1. yolo的txt标注文件转coco的json标注文件
1.1 标注格式
一般来说,现有的标注格式就是xml格式,yolo的txt格式还有coco的json标注特殊,我们使用yolov5项目来说标注文件就是一堆txt文件,文件名是对应的图像名,如下所示:
然后每个txt文件中,就存储着当前图像的标注信息,分别对于的是:类别,归一化后中心点的x坐标,归一化后中心点的y坐标,归一化后的目标框宽度w,归一化后的目标框高度h(此处归一化指的是除以图片宽和高)
0 0.17 0.36678200692041524 0.07 0.09688581314878893 0 0.35625 0.20415224913494812 0.0525 0.08304498269896193 0 0.6375000000000001 0.3788927335640139 0.065 0.10726643598615918 0 0.65 0.19896193771626297 0.03 0.04498269896193772 0 0.6725 0.29584775086505194 0.03 0.04498269896193772 1 0.79 0.32525951557093424 0.07 0.08996539792387544 1 0.91125 0.19377162629757785 0.0625 0.07612456747404844
但是,对于coco的标注格式来说,顺序是:左上角的x坐标,左上角的y坐标,目标框的宽度w,目标框的高度h
所以,对于yolo格式的标注文件,不仅仅要依次的读取每个图像的标注txt信息,还需要对其中的信息进行转换。
下面,需要对coco的json标注格式进行一个简要的说明
1.2 coco字段说明
对于这部分内容,基本是来源于网上资料的,详细可以查看参考资料1,2
不同于voc还有yolo,一张照片对应着一个xml文件或者是一个txt文件,coco是直接将所有图片以及对应的box信息写在了一个json文件里。通常整个coco目录长这样:
coco |______annotations # 存放标注信息 | |__train.json | |__val.json | |__test.json |______trainset # 存放训练集图像 |______valset # 存放验证集图像 |______testset # 存放测试集图像
一个标准的json文件包含如下信息:
{ "info": info, "images": [image], "annotations": [annotation], "licenses": [license], "categories": [categories] } info{ "description": "COCO 2017 Dataset", # 数据集描述 "url": "http://cocodataset.org", # 下载地址 "version": "1.0", # 版本 "year": 2017, # 年份 "contributor": "COCO Consortium", # 提供者 "date_created": "2017/09/01" # 数据创建日期 } image{ "file_name": "000000397133.jpg", # 图片名 "id": 397133 # 图片的ID编号(每张图片ID是唯一的) "height": 427, # 高 "width": 640, # 宽 "license": 4, "coco_url": "http://images.cocodataset.org/val2017/000000397133.jpg",# 网路地址路径 "date_captured": "2013-11-14 17:02:52", # 数据获取日期 "flickr_url": "http://farm7.staticflickr.com/6116/6255196340_da26cf2c9e_z.jpg",# flickr网路地址 } license{ "url": "http://creativecommons.org/licenses/by-nc-sa/2.0/", "id": 1, "name": "Attribution-NonCommercial-ShareAlike License" } categories{ "supercategory": "person", # 主类别 "id": 1, # 类对应的id (0 默认为背景) "name": "person" # 子类别 } annotations{ "id": # 指的是这个annotation的一个id "image_id": # 等同于前面image字段里面的id。 "category_id": # 类别id "segmentation": # 左上-右上-右下-坐下 依次四个点坐标 "area": # 标注区域面积 "bbox": # 标注框,x,y为标注框的左上角坐标。 "iscrowd": # 决定是RLE格式还是polygon格式 }
1.3 yolo转coco脚本
接下来就直接进行转换,代码是我基于参考资料4的基础上修改而来的。
参考代码:
import os import json import random import time from PIL import Image import csv coco_format_save_path = './coco' # 要生成的标准coco格式标签所在文件夹 yolo_format_classes_path = 'annotations.csv' # 类别文件,用csv文件表示,一行一个类 yolo_format_annotation_path = '../dataset/mask/labels/val' # yolo格式标签所在文件夹 img_pathDir = '../dataset/mask/images/val' # 图片所在文件夹 # 类别设置 categories = [] class_names = ['with_mask', 'without_mask', 'mask_weared_incorrect'] for label in class_names: categories.append({'id': class_names.index(label), 'name': label, 'supercategory': ""}) write_json_context = dict() # 写入.json文件的大字典 write_json_context['licenses'] = [{'name': "", 