supervisely 6.73.284__py3-none-any.whl → 6.73.285__py3-none-any.whl

supervisely/nn/inference/tracking/mask_tracking.py

@@ -1,60 +1,401 @@
-import functools
-import json
 import time
+import uuid
 from queue import Queue
 from threading import Event, Thread
-from typing import Any, Dict, List, Optional, Union
+from typing import BinaryIO, Dict, List, Tuple
 
 import numpy as np
-from fastapi import BackgroundTasks, Form, Request, UploadFile
+from pydantic import ValidationError
 
-import supervisely as sly
-import supervisely.nn.inference.tracking.functional as F
+from supervisely._utils import find_value_by_keys
 from supervisely.annotation.label import Geometry, Label
-from supervisely.nn.inference import Inference
-from supervisely.nn.inference.tracking.tracker_interface import TrackerInterface
-from supervisely.nn.prediction_dto import PredictionSegmentation
+from supervisely.annotation.obj_class import ObjClass
+from supervisely.api.api import Api
+from supervisely.api.module_api import ApiField
+from supervisely.api.video.video_figure_api import FigureInfo
+from supervisely.geometry.bitmap import Bitmap
+from supervisely.geometry.helpers import deserialize_geometry
+from supervisely.geometry.polygon import Polygon
+from supervisely.imaging import image as sly_image
+from supervisely.nn.inference.tracking.base_tracking import BaseTracking
+from supervisely.nn.inference.tracking.tracker_interface import (
+    TrackerInterface,
+    TrackerInterfaceV2,
+)
+from supervisely.sly_logger import logger
+from supervisely.task.progress import Progress
+
+
+class MaskTracking(BaseTracking):
+    def _deserialize_geometry(self, data: dict):
+        geometry_type_str = data["type"]
+        geometry_json = data["data"]
+        return deserialize_geometry(geometry_type_str, geometry_json)
 
+    def _inference(self, frames: List[np.ndarray], geometries: List[Geometry]):
+        results = [[] for _ in range(len(frames) - 1)]
+        for geometry in geometries:
+            if not isinstance(geometry, Bitmap) and not isinstance(geometry, Polygon):
+                raise TypeError(f"This app does not support {geometry.geometry_name()} tracking")
 
