yolo_net.py:
輸入圖像尺寸為448，cell_size 為7
self.boundary1 = self.cell_size * self.cell_size * self.num_classes ?# 7 x 7 x 20
self.boundary2 = self.boundary1 + self.cell_size * self.cell_size * self.boxes_per_cell ?
? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ?# 7 x 7 x 20 + 7 x 7 x 2
self.output_size = (self.cell_size * self.cell_size) * (self.num_classes + self.boxes_per_cell * 5)
網(wǎng)絡(luò)輸出predict 為 fc_32層，Labels shape = [batch_size, 7, 7, 25]
類別預(yù)測(cè): ?#[batch_size, 7, 7, 20]
? ? ? ?predict_classes = [self.batch_size, self.cell_size, self.cell_size, self.num_class])
定位預(yù)測(cè): #[batch_size, 7, 7, 2]
? ? ? ?predict_scales = [self.batch_size, self.cell_size, self.cell_size, self.boxes_per_cell]
box大小預(yù)測(cè): #[batch_size, 7, 7, 2, 4]
? ? ? ? predict_boxes = [self.batch_size, self.cell_size, self.cell_size, self.boxes_per_cell, 4]
Label：
類別結(jié)果：
response = [self.batch_size, self.cell_size, self.cell_size, 1] # [batch_size, 7, 7, 1]
定位結(jié)果:
boxes= [self.batch_size, self.cell_size, self.cell_size, 1, 4]) # [batch_size, 7, 7, 1, 4]
box大小結(jié)果：
boxes=tf.tile(boxes, [1, 1, 1, self.boxes_per_cell, 1]) / self.image_size ?# [batch_size, 7, 7, 2, 4]