'id': 0, 'url': ""}] write_json_context['info'] = {'contributor': "", 'date_created': "", 'description': "", 'url': "", 'version': "", 'year': ""} write_json_context['categories'] = categories write_json_context['images'] = [] write_json_context['annotations'] = [] # 接下来的代码主要添加'images'和'annotations'的key值 imageFileList = os.listdir(img_pathDir) # 遍历该文件夹下的所有文件,并将所有文件名添加到列表中 img_id = 0 # 图片编号 anno_id = 0 # 标注标号 for i, imageFile in enumerate(imageFileList): if '_' not in imageFile: img_id += 1 imagePath = os.path.join(img_pathDir, imageFile) # 获取图片的绝对路径 image = Image.open(imagePath) # 读取图片 W, H = image.size # 获取图片的高度宽度 img_context = {} # 使用一个字典存储该图片信息 # img_name=os.path.basename(imagePath) img_context['id'] = img_id # 每张图像的唯一ID索引 img_context['width'] = W img_context['height'] = H img_context['file_name'] = imageFile img_context['license'] = 0 img_context['flickr_url'] = "" img_context['color_url'] = "" img_context['date_captured'] = "" write_json_context['images'].append(img_context) # 将该图片信息添加到'image'列表中 txtFile = imageFile.split('.')[0] + '.txt' # 获取该图片获取的txt文件 with open(os.path.join(yolo_format_annotation_path, txtFile), 'r') as fr: lines = fr.readlines() # 读取txt文件的每一行数据,lines2是一个列表,包含了一个图片的所有标注信息 for j, line in enumerate(lines): anno_id += 1 # 标注的id从1开始 bbox_dict = {} # 将每一个bounding box信息存储在该字典中 class_id, x, y, w, h = line.strip().split(' ') # 获取每一个标注框的详细信息 class_id, x, y, w, h = int(class_id), float(x), float(y), float(w), float(h) # 将字符串类型转为可计算的int和float类型 # 坐标转换 xmin = (x - w / 2) * W ymin = (y - h / 2) * H xmax = (x + w / 2) * W ymax = (y + h / 2) * H w = w * W h = h * H height, width = abs(ymax - ymin), abs(xmax - xmin) # bounding box的坐标信息 bbox_dict['id'] = anno_id # 每个标注信息的索引 bbox_dict['image_id'] = img_id # 当前图像的ID索引 bbox_dict['category_id'] = class_id # 类别信息 bbox_dict['segmentation'] = [[xmin, ymin, xmax, ymin, xmax, ymax, xmin, ymax]] bbox_dict['area'] = height * width bbox_dict['bbox'] = [xmin, ymin, w, h] # 注意目标类别要加一 bbox_dict['iscrowd'] = 0 bbox_dict['attributes'] = "" write_json_context['annotations'].append(bbox_dict) # 将每一个由字典存储的bounding box信息添加到'annotations'列表中 name = os.path.join(coco_format_save_path, "annotations" + '.json') with open(name, 'w') as fw: # 将字典信息写入.json文件中 json.dump(write_json_context, fw, indent=4, ensure_ascii=False)
运行结果:
{ "images": [ { "id": 1, "width": 400, "height": 267, "file_name": "maksssksksss98.png", "license": 0, "flickr_url": "", "color_url": "", "date_captured": "" }, ...... "annotations": [ { "id": 1, "image_id": 1, "category_id": 0, "segmentation": [ [ 196.00000000000003, 43.0, 236.00000000000003, 43.0, 236.00000000000003, 91.0, 196.00000000000003, 91.0 ] ], "area": 1920.0, "bbox": [ 196.00000000000003, 43.0, 40.0, 48.0 ], "iscrowd": 0, "attributes": "" }, { "id": 2, "image_id": 1, "category_id": 0, "segmentation": [ [ 41.0, 73.0, 65.0, 73.0, 65.0, 95.0, 41.0, 95.0 ] ], "area": 528.0, "bbox": [ 41.0, 73.0, 24.0, 22.000000000000004 ], "iscrowd": 0, "attributes": "" }, ...... }