-class MaskTracking(Inference):
-    def __init__(
-        self,
-        model_dir: Optional[str] = None,
-        custom_inference_settings: Optional[Union[Dict[str, Any], str]] = None,
-    ):
-        Inference.__init__(
-            self,
-            model_dir,
-            custom_inference_settings,
-            sliding_window_mode=None,
-            use_gui=False,
+            # combine several binary masks into one multilabel mask
+            i = 0
+            label2id = {}
+
+            original_geometry = geometry.clone()
+
+            # convert polygon to bitmap
+            if isinstance(geometry, Polygon):
+                polygon_obj_class = ObjClass("polygon", Polygon)
+                polygon_label = Label(geometry, polygon_obj_class)
+                bitmap_obj_class = ObjClass("bitmap", Bitmap)
+                bitmap_label = polygon_label.convert(bitmap_obj_class)[0]
+                geometry = bitmap_label.geometry
+            if i == 0:
+                multilabel_mask = geometry.data.astype(int)
+                multilabel_mask = np.zeros(frames[0].shape, dtype=np.uint8)
+                geometry.draw(bitmap=multilabel_mask, color=[1, 1, 1])
+                i += 1
+            else:
+                i += 1
+                geometry.draw(bitmap=multilabel_mask, color=[i, i, i])
+            label2id[i] = {
+                "original_geometry": original_geometry.geometry_name(),
+            }
+            # run tracker
+            tracked_multilabel_masks = self.predict(
+                frames=frames, input_mask=multilabel_mask[:, :, 0]
+            )
+            tracked_multilabel_masks = np.array(tracked_multilabel_masks)
+            # decompose multilabel masks into binary masks
+            for i in np.unique(tracked_multilabel_masks):
+                if i != 0:
+                    binary_masks = tracked_multilabel_masks == i
+                    geometry_type = label2id[i]["original_geometry"]
+                    for j, mask in enumerate(binary_masks[1:]):
+                        # check if mask is not empty
+                        if np.any(mask):
+                            if geometry_type == "polygon":
+                                bitmap_geometry = Bitmap(mask)
+                                bitmap_obj_class = ObjClass("bitmap", Bitmap)
+                                bitmap_label = Label(bitmap_geometry, bitmap_obj_class)
+                                polygon_obj_class = ObjClass("polygon", Polygon)
+                                polygon_labels = bitmap_label.convert(polygon_obj_class)
+                                geometries = [label.geometry for label in polygon_labels]
+                            else:
+                                geometries = [Bitmap(mask)]
+                            results[j].extend(
+                                [
+                                    {"type": geom.geometry_name(), "data": geom.to_json()}
+                                    for geom in geometries
+                                ]
+                            )
+        return results
+
+    def _track(self, api: Api, context: Dict):
+        self.video_interface = TrackerInterface(
+            context=context,
+            api=api,
+            load_all_frames=True,
+            notify_in_predict=True,
+            per_point_polygon_tracking=False,
+            frame_loader=self.cache.download_frame,
+            frames_loader=self.cache.download_frames,
         )
+        range_of_frames = [
+            self.video_interface.frames_indexes[0],
+            self.video_interface.frames_indexes[-1],
+        ]
+
+        if self.cache.is_persistent:
+            # if cache is persistent, run cache task for whole video
+            self.cache.run_cache_task_manually(
+                api,
+                None,
+                video_id=self.video_interface.video_id,
+            )
+        else:
+            # if cache is not persistent, run cache task for range of frames
+            self.cache.run_cache_task_manually(
+                api,
+                [range_of_frames],
+                video_id=self.video_interface.video_id,
+            )
+
+        api.logger.info("Starting tracking process")
+        # load frames
+        frames = self.video_interface.frames
+        # combine several binary masks into one multilabel mask
+        i = 0
+        label2id = {}
+
+        def _upload_loop(q: Queue, stop_event: Event, video_interface: TrackerInterface):
+            try:
+                while True:
+                    items = []
+                    while not q.empty():
+                        items.append(q.get_nowait())
+                    if len(items) > 0:
+                        video_interface.add_object_geometries_on_frames(*list(zip(*items)))
+                        continue
+                    if stop_event.is_set():
+                        video_interface._notify(True, task="stop tracking")
+                        return
+                    time.sleep(1)
+            except Exception as e:
+                api.logger.error("Error in upload loop: %s", str(e), exc_info=True)
+                video_interface._notify(True, task="stop tracking")
+                video_interface.global_stop_indicatior = True
+                raise
+
+        upload_queue = Queue()
+        stop_upload_event = Event()
+        Thread(
+            target=_upload_loop,
+            args=[upload_queue, stop_upload_event, self.video_interface],
+            daemon=True,
+        ).start()
 
