solid state drive SSD OEM solid disks original: SSD: Single Shot MultiBox Detector(disks SSD)
SSD is in Faster Improvements after R-CNN and YOLO solid disks, using arch solid disks, Similar to you dividing the lattice for prediction, it is a one-stage detection method, and the main improvement is the introduction of multi-scale prediction, not only the prediction output at the end of the network. Because the lower layer contains more image detail, using a feature map of the lower layer of the network (shallow layer) can enhance the semantic information and improve the prediction effect.
anchor’s design
The size of the anchor in the SSD is not uniformly scaled according to the size of the feature map. Instead, a linear interpolation transformation is set manually:
set the following values for aspect ratios of anchor: According to a_r, anchor width is calculated as:disks SSD
so for each grid position, a total of 6 anchor pre-check boxes are generated.
in addition, the center coordinate of the anchor is:
compared to the previous practice, the anchor size shapes in the SSD are richer and can better match the objects in the image.
box matches
after generating the anchor preselector box, first for each ground truth box, select and it has the largest IOU The pre-selection box (written as Jaccard overlap) is matched, labeled as a positive sample, for training.
are also all marked as positive samples.
so that aground truth can match multiple pre-selection boxes, resulting in more predictions of higher scores at training time.
(and look at it in detail later)
negative sample equalization
remove the prediction box labeled positive sample, and the other prediction boxes are negative samples. Positive and negative samples produce serious imbalances that are not conducive to training.
SSD is to sort all negative samples by classification loss and select a large part of the sample (difficult sample) as the negative sample for training. The final positive and negative sample ratio is controlled at 1:3.
loss function
the loss function of SSD is also a classification loss and a box regression loss and combination, in the form of:
In general, SSD calculates the box regression loss for all positive samples, weighted sum; Both positive and negative samples of classified losses are involved in the calculation.
network structure
SSD uses VGG-16 as the head network for feature extraction, which has more feature extraction layers than YOLO, such as the last 5\times5， 3×3， 1×1 feature map of size, and the degree of fusion of features is quite high.
for one of the m×n feature plots of size, the corresponding output size is (c+4)km, k is the number of anchor prediction boxes, c+4 for c classification confidences, and 4 box regression parameters. Unlike YOLO’s 1×1 convolution, the output of the predicted values here is obtained by convolution 3×3 convolution.
anchor preselector number of pre-select boxes k slightly adjusted according to the size of the feature map, conv4_3, conv10_2, conv11_2 only produces 4 kinds of pre-select boxes, removed a_r = \frac{1}{3}, 3, the other layers of the preselector box is still 6.
at the end of the network, for the prediction boxes generated by different layers, the unified NMS is used for integration to obtain the prediction results.
depending on the size of the input image (300×300, 512×512), there are two versions of the network, SSD300, and SSD512, SSD512 has an additional convolutional layer conv12_2 for prediction, and the rest of the settings are the same.
results collation
SSD uses a pre-trained VGG16 network initialization during training, fine-tune, and Xavier initialization method for newly added convolutional layers.
ablation experiment:
Ablation experiment for multi-scale prediction:
by removing some of the prediction layers and increasing the number of anchor pre-selectors for other layers, so that the number of pre-selectors in different configurations is the same. The results can see noticeable performance degradation.
SSD uses multi-scale prediction and generates prediction boxes of different sizes on different size feature maps, which divides labor and improves performance.
reference solid-state drive SSD OEM solid disks