这样,就可以将全部的标注txt文件,转化成一个json文件的标注信息
2. 按coco格式获取预测结果的json文件
基于以上的操作,现在已经得到了coco格式的json标注文件。根据API的调用,现在还需要将预测信息整合在一个json文件中,对于每副图像需要获取其所有预测框的类别,边界框的4个坐标,置信度。将所有结果保留为一个列表,输入如下所示:
[ { "image_id": "maksssksksss363", "category_id": 0, "bbox": [ 342.638, 86.238, 36.37, 39.355 ], "score": 0.91578 }, { "image_id": "maksssksksss363", "category_id": 0, "bbox": [ 327.98, 21.8, 38.32, 41.232 ], "score": 0.9059 }, ...... ]
这个预测文件在原本的val.py脚本中,设置--save-json参数基于可以输出
def parse_opt(): parser = argparse.ArgumentParser() parser.add_argument('--save-json', default=True, action='store_true', help='save a COCO-JSON results file') ...... def run(...): # Save JSON if save_jsonand len(jdict): w = Path(weights[0] if isinstance(weights, list) else weights).stem if weights is not None else '' # weights anno_json = str(Path(data.get('path', '../coco')) / 'annotations/instances_val2017.json') # annotations json pred_json = str(save_dir / f"{w}_predictions.json") # predictions json print(f'\nEvaluating pycocotools mAP... saving {pred_json}...') # 保存val的所有预测结果在jdict字典中,然后保存名称为:best_predictions.json with open(pred_json, 'w') as f: json.dump(jdict, f, indent=4, ensure_ascii=False)
输入路径如下所示:
对于jdict字典中的每一个内容,是通过save_one_json函数来保存设置的:
# 将预测信息保存到coco格式的json字典 def save_one_json(predn, jdict, path, class_map): # Save one JSON result {"image_id": 42, "category_id": 18, "bbox": [258.15, 41.29, 348.26, 243.78], "score": 0.236} # 获取图片id image_id = int(path.stem) if path.stem.isnumeric() else path.stem # 获取预测框 并将xyxy转为xywh格式 box = xyxy2xywh(predn[:, :4]) # xywh # 之前的的xyxy格式是左上角右下角坐标 xywh是中心的坐标和宽高 # 而coco的json格式的框坐标是xywh(左上角坐标 + 宽高) # 所以这行代码是将中心点坐标 -> 左上角坐标 box[:, :2] -= box[:, 2:] / 2 # xy center to top-left corner # image_id: 图片id 即属于哪张图片 # category_id: 类别 coco91class()从索引0~79映射到索引0~90 # bbox: 预测框坐标 # score: 预测得分 for p, b in zip(predn.tolist(), box.tolist()): jdict.append({'image_id': image_id, 'category_id': class_map[int(p[5])], 'bbox': [round(x, 3) for x in b], 'score': round(p[4], 5)})
那么,现在有了对val数据集的标注信息json文件,也有了val数据集的预测信息json文件,就可以使用pycocotools.cocoeval工具包来进行map的判断,这样就不需要像yolov5那样写了一大堆复杂的评价函数。
3. 使用coco API评估结果
使用coco api评估当前数据集的map结果非常简单,只需要将coco格式的标注json文件和coco格式的预测json文件同时传入COCOeval函数中即可,代码如下:
from pycocotools.coco import COCO from pycocotools.cocoeval import COCOeval if __name__ == '__main__': anno_json = r'./test/anno_json.json' pred_json = r'./test/pred_json.json' anno = COCO(anno_json) # init annotations api pred = anno.loadRes(pred_json) # init predictions api eval = COCOeval(anno, pred, 'bbox') eval.evaluate() eval.accumulate() eval.summarize() map, map50 = eval.stats[:2] # update results (mAP@0.5:0.95, mAP@0.5) print(eval.stats)
这时候如果直接传入刚刚的两个json文件,是会报错的,错误信息是:AssertionError: Results do not correspond to current coco set。参考资料6.
