OpenSafari FPV Drone Dataset

• OpenSafari is an FPV drone dataset featuring high-dynamic outdoor footage with precise 6-DoF camera trajectories, ideal for long-horizon video synthesis.
• It employs a multi-stage validation pipeline, including geometric and kinematic checks, to ensure robust trajectory estimation and scene coherence.
• The dataset supports rigorous evaluation using metrics like FVD, LPIPS, and trajectory adherence, addressing challenges in controllable video generation.

OpenSafari is a new in-the-wild FPV dataset containing high-dynamic drone videos with verified camera trajectories, introduced in "Captain Safari: A World Engine" (Chou et al., 28 Nov 2025). It is intended for long-horizon, camera-conditioned video generation, trajectory-following evaluation, and multi-view consistency research under aggressive 6-DoF motion and complex outdoor layouts. The dataset pairs 15 s, 720 p, 24 fps clips with camera intrinsics $K$, extrinsics $(R_t,T_t)$, auto-generated captions, and split metadata, and it was constructed through a multi-stage geometric and kinematic validation pipeline (Chou et al., 28 Nov 2025).

1. Provenance and design objective

OpenSafari is derived from in-the-wild first-person-view (FPV) drone footage scraped from AirVuz and public YouTube channels. The selected scenes include urban canyons, parks, fields, building façades, mixed vegetation, and uneven terrain. In the source description, these environments are chosen to expose strong parallax, rapid 6-DoF motion, and complex outdoor layouts (Chou et al., 28 Nov 2025).

This design places OpenSafari within the emerging class of datasets for controllable “world engine” research. The motivating problem is not generic video generation, but the synthesis and evaluation of long, camera-conditioned rollouts in which geometric coherence and adherence to a target path must both be preserved. A plausible implication is that OpenSafari is optimized less for semantic breadth than for the joint stress-testing of camera control, parallax handling, and long-range scene consistency.

2. Data collection and normalization

The data-collection pipeline begins by downloading the highest available resolution for each source URL and discarding clips whose native resolution falls below $720\times1280$. All retained clips are then resampled or padded to 24 fps, subjected to a fixed $16{:}9$ center crop, scaled to 720 p ($1280\times720$), and stripped of black borders or letterbox. Automatic scene-cut detection is applied to isolate single-shot segments, after which the videos are uniformly sliced into fixed-length $T=15$ s clips (Chou et al., 28 Nov 2025).

A subsequent motion-based filter is applied using RAFT optical flow. For each clip, the average flow magnitude is computed, and segments whose mean flow magnitude is below a low threshold are discarded in order to emphasize parallax-rich, high-dynamic sequences. The final output is therefore a large corpus of 15 s, 720 p, 24 fps FPV segments exhibiting aggressive translational and rotational maneuvers under diverse outdoor conditions (Chou et al., 28 Nov 2025).

A common misconception would be to regard OpenSafari as a raw web-video scrape. The stated pipeline indicates otherwise: ingestion, normalization, scene isolation, temporal slicing, and motion filtering are all explicit preprocessing stages. This suggests that the dataset is curated specifically to suppress near-static footage and to privilege clips that are informative for 6-DoF controllability.

3. Geometric and kinematic validation

After preprocessing, each segment undergoes camera-trajectory estimation and a three-stage verification/fix cycle. Trajectory estimation samples frames at 4 fps, detects local features, matches them, and builds a per-clip COLMAP-style SfM model to recover per-frame intrinsics $K$ and extrinsics $(R_t,T_t)$. The description gives SuperPoint and SuperGlue as examples of the local feature detector and matcher, respectively (Chou et al., 28 Nov 2025).

The first verification stage is a database check over SfM statistics. It examines the inlier count $n_{\rm inl}$ and the inlier ratio

$\rho = \frac{n_{\rm inl}}{n_{\rm matches}}.$

Any frame pair whose $(R_t,T_t)$0 falls below $(R_t,T_t)$1 is flagged. The second stage is a geometric check, which recomputes the essential matrix $(R_t,T_t)$2 on each flagged pair and measures the symmetric epipolar error

$(R_t,T_t)$3

Pairs are rejected where $(R_t,T_t)$4 (Chou et al., 28 Nov 2025).

The third stage is a kinematic check. For each time $(R_t,T_t)$5, the translational jump and rotational jump are computed as

$(R_t,T_t)$6

A robust median and MAD are then computed, and any $(R_t,T_t)$7 or $(R_t,T_t)$8 whose scaled deviation exceeds pre-set thresholds is flagged. The pipeline also detects forward-vector flips by testing sign consistency of the camera $(R_t,T_t)$9-axis (Chou et al., 28 Nov 2025).

If flagged errors are sparse, the trajectory is patched locally by linearly interpolating translations,

$720\times1280$0

and by applying SLERP to rotations,

$720\times1280$1

All three checks are then rerun on the interpolated sub-interval. If the patched trajectory passes, the clip is accepted; otherwise, the entire clip is discarded (Chou et al., 28 Nov 2025).

This verification logic is central to OpenSafari’s identity. The dataset is not merely accompanied by estimated trajectories; it is explicitly defined by verified camera trajectories obtained through database, geometric, and kinematic checks, followed by targeted fix and re-validation when possible.

