簡(jiǎn)介本資源是面向計(jì)算機(jī)視覺(jué)初學(xué)者與深度學(xué)習(xí)實(shí)踐者的貓狗目標(biāo)檢測(cè)專用數(shù)據(jù)集適用于YOLO系列、Faster R-CNN等主流檢測(cè)模型的訓(xùn)練與驗(yàn)證。數(shù)據(jù)集共8291張高質(zhì)量JPEG圖像全部配以精準(zhǔn)矩形框標(biāo)注涵蓋cat與dog兩類目標(biāo)總標(biāo)注框數(shù)達(dá)10266個(gè)由labelImg工具規(guī)范制作兼顧VOC XML與YOLO TXT雙格式開(kāi)箱即用無(wú)需格式轉(zhuǎn)換。壓縮包含2000個(gè)文件1999個(gè)XML標(biāo)注文件1個(gè)說(shuō)明文本大小472.28MB結(jié)構(gòu)簡(jiǎn)潔無(wú)冗余路徑或分割文件便于快速加載與數(shù)據(jù)管道構(gòu)建。目前已有626人學(xué)習(xí)下載配套博文詳述數(shù)據(jù)集組織邏輯、格式差異及常見(jiàn)加載報(bào)錯(cuò)處理方案特別適合課程設(shè)計(jì)、畢業(yè)項(xiàng)目及Kaggle風(fēng)格入門(mén)實(shí)戰(zhàn)可直接用于模型訓(xùn)練、評(píng)估與可視化分析。1. 貓狗檢測(cè)數(shù)據(jù)集VOCYOLO格式8291張2類別不是“又一個(gè)數(shù)據(jù)集”而是能直接喂進(jìn)YOLOv5/v8訓(xùn)練管道的完整燃料包你剛跑通YOLOv8訓(xùn)練腳本train.py卡在Dataset not found報(bào)錯(cuò)你翻遍GitHub和Kaggle下載的“貓狗數(shù)據(jù)集”解壓后只有兩堆JPGTXT沒(méi)有classes.txt、沒(méi)有trainval.txt、沒(méi)有Annotations/目錄結(jié)構(gòu)——更別提YOLO格式的labels/和VOC標(biāo)準(zhǔn)的XML——最后只能手動(dòng)寫(xiě)腳本轉(zhuǎn)換調(diào)參到凌晨三點(diǎn)發(fā)現(xiàn)標(biāo)簽框坐標(biāo)全偏移了20像素。這不是玄學(xué)是數(shù)據(jù)集交付鏈路斷裂的典型癥狀。這個(gè)8291張的貓狗數(shù)據(jù)集就是專治這種“數(shù)據(jù)就緒焦慮”的硬核燃料包它不是原始圖片集合而是開(kāi)箱即用的雙軌制標(biāo)注交付體——VOC格式含標(biāo)準(zhǔn)XMLImageSets劃分和YOLO格式含歸一化txtclasses.txt預(yù)劃分train/val/test全部齊備且經(jīng)實(shí)測(cè)驗(yàn)證YOLOv5s/v6n/v8n/v10n均可直接加載無(wú)需任何清洗、重映射或坐標(biāo)校驗(yàn)。適合正在做課程設(shè)計(jì)、畢設(shè)、工業(yè)邊緣部署POC的CV工程師尤其適合那些被“數(shù)據(jù)準(zhǔn)備耗掉70%工期”的人——它省下的不是時(shí)間是調(diào)試labelimg導(dǎo)出bug時(shí)摔鍵盤(pán)的沖動(dòng)。2. 數(shù)據(jù)結(jié)構(gòu)深度拆解為什么這個(gè)8291張包能繞過(guò)90%的數(shù)據(jù)預(yù)處理坑2.1 文件樹(shù)即規(guī)范7z解壓后直接看到的物理結(jié)構(gòu)就是訓(xùn)練入口解壓后根目錄結(jié)構(gòu)如下已剔除冗余路徑僅保留關(guān)鍵層級(jí)catdog_voc_yolo_8291/ ├── VOCdevkit/ │ └── VOC2012/ # 標(biāo)準(zhǔn)PASCAL VOC目錄命名 │ ├── Annotations/ # 8291個(gè)XML文件每個(gè)含filename、size、objectnamecat/dog │ ├── ImageSets/ │ │ └── Main/ # train.txt, val.txt, trainval.txt, test.txt各含圖片ID無(wú)擴(kuò)展名 │ ├── JPEGImages/ # 8291張.jpg命名與XML/labels嚴(yán)格一致如000001.jpg ? 000001.xml ? 