Back to News & Insights
Frontier Observations

Open-source calibration and extrinsics: Kannala-Brandt, intrinsics/extrinsics, and acceptance-ready fields

Explains how the Kannala-Brandt fisheye model, intrinsics, and extrinsics versions in the open-source SDK are incorporated into data cards and acceptance fields, providing reproduction scripts, error reports, and recommendations for regression testing after firmware changes, to help vision and SL

2026-09-08
Chief News Officer, TOBI Technology
Frontier Observations
Multi-camera calibration and wearable capture
Chief Insights Open-Source Frontier 2026-09-13 Chinese original: about 6528 characters

Lead: Calibration Is Not an Accessory, It Is a Gate

If the open-source SDK provides the Kannala-Brandt fisheye model and multi-camera extrinsics, it is equivalent to laying out the geometric ground truth for integrators. The key to acceptance is not whether the parameters can be printed, but whether the fields are complete, the versions are traceable, the errors are reproducible, and regressions occur after firmware changes. This article explains how to turn open-source calibration into a gate that both training and procurement can enforce.

Models and Fields: Write Them Clearly to Reproduce

Kannala-Brandt is suitable for wide fields of view, but the parameter meanings, number of distortion terms, and pixel coordinate origin must be consistent with the implementation. Open-source documentation should provide a field dictionary; integrators must not guess based on experience.

Intrinsics and extrinsics are versioned separately. Updating only one while sharing the old file name is a breeding ground for silent errors.

Coordinate frame definitions include axis orientation, units, and head-mounted rigid-body assumptions. Flexible head-mounted setups must be worn within the constraints of the protocol.

Example parameters are for format illustration only; production is governed by on-site reports.

Error Reports and Spot Checks

The report must include at least reprojection error percentiles, the hash of the image manifest used, the date, the device serial number, and the operator.

Before ingestion, spot-check golden scenarios; if a check fails, reject the entire batch or down-weight it.

Unify protocols and thresholds across sites; any differences must be declared.

Version the spot-check scripts to prevent drift in acceptance criteria.

Multi-camera devices
Figure 1 | Six-camera extrinsic calibration relies on the rigid-body assumption and versioned parameters; open source provides the format, and on-site reports provide the results.

Conditions for triggering re-labeling

Firmware changes, module replacements, severe impacts, long-term deformation, and error drift beyond thresholds all trigger recalibration.

Recalibration is not distrust of open source; it is an acknowledgment that physics changes.

There must be a strategy for handling sessions with old parameters during relabeling: freeze, remap, or isolate.

Post the trigger conditions at the workstation to reduce disputes.

Joint Acceptance with the Timeline

Geometrically correct but temporally wrong is equally untrainable. The relationship between the clock during calibration and during recording must be declared.

In visual-inertial joint calibration, specify the inertial axis orientation and the video time alignment strategy.

The joint acceptance package should include both the calibration report and the synchronization health score.

Gate logic: all conditions must be met to enter the formal dataset.

Calibration spot check for a wearable setup
Figure 2 | Worn-state deformation affects extrinsic calibration; the protocol should specify daily visual inspection and periodic re-calibration.

Calibration Changes in Open-Source Collaboration

If a PR changes a model or field, migration notes and examples must be updated in sync.

Breaking changes bump the major version and are pinned at the top of the release notes.

For security, calibration images may contain environmental privacy; attachments should go through a controlled channel.

Community-contributed calibration tools must be labeled with the applicable device models to prevent misuse.

Acceptable Field Checklist

Treat the following fields as the minimum completeness requirements for open-source calibration delivery.

  • Model type and parameter vector
  • Coordinate frames and units
  • Calibration date and device serial
  • Operator and script version
  • Error quantiles and image manifest hashes
  • Association with firmware/mechanical versions
Binocular geometry
Figure 3 | Stereo rigs also require versioned intrinsics; example parameters must not be used directly as production parameters.

From Calibration to Purchasable Data

Purchasable datasets must inherit the calibration version. Cheap hours without a version number are geometrically unauditable.

When customers mix self-captured and outsourced data, bucket by calibration version.

In content, publishing acceptance fields is more credible than boasting about accuracy numbers.

Mature calibration governance is the bridge from hardware open source to data products.

Summary of Practical Recommendations

Week 1: finalize the field dictionary and spot-check scripts. Week 2: complete training on trigger conditions. Week 3: connect the gate to ingestion.

