# Definitions

| Term | 中文 | Definition |
| --- | --- | --- |
| Pose | 位姿 | Position and orientation considered together. |
| Position | 位置 | The location of an object, camera, person, or agent in a coordinate frame or scene. |
| Orientation | 朝向 | The facing direction or rotational attitude of an object, camera, person, or agent. |
| Motion | 运动 | Change in position, orientation, state, or trajectory over time. |
| Pose Change | 位姿变化 | A change in pose, usually combining translation and/or rotation. |
| Rotation | 旋转 | Change in orientation around one or more axes. |
| Camera Intrinsics | 相机内参 | Camera parameters internal to the camera model, such as focal length, principal point, and distortion. Camera Intrinsics 焦点焦距、曝光时 长 |
| Camera Extrinsic | 相机外参 | Camera pose relative to the world or another camera, usually represented by rotation and translation. |
| Egocentric | 自我中心 | A reference frame centered on the observer, wearer, camera, or acting agent. |
| Allocentric | 非自我中心 / 环境中心 | A reference frame centered on the world, map, scene, or object rather than the observer. |
| Embodied | 具身 | A protocol where an agent, body, robot, vehicle, or wearer is part of the task setting. Use this as cell context unless the evaluated ability is action or interaction itself. |
| Single-view (SV) | 单视角 | A protocol where the model receives one view, image, camera, or observation point for the task. Keep SV in cells when it qualifies a task variant. |
| Multi-view (MV) | 多视角 | A protocol where the model uses multiple views of the same object, scene, route, or event. Keep MV in cells when it qualifies a task variant. |
| Audio-aware Direction | 音频感知方向 | Inferring direction or angular relation from audio plus spatial context. Keep egocentric or allocentric frame information in matrix cells. |
| Audio-aware Distance | 音频感知距离 | Inferring distance or proximity from audio plus spatial context. Keep egocentric or allocentric frame information in matrix cells. |
| Water Level Test | 水平面判断 | Judging the physically horizontal level of liquid or a reference line despite container tilt or viewpoint changes. |
| Line Orientation Judgement | 线方向判断 | Estimating, matching, or comparing the orientation angle of lines or edges. |
| Selective Attention | 选择性注意 | Selecting spatially relevant targets while ignoring distractors, when the spatial layout is part of the task. |
| Spatial Addition | 空间加和 | Composing multiple spatial relations, offsets, or transformations into a resulting spatial state. |
| Spatial Working Memory | 空间工作记忆 | Maintaining and manipulating locations, layouts, routes, or configurations over a short delay. |
| Spatial Correspondence | 空间对应 | Matching entities, regions, sounds, positions, or observations across views, modalities, or reference frames using spatial alignment. This includes viewpoint-adaptive position/object matching. Mirror correspondence is a special case: identifying which real-world objects correspond to objects visible in a mirror. In `../data/matrix.csv`, write this as `Spatial Correspondence (mirror)`. |
| Constrained Localization | 约束定位 | Localizing a target under explicit spatial constraints such as references, paths, occluders, or multi-hop relations. |
| Counterfactual Physical Deduction | 反事实物理推理 | Inferring how a spatial or physical state would change under a hypothetical condition. |
| Trajectory Prediction | 轨迹预测 | Predicting the future path of an object, camera, person, agent, or vehicle. |
| Video/Dynamic (VD) | 视频/动态 | A compact cell tag for video or dynamic-protocol variants inherited from source sheets. Keep it as a protocol tag when the exact source label is useful. |