4. Scale, split structure, and motion profile

OpenSafari contains $720\times1280$2 training segments and $720\times1280$3 test segments, each of length 15 s. At 24 fps, this corresponds to approximately $720\times1280$4 million train frames and approximately $720\times1280$5 test frames (Chou et al., 28 Nov 2025).

The motion distribution is described qualitatively rather than by a full histogram. The dataset states that 100% of clips exceed a minimum optical-flow magnitude, with no static scenes. It also reports a wide variety of 6-DoF maneuvers, including large pitched climbs or descents greater than $720\times1280$6, sharp turns greater than $720\times1280$7, lateral strafing and roll, and speeds up to approximately $720\times1280$8. Strong scene parallax arises from weaving around obstacles and varied depths (Chou et al., 28 Nov 2025).

These properties distinguish OpenSafari from datasets dominated by forward ego-motion or mild viewpoint change. A plausible implication is that models evaluated on OpenSafari cannot rely on conservative motion priors without incurring trajectory-following or consistency failures. The inclusion of building façades, uneven terrain, and mixed vegetation further suggests a deliberate attempt to combine structural regularity with geometric irregularity, thereby increasing failure modes for both SfM and generative rollout.

5. Evaluation protocol and metrics

The recommended evaluation suite spans video quality, 3D consistency, and trajectory adherence. For video quality, the dataset recommends Fréchet Video Distance (FVD) and LPIPS. FVD is defined using Gaussian fits to generated and real video-feature embeddings: $720\times1280$9 Lower values are better in the usual interpretation of the metric (Chou et al., 28 Nov 2025).

For 3D consistency, the suite includes Multi-view Epipolar Thresholded Reprojection error ($16{:}9$0) and reconstruction rate. Given $16{:}9$1 point correspondences and reprojections, with per-point error $16{:}9$2, the thresholded reprojection error is

$16{:}9$3

with lower being better. Reconstruction rate is defined as the fraction of generated frames whose features could be registered into a coherent SfM model (Chou et al., 28 Nov 2025).

For trajectory following, the dataset defines per-frame translation error as

$16{:}9$4

and the accuracy curve as

$16{:}9$5

From this, pose-following AUC@$16{:}9$6 is

$16{:}9$7

typically evaluated at $16{:}9$8 cm or 15 cm. The suite also includes cosine similarity of look-directions, computed from the camera forward $16{:}9$9-axis vectors $1280\times720$0 and $1280\times720$1: $1280\times720$2 Higher is better for AUC@$1280\times720$3 and CosSim (Chou et al., 28 Nov 2025).

The benchmark protocol recommends training each method on the 11,481-clip train split, with 5 s rollout described as typical, generating on the 787 test clips using ground-truth poses as conditions, and evaluating with FVD, LPIPS, $1280\times720$4, reconstruction rate, AUC@15, AUC@30, and CosSim. An optional large-scale human preference study is also suggested (Chou et al., 28 Nov 2025).

6. Release structure, baseline usage, and scope

OpenSafari is served as a single downloadable archive and via Git-LFS for video. The release includes /videos/ with train/*.mp4 and test/*.mp4, /trajectories/ with train/*.json and test/*.json, /captions/ with train/*.txt and test/*.txt, a splits.csv, and LICENSE.txt plus README.md (Chou et al., 28 Nov 2025).

Path	Contents	Description
/videos/	train/*.mp4, test/*.mp4	15 s clips at $1280\times720$5, 24 fps
/trajectories/	train/*.json, test/*.json	Per-frame $1280\times720$6, $1280\times720$7, and timestamp at 24 fps or downsampled 4 fps
/captions/	train/*.txt, test/*.txt	Auto-generated natural language descriptions
splits.csv	Clip metadata	Train/test IDs and mean optical-flow magnitude
LICENSE.txt, README.md	Documentation	Dataset summary, citation request, usage guidelines (CC-BY)

The stated use cases are training and evaluating long-horizon, camera-conditioned video diffusion or transformer models; stress-testing 6-DoF controllability under real-world parallax; multi-view consistency research, including geometry forcing and memory-augmented world engines; and downstream embodied-AI simulation or AR/VR world synthesis from user-specified trajectories (Chou et al., 28 Nov 2025).

The baseline model suite lists Geometry Forcing (Wu et al. 2025), Real-CamI2V (Li et al. 2025), and Wan2.2-5B-Control-Camera (Team Wan et al. 2025), alongside possible ablations such as memory-removed variants (Chou et al., 28 Nov 2025). This positions OpenSafari not only as a dataset but as a benchmark specification with a defined evaluation interface, modality package, and comparison set.

An objective clarification of scope is important. OpenSafari is not presented as a universal benchmark for all video generation problems; it is explicitly centered on high-dynamic outdoor FPV drone footage with verified trajectories. Conversely, it is not limited to raw perceptual realism assessment: its core contribution is the combination of visual data and validated camera motion, which supports simultaneous evaluation of appearance quality, 3D consistency, and path adherence. This suggests that OpenSafari is particularly suitable where geometry preservation and aggressive camera control must be balanced rather than treated as separate objectives.