DJI Terra Video-to-3D Modeling Guide: GPS Metadata, Auto Keyframes, Emergency Reconstruction & Accuracy Limits

01 — Check whether the source video qualifies

The current DJI Terra data-import manual lists .mp4 and .ts video inputs. It names Mavic 3T, Mavic 3E, Matrice 4E, Matrice 4T, Matrice 4D and Matrice 4TD as supported video-source aircraft. This is a documented model list, not a promise that every recording from those aircraft is suitable.

DJI's feature announcement requires GPS metadata. A filename ending in MP4 does not establish that the required flight data survived. A downloaded social clip, screen recording, edited compilation or third-party camera file should not be treated as equivalent to the original aircraft recording.

Keep the original video and matching flight data together

For missing-metadata error 1033, DJI specifically instructs users to check for a same-name .SRT flight-data file in the same directory. If it is missing, recover the complete original video and its matching SRT from the aircraft or SD card. Keep the original base filenames aligned; do not pair a clip with flight data from another recording.

Work from a processing copy and retain an untouched archive. Before importing, play the entire recording, check that the file size is plausible and note where it came from. DJI's troubleshooting guidance warns against third-party compression, transcoding or trimming of the original video. Select the useful segment inside Terra instead.

GPS can also be genuinely absent or unreliable during capture. DJI identifies prolonged GPS-loss environments, including tunnels, indoor scenes and obstructed areas, as unsuitable for this workflow. Recovering a missing companion file may fix an export problem; it cannot recover measurements that were never recorded.

02 — Confirm the edition and processing environment

Separate video import from the reconstruction output you want. DJI's current edition comparison lists video import across the editions in its detailed table. Standard supports visible-light mesh reconstruction and DSM/DOM outputs. Gaussian splatting is listed for Flagship and Cluster, not Standard. Ordinary video-derived mesh reconstruction therefore should not be sold as requiring Flagship solely because the source is video.

The import manual also says extracted frames can enter aerotriangulation, 2D, 3D and Gaussian reconstruction. That describes available downstream workflows, not a guarantee that a particular clip has appropriate coverage or that your licence enables every output. For a 2D map, for example, an oblique orbit may leave ground hidden even if it produces a recognizable building model.

Before committing a full archive, record the installed version, activated edition, desired output and project location. DJI's release notes place video import in v5.3.0, released July 23, 2026; the reviewed downloads and release notes show v5.3.5 dated August 20. Check the current installation rather than assuming an older workstation has the feature.

These are configuration requirements, not a processing-time estimate. A short clip can contain many redundant frames, while a longer clip can contain little usable geometry. Start with a representative segment and observe the actual extracted image count and project resource use. The configuration manual is the appropriate reference when sizing a workstation.

A repeatable first run

03 — Extract frames, then test the reconstruction

The workflow is still photogrammetry: Terra selects still images from a video and uses those images for reconstruction. It does not make a reliable model merely by replaying a clip. The following recipe combines DJI's documented controls with a recommended small-project test before larger processing.

1. Define the required output

   Write down the target area, surfaces that must be visible and intended use. “A scene model for orientation” and “a model for checking a clearance” need different acceptance evidence. Identify the receiving viewer or GIS application before choosing output formats.

2. Create a visible-light project and import the source

   Use Terra's video selection control with the original compatible recording and matching flight data. Keep a record of the project name and storage path. If import fails, capture the error code and use the troubleshooting table below before changing the source.

3. Select one useful segment

   Use the video-trim timeline to choose the start and end. DJI documents one trimmed segment per video. Prefer a continuous, stable section that covers the target from changing positions. Avoid treating take-off, abrupt turns and unrelated transit footage as useful just because they increase clip length.

4. Choose Auto or Manual extraction

   Auto is DJI's recommended starting method: the manual describes removing blurred frames and retaining clearer ones. Manual extraction uses a user-defined frame interval. Use it for a deliberate comparison when needed; there is no documented universal interval that makes every flight reconstruct well.

