tanırım
CompletedBenchmark

In-Browser Hand Tracking Benchmark

Pose Estimation

Context

These measurements were taken while building a browser app that solves a 2×2 Rubik's cube with hand gestures alone. Hand tracking runs on every frame there, so it mattered how smooth it would stay on each device. The same table is in the project's README.

Source code

Finding

The laptop measurements were taken on the same computer and the iPhone measurements on the same phone, so only the effect of the execution path (graphics card or processor) shows. The integrated graphics card, at 20.1 ms, is only 25% slower than the discrete one (16.1 ms). Switching to the processor raises the time to 59.3 ms, about 3 times slower, but 21 fps is still enough for real-time use. On the iPhone 12 Pro the time is 33 ms on the graphics card, but because the graphics card is shared with drawing the screen, it rises to 42 ms on slow frames. On a phone running on the processor, 10 fps leaves the experience at the edge. What really determines smooth tracking on mobile is whether WebGL is available.

Results

Graphics card (WebGL)RTX 3060 (GPU)Intel iGPUiPhone 12 Pro (GPU)
Processor (WASM)i7-11800H (CPU)iPhone 12 Pro (CPU)

Latency (p50) (ms)

RTX 3060 (GPU)
16.1
Intel iGPU
20.1
i7-11800H (CPU)
59.3
iPhone 12 Pro (GPU)
33
iPhone 12 Pro (CPU)
96

Latency (p95) (ms)

RTX 3060 (GPU)
17.3
Intel iGPU
22.8
i7-11800H (CPU)
62.3
iPhone 12 Pro (GPU)
42
iPhone 12 Pro (CPU)
98

Tracking Rate (fps)

RTX 3060 (GPU)
57
Intel iGPU
51
i7-11800H (CPU)
21
iPhone 12 Pro (GPU)
25
iPhone 12 Pro (CPU)
10

Related Technologies

// Help

Frequently Asked Questions

The duration of MediaPipe Hand Landmarker's detectForVideo call was measured on every frame. The measurement lasted 30 seconds, with both hands visible and moving slowly. fps is the effective tracking rate and is capped by the camera frame rate and the display refresh rate. Desktop measurements were taken with Chrome 136, iPhone measurements with Safari.

In the browser, MediaPipe first tries the graphics card through WebGL and falls back to the processor with WASM when WebGL is unavailable. In the measurements the processor path is about 3 times slower, but tracking remains usable at that speed.