Reject "collect first, calibrate later."

Report to management using error distribution plots rather than single-point best values.

Open source lays the model out flat; the team must thicken the governance.

Engineering Deep Dive, On-Site Access Control, Q&A and Appendix Requirements

Around the "field dictionary," it is recommended to track progress and risks separately in the project weekly report: one sentence each for verified facts, unverified assumptions, and next-step experiments. This way, open-source integration will not stop at feeling good, but at inspectable evidence. The same format can also be used to sync with external collaborators, reducing wasted meeting time.

During change windows involving "intrinsic parameter versions," freeze parallel large-scale changes: first complete a small-scope regression, then open up multi-person commits. Open-source hardware is most vulnerable when multiple variables change at once. The regression package should include shortest-path success examples and known failure examples to ensure failures are reproducible.

Add "extrinsic calibration version" to the first-week onboarding tasks: not a slogan to read, but a hands-on operation to complete once and submit a screenshot or log summary of the record. A low pass rate indicates a documentation gap, not merely a gap in the new hire's ability. Documentation gaps must be fed back into the repository or the materials ledger.

Build a counterexample library for "reprojection quantiles": correct practices must be placed alongside incorrect practices and their consequences. Field personnel remember counterexamples more deeply than rules. The counterexample library is pruned and extended once per quarter to stay in sync with the current firmware and toolchain.

If metrics drift in relation to the "image manifest hash," first perform layered diagnosis: device, host, network, procedure, human factors. Jumping directly to changing the model or hardware selection is prohibited. Layered diagnostic records are retained for ninety days for audit and retrospective.

When demonstrating "triggering recalibration" externally, prioritize showing the gate and logs rather than only the final output. Decision-makers trust repeatable processes more. Version the demo script to avoid inconsistent commitments caused by improvisation each time.

The person responsible for the "rigid-body assumption" must have clear veto authority and an escalation path. A quality role without veto authority is merely decorative. The escalation path must be written into the support manual so it can be executed even on holidays.

When measuring "visual-inertial" performance, do not rely solely on a binary success/failure indicator; retain the distribution: the 50th and 95th percentiles and the extreme tail. The tail determines production pain. Distribution plots brought to the weekly meeting drive resource allocation better than averages.

When "golden spot checks" require cross-team collaboration, first write a one-page interface: inputs, outputs, timeouts, failure codes, owner. An interface is cheaper than a consensus meeting. Interface changes go through a short review to prevent private decisions from affecting the whole system.

When reviewing historical incidents of the "cross-site protocol", use a unified template: phenomenon, affected sessions, root cause, fix, and whether re-collection is required. The template forces a distinction between sporadic and systemic issues. Systemic issues must be fed back into firmware, documentation, or procedures — not resolved by individual heroics.

In budget discussions, break down the cost of "example non-production" into tooling time, human-factors time, and rework time. Looking only at the tooling purchase price underestimates total cost of ownership. The disaggregated numbers better support investment in open-source governance.

Keep the minimal runnable example related to "migration notes" always runnable: a broken example equals broken documentation. Run examples nightly in CI, and block merges on failure. Examples are also the fastest path for customer onboarding.

Around the "privacy annex," it is recommended to track progress and risks separately in the project weekly report: one sentence each for verified facts, unverified assumptions, and next-step experiments. This way, open-source integration will not stop at feeling good, but at inspectable evidence. The same format can also be used to sync with external collaborators, reducing wasted meeting time.

During change windows involving "applicable device models," freeze parallel large-scale changes: first complete small-scale regression, then allow multi-user submissions. Hardware open source fears simultaneous movement of multiple variables most. The regression package should include shortest-path success examples and known failure examples to ensure failures are reproducible.

Write "bucketed training" into the new hire's first-week tasks: not reading slogans, but personally completing one operation and submitting a screenshot of the record or a log summary. A low pass rate indicates a documentation gap, not merely the new hire's ability. Documentation gaps must flow back into the repository or the materials ledger.

Build a counterexample library for "error distribution plots": next to each correct practice, place the incorrect practice and its consequences. Field personnel remember counterexamples more deeply than regulations. The counterexample library is pruned and extended once per quarter to stay in sync with the current firmware and toolchain.

If metrics drift in relation to "gate binding," first perform layered diagnosis: device, host, network, procedure, human factors. Jumping directly to changing the model or hardware selection is prohibited. Layered diagnostic records are retained for ninety days for audit and retrospective.

