From b39ee74e999f2bf9866d94fc9d36e2cad61c8d4f Mon Sep 17 00:00:00 2001
From: SikongJueluo <selfconfusion@gmail.com>
Date: Thu, 12 Mar 2026 12:51:04 +0800
Subject: [PATCH] feat(benchmark): add multi-object retrieval benchmark with
 SAM segmentation

---
 .../tasks/multi_object_retrieval.py           | 356 ++++++++++++++++++
 mini-nav/configs/models.py                    |  11 +
 mini-nav/data_loading/insdet_scenes.py        |  64 ++++
 mini-nav/tests/test_multi_object_retrieval.py | 238 ++++++++++++
 mini-nav/tests/test_sam.py                    | 168 +++++++++
 mini-nav/tests/test_scene_scoring.py          | 121 ++++++
 mini-nav/utils/sam.py                         | 100 +++++
 7 files changed, 1058 insertions(+)
 create mode 100644 mini-nav/benchmarks/tasks/multi_object_retrieval.py
 create mode 100644 mini-nav/data_loading/insdet_scenes.py
 create mode 100644 mini-nav/tests/test_multi_object_retrieval.py
 create mode 100644 mini-nav/tests/test_sam.py
 create mode 100644 mini-nav/tests/test_scene_scoring.py
 create mode 100644 mini-nav/utils/sam.py

diff --git a/mini-nav/benchmarks/tasks/multi_object_retrieval.py b/mini-nav/benchmarks/tasks/multi_object_retrieval.py
new file mode 100644
index 0000000..75fafe5
--- /dev/null
+++ b/mini-nav/benchmarks/tasks/multi_object_retrieval.py
@@ -0,0 +1,356 @@
+"""Multi-object retrieval benchmark task.
+
+This benchmark evaluates retrieval accuracy using multiple objects from a cropped
+scene region. It uses SAM for object segmentation, DINO+Hash pipeline for feature
+extraction, and LanceDB for vector storage with scene-level score aggregation.
+"""
+
+import random
+from typing import Any
+
+import lancedb
+import numpy as np
+import pyarrow as pa
+from benchmarks.base import BaseBenchmarkTask
+from benchmarks.tasks.registry import RegisterTask
+from configs.models import BenchmarkTaskConfig
+from rich.progress import track
+from torch import nn
+from torch.utils.data import DataLoader
+from transformers import BitImageProcessorFast
+from utils.feature_extractor import extract_single_image_feature
+from utils.sam import load_sam_model, segment_image
+from utils.common import get_device
+
+
+def _build_object_schema(vector_dim: int) -> pa.Schema:
+    """Build PyArrow schema for object-level vectors.
+
+    Args:
+        vector_dim: Feature vector dimension.
+
+    Returns:
+        PyArrow schema with id, image_id, object_id, category, and vector fields.
+    """
+    return pa.schema(
+        [
+            pa.field("id", pa.int32()),
+            pa.field("image_id", pa.string()),
+            pa.field("object_id", pa.string()),
+            pa.field("category", pa.string()),
+            pa.field("vector", pa.list_(pa.float32(), vector_dim)),
+        ]
+    )
+
+
+def _compute_scene_score(
+    query_object_ids: list[str],
+    retrieved_results: dict[str, list[tuple[float, str]]],
+    gamma: float,
+) -> dict[str, float]:
+    """Compute scene-level scores using co-occurrence penalty.
+
+    Args:
+        query_object_ids: List of query object IDs.
+        retrieved_results: Dict mapping image_id to list of (distance, object_id) results.
+        gamma: Co-occurrence penalty exponent.
+
+    Returns:
+        Dict mapping image_id to computed scene score.
+    """
+    scene_scores: dict[str, float] = {}
+
+    for image_id, results in retrieved_results.items():
+        # Build a set of retrieved object IDs for this scene
+        retrieved_ids = {obj_id for _, obj_id in results}
+
+        # Count how many query objects are found in this scene
+        matched_count = sum(1 for q_id in query_object_ids if q_id in retrieved_ids)
+
+        if matched_count == 0:
+            scene_scores[image_id] = 0.0
+            continue
+
+        # Sum of best similarities (using distance as similarity: smaller = better)
+        # We use 1/(1+distance) to convert distance to similarity
+        similarities = []
+        for dist, obj_id in results:
+            if obj_id in query_object_ids:
+                sim = 1.0 / (1.0 + dist)
+                similarities.append(sim)
+
+        sum_similarity = sum(similarities) if similarities else 0.0
+
+        # Hit rate: ratio of matched objects
+        hit_rate = matched_count / len(query_object_ids)
+
+        # Final score: sum_similarity * (hit_rate)^gamma
+        score = sum_similarity * (hit_rate ** gamma)
+        scene_scores[image_id] = score
+
+    return scene_scores
+
+
+@RegisterTask("multi-object-retrieval")
+class MultiObjectRetrievalTask(BaseBenchmarkTask):
+    """Multi-object retrieval benchmark task."""
