Event

Running object detection on STM32F407 with TFLite Micro

1 reports · 1 sources

Embedded JunG published a hands-on tutorial that walks through the full pipeline for running an object detection model with TFLite Micro on an STM32F407 (168MHz, 192KB SRAM): model conversion, full int8 quantization, porting, and inference. The author reports that a full inference takes about 200ms per frame, and that with an OV7670 camera it can reach 4 to 5 frames per second. The tutorial stresses that full integer quantization is the most critical step for edge deployment: calibration data is randomly drawn from the training set in a batch of 100 to 200 images, returned by a generator in batches, and needs to cover enough variation in brightness, angle, and class; it also recommends first running the same int8 model on a PC to confirm the outputs are normal before porting to the MCU.