The current default, native_intended, uses MegaSAM poses to measure agreement with requested motion directions. The leaderboard averages per-video details.native_intent_accuracy and multiplies by 100. This is a direction-confidence score, not a metric-scale ground-truth trajectory error. The pose-alignment formulas below describe the separate direct pose-GT path: normalize rotation/translation using [0,15] / [0,0.5], take their geometric mean per video, then average videos. Report error_ref values remain compatibility inputs and are never inverted after averaging. Historical published summaries without matching per-video evidence are identified in the score tooltip and are not described as recomputed measurements.
Camera Error quantifies trajectory fidelity between a camera path recovered from a generated video and annotated ground-truth extrinsics. It measures viewpoint control compliance, not object motion or instruction following.
For the direct pose-GT policy, let {Ttgt}t=1N and {Ttpred}t=1N denote 4×4 camera-to-world matrices from annotations and visual odometry (MegaSAM or DROID-SLAM). If the pose counts differ, the GT trajectory is uniformly resampled to the recovered pose count using linear translation interpolation and quaternion SLERP. WorldAtlas then estimates the full positive-scale similarity transform
with Umeyama alignment. Stationary or collinear paths use corresponding camera orientations to resolve the rotational gauge before solving scale and translation. The aligned prediction is R~t=Q⋆Rtpred and t~t=s⋆Q⋆ttpred+u⋆. Mean translation error is
Et=N1t=1∑Nt~t−ttgt2(2)
Mean rotation error (degrees) uses relative rotation Rtrel=R~t(Rtgt)⊤:
WorldAtlas normalizes each direct-GT component to [0,1] with empirical ranges [0,15°] for rotation and [0,0.5] for translation (lower is better), then combines them by the geometric mean. The default native_intended MegaSAM policy instead evaluates recovered motion against the requested camera/action directions and returns its native direction-confidence score; it does not use the direct-GT reducer above. The metric is official when the selected policy has sufficient camera supervision. It is not evaluated or published on image_dynamic.
Those direct-GT empirical ranges predate sim3_v1 and must be revalidated before publishing corrected direct-GT normalized scores. This does not affect the current default native_intended score surface.
