How AUTOIQ.ai Tuned the STM32N6 ISP for a Smart Doorbell — and Made It to CES 2026
- Nihil Saboo
- 3 days ago
- 5 min read
When SIANA Systems set out to build a smart doorbell with real-time presence detection on the STM32N6, they faced a problem that every embedded vision team eventually runs into: the default ISP configuration was good — but not good enough.
The images coming off the VD1943 sensor had a visible green colour cast, poor white balance across lighting conditions, and a colour accuracy error (ΔE) of 15.735 — well above the threshold where colours look visually off. More critically for a presence-detection doorbell, the on-device ML model was missing detections and generating false positives on challenging scenes. The two problems are connected, but solving them simultaneously by hand is practically impossible.
SIANA brought in Emmetra and AUTOIQ.ai to fix it. The result was demonstrated live at STMicroelectronics' CES 2026 booth. SIANA Systems put it plainly on LinkedIn:
Here's what happened under the hood.
The Problem: Two Objectives, One ISP
A smart doorbell lives in two worlds at once. A homeowner glances at a preview and needs colours to look natural. Simultaneously, an on-device model is deciding whether that shape at the door is a person, a parcel, or nothing. These two consumers of the ISP output want different things.
Tuning for human vision — sharpening colour accuracy, correcting white balance — often degrades the contrast and edge characteristics that ML models rely on. Tuning for machine vision tends to produce images that look flat or over-processed to human eyes. Engineers traditionally pick one and compromise on the other, or spend weeks iterating in circles.
AUTOIQ.ai is designed specifically for this situation: joint optimisation of human and machine vision KPIs within a single ISP configuration.
The Setup: STM32N6 + VD1943
The STM32N6 Discovery Kit is ST's latest edge-AI platform, integrating a hardware ISP pipeline alongside an NPU capable of running inference on-device at low power and low latency — exactly the profile a battery-friendly smart doorbell needs.
The camera sensor was a VD1943, a Bayer sensor suited to indoor and outdoor residential scenarios. AUTOIQ.ai targeted 25 tunable parameters across four ISP processing blocks — demosaicing, auto white balance, colour correction matrix, and contrast adjustment — with a 50/50 weighted objective split between human vision quality and machine vision detection performance.
Validation used two real-world image sets captured on the actual hardware:
Set 1 (challenging): cluttered indoor scene, varied lighting — the case where the default ISP struggled most
Set 2 (validation): portrait scenario used to confirm improvements generalised beyond the primary test set
2000 Iterations, 2–4 Hours
AUTOIQ.ai ran fully unattended. Each iteration programmed new ISP parameters onto the N6, captured the test images, computed all KPIs, and fed the scores back into the optimizer — no human in the loop between iterations.
+91.1% | +63.4% | 100% | 50–100× |
WHITE BALANCE | COLOUR ACCURACY | DETECTION RECALL | FASTER THAN MANUAL |
KPI | Default ISP | Tuned ISP | Change |
White Balance (ΔC) | 0.34 | 0.03 | +91.1% |
Colour Accuracy (ΔE) | 15.735 | 5.75 | +63.4% |
SNR (Luminance) | 40.4 dB | 36.9 dB | −0.086% |
Exposure Error | 1.17 | 1.24 | −0.059% |
Set 1 — Missed Detections | 2 | 0 | Recovered |
Set 1 — False Detections | 2 | 0 | Eliminated |
Set 1 — Confidence Stability | Variable (incl. negatives) | Consistent | +55.6% |
Set 2 — Missed Detections | 1 | 0 | Recovered |
A note on scope: These results were validated on two representative image sets captured on the target hardware. They demonstrate AUTOIQ.ai's optimisation capability on real scenes under controlled conditions — not a statistical claim across all possible lighting environments or deployment scenarios. As with any ISP tuning, we recommend validating the output configuration against your own scene library before production deployment.
Human Vision: Before & After
The green cast is gone. ΔC of 0.03 is near-perfect white balance. ΔE of 5.75 sits just above the threshold of perceptibility — practically indistinguishable from accurate colour in real-world use. The colour checker comparison below makes it immediately visible.
Colour checker comparison — Default ISP (left) vs. AUTOIQ tuned (right). Captured on STM32N6 + VD1943.
Machine Vision Set 1: The Challenging Case
On the cluttered indoor office scene, the default configuration produced 4 errors out of 4 expected detections — 2 missed objects and 2 false positives. Confidence scores were highly variable, including negative values from certain detection branches, making downstream filtering unreliable.
After tuning: 4/4 detections, zero false positives, stable confidence scores.
Machine Vision Set 1 — detection overlay comparison. Default (left) vs. tuned (right).
Machine Vision Set 2: Validation
On the portrait validation set, the default ISP missed one person entirely (confidence: 0). The fourth detection failed completely. After optimisation: all 4 subjects detected, with the previously failing detection recovering to 0.73 confidence — confirming the improvements generalise beyond the primary test scenes.
Machine Vision Set 2 — validation portrait set. Default (left) vs. tuned (right).
Why Manual Tuning Wouldn't Have Got There
The 25 parameters in scope are not independent. Demosaicing edge-detection strength affects both ML feature visibility and colour rendering artefacts. White balance gains interact with the colour correction matrix. The contrast curve shapes both aesthetic quality and object detectability. Adjusting any one cascades through the others.
Manual tuning addresses these sequentially — fix white balance, check colour accuracy, adjust edge detection, check ML performance, repeat. Each round risks undoing gains made in the previous round. With 25 interdependent parameters and two competing objective families, a human engineer exploring this space would need weeks and would realistically sample only a tiny fraction of what AUTOIQ.ai evaluated in 2000 iterations.
Manual ISP tuning at this complexity level typically takes 2 to 4 weeks. AUTOIQ.ai completed it in 2 to 4 hours — a 50–100× acceleration.— Emmetra internal benchmark, Dec 2025
Seen Live at CES 2026
The optimised configuration ran on SIANA Systems' smart doorbell demo at STMicroelectronics' CES 2026 stand — real-time edge-AI inference, no cloud dependency, presence and parcel detection on a low-cost MCU. The demo was part of ST's #STPartnerProgram showcase.
Want to see what AUTOIQ.ai can do for your platform?
We work with embedded vision teams to jointly optimise human and machine vision KPIs — on your hardware, with your sensor, in hours instead of weeks.
Written By
Nihil Saboo
Chief Product Officer, Emmetra
Nihil Saboo is the Chief Product Officer at Emmetra, where he leads product strategy with a focus on technological advancement. A self-described tech-savvy leader with a deep interest in robotics, Nihil blends technical curiosity with a tenacious and flexible approach to problem-solving. Known for his diligent coordination and execution, he is dedicated to driving innovation and operational excellence in the tech space.
Comments