DB++转为inference模型后效果很差

[LUXUS1]

DB++在使用infer_det.py进行推理时效果较好，但转为inference模型后效果很差。也尝试进行了参数对齐，但效果并没有改变。 检测结果： 运行指令： python3 tools/infer_det.py -c configs/det/det_r50_db++_td_tr.yml -o Global.infer_img="./images/3.jpg" Global.pretrained_model="./seal_models/DB++/best_accuracy" python3 tools/export_model.py -c configs/det/det_r50_db++_td_tr.yml -o Global.pretrained_model=./seal_models/DB++/best_accuracy Global.save_inference_dir=./inference/det_db++/ python3 tools/infer/predict_det.py --image_dir="./images/3.jpg" --det_model_dir="./inference/det_db++/" --det_algorithm="DB++" --use_gpu=false 参数配置： det_r50_db++_td_tr.yml Global: debug: false use_gpu: False epoch_num: 1000 log_smooth_window: 20 print_batch_step: 10 save_model_dir: ./output/det_r50_td_tr/ save_epoch_step: 200 eval_batch_step:

• 0
• 2000 cal_metric_during_train: false pretrained_model: ./pretrain_models/ResNet50_dcn_asf_synthtext_pretrained checkpoints: null save_inference_dir: null use_visualdl: false infer_img: doc/imgs_en/img_10.jpg save_res_path: ./checkpoints/det_db/predicts_db.txt Architecture: model_type: det algorithm: DB++ Transform: null Backbone: name: ResNet layers: 50 dcn_stage: [False, True, True, True] Neck: name: DBFPN out_channels: 256 use_asf: True Head: name: DBHead k: 50 Loss: name: DBLoss balance_loss: true main_loss_type: BCELoss alpha: 5 beta: 10 ohem_ratio: 3 Optimizer: name: Momentum momentum: 0.9 lr: name: DecayLearningRate learning_rate: 0.007 epochs: 1000 factor: 0.9 end_lr: 0 weight_decay: 0.0001 PostProcess: name: DBPostProcess # 后处理类名 thresh: 0.3 # DBPostProcess中分割图进行二值化的阈值 box_thresh: 0.2 # DBPostProcess中对输出框进行过滤的阈值，低于此阈值的框不会输出 max_candidates: 1000 unclip_ratio: 2.0 det_box_type: 'poly' # 'quad' or 'poly' Metric: name: DetMetric main_indicator: hmean Train: dataset: name: SimpleDataSet data_dir: ./train_data/
  label_file_list:
  • ./train_data/TD_TR/TD500/train_gt_labels.txt
  • ./train_data/TD_TR/TR400/gt_labels.txt
    ratio_list:
  • 1.0
  • 1.0 transforms:
  • DecodeImage: img_mode: BGR channel_first: false
  • DetLabelEncode: null
  • IaaAugment: augmenter_args:
    • type: Fliplr args: p: 0.5
    • type: Affine args: rotate:
      • -10
      • 10
    • type: Resize args: size:
      • 0.5
      • 3
  • EastRandomCropData: size:
    • 640
    • 640 max_tries: 10 keep_ratio: true
  • MakeShrinkMap: shrink_ratio: 0.4 min_text_size: 8
  • MakeBorderMap: shrink_ratio: 0.4 thresh_min: 0.3 thresh_max: 0.7
  • NormalizeImage: scale: 1./255. mean:
    • 0.48109378172549
    • 0.45752457890196
    • 0.40787054090196 std:
    • 1.0
    • 1.0
    • 1.0 order: hwc
  • ToCHWImage: null
  • KeepKeys: keep_keys:
    • image
    • threshold_map
    • threshold_mask
    • shrink_map
    • shrink_mask loader: shuffle: true drop_last: false batch_size_per_card: 4 num_workers: 8 Eval: dataset: name: SimpleDataSet data_dir: ./train_data/
      label_file_list:
  • ./train_data/TD_TR/TD500/test_gt_labels.txt transforms:
  • DecodeImage: img_mode: BGR channel_first: false
  • DetLabelEncode: null
  • DetResizeForTest: image_shape:
    • 736
    • 736 keep_ratio: True
  • NormalizeImage: scale: 1./255. mean:
    • 0.48109378172549
    • 0.45752457890196
    • 0.40787054090196 std:
    • 1.0
    • 1.0
    • 1.0 order: hwc
  • ToCHWImage: null
  • KeepKeys: keep_keys:
    • image
    • shape
    • polys
    • ignore_tags loader: shuffle: false drop_last: false batch_size_per_card: 1 num_workers: 2 profiler_options: null