+
+    def __init__(self, **kwargs: Any):
+        """Initialize multi-object retrieval task.
+
+        Args:
+            **kwargs: Configuration parameters from BenchmarkTaskConfig.
+        """
+        # Use config from kwargs or load default config
+        if kwargs:
+            config_dict = kwargs
+        else:
+            config = BenchmarkTaskConfig(type="multi-object-retrieval")
+            config_dict = config.model_dump()
+
+        super().__init__(**config_dict)
+        self.config = BenchmarkTaskConfig(**config_dict)
+
+        # SAM settings from ModelConfig (passed via kwargs or use defaults)
+        self.sam_model = kwargs.get("sam_model", "facebook/sam2.1-hiera-large")
+        self.min_mask_area = kwargs.get("sam_min_mask_area", 32 * 32)
+        self.max_masks_per_image = kwargs.get("sam_max_masks", 5)
+
+        # Lazy-loaded resources
+        self._sam_model = None
+        self._mask_generator = None
+
+    @property
+    def sam_model(self) -> Any:
+        """Lazy-load SAM model."""
+        if self._sam_model is None:
+            self._sam_model, self._mask_generator = load_sam_model(
+                model_name=self.sam_model,
+                device=str(get_device()),
+            )
+        return self._sam_model
+
+    @property
+    def mask_generator(self) -> Any:
+        """Lazy-load mask generator."""
+        if self._mask_generator is None:
+            self._sam_model, self._mask_generator = load_sam_model(
+                model_name=self.sam_model,
+                device=str(get_device()),
+            )
+        return self._mask_generator
+
+    def build_database(
+        self,
+        model: nn.Module,
+        processor: BitImageProcessorFast,
+        train_dataset: Any,
+        table: lancedb.table.Table,
+        batch_size: int,
+    ) -> None:
+        """Build the evaluation database with object-level vectors.
+
+        Args:
+            model: Feature extraction model.
+            processor: Image preprocessor.
+            train_dataset: Training dataset.
+            table: LanceDB table to store features.
+            batch_size: Batch size for DataLoader.
+        """
+        # Infer vector dimension from a sample
+        sample = train_dataset[0]
+        sample_image = sample["image"]
+
+        # Get vector dimension by running a forward pass
+        vector_dim = self._infer_vector_dim(processor, model, sample_image)
+        expected_schema = _build_object_schema(vector_dim)
+
+        # Check schema compatibility
+        if table.schema != expected_schema:
+            raise ValueError(
+                f"Table schema mismatch. Expected: {expected_schema}, "
+                f"Got: {table.schema}"
+            )
+
+        # Build database: segment each image, extract features per object
+        record_id = 0
+        records = []
+
+        for idx in track(range(len(train_dataset)), description="Building object database"):
+            item = train_dataset[idx]
+            image = item["image"]
+            image_id = item.get("image_id", f"image_{idx}")
+
+            # Segment image using SAM
+            masks = segment_image(
+                self.mask_generator,
+                image,
+                min_area=self.min_mask_area,
+                max_masks=self.max_masks_per_image,
+            )
+
+            if not masks:
+                continue
+
+            # Extract features for each mask
+            for mask_idx, mask_info in enumerate(masks):
+                # Extract masked region
+                masked_image = self._apply_mask(image, mask_info["segment"])
+
+                # Extract feature vector
+                vector = extract_single_image_feature(processor, model, masked_image)
+
+                # Create object ID
+                object_id = f"{image_id}_obj_{mask_idx}"
+                category = mask_info.get("category", "unknown")
+
+                records.append({
+                    "id": record_id,
+                    "image_id": image_id,
+                    "object_id": object_id,
+                    "category": category,
+                    "vector": vector,
+                })
+                record_id += 1
+
+        # Add all records to table
+        if records:
+            table.add(records)
+
+    def evaluate(
+        self,
+        model: nn.Module,
+        processor: BitImageProcessorFast,
+        test_dataset: Any,
+        table: lancedb.table.Table,
+        batch_size: int,
+    ) -> dict[str, Any]:
+        """Evaluate the model on the test dataset.
+
+        Args:
+            model: Feature extraction model.
+            processor: Image preprocessor.
+            test_dataset: Test dataset.
+            table: LanceDB table to search against.
+            batch_size: Batch size for DataLoader.
