umarrasydan/meannit-slide-box

Meannit slide box

The model Meannit runs on the Mac to find the slide in a meeting window: the slide's box, or that the window shows no slide. Meannit checks each slide it captured with it before the summary, on the GPU.

It is DINOv2 ViT-S/14 with registers (Meta, Apache 2.0) with a head of about 1M parameters, trained in MLX on 44,000 meeting and deck windows drawn in headless Chromium. This is round four's weights, averaged over its last 1,000 steps, converted to Core ML at 16 bits.

slide-box-v4.mlmodelc.zip holds the compiled model, slide-box-v4.mlmodelc, for macOS 14 and later.

• Input window: float32, (1, 448, 448, 3), the window letterboxed into the square, values 0 to 255, with bars of (123, 116, 103).
• Output slide: (1), the probability that the window shows a slide.
• Output edges: (1, 4), the slide's left, top, right, and bottom as shares of the square.
• Output peaks: (1, 4), each edge's peak probability. Their product is the box's confidence.

Run it with MLComputeUnits.cpuAndGPU: on the Neural Engine its answers are wrong.