When demonstrating "firmware association" externally, prioritize showing the gate and logs rather than only the final output. Decision-makers trust repeatable processes more. Version the demo script to avoid inconsistent commitments caused by improvisation each time.

Kannala-Brandt is suitable for wide-FOV fisheye lenses, but parameter identifiability depends on calibration target coverage and pose diversity. Open-sourcing the model and extrinsics does not mean on-site inspection can be skipped.

Intrinsic and extrinsic parameters should be versioned in separate files or separate fields. Upgrading only the extrinsics while reusing old intrinsics—or vice versa—will cause hard-to-detect reprojection error drift.

Recommended acceptable fields include: model type, parameter vector, coordinate system definition, calibration date, device serial number, operator, reprojection error quantiles, and the hash of the image list used.

Trigger recalibration after firmware changes, lens module replacement, strong impact, or long-term deformation. Recalibration is not a lack of trust in open source; it is an acknowledgment that the physical world changes.

Extrinsics pin multiple cameras to the same rigid-body assumption. When the headset has too much flexibility, the rigid-body assumption fails, and the procedure must restrict tightening torque and wearing method.

If visual-inertial calibration is performed separately, the relationship to the recording clock must be declared to avoid inconsistent time semantics between the calibration state and the recording state.

Before applying open-source calibration results for training, run a golden-scenario reprojection spot check. If the check fails, reject that batch rather than relying on the network to compensate.

For cross-site projects, the calibration protocol must be unified. Site A cannot use a checkerboard while Site B uses a QR code with different error thresholds and no declaration of the difference.

Example parameters in documentation must not be used directly as production parameters. Examples illustrate format; production follows on-site reports.

Bind the calibration gate to the dataset version: sessions without a qualified calibration report must not enter the purchasable or publishable subsets.

The field dictionary uses Chinese-English bilingual entries to prevent international partners from misreading distortion items.

The sampling ratio for incoming material inspection may be adjusted according to supplier stability, but it cannot be reduced to zero.

After a recalibration is triggered, there must be a written policy on whether old sessions are reprocessed, to prevent geometric fragmentation of the training set.

Normalize headband torque with a simple torque tool to reduce extrinsic drift.

Photograph and archive the golden-scene layout diagram; when changing venues, reproduce the geometry before capture.

If the open-source toolchain outputs synonymous model names, create a synonym table to prevent missed matches in training code.

Report the 95th-percentile error rather than only the mean; the tail determines deployment risk.

When calibration images contain a whiteboard and a human face, follow the privacy process; do not exempt them on the grounds that it is "just calibration."

Cross-platform declaration: human-worn extrinsics are not directly equivalent to robot camera extrinsics; a mapping table is required.

Connect access control to the ingestion API: sessions without a report are rejected at registration.

Show management a control chart of error over time to demonstrate that governance is working.

Community-contributed automatic calibration can only serve as an auxiliary input if its failure modes are not declared; it cannot be the sole basis for gating.

If firmware release notes include camera driver changes, a link to the recalibration checklist is mandatory.

Extrinsic visualization (axes and reprojection) should be generated with one click for manual spot-checking.

The calibration fields of purchasable data packages use the same names as the open-source fields to reduce mixed-training costs.

Q: Can the example parameters be used on the production line? A: No.

Q: Why are intrinsic and extrinsic parameters versioned separately? A: To prevent silent mixing.

Q: Is the mean error sufficient? A: No; look at percentiles and the tail.

Q: Is recalibration required after an impact? A: Yes.

Q: How are flexible head-worn mounts handled? A: The procedure restricts wearing and torque.

Q: Can data without a report be ingested? A: No; the API should reject it.

Q: Different protocols across sites? A: This must be declared, otherwise the results are not comparable.

Q: Does the calibration image contain faces? A: Follow the privacy process.

Q: Can automatic calibration be the sole gate? A: No, unless its failure modes have been validated.

Q: What if the firmware changes the camera driver? A: A link to the mandatory recalibration checklist is required.

Q: What is the visualization for? A: Manual spot-checking of extrinsics and reprojection.

Q: What is the relationship to robot extrinsics? A: A mapping table is required; they are not directly equivalent.

Q: Should fields have Chinese-English bilingual names? A: Recommended, to reduce misreading.

