Mimir/docs/native-capture-cadence.md

Native Capture Cadence

This file records measured driver-path camera cadence. It is evidence, not a promise.

2026-05-21: PS3 Eye Raw USB Probe

Probe location: temp workspace only, outside this repo.

Probe path:

• raw PS3 Eye userspace driver reference;
• libusb/WinUSB interface 0;
• Bayer frame output;
• no DirectShow;
• no Media Foundation;
• no frame processing;
• five-second measurement window.

Installed device state:

• two PS3 Eyes present as VID_1415&PID_2000;
• interface 0 is available through the WinUSB/libusb-style path;
• interface 1 remains the media/audio side and is not the capture path measured here.

Sequential results:

Camera	Requested Mode	Frames	Elapsed	Delivered FPS	Avg Delta
0	320x240 @ 187	934	5.002s	186.72	5.355 ms
1	320x240 @ 187	917	5.001s	183.35	5.454 ms
0	640x480 @ 60	292	5.006s	58.33	17.118 ms
1	640x480 @ 60	290	5.010s	57.89	17.237 ms

Dual-camera results in one process:

Camera	Requested Mode	Frames	Elapsed	Delivered FPS	Avg Delta
0	320x240 @ 187	915	5.005s	182.81	5.476 ms
1	320x240 @ 187	916	5.005s	183.01	5.460 ms
0	640x480 @ 60	294	5.009s	58.69	17.068 ms
1	640x480 @ 60	286	5.009s	57.09	17.478 ms

Conclusion: the two PS3 Eyes can be treated as roughly 183 fps timing/marker witnesses at 320x240 on the current USB topology when using the raw USB path. At 640x480 they are roughly 57-59 fps.

The probe reported one unrelated invalid USB descriptor from USB\VID_0000&PID_0002; it did not prevent PS3 Eye capture.

2026-05-21: LeapUVC Kernel Streaming Probe

The Leap Motion Controller is currently present as VID_F182&PID_0003 and bound to usbvideo.sys, which means the active mode is LeapUVC/UVC. That mode is mutually exclusive with the Ultraleap SDK path for this device.

Probe path:

• Windows Kernel Streaming capture interface for KSCATEGORY_CAPTURE;
• LeapUVC through usbvideo.sys;
• no Media Foundation;
• no DirectShow graph;
• no OpenCV;
• no frame processing;
• five-second measurement window per advertised mode.

Probe source:

• native/probes/ks_camera_cadence

Advertised modes and measured queued KS pull rates:

Mode	Advertised FPS	Frames	Elapsed	Delivered FPS	Bytes / Frame
752x480 YUY2	50.00	173	5.008s	34.55	721,920
640x480 YUY2	57.50	204	5.017s	40.66	614,400
640x240 YUY2	115.00	415	5.004s	82.93	307,200
640x120 YUY2	214.00	777	5.000s	155.39	153,600
752x240 YUY2	100.00	308	5.013s	61.44	360,960
752x120 YUY2	190.00	589	5.008s	117.62	180,480

Conclusion: the current direct KS probe can pull real frames from LeapUVC, but even eight queued asynchronous reads do not reach the device's advertised cadence. The useful stereo IR mode is 640x240 YUY2, advertised at 115 fps and currently measured at about 83 fps. The next cut should either tune LeapUVC controls that affect exposure/frame interval or deliberately rebind to a WinUSB/libusb UVC path if we decide the Windows UVC stack is the bottleneck.

Follow-up control sweep:

• the probe now reports known LeapUVC-ish controls directly through KS;
• it discards a two-second warm-up before the five-second measurement window;
• 640x240 YUY2 steady-state cadence lands at about 110.7 fps;
• 640x120 YUY2 steady-state cadence lands at about 205.9 fps.

Steady-state 640x240 YUY2 control sweep:

Scenario	Result
baseline	554 frames / 5.007s = 110.65 fps
exposure 10us	554 frames / 5.006s = 110.68 fps
gamma off	554 frames / 5.002s = 110.76 fps
HDR off	554 frames / 5.003s = 110.73 fps
LEDs on	554 frames / 5.003s = 110.72 fps
LEDs off	554 frames / 5.006s = 110.68 fps
gain minimum	554 frames / 5.007s = 110.64 fps
dark-frame interval 0	555 frames / 5.009s = 110.80 fps
fast combined	554 frames / 5.001s = 110.77 fps

Rejected controls:

• digital gain via KSPROPERTY_VIDEOPROCAMP_BRIGHTNESS = 0 returned ERROR_INVALID_PARAMETER;
• the old LeapUVC FPS-ratio selector through KSPROPERTY_VIDEOPROCAMP_GAIN returned ERROR_INVALID_PARAMETER.

Conclusion: the earlier 83 fps number was mostly startup drag inside the short measurement window. Once the Leap is hot, the useful stereo mode is close to its advertised cadence through the current KS path. The half-height mode can beat the PS3 Eyes on raw frame rate, but the PS3 Eyes still win when comparing their full 320x240 frames against Leap's useful 640x240 stereo stream.

2026-05-21: Razer Kiyo Kernel Streaming Probe

Installed device state:

• Razer Kiyo camera: VID_1532&PID_0E03&MI_00;
• Razer Kiyo Pro camera: VID_1532&PID_0E05&MI_00;
• both expose UVC camera interfaces through the Windows Kernel Streaming capture category.

Probe path:

• Windows Kernel Streaming capture interface for KSCATEGORY_CAPTURE;
• no Media Foundation;
• no DirectShow graph;
• no OpenCV;
• eight queued asynchronous reads;
• two-second warm-up discarded before each five-second measurement.

Kiyo Pro advertised MJPG modes and measured cadence:

Device	Mode	Advertised FPS	Best Measured FPS	Note
Kiyo Pro	1280x720 MJPG	60.00	25.20	manual exposure did not improve cadence
Kiyo Pro	1920x1080 MJPG	60.00	25.18	manual exposure did not improve cadence

Plain Kiyo advertised MJPG modes and measured cadence:

Device	Mode	Advertised FPS	Best Measured FPS	Best Setting
Kiyo	1920x1080 MJPG	30.00	30.27	manual exposure -9 or -10
Kiyo	1280x720 MJPG	60.00	51.51	gain minimum after manual exposure sweep

Conclusion: plain Kiyo is exposure-limited by default and becomes usable once manual exposure is forced shorter. Kiyo Pro advertises 60 fps modes but the current KS path delivers about 25 fps regardless of manual exposure, so it needs a separate driver/control investigation before it should be trusted for high cadence capture.