 Hi, welcome back to the part 4 of this video series. In this video, we will go through the execution performance and memory footprint of the foot recognition application. Before we start, please know that all the measurements in this video are based on FPEI Vision 1 version 2.1 using another version of camera board, the STM32-F4 discount. If you have the latest results using the new B-CAMS OMV camera board, please refer to the UM2611, the user manual for the function pack. First, let's see the frame capture and the pre-processing time. This stage includes camera frame capture, image rescaling, pixel color conversion, and the pixel format adaptation. The measurements were taken with STM32-H747 M7 core running at 400 MHz and the code was compiled with E-Warm version 8.40.2 with Option-03. As you can see in the table, VGA resolution surely needs more time than the QVGA, and the camera frame capture takes most of the time at this stage. But please note, this value depends on the lighting conditions since the exposure time may be different. This is the same table we have seen in the Part 3 video. It lists the various memory configurations and their corresponding boundaries. Now let's see the execution performance of these different configurations. First, for the float model, all configurations use VGA resolution, and we have fully external and split schemes for the volatile memory. As a reminder, the fully external scheme is to place all buffers in the external SD RAM. The split scheme is to place activation buffer and input buffer in the internal SD RAM and camera buffers in the external SD RAM. So as you can see, the split memory scheme has better performance. Regarding the non-volatile memory that contains weights and biases, of course, internal flash gives the best performance, then is external SD RAM, and QSVI external flash is the slowest among the three. For the quantized model, the performance is about 2-3 times better than the float model. With VGA resolution, the same as float model, split scheme performs better. With QVGA resolution, as we mentioned in Part 3, there are two different optimized schemes. They have the same inference time as the VGA split case, but the FPS optimized scheme has a greater FPS. Finally, the quantized model has almost the same accuracy as the float model, but has much less inference time. In addition, it requires a smaller memory space, which we will talk about in the following slides. Still, the same table, the same memory configurations, and here are the memory footprint. The code size of the quantized model is 20 kilobytes more than the float model, but the weight and bias table is less than one third of the float model. And for the RAM size, in fully external scheme, the quantized model saved 400 kilobytes in external SD RAM. In the split case, it saved 250 kilobytes in the internal SD RAM. For both models, in VGA resolution, the external SD RAM takes a lot of memory because it is used for the camera buffers. But it's zero for quantized model in the QVGA solution because all the buffers are overlaid, which means a unique physical memory space in the internal SD RAM is used for all the buffers. Okay, this is all for this part. Thank you for taking your time to watch this video. In the next video, I will show you a hands-on demo of how to run a food recognition application using the QSPI external flash on the SDM32 H747i Discovery board. See you in the next video.