Feat: MOT dataset loader

trackers/dataset/base.py

@@ -0,0 +1,87 @@
+import abc
+from typing import Any, Dict, Iterator, List, Optional, Tuple
+
+import supervision as sv
+
+
+# --- Base Dataset ---
+class Dataset(abc.ABC):
+    """Abstract base class for datasets used in tracking evaluation."""
+
+    @abc.abstractmethod
+    def load_ground_truth(self, sequence_name: str) -> Optional[sv.Detections]:
+        """
+        Loads ground truth data for a specific sequence.
+
+        Args:
+            sequence_name: The name of the sequence.
+
+        Returns:
+            An sv.Detections object containing ground truth annotations, or None
+            if ground truth is not available or cannot be loaded. The Detections
+            object should ideally include `tracker_id` and `data['frame_idx']`.
+        """
+        pass
+
+    @abc.abstractmethod
+    def get_sequence_names(self) -> List[str]:
+        """Returns a list of sequence names available in the dataset."""
+        pass
+
+    @abc.abstractmethod
+    def get_sequence_info(self, sequence_name: str) -> Dict[str, Any]:
+        """
+        Returns metadata for a specific sequence.
+
+        Args:
+            sequence_name: The name of the sequence.
+
+        Returns:
+            A dictionary containing sequence information (e.g., 'frame_rate',
+            'seqLength', 'img_width', 'img_height', 'img_dir'). Keys and value
+            types may vary depending on the dataset format.
+        """
+        pass
+
+    @abc.abstractmethod
+    def get_frame_iterator(self, sequence_name: str) -> Iterator[Dict[str, Any]]:
+        """
+        Returns an iterator over frame information dictionaries for a sequence.
+
+        Args:
+            sequence_name: The name of the sequence.
+
+        Yields:
+            Dictionaries, each representing a frame. Each dictionary should
+            contain at least 'frame_idx' (int, typically 1-based) and
+            'image_path' (str, absolute path recommended).
+        """
+        pass
+
+    @abc.abstractmethod
+    def preprocess(
+        self,
+        ground_truth: sv.Detections,
+        predictions: sv.Detections,
+        iou_threshold: float = 0.5,
+        remove_distractor_matches: bool = True,
+    ) -> Tuple[sv.Detections, sv.Detections]:
+        """
+        Applies dataset-specific preprocessing steps to ground truth and predictions.
+
+        This typically involves filtering unwanted annotations (e.g., distractors,
+        zero-marked GTs) and potentially relabeling IDs.
+
+        Args:
+            ground_truth (sv.Detections): Raw ground truth detections for a sequence.
+            predictions (sv.Detections): Raw prediction detections for a sequence.
+            iou_threshold (float): IoU threshold used for matching during preprocessing
+                                 (e.g., for removing predictions matching distractors).
+            remove_distractor_matches (bool): Flag indicating whether to remove
+                                              predictions matched to distractors.
+
+        Returns:
+            Tuple[sv.Detections, sv.Detections]: A tuple containing the processed
+            ground truth detections and processed prediction detections.
+        """
+        pass