Q: Can incoming-material sampling inspection be zero? A: No.

Q: Re-annotation strategy for old sessions? A: Written freeze, remapping, or isolation.

Q: An open-source PR changes the model? A: A migration note and examples are required.

Q: Who reviews the control charts? A: Management, to see whether governance is working.

Q: Should the data card use the same scene names? A: Using the same names as the purchasable set reduces mixed-training costs.

Q: What if the golden scene changes venue? A: Reproduce the geometry first.

Q: What about synonymous model names? A: Maintain a synonym table.

Q: Who has veto power? A: The calibration lead has veto power over ingestion.

Q: How often should periodic spot-checks occur? A: Determined by risk and documented in the procedure.

Q: Visual-inertial fusion? A: Declare the clock relationship.

Q: Which device models do the community tools apply to? A: Must be labeled.

Q: Can accuracy figures be advertised externally? A: Publishing acceptance fields is more credible.

Q: What is the parameter file naming convention? A: Include the version and device serial number.

Q: Where is the gate connected? A: The ingestion registration API.

Q: What is the value of open-sourcing? A: It lays out the format; governance still depends on the organization.

When implementing "open"-related capabilities, provide a Chinese-English bilingual field dictionary with versioning. In addition, the incoming-material spot-check ratio must be adjustable but never zero. It is recommended to write these requirements into the current week's checklist and review at next week's stand-up whether they have been implemented; any failure to implement should be recorded as a process defect, not an individual attitude issue. The same checklist can also be shared with external collaborators to reduce memory bias about what has already been said, keeping open-source integration and delivery acceptance always traceable.

When implementing "open"-related capabilities, document the old-session policy after recalibration. In addition, include the headband torque tool in the pre-shift check. It is recommended to write these requirements into the current week's checklist and review at next week's stand-up whether they have been implemented; any failure to implement should be recorded as a process defect, not an individual attitude issue. The same checklist can also be shared with external collaborators to reduce memory bias about what has already been said, keeping open-source integration and delivery acceptance always traceable.

When implementing "open"-related capabilities, golden scene layout diagrams should be archived and searchable. At the same time, a synonym model name table should be integrated into the training code. It is recommended to write the above requirements into the weekly checklist and review at next week's stand-up whether they have been properly executed; if not, record it as a process deficiency rather than an individual attitude issue. The same checklist can also be shared with external collaborators to reduce memory bias about "already mentioned" on both sides, keeping open-source integration and delivery acceptance always traceable.

When implementing "open"-related capabilities, include error control charts in the monthly report. In addition, calibration privacy attachments must go through a controlled channel. It is recommended to write these requirements into the current week's checklist and review at next week's stand-up whether they have been implemented; any failure to implement should be recorded as a process defect, not an individual attitude issue. The same checklist can also be shared with external collaborators to reduce memory bias about what has already been said, keeping open-source integration and delivery acceptance always traceable.

When implementing "open"-related capabilities, deliver the cross-platform mapping table together with the extrinsic parameters. At the same time, the ingestion API should reject requests that lack a report. It is recommended to write the above requirements into that week's checklist and review at the next stand-up whether they have been implemented; if not, record it as a process deficiency rather than an individual attitude issue. The same checklist can also be shared with external collaborators to reduce both sides' memory bias about "what has already been said," keeping open-source integration and delivery acceptance always traceable and documented.

When implementing "open"-related capabilities, automatic calibration failure modes must be disclosed. In addition, firmware notes must include a link to the mandatory recalibration checklist. It is recommended to write these requirements into the current week's checklist and review at next week's stand-up whether they have been implemented; any failure to implement should be recorded as a process defect, not an individual attitude issue. The same checklist can also be shared with external collaborators to reduce memory bias about what has already been said, keeping open-source integration and delivery acceptance always traceable.

When implementing "open"-related capabilities, extrinsic visualization should be generated with one click. In addition, purchasable fields should use the same names as the open-source fields. It is recommended to write these requirements into the current week's checklist and review at next week's stand-up whether they have been implemented; any failure to implement should be recorded as a process defect, not an individual attitude issue. The same checklist can also be shared with external collaborators to reduce memory bias about what has already been said, keeping open-source integration and delivery acceptance always traceable.

