Event

RDK X5 binocular depth estimation only has StereoNet; for a lighter model, you can only train one yourself

1 reports · 1 sources

In a developer community reply, Guazai stated that the only official binocular depth estimation model on RDK X5 is StereoNet (IGEV/GRU architecture), and there is no smaller ready-made network to choose from. The available options are different compiled versions of the same model: V2.4_int8 (640x352x3x2, 23fps) and V2.4_int16 (same resolution, 15fps); the documentation runs V2.4_int8 by default. If a smaller and faster binocular network is needed, according to Guazai the only option is to train and quantize a model yourself using the toolchain—HAT provides the StereoNet training pipeline, and after PTQ quantization it can be deployed on the board; the official team will not release a smaller ready-made model.