YOLO 在自定义数据集上的表现非常糟糕

如何解决YOLO 在自定义数据集上的表现非常糟糕

我正在使用自定义数据集来检测游戏内横幅内的徽标，但该模型的表现非常糟糕。最后，它无法检测到任何东西（没有 boundinb 框）。我有 6 个类的 143 个图像。这是训练输出的示例：

(next mAP calculation at 3000 iterations) 
 Last accuracy mAP@0.5 = 0.64 %,best = 25.44 % 
 2991: 1.685981,1.863738 avg loss,0.010000 rate,2.355753 seconds,47856 images,5.858397 hours left
Loaded: 0.000069 seconds
v3 (giou loss,Normalizer: (iou: 0.50,obj: 1.00,cls: 1.00) Region 30 Avg (IOU: 0.000000),count: 1,class_loss = 0.184588,iou_loss = 0.000000,total_loss = 0.184588 
v3 (giou loss,cls: 1.00) Region 37 Avg (IOU: 0.001107),count: 2,class_loss = 1.428874,iou_loss = 0.006507,total_loss = 1.435380 
 total_bbox = 108594,rewritten_bbox = 0.000000 % 
v3 (giou loss,class_loss = 0.248431,total_loss = 0.248431 
v3 (giou loss,cls: 1.00) Region 37 Avg (IOU: 0.062057),count: 5,class_loss = 3.823031,iou_loss = 782.007629,total_loss = 785.830688 
 total_bbox = 108599,cls: 1.00) Region 30 Avg (IOU: 0.707677),class_loss = 1.166240,iou_loss = 1227.803467,total_loss = 1228.969727 
v3 (giou loss,cls: 1.00) Region 37 Avg (IOU: 0.706214),class_loss = 1.750910,iou_loss = 1684.371338,total_loss = 1686.122192 
 total_bbox = 108603,cls: 1.00) Region 30 Avg (IOU: 0.668676),class_loss = 4.053865,iou_loss = 3389.313232,total_loss = 3393.367188 
v3 (giou loss,cls: 1.00) Region 37 Avg (IOU: 0.386856),count: 8,class_loss = 6.107755,iou_loss = 2782.196289,total_loss = 2788.303955 
 total_bbox = 108616,cls: 1.00) Region 30 Avg (IOU: 0.855742),class_loss = 0.354011,iou_loss = 1362.482300,total_loss = 1362.836304 
v3 (giou loss,cls: 1.00) Region 37 Avg (IOU: 0.344064),class_loss = 1.351954,iou_loss = 892.178528,total_loss = 893.530457 
 total_bbox = 108619,class_loss = 0.078801,total_loss = 0.078801 
v3 (giou loss,cls: 1.00) Region 37 Avg (IOU: 0.451274),class_loss = 0.938443,iou_loss = 2481.435547,total_loss = 2482.374023 
 total_bbox = 108620,cls: 1.00) Region 30 Avg (IOU: 0.605520),count: 3,class_loss = 1.935990,iou_loss = 1187.735596,total_loss = 1189.671631 
v3 (giou loss,cls: 1.00) Region 37 Avg (IOU: 0.402832),count: 7,class_loss = 5.597641,iou_loss = 6449.970703,total_loss = 6455.568359 
 total_bbox = 108630,rewritten_bbox = 0.000000 % 

这是关于我的批量大小和学习率的信息。我选择了较小的批量大小，因为我的分辨率很高。我的分辨率很高，因为我检测到的徽标非常小。

batch=16
subdivisions=8
width=2048
height=2048
channels=3
momentum=0.9
decay=0.0005
angle=0
saturation = 1.5
exposure = 1.5
hue=.1
 
learning_rate=0.01

我不明白为什么模型的表现真的很差？是不是因为我的对象太小了？我的问题究竟在哪里？我无法理解训练输出。