VoxHOD / roadmap-voxhod-light.svg
Card: held-out validation split, capacity check, NOMAD empty-frame experiment (2026-11-10); nothing published
4ad4462 verified 5 days ago
4.7 kB
<svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 860 409" width="860" height="409" font-family="-apple-system, BlinkMacSystemFont, 'Segoe UI', Helvetica, Arial, sans-serif"><rect x="0.5" y="0.5" width="859" height="408" rx="6" fill="#fcfcfb" stroke="#e1e0d9"/><text x="20" y="30" font-size="15" font-weight="600" fill="#0b0b0b">VoxHOD roadmap</text><text x="20" y="49" font-size="12" fill="#7d7b75">Rule: the device model never grows; larger models are only teachers. A step ships only if it passes its gate.</text><text x="122" y="100" font-size="11.5" font-weight="600" letter-spacing="0.6" fill="#7d7b75">NEXT</text><text x="365" y="100" font-size="11.5" font-weight="600" letter-spacing="0.6" fill="#7d7b75">THIS WEEK</text><text x="607" y="100" font-size="11.5" font-weight="600" letter-spacing="0.6" fill="#7d7b75">LATER</text><line x1="20" x2="840" y1="108" y2="108" stroke="#e1e0d9"/><text x="20" y="130" font-size="15" font-weight="700" fill="#0b0b0b">VoxHOD</text><rect x="122" y="116" width="233" height="135" rx="7" fill="#fcfcfb" stroke="#eb6834" stroke-width="1.6"/><rect x="122" y="116" width="4" height="135" rx="2" fill="#eb6834"/><text x="134" y="138" font-size="12.5" font-weight="600" fill="#0b0b0b">Retrain v3 and nano on the</text><text x="134" y="155" font-size="12.5" font-weight="600" fill="#0b0b0b">held-out split</text><rect x="130" y="166" width="217" height="83" rx="4" fill="#f1f0ec"/><text x="134" y="181" font-size="11.5" fill="#52514e">New validation (2026-11-10, 739</text><text x="134" y="196" font-size="11.5" fill="#52514e">people from unseen locations)</text><text x="134" y="211" font-size="11.5" fill="#52514e">halves the spread to 5 points;</text><text x="134" y="226" font-size="11.5" fill="#52514e">published files must be retrained</text><text x="134" y="241" font-size="11.5" fill="#52514e">to be compared on it</text><rect x="122" y="259" width="233" height="118" rx="7" fill="#fcfcfb" stroke="#e1e0d9" stroke-width="1"/><text x="134" y="281" font-size="12.5" font-weight="600" fill="#0b0b0b">Fewer false alarms, same recall</text><rect x="130" y="292" width="217" height="83" rx="4" fill="#f1f0ec"/><text x="134" y="307" font-size="11.5" fill="#52514e">Tried 2026-11-08/09: hard</text><text x="134" y="322" font-size="11.5" fill="#52514e">negatives, 1.5x weight, extra</text><text x="134" y="337" font-size="11.5" fill="#52514e">training, pruning and a QAT nano</text><text x="134" y="352" font-size="11.5" fill="#52514e">all slid along the same</text><text x="134" y="367" font-size="11.5" fill="#52514e">trade-off. Next: a larger teacher</text><rect x="365" y="116" width="233" height="103" rx="7" fill="#fcfcfb" stroke="#e1e0d9" stroke-width="1"/><text x="377" y="138" font-size="12.5" font-weight="600" fill="#0b0b0b">Empty frames from other cameras</text><rect x="373" y="149" width="217" height="68" rx="4" fill="#f1f0ec"/><text x="377" y="164" font-size="11.5" fill="#52514e">2026-11-10: empty NOMAD frames</text><text x="377" y="179" font-size="11.5" fill="#52514e">halve NOMAD false alarms at the</text><text x="377" y="194" font-size="11.5" fill="#52514e">same recall, but only for that</text><text x="377" y="209" font-size="11.5" fill="#52514e">camera; WiSARD unchanged</text><rect x="365" y="227" width="233" height="73" rx="7" fill="#fcfcfb" stroke="#e1e0d9" stroke-width="1"/><text x="377" y="249" font-size="12.5" font-weight="600" fill="#0b0b0b">Find small people more often</text><rect x="373" y="260" width="217" height="38" rx="4" fill="#f1f0ec"/><text x="377" y="275" font-size="11.5" fill="#52514e">Gate: ≥ 85% of 8-16 px people,</text><text x="377" y="290" font-size="11.5" fill="#52514e">same file size</text><rect x="607" y="116" width="233" height="90" rx="7" fill="#fcfcfb" stroke="#e1e0d9" stroke-width="1"/><text x="619" y="138" font-size="12.5" font-weight="600" fill="#0b0b0b">More varied aerial data (full</text><text x="619" y="155" font-size="12.5" font-weight="600" fill="#0b0b0b">WiSARD, SeaDronesSee)</text><rect x="615" y="166" width="217" height="38" rx="4" fill="#f1f0ec"/><text x="619" y="181" font-size="11.5" fill="#52514e">Gate: WiSARD recall &gt; 61.3%</text><text x="619" y="196" font-size="11.5" fill="#52514e">without losing HERIDAL</text><rect x="607" y="214" width="233" height="73" rx="7" fill="#fcfcfb" stroke="#e1e0d9" stroke-width="1"/><text x="619" y="236" font-size="12.5" font-weight="600" fill="#0b0b0b">Measure on an ESP32-S3</text><rect x="615" y="247" width="217" height="38" rx="4" fill="#f1f0ec"/><text x="619" y="262" font-size="11.5" fill="#52514e">P4 checked 2026-11-09: same</text><text x="619" y="277" font-size="11.5" fill="#52514e">recall as the PC on 163 tiles</text></svg>