralscapes-1024-1024")

label_pred = segment_image(image, preprocessor, model)

训练与评估

数据

使用通用珊瑚礁语义分割数据集Coralscapes进行训练和评估。

训练流程

遵循SegFormer原论文实现，采用：

• 批量大小8训练265轮次
• AdamW优化器(初始学习率6e-5)
• 权重衰减1e-2
• 多项式学习率调度器(幂次1)
• 数据增强：随机缩放(1-2倍)、水平翻转(0.5概率)、随机裁剪至1024×1024
• 评估采用非重叠1024x1024滑动窗口策略

性能指标

• 测试准确率: 80.904
• 测试平均交并比: 54.682

引用

若使用本模型请引用：

@misc{sauder2025coralscapesdatasetsemanticscene,
        title={珊瑚景观数据集：珊瑚礁语义场景理解}, 
        author={Jonathan Sauder and Viktor Domazetoski and Guilhem Banc-Prandi and Gabriela Perna and Anders Meibom and Devis Tuia},
        year={2025},
        eprint={2503.20000},
        archivePrefix={arXiv},
        primaryClass={cs.CV},
        url={https://arxiv.org/abs/2503.20000}, 
  }