我使用tensorflow对象检测api(https://github.com/tensorflow/models/tree/master/research/object_detection)使用voc 2007 + 2012 trainval数据集来训练rfcn模型,并在voc 2007测试中进行了测试。与caffe版本相比,MAP @ 0.5低得多。 caffe版本训练了110000次迭代,而Tensorflow版本训练了140000次迭代。经过预训练的resnet-v1-50模块,用于初始化主干特征提取器。配置文件如下:
#pascal_voc_resnet50_rfcn.config:
model {
faster_rcnn {
num_classes: 20
image_resizer {
keep_aspect_ratio_resizer {
min_dimension: 600
max_dimension: 1024
}
}
feature_extractor {
type: 'faster_rcnn_resnet50'
first_stage_features_stride: 16
}
first_stage_anchor_generator {
grid_anchor_generator {
scales: [0.5, 1.0, 2.0]
aspect_ratios: [0.5, 1.0, 2.0]
height_stride: 16
width_stride: 16
}
}
first_stage_box_predictor_conv_hyperparams {
op: CONV
regularizer {
l2_regularizer {
weight: 0.0005
}
}
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: 300
first_stage_localization_loss_weight: 1.0
first_stage_objectness_loss_weight: 1.0
second_stage_box_predictor {
rfcn_box_predictor {
conv_hyperparams {
op: CONV
regularizer {
l2_regularizer {
weight: 0.0005
}
}
initializer {
truncated_normal_initializer {
stddev: 0.01
}
}
}
crop_height: 18
crop_width: 18
num_spatial_bins_height: 3
num_spatial_bins_width: 3
}
}
second_stage_post_processing {
batch_non_max_suppression {
score_threshold: 0.0
iou_threshold: 0.7
max_detections_per_class: 100
max_total_detections: 300
}
score_converter: SOFTMAX
}
second_stage_localization_loss_weight: 1.0
second_stage_classification_loss_weight: 1.0
}
}
train_config: {
batch_size: 1
optimizer {
momentum_optimizer: {
learning_rate: {
manual_step_learning_rate {
initial_learning_rate: 0.001
schedule {
step: 0
learning_rate: .001
}
schedule {
step: 900000
learning_rate: .0001
}
schedule {
step: 1200000
learning_rate: .00001
}
}
}
momentum_optimizer_value: 0.9
}
use_moving_average: false
}
gradient_clipping_by_norm: 10.0
fine_tune_checkpoint: "resnet_v1_50/resnet_v1_50.ckpt"
from_detection_checkpoint: false
# Note: The below line limits the training process to 200K steps, which we
# empirically found to be sufficient enough to train the pets dataset. This
# effectively bypasses the learning rate schedule (the learning rate will
# never decay). Remove the below line to train indefinitely.
num_steps: 1500000
data_augmentation_options {
random_horizontal_flip {
}
}
}
train_input_reader: {
tf_record_input_reader {
input_path: "voc_dataset/trainval.tfrecords"
}
label_map_path: "object_detection/data/pascal_label_map.pbtxt"
}
eval_config: {
# num_examples: 8000
num_examples: 4952
num_visualizations: 4952
# Note: The below line limits the evaluation process to 10 evaluations.
# Remove the below line to evaluate indefinitely.
max_evals: 1
visualization_export_dir: 'outputs_eval_imgs'
metrics_set: 'pascal_voc_metrics'
}
eval_input_reader: {
tf_record_input_reader {
input_path: "voc_dataset/test.tfrecords"
}
label_map_path: "object_detection/data/pascal_label_map.pbtxt"
shuffle: false
num_readers: 1
num_epochs: 1
}
最终结果是: PascalBoxes_PerformanceByCategory/AP@0.5IOU/aeroplane:0.701776 PascalBoxes_PerformanceByCategory/AP@0.5IOU/bicycle:0.742742 PascalBoxes_PerformanceByCategory/AP@0.5IOU/bird:0.723409 PascalBoxes_PerformanceByCategory/AP@0.5IOU/boat:0.513328 PascalBoxes_PerformanceByCategory/AP@0.5IOU/bottle:0.531051 PascalBoxes_PerformanceByCategory/AP@0.5IOU/bus:0.769170 PascalBoxes_PerformanceByCategory/AP@0.5IOU/car:0.811411 PascalBoxes_PerformanceByCategory/AP@0.5IOU/cat:0.831349 PascalBoxes_PerformanceByCategory/AP@0.5IOU/chair:0.472102 PascalBoxes_PerformanceByCategory/AP@0.5IOU/cow:0.790175 PascalBoxes_PerformanceByCategory/AP@0.5IOU/diningtable:0.483809 PascalBoxes_PerformanceByCategory/AP@0.5IOU/dog:0.819959 PascalBoxes_PerformanceByCategory/AP@0.5IOU/马:0.838640 PascalBoxes_PerformanceByCategory/AP@0.5IOU/motor:0.733901 PascalBoxes_PerformanceByCategory/AP@0.5IOU/人:0.765344 PascalBoxes_PerformanceByCategory/AP@0.5IOU/盆栽植物:0.379224 PascalBoxes_PerformanceByCategory/AP@0.5IOU/sheep:0.719418 PascalBoxes_PerformanceByCategory/AP@0.5IOU/sofa:0.576437 PascalBoxes_PerformanceByCategory/AP@0.5IOU/train:0.726485 PascalBoxes_PerformanceByCategory/AP@0.5IOU/tvmonitor:0.683094 PascalBoxes_Precision/mAP@0.5IOU:0.680641
但是,当我使用原始版本(基于caffe)时,mAP为0.746,详细信息如下: PascalBoxes_PerformanceByCategory/AP@0.5IOU/aeroplane:0.781 PascalBoxes_PerformanceByCategory/AP@0.5IOU/bicycle:0.793 PascalBoxes_PerformanceByCategory/AP@0.5IOU/bird:0.756 PascalBoxes_PerformanceByCategory/AP@0.5IOU/boat:0.652 PascalBoxes_PerformanceByCategory/AP@0.5IOU/bottle:0.578 PascalBoxes_PerformanceByCategory/AP@0.5IOU/bus:0.843 PascalBoxes_PerformanceByCategory/AP@0.5IOU/car:0.846 PascalBoxes_PerformanceByCategory/AP@0.5IOU/cat:0.889 PascalBoxes_PerformanceByCategory/AP@0.5IOU/chair:0.565 PascalBoxes_PerformanceByCategory/AP@0.5IOU/cow:0.835 PascalBoxes_PerformanceByCategory/AP@0.5IOU/diningtable:0.658 PascalBoxes_PerformanceByCategory/AP@0.5IOU/dog:0.867 PascalBoxes_PerformanceByCategory/AP@0.5IOU/马:0.857 PascalBoxes_PerformanceByCategory/AP@0.5IOU/motor:0.792 PascalBoxes_PerformanceByCategory/AP@0.5IOU/person:0.778 PascalBoxes_PerformanceByCategory/AP@0.5IOU/pottedplant:0.412 PascalBoxes_PerformanceByCategory/AP@0.5IOU/sheep:0.757 PascalBoxes_PerformanceByCategory/AP@0.5IOU/sofa:0.723 PascalBoxes_PerformanceByCategory/AP@0.5IOU/train:0.846 PascalBoxes_PerformanceByCategory/AP@0.5IOU/tvmonitor:0.684 PascalBoxes_Precision/mAP@0.5IOU:0.746