首页 PaddleCV 帖子详情
同样数据集与配置,评价标准不同 ,结果为何不同
收藏
快速回复
PaddleCV 问答目标检测 945 1
同样数据集与配置,评价标准不同 ,结果为何不同
收藏
快速回复
PaddleCV 问答目标检测 945 1

同样的数据集(分别组织成coco格式与VOC格式),同样的配置,用coco(看AP0.5吧)评价标准与用VOC评价标准 ,结果也不一样?   0.871(87.1%) VS 85.19%

yolov3_darknet_TGRS

Average Precision (AP) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.524
Average Precision (AP) @[ IoU=0.50 | area= all | maxDets=100 ] = 0.871
Average Precision (AP) @[ IoU=0.75 | area= all | maxDets=100 ] = 0.557

yolov3_darknet_voc_TGRS

mAP(0.50, 11point) = 85.19%

 

architecture: YOLOv3
use_gpu: true
#max_iters: 500000
max_iters: 70000
log_smooth_window: 20
save_dir: output
#snapshot_iter: 10000
snapshot_iter: 2000
metric: COCO
pretrain_weights: https://paddle-imagenet-models-name.bj.bcebos.com/DarkNet53_pretrained.tar
weights: output/yolov3_darknet_TGRS/model_final
num_classes: 13
use_fine_grained_loss: false

YOLOv3:
  backbone: DarkNet
  yolo_head: YOLOv3Head

DarkNet:
  norm_type: sync_bn
  norm_decay: 0.
  depth: 53

YOLOv3Head:
  anchor_masks: [[6, 7, 8], [3, 4, 5], [0, 1, 2]]
  anchors: [[10, 13], [16, 30], [33, 23],
            [30, 61], [62, 45], [59, 119],
            [116, 90], [156, 198], [373, 326]]
  norm_decay: 0.
  yolo_loss: YOLOv3Loss
  nms:
    background_label: -1
    keep_top_k: 100
    nms_threshold: 0.45
    nms_top_k: 1000
    normalized: false
    score_threshold: 0.01

YOLOv3Loss:
  ignore_thresh: 0.7
  #label_smooth: true  #yolov3_darknet_voc_TGRS.yml中 是false
  label_smooth: false

LearningRate:
  base_lr: 0.001
  schedulers:
  - !PiecewiseDecay
    gamma: 0.1
  #  milestones:
  #  - 400000
  #  - 450000
    milestones:
    - 55000
    - 62000
  - !LinearWarmup
    start_factor: 0.
    #steps: 4000
    steps: 1000

OptimizerBuilder:
  optimizer:
    momentum: 0.9
    type: Momentum
  regularizer:
    factor: 0.0005
    type: L2

_READER_: 'yolov3_reader.yml'

 去掉了mixup

architecture: YOLOv3
use_gpu: true
max_iters: 70000
log_smooth_window: 20
save_dir: output
snapshot_iter: 2000
metric: VOC
map_type: 11point
pretrain_weights: https://paddle-imagenet-models-name.bj.bcebos.com/DarkNet53_pretrained.tar
weights: output/yolov3_darknet_voc_TGRS/model_final
num_classes: 13
use_fine_grained_loss: false

YOLOv3:
  backbone: DarkNet
  yolo_head: YOLOv3Head

DarkNet:
  norm_type: sync_bn
  norm_decay: 0.
  depth: 53

YOLOv3Head:
  anchor_masks: [[6, 7, 8], [3, 4, 5], [0, 1, 2]]
  anchors: [[10, 13], [16, 30], [33, 23],
            [30, 61], [62, 45], [59, 119],
            [116, 90], [156, 198], [373, 326]]
  norm_decay: 0.
  yolo_loss: YOLOv3Loss
  nms:
    background_label: -1
    keep_top_k: 100
    nms_threshold: 0.45
    nms_top_k: 1000
    normalized: false
    score_threshold: 0.01

YOLOv3Loss:
  ignore_thresh: 0.7
  label_smooth: false

LearningRate:
  base_lr: 0.001
  schedulers:
  - !PiecewiseDecay
    gamma: 0.1
    milestones:
    - 55000
    - 62000
  - !LinearWarmup
    start_factor: 0.
    steps: 1000

OptimizerBuilder:
  optimizer:
    momentum: 0.9
    type: Momentum
  regularizer:
    factor: 0.0005
    type: L2

_READER_: 'yolov3_reader.yml'

同样去掉了mixup

 

 

 

 

 

TrainReader:
  inputs_def:
    fields: ['image', 'gt_bbox', 'gt_class', 'gt_score']
    num_max_boxes: 50
  dataset:
    !VOCDataSet
    dataset_dir: data/TGRS_yuan
    anno_path: trainval.txt
    use_default_label: false
    with_background: false

EvalReader:
  inputs_def:
    fields: ['image', 'im_size', 'im_id', 'gt_bbox', 'gt_class', 'is_difficult']
    num_max_boxes: 50
  dataset:
    !VOCDataSet
    dataset_dir: data/TGRS_yuan
    anno_path: test.txt
    use_default_label: false
    with_background: false

TestReader:
  dataset:
    !ImageFolder
    anno_path: label_list.txt
    use_default_label: false
    with_background: false

0
收藏
回复
全部评论(1)
时间顺序
AIStudio810260
#2 回复于2020-11

coco和voc评价标准本来就不一样的……

0
回复
需求/bug反馈?一键提issue告诉我们
发现bug?如果您知道修复办法,欢迎提pr直接参与建设飞桨~
在@后输入用户全名并按空格结束,可艾特全站任一用户