vision-agent 0.2.224__py3-none-any.whl → 0.2.225__py3-none-any.whl

vision_agent/tools/tools.py

@@ -6,7 +6,6 @@ import tempfile
 import urllib.request
 from base64 import b64encode
 from concurrent.futures import ThreadPoolExecutor, as_completed
-from enum import Enum
 from importlib import resources
 from pathlib import Path
 from typing import Any, Dict, List, Optional, Tuple, Union, cast
@@ -54,6 +53,13 @@ from vision_agent.utils.video import (
     frames_to_bytes,
     video_writer,
 )
+from vision_agent.utils.video_tracking import (
+    ODModels,
+    merge_segments,
+    post_process,
+    process_segment,
+    split_frames_into_segments,
+)
 
 register_heif_opener()
 
@@ -224,12 +230,6 @@ def sam2(
     return ret["return_data"]  # type: ignore
 
 
-class ODModels(str, Enum):
-    COUNTGD = "countgd"
-    FLORENCE2 = "florence2"
-    OWLV2 = "owlv2"
-
-
 def od_sam2_video_tracking(
     od_model: ODModels,
     prompt: str,
@@ -237,105 +237,92 @@ def od_sam2_video_tracking(
     chunk_length: Optional[int] = 10,
     fine_tune_id: Optional[str] = None,
 ) -> Dict[str, Any]:
-    results: List[Optional[List[Dict[str, Any]]]] = [None] * len(frames)
+    SEGMENT_SIZE = 50
+    OVERLAP = 1  # Number of overlapping frames between segments
 
-    if chunk_length is None:
-        step = 1  # Process every frame
-    elif chunk_length <= 0:
-        raise ValueError("chunk_length must be a positive integer or None.")
-    else:
-        step = chunk_length  # Process frames with the specified step size
+    image_size = frames[0].shape[:2]
+
+    # Split frames into segments with overlap
+    segments = split_frames_into_segments(frames, SEGMENT_SIZE, OVERLAP)
+
+    def _apply_object_detection(  # inner method to avoid circular importing issues.
+        od_model: ODModels,
+        prompt: str,
+        segment_index: int,
+        frame_number: int,
+        fine_tune_id: str,
+        segment_frames: list,
+    ) -> tuple:
+        """
+        Applies the specified object detection model to the given image.
+
+        Args:
+            od_model: The object detection model to use.
+            prompt: The prompt for the object detection model.
+            segment_index: The index of the current segment.
+            frame_number: The number of the current frame.
+            fine_tune_id: Optional fine-tune ID for the model.
+            segment_frames: List of frames for the current segment.
+
+        Returns:
+            A tuple containing the object detection results and the name of the function used.
+        """
 
-    for idx in range(0, len(frames), step):
         if od_model == ODModels.COUNTGD:
-            results[idx] = countgd_object_detection(prompt=prompt, image=frames[idx])
+            segment_results = countgd_object_detection(
+                prompt=prompt, image=segment_frames[frame_number]
+            )
             function_name = "countgd_object_detection"
+
         elif od_model == ODModels.OWLV2:
-            results[idx] = owlv2_object_detection(
-                prompt=prompt, image=frames[idx], fine_tune_id=fine_tune_id
+            segment_results = owlv2_object_detection(
+                prompt=prompt,
+                image=segment_frames[frame_number],
+                fine_tune_id=fine_tune_id,
             )
             function_name = "owlv2_object_detection"
+
         elif od_model == ODModels.FLORENCE2:
-            results[idx] = florence2_object_detection(
-                prompt=prompt, image=frames[idx], fine_tune_id=fine_tune_id
+            segment_results = florence2_object_detection(
+                prompt=prompt,
+                image=segment_frames[frame_number],
+                fine_tune_id=fine_tune_id,
             )
             function_name = "florence2_object_detection"
+
         else:
             raise NotImplementedError(
                 f"Object detection model '{od_model}' is not implemented."
             )
 