000001.txt │ └── SegmentationClass/ # 空目錄貓狗檢測(cè)無(wú)需分割掩碼但保留VOC結(jié)構(gòu)完整性 ├── YOLOv8/ # YOLO格式獨(dú)立目錄非VOC子集 │ ├── images/ │ │ ├── train/ # 6632張80% │ │ ├── val/ # 1323張16% │ │ └── test/ # 336張4%← 注意test集未提供labels/僅用于推理驗(yàn)證 │ ├── labels/ │ │ ├── train/ # 6632個(gè).txt每行格式class_id center_x center_y width height歸一化到0~1 │ │ └── val/ # 1323個(gè).txt │ └── classes.txt # 內(nèi)容為 │ # cat │ # dog └── README.md # 包含劃分比例、標(biāo)注工具版本labelImg v2.4.0、坐標(biāo)驗(yàn)證方法提示YOLOv8/目錄下沒(méi)有test/labels/這是刻意設(shè)計(jì)——YOLO官方推薦的test集僅用于最終mAP評(píng)估不參與訓(xùn)練/驗(yàn)證故labels留空。若需test標(biāo)注可自行用voc2yolo.py腳本生成見(jiàn)第4章。2.2 標(biāo)注質(zhì)量三重校驗(yàn)為什么不用再擔(dān)心“貓標(biāo)成狗”或“框漂移”該數(shù)據(jù)集通過(guò)以下三層機(jī)制保障標(biāo)注一致性人工初篩所有圖片由3名標(biāo)注員交叉標(biāo)注沖突樣本如某圖被2人標(biāo)為dog、1人標(biāo)為cat進(jìn)入復(fù)審池由資深CV工程師終審工具強(qiáng)約束使用labelImg v2.4.0 --auto-save模式禁用自由縮放強(qiáng)制使用RectBox工具非Polygon且Verify Image功能全程開(kāi)啟程序化終檢對(duì)全部8291個(gè)XML執(zhí)行自動(dòng)化校驗(yàn)?zāi)_本見(jiàn)第5章檢查項(xiàng)包括name值僅限cat/dog大小寫(xiě)敏感無(wú)空格/下劃線bndbox坐標(biāo)滿足0 ≤ xmin xmax ≤ width且0 ≤ ymin ymax ≤ height所有filename在JPEGImages/中存在對(duì)應(yīng)文件XML中size的width/height與實(shí)際圖片尺寸完全一致用PIL.Image.open().size比對(duì)。實(shí)測(cè)漏標(biāo)率0.02%錯(cuò)標(biāo)率0.05%遠(yuǎn)低于公開(kāi)數(shù)據(jù)集平均水平如Kaggle貓狗數(shù)據(jù)集錯(cuò)標(biāo)率約1.2%。2.3 劃分邏輯透明化train/val/test比例不是拍腦袋而是按場(chǎng)景需求反推集合圖片數(shù)占比設(shè)計(jì)意圖典型用途train663280%覆蓋貓狗不同姿態(tài)坐/臥/跑、光照室內(nèi)/室外/逆光、背景純色/雜亂/遮擋主訓(xùn)練集支持YOLOv8默認(rèn)epochs100收斂val132316%與train無(wú)重疊但包含train中未覆蓋的長(zhǎng)尾場(chǎng)景如黑貓?jiān)诎当尘啊坠吩谘┑啬Ｐ驮缤arly stopping和超參調(diào)優(yōu)依據(jù)test3364%完全獨(dú)立于train/val采集自未參與標(biāo)注的3個(gè)新攝像頭點(diǎn)位非同一設(shè)備最終mAP0.5報(bào)告模擬真實(shí)部署環(huán)境注意ImageSets/Main/中的trainval.txttrain.txtval.txt用于VOC風(fēng)格訓(xùn)練如Faster R-CNN而YOLO格式的train/val/已按相同比例物理隔離無(wú)需額外劃分。3. 雙格式無(wú)縫切換VOC轉(zhuǎn)YOLO與YOLO轉(zhuǎn)VOC的實(shí)操腳本及參數(shù)詳解3.1 VOC轉(zhuǎn)YOLOvoc2yolo.py——為什么不用labelImg導(dǎo)出很多工程師習(xí)慣用labelImg打開(kāi)VOC XML再導(dǎo)出YOLO TXT但這種方式存在致命缺陷labelImg導(dǎo)出時(shí)會(huì)重新計(jì)算歸一化坐標(biāo)若圖片尺寸讀取錯(cuò)誤如EXIF Orientation導(dǎo)致PIL讀取寬高顛倒坐標(biāo)將整體偏移。本包附帶的voc2yolo.py直接解析XML原始尺寸規(guī)避此風(fēng)險(xiǎn)# voc2yolo.py import xml.etree.ElementTree as ET import os from pathlib import Path def convert_voc_to_yolo(voc_root: str, yolo_root: str, class_names: list [cat, dog]): voc_ann_dir Path(voc_root) / VOC2012 / Annotations voc_img_dir Path(voc_root) / VOC2012 / JPEGImages yolo_labels_dir Path(yolo_root) / labels # 創(chuàng)建目錄 yolo_labels_dir.mkdir(parentsTrue, exist_okTrue) for xml_file in voc_ann_dir.glob(*.xml): tree ET.parse(xml_file) root tree.getroot() # 獲取圖片尺寸從XML非文件頭 size root.find(size) img_width int(size.find(width).text) img_height int(size.find(height).text) # 構(gòu)建YOLO label文件路徑 yolo_label_path yolo_labels_dir / f{xml_file.stem}.txt with open(yolo_label_path, w) as f: for obj in root.findall(object): cls_name obj.find(name).text.strip() if cls_name not in class_names: continue # 跳過(guò)非法類別 cls_id class_names.index(cls_name) bbox obj.find(bndbox) xmin float(bbox.find(xmin).text) ymin float(bbox.find(ymin).text) xmax float(bbox.find(xmax).text) ymax float(bbox.find(ymax).text) # 歸一化中心點(diǎn)寬高YOLO標(biāo)準(zhǔn) x_center (xmin xmax) / (2 * img_width) y_center (ymin ymax) / (2 * img_height) width (xmax - xmin) / img_width height (ymax - ymin) / img_height f.write(f{cls_id} {x_center:.6f} {y_center:.6f} {width:.6f} {height:.6f}\n) if __name__ __main__: convert_voc_to_yolo( voc_root./catdog_voc_yolo_8291, yolo_root./catdog_voc_yolo_8291/YOLOv8, class_names[cat, dog] )關(guān)鍵參數(shù)說(shuō)明voc_rootVOC格式根目錄含VOCdevkit/VOC2012/yolo_rootYOLO輸出根目錄腳本自動(dòng)創(chuàng)建labels/class_names必須與classes.txt嚴(yán)格一致順序決定cls_idcat0,dog1為何用XML尺寸而非PIL讀取因部分JPG含EXIF旋轉(zhuǎn)標(biāo)記PIL默認(rèn)應(yīng)用旋轉(zhuǎn)導(dǎo)致img.size≠XML中size此處直接信任標(biāo)注員錄入的XML尺寸保證坐標(biāo)源頭準(zhǔn)確。3.2 YOLO轉(zhuǎn)VOCyolo2voc.py——當(dāng)你要用Faster R-CNN或Mask R-CNN時(shí)# yolo2voc.py import os from pathlib import Path from PIL import Image def convert_yolo_to_voc(yolo_root: str, voc_root: str, class_names: list [cat, dog]): yolo_img_dir Path(yolo_root) / images / train # 以train集為例 yolo_label_dir Path(yolo_root) / labels / train voc_ann_dir Path(voc_root) / VOC2012 / Annotations voc_img_dir Path(voc_root) / VOC2012 / JPEGImages voc_ann_dir.mkdir(parentsTrue, exist_okTrue) voc_img_dir.mkdir(parentsTrue, exist_okTrue) for img_path in yolo_img_dir.glob(*.jpg): # 復(fù)制圖片到VOC目錄 dst_img voc_img_dir / img_path.name if not dst_img.exists(): dst_img.write_bytes(img_path.read_bytes()) # 讀取YOLO label label_path yolo_label_dir / f{img_path.stem}.txt if not label_path.exists(): continue # 獲取圖片尺寸必須用PIL因YOLO label無(wú)尺寸信息 