+
+        Returns:
+            Dictionary containing evaluation results with keys:
+                - accuracy: Recall@K accuracy (0.0 ~ 1.0)
+                - correct: Number of correct predictions
+                - total: Total number of test samples
+                - top_k: The K value used
+        """
+        top_k = self.config.top_k_per_object
+
+        correct = 0
+        total = 0
+
+        for idx in track(range(len(test_dataset)), description=f"Evaluating Recall@{top_k}"):
+            item = test_dataset[idx]
+            image = item["image"]
+            target_image_id = item.get("image_id", f"image_{idx}")
+
+            # Segment query image
+            masks = segment_image(
+                self.mask_generator,
+                image,
+                min_area=self.min_mask_area,
+                max_masks=self.max_masks_per_image,
+            )
+
+            if not masks:
+                continue
+
+            # Randomly sample query objects
+            num_query = min(self.config.num_query_objects, len(masks))
+            query_masks = random.sample(masks, num_query)
+
+            # Extract features and search for each query object
+            retrieved_results: dict[str, list[tuple[float, str]]] = {}
+
+            for mask_info in query_masks:
+                # Extract masked region
+                masked_image = self._apply_mask(image, mask_info["segment"])
+
+                # Extract feature vector
+                vector = extract_single_image_feature(processor, model, masked_image)
+
+                # Search in LanceDB
+                results = (
+                    table.search(vector)
+                    .select(["image_id", "object_id", "_distance"])
+                    .limit(top_k)
+                    .to_polars()
+                )
+
+                # Aggregate results by scene
+                for row in results.iter_rows():
+                    image_id = row["image_id"]
+                    object_id = row["object_id"]
+                    distance = row["_distance"]
+
+                    if image_id not in retrieved_results:
+                        retrieved_results[image_id] = []
+                    retrieved_results[image_id].append((distance, object_id))
+
+            # Compute scene scores
+            query_object_ids = [m.get("object_id", f"query_obj_{i}") for i, m in enumerate(query_masks)]
+            scene_scores = _compute_scene_score(
+                query_object_ids,
+                retrieved_results,
+                self.config.gamma,
+            )
+
+            # Rank scenes by score
+            ranked_scenes = sorted(scene_scores.items(), key=lambda x: x[1], reverse=True)
+
+            # Check if target is in top-K
+            top_k_scenes = [scene_id for scene_id, _ in ranked_scenes[:top_k]]
+            if target_image_id in top_k_scenes:
+                correct += 1
+            total += 1
+
+        accuracy = correct / total if total > 0 else 0.0
+
+        return {
+            "accuracy": accuracy,
+            "correct": correct,
+            "total": total,
+            "top_k": top_k,
+        }
+
+    def _infer_vector_dim(
+        self,
+        processor: BitImageProcessorFast,
+        model: nn.Module,
+        sample_image: Any,
+    ) -> int:
+        """Infer vector dimension from model output."""
+        vector = extract_single_image_feature(processor, model, sample_image)
+        return len(vector)
+
+    def _apply_mask(self, image: Any, mask: np.ndarray) -> Any:
+        """Apply mask to image and return masked image.
+
+        Args:
+            image: PIL Image.
+            mask: Binary mask as numpy array.
+
+        Returns:
+            Masked PIL Image.
+        """
+        import numpy as np
+        from PIL import Image
+
+        image_np = np.array(image.convert("RGB"))
+        # Ensure mask is the right shape
+        if mask.shape != image_np.shape[:2]:
+            from skimage.transform import resize
+            mask_resized = resize(mask, image_np.shape[:2], order=0, anti_aliasing=False)
+        else:
+            mask_resized = mask
+
+        # Apply mask
+        masked_np = image_np * mask_resized[:, :, np.newaxis]
+        return Image.fromarray(masked_np.astype(np.uint8))
diff --git a/mini-nav/configs/models.py b/mini-nav/configs/models.py
index b3e244c..7109178 100644
--- a/mini-nav/configs/models.py
+++ b/mini-nav/configs/models.py
@@ -118,6 +118,17 @@ class BenchmarkTaskConfig(BaseModel):
     type: str = Field(default="retrieval", description="Task type")
     top_k: int = Field(default=10, gt=0, description="Top K for recall evaluation")
 
+    # Multi-object retrieval specific settings
+    gamma: float = Field(
+        default=1.0, ge=0, description="Co-occurrence penalty exponent"
+    )
+    top_k_per_object: int = Field(
+        default=50, gt=0, description="Top K results per object query"
+    )
+    num_query_objects: int = Field(
+        default=3, gt=0, description="Number of objects to sample from query image"
+    )
+
 
 class BenchmarkConfig(BaseModel):
     """Configuration for benchmark evaluation."""
diff --git a/mini-nav/data_loading/insdet_scenes.py b/mini-nav/data_loading/insdet_scenes.py
new file mode 100644
index 0000000..e335857
--- /dev/null
+++ b/mini-nav/data_loading/insdet_scenes.py
@@ -0,0 +1,64 @@
+"""InsDet Scenes dataset for multi-object retrieval benchmark."""
+
+from pathlib import Path
+from typing import Any
+
+from benchmarks.base import BaseDataset
+from data_loading.loader import load_val_dataset
+
+
+class InsDetScenesDataset(BaseDataset):
+    """InsDet-FULL/Scenes dataset with easy/hard splits.