5. Inspect the extracted images

   Confirm extraction, then review the photos in project management. The documented storage path is image/keyframes/[video name]/ inside the project directory. Check actual sharpness, exposure and viewpoint changes before launching a larger reconstruction. Save the selected segment and extraction settings in the project record.

6. Run a representative pilot and examine the reports

   Process the relevant reconstruction stages for the intended output. Review aerotriangulation, image positions and the resulting geometry. Record failures and settings. Expand to other segments only after the pilot shows that this source can answer the original question.

Change one meaningful variable at a time when comparing runs: the chosen segment, extraction method or reconstruction setting. Keep the source fixed and name the outputs clearly. Otherwise, a cleaner-looking model may be attributed to the wrong change, and another operator will struggle to reproduce it.

04 — Inspect the frames and the camera geometry

Automatic frame selection reduces manual sorting; it cannot invent viewpoints. A stationary pan can show a wide field of view without providing the positional change needed for robust depth. Similarly, repeated frames taken from almost the same position may add processing work without covering a hidden wall or connecting separate parts of the scene.

Check the surfaces that matter, not just the frame count

Open representative extracted images at a useful inspection scale. Look for motion blur, crushed shadows, washed-out surfaces and abrupt differences in zoom or framing. Then follow the subject through the sequence: does it remain visible across changing camera positions, or disappear behind an obstruction? Are the required sides and ground transitions actually recorded?

Feature-poor snow, water, reflective surfaces and repetitive patterns can make matching difficult. Moving vehicles or vegetation may be inconsistent between frames. DJI's general reconstruction guidance discusses these image-quality and texture issues; it does not provide a video-specific threshold that makes such scenes safe for measurement. A large number of extracted images is not a substitute for connected, useful observations.

Use the AT report as a diagnostic, not an accuracy certificate

DJI's aerotriangulation report guidance identifies calibrated-image count, connected components, reprojection error and georeferencing RMSE among the review parameters. Uncalibrated images or multiple components warrant investigation. A coherent scene should not be silently accepted if the target area sits in a disconnected fragment.

Distinguish internal image fitting from external accuracy. A low reprojection error indicates image observations fit the estimated geometry relatively well; it does not independently prove the model's absolute coordinates or elevations. Likewise, georeferencing RMSE against recorded image positions is not the same as a checkpoint survey outside the reconstruction.

Rolling-shutter optimization is a separate Terra capability. Its announcement alongside video extraction does not establish that it is automatically applied to every extracted frame. Do not describe either feature as a cure for blurred footage, absent GPS, missing coverage or an undocumented camera source.

Keep the original files intact

05 — Troubleshoot the failure you actually have

Use the failed step and error code to choose the next action. Repeatedly changing image intervals will not repair a corrupted file, and a GPU upgrade will not supply missing flight data. The following table paraphrases DJI's current reconstruction exception guidance; it is not a complete list of Terra errors.

For support escalation, collect the installed Terra version, edition, error code, failed stage, extraction settings and the smallest original sample that still demonstrates the problem. Include relevant logs and a short account of what you have already tried. Check organizational permissions before sharing footage of people, sensitive assets or an incident scene.

Preserve the failed project and create a separately named retry. An intact record helps distinguish an export failure from a capture limitation, and prevents a later reviewer from mistaking a manually altered source for the original evidence.

06 — Decide between archived video and planned photos

Use video when existing footage is the available source and a pilot shows useful coverage for the intended decision. Plan a dedicated photo mission when you control acquisition and need repeatable geometry, deliberate surface coverage or a specified measurement deliverable. The distinction is acquisition control, not a claim that one mode wins every reconstruction.

A hybrid approach can be useful for evaluation, but do not merge unrelated dates or changed scenes without a reasoned reconstruction plan. An archive model may establish what was visible on one date, while a new mission measures current conditions. Those are different records and should remain distinguishable.