| Generation (Gen) | 生成 | A compact cell tag for generation-evaluation variants. |
| VQA | 视觉问答 | A compact cell tag for visual question-answering variants. |
| BEV | 鸟瞰视角 | Bird-eye-view protocol or task variant. |
| FPV | 第一人称视角 | First-person-view protocol or task variant. |
| Paper Task Code | 论文任务代码 | A paper-internal task code such as Q1, SRO.2, DIR-Rec, or SSP. Keep only when it is needed for traceability; otherwise use the short ability phrase in matrix cells. |
| Shape Finding | 形状识别 | Finding or inferring 2D/3D shape structure, including cross-section style tasks. 2D/3D Shape Finding (截面解构) |
| Spatial Folding | 空间折叠 | Reasoning about folding, unfolding, paper folding, cube nets, or other 2D/3D fold transformations. |
| Object Detection | 物体检测 / 定位 | Detecting, recognizing, referring to, or grounding objects in 2D, 3D, image, video, or scene inputs. |
| Projection | 投影 | Reasoning about projection between 3D structures, 2D views, shape projections, cross-sections, or multi-view projections. |
| Spatial Localization | 空间定位 | Locating objects, pieces, positions, UI targets, or constrained targets in a spatial scene or representation. |
| 3D Pose | 三维位姿 | Estimating a static object pose or 3D angular attitude such as pitch, yaw, or roll. 3D Pose 物体静态情况下的 俯仰角/... |
| Absolute Size | 绝对尺寸 | Estimating absolute object, room, or region size, including area, height, length, width, thickness, diameter, and volume. Absolute Area 占地面积，也是 size的一种（比如 vsi测room面积） spatialbench也是 |
| Absolute Direction | 绝对方向 | Estimating direction relative to a camera, agent, or world frame without comparing two objects. Absolute Direction(前后左 右,相对相机)动态 具身 |
| Absolute Orientation | 绝对朝向 | Estimating an object, camera, or agent orientation such as cardinal direction or facing direction. Absolute Orientation 朝向，可以不需要 第二个物体，东南 西北 |
| Acceleration | 加速度 | Estimating or comparing acceleration values or acceleration state. Acceleration 数值 估计 |
| Affordance | 可供性 | Recognizing the function or action possibility of an object, possibly combined with visual attributes. Affordance 识别物体功能/可供 性，再结合视觉属 性回答问题 例：能被骑乘的物 体是什么颜色？ |
| Adjacency | 邻接 | Determining whether two entities, regions, or parts are immediately adjacent or touching. Adjacency(是否紧 贴LEGO) |
| Appearance | 外观 | Visual appearance attributes such as shape, color, texture, side, or visible attribute state. Appearance 外观 细分：Shape / ... |
| Pattern Reasoning | 图案规律推理 | Recognizing or completing visual-spatial patterns, including alignment, style, quantity, attribute, and location-transformation variants. |
| Object Attribute Recognition | 物体属性识别 | Recognizing object attributes such as shape, side, color, material, category, or other semantic-visual properties. |
| Azimuth Transfer | 方位角迁移 | Transferring or tracking azimuth/direction across first-person or changed viewpoints. Azimuth Transfer(强调第一 视角跟踪) |
| Best-view Selection | 最佳视角选择 | Selecting the most informative view or frame for a spatial task such as counting or localization. Best-view Selection(算出 object counting的 同时确定最 informative的 frame) |
| Betweenness Relationship | 中间关系 | Identifying objects or regions located between two references. Betweenness Relationship(识别 出2 reference中间 区域的物体) |
| Camera Direction | 相机方向 | Inferring camera heading or viewing direction; protocol qualifiers such as passive are kept in matrix cells. |
| Center of Mass - Position | 质心位置 | Reasoning about where the center of mass lies relative to a support or reference point. Center of Mass - Position 判断悬挂点相对质 心的位置 |
| Center of Mass - Rotation | 质心旋转 | Predicting rotation direction or stability caused by center-of-mass placement. Center of Mass - Rotation 判断释放后物体绕 悬挂点的旋转方向 |