--perdict.py

[andyjiang1116]

训练和预测使用的shape要保持一致

[LDOUBLEV]

预测时在这里加个参数 "image_shape": [736, 736] 和训练配置里的shape保持一致 https://github.com/PaddlePaddle/PaddleOCR/blob/0ed9d8889f7ba90949e64fa6e3240df282983a77/tools/infer/predict_det.py#L44

代码里默认都是动态shape预测，所以没有设置固定shape预测的参数，DetResizeForTest这个函数支持传入自定义shape

[LUXUS1]

感谢，不过又遇到了新的问题，

1. 预测同一张图像"--det_box_type", type=str, default='poly'会显示如下错误：

改为"--det_box_type", type=str, default='quad'错误消失但只能识别四边形文本，如下所示：

2. "--det_box_type", type=str, default='poly'时个别图像可以正常检测出来，如下： 但大多数图像无法检测

[LDOUBLEV]

训练模型预测会报错吗？

在报错前打印dt_boxes_new的shape看下，是不是第一个维度为0

[LUXUS1]

预训练模型和中文的这个都会报错

1. 预训练模型

2. 中文模型

3. 中文个别可以检测的图像输出如下：

[LDOUBLEV]

dt_boxes = np.array(dt_boxes_new) 这一行删掉，直接return dt_boxes_new

原因是预测的box的点数不一致，不能直接组成一个数组；

[LUXUS1]

感谢！

[LUXUS1]

在上个问题中，删掉dt_boxes = np.array(dt_boxes_new)，直接返回dt_boxes_new确实可以解决问题。但使用predict_system.py将DB++与SVTR进行串联使用时，改动的位置便会报错，请问如何解决？感谢

[LUXUS1]

在上个问题中，删掉dt_boxes = np.array(dt_boxes_new)，直接返回dt_boxes_new确实可以解决问题。但使用predict_system.py将DB++与SVTR进行串联使用时，改动的位置便会报错，请问如何解决？感谢

num_boxes = dt_boxes.shape[0]修改为num_boxes = len(dt_boxes)

[papersuper]

预测时在这里加个参数 "image_shape": [736, 736] 和训练配置里的shape保持一致

https://github.com/PaddlePaddle/PaddleOCR/blob/0ed9d8889f7ba90949e64fa6e3240df282983a77/tools/infer/predict_det.py#L44

代码里默认都是动态shape预测，所以没有设置固定shape预测的参数，DetResizeForTest这个函数支持传入自定义shape 你好，请问一下，是在PaddleOCR/tools/infer/predict_det.py的DetResizeForTest里面加入“image_shape”:[512,512]是吗？我的会报错呢

[Homura852]

@LUXUS1 兄弟您好，请问可以添加个联系方式交流一下DBnet++的训练推理吗

[kkiskkk]

DB++在使用infer_det.py进行推理时效果较好，但转为inference模型后效果很差。也尝试进行了参数对齐，但效果并没有改变。 检测结果： 运行指令： python3 tools/infer_det.py -c configs/det/det_r50_db++_td_tr.yml -o Global.infer_img="./images/3.jpg" Global.pretrained_model="./seal_models/DB++/best_accuracy" python3 tools/export_model.py -c configs/det/det_r50_db++_td_tr.yml -o Global.pretrained_model=./seal_models/DB++/best_accuracy Global.save_inference_dir=./inference/det_db++/ python3 tools/infer/predict_det.py --image_dir="./images/3.jpg" --det_model_dir="./inference/det_db++/" --det_algorithm="DB++" --use_gpu=false 参数配置： det_r50_db++_td_tr.yml Global: debug: false use_gpu: False epoch_num: 1000 log_smooth_window: 20 print_batch_step: 10 save_model_dir: ./output/det_r50_td_tr/ save_epoch_step: 200 eval_batch_step:

