Run in Google Colab

View source on GitHub

Download notebook

# Step 3: Smart Sample Selection Random sampling wastes labels on redundant near-duplicates. This step uses **diversity-based selection** to pick high-value scenes that cover your data distribution efficiently. We’ll use **ZCore (Zero-Shot Coreset Selection)** to score samples based on: - **Coverage**: How much of the embedding space does this sample represent? - **Redundancy**: How many near-duplicates exist? High ZCore score = valuable for labeling. Low score = redundant, skip it. > **Note:** For grouped datasets, we compute embeddings on the **left camera slice** and select at the **group level** (scene). ## Compute Embeddings on Left Camera Slice For diversity selection, we need embeddings. We compute them on the **left camera images** since that is our primary 2D annotation target. > **Dependencies:** This step requires `torch` and `umap-learn`. Install them if needed: > ```bash > pip install torch torchvision umap-learn > ``` ## ZCore: Zero-Shot Coreset Selection ZCore scores each sample by iteratively: 1. Sampling random points in embedding space 2. Finding the nearest data point (coverage bonus) 3. Penalizing nearby neighbors (redundancy penalty) The result: samples covering unique regions score high; redundant samples score low. ## Select at the Group Level We computed scores on individual samples (left camera), but we need to select **groups** (scenes). Each group includes all slices (left, right, pcd). Selection strategy: Use the ZCore score from the left camera sample to rank groups. In the App: 1. Open the **Embeddings** panel to see the 2D projection 2. Color by `zcore` to see score distribution 3. Filter by `batch:v0` tag to see selected groups 4. Verify high-ZCore samples are spread across clusters (good coverage)  ## Why Diversity Sampling Beats Random | Method | What it does | Result | |------------|---------------------------------|-----------------------------------------------| | **Random** | Picks samples uniformly | Over-samples dense regions, misses rare cases | | **ZCore** | Balances coverage vs redundancy | Maximizes diversity, fewer wasted labels | Research shows diversity-based selection can significantly reduce labeling requirements while maintaining model performance. ## Summary You selected a diverse batch using ZCore: - Computed embeddings on **left camera slice** - Ran ZCore to score coverage vs redundancy - Selected top-scoring **groups** (scenes) - Tagged all slices (left, right, pcd) for annotation **Artifacts:** - `embeddings` field on left camera samples - `zcore` field with selection scores - `batch_v0` saved view (all slices for selected groups) - Tags: `batch:v0`, `to_annotate` **Next:** Step 4 - 2D Annotation + QA