| Center of Mass | 质心 | Reasoning about the mass-balance point of an object or generated state. |
| Corner/Angle Position | 角点位置 | Identifying important objects or regions at corners, angles, or boundary positions. Corner/Angle Position(重要位置 的物体识别) |
| Distance-Size Correlation | 距离-大小关联 | Reasoning about apparent size changes caused by distance, such as nearer objects appearing larger. Distance-Size Correlation(近大 远小) |
| Layout Spatiotemporal Transition | 布局时空转换 | Tracking or predicting layout changes over time. Layout spatiotemporal transition 布局转换 |
| Lever Equilibrium | 杠杆平衡 | Judging the final equilibrium state of a lever after release. Level Equilibrium - Equilibrium 根据杠杆两侧物体的质量、到支点距 离，判断释放后的最终状态 示例选项：保持水平、左端下沉、右 端下沉、无法判断 |
| Lever Balance Adjustment | 杠杆平衡调整 | Judging how a lever state changes after a mass or distance adjustment. Level Equilibrium - Balance Adjustment 在平衡基础上给一个调整，让模型 判断是否会发生变化 例如：如果左侧物体向支点移动一 格，杠杆会变成什么状态 |
| Newton's First Law - Object Position | 牛顿第一定律-位置 | Predicting whether an object position changes under inertia-based setups. Newton’s First Law - Object Object Position 快速抽走纸张，判断上面的物体会 不会动 本质是测平动惯性 |
| Newton's First Law - Object Stability | 牛顿第一定律-稳定性 | Predicting whether an object remains stable or tips under inertia-based setups. Newton’s First Law - Object Stability 快速抽走纸张，判断上面的圆柱体 会不会倒 本质是测转动惯性 |
| Occlusion-Aware Relative Direction | 遮挡感知相对方向 | Determining relative direction while accounting for occlusion or axis ambiguity. Occlusion-Aware Relative Direction(B在A的 哪里,加上y/z轴的 干扰) |
| Occlusion-Aware Direction Existence | 遮挡感知方向存在性 | Determining whether a direction relation exists under occlusion. Occlusion-Aware Relative Direction Existence(是否存 在) |
| Occlusion-Aware Directional Attribute | 遮挡感知方向属性 | Identifying which object satisfies a directional relation under occlusion. Occlusion-Aware Directional Attribute(谁在A左 边) |
| Counting | 计数 | Counting objects, events, regions, or modality-specific items; relative, depth-aware, grid, occlusion, modality, and rule variants belong in matrix cells. |
| Relative Size | 相对尺寸 | Comparing relative size, area, height, length, width, volume, or scale. Relative Area 相对面积 |
| Relative Distance | 相对距离 | Judging, comparing, localizing, or estimating distance between objects, regions, cameras, or agents. |
| Relative Direction | 相对方向 | Judging, comparing, localizing, or estimating relative direction, orientation, or left/right/front/back relation. |
| Spatial Relations | 空间关系 | Reasoning about spatial relations across object-object, camera-object, camera-region, object-region, and region-region pairs. |
| Route Recall | 路线回忆 | Recalling observed route regions, order, or trajectory from prior observations. Route Recall 回忆视频实际经过 了哪些区域、顺序 是什么，需要记住 已观察到的轨迹 |
| Rotation Difference | 旋转差异 | Recognizing whether two views correspond after rotation or identifying a rotation mismatch. Rotation Diff. (MCA) 同一个场景转了一 个角度，选择题混 入一个非该场景的 旋转图，看能不能 识别出来 |
| Scene Consistency | 场景一致性 | Judging whether a scene type, layout, or state remains consistent over time. Scene Consistency 判断某个对场景类 型/布局的描述是否 和视频里持续一致 |
| Scale Consistency | 尺度一致性 | Judging whether relative object scale remains plausible or consistent. Scale Consistency(大小 尝试e.g.椅子比老 鼠大) |
| Scene Classification | 场景分类 | Classifying a room, environment, or scene type from visual-spatial evidence. Scene Classification 根据视频拍到的环境，判断某个房 间/场景的类别或功能 例：根据它和地下室走廊的连接、 可见设施，包含 large white basin 的房间主要功能是什么？ 选项： mechanical utility room / secondary kitchen / laundry room / fitness center |
| Tool Handedness | 工具手性 | Inferring handedness or orientation from asymmetry. Tool Handedness(可通 过不对称性确认) |