出现这个问题的原因有两个:
- 图像id和标注的id数量不对应。也就是说出现了一些没有标注的图像信息,在image列表中出现,但是却没有在annotations中出现,也就是有点图像没有目标没有标注。
- image_id 类型出现错误,image_id 必须为 int类型,不能是字符串
随后,我检查了一下txt标注文件,发现所有的图像都有目标,都有标注,也就排除了第一个问题。(假如是因为第一个问题,需要把标注信息为空的图像进行删除,这个操作其实挺不合理的)。那么,就是第二个问题了。
然后,我们的预测json文件中,image_id 是图像的文件名。image_id 必须为 int类型,不能是字符串。为什么会出现这个错误?原因是在save_one_json()函数主要注意image_id = int(path.stem) if path.stem.isnumeric() else path.stem这一句出现的了问题,因为我们传入的 path.stem 本身就是一个字符串。
path.stem是指验证集图片名,如host0000001.jpg
那么path.stem为host0000001,则取数字部分:path.stem[5:] #为0000001
由于本身就是字符串,所以判断后的image_id 传入还是字符串,导致了这个错误。同样的,在标注信息的json文件中,也出现了这个错误。
- annotations.json的错误:
"annotations": [ { "id": 1, "image_id": "maksssksksss98", # 错误,需要是int类型,和image信息相匹配 "category_id": 0, "segmentation": [ [ 196.00000000000003, 43.0, 236.00000000000003, 43.0, 236.00000000000003, 91.0, 196.00000000000003, 91.0 ] ], "area": 1920.0, "bbox": [ 196.00000000000003, 43.0, 40.0, 48.0 ], "iscrowd": 0, "attributes": "" },
- best_preditions.json的错误:
{ "image_id": "maksssksksss363", # 错误,需要是int类型,和image信息相匹配 "category_id": 0, "bbox": [ 342.638, 86.238, 36.37, 39.355 ], "score": 0.91578 },
那么,现在知道了错误的原因,就需要将问题改正。对于这些字符串,我们需要和annotations.json字典中的images信息来进行匹配,在对应的地方转为id,而不是图像名。比如:
"images": [ { "id": 1, "width": 400, "height": 267, "file_name": "maksssksksss98.png", "license": 0, "flickr_url": "", "color_url": "", "date_captured": "" },
也就是说,将原本image_id为maksssksksss98的内容,改为1,因为匹配的是id是1。基于这一点,下面就写了一个修正脚本:
''' 修正脚本:对预测的json文件还有标注的json文件的id信息根据标注文件的image来命名 ''' import json import os from collections import OrderedDict # 获取标注文件图像id与图像名字的字典 def get_name2id_map(image_dict): name2id_dict = OrderedDict() for image in image_dict: file_name = image['file_name'].split('.')[0] # maksssksksss98.png -> maksssksksss98 id = image['id'] name2id_dict[file_name] = id return name2id_dict if __name__ == '__main__': anno_json = r'./coco/annotations.json' pred_json = r'../runs/val/mask/best_predictions.json' with open(pred_json, 'r') as fr: pred_dict = json.load(fr) with open(anno_json, 'r') as fr: anno_dict = json.load(fr) name2id_dict = get_name2id_map(anno_dict['images']) # 对标注文件annotations的image_id进行更改 for annotations in anno_dict['annotations']: image_id = annotations['image_id'] annotations['image_id'] = int(name2id_dict[image_id]) # 对预测文件的image_id同样进行更改 for predictions in pred_dict: image_id = predictions['image_id'] predictions['image_id'] = int(name2id_dict[image_id]) # 分别保存更改后的标注文件和预测文件 with open('anno_json.json', 'w') as fw: json.dump(anno_dict, fw, indent=4, ensure_ascii=False) with open('pred_json.json', 'w') as fw: json.dump(pred_dict, fw, indent=4, ensure_ascii=False)
输出两个修正后的json文件:
现在重新查看修正后的标注信息:
# pred_json.json { "image_id": 112, # 这里需要修改为图像的的ID索引 "category_id": 0, "bbox": [ 342.638, 86.238, 36.37, 39.355 ], "score": 0.91578 }, ... # anno_json.json "annotations": [ { "id": 1, "image_id": 1, # 由于图像的读取顺序是固定的,所以这里的image_id其实也就是id "category_id": 0, "segmentation": [ [ 196.00000000000003, 43.0, 236.00000000000003, 43.0, 236.00000000000003, 91.0, 196.00000000000003, 91.0 ] ], "area": 1920.0, "bbox": [ 196.00000000000003, 43.0, 40.0, 48.0 ], "iscrowd": 0, "attributes": "" },
经过如此修正之后,就可以正常的调用coco的api了。
- COCO API评估代码:
from pycocotools.coco import COCO from pycocotools.cocoeval import COCOeval if __name__ == '__main__': anno_json = r'./test/anno_json.json' pred_json = r'./test/pred_json.json' anno = COCO(anno_json) # init annotations api pred = anno.loadRes(pred_json) # init predictions api eval = COCOeval(anno, pred, 'bbox') eval.evaluate() eval.accumulate() eval.summarize() map, map50 = eval.stats[:2] # update results (mAP@0.5:0.95, mAP@0.5) print(eval.stats)