         try:
-            self.load_on_device(model_dir, "cuda")
-        except RuntimeError:
-            self.load_on_device(model_dir, "cpu")
-            sly.logger.warn("Failed to load model on CUDA device.")
-
-        sly.logger.debug(
-            "Smart cache params",
-            extra={
-                "ttl": sly.env.smart_cache_ttl(),
-                "maxsize": sly.env.smart_cache_size(),
-            },
-        )
+            for (fig_id, geometry), obj_id in zip(
+                self.video_interface.geometries.items(),
+                self.video_interface.object_ids,
+            ):
+                original_geometry = geometry.clone()
+                if not isinstance(geometry, Bitmap) and not isinstance(geometry, Polygon):
+                    stop_upload_event.set()
+                    raise TypeError(
+                        f"This app does not support {geometry.geometry_name()} tracking"
+                    )
+                # convert polygon to bitmap
+                if isinstance(geometry, Polygon):
+                    polygon_obj_class = ObjClass("polygon", Polygon)
+                    polygon_label = Label(geometry, polygon_obj_class)
+                    bitmap_obj_class = ObjClass("bitmap", Bitmap)
+                    bitmap_label = polygon_label.convert(bitmap_obj_class)[0]
+                    geometry = bitmap_label.geometry
+                if i == 0:
+                    multilabel_mask = geometry.data.astype(int)
+                    multilabel_mask = np.zeros(frames[0].shape, dtype=np.uint8)
+                    geometry.draw(bitmap=multilabel_mask, color=[1, 1, 1])
+                    i += 1
+                else:
+                    i += 1
+                    geometry.draw(bitmap=multilabel_mask, color=[i, i, i])
+                label2id[i] = {
+                    "fig_id": fig_id,
+                    "obj_id": obj_id,
+                    "original_geometry": original_geometry.geometry_name(),
+                }
+            # run tracker
+            tracked_multilabel_masks = self.predict(
+                frames=frames, input_mask=multilabel_mask[:, :, 0]
+            )
+            tracked_multilabel_masks = np.array(tracked_multilabel_masks)
+            # decompose multilabel masks into binary masks
+            for i in np.unique(tracked_multilabel_masks):
+                if i != 0:
+                    binary_masks = tracked_multilabel_masks == i
+                    fig_id = label2id[i]["fig_id"]
+                    obj_id = label2id[i]["obj_id"]
+                    geometry_type = label2id[i]["original_geometry"]
+                    for j, mask in enumerate(binary_masks[1:]):
+                        # check if mask is not empty
+                        if not np.any(mask):
+                            api.logger.info(
+                                f"Skipping empty mask on frame {self.video_interface.frame_index + j + 1}"
+                            )
+                            # update progress bar anyway (otherwise it will not be finished)
+                            self.video_interface._notify(task="add geometry on frame")
+                        else:
+                            if geometry_type == "polygon":
+                                bitmap_geometry = Bitmap(mask)
+                                bitmap_obj_class = ObjClass("bitmap", Bitmap)
+                                bitmap_label = Label(bitmap_geometry, bitmap_obj_class)
+                                polygon_obj_class = ObjClass("polygon", Polygon)
+                                polygon_labels = bitmap_label.convert(polygon_obj_class)
+                                geometries = [label.geometry for label in polygon_labels]
+                            else:
+                                geometries = [Bitmap(mask)]
+                            for l, geometry in enumerate(geometries):
+                                if l == len(geometries) - 1:
+                                    notify = True
+                                else:
+                                    notify = False
+                                upload_queue.put(
+                                    (
+                                        geometry,
+                                        obj_id,
+                                        self.video_interface.frames_indexes[j + 1],
+                                        notify,
+                                    )
+                                )
+                    if self.video_interface.global_stop_indicatior:
+                        stop_upload_event.set()
+                        return
+                    api.logger.info(f"Figure with id {fig_id} was successfully tracked")
+        except Exception:
+            stop_upload_event.set()
+            raise
+        stop_upload_event.set()
+
+    def _track_async(self, api: Api, context: dict, inference_request_uuid: str = None):
+        inference_request = self._inference_requests[inference_request_uuid]
+        tracker_interface = TrackerInterfaceV2(api, context, self.cache)
+        progress: Progress = inference_request["progress"]
+        frames_count = tracker_interface.frames_count
+        figures = tracker_interface.figures
+        progress_total = frames_count * len(figures)
+        progress.total = progress_total
+
+        def _upload_f(items: List[Tuple[FigureInfo, bool]]):
+            with inference_request["lock"]:
+                inference_request["pending_results"].extend([item[0] for item in items])
+
+        def _notify_f(items: List[Tuple[FigureInfo, bool]]):
+            items_by_object_id: Dict[int, List[Tuple[FigureInfo, bool]]] = {}
+            for item in items:
+                items_by_object_id.setdefault(item[0].object_id, []).append(item)
+
+            for object_id, object_items in items_by_object_id.items():
+                frame_range = [
+                    min(item[0].frame_index for item in object_items),
+                    max(item[0].frame_index for item in object_items),
+                ]
+                progress.iters_done_report(sum(1 for item in object_items if item[1]))
+                tracker_interface.notify_progress(progress.current, progress.total, frame_range)
 