+
+    This dataset provides scene images with object annotations from the
+    Instance Detection (InsDet) dataset, supporting easy and hard splits.
+    """
+
+    def __init__(
+        self,
+        scenes_dir: Path | str,
+        split: str = "easy",
+    ):
+        """Initialize InsDet Scenes dataset.
+
+        Args:
+            scenes_dir: Path to the InsDet-FULL/Scenes directory.
+            split: Scene split to use ('easy' or 'hard').
+        """
+        self.scenes_dir = Path(scenes_dir)
+        self.split = split
+        self._dataset = load_val_dataset(self.scenes_dir, split)
+
+    def get_train_split(self) -> Any:
+        """Get training split (same as test for this dataset).
+
+        Returns:
+            HuggingFace Dataset for training.
+        """
+        return self._dataset
+
+    def get_test_split(self) -> Any:
+        """Get test/evaluation split.
+
+        Returns:
+            HuggingFace Dataset for testing.
+        """
+        return self._dataset
+
+    def __len__(self) -> int:
+        """Get dataset length."""
+        return len(self._dataset)
+
+    def __getitem__(self, idx: int) -> dict[str, Any]:
+        """Get a single item from the dataset.
+
+        Args:
+            idx: Index of the item.
+
+        Returns:
+            Dictionary containing:
+                - image: PIL Image
+                - image_id: Scene identifier
+                - objects: dict with bbox, category, area, id
+        """
+        return self._dataset[idx]
diff --git a/mini-nav/tests/test_multi_object_retrieval.py b/mini-nav/tests/test_multi_object_retrieval.py
new file mode 100644
index 0000000..7d71ac2
--- /dev/null
+++ b/mini-nav/tests/test_multi_object_retrieval.py
@@ -0,0 +1,238 @@
+"""Integration tests for multi-object retrieval benchmark pipeline.
+
+These tests verify the end-to-end functionality of the multi-object retrieval
+benchmark, including schema building, database population, and evaluation.
+"""
+
+import numpy as np
+import pytest
+from unittest.mock import Mock, patch, MagicMock
+from PIL import Image
+
+
+class TestMultiObjectRetrievalIntegration:
+    """Integration tests for multi-object retrieval benchmark."""
+
+    @pytest.fixture
+    def mock_model_processor(self):
+        """Create mock model and processor."""
+        mock_model = Mock()
+        mock_processor = Mock()
+
+        # Mock the feature extraction to return a fixed-size vector
+        def mock_extract(processor, model, image):
+            return [0.1] * 256  # 256-dim vector
+        mock_processor.images = mock_extract
+
+        return mock_model, mock_processor
+
+    @pytest.fixture
+    def mock_dataset(self):
+        """Create a mock dataset with images and annotations."""
+        # Create mock items
+        items = []
+        for i in range(3):
+            item = {
+                "image": Image.new("RGB", (224, 224), color=(i * 50, 100, 150)),
+                "image_id": f"scene_{i}",
+                "objects": {
+                    "bbox": [[10, 10, 50, 50], [60, 60, 40, 40]],
+                    "category": ["object_a", "object_b"],
+                    "area": [2500, 1600],
+                    "id": [0, 1],
+                },
+            }
+            items.append(item)
+
+        mock_dataset = Mock()
+        mock_dataset.__len__ = Mock(return_value=len(items))
+        mock_dataset.__getitem__ = lambda self, idx: items[idx]
+        mock_dataset.with_format = lambda fmt: mock_dataset
+
+        return mock_dataset
+
+    def test_build_object_schema(self):
+        """Test that object schema is built correctly."""
+        from benchmarks.tasks.multi_object_retrieval import _build_object_schema
+        import pyarrow as pa
+
+        vector_dim = 256
+        schema = _build_object_schema(vector_dim)
+
+        assert isinstance(schema, pa.Schema)
+        assert "id" in schema.names
+        assert "image_id" in schema.names
+        assert "object_id" in schema.names
+        assert "category" in schema.names
+        assert "vector" in schema.names
+
+        # Check vector field has correct dimension
+        vector_field = schema.field("vector")
+        assert isinstance(vector_field.type, pa.List)
+        assert vector_field.type.value_type == pa.float32()
+
+    @patch("benchmarks.tasks.multi_object_retrieval.load_sam_model")
+    @patch("benchmarks.tasks.multi_object_retrieval.segment_image")
+    def test_build_database_with_mocked_sam(
+        self,
+        mock_segment,
+        mock_load_sam,
+        mock_model_processor,
+        mock_dataset,
+    ):
+        """Test database building with mocked SAM segmentation."""