| Maze Navigation | 迷宫导航 | Planning or identifying a path through a static maze or grid-like environment. maze navigation 静态单图迷宫，规划路径 |
| Cube Layer Rotation | 魔方层旋转 | Reasoning about local cube-layer rotation states. *魔方局部层旋转 |
| Geometric Assembly | 几何组合 / 拼装 | Combining, assembling, or arranging 2D/3D shapes, blocks, boards, or geometric parts into a target configuration. |
| Absolute Speed | 绝对速度 | Estimating or comparing absolute speed or velocity values. Velocity 数值估计 |
| Region Recognition | 区域识别 | Recognizing or segmenting functional spatial regions. Region Recognition区域 划分 |
| Vertical Spatial Relationships | 垂直空间关系 | Determining above/below or vertical alignment relations. Vertical Spatial Relationships(是 否是在正上/下方) |
| Viewpoint Count | 视角条件计数 | Counting objects or regions under a specified observed or imagined viewpoint. Keep egocentric, allocentric, or hypothetical viewpoint information in matrix cells. |
| Viewpoint Apparent Size | 视角表观大小 | Judging apparent size as perceived from a specified observed or imagined viewpoint. |
| Viewpoint Direction | 视角条件方向 | Inferring direction or relative direction under a specified observed or imagined viewpoint. |
| Viewpoint Object Order | 视角条件物体顺序 | Determining perceived object order from a specified observed or imagined viewpoint. |
| Viewpoint Perceived Distance | 视角条件感知距离 | Judging perceived distance from a specified observed or imagined viewpoint. |
| Viewpoint-Adaptive Position Matching | 视角自适应位置匹配 | Matching an object or position after a viewpoint change. Viewpoint- Adaptive Position/Position Matching(变换视 角后确定同一个物 体) |
| Center of Mass - Shape | 质心形状一致性 | Evaluating whether generated or predicted object shape remains consistent around the center of mass. Center of Mass - shape 评估生成视频是否 保持物体的形状 指标： Segmentation Mask IoU，重合区 域在Union中的占 比 |
| Center of Mass - Location | 质心位置一致性 | Evaluating center-of-mass location error or displacement. Center of Mass - Location 评估生成视频中物 体的质心位置相对 ground truth是否 偏移 指标：mask几何 中心和ground truth几何中心坐标 之间的欧式距离 |
| Lever Equilibrium Consistency | 杠杆平衡一致性 | Evaluating whether a generated or predicted lever state preserves the expected final equilibrium. |
| Trajectory RMSE | 轨迹均方根误差 | A trajectory metric comparing predicted and ground-truth paths frame by frame. 1. Trajectory RMSE →逐帧比较 Ground Truth 和 pred 的 tracted point →越小越好 |
| Final Position Error | 最终位置误差 | A trajectory metric measuring final predicted position error, often normalized by path length. 2. Final Position Error，最终位置误 差，按GT路径长度 归一化 → 越小越好 |
| Speed Similarity | 速度相似度 | A trajectory metric comparing predicted and ground-truth speed vectors. 3. Speed Similarity:逐帧比 较速度是否一致， 通过比速度向量的 cosine similarity →越大越好 |
| Acceleration Similarity | 加速度相似度 | A trajectory metric comparing predicted and ground-truth acceleration vectors. 4. acceleration similarity,加速度向 量余弦相似度→越 大越好 |
| Directional Consistency | 方向一致性 | A trajectory metric comparing overall displacement direction. 5. Directional Consistency→起 点到终点整体位移 方向是否一致 →越 大越好 |
| Rigid Containment | 刚体容纳 | Planning placement of multiple rigid objects inside containers. |
| Liquid Volume | 液体体积 | Reasoning about liquid volume or relative volume, including embodied variants. Liquid Volume / relative volume（具身） |
| Deformable Object Reasoning | 可变形物体推理 | Inferring structure or hidden contents from deformable objects. Deformable 从可变形的物体推断结构 例如：毛衣下面是苹果还是 杯子 |
| Inclined Plane Reasoning | 斜面推理 | Predicting object motion or stability on slopes. |
| Reflection Authoring | 反射生成 | Distinguishing or generating real objects and their mirror reflections. |
| Mirror Spatial Relations | 镜面空间关系 | Inferring spatial relations across mirror and real-world views. |