-    def _deserialize_geometry(self, data: dict):
-        geometry_type_str = data["type"]
-        geometry_json = data["data"]
-        return sly.deserialize_geometry(geometry_type_str, geometry_json)
+        # run tracker
+        frame_index = tracker_interface.frame_index
+        direction_n = tracker_interface.direction_n
+        api.logger.info("Start tracking.")
+        try:
+            with tracker_interface(_upload_f, _notify_f):
+                # combine several binary masks into one multilabel mask
+                i = 0
+                label2id = {}
+                # load frames
+                frames = tracker_interface.load_all_frames()
+                frames = [frame.image for frame in frames]
+                for figure in figures:
+                    figure = api.video.figure._convert_json_info(figure)
+                    fig_id = figure.id
+                    obj_id = figure.object_id
+                    geometry = deserialize_geometry(figure.geometry_type, figure.geometry)
+                    original_geometry = geometry.clone()
+                    if not isinstance(geometry, (Bitmap, Polygon)):
+                        raise TypeError(
+                            f"This app does not support {geometry.geometry_name()} tracking"
+                        )
+                    # convert polygon to bitmap
+                    if isinstance(geometry, Polygon):
+                        polygon_obj_class = ObjClass("polygon", Polygon)
+                        polygon_label = Label(geometry, polygon_obj_class)
+                        bitmap_obj_class = ObjClass("bitmap", Bitmap)
+                        bitmap_label = polygon_label.convert(bitmap_obj_class)[0]
+                        geometry = bitmap_label.geometry
+                    if i == 0:
+                        multilabel_mask = geometry.data.astype(int)
+                        multilabel_mask = np.zeros(frames[0].shape, dtype=np.uint8)
+                        geometry.draw(bitmap=multilabel_mask, color=[1, 1, 1])
+                        i += 1
+                    else:
+                        i += 1
+                        geometry.draw(bitmap=multilabel_mask, color=[i, i, i])
+                    label2id[i] = {
+                        "fig_id": fig_id,
+                        "obj_id": obj_id,
+                        "original_geometry": original_geometry.geometry_name(),
+                    }
+                    if inference_request["cancel_inference"]:
+                        return
+                    if tracker_interface.is_stopped():
+                        reason = tracker_interface.stop_reason()
+                        if isinstance(reason, Exception):
+                            raise reason
+                        return
+
+                # predict
+                tracked_multilabel_masks = self.predict(
+                    frames=frames, input_mask=multilabel_mask[:, :, 0]
+                )
+                tracked_multilabel_masks = np.array(tracked_multilabel_masks)
 
-    def get_info(self):
-        info = super().get_info()
-        info["task type"] = "tracking"
-        return info
+                # decompose multilabel masks into binary masks
+                for i in np.unique(tracked_multilabel_masks):
+                    if inference_request["cancel_inference"]:
+                        return
+                    if tracker_interface.is_stopped():
+                        reason = tracker_interface.stop_reason()
+                        if isinstance(reason, Exception):
+                            raise reason
+                        return
+                    if i != 0:
+                        binary_masks = tracked_multilabel_masks == i
+                        fig_id = label2id[i]["fig_id"]
+                        obj_id = label2id[i]["obj_id"]
+                        geometry_type = label2id[i]["original_geometry"]
+                        for j, mask in enumerate(binary_masks[1:], 1):
+                            if inference_request["cancel_inference"]:
+                                return
+                            if tracker_interface.is_stopped():
+                                reason = tracker_interface.stop_reason()
+                                if isinstance(reason, Exception):
+                                    raise reason
+                                return
+                            this_figure_index = frame_index + j * direction_n
+                            # check if mask is not empty
+                            if not np.any(mask):
+                                api.logger.info(f"Skipping empty mask on frame {this_figure_index}")
+                                # update progress bar anyway (otherwise it will not be finished)
+                                progress.iter_done_report()
+                            else:
+                                if geometry_type == "polygon":
+                                    bitmap_geometry = Bitmap(mask)
+                                    bitmap_obj_class = ObjClass("bitmap", Bitmap)
+                                    bitmap_label = Label(bitmap_geometry, bitmap_obj_class)
+                                    polygon_obj_class = ObjClass("polygon", Polygon)
+                                    polygon_labels = bitmap_label.convert(polygon_obj_class)
+                                    geometries = [label.geometry for label in polygon_labels]
+                                else:
+                                    geometries = [Bitmap(mask)]
+                                for l, geometry in enumerate(geometries):
+                                    figure_id = uuid.uuid5(
+                                        namespace=uuid.NAMESPACE_URL, name=f"{time.time()}"
+                                    ).hex
+                                    result_figure = api.video.figure._convert_json_info(
+                                        {
+                                            ApiField.ID: figure_id,
+                                            ApiField.OBJECT_ID: obj_id,
+                                            "meta": {"frame": this_figure_index},
+                                            ApiField.GEOMETRY_TYPE: geometry.geometry_name(),
+                                            ApiField.GEOMETRY: geometry.to_json(),
+                                            ApiField.TRACK_ID: tracker_interface.track_id,
+                                        }
+                                    )
+                                    should_notify = l == len(geometries) - 1
+                                    tracker_interface.add_prediction((result_figure, should_notify))
+                        api.logger.info(
+                            "Figure [%d, %d] tracked.",
+                            i,
+                            len(figures),
+                            extra={"figure_id": figure.id},
+                        )
+        except Exception:
+            progress.message = "Error occured during tracking"
+            raise
+        else:
+            progress.message = "Ready"
+        finally:
+            progress.set(current=0, total=1, report=True)
+
+    # Implement the following methods in the derived class
+    def track(self, api: Api, state: Dict, context: Dict):
+        fn = self.send_error_data(api, context)(self._track)
+        self.schedule_task(fn, api, context)
+        return {"message": "Tracking has started."}
 