+        from benchmarks.tasks.multi_object_retrieval import (
+            MultiObjectRetrievalTask,
+            _build_object_schema,
+        )
+
+        mock_model, mock_processor = mock_model_processor
+
+        # Mock SAM
+        mock_load_sam.return_value = (Mock(), Mock())
+        mock_segment.return_value = [
+            {
+                "segment": np.ones((224, 224), dtype=bool),
+                "area": 50000,
+                "bbox": [0, 0, 224, 224],
+            }
+        ]
+
+        # Create task with config
+        task = MultiObjectRetrievalTask(
+            sam_model="facebook/sam2.1-hiera-large",
+            min_mask_area=1024,
+            max_masks_per_image=5,
+            gamma=1.0,
+            top_k_per_object=50,
+            num_query_objects=3,
+        )
+
+        # Create mock table
+        mock_table = Mock()
+        mock_table.schema = _build_object_schema(256)
+
+        # Build database (this should not raise)
+        task.build_database(mock_model, mock_processor, mock_dataset, mock_table, batch_size=1)
+
+        # Verify table.add was called
+        assert mock_table.add.called
+
+    @patch("benchmarks.tasks.multi_object_retrieval.load_sam_model")
+    @patch("benchmarks.tasks.multi_object_retrieval.segment_image")
+    def test_evaluate_with_mocked_sam(
+        self,
+        mock_segment,
+        mock_load_sam,
+        mock_model_processor,
+        mock_dataset,
+    ):
+        """Test evaluation with mocked SAM segmentation."""
+        from benchmarks.tasks.multi_object_retrieval import (
+            MultiObjectRetrievalTask,
+            _build_object_schema,
+        )
+
+        mock_model, mock_processor = mock_model_processor
+
+        # Mock SAM
+        mock_load_sam.return_value = (Mock(), Mock())
+        mock_segment.return_value = [
+            {
+                "segment": np.ones((224, 224), dtype=bool),
+                "area": 50000,
+                "bbox": [0, 0, 224, 224],
+                "object_id": "query_obj_0",
+            }
+        ]
+
+        # Create mock table with search results
+        mock_table = Mock()
+        mock_table.schema = _build_object_schema(256)
+
+        # Mock search to return matching result
+        mock_result = Mock()
+        mock_result.to_polars.return_value = {
+            "image_id": ["scene_0"],
+            "object_id": ["scene_0_obj_0"],
+            "_distance": [0.1],
+        }
+
+        mock_table.search.return_value.select.return_value.limit.return_value = mock_result
+
+        # Create task
+        task = MultiObjectRetrievalTask(
+            sam_model="facebook/sam2.1-hiera-large",
+            min_mask_area=1024,
+            max_masks_per_image=5,
+            gamma=1.0,
+            top_k_per_object=50,
+            num_query_objects=1,
+        )
+
+        # Evaluate
+        results = task.evaluate(mock_model, mock_processor, mock_dataset, mock_table, batch_size=1)
+
+        # Verify results structure
+        assert "accuracy" in results
+        assert "correct" in results
+        assert "total" in results
+        assert "top_k" in results
+        assert results["top_k"] == 50
+
+    def test_task_initialization_with_config(self):
+        """Test task initialization with custom config."""
+        from benchmarks.tasks.multi_object_retrieval import MultiObjectRetrievalTask
+
+        task = MultiObjectRetrievalTask(
+            sam_model="facebook/sam2.1-hiera-small",
+            min_mask_area=500,
+            max_masks_per_image=3,
+            gamma=0.5,
+            top_k_per_object=100,
+            num_query_objects=5,
+        )
+
+        assert task.sam_model == "facebook/sam2.1-hiera-small"
+        assert task.min_mask_area == 500
+        assert task.max_masks_per_image == 3
+        assert task.config.gamma == 0.5
+        assert task.config.top_k_per_object == 100
+        assert task.config.num_query_objects == 5
+
+    def test_task_initialization_defaults(self):
+        """Test task initialization with default config."""
+        from benchmarks.tasks.multi_object_retrieval import MultiObjectRetrievalTask
+
+        task = MultiObjectRetrievalTask()
+
+        # Check defaults from BenchmarkTaskConfig
+        assert task.config.gamma == 1.0
+        assert task.config.top_k_per_object == 50
+        assert task.config.num_query_objects == 3
+        # SAM settings from ModelConfig defaults
+        assert task.sam_model == "facebook/sam2.1-hiera-large"
+        assert task.min_mask_area == 1024
+        assert task.max_masks_per_image == 5
+
+
+class TestInsDetScenesDataset:
+    """Tests for InsDetScenesDataset class."""
+
+    def test_dataset_class_exists(self):
+        """Test that InsDetScenesDataset can be imported."""
+        from data_loading.insdet_scenes import InsDetScenesDataset
+
+        assert InsDetScenesDataset is not None
+
+    @patch("data_loading.insdet_scenes.load_val_dataset")
+    def test_dataset_loads_correct_split(self, mock_load):
+        """Test dataset loads correct split."""