输出信息:
loading annotations into memory... Done (t=0.00s) creating index... index created! Loading and preparing results... DONE (t=0.01s) creating index... index created! Running per image evaluation... Evaluate annotation type *bbox* DONE (t=0.45s). Accumulating evaluation results... DONE (t=0.05s). Average Precision (AP) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.494 Average Precision (AP) @[ IoU=0.50 | area= all | maxDets=100 ] = 0.764 Average Precision (AP) @[ IoU=0.75 | area= all | maxDets=100 ] = 0.545 Average Precision (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.392 Average Precision (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.680 Average Precision (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.853 Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 1 ] = 0.269 Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 10 ] = 0.565 Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.591 Average Recall (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.503 Average Recall (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.755 Average Recall (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.868 Process finished with exit code 0
普通执行val.py函数的预测信息:
(yolov5) [fs@localhost yolov5-6.0]$ python val.py val: data=./dataset/mask/mask.yaml, weights=./runs/train/mask/weights/best.pt, batch_size=32, imgsz=640, conf_thres=0.001, iou_thres=0.6, task=val, device=cpu, single_cls=False, augment=False, verbose=False, save_txt=False, save_hybrid=False, save_conf=False, save_json=True, project=runs/val, name=exp, exist_ok=False, half=False YOLOv5 🚀 2022-6-10 torch 1.9.1 CPU Fusing layers... Model Summary: 213 layers, 7018216 parameters, 0 gradients, 15.8 GFLOPs val: Scanning 'dataset/mask/labels/val.cache' images and labels... 171 found, 0 missing, 0 empty, 0 corrupted: 100%|█| 171/171 [00:00< Class Images Labels P R mAP@.5 mAP@.5:.95: 100%|███████████| 6/6 [00:06<00:00, 1.11s/it] all 171 754 0.808 0.708 0.767 0.501 with_mask 171 630 0.96 0.881 0.943 0.656 without_mask 171 104 0.799 0.841 0.872 0.552 mask_weared_incorrect 171 20 0.666 0.4 0.486 0.296 Speed: 0.3ms pre-process, 26.4ms inference, 0.5ms NMS per image at shape (32, 3, 640, 640) Evaluating pycocotools mAP... saving runs/val/exp3/best_predictions.json... loading annotations into memory... pycocotools unable to run: [Errno 2] No such file or directory: 'dataset/mask/annotations/instances_val2017.json' Results saved to runs/val/exp3
4. val脚本简化
到了这里,就可以直接将cal脚本给简化了。现在,我再总结一下,需要两个步骤。
第一步:首选需要将yolo的txt目录转换成coco的json文件,参考代码:
yolo2coco.py:
import os import json import random import time from PIL import Image import csv coco_format_save_path = './coco' # 要生成的标准coco格式标签所在文件夹 yolo_format_classes_path = 'annotations.csv' # 类别文件,用csv文件表示,一行一个类 yolo_format_annotation_path = '../dataset/mask/labels/val' # yolo格式标签所在文件夹 img_pathDir = '../dataset/mask/images/val' # 图片所在文件夹 # with open(yolo_format_classes_path, 'r') as f: # reader = csv.reader(f) # for label in reader: # print(label) # categories = [] # for i in label: # categories.append({'id': label.index(i) + 1, 'name': i, 'supercategory': ""}) # 存储类别 categories = [] class_names = ['with_mask', 'without_mask', 'mask_weared_incorrect'] for label in class_names: categories.append({'id': class_names.index(label), 'name': label, 'supercategory': ""}) write_json_context = dict() # 写入.json文件的大字典 write_json_context['licenses'] = [{'name': "", 'id': 0, 