-    def _track_api(self, api: sly.Api, context: dict):
+    def track_api(self, api: Api, state: Dict, context: Dict):
         # unused fields:
         context["trackId"] = "auto"
         context["objectIds"] = []
@@ -104,13 +445,13 @@ class MaskTracking(Inference):
 
         for input_geom in input_geometries:
             geometry = self._deserialize_geometry(input_geom)
-            if not isinstance(geometry, sly.Bitmap) and not isinstance(geometry, sly.Polygon):
+            if not isinstance(geometry, Bitmap) and not isinstance(geometry, Polygon):
                 raise TypeError(f"This app does not support {geometry.geometry_name()} tracking")
             # convert polygon to bitmap
-            if isinstance(geometry, sly.Polygon):
-                polygon_obj_class = sly.ObjClass("polygon", sly.Polygon)
-                polygon_label = sly.Label(geometry, polygon_obj_class)
-                bitmap_obj_class = sly.ObjClass("bitmap", sly.Bitmap)
+            if isinstance(geometry, Polygon):
+                polygon_obj_class = ObjClass("polygon", Polygon)
+                polygon_label = Label(geometry, polygon_obj_class)
+                bitmap_obj_class = ObjClass("bitmap", Bitmap)
                 bitmap_label = polygon_label.convert(bitmap_obj_class)[0]
                 geometry = bitmap_label.geometry
             if i == 0:
@@ -143,7 +484,7 @@ class MaskTracking(Inference):
                         predictions_for_label.append(None)
                     else:
                         # frame_idx = j + 1
-                        geometry = sly.Bitmap(mask)
+                        geometry = Bitmap(mask)
                         predictions_for_label.append(
                             {"type": geometry.geometry_name(), "data": geometry.to_json()}
                         )
@@ -153,295 +494,45 @@ class MaskTracking(Inference):
         predictions = list(map(list, zip(*predictions)))
         return predictions
 
