使用coco数据集进行语义分割：数据预处理与损失函数

如何coco数据集进行目标检测的介绍已经有很多了，但是关于语义分割几乎没有。本文旨在说明如何处理 stuff_train2017.json stuff_val2017.json panoptic_train2017.json panoptic_val2017.json，将上面那些json中的dict转化为图片的label mask，也就是制作图片像素的标签的ground truth。

首先下载图片和annotation文件（也就是json文件）

2017 Train Images和2017 Val Images解压得到这两个文件夹，里面放着的是所有图片。

annotation下载下来有4个json文件

以及两个压缩包

这两个压缩包里面是panoptic segmentation的mask，用于制作ground truth。

首先说明一下那四个json文件的含义，stuff是语义分割，panoptic是全景分割。二者的区别在与

语义分割只区别种类，全景分割还区分个体。上图中，中间的语义分割将所有的人类视作同一个种类看待，而右边的全景分割还将每一个个体区分出来。

做语义分割，可以使用stuff_train2017.json生成ground truth。但是即便是只做语义分割，不做全景分割，在只看语义分割的情况下，stuff的划分精度不如panoptic，因此建立即使是做语义分割也用panoptic_train2017.json。本文中将两种json都进行说明。

stuff_train2017.json

下面这段代码是处理stuff_train2017.json生成ground truth

from pycocotools.coco import COCO
import os
from PIL import Image
import numpy as np
from matplotlib import pyplot as pltdef convert_coco2mask_show(image_id):print(image_id)img = coco.imgs[image_id]image = np.array(Image.open(os.path.join(img_dir, img['file_name'])))plt.imshow(image, interpolation='nearest')# cat_ids = coco.getCatIds()cat_ids = list(range(183))anns_ids = coco.getAnnIds(imgIds=img['id'], catIds=cat_ids, iscrowd=None)anns = coco.loadAnns(anns_ids)mask = np.zeros((image.shape[0], image.shape[1]))for i in range(len(anns)):tmp = coco.annToMask(anns[i]) * anns[i]["category_id"]mask += tmpprint(np.max(mask), np.min(mask))# 绘制二维数组对应的颜色图plt.figure(figsize=(8, 6))plt.imshow(mask, cmap='viridis', vmin=0, vmax=182)plt.colorbar(ticks=np.linspace(0, 182, 6), label='Colors')  # 添加颜色条# plt.savefig(save_dir + str(image_id) + ".jpg")plt.show()if __name__ == '__main__':Dataset_dir = "/home/xxxx/Downloads/coco2017/"coco = COCO("/home/xxxx/Downloads/coco2017/annotations/stuff_train2017.json")# createIndex# coco = COCO("/home/robotics/Downloads/coco2017/annotations/annotations/panoptic_val2017.json")img_dir = os.path.join(Dataset_dir, 'train2017')save_dir = os.path.join(Dataset_dir, "Mask/stuff mask/train")if not os.path.isdir(save_dir):os.makedirs(save_dir)image_id = 9convert_coco2mask_show(image_id)# for keyi, valuei in coco.imgs.items():#     image_id = valuei["id"]#     convert_coco2mask_show(image_id)

解释一下上面的代码。

coco = COCO("/home/xxxx/Downloads/coco2017/annotations/stuff_train2017.json")

从coco的官方库中导入工具，读取json。会得到这样的字典结构

在 coco.imgs ，找到 "id" 对应的value，这个就是每个图片唯一的编号，根据这个编号，到 coco.anns 中去索引 segmentation

注意 coco.imgs 的 id 对应的是 anns 中的 image_id，这个也是 train2017 文件夹中图片的文件名。

# cat_ids = coco.getCatIds()
cat_ids = list(range(183))

网上的绝大多数教程都写的是上面我注释的那行代码，那行代码会得到91个类，那个只能用于图像检测，但对于图像分割任务，包括的种类是182个和一个未归类的0类一共183个类。

tmp = coco.annToMask(anns[i]) * anns[i]["category_id"]

上面那行代码就是提取每一个类所对应的id，将所有的类都加进一个二维数组，就制作完成了ground truth。

panoptic_train2017.json

上面完成了通过stuff的json文件自作ground truth，那么用panoptic的json文件制作ground truth，是不是只要把

coco = COCO("/home/xxxx/Downloads/coco2017/annotations/stuff_train2017.json")

换为

coco = COCO("/home/xxxx/Downloads/coco2017/annotations/annotations/panoptic_val2017.json")