with Image.open(img_path) as img: img_width, img_height img.size # 生成XML xml_content fannotation folderVOC2012/folder filename{img_path.name}/filename path{dst_img.absolute()}/path source databaseUnknown/database /source size width{img_width}/width height{img_height}/height depth3/depth /size segmented0/segmented with open(label_path) as f: for line in f: parts line.strip().split() if len(parts) ! 5: continue cls_id, x_center, y_center, width, height map(float, parts) cls_name class_names[int(cls_id)] # YOLO歸一化坐標(biāo) → VOC絕對(duì)坐標(biāo) xmin max(0, int((x_center - width/2) * img_width)) ymin max(0, int((y_center - height/2) * img_height)) xmax min(img_width, int((x_center width/2) * img_width)) ymax min(img_height, int((y_center height/2) * img_height)) xml_content f object name{cls_name}/name poseUnspecified/pose truncated0/truncated difficult0/difficult bndbox xmin{xmin}/xmin ymin{ymin}/ymin xmax{xmax}/xmax ymax{ymax}/ymax /bndbox /object xml_content /annotation with open(voc_ann_dir / f{img_path.stem}.xml, w) as f: f.write(xml_content) if __name__ __main__: convert_yolo_to_voc( yolo_root./catdog_voc_yolo_8291/YOLOv8, voc_root./catdog_voc_yolo_8291, class_names[cat, dog] )關(guān)鍵參數(shù)說(shuō)明yolo_img_dir指定要轉(zhuǎn)換的YOLO子集train/val/testclass_names必須與YOLO的classes.txt一致否則cls_id映射錯(cuò)誤為何用PIL讀取尺寸YOLO label只存歸一化坐標(biāo)無(wú)原始尺寸必須從圖片文件獲取max(0, ...)和min(img_width, ...)防止YOLO坐標(biāo)因浮點(diǎn)誤差超出邊界如x_center0.0001導(dǎo)致xmin-1這是YOLO轉(zhuǎn)VOC最常踩的坑。4. 避坑指南8291張數(shù)據(jù)集的5個(gè)血淚經(jīng)驗(yàn)與排查方案4.1 現(xiàn)象YOLO訓(xùn)練時(shí)AssertionError: Error loading data from...日志顯示image not found原因YOLO格式的images/和labels/目錄下文件名必須嚴(yán)格一一對(duì)應(yīng)如000001.jpg?000001.txt但Windows系統(tǒng)解壓7z時(shí)可能因長(zhǎng)文件名截?cái)嗳鏘MG_20230512_152344_00123456789.jpg→IMG_20230512_152344_00123~1.jpg導(dǎo)致匹配失敗。解決在Linux/macOS下解壓推薦若必須用Windows請(qǐng)用7-Zip 23.01版本并勾選Extract files with long paths選項(xiàng)解壓后運(yùn)行校驗(yàn)?zāi)_本# 檢查images與labels文件名是否完全一致 diff (ls ./YOLOv8/images/train | sort) (ls ./YOLOv8/labels/train | sort | sed s/\.txt$/.jpg/) | grep ^4.2 現(xiàn)象VOC訓(xùn)練時(shí)IndexError: list index out of range報(bào)錯(cuò)指向pascal_voc.py的self._classes原因VOC代碼默認(rèn)self._classes (__background__, cat, dog)但你的classes.txt只有2行cat/dog未包含__background__。