-    image_size = frames[0].shape[:2]
-
-    def _transform_detections(
-        input_list: List[Optional[List[Dict[str, Any]]]],
-    ) -> List[Optional[Dict[str, Any]]]:
-        output_list: List[Optional[Dict[str, Any]]] = []
-
-        for _, frame in enumerate(input_list):
-            if frame is not None:
-                labels = [detection["label"] for detection in frame]
-                bboxes = [
-                    denormalize_bbox(detection["bbox"], image_size)
-                    for detection in frame
-                ]
-
-                output_list.append(
-                    {
-                        "labels": labels,
-                        "bboxes": bboxes,
-                    }
-                )
-            else:
-                output_list.append(None)
-
-        return output_list
+        return segment_results, function_name
+
+    # Process each segment and collect detections
+    detections_per_segment: List[Any] = []
+    for segment_index, segment in enumerate(segments):
+        segment_detections = process_segment(
+            segment_frames=segment,
+            od_model=od_model,
+            prompt=prompt,
+            fine_tune_id=fine_tune_id,
+            chunk_length=chunk_length,
+            image_size=image_size,
+            segment_index=segment_index,
+            object_detection_tool=_apply_object_detection,
+        )
+        detections_per_segment.append(segment_detections)
 
-    output = _transform_detections(results)
+    merged_detections = merge_segments(detections_per_segment)
+    post_processed = post_process(merged_detections, image_size)
 
     buffer_bytes = frames_to_bytes(frames)
     files = [("video", buffer_bytes)]
-    payload = {"bboxes": json.dumps(output), "chunk_length_frames": chunk_length}
-    metadata = {"function_name": function_name}
-
-    detections = send_task_inference_request(
-        payload,
-        "sam2",
-        files=files,
-        metadata=metadata,
-    )
 
-    return_data = []
-    for frame in detections:
-        return_frame_data = []
-        for detection in frame:
-            mask = rle_decode_array(detection["mask"])
-            label = str(detection["id"]) + ": " + detection["label"]
-            return_frame_data.append(
-                {"label": label, "mask": mask, "score": 1.0, "rle": detection["mask"]}
-            )
-        return_data.append(return_frame_data)
-    return_data = add_bboxes_from_masks(return_data)
-    return_data = nms(return_data, iou_threshold=0.95)
-
-    # We save the RLE for display purposes, re-calculting RLE can get very expensive.
-    # Deleted here because we are returning the numpy masks instead
-    display_data = []
-    for frame in return_data:
-        display_frame_data = []
-        for obj in frame:
-            display_frame_data.append(
-                {
-                    "label": obj["label"],
-                    "score": obj["score"],
-                    "bbox": denormalize_bbox(obj["bbox"], image_size),
-                    "mask": obj["rle"],
-                }
-            )
-            del obj["rle"]
-        display_data.append(display_frame_data)
-
-    return {"files": files, "return_data": return_data, "display_data": detections}
+    return {
+        "files": files,
+        "return_data": post_processed["return_data"],
+        "display_data": post_processed["display_data"],
+    }
 
