安装labelme2coco，执行如下命令：
bash
pip install labelme2coco将使用LabelMe生成的.json文件转换成coco支持的json文件，执行以下命令：
bash
labelme2coco melon_detect/train --category_id_start 1 # 生成的dataset.json重命名为train.json
labelme2coco melon_detect/val --category_id_start 1 # 生成的dataset.json重命名为val.json结果如下所示：
bash
{
"images":
[{"height":340,"width":437,"id":1,"file_name":"melon_detect\\val\\0940ff7f2754e74b5ad6d13495cd13e1_t.jpeg"},
{"height":534,"width":800,"id":2,"file_name":"melon_detect\\val\\14896085.webp"},
{"height":693,"width":1024,"id":3,"file_name":"melon_detect\\val\\21319742_204241990080_2.jpg"},
{"height":410,"width":600,"id":4,"file_name":"melon_detect\\val\\213511.jpg"},
{"height":1157,"width":1600,"id":5,"file_name":"melon_detect\\val\\222859661.webp"}],
"annotations":
[{"iscrowd":0,"image_id":1,"bbox":[35.65817694369973,171.87667560321717,234.85254691689005,133.51206434316356],"segmentation":[],"category_id":1,"id":1,"area":31355},
{"iscrowd":0,"image_id":1,"bbox":[168.36595174262735,32.46648793565684,250.40214477211802,259.51742627345845],"segmentation":[],"category_id":1,"id":2,"area":64983},
{"iscrowd":0,"image_id":2,"bbox":[105.90717299578058,250.12236286919833,421.5189873417721,232.48945147679322],"segmentation":[],"category_id":1,"id":3,"area":97998},
{"iscrowd":0,"image_id":2,"bbox":[266.2447257383966,39.99578059071733,385.65400843881855,355.27426160337546],"segmentation":[],"category_id":1,"id":4,"area":137012},
{"iscrowd":0,"image_id":3,"bbox":[205.44262295081967,17.53825136612023,618.5792349726776,598.9071038251367],"segmentation":[],"category_id":1,"id":5,"area":370471}],
"categories":
[{"id":1,"name":"watermelon","supercategory":"watermelon"},
{"id":2,"name":"wintermelon","supercategory":"wintermelon"}]
}注：
1.LabelMe的label索引是从0开始的，而COCO类别ID必须从1开始，所以必须要加：--category_id_start 1
2.LabelMe的bbox为{x1,y1,x2,y2}，即左上角到右下角的两个点坐标。labelme2coco转换为COCO格式时将bbox自动转为{x1,y1,w,h}，但PyTorch中的目标检测需要{x1,y1,x2,y2}，所以实现中需要：boxes.append([x, y, x + w, y + h])
CocoDataset类实现如下：
python
import argparse
import colorama
import json
from torch.utils.data import Dataset, DataLoader
from torchvision import transforms
from PIL import Image
from pathlib import Path
def parse_args():
parser = argparse.ArgumentParser(description="labelme to coco")
parser.add_argument("--dataset_path", type=str, help="image dataset path")
parser.add_argument("--json_file", type=str, help="image dataset json file, coco format")
args = parser.parse_args()
return args
class CocoDataset(Dataset):
def __init__(self, json_file, image_dir, transforms=None):
self.image_dir = image_dir
self.transforms = transforms
with open(json_file, mode="r", encoding="utf-8") as f:
coco = json.load(f)
self.image_infos = {img["id"]: img for img in coco["images"]}
self.image_ids = list(self.image_infos.keys())
self.annotations = {}
for ann in coco["annotations"]:
img_id = ann["image_id"]
self.annotations.setdefault(img_id, []).append(ann)
def __len__(self):
return len(self.image_ids)
def __getitem__(self, idx):
img_id = self.image_ids[idx]
img_info = self.image_infos[img_id]
file_path = Path(self.image_dir) / Path(img_info["file_name"]).name
if not file_path.exists():
raise FileNotFoundError(colorama.Fore.RED + f"image not found: {file_path}")
image = Image.open(file_path).convert("RGB")
anns = self.annotations.get(img_id, [])
boxes, labels, areas, iscrowd = [], [], [], []
for ann in anns:
x, y, w, h = ann["bbox"]
if w <= 0 or h <= 0:
continue
# COCO bbox:[xmin, ymin, width, height]; PyTorch need:[xmin, ymin, xmax, ymax]
boxes.append([x, y, x + w, y + h])
labels.append(ann["category_id"])
areas.append(ann.get("area", w * h))
iscrowd.append(ann.get("iscrowd", 0))
target = {
"boxes": boxes,
"labels": labels,
"image_id": img_id,
"areas": areas,
"iscrowd": iscrowd,
}
if self.transforms is not None:
image = self.transforms(image)
return image, target
# custom batch processing function to process different numbers of targets
def collate_fn(batch):
return tuple(zip(*batch))
if __name__ == "__main__":
colorama.init(autoreset=True)
args = parse_args()
transform = transforms.Compose([
transforms.ToTensor()
])
dataset = CocoDataset(args.json_file, args.dataset_path, transforms=transform)
data_loader = DataLoader(dataset, batch_size=2, shuffle=True, num_workers=2, collate_fn=collate_fn)
for idx, (images, targets) in enumerate(data_loader):
print(f"idx: {idx}")
for i in range(len(images)):
print(f"image{i} shape: {images[i].shape}")
print(f"boxes{i}: {targets[i]['boxes']}")
print(f"labels{i}: {targets[i]['labels']}")
print(colorama.Fore.GREEN + "====== execution completed ======")执行结果如下图所示：
DataLoader中的:
1.collate_fn参数：当每个样本(image)中的target数量不一致或target的shape不一致或样本shape不一致时必须指定collate_fn。
2.num_workers参数：当数据加载速度跟不上训练速度时，需指定num_workers>0，控制后台使用多少个子进程(并行)来加载数据。