refactor(compressors): switch SAM from automatic mask generation to bbox-prompted segmentation

- Replace SAM2AutomaticMaskGenerator pipeline with Sam2Processor+Sam2Model
- Freeze SAM model parameters at load time, removing torch.no_grad() at call sites
- Rewrite proposal/core.py to use bbox prompts instead of automatic point sampling
- Add bboxes parameter to all HashPipeline public methods (forward, forward_dataset, extract_features, extract_features_dataset)
- Extract mask filtering logic (_filter_masks) from proposal into pipeline
- Rename object_score/ to filter/
- Add load_owlv2_model to model_loader
- Rename notebooks/test.py to habitat_sim_setup.py

mini-nav/compressors/model_loader.py

@@ -1,11 +1,17 @@
-"""Model loading utilities for DINO, SAM2 and HashCompressor."""
+"""Model loading utilities for DINO, SAM2, OWLv2 and HashCompressor."""
 
 from typing import TYPE_CHECKING, Any
 
-
 import torch
-from transformers import AutoImageProcessor, AutoModel, pipeline, MaskGenerationPipeline
-
+from transformers import (
+    AutoImageProcessor,
+    AutoModel,
+    Owlv2ForObjectDetection,
+    Owlv2Model,
+    Owlv2Processor,
+    Sam2Model,
+    Sam2Processor,
+)
 from utils import get_device
 
 if TYPE_CHECKING:
@@ -14,14 +20,17 @@ if TYPE_CHECKING:
 
 def load_sam_model(
     model_name: str = "facebook/sam2.1-hiera-large",
-) -> MaskGenerationPipeline:
+) -> tuple[Sam2Processor, Sam2Model]:
+    """Load SAM2 processor and model with frozen parameters."""
     device = get_device()
 
-    return pipeline(
-        task="mask-generation",
-        model=model_name,
-        device=device,
-    )
+    processor = Sam2Processor.from_pretrained(model_name)
+    model = Sam2Model.from_pretrained(model_name).to(device)
+    model.eval()
+    for param in model.parameters():
+        param.requires_grad = False
+
+    return processor, model
 
 
 def load_dino_model(
@@ -36,6 +45,18 @@ def load_dino_model(
     return processor, dino
 
 
+def load_owlv2_model(
+    model_name: str = "google/owlv2-base-patch16-ensemble",
+) -> tuple[Owlv2Processor, Owlv2Model]:
+    device = get_device()
+
+    processor = Owlv2Processor.from_pretrained(model_name)
+    model = Owlv2ForObjectDetection.from_pretrained(model_name).to(device)
+    model.eval()
+
+    return processor, model
+
+
 def get_dino_dim(model_name: str) -> int:
     if "large" in model_name.lower():
         return 1024