> For the complete documentation index, see [llms.txt](https://jtheta-ai.gitbook.io/docs.jtheta.ai/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://jtheta-ai.gitbook.io/docs.jtheta.ai/domain-specific-workflow/lidar-image-annotation/annotation-platform.md). # Annotation Platform Click **Annotate** to launch the 3D annotation workspace. ## **Editor UI Components** LiDAR Annotation Platform – jtheta.ai This section describes the components and functionality available within the LiDAR annotation editor interface. *** ### **A. Multi-View LiDAR Visualization Panels** The editor provides synchronized spatial views to support accurate 3D annotation. The interface includes: * **Main 3D Point Cloud View (Center)** * **Top View** * **Side View** * **Back View** * **Camera Views (Synchronized)** All views remain synchronized during navigation and editing. Any modification made in one view is reflected consistently across all panels. These views assist annotators in verifying spatial alignment, object dimensions, and orientation. *** ### **B. BEV (Bird’s Eye View) Panel** The BEV panel provides a top-down orthographic representation of the LiDAR scene for precise ground-plane alignment and spatial validation. #### **1. BEV Toggle** * Enables or disables the top-down orthographic view. * Switches between **Perspective (3D)** mode and **BEV** mode. * Maintains object position consistency across view modes.
#### **2. Orthographic View** * Removes perspective distortion. * Ensures uniform spatial scaling across the scene. * Improves accuracy when adjusting: * Length * Width * Orientation (Yaw) This view is recommended for lane-level alignment and spacing validation. #### **3. Interaction Controls** While in BEV mode: * Zoom in/out is available for detailed inspection. * Rotation is restricted to the LiDAR draft it will rotated on X-axis (left and right) only, The cuboid Editing and rotation in the BEV view work normally as in the 3D view. * Unintended panning is minimized to prevent accidental misalignment. #### **4. Annotation Support** The BEV panel supports: * Accurate ground-plane alignment. * Verification of object spacing and overlap. * Precise yaw/orientation correction. *** ### **C. Class, Attributes & Objects Panel (Left Panel)** The left-side panel manages object classification, attributes, and frame-level object control. #### **1. Class List** * Displays all predefined object classes (e.g., Car, Pedestrian, Cyclist). * Determines the category assigned to newly created annotations. * Class definitions are configured at the project level.
#### **2. Attributes Panel** * Allows selection of object-level attributes such as: * State * Visibility * Occlusion * Motion * Attribute fields are configurable according to project schema. * Updates dynamically based on the selected object. * Ensures structured labeling for downstream model training.
**3. Objects List** * Displays all annotated objects within the current frame. * Provides the following actions: * Select * Edit * Hide/Show * Delete * Displays object count per class. #### **4. Status Indicators** Displays the annotation workflow state: * **Submitted** * **In Review** * **Done** *** ### **D. Annotation Tools** The annotation toolbar enables the following operations: * Create 3D bounding boxes (Cuboids) * Modify dimensions (Length, Width, Height) * Rotate along X, Y, Z axes * Reposition objects * Duplicate annotations * Delete annotations All adjustments are reflected immediately in synchronized views. *** ### **E. Interpolation Panel (Across Frames)** The Interpolation Panel enables consistent object tracking across sequential frames. #### **1. Track ID Selection** * Select a specific object using its Track ID. * Interpolation is applied only to the selected Track ID. * Ensures identity consistency across frames. #### **2. Frame Range Selection** * Define start and end frames for interpolation. * The system validates the frame range (Start frame must be less than End frame). * The selected range is visually highlighted in the frame timeline.
#### **3. Automatic Position Interpolation** * Performs linear interpolation of position values (X, Y, Z) between key frames. * Generates smooth object motion across intermediate frames. * Preserves spatial continuity. #### **4. Size & Orientation Interpolation** * Interpolates cuboid dimensions (Width, Height, Depth). * Interpolates yaw/orientation values. * Maintains geometric consistency across frames. #### **5. Data Update & Storage** * Replaces existing annotations for the selected Track ID within the specified frame range. * Updates are stored in local session data. * Changes are reflected immediately in the visualization panels.
*** ### **F. AI Assist Tools** AI Assist provides automated object detection based on predefined classes. * Generates initial bounding box suggestions. * Accelerates large-scale annotation workflows. * All AI-generated annotations must be reviewed before submission. * Suggestions may be accepted, modified, or rejected.
*** ### **G. Ground Snap** #### **1. Overview** Ground Snap automatically aligns selected 3D cuboids to the ground surface using the LiDAR point cloud. Instead of assuming a flat scene, the system detects the local ground level beneath each cuboid for accurate placement on uneven terrain, slopes, and ramps. #### **2. How to Use** 1. Select one or more cuboids. 2. Press G or click the Ground Snap button in the toolbar. 3. The selected cuboid(s) will automatically snap to the detected ground surface. Note: Ground Snap is disabled for submitted frames. #### **3. Key Behaviours** * Local Ground Detection\ Each cuboid independently detects the ground beneath its footprint. * Noise Handling\ Uses robust ground estimation to avoid incorrect placement caused by noisy LiDAR points. * Yaw-Aware Sampling\ Supports rotated cuboids and angled vehicles. * Batch Support\ Multiple cuboids can be snapped in a single operation. * Sparse Point Fallback\ If insufficient ground points are available, the cuboid falls back to Z = 0 and a warning notification is displayed. * No-Op Safety\ No adjustment is made if the cuboid is already aligned with the ground. #### **4. Toast Messages** | Scenario | Message | | ------------------------ | ----------------------------------------------------------------- | | Single cuboid snapped | Ground snapped: Z = X.XXX m | | Sparse ground points | Ground snapped (sparse): Z = X.XXX m | | Multiple cuboids snapped | Ground snapped: N cuboids | | Submitted frame | Already submitted. Drawing, deletion, and rename are not allowed. | #### **5. Why It Matters** Ground Snap reduces manual Z-axis adjustment effort and improves annotation consistency across annotators, especially in complex or uneven environments. *** ### **H.**AI Assist Score Threshold #### **1. Overview** The Score Threshold setting controls the minimum confidence score required for AI-generated detections to appear during AI Assist operations. This helps users balance between: * Detecting more objects (higher recall) * Reducing false positives (higher precision) #### **2. Location** Navigate to: Settings → Score Threshold Adjust the threshold using the slider. #### **3. Default Value** Recommended default: 0.50 #### **4. How It Works** Only predictions with confidence scores greater than or equal to the selected threshold are accepted. Prediction Score >= Threshold → Accepted Prediction Score < Threshold → Ignored #### **5. Example** |