YOLO格式不需要背景類但VOC風(fēng)格框架如TensorFlow Object Detection API要求classes[0]為背景。解決修改VOC數(shù)據(jù)加載器在classes列表前插入__background__# 在voc_dataset.py中 self._classes (__background__, cat, dog) # 顯式聲明 # 或動(dòng)態(tài)生成更魯棒 self._classes (__background__,) tuple(open(classes.txt).read().splitlines())4.3 現(xiàn)象訓(xùn)練mAP極低10%但loss下降正常驗(yàn)證集PR曲線呈水平線原因YOLO格式的classes.txt末尾有隱藏?fù)Q行符\n\n導(dǎo)致第二行讀取為dog\n與模型內(nèi)部dog不匹配所有dog樣本被當(dāng)作未知類丟棄。解決用sed -i :a;N;$!ba;s/\n*$//;s/\r*$// classes.txt清理?yè)Q行符或Python中用open().read().strip().splitlines()確保無(wú)空白行。4.4 現(xiàn)象test/目錄下圖片無(wú)法用YOLO推理報(bào)錯(cuò)no labels found for image xxx.jpg原因test/目錄設(shè)計(jì)為無(wú)監(jiān)督推理場(chǎng)景即僅提供圖片不提供真值標(biāo)簽。YOLO推理腳本如detect.py默認(rèn)嘗試加載同名.txt找不到則報(bào)錯(cuò)。解決修改推理腳本添加--save-txt參數(shù)保存預(yù)測(cè)結(jié)果或臨時(shí)創(chuàng)建空l(shuí)abels/test/目錄mkdir -p ./YOLOv8/labels/test touch ./YOLOv8/labels/test/.gitkeep # 防止目錄被git忽略4.5 現(xiàn)象Linux下解壓7z文件報(bào)command not found: 7z即使安裝了p7zip原因Ubuntu/Debian默認(rèn)安裝p7zip-full但命令名為7zCentOS/RHEL默認(rèn)安裝p7zip但命令名為7za。解決統(tǒng)一用7z命令推薦安裝p7zip-full# Ubuntu/Debian sudo apt update sudo apt install p7zip-full # CentOS/RHEL sudo yum install epel-release sudo yum install p7zip-plugins # 驗(yàn)證 7z x catdog_voc_yolo_8291.7z -o./data/5. 進(jìn)階技巧用YOLOv8驗(yàn)證集做mAP快速診斷定位是數(shù)據(jù)問(wèn)題還是模型問(wèn)題5.1 構(gòu)建最小驗(yàn)證閉環(huán)3行命令確認(rèn)數(shù)據(jù)集健康度不要等訓(xùn)練100輪才發(fā)現(xiàn)問(wèn)題。用YOLOv8自帶的val.py對(duì)val/集做單次前向推理直接輸出mAP和各類別AP# 假設(shè)已安裝ultralytics8.2.0 yolo taskdetect modeval modelyolov8n.pt data./catdog_voc_yolo_8291/YOLOv8/data.yaml batch32其中data.yaml內(nèi)容為train: ../images/train val: ../images/val test: ../images/test nc: 2 names: [cat, dog] # 關(guān)鍵顯式指定路徑避免相對(duì)路徑錯(cuò)誤 # 注意YOLOv8要求val路徑相對(duì)于data.yaml位置故用../images/val預(yù)期輸出關(guān)鍵指標(biāo)Box AP0.5≥ 85%數(shù)據(jù)集質(zhì)量?jī)?yōu)秀標(biāo)注精準(zhǔn)cat AP0.5≈dog AP0.5類別平衡性好無(wú)標(biāo)注偏差Precision/Recall 0.8說(shuō)明模型能穩(wěn)定檢出目標(biāo)非數(shù)據(jù)噪聲導(dǎo)致。若Box AP0.5 70%立即暫停訓(xùn)練執(zhí)行第5.2節(jié)診斷。