-    def _inference(self, frames: List[np.ndarray], geometries: List[Geometry]):
-        results = [[] for _ in range(len(frames) - 1)]
-        for geometry in geometries:
-            if not isinstance(geometry, sly.Bitmap) and not isinstance(geometry, sly.Polygon):
-                raise TypeError(f"This app does not support {geometry.geometry_name()} tracking")
-
-            # combine several binary masks into one multilabel mask
-            i = 0
-            label2id = {}
-
-            original_geometry = geometry.clone()
-
-            # convert polygon to bitmap
-            if isinstance(geometry, sly.Polygon):
-                polygon_obj_class = sly.ObjClass("polygon", sly.Polygon)
-                polygon_label = sly.Label(geometry, polygon_obj_class)
-                bitmap_obj_class = sly.ObjClass("bitmap", sly.Bitmap)
-                bitmap_label = polygon_label.convert(bitmap_obj_class)[0]
-                geometry = bitmap_label.geometry
-            if i == 0:
-                multilabel_mask = geometry.data.astype(int)
-                multilabel_mask = np.zeros(frames[0].shape, dtype=np.uint8)
-                geometry.draw(bitmap=multilabel_mask, color=[1, 1, 1])
-                i += 1
-            else:
-                i += 1
-                geometry.draw(bitmap=multilabel_mask, color=[i, i, i])
-            label2id[i] = {
-                "original_geometry": original_geometry.geometry_name(),
-            }
-            # run tracker
-            tracked_multilabel_masks = self.predict(
-                frames=frames, input_mask=multilabel_mask[:, :, 0]
-            )
-            tracked_multilabel_masks = np.array(tracked_multilabel_masks)
-            # decompose multilabel masks into binary masks
-            for i in np.unique(tracked_multilabel_masks):
-                if i != 0:
-                    binary_masks = tracked_multilabel_masks == i
-                    geometry_type = label2id[i]["original_geometry"]
-                    for j, mask in enumerate(binary_masks[1:]):
-                        # check if mask is not empty
-                        if np.any(mask):
-                            if geometry_type == "polygon":
-                                bitmap_geometry = sly.Bitmap(mask)
-                                bitmap_obj_class = sly.ObjClass("bitmap", sly.Bitmap)
-                                bitmap_label = sly.Label(bitmap_geometry, bitmap_obj_class)
-                                polygon_obj_class = sly.ObjClass("polygon", sly.Polygon)
-                                polygon_labels = bitmap_label.convert(polygon_obj_class)
-                                geometries = [label.geometry for label in polygon_labels]
-                            else:
-                                geometries = [sly.Bitmap(mask)]
-                            results[j].extend(
-                                [
-                                    {"type": geom.geometry_name(), "data": geom.to_json()}
-                                    for geom in geometries
-                                ]
-                            )
-        return results
-
-    def _track_api_cached(self, request: Request, context: dict):
-        sly.logger.info(f"Start tracking with settings: {context}.")
-        video_id = context["video_id"]
-        frame_indexes = list(
-            range(context["frame_index"], context["frame_index"] + context["frames"] + 1)
-        )
-        geometries = map(self._deserialize_geometry, context["input_geometries"])
-        frames = self.cache.get_frames_from_cache(video_id, frame_indexes)
-        return self._inference(frames, geometries)
-
-    def _track_api_files(
-        self, request: Request, files: List[UploadFile], settings: str = Form("{}")
+    def track_api_files(
+        self,
+        files: List[BinaryIO],
+        settings: Dict,
     ):
-        state = json.loads(settings)
-        sly.logger.info(f"Start tracking with settings: {state}.")
-        video_id = state["video_id"]
+        logger.info(f"Start tracking with settings:", extra={"settings": settings})
+        frame_index = find_value_by_keys(settings, ["frame_index", "frameIndex"])
+        frames_count = find_value_by_keys(settings, ["frames", "framesCount"])
+        input_geometries = find_value_by_keys(settings, ["input_geometries", "inputGeometries"])
+        direction = settings.get("direction", "forward")
+        direction_n = 1 if direction == "forward" else -1
         frame_indexes = list(
-            range(state["frame_index"], state["frame_index"] + state["frames"] + 1)
+            range(frame_index, frame_index + frames_count * direction_n + direction_n, direction_n)
         )
-        geometries = map(self._deserialize_geometry, state["input_geometries"])
+        geometries = map(self._deserialize_geometry, input_geometries)
         frames = []
         for file, frame_idx in zip(files, frame_indexes):
-            img_bytes = file.file.read()
-            frame = sly.image.read_bytes(img_bytes)
+            img_bytes = file.read()
+            frame = sly_image.read_bytes(img_bytes)
             frames.append(frame)
-        sly.logger.info("Start tracking.")
+        logger.info("Start tracking.")
         return self._inference(frames, geometries)
 