'url': ""}] write_json_context['info'] = {'contributor': "", 'date_created': "", 'description': "", 'url': "", 'version': "", 'year': ""} write_json_context['categories'] = categories write_json_context['images'] = [] write_json_context['annotations'] = [] # 接下来的代码主要添加'images'和'annotations'的key值 imageFileList = os.listdir(img_pathDir) # 遍历该文件夹下的所有文件,并将所有文件名添加到列表中 img_id = 0 # 图片编号 anno_id = 0 # 标注标号 for i, imageFile in enumerate(imageFileList): if '_' not in imageFile: img_id += 1 imagePath = os.path.join(img_pathDir, imageFile) # 获取图片的绝对路径 image = Image.open(imagePath) # 读取图片 W, H = image.size # 获取图片的高度宽度 img_context = {} # 使用一个字典存储该图片信息 # img_name=os.path.basename(imagePath) img_context['id'] = img_id # 每张图像的唯一ID索引 img_context['width'] = W img_context['height'] = H img_context['file_name'] = imageFile img_context['license'] = 0 img_context['flickr_url'] = "" img_context['color_url'] = "" img_context['date_captured'] = "" write_json_context['images'].append(img_context) # 将该图片信息添加到'image'列表中 txtFile = imageFile.split('.')[0] + '.txt' # 获取该图片获取的txt文件 with open(os.path.join(yolo_format_annotation_path, txtFile), 'r') as fr: lines = fr.readlines() # 读取txt文件的每一行数据,lines2是一个列表,包含了一个图片的所有标注信息 for j, line in enumerate(lines): anno_id += 1 # 标注的id从1开始 bbox_dict = {} # 将每一个bounding box信息存储在该字典中 class_id, x, y, w, h = line.strip().split(' ') # 获取每一个标注框的详细信息 class_id, x, y, w, h = int(class_id), float(x), float(y), float(w), float(h) # 将字符串类型转为可计算的int和float类型 # 坐标转换 xmin = (x - w / 2) * W ymin = (y - h / 2) * H xmax = (x + w / 2) * W ymax = (y + h / 2) * H w = w * W h = h * H height, width = abs(ymax - ymin), abs(xmax - xmin) # bounding box的坐标信息 bbox_dict['id'] = anno_id # 每个标注信息的索引 bbox_dict['image_id'] = img_id # 当前图像的ID索引 bbox_dict['category_id'] = class_id # 类别信息 bbox_dict['segmentation'] = [[xmin, ymin, xmax, ymin, xmax, ymax, xmin, ymax]] bbox_dict['area'] = height * width bbox_dict['bbox'] = [xmin, ymin, w, h] # 注意目标类别要加一 bbox_dict['iscrowd'] = 0 bbox_dict['attributes'] = "" write_json_context['annotations'].append(bbox_dict) # 将每一个由字典存储的bounding box信息添加到'annotations'列表中 name = os.path.join(coco_format_save_path, "annotations" + '.json') with open(name, 'w') as fw: # 将字典信息写入.json文件中 json.dump(write_json_context, fw, indent=4, ensure_ascii=False)
将val数据集的所以txt信息,就可以转换成一个json文件了,获得annotations.json文件
第二步:遍历带检测目录下的全部头像,依次检测每张图像,将每一个预测结果全部依次添加在一个列表中,同样构建一个预测的json文件,获得preditions.json文件。
那么,根据标注信息的annotations.json文件和预测结果的preditions.json文件就可以调用coco的api完成一个简单的处理。
val_simplify.py:
import torch import cv2 import numpy as np import os import json from tqdm import tqdm from models.experimental import attempt_load from utils.augmentations import letterbox from utils.general import check_img_size, non_max_suppression, scale_coords, xyxy2xywh from pycocotools.coco import COCO from pycocotools.cocoeval import COCOeval # 获取标注文件图像id与图像名字的字典 def get_name2id_map(): # 获取标注文件的images标注信息 anno_json = r'./test/coco/annotations.json' with open(anno_json, 'r') as fr: anno_dict = json.load(fr) image_dict = anno_dict['images'] # 构建图像名称与索引的字典对 name2id_dict = {} for image in image_dict: # file_name = image['file_name'].split('.')