 
 # Owl V2 Tools

vision_agent/utils/video_tracking.py

@@ -0,0 +1,305 @@
+import json
+from enum import Enum
+from typing import Any, Callable, Dict, List, Optional, Tuple
+
+import numpy as np
+
+from vision_agent.tools.tool_utils import (
+    add_bboxes_from_masks,
+    nms,
+    send_task_inference_request,
+)
+from vision_agent.utils.image_utils import denormalize_bbox, rle_decode_array
+from vision_agent.utils.video import frames_to_bytes
+
+
+class ODModels(str, Enum):
+    COUNTGD = "countgd"
+    FLORENCE2 = "florence2"
+    OWLV2 = "owlv2"
+
+
+def split_frames_into_segments(
+    frames: List[np.ndarray], segment_size: int = 50, overlap: int = 1
+) -> List[List[np.ndarray]]:
+    """
+    Splits the list of frames into segments with a specified size and overlap.
+
+    Args:
+        frames (List[np.ndarray]): List of video frames.
+        segment_size (int, optional): Number of frames per segment. Defaults to 50.
+        overlap (int, optional): Number of overlapping frames between segments. Defaults to 1.
+
+    Returns:
+        List[List[np.ndarray]]: List of frame segments.
+    """
+    segments = []
+    start = 0
+    segment_count = 0
+    while start < len(frames):
+        end = start + segment_size
+        if end > len(frames):
+            end = len(frames)
+        if start != 0:
+            # Include the last frame of the previous segment
+            segment = frames[start - overlap : end]
+        else:
+            segment = frames[start:end]
+        segments.append(segment)
+        start += segment_size
+        segment_count += 1
+    return segments
+
+
+def process_segment(
+    segment_frames: List[np.ndarray],
+    od_model: ODModels,
+    prompt: str,
+    fine_tune_id: Optional[str],
+    chunk_length: Optional[int],
+    image_size: Tuple[int, ...],
+    segment_index: int,
+    object_detection_tool: Callable,
+) -> Any:
+    """
+    Processes a segment of frames with the specified object detection model.
+
+    Args:
+        segment_frames (List[np.ndarray]): Frames in the segment.
+        od_model (ODModels): Object detection model to use.
+        prompt (str): Prompt for the model.
+        fine_tune_id (Optional[str]): Fine-tune model ID.
+        chunk_length (Optional[int]): Chunk length for processing.
+        image_size (Tuple[int, int]): Size of the images.
+        segment_index (int): Index of the segment.
+        object_detection_tool (Callable): Object detection tool to use.
+
+    Returns:
+       Any: Detections for the segment.
+    """
+    segment_results: List[Optional[List[Dict[str, Any]]]] = [None] * len(segment_frames)
+
+    if chunk_length is None:
+        step = 1
+    elif chunk_length <= 0:
+        raise ValueError("chunk_length must be a positive integer or None.")
+    else:
+        step = chunk_length
+
+    function_name = ""
+
+    for idx in range(0, len(segment_frames), step):
+        frame_number = idx
+        segment_results[idx], function_name = object_detection_tool(
+            od_model, prompt, segment_index, frame_number, fine_tune_id, segment_frames
+        )
+
+    transformed_detections = transform_detections(
+        segment_results, image_size, segment_index
+    )
+
+    buffer_bytes = frames_to_bytes(segment_frames)
+    files = [("video", buffer_bytes)]
+    payload = {
+        "bboxes": json.dumps(transformed_detections),
+        "chunk_length_frames": chunk_length,
+    }
+    metadata = {"function_name": function_name}
+
+    segment_detections = send_task_inference_request(
+        payload,
+        "sam2",
+        files=files,
+        metadata=metadata,
+    )
+
+    return segment_detections
+
+
+def transform_detections(
+    input_list: List[Optional[List[Dict[str, Any]]]],
+    image_size: Tuple[int, ...],
+    segment_index: int,
+) -> List[Optional[Dict[str, Any]]]:
+    """
+    Transforms raw detections into a standardized format.
+
+    Args:
+        input_list (List[Optional[List[Dict[str, Any]]]]): Raw detections.
+        image_size (Tuple[int, int]): Size of the images.
+        segment_index (int): Index of the segment.
+
+    Returns:
+        List[Optional[Dict[str, Any]]]: Transformed detections.
+    """
+    output_list: List[Optional[Dict[str, Any]]] = []
+    for frame_idx, frame in enumerate(input_list):
+        if frame is not None:
+            labels = [detection["label"] for detection in frame]
+            bboxes = [
+                denormalize_bbox(detection["bbox"], image_size) for detection in frame
+            ]
+
+            output_list.append(
+                {
+                    "labels": labels,
+                    "bboxes": bboxes,
+                }
+            )
+        else:
+            output_list.append(None)
+    return output_list
+
+
+def _calculate_mask_iou(mask1: np.ndarray, mask2: np.ndarray) -> float:
+    mask1 = mask1.astype(bool)