• 0

• 2000 cal_metric_during_train: false pretrained_model: ./pretrain_models/ResNet50_dcn_asf_synthtext_pretrained checkpoints: null save_inference_dir: null use_visualdl: false infer_img: doc/imgs_en/img_10.jpg save_res_path: ./checkpoints/det_db/predicts_db.txt Architecture: model_type: det algorithm: DB++ Transform: null Backbone: name: ResNet layers: 50 dcn_stage: [False, True, True, True] Neck: name: DBFPN out_channels: 256 use_asf: True Head: name: DBHead k: 50 Loss: name: DBLoss balance_loss: true main_loss_type: BCELoss alpha: 5 beta: 10 ohem_ratio: 3 Optimizer: name: Momentum momentum: 0.9 lr: name: DecayLearningRate learning_rate: 0.007 epochs: 1000 factor: 0.9 end_lr: 0 weight_decay: 0.0001 PostProcess: name: DBPostProcess # 后处理类名 thresh: 0.3 # DBPostProcess中分割图进行二值化的阈值 box_thresh: 0.2 # DBPostProcess中对输出框进行过滤的阈值，低于此阈值的框不会输出 max_candidates: 1000 unclip_ratio: 2.0 det_box_type: 'poly' # 'quad' or 'poly' Metric: name: DetMetric main_indicator: hmean Train: dataset: name: SimpleDataSet data_dir: ./train_data/ label_file_list:

  • ./train_data/TD_TR/TD500/train_gt_labels.txt

  • ./train_data/TD_TR/TR400/gt_labels.txt ratio_list:

  • 1.0

  • 1.0 transforms:

  • DecodeImage: img_mode: BGR channel_first: false

  • DetLabelEncode: null

  • IaaAugment: augmenter_args:

  • type: Fliplr args: p: 0.5

  • type: Affine args: rotate:

    • -10
    • 10

  • type: Resize args: size:

    • 0.5
    • 3

  • EastRandomCropData: size:

  • 640

  • 640 max_tries: 10 keep_ratio: true

  • MakeShrinkMap: shrink_ratio: 0.4 min_text_size: 8

  • MakeBorderMap: shrink_ratio: 0.4 thresh_min: 0.3 thresh_max: 0.7

  • NormalizeImage: scale: 1./255. mean:

  • 0.48109378172549

  • 0.45752457890196

  • 0.40787054090196 std:

  • 1.0

  • 1.0

  • 1.0 order: hwc

  • ToCHWImage: null

  • KeepKeys: keep_keys:

  • image

  • threshold_map

  • threshold_mask

  • shrink_map

  • shrink_mask loader: shuffle: true drop_last: false batch_size_per_card: 4 num_workers: 8 Eval: dataset: name: SimpleDataSet data_dir: ./train_data/ label_file_list:

  • ./train_data/TD_TR/TD500/test_gt_labels.txt transforms:

  • DecodeImage: img_mode: BGR channel_first: false

  • DetLabelEncode: null

  • DetResizeForTest: image_shape:

  • 736

  • 736 keep_ratio: True

  • NormalizeImage: scale: 1./255. mean:

  • 0.48109378172549

  • 0.45752457890196

  • 0.40787054090196 std:

  • 1.0

  • 1.0

  • 1.0 order: hwc

  • ToCHWImage: null

  • KeepKeys: keep_keys:

  • image

  • shape

  • polys

  • ignore_tags loader: shuffle: false drop_last: false batch_size_per_card: 1 num_workers: 2 profiler_options: null

--perdict.py

可以分享下这个权重吗？ 我在官方下载的就算把det_box_type: 'poly' 设置为poly 检测结果还是只有矩形框。