feat(compressors): add OWLv2 bbox crop to HashPipeline and refactor image utilities

- Add Owlv2ForObjectDetection and Owlv2Processor imports to model_loader
- Refactor load_dino_model to return tuple of processor and model
- Rewrite generate_proposals_batch to group images by bbox count for efficient batching
- Add _normalize_single_bbox_list helper for bbox normalization
- Update verification.py to use new pipeline architecture with detect/segment/filter/crop steps

mini-nav/compressors/model_loader.py

@@ -6,6 +6,7 @@ import torch
 from transformers import (
     AutoImageProcessor,
     AutoModel,
+    BitImageProcessor,
     Dinov2Model,
     Owlv2ForObjectDetection,
     Owlv2Processor,
@@ -35,7 +36,7 @@ def load_sam_model(
 
 def load_dino_model(
     model_name: str = "facebook/dinov2-large",
-) -> tuple[AutoImageProcessor, Dinov2Model]:
+) -> tuple[BitImageProcessor, Dinov2Model]:
     device = get_device()
 
     processor = AutoImageProcessor.from_pretrained(model_name)

mini-nav/compressors/proposal/core.py

@@ -42,12 +42,13 @@ def generate_proposals(
             - predicted_iou: Model's confidence in the mask
             - stability_score: Stability score for the mask
     """
-    if not bboxes:
+    normalized_bboxes = _normalize_single_bbox_list(bboxes)
+    if not normalized_bboxes:
         return []
 
     device = get_device()
     image_rgb = image.convert("RGB")
-    input_boxes = [bboxes]
+    input_boxes = [normalized_bboxes]
     inputs = processor(
         images=image_rgb,
         input_boxes=input_boxes,
@@ -84,22 +85,68 @@ def generate_proposals_batch(
     if not image_list:
         return []
 
+    if len(image_list) != len(bboxes_per_image):
+        raise ValueError(
+            f"Length mismatch: {len(image_list)} images, {len(bboxes_per_image)} bbox groups"
+        )
+
+    normalized_bboxes_per_image = [
+        _normalize_single_bbox_list(bboxes) for bboxes in bboxes_per_image
+    ]
+
+    proposals_per_image: list[list[dict[str, Any]]] = [[] for _ in image_list]
+    valid_indices = [
+        index for index, bboxes in enumerate(normalized_bboxes_per_image) if bboxes
+    ]
+    if not valid_indices:
+        return proposals_per_image
+
     device = get_device()
-    image_rgb_list = [img.convert("RGB") for img in image_list]
+    grouped_indices: dict[int, list[int]] = {}
+    for image_index in valid_indices:
+        box_count = len(normalized_bboxes_per_image[image_index])
+        grouped_indices.setdefault(box_count, []).append(image_index)
 
-    inputs = processor(
-        images=image_rgb_list,
-        input_boxes=bboxes_per_image,
-        return_tensors="pt",
-    ).to(device)
+    for grouped_image_indices in grouped_indices.values():
+        image_rgb_list = [
+            image_list[image_index].convert("RGB")
+            for image_index in grouped_image_indices
+        ]
+        batched_input_boxes = [
+            normalized_bboxes_per_image[image_index]
+            for image_index in grouped_image_indices
+        ]
 
-    outputs = model(**inputs, multimask_output=False)
-    all_masks = processor.post_process_masks(
-        outputs.pred_masks.cpu(),
-        inputs["original_sizes"],
-    )
+        inputs = processor(
+            images=image_rgb_list,
+            input_boxes=batched_input_boxes,
+            return_tensors="pt",
+        ).to(device)
 
-    return [_masks_to_proposals(image_masks) for image_masks in all_masks]
+        outputs = model(**inputs, multimask_output=False)
+        all_masks = processor.post_process_masks(
+            outputs.pred_masks.cpu(),
+            inputs["original_sizes"],
+        )
+
+        batched_proposals = [
+            _masks_to_proposals(image_masks) for image_masks in all_masks
+        ]
+        for output_index, image_index in enumerate(grouped_image_indices):
+            proposals_per_image[image_index] = batched_proposals[output_index]
+
+    return proposals_per_image
+
+
+def _normalize_single_bbox_list(bboxes: Sequence[Sequence[float]]) -> list[list[float]]:
+    normalized: list[list[float]] = []
+    for bbox in bboxes:
+        bbox_list = list(bbox)
+        if len(bbox_list) != 4:
+            continue
+        normalized.append([float(value) for value in bbox_list])
+
+    return normalized
 
 
 def detect_objects(

mini-nav/configs/config.yaml

@@ -1,35 +1,35 @@
 model:
-  dino_model: "facebook/dinov2-large"
-  compression_dim: 512
-  device: "cuda:3" # auto-detect GPU
-  sam_model: "facebook/sam2.1-hiera-large" # SAM model name
-  sam_min_mask_area: 100 # Minimum mask area threshold
-  sam_max_masks: 10 # Maximum number of masks to keep
-  sam_points_per_batch: 64
-  compressor_path: null # Path to trained HashCompressor weights (optional)
+    dino_model: "facebook/dinov2-large"
+    compression_dim: 512
+    device: "cuda:2" # auto-detect GPU
+    sam_model: "facebook/sam2.1-hiera-large" # SAM model name
+    sam_min_mask_area: 100 # Minimum mask area threshold
+    sam_max_masks: 10 # Maximum number of masks to keep
+    sam_points_per_batch: 64
+    compressor_path: null # Path to trained HashCompressor weights (optional)
 
 output:
-  directory: "./outputs"
+    directory: "./outputs"
 
 dataset:
-  dataset_root: "datasets/InsDet-FULL"
-  output_dir: "datasets/InsDet-FULL/Synthesized"
-  num_objects_range: [3, 8]
-  num_scenes: 1000
-  object_scale_range: [0.1, 0.4]
-  rotation_range: [-30, 30]
-  overlap_threshold: 0.3
-  seed: 42
+    dataset_root: "datasets/InsDet-FULL"
+    output_dir: "datasets/InsDet-FULL/Synthesized"
+    num_objects_range: [3, 8]
+    num_scenes: 1000
+    object_scale_range: [0.1, 0.4]
+    rotation_range: [-30, 30]
+    overlap_threshold: 0.3
+    seed: 42
 
 benchmark:
-  dataset:
-    source_type: "huggingface"
-    path: "uoft-cs/cifar10"
-    img_column: "img"
-    label_column: "label"
-  task:
-    name: "recall_at_k"
-    type: "retrieval"
-    top_k: 1
-  batch_size: 64
-  model_table_prefix: "benchmark"
+    dataset:
+        source_type: "huggingface"
+        path: "uoft-cs/cifar10"
+        img_column: "img"
+        label_column: "label"
+    task:
+        name: "recall_at_k"
+        type: "retrieval"
+        top_k: 1
+    batch_size: 64
+    model_table_prefix: "benchmark"