-    def serve(self):
-        super().serve()
-        server = self._app.get_server()
-        self.cache.add_cache_endpoint(server)
-        self.cache.add_cache_files_endpoint(server)
-
-        @server.post("/track")
-        def start_track(request: Request, task: BackgroundTasks):
-            task.add_task(track, request)
-            return {"message": "Track task started."}
-
-        @server.post("/track-api")
-        def track_api(request: Request):
-            return self._track_api(request.state.api, request.state.context)
-
-        @server.post("/track-api-files")
-        def track_api_files(
-            request: Request,
-            files: List[UploadFile],
-            settings: str = Form("{}"),
-        ):
-            return self._track_api_files(request, files, settings)
-
-        def send_error_data(func):
-            @functools.wraps(func)
-            def wrapper(*args, **kwargs):
-                value = None
-                try:
-                    value = func(*args, **kwargs)
-                except Exception as exc:
-                    # print("An error occured:")
-                    # print(traceback.format_exc())
-                    request: Request = args[0]
-                    context = request.state.context
-                    api: sly.Api = request.state.api
-                    track_id = context["trackId"]
-                    api.logger.error(f"An error occured: {repr(exc)}")
-
-                    api.post(
-                        "videos.notify-annotation-tool",
-                        data={
-                            "type": "videos:tracking-error",
-                            "data": {
-                                "trackId": track_id,
-                                "error": {"message": repr(exc)},
-                            },
-                        },
-                    )
-                return value
-
-            return wrapper
-
-        @send_error_data
-        def track(request: Request):
-            context = request.state.context
-            api: sly.Api = request.state.api
-
-            self.video_interface = TrackerInterface(
-                context=context,
-                api=api,
-                load_all_frames=True,
-                notify_in_predict=True,
-                per_point_polygon_tracking=False,
-                frame_loader=self.cache.download_frame,
-                frames_loader=self.cache.download_frames,
+    def track_async(self, api: Api, state: Dict, context: Dict):
+        batch_size = context.get("batch_size", self.get_batch_size())
+        if self.max_batch_size is not None and batch_size > self.max_batch_size:
+            raise ValidationError(
+                f"Batch size should be less than or equal to {self.max_batch_size} for this model."
             )
-            range_of_frames = [
-                self.video_interface.frames_indexes[0],
-                self.video_interface.frames_indexes[-1],
-            ]
-
-            if self.cache.is_persistent:
-                # if cache is persistent, run cache task for whole video
-                self.cache.run_cache_task_manually(
-                    api,
-                    None,
-                    video_id=self.video_interface.video_id,
-                )
-            else:
-                # if cache is not persistent, run cache task for range of frames
-                self.cache.run_cache_task_manually(
-                    api,
-                    [range_of_frames],
-                    video_id=self.video_interface.video_id,
-                )
 
-            api.logger.info("Starting tracking process")
-            # load frames
-            frames = self.video_interface.frames
-            # combine several binary masks into one multilabel mask
-            i = 0
-            label2id = {}
+        inference_request_uuid = uuid.uuid5(namespace=uuid.NAMESPACE_URL, name=f"{time.time()}").hex
+        fn = self.send_error_data(api, context)(self._track_async)
+        self.schedule_task(fn, api, context, inference_request_uuid=inference_request_uuid)
 
