File input and output can dominate measured pipeline performance. This is especially common when reading from or writing to network-mounted storage, or when writing large raw output frames for every buffer.
When measuring inference performance:
Use local board storage for input and output when possible.
Use
fakesinkwhile isolating inference throughput.Measure overlay, display, encode, and file output separately when they are part of the final application.
Avoid comparing pipelines that write different amounts of output data.