Content Detail

Camera2 YUV Input Review: rowStride, pixelStride and Frame Lifetime

Check rowStride, pixelStride and Image release timing for Camera2 YUV_420_888 input. Use the vision SDK case and record fields for format, orientation and buffer ownership.

Industry Updates 2026-10-17 Winge Technology
Article Content

Technical review

For Camera2 integration, establish how buffers are arranged, when they are valid and when they are released. Do not read YUV_420_888 as contiguous NV21 based only on dimensions. Include rowStride, pixelStride, the image cropping rectangle, image orientation and ownership during asynchronous processing in the SDK input contract.

What rowStride and pixelStride mean

rowStride is the byte distance between the starts of adjacent rows. pixelStride is the byte distance between adjacent pixel samples within a row. Neither is simply the image width. Read the actual plane information returned by the target device to avoid treating padding or a different plane layout as valid pixels.

Android's ImageFormat reference and Image.Plane interface define formats and strides. This review applies those interfaces to integration checks; each target device still needs relevant input cases.

Check the input before passing it to the SDK

Swipe the table horizontally to view all columns.

Item Information to record Related checks
Image format and image planes Format, dimensions and each plane's rowStride and pixelStride Row/color errors, bounds and image cropping rectangle
Orientation and mirroring Sensor/display orientation and front-camera mirroring Portrait/landscape, left/right and coordinate transforms
Lifetime Image owner, copying policy and asynchronous completion Invalid reads, missing releases and blocked queues
Load and queue Capture/submission rates, backlog and dropping policy Latency and resource use under sustained input
Output coordinates Model input, image cropping rectangle and preview region Boxes compared with known positions

Assign ownership for asynchronous processing

The validity of Image data and its planes must cover the asynchronous task. If Image is released when a callback ends, the later task needs valid data obtained through the agreed method. If release is delayed, define who releases it after completion, cancellation or failure. Check sustained capture and failure paths as well as a successful single frame.

Use these fields for an integration issue record. No device measurements are preset.

Swipe the table horizontally to view all columns.

Record field Contents Purpose
Environment identity Device, Android version, camera and SDK revision Distinguish hardware and code changes
Input metadata Format, dimensions, image cropping rectangle and plane strides Recheck conversion assumptions
Ownership and timing Callback, submission, completion and release responsibility Diagnose invalid reads and backlog
Validation sample Publishable pattern, orientation and expected coordinates Compare conversion and coordinate results
Failure paths Cancellation, timeouts, disconnection and recovery Check agreed resource cleanup

Project scope and integration inputs

The RK3588 offline face-detection and vision SDK case, in Chinese covers Android, Camera2 and an offline CPU path. Face detection and identity recognition are different tasks; NPU acceleration and other devices need separate evidence. For RK3588 Android adaptation, in Chinese, provide hardware, system, samples, existing interfaces and the record above before defining the device matrix and tests.

Submit a Project Requirement

Leave your contact details and a brief requirement so we can review the technical direction and scope.

Online
Phone
13910119357
WeChat
WhatsApp
Winge Technology WhatsApp QR code Scan or click to contact us
Top