Skip to content

Detection Smoother

supervision.detection.tools.smoother.DetectionsSmoother

A utility class for smoothing detections over multiple frames in video tracking. It maintains a history of detections for each track and provides smoothed predictions based on these histories.

Warning

  • DetectionsSmoother requires the tracker_id for each detection. Refer to Roboflow Trackers for information on integrating tracking into your inference pipeline.
  • This class is not compatible with segmentation models.
  • When detections in a frame disagree on confidence presence — some tracks carry confidence scores and others do not — confidence is set to None for all smoothed detections in that frame.
Example 
>>> import numpy as np
>>> import supervision as sv
>>> smoother = sv.DetectionsSmoother(length=3)
>>> detections_1 = sv.Detections(
...     xyxy=np.array([[0, 0, 10, 10]]),
...     confidence=np.array([0.5]),
...     tracker_id=np.array([1])
... )
>>> detections_2 = sv.Detections(
...     xyxy=np.array([[2, 2, 12, 12]]),
...     confidence=np.array([0.7]),
...     tracker_id=np.array([1])
... )
>>> smoothed = smoother.update_with_detections(detections_1)
>>> smoothed.xyxy
array([[ 0.,  0., 10., 10.]])
>>> smoothed = smoother.update_with_detections(detections_2)
>>> smoothed.xyxy
array([[ 1.,  1., 11., 11.]])
>>> smoothed.confidence
array([0.6])

import supervision as sv

from ultralytics import YOLO

video_info = sv.VideoInfo.from_video_path(video_path="<SOURCE_FILE_PATH>")
frame_generator = sv.get_video_frames_generator(
    source_path="<SOURCE_FILE_PATH>")

model = YOLO("<MODEL_PATH>")
tracker = sv.ByteTrack(frame_rate=video_info.fps)
smoother = sv.DetectionsSmoother()

box_annotator = sv.BoxAnnotator()

with sv.VideoSink("<TARGET_FILE_PATH>", video_info=video_info) as sink:
    for frame in frame_generator:
        result = model(frame)[0]
        detections = sv.Detections.from_ultralytics(result)
        detections = tracker.update_with_detections(detections)
        detections = smoother.update_with_detections(detections)

        annotated_frame = box_annotator.annotate(frame.copy(), detections)
        sink.write_frame(annotated_frame)
Source code in src/supervision/detection/tools/smoother.py 
 11
 12
 13
 14
 15
 16
 17
 18
 19
 20
 21
 22
 23
 24
 25
 26
 27
 28
 29
 30
 31
 32
 33
 34
 35
 36
 37
 38
 39
 40
 41
 42
 43
 44
 45
 46
 47
 48
 49
 50
 51
 52
 53
 54
 55
 56
 57
 58
 59
 60
 61
 62
 63
 64
 65
 66
 67
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
class DetectionsSmoother:
    """
    A utility class for smoothing detections over multiple frames in video tracking.
    It maintains a history of detections for each track and provides smoothed
    predictions based on these histories.

    <video controls>
        <source
            src="https://media.roboflow.com/supervision-detection-smoothing.mp4"
            type="video/mp4">
    </video>

    !!! warning

        - `DetectionsSmoother` requires the `tracker_id` for each detection. Refer to
          [Roboflow Trackers](/latest/trackers/) for
          information on integrating tracking into your inference pipeline.
        - This class is not compatible with segmentation models.
        - When detections in a frame disagree on confidence presence — some tracks
          carry confidence scores and others do not — `confidence` is set to `None`
          for all smoothed detections in that frame.

