# Step 2: Setup Data Splits Before iterating on annotations, you need proper data splits. Without them, you’ll contaminate your evaluation and build a model that only looks good on paper. This step uses the **quickstart-groups** dataset (KITTI multimodal data with left/right cameras and point clouds) and creates: - **Test set (15%)** - Frozen. Never used for selection or training. Final evaluation only. - **Validation set (15%)** - For iteration decisions. Used to evaluate between training rounds. - **Golden QA set (5%)** - Small, heavily reviewed. Detects label drift. - **Pool (65%)** - Active learning pool. All new labels come from here. > **Critical:** Splits are created at the **group level** (scene), not sample level. This ensures the same scene stays together across all slices (left, right, pcd), preventing data leakage. ## Load or Create the Dataset We clone `quickstart-groups` to a persistent working dataset. This keeps your annotations separate from the zoo dataset. ## Understand the Grouped Structure The `quickstart-groups` dataset is a **grouped dataset** from KITTI: | Slice | Content | Purpose | |---------|---------------------|----------------------------| | `left` | Left camera images | 2D detection annotation | | `right` | Right camera images | Stereo pair (optional use) | | `pcd` | Point cloud data | 3D cuboid annotation | Each **group** represents one scene/frame with synchronized data across all sensors. ## Create Splits at the Group Level **Why group-level splits?** If we split at the sample level, the same scene could end up in both train and test (just different slices). This causes data leakage - the model “sees” scenes at training time that appear in evaluation. By splitting at the **group level**, we ensure: - All slices from the same scene stay together - No information leaks between splits ## Launch the App Explore your grouped dataset in the App. Notice: - The **group mode** shows synchronized samples - Use the **slice selector** to switch between left, right, and pcd - Filter by split tags to see each partition ## Summary You created four data splits with clear purposes: - Test (frozen), Val (iteration), Golden (QA), Pool (labeling source) - **Splits are at the group level** - same scene = same split across all slices **Artifacts:** - `annotation_tutorial` dataset (persistent clone of quickstart-groups) - Split tags: `split:test`, `split:val`, `split:golden`, `split:pool` - Saved views: `test_set`, `val_set`, `golden_qa`, `pool` - `annotation_status` field for tracking progress **Next:** Step 3 - Smart Sample Selection