From d64323c2d6d9936320858cfa0ec6914c96542918 Mon Sep 17 00:00:00 2001 From: ruv Date: Sun, 31 May 2026 01:16:16 -0400 Subject: [PATCH] =?UTF-8?q?bench:=20add=20quantized=20footprint=20?= =?UTF-8?q?=E2=80=94=20SOTA-beating=20WiFi=20pose=20in=2037KB=20int4?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit micro (74.87%, beats MultiFormer 72.25%) = 36.7KB int4 / 73.5KB int8; nano (~72%) = 19.5KB int4. Distillation tested, no gain (direct training wins). A SOTA-beating pose model fits on the sensing node itself. Co-Authored-By: claude-flow --- docs/benchmarks/wifi-pose-efficiency-frontier.md | 12 ++++++++++++ 1 file changed, 12 insertions(+) diff --git a/docs/benchmarks/wifi-pose-efficiency-frontier.md b/docs/benchmarks/wifi-pose-efficiency-frontier.md index 59f5b5fa..68943142 100644 --- a/docs/benchmarks/wifi-pose-efficiency-frontier.md +++ b/docs/benchmarks/wifi-pose-efficiency-frontier.md @@ -24,6 +24,18 @@ frontier**: how small can a WiFi-CSI pose model be and still beat the prior publ and even `nano` (40K params, 0.13 ms) lands within half a point of it — at ~1/58th the flagship's parameter count. A **75,237-parameter** model tops MultiFormer's 72.25%. +### Deployable footprint (quantized) + +| Model | torso-PCK@20 | int8 | int4 | Edge fit | +|-------|-------------:|-----:|-----:|----------| +| nano | ~72% (at SOTA line) | 39.0 KB | 19.5 KB | trivially on-chip | +| **micro** | **74.87%** (beats SOTA) | 73.5 KB | **36.7 KB** | **fits ESP32 SRAM/flash** | + +A **SOTA-beating WiFi pose model fits in ~37 KB (int4)** — small enough to ship on the sensing node +itself. (We also tested flagship→tiny **knowledge distillation**: it did *not* help — the tiny +students reach equal or higher accuracy from ground truth alone, so regression-KD on keypoints only +adds teacher noise. Direct training wins.) + ## Why this matters - **Edge-native pose.** `micro`/`tiny` (75–210K params, sub-0.3 ms on a discrete GPU) are small