我试图准确地了解tensorflow对象检测配置字段。
根据这篇文章(https://medium.com/@jonathan_hui/object-detection-speed-and-accuracy-comparison-faster-r-cnn-r-fcn-ssd-and-yolo-5425656ae359),为了在准确性和速度之间取得良好的平衡,我将first_stage_max_proposals从原点100更改为50。
好消息是,它确实将推理延迟(从每张图像的4.2秒减少到2.2秒),但是,坏消息是,它也降低了准确性。
然后,我将提案总数从50个更改为70个,准确性更高。
因此,我想确切地了解最高提案的控制范围。它是否与其他任何配置(例如max_detections_per_class或max_total_detections .etc)相关?
我在Google上搜索了很多,但对这个家伙的兴趣似乎有所减少。 我使用python3.6.4和tensorflow 1.8.0,这是我的模型配置:
model {
faster_rcnn {
num_classes: 3
image_resizer {
keep_aspect_ratio_resizer {
min_dimension: 670
max_dimension: 1013
}
}
feature_extractor {
type: "faster_rcnn_resnet101"
first_stage_features_stride: 16
}
first_stage_anchor_generator {
grid_anchor_generator {
height_stride: 16
width_stride: 16
scales: 0.25
scales: 0.5
scales: 1.0
scales: 2.0
aspect_ratios: 0.5
aspect_ratios: 1.0
aspect_ratios: 2.0
}
}
first_stage_box_predictor_conv_hyperparams {
op: CONV
regularizer {
l2_regularizer {
weight: 0.0
}
}
initializer {
truncated_normal_initializer {
stddev: 0.01
}
}
}
first_stage_nms_score_threshold: 0.0
first_stage_nms_iou_threshold: 0.7
first_stage_max_proposals: 70
first_stage_localization_loss_weight: 2.0
first_stage_objectness_loss_weight: 1.0
initial_crop_size: 14
maxpool_kernel_size: 2
maxpool_stride: 2
second_stage_box_predictor {
mask_rcnn_box_predictor {
fc_hyperparams {
op: FC
regularizer {
l2_regularizer {
weight: 0.0
}
}
initializer {
variance_scaling_initializer {
factor: 1.0
uniform: true
mode: FAN_AVG
}
}
}
use_dropout: false
dropout_keep_probability: 1.0
}
}
second_stage_post_processing {
batch_non_max_suppression {
score_threshold: 0.3
iou_threshold: 0.6
max_detections_per_class: 30
max_total_detections: 30
}
score_converter: SOFTMAX
}
second_stage_localization_loss_weight: 2.0
second_stage_classification_loss_weight: 1.0
second_stage_batch_size: 70
}
}
train_config {
batch_size: 1
data_augmentation_options {
random_horizontal_flip {
}
}
optimizer {
momentum_optimizer {
learning_rate {
exponential_decay_learning_rate {
initial_learning_rate: 0.0003
decay_steps: 2000
decay_factor: 0.95
}
}
momentum_optimizer_value: 0.9
}
use_moving_average: false
}
gradient_clipping_by_norm: 10.0
fine_tune_checkpoint: "d:/od/tool/faster_rcnn3/model.ckpt"
from_detection_checkpoint: true
}
train_input_reader {
label_map_path: "d:/od/project/train_allinone/file/labelmap.pbtxt"
tf_record_input_reader {
input_path: "d:/od/project/train_allinone/file/tf.record"
}
}
对此有任何解释,我们将不胜感激。
谢谢。
答案 0 :(得分:0)
// Naming conventions:
// Faster R-CNN models have two stages: a first stage region proposal network
// (or RPN) and a second stage box classifier. We thus use the prefixes
// `first_stage_` and `second_stage_` to indicate the stage to which each
// parameter pertains when relevant
等等:
// Maximum number of RPN proposals retained after first stage postprocessing.
optional int32 first_stage_max_proposals = 15 [default=300];
更快的R-CNN具有两个网络,第一个提议可以发现对象的区域,第二个试图检测那些对象中的对象。由于第二网络必须在更多的潜在区域中进行搜索,因此第一个网络增加的投标数量会提高准确性,但意味着要进行更多的计算工作。有关Faster R-CNN的工作原理的简要说明,请查看Faster R-CNN Explained,如果要查看完整图片,可以查看原始出版物:Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks。