English is currently available.
Renesas Microcontrollers DISCUSSION
A vision model gives sensible results on test images but becomes unreliable with the live RA8P1 camera pipeline. How can the input path be checked before collecting a larger training set or changing the network?
Reference: Renesas: RA8P1 AI microcontroller (https://www.renesas.com/en/products/ra8p1).
Capture the exact tensor presented to inference and compare it with the desktop preprocessing output for the same frame. Check color order, crop, resize method, rotation, normalization and quantization scale. A correct model can produce poor results from a numerically different input pipeline.
Inspect stride and buffer lifetime as well as pixel values. Camera and display buffers may contain padding or be updated while another component reads them. Freeze one frame and run the complete processing sequence deterministically to separate data corruption from image-quality differences.
After the pipelines agree, test exposure, lighting, motion blur and camera placement across the intended conditions. Report accuracy and latency on live data. Saved images can omit acquisition artifacts, so they are a useful baseline rather than a complete deployment validation.
Back to Renesas Microcontrollers community
