Differences
This shows you the differences between two versions of the page.
| Both sides previous revision Previous revision Next revision | Previous revision | ||
|
Member:sungbeanJo_tran [2021/01/17 14:38] sungbean |
Member:sungbeanJo_tran [2021/01/18 16:21] (current) sungbean |
||
|---|---|---|---|
| Line 1: | Line 1: | ||
| - | We use an input image that is resized to 288×512×3 with | + | Choosing scales and aspect ratios for default boxes To handle different object scales, |
| - | bilinear interpolation algorithm, hence to reduce computational burdens for a real-time detection. For the images with | + | some methods [4,9] suggest processing the image at different sizes and combining the |
| - | different aspect ratios, we cropped the height to match the | + | results afterwards. However, by utilizing feature maps from several different layers in a |
| - | ratio. Following a common practice in image classification | + | single network for prediction we can mimic the same effect, while also sharing parameters across all object scales. Previous works [10,11] have shown that using feature maps |
| - | tasks, we subtracted the mean RGB value to achieve zerocentered inputs, which are originally in different scales. Note | + | from the lower layers can improve semantic segmentation quality because the lower |
| - | that our dataset contains images where the camera gains | + | layers capture more fine details of the input objects. Similarly, [12] showed that adding |
| - | are automatically calibrated to obtain high-quality images. | + | global context pooled from a feature map can help smooth the segmentation results. |
| - | During the testing process, we also used a cropped image | + | Motivated by these methods, we use both the lower and upper feature maps for detection. Figure 1 shows two exemplar feature maps (8×8 and 4×4) which are used in the |
| - | in the center part of the image, where traffic lights are | + | framework. In practice, we can use many more with small computational overhead. |
| - | commonly observed in that area. Thus, a batch of two images | + | Feature maps from different levels within a network are known to have different |
| - | (i.e., whole and cropped images) are fed into our detector | + | (empirical) receptive field sizes [13]. Fortunately, within the SSD framework, the default boxes do not necessary need to correspond to the actual receptive fields of each |
| + | layer. We design the tiling of default boxes so that specific feature maps learn to be | ||
| + | responsive to particular scales of the objects. Suppose we want to use m feature maps | ||
| + | for prediction. The scale of the default boxes for each feature map is computed as: | ||
