AVION 0.1: a perception front end for robot brains, in one C99 header
Most vision stacks feed a model frames that were decoded, copied and allocated for. AVION does the work in front of the model on the GPU, with zero heap allocations while it runs, and hands the network a foveated NCHW tensor (fp32/fp16) with no copy.
What it gives a model, from an ordinary camera:
• A virtual IMU: the camera’s shake read from pixels, 0.031 px rms off the truth
• A stabilised event-camera (DVS) stream: a shaken clip’s 17.2M events become 1.71M, only what moves
• Time to contact: 1.24 s read for a true 1.33 s
• A pulse through camera shake: 71.98 BPM; 72.00 from Cr chroma alone
• Scene cuts from H.264/HEVC slice headers before decoding: 0.23 µs a 4K frame
Head to head on one machine, one allocation counter: 0 allocations per 1,000 frames against 132,292 (ffmpeg) and 20,000 (OpenCV DIS). And where we lose, we say so: VLC plays 720p with less CPU.
Every picture is a snapshot the engine wrote itself, against a truth we set.
Model card: [alexandre5645/avion-0.1 · Hugging Face](https://huggingface.co/alexandre5645/avion-0.1)
Site, method and evaluation bundle: [AVION 0.1](https://5645hm-a11y.github.io/avion/)