+        from data_loading.insdet_scenes import InsDetScenesDataset
+
+        mock_load.return_value = Mock()
+
+        dataset = InsDetScenesDataset("/path/to/scenes", split="easy")
+
+        mock_load.assert_called_once_with("/path/to/scenes", "easy")
+        assert dataset.split == "easy"
diff --git a/mini-nav/tests/test_sam.py b/mini-nav/tests/test_sam.py
new file mode 100644
index 0000000..3d9ba2c
--- /dev/null
+++ b/mini-nav/tests/test_sam.py
@@ -0,0 +1,168 @@
+"""Tests for SAM segmentation utilities.
+
+Note: These tests mock the SAM model loading since SAM requires
+heavy model weights. The actual SAM integration should be tested
+separately in integration tests.
+"""
+
+import numpy as np
+import pytest
+from unittest.mock import Mock, patch
+from PIL import Image
+
+
+class TestSAMSegmentation:
+    """Test suite for SAM segmentation utilities."""
+
+    def test_segment_image_empty_masks(self):
+        """Test segment_image returns empty list when no masks generated."""
+        from utils.sam import segment_image
+
+        # Create mock mask generator that returns empty list
+        mock_generator = Mock()
+        mock_generator.generate.return_value = []
+
+        result = segment_image(mock_generator, Image.new("RGB", (100, 100)))
+
+        assert result == []
+
+    def test_segment_image_filters_small_masks(self):
+        """Test segment_image filters masks below min_area threshold."""
+        from utils.sam import segment_image
+
+        # Create mock masks with different areas
+        small_mask = {
+            "segment": np.zeros((10, 10), dtype=bool),
+            "area": 50,  # Below 32*32 = 1024
+            "bbox": [0, 0, 10, 10],
+            "predicted_iou": 0.9,
+            "stability_score": 0.8,
+        }
+        large_mask = {
+            "segment": np.ones((100, 100), dtype=bool),
+            "area": 10000,  # Above threshold
+            "bbox": [0, 0, 100, 100],
+            "predicted_iou": 0.95,
+            "stability_score": 0.9,
+        }
+
+        mock_generator = Mock()
+        mock_generator.generate.return_value = [small_mask, large_mask]
+
+        result = segment_image(
+            mock_generator,
+            Image.new("RGB", (100, 100)),
+            min_area=32 * 32,
+            max_masks=5,
+        )
+
+        # Should only return the large mask
+        assert len(result) == 1
+        assert result[0]["area"] == 10000
+
+    def test_segment_image_limits_max_masks(self):
+        """Test segment_image limits to max_masks largest masks."""
+        from utils.sam import segment_image
+
+        # Create 10 masks with different areas
+        masks = [
+            {
+                "segment": np.ones((i + 1, i + 1), dtype=bool),
+                "area": (i + 1) * (i + 1),
+                "bbox": [0, 0, i + 1, i + 1],
+                "predicted_iou": 0.9,
+                "stability_score": 0.8,
+            }
+            for i in range(10)
+        ]
+
+        mock_generator = Mock()
+        mock_generator.generate.return_value = masks
+
+        result = segment_image(
+            mock_generator,
+            Image.new("RGB", (100, 100)),
+            min_area=1,
+            max_masks=3,
+        )
+
+        # Should only return top 3 largest masks
+        assert len(result) == 3
+        # Check they are sorted by area (largest first)
+        areas = [m["area"] for m in result]
+        assert areas == sorted(areas, reverse=True)
+
+    def test_segment_image_sorted_by_area(self):
+        """Test segment_image returns masks sorted by area descending."""
+        from utils.sam import segment_image
+
+        # Create masks with known areas (unordered)
+        mask1 = {"segment": np.ones((5, 5), dtype=bool), "area": 25, "bbox": [0, 0, 5, 5]}
+        mask2 = {"segment": np.ones((10, 10), dtype=bool), "area": 100, "bbox": [0, 0, 10, 10]}
+        mask3 = {"segment": np.ones((3, 3), dtype=bool), "area": 9, "bbox": [0, 0, 3, 3]}
+
+        mock_generator = Mock()
+        mock_generator.generate.return_value = [mask1, mask2, mask3]
+
+        result = segment_image(
+            mock_generator,
+            Image.new("RGB", (100, 100)),
+            min_area=1,
+            max_masks=10,
+        )
+
+        # Should be sorted by area descending
+        assert result[0]["area"] == 100
+        assert result[1]["area"] == 25
+        assert result[2]["area"] == 9
+
+
+class TestExtractMaskedRegion:
+    """Test suite for extracting masked regions from images."""
+
+    def test_extract_masked_region_binary(self):
+        """Test extracting masked region with binary mask."""