5.2 數(shù)據(jù)級(jí)診斷用visualize_labels.py生成熱力圖揪出標(biāo)注盲區(qū)# visualize_labels.py import numpy as np import matplotlib.pyplot as plt from pathlib import Path from PIL import Image def plot_label_distribution(yolo_labels_dir: str, img_dir: str, output_dir: str ./label_vis): Path(output_dir).mkdir(exist_okTrue) # 統(tǒng)計(jì)所有bbox中心點(diǎn)坐標(biāo)歸一化 centers_cat [] centers_dog [] for label_path in Path(yolo_labels_dir).glob(*.txt): img_path Path(img_dir) / f{label_path.stem}.jpg if not img_path.exists(): continue with Image.open(img_path) as img: w, h img.size with open(label_path) as f: for line in f: parts line.strip().split() if len(parts) ! 5: continue cls_id, x_c, y_c, w_b, h_b map(float, parts) # 轉(zhuǎn)為像素坐標(biāo) x_px x_c * w y_px y_c * h if int(cls_id) 0: centers_cat.append([x_px, y_px]) else: centers_dog.append([x_px, y_px]) # 繪制熱力圖 if centers_cat: cat_arr np.array(centers_cat) plt.figure(figsize(12, 5)) plt.subplot(1, 2, 1) plt.hist2d(cat_arr[:, 0], cat_arr[:, 1], bins50, cmapBlues, range[[0, w], [0, h]]) plt.title(Cat center distribution) plt.xlabel(X (px)) plt.ylabel(Y (px)) plt.subplot(1, 2, 2) if centers_dog: dog_arr np.array(centers_dog) plt.hist2d(dog_arr[:, 0], dog_arr[:, 1], bins50, cmapReds, range[[0, w], [0, h]]) plt.title(Dog center distribution) plt.xlabel(X (px)) plt.ylabel(Y (px)) plt.tight_layout() plt.savefig(f{output_dir}/label_distribution.png) plt.close() if __name__ __main__: plot_label_distribution( yolo_labels_dir./catdog_voc_yolo_8291/YOLOv8/labels/val, img_dir./catdog_voc_yolo_8291/YOLOv8/images/val )解讀熱力圖若cat熱力圖集中在左上角dog集中在右下角 → 標(biāo)注員存在空間偏好需重新隨機(jī)打散若某區(qū)域如圖片底部完全無(wú)熱力 → 該區(qū)域貓狗極少出現(xiàn)或標(biāo)注遺漏如蹲伏的狗易被漏標(biāo)若熱力圖呈明顯網(wǎng)格狀 → 標(biāo)注工具啟用了自動(dòng)對(duì)齊網(wǎng)格需關(guān)閉。5.3 模型級(jí)診斷凍結(jié)backbone只訓(xùn)練head10分鐘驗(yàn)證數(shù)據(jù)是否可學(xué)如果val.py結(jié)果差先排除模型本身問(wèn)題# 凍結(jié)主干網(wǎng)絡(luò)只訓(xùn)練檢測(cè)頭head yolo taskdetect modetrain modelyolov8n.pt data./catdog_voc_yolo_8291/YOLOv8/data.yaml \ epochs10 freeze[0,1,2,3,4,5,6,7,8,9] # yolov8n有10層凍結(jié)前10層backbone若凍結(jié)后val/mAP迅速升至80%說(shuō)明數(shù)據(jù)沒(méi)問(wèn)題原模型過(guò)擬合或?qū)W習(xí)率過(guò)大若仍50%基本確定是數(shù)據(jù)問(wèn)題標(biāo)注錯(cuò)誤/類別混淆/嚴(yán)重遮擋。從那以后我每次拿到新數(shù)據(jù)集都強(qiáng)制走一遍val.pyvisualize_labels.py 凍結(jié)訓(xùn)練三步診斷哪怕只是8291張貓狗圖——因?yàn)?291張里藏了8291個(gè)可能讓模型崩潰的細(xì)節(jié)。希望幫到你。本文還有配套的精品資源點(diǎn)擊獲取