    Example:
        ```pycon
        >>> import numpy as np
        >>> import supervision as sv
        >>> smoother = sv.DetectionsSmoother(length=3)
        >>> detections_1 = sv.Detections(
        ...     xyxy=np.array([[0, 0, 10, 10]]),
        ...     confidence=np.array([0.5]),
        ...     tracker_id=np.array([1])
        ... )
        >>> detections_2 = sv.Detections(
        ...     xyxy=np.array([[2, 2, 12, 12]]),
        ...     confidence=np.array([0.7]),
        ...     tracker_id=np.array([1])
        ... )
        >>> smoothed = smoother.update_with_detections(detections_1)
        >>> smoothed.xyxy
        array([[ 0.,  0., 10., 10.]])
        >>> smoothed = smoother.update_with_detections(detections_2)
        >>> smoothed.xyxy
        array([[ 1.,  1., 11., 11.]])
        >>> smoothed.confidence
        array([0.6])

        ```


        ```python
        import supervision as sv

        from ultralytics import YOLO

        video_info = sv.VideoInfo.from_video_path(video_path="<SOURCE_FILE_PATH>")
        frame_generator = sv.get_video_frames_generator(
            source_path="<SOURCE_FILE_PATH>")

        model = YOLO("<MODEL_PATH>")
        tracker = sv.ByteTrack(frame_rate=video_info.fps)
        smoother = sv.DetectionsSmoother()

        box_annotator = sv.BoxAnnotator()

        with sv.VideoSink("<TARGET_FILE_PATH>", video_info=video_info) as sink:
            for frame in frame_generator:
                result = model(frame)[0]
                detections = sv.Detections.from_ultralytics(result)
                detections = tracker.update_with_detections(detections)
                detections = smoother.update_with_detections(detections)

                annotated_frame = box_annotator.annotate(frame.copy(), detections)
                sink.write_frame(annotated_frame)
        ```
    """

    def __init__(self, length: int = 5) -> None:
        """
        Args:
            length: The maximum number of frames to consider for smoothing
                detections. Defaults to 5.
        """
        self.tracks: defaultdict[int, deque[Detections | None]] = defaultdict(
            lambda: deque(maxlen=length)
        )

    def reset(self) -> None:
        """
        Clears the per-track detection history so the smoother can be reused
        across independent streams without carrying over frames from a
        previous stream. The configured window `length` is preserved.

        Examples:
            ```pycon
            >>> import numpy as np
            >>> import supervision as sv
            >>> smoother = sv.DetectionsSmoother()
            >>> detections = sv.Detections(
            ...     xyxy=np.array([[0, 0, 10, 10]]),
            ...     confidence=np.array([0.5]),
            ...     tracker_id=np.array([1])
            ... )
            >>> _ = smoother.update_with_detections(detections)
            >>> len(smoother.tracks)
            1
            >>> smoother.reset()
            >>> len(smoother.tracks)
            0

            ```
        """
        self.tracks.clear()

    def update_with_detections(self, detections: Detections) -> Detections:
        """
        Updates the smoother with a new set of detections from a frame.

        Args:
            detections: The detections to add to the smoother.
        """

        if detections.tracker_id is None:
            warnings.warn(
                "Smoothing skipped. DetectionsSmoother requires tracker_id. Refer to "
                "https://supervision.roboflow.com/latest/trackers for more "
                "information.",
                category=SupervisionWarnings,
            )
            return detections

        for detection_idx in range(len(detections)):
            tracker_id_value = detections.tracker_id[detection_idx]
            tracker_id = int(tracker_id_value)

            self.tracks[tracker_id].append(detections.select(detection_idx))

        for track_id in self.tracks.keys():
            if track_id not in detections.tracker_id:
                self.tracks[track_id].append(None)

        for track_id in list(self.tracks.keys()):
            if all([d is None for d in self.tracks[track_id]]):
                del self.tracks[track_id]

        current_track_ids = {int(track_id) for track_id in detections.tracker_id}
        return self.get_smoothed_detections(track_ids=current_track_ids)

    def get_track(self, track_id: int) -> Detections | None:
        """Return the smoothed `Detections` for a single track.

        Averages `xyxy` over all valid (non-`None`) frames in the track window.
        `confidence` is averaged only over frames that carry it; frames with
        `confidence=None` are excluded. Returns `None` when the track is unknown
        or its entire window is empty.

        Args:
            track_id: The tracker ID whose smoothed detection to retrieve.

