使用TensorFlow Object Detection API检测大量目标时存在遗漏问题求助

我使用TensorFlow Object Detection API搭配EfficientDet D0模型检测单张图片中的大量目标（超过100个）时，大部分目标能被正常检测，但训练数据集中部分目标被完全忽略。经观察发现，被忽略的均为使用LabelImg最后标注的目标：

• 第一张图：先标注左框，再标注中框，最后标注右框，右框被模型忽略
• 第二张图：先标注左框，再标注右框，最后标注中框，中框被模型忽略

（注：截图中模型整体表现较差是因为我专注于解决当前问题，尚未完成完整训练）

我曾怀疑是pipeline.config中的参数问题，尤其是max_detections_per_class、max_total_detections和max_number_of_boxes，于是将这些参数调至很高值（max_detections_per_class:3000，max_total_detections:60000，max_number_of_boxes:60000），但结果并未改变，因此我怀疑在配置文件或Object Detection API中存在隐藏的目标数量限制。

以下是我的pipeline.config配置：

model {
  ssd {
    num_classes: 90
    image_resizer {
      keep_aspect_ratio_resizer {
        min_dimension: 512
        max_dimension: 512
        pad_to_max_dimension: true
      }
    }
    feature_extractor {
      type: "ssd_efficientnet-b0_bifpn_keras"
      conv_hyperparams {
        regularizer {
          l2_regularizer {
            weight: 3.9999998989515007e-05
          }
        }
        initializer {
          truncated_normal_initializer {
            mean: 0.0
            stddev: 0.029999999329447746
          }
        }
        activation: SWISH
        batch_norm {
          decay: 0.9900000095367432
          scale: true
          epsilon: 0.0010000000474974513
        }
        force_use_bias: true
      }
      bifpn {
        min_level: 3
        max_level: 7
        num_iterations: 3
        num_filters: 64
      }
    }
    box_coder {
      faster_rcnn_box_coder {
        y_scale: 1.0
        x_scale: 1.0
        height_scale: 1.0
        width_scale: 1.0
      }
    }
    matcher {
      argmax_matcher {
        matched_threshold: 0.5
        unmatched_threshold: 0.5
        ignore_thresholds: false
        negatives_lower_than_unmatched: true
        force_match_for_each_row: true
        use_matmul_gather: true
      }
    }
    similarity_calculator {
      iou_similarity {
      }
    }
    box_predictor {
      weight_shared_convolutional_box_predictor {
        use_dropout: True
        dropout_keep_probability: 0.7
        conv_hyperparams {
          regularizer {
            l2_regularizer {
              weight: 3.9999998989515007e-05
            }
          }
          initializer {
            random_normal_initializer {
              mean: 0.0
              stddev: 0.009999999776482582
            }
          }
          activation: SWISH
          batch_norm {
            decay: 0.9900000095367432
            scale: true
            epsilon: 0.0010000000474974513
          }
          force_use_bias: true
        }
        depth: 64
        num_layers_before_predictor: 3
        kernel_size: 3
        class_prediction_bias_init: -4.599999904632568
        use_depthwise: true
      }
    }
    anchor_generator {
      multiscale_anchor_generator {
        min_level: 3
        max_level: 7
        anchor_scale: 4.0
        aspect_ratios: 1.0
        aspect_ratios: 2.0
        aspect_ratios: 5.0
        aspect_ratios: 0.4
        scales_per_octave: 3
      }
    }
    post_processing {
      batch_non_max_suppression {
        score_threshold: 0.2
        iou_threshold: 0.5
        max_detections_per_class: 3000
        max_total_detections: 60000
        }
      score_converter: SIGMOID
    }
    normalize_loss_by_num_matches: true
    loss {
      localization_loss {
        weighted_smooth_l1 {
        }
      }
      classification_loss {
        weighted_sigmoid_focal {
          gamma: 1.5
          alpha: 0.25
        }
      }
      classification_weight: 1.0
      localization_weight: 1.0
    }
    encode_background_as_zeros: true
    normalize_loc_loss_by_codesize: true
    inplace_batchnorm_update: true
    freeze_batchnorm: false
    add_background_class: false
  }
}
train_config {
  batch_size: 4
  data_augmentation_options {
    random_horizontal_flip {
    }
    random_patch_gaussian{
    }
    random_adjust_brightness{    
    }
    random_adjust_contrast{
    }
    random_rgb_to_gray{
      probability: 0.2
    }
    random_scale_crop_and_pad_to_square {
      output_size: 512
      scale_min: 0.8
      scale_max: 1.2
    }
  }


  sync_replicas: true
  optimizer {
    momentum_optimizer {
      learning_rate {
        cosine_decay_learning_rate {
          learning_rate_base: 0.07999999821186066
          total_steps: 300000
          warmup_learning_rate: 0.0010000000474974513
          warmup_steps: 2500
        }
      }
      momentum_optimizer_value: 0.8999999761581421
    }
    use_moving_average: false
  }
  fine_tune_checkpoint: "C:\Tensorflow\workspace\pre_trained_models\efficientdet_d0_coco17_tpu-32\checkpoint\ckpt-0"
  num_steps: 300000
  startup_delay_steps: 0.0
  replicas_to_aggregate: 8
  max_number_of_boxes: 60000
  unpad_groundtruth_tensors: false
  fine_tune_checkpoint_type: "detection"
  use_bfloat16: false
  fine_tune_checkpoint_version: V2
}
train_input_reader: {
  label_map_path: "C:\Tensorflow\Dataset\label_map.pbtxt"
  tf_record_input_reader {
    input_path: "C:\Tensorflow\Dataset\train4.record"
  }
}

eval_config: {
  metrics_set: "coco_detection_metrics"
  use_moving_averages: false
  batch_size: 1;
}

eval_input_reader: {
  label_map_path: "C:\Tensorflow\Dataset\label_map.pbtxt"
  shuffle: false
  num_epochs: 1
  tf_record_input_reader {
    input_path: "C:\Tensorflow\workspace\data\val.record"
  }
}

内容的提问来源于stack exchange，提问作者haghehog