就行了呢？

答案是不行！会报错

Traceback (most recent call last):File "/home/xxxx/.local/lib/python3.8/site-packages/IPython/core/interactiveshell.py", line 3508, in run_codeexec(code_obj, self.user_global_ns, self.user_ns)File "<ipython-input-4-66f5c6e7dbe4>", line 1, in <module>coco = COCO("/home/xxxx/Downloads/coco2017/annotations/panoptic_val2017.json")File "/home/robotics/.local/lib/python3.8/site-packages/pycocotools/coco.py", line 86, in __init__self.createIndex()File "/home/xxxx/.local/lib/python3.8/site-packages/pycocotools/coco.py", line 96, in createIndexanns[ann['id']] = ann
KeyError: 'id'

用官方的库读官方的json还报错，真是巨坑！关键还没有官方的文档教你怎么用，只能一点点摸索，非常浪费时间精力。

制作panoptic的ground truth，用到了Meta公司的detectron2这个库

# Copyright (c) Facebook, Inc. and its affiliates.
import copy
import json
import osfrom detectron2.data import MetadataCatalog
from detectron2.utils.file_io import PathManagerdef load_coco_panoptic_json(json_file, image_dir, gt_dir, meta):"""Args:image_dir (str): path to the raw dataset. e.g., "~/coco/train2017".gt_dir (str): path to the raw annotations. e.g., "~/coco/panoptic_train2017".json_file (str): path to the json file. e.g., "~/coco/annotations/panoptic_train2017.json".Returns:list[dict]: a list of dicts in Detectron2 standard format. (See`Using Custom Datasets </tutorials/datasets.html>`_ )"""def _convert_category_id(segment_info, meta):if segment_info["category_id"] in meta["thing_dataset_id_to_contiguous_id"]:segment_info["category_id"] = meta["thing_dataset_id_to_contiguous_id"][segment_info["category_id"]]segment_info["isthing"] = Trueelse:segment_info["category_id"] = meta["stuff_dataset_id_to_contiguous_id"][segment_info["category_id"]]segment_info["isthing"] = Falsereturn segment_infowith PathManager.open(json_file) as f:json_info = json.load(f)ret = []for ann in json_info["annotations"]:image_id = int(ann["image_id"])# TODO: currently we assume image and label has the same filename but# different extension, and images have extension ".jpg" for COCO. Need# to make image extension a user-provided argument if we extend this# function to support other COCO-like datasets.image_file = os.path.join(image_dir, os.path.splitext(ann["file_name"])[0] + ".jpg")label_file = os.path.join(gt_dir, ann["file_name"])segments_info = [_convert_category_id(x, meta) for x in ann["segments_info"]]ret.append({"file_name": image_file,"image_id": image_id,"pan_seg_file_name": label_file,"segments_info": segments_info,})assert len(ret), f"No images found in {image_dir}!"assert PathManager.isfile(ret[0]["file_name"]), ret[0]["file_name"]assert PathManager.isfile(ret[0]["pan_seg_file_name"]), ret[0]["pan_seg_file_name"]return retif __name__ == "__main__":from detectron2.utils.logger import setup_loggerfrom detectron2.utils.visualizer import Visualizerimport detectron2.data.datasets  # noqa # add pre-defined metadatafrom PIL import Imageimport matplotlib.pyplot as pltimport numpy as nplogger = setup_logger(name=__name__)meta = MetadataCatalog.get("coco_2017_train_panoptic")dicts = load_coco_panoptic_json("/home/xxxx/Downloads/coco2017/annotations/panoptic_train2017.json","/home/xxxx/Downloads/coco2017/train2017","/home/xxxx/Downloads/coco2017/Mask/panoptic mask/panoptic_train2017", meta.as_dict())logger.info("Done loading {} samples.".format(len(dicts)))dirname = "coco-data-vis"os.makedirs(dirname, exist_ok=True)new_dic = {}num_imgs_to_vis = 100for i, d in enumerate(dicts):img = np.array(Image.open(d["file_name"]))visualizer = Visualizer(img, metadata=meta)pan_seg, segments_info = visualizer.draw_dataset_dict(d)seg_cat = {0: 0}for segi in segments_info:seg_cat[segi["id"]] = segi["category_id"]mapped_seg = np.vectorize(seg_cat.get)(pan_seg)# 保存数组为txt文件save_name = "/home/xxxx/Downloads/coco2017/Mask/panoptic label/train/" + str(d["image_id"]) + ".txt"np.savetxt(save_name, mapped_seg, fmt='%i')new_dic[d["image_id"]] = {"image_name": d["file_name"], "label_name": save_name}# # 将numpy数组转换为PIL Image对象# img1 = Image.fromarray(np.uint8(mapped_seg))# # 缩放图片# img_resized = img1.resize((224, 224))# # 将PIL Image对象转换回numpy数组# mapped_seg_resized = np.array(img_resized)# # 创建一个新的图形# plt.figure(figsize=(6, 8))# # 使用imshow函数来显示数组，并使用cmap参数来指定颜色映射# plt.imshow(mapped_seg_resized, cmap='viridis')# # 显示图形# plt.show()# fpath = os.path.join(dirname, os.path.basename(d["file_name"]))# # vis.save(fpath)if i + 1 >= num_imgs_to_vis:# 将字典转换为json字符串json_data = json.dumps(new_dic)# 将json字符串写入文件with open('/home/xxxx/Downloads/coco2017/data_for_train.json', 'w') as f:f.write(json_data)break