Prediction

|

Confidence

|

Result

| | ------------------- | ------------------- | --------------- | | Car | 0.92 | Accepted | | Pedestrian | 0.71 | Accepted | | Cyclist | 0.43 | Rejected | #### **6. Usage Guidance** * Lower Thresholds\ Show more detections but may increase false positives. * Higher Thresholds\ Reduce noisy predictions but may miss low-confidence objects. #### **7. Why It Matters** The Score Threshold feature allows users to customize AI Assist behavior based on dataset quality, scene complexity, and annotation requirements. *** ### **I. Submitting Annotations** To complete annotation: 1. Click **Submit**. 2. The item status updates as follows: * **In Review** (if reviewers are assigned) * **Done** (if no reviewers are assigned) After submission, the system redirects to the Project Dashboard. *** ### **🔍 Understanding 3D Bounding Box Properties (LiDAR Annotation)** #### 🧱 1. Dimensions (L, W, H) Every 3D cuboid has three size measurements: #### **Length (L)** * Measurement from front to back of the object * Example: the length of a car or truck #### **Width (W)** * Measurement from side to side * Example: door-to-door width of a car #### **Height (H)** * Measurement from bottom to top * Example: height of a pedestrian or vehicle roof #### **How annotators use dimensions** * Resize the cuboid to tightly fit the object in 3D * Represent the real-world physical size accurately * Avoid unrealistically large or small values #### **Example (Typical Sedan):** * Length ≈ 4.5 m * Width ≈ 1.8 m * Height ≈ 1.5 m #### 🎯 2. Rotation (Yaw, Pitch, Roll) Rotation defines how the cuboid is oriented in 3D space. #### **A. Yaw (Rotation around Z-axis)** * Represents left/right turning direction * Most important rotation in automotive LiDAR * Matches vehicle heading Examples: * Car facing north → yaw = 0° * Car facing east → yaw = 90° Yaw may appear in degrees in the UI; internally stored in radians. #### **B. Pitch (Rotation around Y-axis)** * Represents up/down tilt * Relevant for slopes, drones, aerial LiDAR * Usually 0° for automotive ground scenes #### **C. Roll (Rotation around X-axis)** * Represents side tilt * Used for fallen or tilted objects * Typically 0° in road environments
#### **3. Position (X, Y, Z)** Defines the center location of the cuboid in 3D space. * **X →** Left / right distance from LiDAR sensor * **Y →** Forward / backward distance * **Z →** Height above ground Annotators adjust position to: * Correct misalignment * Ensure accurate object centering * Place the cuboid correctly relative to ground #### **4. Axis Alignment (is\_axis\_aligned)** System-generated property indicating whether the cuboid is rotated. * true → yaw = 0° * false → yaw ≠ 0° Why it matters * Helps reviewers identify rotation issues * Supports QA and analytics * Not editable by annotators
| Property | Represents | Important For | Typical Use | | --------------------- | ------------------- | -------------- | ------------------- | | Length, Width, Height | Object size | All cuboids | Accurate fitting | | Yaw | Horizontal rotation | Vehicles | Direction matching | | Pitch | Up/down tilt | Slopes, drones | Rare automotive use | | Roll | Side tilt | Fallen objects | Rare automotive use | | X, Y, Z | Cuboid center | All objects | Precise placement | | is\_axis\_aligned | Rotation flag | QA & analytics | Auto-calculated |