| View Hallucination | 视角幻觉 | Detecting objects whose visibility changes critically with viewpoint. |
| Material Transparency | 透明材质推理 | Reasoning about objects seen through transparent materials. |
| Dimensional Size | 维度大小 | Comparing relative sizes across dimensions or vantage points. |
| Spatial Distance | 空间距离 | Comparing relative distances with respect to a reference object. |
| Traversable Passage | 可通行通道 | Identifying navigable corridors or passageways between regions. |
| Regional Boundary | 区域边界 | Identifying or delineating functional spatial region boundaries. |
| Long-Term Navigation | 长期导航 | Planning multi-step navigation toward a distant goal. |
| Counting with Occlusion | 遮挡计数 | Counting objects partially or fully obscured by other scene elements. |
| Spatial Segmentation | 空间分区计数 | Counting or separating objects across distinct spatial regions. |
| Category Ambiguity | 类别歧义 | Counting or identifying visually similar objects that require fine-grained distinction. |
| Illumination Variability | 光照变化 | Reasoning under challenging or non-uniform lighting. |
| Structural Enclosure | 结构封闭 | Reasoning about objects hidden within enclosed or covered structures. |
| Unobserved State Change | 未观测状态变化 | Inferring changes that occur during an unobserved interval. |
| Perception-Action Loop | 感知-动作循环 | An active embodied protocol where agent actions change the viewpoint, observations, or world state. perception-action loop 用 于 embodied/agentic navigation 具身导航，agent选择的移 动能让它真的移动并看到新 的视角，直到agent选择 stop，如果它停在终点，则 成功 |
| Judgement | 判断 | A qualitative decision task, usually selecting or classifying a spatial state rather than estimating a numeric value. ps：ost-bench把 定性和定量写作 judgement和 estimation，可供 参考 |
| Estimation | 估计 | A quantitative or approximate-value task, such as estimating distance, angle, speed, or position. ps：ost-bench把 定性和定量写作 judgement和 estimation，可供 参考 |
| Ego-Exo | 自我-外部视角 | A relation or transition between egocentric and exocentric observations. |
| Crash Prediction | 碰撞预测 | Predicting whether a collision or crash will occur. |
| Distance Description/Identification | 距离描述/识别 | Describing or identifying distance categories or relations. |
| Valid Path Generation | 有效路径生成 | Generating a feasible path from a start state to a goal state. |
| Lane Designation Recognition | 车道指定识别 | Recognizing lane designation or lane-region identity, often under BEV/FPV views. |
| Object Dimensions | 物体尺寸 | Object size attributes such as height, length, width, volume, radius, or diameter. |
| Street-View to Map Alignment | 街景-地图对齐 | Aligning an egocentric street-level view with a north-up or allocentric map. |
| Object/Room Size | 物体/房间大小 | A compact cell phrase for size-related absolute-area or scale tasks involving objects or rooms. |

## Header Notes Moved From `../data/matrix.csv`

从 `../data/matrix.csv` 第一行移出的说明，方便 CSV 表头保持简明。

| Column | Definition |
| --- | --- |
| `2D/3D Shape Finding` | 截面解构。 |
| `3D Pose` | 物体静态情况下的俯仰角等。 |
| `Absolute Area` | 占地面积，也是 size 的一种（比如 VSI 测 room 面积）；SpatialBench 也是。 |
| `Absolute Direction` | 前后左右，相对相机；动态、具身。 |
| `Absolute Orientation` | 朝向，可以不需要第二个物体，东南西北。 |
| `Acceleration` | 数值估计。 |
| `Affordance` | 识别物体功能/可供性，再结合视觉属性回答问题。例：能被骑乘的物体是什么颜色？ |
| `Adjacency` | 是否紧贴 LEGO。 |
| `Appearance` | 外观；细分：Shape / ... |
| `Azimuth Transfer` | 强调第一视角跟踪。 |
| `Best-view Selection` | 在算出 object counting 的同时确定最 informative 的 frame。 |
| `Betweenness Relationship` | 识别出两个 reference 中间区域的物体。 |
| `Camera Intrinsics` | 焦点/焦距、曝光时长。 |
| `Center of Mass - Position` | 判断悬挂点相对质心的位置。 |
| `Center of Mass - Rotation` | 判断释放后物体绕悬挂点的旋转方向。 |