+    mask2 = mask2.astype(bool)
+
+    intersection = np.sum(np.logical_and(mask1, mask2))
+    union = np.sum(np.logical_or(mask1, mask2))
+
+    if union == 0:
+        iou = 0.0
+    else:
+        iou = intersection / union
+
+    return iou
+
+
+def _match_by_iou(
+    first_param: List[Dict],
+    second_param: List[Dict],
+    iou_threshold: float = 0.8,
+) -> Tuple[List[Dict], Dict[int, int]]:
+    max_id = max((item["id"] for item in first_param), default=0)
+
+    matched_new_item_indices = set()
+    id_mapping = {}
+
+    for new_index, new_item in enumerate(second_param):
+        matched_id = None
+
+        for existing_item in first_param:
+            iou = _calculate_mask_iou(
+                existing_item["decoded_mask"], new_item["decoded_mask"]
+            )
+            if iou > iou_threshold:
+                matched_id = existing_item["id"]
+                matched_new_item_indices.add(new_index)
+                id_mapping[new_item["id"]] = matched_id
+                break
+
+        if matched_id:
+            new_item["id"] = matched_id
+        else:
+            max_id += 1
+            id_mapping[new_item["id"]] = max_id
+            new_item["id"] = max_id
+
+    unmatched_items = [
+        item for i, item in enumerate(second_param) if i not in matched_new_item_indices
+    ]
+    combined_list = first_param + unmatched_items
+
+    return combined_list, id_mapping
+
+
+def _update_ids(detections: List[Dict], id_mapping: Dict[int, int]) -> None:
+    for inner_list in detections:
+        for detection in inner_list:
+            if detection["id"] in id_mapping:
+                detection["id"] = id_mapping[detection["id"]]
+            else:
+                max_new_id = max(id_mapping.values(), default=0)
+                detection["id"] = max_new_id + 1
+                id_mapping[detection["id"]] = detection["id"]
+
+
+def _convert_to_2d(detections_per_segment: List[Any]) -> List[Any]:
+    result = []
+    for i, segment in enumerate(detections_per_segment):
+        if i == 0:
+            result.extend(segment)
+        else:
+            result.extend(segment[1:])
+    return result
+
+
+def merge_segments(detections_per_segment: List[Any]) -> List[Any]:
+    """
+    Merges detections from all segments into a unified result.
+
+    Args:
+        detections_per_segment (List[Any]): List of detections per segment.
+
+    Returns:
+        List[Any]: Merged detections.
+    """
+    for segment in detections_per_segment:
+        for detection in segment:
+            for item in detection:
+                item["decoded_mask"] = rle_decode_array(item["mask"])
+
+    for segment_idx in range(len(detections_per_segment) - 1):
+        combined_detection, id_mapping = _match_by_iou(
+            detections_per_segment[segment_idx][-1],
+            detections_per_segment[segment_idx + 1][0],
+        )
+        _update_ids(detections_per_segment[segment_idx + 1], id_mapping)
+
+    merged_result = _convert_to_2d(detections_per_segment)
+
+    return merged_result
+
+
+def post_process(
+    merged_detections: List[Any],
+    image_size: Tuple[int, ...],
+) -> Dict[str, Any]:
+    """
+    Performs post-processing on merged detections, including NMS and preparing display data.
+
+    Args:
+        merged_detections (List[Any]): Merged detections from all segments.
+        image_size (Tuple[int, int]): Size of the images.
+
+    Returns:
+        Dict[str, Any]: Post-processed data including return_data and display_data.
+    """
+    return_data = []
+    for frame_idx, frame in enumerate(merged_detections):
+        return_frame_data = []
+        for detection in frame:
+            label = f"{detection['id']}: {detection['label']}"
+            return_frame_data.append(
+                {
+                    "label": label,
+                    "mask": detection["decoded_mask"],
+                    "rle": detection["mask"],
+                    "score": 1.0,
+                }
+            )
+            del detection["decoded_mask"]
+        return_data.append(return_frame_data)
+
+    return_data = add_bboxes_from_masks(return_data)
+    return_data = nms(return_data, iou_threshold=0.95)
+
+    # We save the RLE for display purposes, re-calculting RLE can get very expensive.
+    # Deleted here because we are returning the numpy masks instead
+    display_data = []
+    for frame in return_data:
+        display_frame_data = []
+        for obj in frame:
+            display_frame_data.append(
+                {
+                    "label": obj["label"],
+                    "bbox": denormalize_bbox(obj["bbox"], image_size),
+                    "mask": obj["rle"],
+                    "score": obj["score"],
+                }
+            )
+            del obj["rle"]
+        display_data.append(display_frame_data)
+
+    return {"return_data": return_data, "display_data": display_data}