If the deliverable is stockpile volume, use the dedicated stockpile measurement workflow to address coverage, base-surface assumptions and validation. Video import is an input capability, not a replacement for that measurement methodology.

07 — Validate the output, restrict its use or re-fly

Define acceptance before inspecting a polished rendering. For visual orientation, the relevant checks may be scene completeness, recognizable features and clearly labelled blind areas. For an engineering or financial measurement, acceptance needs the project's coordinate requirements, independent reference data and an agreed tolerance. A model that passes the first use may fail the second.

DJI's accuracy-check methodology describes comparisons with higher-accuracy, independent sources and RMSE-based assessment. For a consequential deliverable, agree which features or checkpoints will be checked, how their reference values were measured and how errors will be reported. Do not use a point fitted as control as the sole independent test of the resulting model.

Inspect local geometry as well as summary statistics. A wall can be incomplete, a roof edge distorted or a region weakly connected even when the overall model looks plausible. A single known distance is a useful diagnostic but cannot certify every dimension, surface or coordinate in the scene.

Re-flight should address the failure, not simply repeat the original route. If the issue is missing views, plan additional coverage. If it is motion blur or feature-poor conditions, change the acquisition approach. If the geometry is adequate but reference checks are missing, resolve the validation plan before assuming a new clip will help.

Canadian flight requirements, site access and scene restrictions remain separate from software capability. Processing footage does not authorize another flight or guarantee that a result is acceptable to a regulator, engineer, insurer or court. Obtain the applicable operational permissions and professional review for the project.

08 — Apply it carefully to emergencies and archives

DJI positions video frame extraction for rapid use of video archives, including emergencies where individual photo capture was not practical. The advantage is access to an existing source—not an established processing-time or accuracy guarantee. Run a small feasibility test before promising a complete scene model to another team.

Emergency scene orientation

An illustrative response workflow is to preserve the original recording, build a limited model from qualifying footage and label the visible area and acquisition time. It may help a briefing team understand spatial relationships after initial capture. Keep the original video available so reviewers can inspect objects or motion that a static mesh fails to represent.

Do not replace time-critical observation with a reconstructed scene or describe a rapid model as a certified accident survey. Moving people and vehicles, changed conditions and hidden surfaces limit what the mesh can establish. The receiving team needs to know what was reconstructed, what was not and which decisions remain unsupported.

Historical asset review

For an inspection archive, first ask whether recovering a model would answer a specific question: viewing a recorded facade, orienting a later inspection or reviewing documented site context. If the question instead requires an exact dimension or a change measurement, establish independent validation and comparable dates. Two visually similar models are not automatically a defensible change analysis.

Keep source date, selected segment and exclusions attached to every exported view. These examples are recommended applications to evaluate, not SpeedyDrone field-test results. A failed pilot can still be useful: it identifies whether the archive needs better file retention, a different capture plan or a dedicated future survey.

A reusable project record

09 — Deliver the model with its evidence and limits

A standalone mesh rarely tells the next operator how it was made. Include a short record linking the exported result to its source, processing settings and acceptance decision. The checklist below is a recommended team template, not a DJI-mandated form. Its purpose is to make the work reviewable without relying on the original processor's memory.

DJI documents mesh export choices including OBJ and PLY, with tiled options for other viewers. Choose according to the receiving application and test the actual exported bundle before delivery. Confirm textures, orientation, scale and coordinate interpretation; do not assume one successfully opened preview proves every exported format is usable.

If the next step is model editing, use the Terra-to-Modify cleanup workflow as a separate reference. Cleaning a mesh can improve presentation, but removing an artifact or filling a hole does not create independent evidence of the missing real-world surface. Keep raw and edited results distinguishable.

10 — Match the software and acquisition plan to the job

First establish whether the existing footage is usable. Then decide whether your team needs software access, workstation changes or a new acquisition plan. Buying a higher edition or a different aircraft before identifying the failed gate can leave the original problem unresolved.