+        from utils.sam import extract_masked_region
+
+        # Create a simple image
+        image = Image.new("RGB", (10, 10), color=(255, 0, 0))
+
+        # Create a binary mask (half kept, half masked)
+        mask = np.zeros((10, 10), dtype=bool)
+        mask[:, :5] = True
+
+        result = extract_masked_region(image, mask)
+
+        # Check that left half is red, right half is black
+        result_np = np.array(result)
+        left_half = result_np[:, :5, :]
+        right_half = result_np[:, 5:, :]
+
+        assert np.all(left_half == [255, 0, 0])
+        assert np.all(right_half == [0, 0, 0])
+
+    def test_extract_masked_region_all_masked(self):
+        """Test extracting when entire image is masked."""
+        from utils.sam import extract_masked_region
+
+        image = Image.new("RGB", (10, 10), color=(255, 0, 0))
+        mask = np.ones((10, 10), dtype=bool)
+
+        result = extract_masked_region(image, mask)
+        result_np = np.array(result)
+
+        # Entire image should be preserved
+        assert np.all(result_np == [255, 0, 0])
+
+    def test_extract_masked_region_all_zero_mask(self):
+        """Test extracting when mask is all zeros."""
+        from utils.sam import extract_masked_region
+
+        image = Image.new("RGB", (10, 10), color=(255, 0, 0))
+        mask = np.zeros((10, 10), dtype=bool)
+
+        result = extract_masked_region(image, mask)
+        result_np = np.array(result)
+
+        # Entire image should be black
+        assert np.all(result_np == [0, 0, 0])
diff --git a/mini-nav/tests/test_scene_scoring.py b/mini-nav/tests/test_scene_scoring.py
new file mode 100644
index 0000000..de12468
--- /dev/null
+++ b/mini-nav/tests/test_scene_scoring.py
@@ -0,0 +1,121 @@
+"""Tests for scene scoring algorithm in multi-object retrieval."""
+
+import pytest
+from benchmarks.tasks.multi_object_retrieval import _compute_scene_score
+
+
+class TestSceneScoringAlgorithm:
+    """Test suite for scene scoring with co-occurrence penalty."""
+
+    def test_scene_score_basic(self):
+        """Test basic scene scoring with single match."""
+        query_object_ids = ["obj_1", "obj_2", "obj_3"]
+
+        # Scene A has obj_1
+        retrieved_results = {
+            "scene_A": [("distance_1", "obj_1")],
+        }
+
+        scores = _compute_scene_score(query_object_ids, retrieved_results, gamma=1.0)
+
+        # Hit rate = 1/3, similarity = 1/(1+distance_1)
+        assert "scene_A" in scores
+        assert scores["scene_A"] > 0
+
+    def test_scene_score_no_match(self):
+        """Test scene scoring when no objects match."""
+        query_object_ids = ["obj_1", "obj_2", "obj_3"]
+
+        retrieved_results = {
+            "scene_A": [("distance_1", "other_obj")],
+        }
+
+        scores = _compute_scene_score(query_object_ids, retrieved_results, gamma=1.0)
+
+        assert scores["scene_A"] == 0.0
+
+    def test_scene_score_multiple_scenes(self):
+        """Test scoring across multiple scenes."""
+        query_object_ids = ["obj_1", "obj_2"]
+
+        retrieved_results = {
+            "scene_A": [("0.1", "obj_1")],
+            "scene_B": [("0.1", "obj_2")],
+            "scene_C": [("0.1", "other")],
+        }
+
+        scores = _compute_scene_score(query_object_ids, retrieved_results, gamma=1.0)
+
+        # Scenes with matches should have positive scores
+        assert scores["scene_A"] > 0
+        assert scores["scene_B"] > 0
+        # Scene C has no match, score should be 0
+        assert scores["scene_C"] == 0.0
+
+    def test_scene_score_gamma_zero(self):
+        """Test scoring with gamma=0 (no penalty)."""
+        query_object_ids = ["obj_1", "obj_2", "obj_3", "obj_4", "obj_5"]
+
+        retrieved_results = {
+            "scene_A": [("0.1", "obj_1")],
+        }
+
+        scores_gamma_0 = _compute_scene_score(query_object_ids, retrieved_results, gamma=0.0)
+        scores_gamma_1 = _compute_scene_score(query_object_ids, retrieved_results, gamma=1.0)
+
+        # With gamma=0, hit_rate^0 = 1, so score = similarity
+        # With gamma=1, hit_rate^1 = 1/5, so score = similarity * 1/5
+        # scores_gamma_0 should be larger
+        assert scores_gamma_0["scene_A"] > scores_gamma_1["scene_A"]
+
+    def test_scene_score_multiple_matches(self):
+        """Test scoring when scene has multiple matching objects."""
