GVHMR / 20-clip field evaluation

Pinned commit 6ec3ca3 · job 5222 · 4× NVIDIA GB200 · 86 seconds · research-only evaluation

20/20rendered
4QA pass
5QA fail
0runtime failures
Evaluation boundary. All 20 clips use --static_cam, so this measures single-subject body and pose recovery, not camera trajectory estimation. Double-person and anime clips are deliberate out-of-distribution stress cases. The pinned source has narrow ARM64/Python 3.12 compatibility patches for optional pycolmap, headless import, and legacy chumpy APIs.

real-stage-a-full

Positive control: distant full body on a static stage.

pass
real_single_fullbody single static camera exit 0
Input
In-camera mesh
Global motion
QA

全段主体锁定稳定，整体姿态与手臂变化合理，root基本固定在舞台原位，双脚接地一致；主要局限是远景导致手部与细微重心变化被平滑。

real-stage-a-early

Short-clip stability on the first movement phrase.

pass
real_single_fullbody single static camera exit 0
Input
In-camera mesh
Global motion
QA

远景小人仍能稳定跟踪，躯干朝向、手臂手势和站姿基本吻合；root位置连续，双脚落地稳定，仅手部细节和轻微屈膝不够精确。

real-stage-b-full

Positive control: second distant full-body performer.

pass
real_single_fullbody single static camera exit 0
Input
In-camera mesh
Global motion
QA

人物跟踪贯穿全段，抬手触脸等关键姿态被捕捉，root无明显漂移，双脚持续落地；手臂和手腕细节略粗，但不影响作为低动态全身训练数据使用。

real-stage-b-late

Short-clip stability on the later movement phrase.

pass
real_single_fullbody single static camera exit 0
Input
In-camera mesh
Global motion
QA

后段跟踪稳定，手从脸侧回落及上身姿态变化连贯，root位置稳定、双脚接地正常；仅有远景下手部姿态简化和轻微身体朝向误差。

real-double-raw

Expose the one-track detector choice in a two-person shot.

partial
real_double double static camera exit 0
Input
In-camera mesh
Global motion
QA

三处采样均跟踪前景人物，整体姿态连续；双人场景存在选轨歧义，字幕长期遮挡下肢，root/脚步主要依赖补全，不宜作为高质量全身监督。

real-double-left-crop

Test whether a person-biased crop changes the selected track.

partial
real_double_crop double static camera exit 0
Input
In-camera mesh
Global motion
QA

裁剪后持续锁定左侧前景人物，躯干姿态和前向位移较稳定；画面低清且字幕遮挡腿脚，步态、root 深度和脚部接触可信度有限。

real-podium-opening

Crowd and podium occlusion stress test.

partial
real_multi_occlusion multi static camera exit 0
Input
In-camera mesh
Global motion
QA

主体锁定稳定，头肩和轻微上身动作基本一致；前景手机与讲台遮挡双手及下半身，模型给出的屈臂、root 和双脚接地无法由画面验证，姿态偏模板化。

real-podium-late

Track continuity after a shot change in a crowded scene.

partial
real_multi_occlusion multi static camera exit 0
Input
In-camera mesh
Global motion
QA

人物跟踪连续，上身朝向及双臂开合大致准确，root 基本稳定；讲台长期遮住下身，腿脚和站姿属于无观测补全，手指与手势也较粗，不适合作为全身监督。

real-interview-halfbody

Upper-body reconstruction with legs outside frame.

partial
real_single_partialbody single static camera exit 0
Input
In-camera mesh
Global motion
QA

上半身跟踪良好，头部转向、躯干前倾和手臂手势基本合理；腿脚始终不可见，模型补出的站姿、root 和接地无视觉依据，仅适合上半身姿态数据。

real-pregnant-upperbody

Body-shape and partial-body stress case.

fail
real_single_shape_stress single static camera exit 0
Input
In-camera mesh
Global motion
QA

