Chen & Brat expose fake 0ms latency in browser webcam gaze trackers, release open benchmark

Browser-based webcam gaze tracking is rapidly expanding into crowd-scale data collection and clinical settings where lab-grade eye trackers aren't practical. But according to a new paper by Chi-Sheng Chen and Gabriel A. Brat, accepted at the DEMI 2026 workshop (MICCAI), the latency numbers reported by these systems are often misleading. The root cause: most trackers timestamp each gaze sample when it is emitted by the inference engine, not when the source video frame was actually captured. This makes measured inference latency appear to be roughly 0ms—even when the underlying engine requires tens or hundreds of milliseconds to process a frame. The paper calls this out as a serious integrity problem for anyone building on these tools.

To fix this, Chen and Brat propose a new “capture-clock” methodology that reconstructs the true timing of each webcam frame. They leverage the browser's requestVideoFrameCallback (rVFC) API, which exposes a captureTime for local camera streams—or, when unavailable, a presentationTime that yields a verifiable lower bound on latency. For engines that expose their inference pipeline, the method enables exact source-frame pairing via a per-frame queue. For closed systems like WebGazer, it still produces a rigorous lower bound. The authors release an open TypeScript implementation and benchmark harness, and demonstrate it on two interchangeable engines: WebGazer and a new FaceMesh+KRR pipeline. This allows developers to compare real-world performance honestly, not just headline numbers.