dicts = load_coco_panoptic_json("/home/xxxx/Downloads/coco2017/annotations/panoptic_train2017.json","/home/xxxx/Downloads/coco2017/train2017","/home/xxxx/Downloads/coco2017/Mask/panoptic mask/panoptic_train2017", meta.as_dict())

load_coco_panoptic_json的第三个参数，就是下载panoptic的annotations时，里面包含的两个压缩包。

上面这段代码，还需要修改一下detectron2库内部的文件

pan_seg, segments_info = visualizer.draw_dataset_dict(d)

进入 draw_dataset_dict这个函数内部，将两个中间结果拿出来

603行和604行是我自己加的代码。

这两个中间结果拿出来后，pan_seg是图片每个像素所属的种类，但是这个种类不是coco分类的那183个类

pan_seg中的数，要映射到segments_info中的 category_id，这个才是coco数据集所规定的183个类。下图节选了183个中的前10个展示

mapped_seg = np.vectorize(seg_cat.get)(pan_seg)

上面这行代码就是完成pan_seg中的数，映射到segments_info中的 category_id。不要用两层的for循环，太低效了，numpy中有函数可以完成数组的映射。

这样就得到了语义分割的ground truth，也就是每个像素所属的种类。

通过以下代码训练网络

import argparse
import time
import os
import matplotlib.pyplot as plt
import numpy as np
from PIL import Image
import torch
import torch.optim as optim
from config import get_config
from build import build_model, build_transform, build_criterion
from utils import load_pretrained
from prepare_data import get_dataloaderCHECKPOINT_PATH = os.environ.get("PATH_CHECKPOINT", "saved_models/my_model/")
os.makedirs(CHECKPOINT_PATH, exist_ok=True)def parse_option():parser = argparse.ArgumentParser("Swin Backbone", add_help=False)parser.add_argument("--cfg", type=str, default="configs/swin/swin_tiny_patch4_window7_224_22k.yaml")parser.add_argument("--pretrained", default="checkpoint/swin_tiny_patch4_window7_224_22k.pth")parser.add_argument("--device", default="cuda")args = parser.parse_args()config = get_config(args)return configif __name__ == "__main__":device = torch.device("cuda" if torch.cuda.is_available() else "cpu")config = parse_option()model = build_model(config).to(device)load_pretrained(config, model)transform = build_transform(config)# img_addr = "/home/xxxx/Downloads/coco2017/train2017/000000000009.jpg"# img = Image.open(img_addr)# img = transform(img)# # img = img.numpy().transpose(1, 2, 0)# # plt.imshow(img)# # plt.show()# img = img.unsqueeze(0).cuda()# output = model(img)train_loader = get_dataloader(transform)criterion = build_criterion()num_epochs = 500for epoch in range(num_epochs):for batch_idx, (data, target) in enumerate(train_loader):data, target = data.to(device), target.to(device)output = model(data)output = output.permute(0, 2, 3, 1).contiguous()features = output.shape[-1]output = output.view(-1, features)  # (batchsize*width*height, features)target = target.view(-1)  # (batchsize*width*height)loss = criterion(output, target)optimizer = optim.AdamW(model.parameters(), lr=0.0001)optimizer.zero_grad()loss.backward()optimizer.step()if batch_idx % 10 == 0:print('Epoch [{}/{}], Step [{}/{}], Loss: {:.4f}'.format(epoch+1, num_epochs, batch_idx+1, len(train_loader), loss.item()))if (epoch + 1) % 10 == 0:torch.save(model.state_dict(), CHECKPOINT_PATH)print(f"Model weights saved at epoch {epoch+1}.")

网络输出的tensor的形状是（batch_size, 特征数，width, height）。将输出reshape成（batch_size*width*height, 特征数），将ground truth形状变成（batch_size*width*height，）这样的一个一维数组，数组中每个元素就是种类的index，代表该像素属于什么类。通过