When implementing "open"-related capabilities, the calibration lead has veto power over ingestion. In addition, the periodic spot-check frequency should be tiered by risk. It is recommended to write these requirements into the current week's checklist and review at next week's stand-up whether they have been implemented; any failure to implement should be recorded as a process defect, not an individual attitude issue. The same checklist can also be shared with external collaborators to reduce memory bias about what has already been said, keeping open-source integration and delivery acceptance always traceable.

When implementing "open"-related capabilities, declare the clock relationship for visual-inertial joint statements. At the same time, community tools should annotate applicable models. It is recommended to write the above requirements into the weekly checklist and review at the next stand-up meeting whether they have been implemented; if not implemented, record it as a process defect rather than a personal attitude issue. The same checklist can also be shared with external collaborators to reduce both sides' memory bias about "already said," ensuring that open-source integration and delivery acceptance are always traceable.

When implementing "open"-related capabilities, external communication should emphasize acceptance fields. At the same time, parameter file names should include version and sequence. It is recommended to write the above requirements into the weekly checklist and review at the next standup whether they have been properly executed; if not, record it as a process deficiency rather than an individual attitude issue. External collaborators can also share the same checklist, reducing both sides' memory bias about "what was already said," so that open-source integration and delivery acceptance always have a verifiable record.

Additional requirements: bilingual field dictionary with versioning. Headband torque tool included in pre-shift checks. Calibration privacy attachment goes through a controlled channel. If conflicts are found during execution, the written specification and current frozen version take precedence, and a documentation revision should be proposed within twenty-four hours to avoid relying on verbal ad-hoc explanations for long-term operation on site.

Additional requirements: the incoming-material spot-check ratio must be adjustable but never zero. Golden-scene layout diagrams must be archived and searchable. The cross-platform mapping table must be delivered together with the extrinsics. If conflicts are found during execution, the written specification and the current frozen version take precedence, and a documentation revision must be proposed within twenty-four hours, avoiding long-term reliance on ad hoc verbal explanations on site.

Additional requirements: the old-session policy after recalibration must be documented. The synonym table for model names must be integrated into the training code. The ingestion API must reject entries missing a report. If conflicts are found during execution, the written specification and the current frozen version take precedence, and a documentation revision must be proposed within twenty-four hours, avoiding long-term reliance on ad hoc verbal explanations on site.

Additional requirements: the headband torque tool must be included in the pre-shift check. Error control charts must be included in the monthly report. Automatic calibration failure modes must be disclosed. If conflicts are found during execution, the written specification and the current frozen version take precedence, and a documentation revision must be proposed within twenty-four hours, avoiding long-term reliance on ad hoc verbal explanations on site.

Additional requirements: golden scene layout diagrams must be archived and searchable. Calibration privacy attachments go through a controlled channel. Firmware documentation must include a mandatory re-calibration check link. If conflicts are found during execution, the written specifications and the current frozen version prevail, and a documentation revision must be proposed within twenty-four hours to avoid relying on verbal on-site explanations for long-term operation.

Additional requirements: Integrate the synonym model-name table into the training code. Deliver the cross-embodiment mapping table together with the extrinsic parameters. Generate extrinsic visualizations with one click. If conflicts are discovered during execution, the written specification and the current frozen version take precedence, and a documentation revision must be proposed within twenty-four hours to avoid relying on ad hoc verbal explanations on site for long-term operation.

Supplementary requirements: the error control chart enters the monthly report. Ingestion API rejects requests missing the report. Procurement-available fields share names with the open-source ones. If conflicts are found during execution, the written specification and the current frozen version prevail, and a documentation revision should be raised within twenty-four hours to avoid relying on verbal on-site explanations for long-term operation.

Additional requirements: Calibration privacy attachments must be delivered through a controlled channel. Automatic calibration failure modes must be disclosed. The calibration lead holds veto authority over repository admission. If conflicts are discovered during execution, the written specification and the current frozen version take precedence, and a documentation revision must be proposed within twenty-four hours to avoid relying on ad hoc verbal explanations on site for long-term operation.

Additional requirements: deliver the cross-platform mapping table together with the extrinsics. Firmware documentation must include a re-labeling check link. Periodic spot-check frequency is risk-tiered. If conflicts are found during execution, the written specification and the current frozen version take precedence, and a documentation revision should be filed within twenty-four hours, avoiding reliance on ad-hoc verbal explanations on site for long-running operations.