上半身跟踪尚稳定，手臂位置基本可用；但网格完全丢失孕腹这一关键体型特征，并凭空补全未入镜的腿脚与站立 root，身体形状和下肢监督均不可信。

real-curly-upperbody-a

Upper-body gesture reconstruction.

partial
real_single_partialbody single static camera exit 0
Input
In-camera mesh
Global motion
QA

人物跟踪连续，手靠头发的主要动作和上身朝向基本保留，但双臂、手腕及手部位置偏差较大；root基本锁定，下肢和脚为画外推断，只适合较粗的上半身动作监督。

real-curly-upperbody-b

Later gesture phrase and temporal consistency.

partial
real_single_partialbody single static camera exit 0
Input
In-camera mesh
Global motion
QA

主体跟踪稳定，头肩和双臂姿态基本连贯；下半身完全出框，root 与双脚均为推测，脚部接地和全身运动不可作为可靠监督。

real-selfie-handheld

Handheld camera and partial second-person hard case.

fail
real_handheld_partialbody single_plus_partial static camera exit 0
Input
In-camera mesh
Global motion
QA

人物尚未入镜时网格已持续出现在空路面，存在严重误检和时序目标关联错误；手持镜头运动被混入 root/全局运动，后段近景姿态虽勉强对应，但双脚与位移完全不可用。

real-chestup-interview

Chest-up input with substantial leg hallucination required.

partial
real_single_partialbody single static camera exit 0
Input
In-camera mesh
Global motion
QA

主体跟踪稳定，头肩和静态上身姿态大致合理，但手臂细节带有明显臆测；胸部以下不可见，root、腿和脚全部是稳定但无证据的补全，不适合监督全身或全局运动。

real-movie-opening

Cinematic multi-person shot and camera-motion stress.

partial
real_cinematic_multi multi static camera exit 0
Input
In-camera mesh
Global motion
QA

主体在多人远景到近景切换后仍基本跟住，躯干和抬手姿态可辨；但多人遮挡下目标关联不够可靠，global root 前倾和位移变化偏大，双脚接地与步态缺乏可信对应。

real-movie-late

Later cinematic shot and cut sensitivity.

fail
real_cinematic_single single_or_multi static camera exit 0
Input
In-camera mesh
Global motion
QA

镜头切换后人物外观与构图发生突变，虽能逐帧贴合上半身，但跨帧身份和姿态轨迹不连续；下肢多处出框且后段出现跪姿式错误补全，root/脚部运动不可用。

anime-cropped-a

Stylized cropped character, intentionally out of distribution.

partial
anime_partialbody single static camera exit 0
Input
In-camera mesh
Global motion
QA

人物跟踪基本连续，头身朝向和大幅上肢动作可用，但前伸手的透视与双手位置明显失真；下半身完全由裁切画面外推，root近乎静止且脚部落地可信度低。

anime-cropped-b

Second stylized segment for detector consistency.

fail
anime_partialbody single static camera exit 0
Input
In-camera mesh
Global motion
QA

遮挡和近景切换下跟踪不稳，躯干朝向、手臂与手部姿态多处错配，动作连续性较差；root与双腿均属无依据补全，脚位和支撑关系不可用。

anime-double-dance

Hard anime-domain test with synchronized motion; GVHMR is expected to choose only one track.

fail
anime_double_fullbody double static camera exit 0
Input
In-camera mesh
Global motion
QA

双人交叠时疑似发生选人或身份切换，单帧能抓到部分舞姿，但跨帧姿态与朝向不连续；全局位移漂移明显，脚部滑动、悬空和落地关系均不可信。

synthetic-real-dance

Strong generated-video positive: frontal full-body dance with visible feet.

partial
synthetic_real_single_fullbody single static camera exit 0
Input
In-camera mesh
Global motion
QA

上身与双臂舞姿跟踪较好，腿部交叉动作大体可辨；但global motion出现明显向镜头方向的root漂移，和画面内近乎原地舞动不符，脚部接触也有滑步/悬浮感，仅适合保留姿态监督而不宜直接用于全局轨迹训练。