# Step 5: 3D Annotation Now we annotate 3D cuboids on the point cloud slice. This step covers: 1. Setting up a 3D annotation schema 2. Using the 3D annotation tools (cuboids, transform controls) 3. Understanding the annotation plane concept 4. Viewing 3D labels projected onto 2D camera images > **Tip:** Complete Step 4 (2D annotation) first. Having 2D labels as reference helps with 3D annotation consistency. ## Define Your 3D Schema For 3D cuboids, we use a subset of KITTI classes - focusing on objects that have clear 3D extent in point clouds. ## 3D Annotation in the App ### Getting to the 3D View 1. Launch the App with your batch 2. Click a sample to open the modal 3. **Select the \`\`pcd\`\` slice** from the slice dropdown 4. The 3D visualizer will load the point cloud ### 3D Navigation - **Rotate**: Left-click and drag - **Pan**: Right-click and drag (or Shift + left-click) - **Zoom**: Scroll wheel - **Preset views**: Press `1`, `2`, `3`, `4` for top/right/front/annotation-plane views ## Creating 3D Cuboids ### Enter Annotate Mode 1. Click the **Annotate** tab (pencil icon) 2. Click **Schema** -> **New Field** -> name it `human_cuboids` 3. Set type to **Detections** and add the classes above ### Understanding the Annotation Plane The **annotation plane** is a virtual surface that determines where your clicks place vertices. By default, it’s the XY plane (ground level). - **Moving the plane**: Reposition to place vertices at different heights - **Why it matters**: Cuboid corners snap to this plane when you click ### Drawing a Cuboid 1. Click the **Cuboid** tool in the left toolbar 2. Click to place the **first corner** on the annotation plane 3. Click to place the **opposite corner** (defines the base rectangle) 4. The cuboid is created with a default height 5. Select a class from the dropdown ### Transform Controls After creating a cuboid, use transform controls to refine it: | Control | What it does | |-----------------|------------------------------------------| | **Translation** | Move along X/Y/Z axes or XY/XZ/YZ planes | | **Rotation** | Rotate around X/Y/Z axes | | **Scaling** | Resize along X/Y/Z axes | Click on a cuboid to select it, then use the transform handles. ## Camera Projections One of FiftyOne’s key 3D features is **camera projections**: ### Point Cloud Projections - Flatten the 3D view to 2D planes (top-down, side views) - Useful for accurate positioning ### 2D Image Projections - See the camera images in the 3D viewer dropdown - Your 3D cuboids are **projected onto the 2D images in real-time** - This helps verify that your 3D labels align with the 2D scene To use camera projections: 1. Look for the **projection dropdown** in the 3D viewer 2. Select a camera (e.g., `left`) 3. See your cuboids rendered on the 2D image > **Note:** Camera projections require camera intrinsics/extrinsics to be defined in the dataset. The KITTI data in quickstart-groups should have these. ## Annotation Guidelines for 3D ### Positioning - Center the cuboid on the point cloud cluster representing the object - The base should touch the ground plane - Include all points belonging to the object ### Orientation - Align the cuboid’s longest axis with the object’s heading direction - For vehicles, the front should point in the driving direction ### Sizing - Tightly fit the cuboid to the point cloud extent - Don’t include points from other objects or ground ### Consistency with 2D - Objects labeled in 2D should also be labeled in 3D (if visible in point cloud) - Use the same class for the same object across both modalities --- ## Fast-Forward Option If you want to skip manual 3D labeling, set `FAST_FORWARD = True` below. ## QA Checks for 3D ## Summary You annotated 3D cuboids on the point cloud slice: - Defined a 3D schema (subset of KITTI classes) - Used the annotation plane and transform controls - Verified alignment using camera projections - Ran QA checks for coverage and cross-modality consistency **Artifacts:** - `human_cuboids` field with 3D cuboid annotations - `annotated_3d:v0` tag on point cloud samples with labels **Key Concept:** The 3D→2D camera projections let you verify that your 3D labels align with the 2D scene. This cross-modal validation is a key differentiator for multimodal annotation workflows. **Next:** Step 6 - Train + Evaluate