Dev hits 60fps real-time depth estimation, targets 20ms end-to-end latency

yacineMTB (Yacine Atie) demonstrated a real-time stereo depth estimation system built on Fast Foundation Stereo: depth maps are generated at 60fps with pixel data streamed wirelessly over WiFi UDP, current inference takes only 10ms, and the next goal is to push end-to-end latency down to 20ms.

Confirmed

Depth estimation runs at 60fps, with pixel data streamed over WiFi UDP.
A single inference takes about 10ms, with an end-to-end target of 20ms.
After aggressive optimization (mainly targeting inefficiencies outside of inference), system latency has dropped to a level nearly imperceptible to the eye; multiple reposts mention a p99 latency of around 80 milliseconds.
The author clarified that the stuttering and tearing in the demo video were caused by screen-recording software interfering with the renderer, not by the depth generation itself (Twitter's encoding compression also affected how it looked).

Why it matters

The author's key takeaway: most of the latency came from various low-level engineering issues (what he called "stupid mistakes"), not from model inference itself. This is broadly instructive for developers working on real-time robotic vision — optimizing pipeline engineering often pays off more than swapping in a faster model.
The author plans to add event-triggered functionality for mosquito detection, showing the system is moving toward practical robotic perception.

2026-08-27 ~ 2026-08-27 · 6 related posts

Primary sources

60fps depth estimation achieved targeting 20ms end-to-end latency — yacineMTB · 2026-08-27 01:38 (UTC+08:00)
yacine gets real-time stereo depth latency down to imperceptible levels — yacineMTB · 2026-08-27 21:28 (UTC+08:00)
Yacine demos depth generation video; stuttering is his recorder, not the model — yacineMTB · 2026-08-27 21:29 (UTC+08:00)

[source] 60fps depth estimation achieved targeting 20ms end-to-end latency — yacineMTB · 2026-08-27
60fps depth streamed over WiFi UDP at 10ms inference, targeting 20ms end-to-end — astridwilde1 · 2026-08-27
Latency reduced to imperceptible levels via aggressive non-inference optimization — yacineMTB · 2026-08-27
[source] yacine gets real-time stereo depth latency down to imperceptible levels — yacineMTB · 2026-08-27
[source] Yacine demos depth generation video; stuttering is his recorder, not the model — yacineMTB · 2026-08-27
Robot Vision Latency Squeezed to 80ms p99 — Most Came From Dumb Things, Not Inference — REVOLVO_OCELOTS · 2026-08-27