-            def _upload_loop(q: Queue, stop_event: Event, video_interface: TrackerInterface):
-                try:
-                    while True:
-                        items = []
-                        while not q.empty():
-                            items.append(q.get_nowait())
-                        if len(items) > 0:
-                            video_interface.add_object_geometries_on_frames(*list(zip(*items)))
-                            continue
-                        if stop_event.is_set():
-                            video_interface._notify(True, task="stop tracking")
-                            return
-                        time.sleep(1)
-                except Exception as e:
-                    api.logger.error("Error in upload loop: %s", str(e), exc_info=True)
-                    video_interface._notify(True, task="stop tracking")
-                    video_interface.global_stop_indicatior = True
-                    raise
-
-            upload_queue = Queue()
-            stop_upload_event = Event()
-            Thread(
-                target=_upload_loop,
-                args=[upload_queue, stop_upload_event, self.video_interface],
-                daemon=True,
-            ).start()
-
-            try:
-                for (fig_id, geometry), obj_id in zip(
-                    self.video_interface.geometries.items(),
-                    self.video_interface.object_ids,
-                ):
-                    original_geometry = geometry.clone()
-                    if not isinstance(geometry, sly.Bitmap) and not isinstance(
-                        geometry, sly.Polygon
-                    ):
-                        stop_upload_event.set()
-                        raise TypeError(
-                            f"This app does not support {geometry.geometry_name()} tracking"
-                        )
-                    # convert polygon to bitmap
-                    if isinstance(geometry, sly.Polygon):
-                        polygon_obj_class = sly.ObjClass("polygon", sly.Polygon)
-                        polygon_label = sly.Label(geometry, polygon_obj_class)
-                        bitmap_obj_class = sly.ObjClass("bitmap", sly.Bitmap)
-                        bitmap_label = polygon_label.convert(bitmap_obj_class)[0]
-                        geometry = bitmap_label.geometry
-                    if i == 0:
-                        multilabel_mask = geometry.data.astype(int)
-                        multilabel_mask = np.zeros(frames[0].shape, dtype=np.uint8)
-                        geometry.draw(bitmap=multilabel_mask, color=[1, 1, 1])
-                        i += 1
-                    else:
-                        i += 1
-                        geometry.draw(bitmap=multilabel_mask, color=[i, i, i])
-                    label2id[i] = {
-                        "fig_id": fig_id,
-                        "obj_id": obj_id,
-                        "original_geometry": original_geometry.geometry_name(),
-                    }
-                # run tracker
-                tracked_multilabel_masks = self.predict(
-                    frames=frames, input_mask=multilabel_mask[:, :, 0]
-                )
-                tracked_multilabel_masks = np.array(tracked_multilabel_masks)
-                # decompose multilabel masks into binary masks
-                for i in np.unique(tracked_multilabel_masks):
-                    if i != 0:
-                        binary_masks = tracked_multilabel_masks == i
-                        fig_id = label2id[i]["fig_id"]
-                        obj_id = label2id[i]["obj_id"]
-                        geometry_type = label2id[i]["original_geometry"]
-                        for j, mask in enumerate(binary_masks[1:]):
-                            # check if mask is not empty
-                            if not np.any(mask):
-                                api.logger.info(
-                                    f"Skipping empty mask on frame {self.video_interface.frame_index + j + 1}"
-                                )
-                                # update progress bar anyway (otherwise it will not be finished)
-                                self.video_interface._notify(task="add geometry on frame")
-                            else:
-                                if geometry_type == "polygon":
-                                    bitmap_geometry = sly.Bitmap(mask)
-                                    bitmap_obj_class = sly.ObjClass("bitmap", sly.Bitmap)
-                                    bitmap_label = sly.Label(bitmap_geometry, bitmap_obj_class)
-                                    polygon_obj_class = sly.ObjClass("polygon", sly.Polygon)
-                                    polygon_labels = bitmap_label.convert(polygon_obj_class)
-                                    geometries = [label.geometry for label in polygon_labels]
-                                else:
-                                    geometries = [sly.Bitmap(mask)]
-                                for l, geometry in enumerate(geometries):
-                                    if l == len(geometries) - 1:
-                                        notify = True
-                                    else:
-                                        notify = False
-                                    upload_queue.put(
-                                        (
-                                            geometry,
-                                            obj_id,
-                                            self.video_interface.frames_indexes[j + 1],
-                                            notify,
-                                        )
-                                    )
-                        if self.video_interface.global_stop_indicatior:
-                            stop_upload_event.set()
-                            return
-                        api.logger.info(f"Figure with id {fig_id} was successfully tracked")
-            except Exception:
-                stop_upload_event.set()
-                raise
-            stop_upload_event.set()
+        logger.debug(
+            "Inference has scheduled from 'track_async' endpoint",
+            extra={"inference_request_uuid": inference_request_uuid},
+        )
+        return {
+            "message": "Inference has started.",
+            "inference_request_uuid": inference_request_uuid,
+        }