{vision_agent-0.2.224.dist-info → vision_agent-0.2.225.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: vision-agent
-Version: 0.2.224
+Version: 0.2.225
 Summary: Toolset for Vision Agent
 Author: Landing AI
 Author-email: dev@landing.ai

{vision_agent-0.2.224.dist-info → vision_agent-0.2.225.dist-info}/RECORD

@@ -31,7 +31,7 @@ vision_agent/tools/meta_tools.py,sha256=TPeS7QWnc_PmmU_ndiDT03dXbQ5yDSP33E7U8cSj
 vision_agent/tools/planner_tools.py,sha256=CvaJ2vGM8O_CYvsoSk1avxAMqpIu3tv4C2bY0p1X-X4,13519
 vision_agent/tools/prompts.py,sha256=V1z4YJLXZuUl_iZ5rY0M5hHc_2tmMEUKr0WocXKGt4E,1430
 vision_agent/tools/tool_utils.py,sha256=q9cqXO2AvigUdO1krjnOy8o0goYhgS6eILl6-F5Kxyk,10211
-vision_agent/tools/tools.py,sha256=60S5ItFG9yKzVb8FU8oLFj_aouDg2-4vlieDbSgfPdQ,91306
+vision_agent/tools/tools.py,sha256=cQYO1TfWhm9C_KaU201aTYec-w0m9QoQMzqjxWvQWGU,90770
 vision_agent/tools/tools_types.py,sha256=8hYf2OZhI58gvf65KGaeGkt4EQ56nwLFqIQDPHioOBc,2339
 vision_agent/utils/__init__.py,sha256=QKk4zVjMwGxQI0MQ-aZZA50N-qItxRY4EB9CwQkZ2HY,185
 vision_agent/utils/exceptions.py,sha256=booSPSuoULF7OXRr_YbC4dtKt6gM_HyiFQHBuaW86C4,2052
@@ -40,7 +40,8 @@ vision_agent/utils/image_utils.py,sha256=z_ONgcza125B10NkoGwPOzXnL470bpTWZbkB16N
 vision_agent/utils/sim.py,sha256=znsInUDrsyBi3OlgAlV3rDn5UQQRfJAWXTXm7D7eJA8,9125
 vision_agent/utils/type_defs.py,sha256=BE12s3JNQy36QvauXHjwyeffVh5enfcvd4vTzSwvEZI,1384
 vision_agent/utils/video.py,sha256=e1VwKhXzzlC5LcFMyrcQYrPnpnX4wxDpnQ-76sB4jgM,6001
-vision_agent-0.2.224.dist-info/LICENSE,sha256=xx0jnfkXJvxRnG63LTGOxlggYnIysveWIZ6H3PNdCrQ,11357
-vision_agent-0.2.224.dist-info/METADATA,sha256=wT49_byW9-Oz6-1eSlP3cW_AFGbWaxtKrYsGB4nT62o,20039
-vision_agent-0.2.224.dist-info/WHEEL,sha256=7Z8_27uaHI_UZAc4Uox4PpBhQ9Y5_modZXWMxtUi4NU,88
-vision_agent-0.2.224.dist-info/RECORD,,
+vision_agent/utils/video_tracking.py,sha256=EeOiSY8gjvvneuAnv-BO7yOyMBF_-1Irk_lLLOt3bDM,9452
+vision_agent-0.2.225.dist-info/LICENSE,sha256=xx0jnfkXJvxRnG63LTGOxlggYnIysveWIZ6H3PNdCrQ,11357
+vision_agent-0.2.225.dist-info/METADATA,sha256=PzehPaQUIj_3TImCmj1YEFMI1rPkrd6FqcIlXhmWjLE,20039
+vision_agent-0.2.225.dist-info/WHEEL,sha256=7Z8_27uaHI_UZAc4Uox4PpBhQ9Y5_modZXWMxtUi4NU,88
+vision_agent-0.2.225.dist-info/RECORD,,