        Returns:
            Smoothed `Detections` for the track, or `None` if the track is
            unknown or all frames in its window are empty.
        """
        track = self.tracks.get(track_id, None)
        if track is None:
            return None

        valid: list[Detections] = [d for d in track if d is not None]
        if len(valid) == 0:
            return None

        ret = deepcopy(valid[0])
        ret.xyxy = np.mean(np.stack([d.xyxy for d in valid], axis=0), axis=0)
        # Average confidence only over frames that carry it; frames with
        # confidence=None contribute nothing to the mean. Retain None when
        # no frame in the window carries confidence.
        confidences = [d.confidence for d in valid if d.confidence is not None]
        ret.confidence = np.mean(np.array(confidences), axis=0) if confidences else None

        return ret

    def get_smoothed_detections(self, track_ids: set[int] | None = None) -> Detections:
        """Return the smoothed detections for the requested active tracks.

        Args:
            track_ids: Optional set of track IDs to include in the output. When
                provided, tracks absent from the current frame are excluded from the
                emitted detections but their history stays cached.
        """
        tracked_detections = []
        for track_id in self.tracks:
            if track_ids is not None and track_id not in track_ids:
                continue
            track = self.get_track(track_id)
            if track is not None:
                tracked_detections.append(track)

        # Detections.merge requires all-or-none for optional fields.
        # When tracks disagree on confidence presence, drop it from all to
        # prevent ValueError inside Detections.merge (stack_or_none invariant).
        if tracked_detections and any(d.confidence is None for d in tracked_detections):
            for d in tracked_detections:
                d.confidence = None

        detections = Detections.merge(tracked_detections)
        if len(detections) == 0:
            detections.tracker_id = np.array([], dtype=int)

        return detections

Methods:

__init__(length: int = 5) -> None

Parameters:

Name Type Description Default
length
int

The maximum number of frames to consider for smoothing detections. Defaults to 5.

 5
Source code in src/supervision/detection/tools/smoother.py 
87
88
89
90
91
92
93
94
95
def __init__(self, length: int = 5) -> None:
    """
    Args:
        length: The maximum number of frames to consider for smoothing
            detections. Defaults to 5.
    """
    self.tracks: defaultdict[int, deque[Detections | None]] = defaultdict(
        lambda: deque(maxlen=length)
    )

get_smoothed_detections(track_ids: set[int] | None = None) -> Detections

Return the smoothed detections for the requested active tracks.

Parameters:

Name Type Description Default
track_ids
set[int] | None

Optional set of track IDs to include in the output. When provided, tracks absent from the current frame are excluded from the emitted detections but their history stays cached.

 None
Source code in src/supervision/detection/tools/smoother.py 
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
def get_smoothed_detections(self, track_ids: set[int] | None = None) -> Detections:
    """Return the smoothed detections for the requested active tracks.

    Args:
        track_ids: Optional set of track IDs to include in the output. When
            provided, tracks absent from the current frame are excluded from the
            emitted detections but their history stays cached.
    """
    tracked_detections = []
    for track_id in self.tracks:
        if track_ids is not None and track_id not in track_ids:
            continue
        track = self.get_track(track_id)
        if track is not None:
            tracked_detections.append(track)

    # Detections.merge requires all-or-none for optional fields.
    # When tracks disagree on confidence presence, drop it from all to
    # prevent ValueError inside Detections.merge (stack_or_none invariant).
    if tracked_detections and any(d.confidence is None for d in tracked_detections):
        for d in tracked_detections:
            d.confidence = None

    detections = Detections.merge(tracked_detections)
    if len(detections) == 0:
        detections.tracker_id = np.array([], dtype=int)

    return detections

get_track(track_id: int) -> Detections | None

Return the smoothed Detections for a single track.

Averages xyxy over all valid (non-None) frames in the track window. confidence is averaged only over frames that carry it; frames with confidence=None are excluded. Returns None when the track is unknown or its entire window is empty.

