Model Conversion Pre-Processing Configuration

As shown in Figure 1, the cropped or resized image are aligned. As a result, some areas are padded and are not the input required by the original model. To obtain the required image, you can copy the data to a new buffer and input the data to the model inference module. However, this process causes overheads. To reduce such overheads, the Matrix module provides a mechanism that allows padded areas in input images to the model manager (modelManager). The AIPP module of the model manager crops the image based on the user-defined width and height to generate an image that meets the model input requirements. When the image generated by the AIPP module is used, the inference module does not need to copy data, greatly improving the performance.

In the following example, the input image size for module inference is 224 pixels x 224 pixels, and the data obtained from DVPP is 128 pixels x 16 pixels aligned (the aligned image size is 256 pixels x 224 pixels).