| `Corner/Angle Position` | 重要位置的物体识别。 |
| `Distance-Size Correlation` | 近大远小。 |
| `Layout Spatiotemporal Transition` | 布局转换。 |
| `Level Equilibrium - Equilibrium` | 根据杠杆两侧物体的质量、到支点距离，判断释放后的最终状态。示例选项：保持水平、左端下沉、右端下沉、无法判断。 |
| `Level Equilibrium - Balance Adjustment` | 在平衡基础上给一个调整，让模型判断是否会发生变化。例如：如果左侧物体向支点移动一格，杠杆会变成什么状态。 |
| `Newton’s First Law - Object Position` | 快速抽走纸张，判断上面的物体会不会动；本质是测平动惯性。 |
| `Newton’s First Law - Object Stability` | 快速抽走纸张，判断上面的圆柱体会不会倒；本质是测转动惯性。 |
| `Occlusion-Aware Relative Direction` | B 在 A 的哪里，加上 y/z 轴的干扰。 |
| `Occlusion-Aware Relative Direction Existence` | 是否存在。 |
| `Occlusion-Aware Directional Attribute` | 谁在 A 左边。 |
| `Relative Area` | 相对面积。 |
| `Judgement/Estimation Note` | OST-Bench 把定性和定量写作 judgement 和 estimation，可供参考。 |
| `Route Recall` | 回忆视频实际经过了哪些区域、顺序是什么，需要记住已观察到的轨迹。 |
| `Rotation Diff. (MCA)` | 同一个场景转了一个角度，选择题混入一个非该场景的旋转图，看能不能识别出来。 |
| `Scene Consistency` | 判断某个对场景类型/布局的描述是否和视频里持续一致。 |
| `Scale Consistency` | 大小常识，例如椅子比老鼠大。 |
| `Scene Classification` | 根据视频拍到的环境，判断某个房间/场景的类别或功能。例：根据它和地下室走廊的连接、可见设施，包含 large white basin 的房间主要功能是什么？选项：mechanical utility room / secondary kitchen / laundry room / fitness center。 |
| `Tool Handedness` | 可通过不对称性确认。 |
| `Maze Navigation` | 静态单图迷宫，规划路径。 |
| `Velocity` | 数值估计。 |
| `Region Recognition` | 区域划分。 |
| `Vertical Spatial Relationships` | 是否是在正上/下方。 |
| `Viewpoint-Adaptive Position/Position Matching` | 变换视角后确定同一个物体。 |
| `Center of Mass - Shape` | 评估生成视频是否保持物体的形状。指标：Segmentation Mask IoU，重合区域在 Union 中的占比。 |
| `Center of Mass - Location` | 评估生成视频中物体的质心位置相对 ground truth 是否偏移。指标：mask 几何中心和 ground truth 几何中心坐标之间的欧式距离。 |
| `Trajectory RMSE` | 逐帧比较 Ground Truth 和 pred 的 tracked point；越小越好。 |
| `Final Position Error` | 最终位置误差，按 GT 路径长度归一化；越小越好。 |
| `Speed Similarity` | 逐帧比较速度是否一致，通过比速度向量的 cosine similarity；越大越好。 |
| `Acceleration Similarity` | 加速度向量余弦相似度；越大越好。 |
| `Directional Consistency` | 起点到终点整体位移方向是否一致；越大越好。 |
| `Rigid Containment` | Plan the placement of multiple objects across multiple containers. |
| `Liquid Volume / Relative Volume` | 具身。 |
| `Deformable` | 从可变形的物体推断结构。例如：毛衣下面是苹果还是杯子。 |
| `Inclined Plane` | Predict object motion and stability on slopes. |
| `Reflection Authoring` | Distinguish real objects from mirror reflections. |
| `Spatial Relations` | Infer relations across mirror and real-world views. |
| `Correspondence` | Identify which objects appear in the mirror given the real scene. |
| `View Hallucination` | Detect objects whose visibility changes critically with viewing angle. |
| `Material Transparency` | Reason about objects seen through transparent surfaces. |
| `Dimensional Size` | Compare relative sizes of objects across vantage points. |
| `Spatial Distance` | Compare relative distances with respect to a reference object. |
| `Traversable Passage` | Identify navigable corridors or passageways between regions. |
| `Regional Boundary` | Identify and delineate distinct functional spatial regions. |
| `Long-Term Navigation` | Plan multi-step navigation toward a distant goal. |
| `Counting w/ Occlusion` | Count objects partially obscured by other scene elements. |
| `Spatial Segmentation` | Count objects separated across distinct spatial regions. |
| `Category Ambiguity` | Count visually similar objects requiring fine-grained distinction. |
| `Illumination Variability` | Count objects under challenging or non-uniform lighting. |
| `Structural Enclosure` | Count objects hidden within enclosed or covered spaces. |
| `Unobserved State Change` | Infer scene changes during an unobserved interval. |
| `Perception-Action Loop` | 用于 embodied/agentic navigation：具身导航中，agent 选择的移动能让它真的移动并看到新的视角，直到 agent 选择 stop；如果它停在终点，则成功。 |