+        query_object_ids = ["obj_1", "obj_2"]
+
+        retrieved_results = {
+            "scene_A": [("0.1", "obj_1"), ("0.2", "obj_2")],
+        }
+
+        scores = _compute_scene_score(query_object_ids, retrieved_results, gamma=1.0)
+
+        # Both objects match, hit_rate = 2/2 = 1.0
+        # Score = (1/(1+0.1) + 1/(1+0.2)) * 1.0
+        expected_similarity = 1.0 / 1.1 + 1.0 / 1.2
+        assert abs(scores["scene_A"] - expected_similarity) < 0.01
+
+    def test_scene_score_distance_to_similarity(self):
+        """Test that smaller distance yields higher score."""
+        query_object_ids = ["obj_1"]
+
+        retrieved_results = {
+            "scene_close": [("0.01", "obj_1")],
+            "scene_far": [("10.0", "obj_1")],
+        }
+
+        scores = _compute_scene_score(query_object_ids, retrieved_results, gamma=1.0)
+
+        # Closer scene should have higher score
+        assert scores["scene_close"] > scores["scene_far"]
+
+    def test_scene_score_empty_results(self):
+        """Test scoring with empty retrieved results."""
+        query_object_ids = ["obj_1", "obj_2"]
+
+        retrieved_results = {}
+
+        scores = _compute_scene_score(query_object_ids, retrieved_results, gamma=1.0)
+
+        assert scores == {}
+
+    def test_scene_score_empty_query(self):
+        """Test scoring with empty query objects."""
+        query_object_ids = []
+
+        retrieved_results = {
+            "scene_A": [("0.1", "obj_1")],
+        }
+
+        scores = _compute_scene_score(query_object_ids, retrieved_results, gamma=1.0)
+
+        # With empty query, no scenes should have positive score
+        assert all(score == 0.0 for score in scores.values())
diff --git a/mini-nav/utils/sam.py b/mini-nav/utils/sam.py
new file mode 100644
index 0000000..a896d90
--- /dev/null
+++ b/mini-nav/utils/sam.py
@@ -0,0 +1,100 @@
+"""SAM (Segment Anything Model) utilities for object segmentation."""
+
+from pathlib import Path
+from typing import Any
+
+import numpy as np
+import torch
+from PIL import Image
+from sam2.build_sam import build_sam2
+from sam2.automatic_mask_generator import SAM2AutomaticMaskGenerator
+
+
+def load_sam_model(
+    model_name: str = "facebook/sam2.1-hiera-large",
+    device: str = "cuda",
+    checkpoint_dir: Path | None = None,
+) -> tuple[Any, Any]:
+    """Load SAM 2.1 model and mask generator.
+
+    Args:
+        model_name: SAM model name (currently supports facebook/sam2.1-hiera-*).
+        device: Device to load model on (cuda or cpu).
+        checkpoint_dir: Optional directory for model checkpoint cache.
+
+    Returns:
+        Tuple of (sam_model, mask_generator).
+    """
+    if device == "cuda" and not torch.cuda.is_available():
+        device = "cpu"
+
+    # Build SAM2 model
+    sam_model = build_sam2(model_name, device=device)
+
+    # Create automatic mask generator
+    mask_generator = SAM2AutomaticMaskGenerator(sam_model)
+
+    return sam_model, mask_generator
+
+
+def segment_image(
+    mask_generator: Any,
+    image: Image.Image,
+    min_area: int = 32 * 32,
+    max_masks: int = 5,
+) -> list[dict[str, Any]]:
+    """Segment image using SAM to extract object masks.
+
+    Args:
+        mask_generator: SAM2AutomaticMaskGenerator instance.
+        image: PIL Image to segment.
+        min_area: Minimum mask area threshold in pixels.
+        max_masks: Maximum number of masks to return.
+
+    Returns:
+        List of mask dictionaries with keys:
+            - segment: Binary mask (numpy array)
+            - area: Mask area in pixels
+            - bbox: Bounding box [x, y, width, height]
+            - predicted_iou: Model's confidence in the mask
+            - stability_score: Stability score for the mask
+    """
+    # Convert PIL Image to numpy array
+    image_np = np.array(image.convert("RGB"))
+
+    # Generate masks
+    masks = mask_generator.generate(image_np)
+
+    if not masks:
+        return []
+
+    # Filter by minimum area
+    filtered_masks = [m for m in masks if m["area"] >= min_area]
+
+    if not filtered_masks:
+        return []
+
+    # Sort by area (largest first) and limit to max_masks
+    sorted_masks = sorted(filtered_masks, key=lambda x: x["area"], reverse=True)
+    return sorted_masks[:max_masks]
+
+
+def extract_masked_region(
+    image: Image.Image,
+    mask: np.ndarray,
+) -> Image.Image:
+    """Extract masked region from image.
+
+    Args:
+        image: Original PIL Image.
+        mask: Binary mask as numpy array (True = keep).
+
+    Returns:
+        PIL Image with only the masked region visible.
+    """
+    image_np = np.array(image.convert("RGB"))
+
+    # Apply mask
+    masked_np = image_np * mask[:, :, np.newaxis]
+
+    return Image.fromarray(masked_np.astype(np.uint8))