Parameters:

Name Type Description Default
track_id
int

The tracker ID whose smoothed detection to retrieve.

 required

Returns:

Type Description
Detections | None

Smoothed Detections for the track, or None if the track is
Detections | None

unknown or all frames in its window are empty.
Source code in src/supervision/detection/tools/smoother.py 
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
def get_track(self, track_id: int) -> Detections | None:
    """Return the smoothed `Detections` for a single track.

    Averages `xyxy` over all valid (non-`None`) frames in the track window.
    `confidence` is averaged only over frames that carry it; frames with
    `confidence=None` are excluded. Returns `None` when the track is unknown
    or its entire window is empty.

    Args:
        track_id: The tracker ID whose smoothed detection to retrieve.

    Returns:
        Smoothed `Detections` for the track, or `None` if the track is
        unknown or all frames in its window are empty.
    """
    track = self.tracks.get(track_id, None)
    if track is None:
        return None

    valid: list[Detections] = [d for d in track if d is not None]
    if len(valid) == 0:
        return None

    ret = deepcopy(valid[0])
    ret.xyxy = np.mean(np.stack([d.xyxy for d in valid], axis=0), axis=0)
    # Average confidence only over frames that carry it; frames with
    # confidence=None contribute nothing to the mean. Retain None when
    # no frame in the window carries confidence.
    confidences = [d.confidence for d in valid if d.confidence is not None]
    ret.confidence = np.mean(np.array(confidences), axis=0) if confidences else None

    return ret

reset() -> None

Clears the per-track detection history so the smoother can be reused across independent streams without carrying over frames from a previous stream. The configured window length is preserved.

Examples:

>>> import numpy as np
>>> import supervision as sv
>>> smoother = sv.DetectionsSmoother()
>>> detections = sv.Detections(
...     xyxy=np.array([[0, 0, 10, 10]]),
...     confidence=np.array([0.5]),
...     tracker_id=np.array([1])
... )
>>> _ = smoother.update_with_detections(detections)
>>> len(smoother.tracks)
1
>>> smoother.reset()
>>> len(smoother.tracks)
0
Source code in src/supervision/detection/tools/smoother.py 
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
def reset(self) -> None:
    """
    Clears the per-track detection history so the smoother can be reused
    across independent streams without carrying over frames from a
    previous stream. The configured window `length` is preserved.

    Examples:
        ```pycon
        >>> import numpy as np
        >>> import supervision as sv
        >>> smoother = sv.DetectionsSmoother()
        >>> detections = sv.Detections(
        ...     xyxy=np.array([[0, 0, 10, 10]]),
        ...     confidence=np.array([0.5]),
        ...     tracker_id=np.array([1])
        ... )
        >>> _ = smoother.update_with_detections(detections)
        >>> len(smoother.tracks)
        1
        >>> smoother.reset()
        >>> len(smoother.tracks)
        0

        ```
    """
    self.tracks.clear()

update_with_detections(detections: Detections) -> Detections

Updates the smoother with a new set of detections from a frame.

Parameters:

Name Type Description Default
detections
Detections

The detections to add to the smoother.

 required
Source code in src/supervision/detection/tools/smoother.py 
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
def update_with_detections(self, detections: Detections) -> Detections:
    """
    Updates the smoother with a new set of detections from a frame.

    Args:
        detections: The detections to add to the smoother.
    """

    if detections.tracker_id is None:
        warnings.warn(
            "Smoothing skipped. DetectionsSmoother requires tracker_id. Refer to "
            "https://supervision.roboflow.com/latest/trackers for more "
            "information.",
            category=SupervisionWarnings,
        )
        return detections

    for detection_idx in range(len(detections)):
        tracker_id_value = detections.tracker_id[detection_idx]
        tracker_id = int(tracker_id_value)

        self.tracks[tracker_id].append(detections.select(detection_idx))

    for track_id in self.tracks.keys():
        if track_id not in detections.tracker_id:
            self.tracks[track_id].append(None)

    for track_id in list(self.tracks.keys()):
        if all([d is None for d in self.tracks[track_id]]):
            del self.tracks[track_id]

    current_track_ids = {int(track_id) for track_id in detections.tracker_id}
    return self.get_smoothed_detections(track_ids=current_track_ids)

Comments