tracklab.engine package

Submodules

tracklab.engine.batch module

tracklab.engine.engine module

class tracklab.engine.engine.TrackingEngine(modules: Pipeline, tracker_state: TrackerState, num_workers: int, callbacks: Dict[Callback] = None)[source]

Bases: ABC

Manages the full tracking pipeline.

After initializing the TrackingEngine, users should call track_dataset which will track each video in turn. The call stack looks like :

track_dataset | video_step | -> detect_multi_step -> detect_single_step -> reid_step -> track_step

Implementors of TrackingEngine will need to at least implement video_loop(). for example, an online engine will simply call each step in turn for every image in a video. An offline engine might instead call each step for all the images before doing the next step in the pipeline.

You should take care to implement the different callback hooks, by calling:

self.fabric.call("a_callback_function", *args, **kwargs)
Parameters:

  • detect_multi_model – The bbox/pose detection model

  • detect_single_model – The pose detection model

  • reid_model – Reid model

  • track_model – tracking model

  • tracker_state – contains inputs and outputs

  • callbacks – called at different steps

  • num_workers – number of workers for preprocessing

default_step(batch: Any, task: str, detections: DataFrame, image_pred: DataFrame, **kwargs)[source]
track_dataset()[source]

Run tracking on complete dataset.

abstract video_loop(tracker_state: TrackerState, video_metadata: Series, video_id: int) → DataFrame[source]

Run tracking on one video.

The pipeline for each video looks like :

detect_multi -> (detect_single) -> reid -> track

Parameters:

  • tracker_state (TrackerState) – tracker state object

  • video_metadata (pd.Series) – metadata for the video

  • video_id (int) – id of the video

Returns:

a dataframe of all detections

Return type:

detections

tracklab.engine.engine.merge_dataframes(main_df, appended_piece)[source]

tracklab.engine.offline module

class tracklab.engine.offline.OfflineTrackingEngine(modules: Pipeline, tracker_state: TrackerState, num_workers: int, callbacks: Dict[Callback] = None)[source]

Bases: TrackingEngine

video_loop(tracker_state, video, video_id)[source]

Run tracking on one video.

The pipeline for each video looks like :

detect_multi -> (detect_single) -> reid -> track

Parameters:

  • tracker_state (TrackerState) – tracker state object

  • video_metadata (pd.Series) – metadata for the video

  • video_id (int) – id of the video

Returns:

a dataframe of all detections

Return type:

detections

tracklab.engine.pipelined module

class tracklab.engine.pipelined.PipelinedTrackingEngine(**kwargs)[source]

Bases: TrackingEngine

Pipelined implementation of an online tracking engine.

video_loop(video, video_id) → DataFrame[source]

Run tracking on one video.

The pipeline for each video looks like :

detect_multi -> (detect_single) -> reid -> track

Parameters:

  • tracker_state (TrackerState) – tracker state object

  • video_metadata (pd.Series) – metadata for the video

  • video_id (int) – id of the video

Returns:

a dataframe of all detections

Return type:

detections