Additional requirements: the ingestion API rejects entries missing a report. Extrinsic visualization is generated with one click. Visual-inertial joint declaration of clock relationships. If conflicts are found during execution, the written specification and the current frozen version take precedence, and a documentation revision must be proposed within twenty-four hours, avoiding long-term operation relying on ad hoc verbal explanations on site.

Additional requirements: Automatic calibration failure modes must be disclosed. Purchasable fields must have the same names as the open-source ones. Community tools must be annotated with the applicable models. If conflicts are found during execution, the written specification and the current frozen version take precedence, and a documentation revision must be raised within twenty-four hours, to avoid relying on verbal on-the-spot explanations for long-term operation.

Additional requirements: Firmware documentation must include a mandatory recalibration check link. The calibration lead holds veto authority over repository admission. External communications must emphasize acceptance fields. If conflicts are discovered during execution, the written specification and the current frozen version take precedence, and a documentation revision must be proposed within twenty-four hours to avoid relying on ad hoc verbal explanations on site for long-term operation.

Additional requirements: one-click generation of extrinsic visualization. Periodic spot-check frequency graded by risk. Parameter file names include version and serial number. If conflicts are found during execution, the written specification and the current frozen version prevail, and a documentation revision must be proposed within 24 hours, avoiding relying on verbal on-site explanations for long-term operation.

Supplementary requirements: purchasable fields must match the open-source names. Visual-inertial joint declaration of clock relationships. Field dictionary in Chinese-English bilingual and versioned. If conflicts are found during execution, the written specification and the current frozen version take precedence, and a document revision must be proposed within twenty-four hours to avoid relying on ad-hoc verbal explanations on site for long-term operation.

Supplementary requirements: The calibration lead holds veto authority over data ingestion. Community tools annotate applicable models. Incoming material sampling ratio is adjustable but must not be zero. If conflicts are discovered during execution, the written specification and the current frozen version shall prevail, and a documentation revision shall be proposed within twenty-four hours, to avoid relying on ad hoc verbal explanations on site for long-term operation.

Additional requirements: Periodic spot-check frequency shall be tiered by risk level. External communications shall emphasize acceptance fields. After re-calibration, the legacy session strategy shall be documented in writing. If conflicts are discovered during execution, the written specification and the current frozen version shall prevail, and a documentation revision shall be proposed within twenty-four hours, avoiding reliance on ad hoc verbal explanations in the field for long-term operation.

Additional requirements: Visual-inertial joint statements must declare clock relationships. Parameter file names must include version and sequence. The headband torque tool must be included in the pre-shift checklist. If conflicts are discovered during execution, the written specification and the current frozen version take precedence, and a documentation revision must be proposed within twenty-four hours to avoid relying on ad hoc verbal explanations on site for long-term operation.

Additional requirements: Community tools must be annotated with applicable device models. The field dictionary must be bilingual (Chinese-English) and versioned. Golden-scene layout diagrams must be archived in a searchable manner. If conflicts are discovered during execution, the written specification and the current frozen version take precedence, and a documentation revision must be proposed within twenty-four hours to avoid relying on ad hoc verbal explanations on site for long-term operation.

Additional requirements: external communications should emphasize acceptance fields. Incoming material spot-check ratios may be adjusted but must not be zero. A table of synonymous model names must be integrated into the training code. If conflicts are found during execution, the written specification and the current frozen version take precedence, and a documentation revision must be proposed within twenty-four hours, avoiding reliance on ad hoc verbal explanations on site for long-term operation.

Additional requirements: Parameter file names must include version and sequence. Post-recalibration policies for legacy sessions must be documented in writing. Error control charts must be included in the monthly report. If conflicts are discovered during execution, the written specification and the current frozen version take precedence, and a documentation revision must be proposed within twenty-four hours to avoid relying on ad hoc verbal explanations on site for long-term operation.

Next Step: Moving from an Open-Source SDK to an Acceptable Joint Integration

If your team is evaluating open-source calibration and Kannala-Brandt extrinsics, first complete cloning, package verification, and host-side integration based on the GitHub README and the delivery package, then document synchronization, upload, calibration, and licensing boundaries in the project acceptance appendix. The open-source repository helps you get started quickly; the written specification determines the delivery scope.

Go to Developer Docs · E2 / E6 Device SDK Get the clone command, Drive / Baidu materials, and Linux integration instructions, or directly Contact Technical Integration Align versions, samples, and acceptance fields.