[0] # maksssksksss98.png -> maksssksksss98 file_name = image['file_name'] id = image['id'] name2id_dict[file_name] = id return name2id_dict # 功能:单图像推理 def val(image_dir, img_size=640, stride=32, augment=False, visualize=False): device = 'cpu' weights = r'./runs/train/mask/weights/best.pt' # 权重路径 anno_json = r'./test/coco/annotations.json' # 已处理的标注信息json文件 pred_json = 'preditions.json' # 带保存的预测信息json文件 # 导入模型 model = attempt_load(weights, map_location=device) img_size = check_img_size(img_size, s=stride) # names = model.names jdict = [] name2id_dict = get_name2id_map() image_list = os.listdir(image_dir) # 依次预测每张图像,将预测信息全部保存在json文件中 for image_name in tqdm(image_list, desc='val image'): # Padded resize image_path = image_dir + os.sep + image_name img0 = cv2.imread(image_path) img = letterbox(img0, img_size, stride=stride, auto=True)[0] # Convert img = img.transpose((2, 0, 1))[::-1] # HWC to CHW, BGR to RGB img = np.ascontiguousarray(img) img = torch.from_numpy(img).to(device) img = img.float() / 255.0 # 0 - 255 to 0.0 - 1.0 img = img[None] # [h w c] -> [1 h w c] # inference pred = model(img, augment=augment, visualize=visualize)[0] pred = non_max_suppression(pred, conf_thres=0.25, iou_thres=0.45, max_det=1000) # plot label det = pred[0] # annotator = Annotator(img0.copy(), line_width=3, example=str(names)) if len(det): # (xyxy, conf, cls) det[:, :4] = scale_coords(img.shape[2:], det[:, :4], img0.shape).round() # bbox格式需要转换: xyxy -> (x_min, y_min, width, height) bbox = xyxy2xywh(det[:, :4]) bbox[:, :2] -= bbox[:, 2:] / 2 # xy center to top-left corner score = det[:, 4] category_id = det[:, -1] for box, src, cls in zip(bbox, score, category_id): jdict.append( {'image_id': name2id_dict[image_name], 'category_id': int(cls), 'bbox': box.tolist(), 'score': float(src)} ) # 保存预测好的json文件 with open(pred_json, 'w') as fw: json.dump(jdict, fw, indent=4, ensure_ascii=False) # 使用coco api评价指标 anno = COCO(anno_json) # init annotations api pred = anno.loadRes(pred_json) # init predictions api eval = COCOeval(anno, pred, 'bbox') eval.evaluate() eval.accumulate() eval.summarize() if __name__ == '__main__': image_dir = r'./dataset/mask/images/val' val(image_dir=image_dir)
输出结果:
val image: 100%|██████████████████████████████| 171/171 [00:06<00:00, 24.88it/s] loading annotations into memory... Done (t=0.00s) creating index... index created! Loading and preparing results... DONE (t=0.00s) creating index... index created! Running per image evaluation... Evaluate annotation type *bbox* DONE (t=0.15s). Accumulating evaluation results... DONE (t=0.02s). Average Precision (AP) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.465 Average Precision (AP) @[ IoU=0.50 | area= all | maxDets=100 ] = 0.701 Average Precision (AP) @[ IoU=0.75 | area= all | maxDets=100 ] = 0.525 Average Precision (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.360 Average Precision (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.648 Average Precision (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.848 Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 1 ] = 0.247 Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 10 ] = 0.493 Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.514 Average Recall (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.409 Average Recall (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.713 Average Recall (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.867 Process finished with exit code 0
最后,简化成这样的直观的脚本就是我最后的目标,比一开始的代码容易看多了。最后的输出结果也比较清晰明了。但是,缺点是相比与yolov5原始的验证指标,这里没有为每一个类单独的输出map结果,只是对全部的所有类进行的测试,这是不足之处。
- 后续:
以上,我已经分别的简化的yolov5项目的val脚本与detect脚本,同时也对yolov5的网络结构,训练策略以及最基本的使用方法一一说明,那么yolov5项目的学习就到此结束了。这应该是这个专栏的最后一篇博文。感谢大家的关注与支持。
参考资料:
1. COCO数据集标注格式及意义
2. COCO数据集的标注格式
3. 如何将VOC XML文件转化成COCO数据格式
4. python 图像检测之yolo txt格式转成coco json格式
5. 【目标检测】coco工具包验证时错误Results do not correspond to current coco set
6. yolov5 调用cocotools 评价自己的模型和数据集(AP低的问题已解决)